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

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

Version: ~ [ linux-5.3-rc4 ] ~ [ linux-5.2.8 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.66 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.138 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.189 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.189 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.71 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
  3  *              operating system.  INET is implemented using the  BSD Socket
  4  *              interface as the means of communication with the user level.
  5  *
  6  *              Implementation of the Transmission Control Protocol(TCP).
  7  *
  8  * Authors:     Ross Biro
  9  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
 10  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
 11  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
 12  *              Florian La Roche, <flla@stud.uni-sb.de>
 13  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
 14  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
 15  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
 16  *              Matthew Dillon, <dillon@apollo.west.oic.com>
 17  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
 18  *              Jorge Cwik, <jorge@laser.satlink.net>
 19  */
 20 
 21 /*
 22  * Changes:     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 two TSO segments */
 54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
 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 int sysctl_tcp_mtu_probing __read_mostly = 0;
 63 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
 64 
 65 /* By default, RFC2861 behavior.  */
 66 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
 67 
 68 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
 69                            int push_one, gfp_t gfp);
 70 
 71 /* Account for new data that has been sent to the network. */
 72 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
 73 {
 74         struct inet_connection_sock *icsk = inet_csk(sk);
 75         struct tcp_sock *tp = tcp_sk(sk);
 76         unsigned int prior_packets = tp->packets_out;
 77 
 78         tcp_advance_send_head(sk, skb);
 79         tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
 80 
 81         tp->packets_out += tcp_skb_pcount(skb);
 82         if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 83             icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
 84                 tcp_rearm_rto(sk);
 85         }
 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 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
139 {
140         struct tcp_sock *tp = tcp_sk(sk);
141         s32 delta = tcp_time_stamp - tp->lsndtime;
142         u32 restart_cwnd = tcp_init_cwnd(tp, dst);
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         const struct dst_entry *dst = __sk_dst_get(sk);
164 
165         if (sysctl_tcp_slow_start_after_idle &&
166             (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
167                 tcp_cwnd_restart(sk, __sk_dst_get(sk));
168 
169         tp->lsndtime = now;
170 
171         /* If it is a reply for ato after last received
172          * packet, enter pingpong mode.
173          */
174         if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
175             (!dst || !dst_metric(dst, RTAX_QUICKACK)))
176                         icsk->icsk_ack.pingpong = 1;
177 }
178 
179 /* Account for an ACK we sent. */
180 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
181 {
182         tcp_dec_quickack_mode(sk, pkts);
183         inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
184 }
185 
186 
187 u32 tcp_default_init_rwnd(u32 mss)
188 {
189         /* Initial receive window should be twice of TCP_INIT_CWND to
190          * enable proper sending of new unsent data during fast recovery
191          * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
192          * limit when mss is larger than 1460.
193          */
194         u32 init_rwnd = TCP_INIT_CWND * 2;
195 
196         if (mss > 1460)
197                 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
198         return init_rwnd;
199 }
200 
201 /* Determine a window scaling and initial window to offer.
202  * Based on the assumption that the given amount of space
203  * will be offered. Store the results in the tp structure.
204  * NOTE: for smooth operation initial space offering should
205  * be a multiple of mss if possible. We assume here that mss >= 1.
206  * This MUST be enforced by all callers.
207  */
208 void tcp_select_initial_window(int __space, __u32 mss,
209                                __u32 *rcv_wnd, __u32 *window_clamp,
210                                int wscale_ok, __u8 *rcv_wscale,
211                                __u32 init_rcv_wnd)
212 {
213         unsigned int space = (__space < 0 ? 0 : __space);
214 
215         /* If no clamp set the clamp to the max possible scaled window */
216         if (*window_clamp == 0)
217                 (*window_clamp) = (65535 << 14);
218         space = min(*window_clamp, space);
219 
220         /* Quantize space offering to a multiple of mss if possible. */
221         if (space > mss)
222                 space = (space / mss) * mss;
223 
224         /* NOTE: offering an initial window larger than 32767
225          * will break some buggy TCP stacks. If the admin tells us
226          * it is likely we could be speaking with such a buggy stack
227          * we will truncate our initial window offering to 32K-1
228          * unless the remote has sent us a window scaling option,
229          * which we interpret as a sign the remote TCP is not
230          * misinterpreting the window field as a signed quantity.
231          */
232         if (sysctl_tcp_workaround_signed_windows)
233                 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
234         else
235                 (*rcv_wnd) = space;
236 
237         (*rcv_wscale) = 0;
238         if (wscale_ok) {
239                 /* Set window scaling on max possible window
240                  * See RFC1323 for an explanation of the limit to 14
241                  */
242                 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
243                 space = min_t(u32, space, *window_clamp);
244                 while (space > 65535 && (*rcv_wscale) < 14) {
245                         space >>= 1;
246                         (*rcv_wscale)++;
247                 }
248         }
249 
250         if (mss > (1 << *rcv_wscale)) {
251                 if (!init_rcv_wnd) /* Use default unless specified otherwise */
252                         init_rcv_wnd = tcp_default_init_rwnd(mss);
253                 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
254         }
255 
256         /* Set the clamp no higher than max representable value */
257         (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
258 }
259 EXPORT_SYMBOL(tcp_select_initial_window);
260 
261 /* Chose a new window to advertise, update state in tcp_sock for the
262  * socket, and return result with RFC1323 scaling applied.  The return
263  * value can be stuffed directly into th->window for an outgoing
264  * frame.
265  */
266 static u16 tcp_select_window(struct sock *sk)
267 {
268         struct tcp_sock *tp = tcp_sk(sk);
269         u32 cur_win = tcp_receive_window(tp);
270         u32 new_win = __tcp_select_window(sk);
271 
272         /* Never shrink the offered window */
273         if (new_win < cur_win) {
274                 /* Danger Will Robinson!
275                  * Don't update rcv_wup/rcv_wnd here or else
276                  * we will not be able to advertise a zero
277                  * window in time.  --DaveM
278                  *
279                  * Relax Will Robinson.
280                  */
281                 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
282         }
283         tp->rcv_wnd = new_win;
284         tp->rcv_wup = tp->rcv_nxt;
285 
286         /* Make sure we do not exceed the maximum possible
287          * scaled window.
288          */
289         if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
290                 new_win = min(new_win, MAX_TCP_WINDOW);
291         else
292                 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
293 
294         /* RFC1323 scaling applied */
295         new_win >>= tp->rx_opt.rcv_wscale;
296 
297         /* If we advertise zero window, disable fast path. */
298         if (new_win == 0)
299                 tp->pred_flags = 0;
300 
301         return new_win;
302 }
303 
304 /* Packet ECN state for a SYN-ACK */
305 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
306 {
307         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
308         if (!(tp->ecn_flags & TCP_ECN_OK))
309                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
310 }
311 
312 /* Packet ECN state for a SYN.  */
313 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
314 {
315         struct tcp_sock *tp = tcp_sk(sk);
316 
317         tp->ecn_flags = 0;
318         if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
319                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
320                 tp->ecn_flags = TCP_ECN_OK;
321         }
322 }
323 
324 static __inline__ void
325 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
326 {
327         if (inet_rsk(req)->ecn_ok)
328                 th->ece = 1;
329 }
330 
331 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
332  * be sent.
333  */
334 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
335                                 int tcp_header_len)
336 {
337         struct tcp_sock *tp = tcp_sk(sk);
338 
339         if (tp->ecn_flags & TCP_ECN_OK) {
340                 /* Not-retransmitted data segment: set ECT and inject CWR. */
341                 if (skb->len != tcp_header_len &&
342                     !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
343                         INET_ECN_xmit(sk);
344                         if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
345                                 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
346                                 tcp_hdr(skb)->cwr = 1;
347                                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
348                         }
349                 } else {
350                         /* ACK or retransmitted segment: clear ECT|CE */
351                         INET_ECN_dontxmit(sk);
352                 }
353                 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
354                         tcp_hdr(skb)->ece = 1;
355         }
356 }
357 
358 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
359  * auto increment end seqno.
360  */
361 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
362 {
363         skb->ip_summed = CHECKSUM_PARTIAL;
364         skb->csum = 0;
365 
366         TCP_SKB_CB(skb)->tcp_flags = flags;
367         TCP_SKB_CB(skb)->sacked = 0;
368 
369         skb_shinfo(skb)->gso_segs = 1;
370         skb_shinfo(skb)->gso_size = 0;
371         skb_shinfo(skb)->gso_type = 0;
372 
373         TCP_SKB_CB(skb)->seq = seq;
374         if (flags & (TCPHDR_SYN | TCPHDR_FIN))
375                 seq++;
376         TCP_SKB_CB(skb)->end_seq = seq;
377 }
378 
379 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
380 {
381         return tp->snd_una != tp->snd_up;
382 }
383 
384 #define OPTION_SACK_ADVERTISE   (1 << 0)
385 #define OPTION_TS               (1 << 1)
386 #define OPTION_MD5              (1 << 2)
387 #define OPTION_WSCALE           (1 << 3)
388 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
389 
390 struct tcp_out_options {
391         u16 options;            /* bit field of OPTION_* */
392         u16 mss;                /* 0 to disable */
393         u8 ws;                  /* window scale, 0 to disable */
394         u8 num_sack_blocks;     /* number of SACK blocks to include */
395         u8 hash_size;           /* bytes in hash_location */
396         __u8 *hash_location;    /* temporary pointer, overloaded */
397         __u32 tsval, tsecr;     /* need to include OPTION_TS */
398         struct tcp_fastopen_cookie *fastopen_cookie;    /* Fast open cookie */
399 };
400 
401 /* Write previously computed TCP options to the packet.
402  *
403  * Beware: Something in the Internet is very sensitive to the ordering of
404  * TCP options, we learned this through the hard way, so be careful here.
405  * Luckily we can at least blame others for their non-compliance but from
406  * inter-operatibility perspective it seems that we're somewhat stuck with
407  * the ordering which we have been using if we want to keep working with
408  * those broken things (not that it currently hurts anybody as there isn't
409  * particular reason why the ordering would need to be changed).
410  *
411  * At least SACK_PERM as the first option is known to lead to a disaster
412  * (but it may well be that other scenarios fail similarly).
413  */
414 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
415                               struct tcp_out_options *opts)
416 {
417         u16 options = opts->options;    /* mungable copy */
418 
419         if (unlikely(OPTION_MD5 & options)) {
420                 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
421                                (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
422                 /* overload cookie hash location */
423                 opts->hash_location = (__u8 *)ptr;
424                 ptr += 4;
425         }
426 
427         if (unlikely(opts->mss)) {
428                 *ptr++ = htonl((TCPOPT_MSS << 24) |
429                                (TCPOLEN_MSS << 16) |
430                                opts->mss);
431         }
432 
433         if (likely(OPTION_TS & options)) {
434                 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
435                         *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
436                                        (TCPOLEN_SACK_PERM << 16) |
437                                        (TCPOPT_TIMESTAMP << 8) |
438                                        TCPOLEN_TIMESTAMP);
439                         options &= ~OPTION_SACK_ADVERTISE;
440                 } else {
441                         *ptr++ = htonl((TCPOPT_NOP << 24) |
442                                        (TCPOPT_NOP << 16) |
443                                        (TCPOPT_TIMESTAMP << 8) |
444                                        TCPOLEN_TIMESTAMP);
445                 }
446                 *ptr++ = htonl(opts->tsval);
447                 *ptr++ = htonl(opts->tsecr);
448         }
449 
450         if (unlikely(OPTION_SACK_ADVERTISE & options)) {
451                 *ptr++ = htonl((TCPOPT_NOP << 24) |
452                                (TCPOPT_NOP << 16) |
453                                (TCPOPT_SACK_PERM << 8) |
454                                TCPOLEN_SACK_PERM);
455         }
456 
457         if (unlikely(OPTION_WSCALE & options)) {
458                 *ptr++ = htonl((TCPOPT_NOP << 24) |
459                                (TCPOPT_WINDOW << 16) |
460                                (TCPOLEN_WINDOW << 8) |
461                                opts->ws);
462         }
463 
464         if (unlikely(opts->num_sack_blocks)) {
465                 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
466                         tp->duplicate_sack : tp->selective_acks;
467                 int this_sack;
468 
469                 *ptr++ = htonl((TCPOPT_NOP  << 24) |
470                                (TCPOPT_NOP  << 16) |
471                                (TCPOPT_SACK <<  8) |
472                                (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
473                                                      TCPOLEN_SACK_PERBLOCK)));
474 
475                 for (this_sack = 0; this_sack < opts->num_sack_blocks;
476                      ++this_sack) {
477                         *ptr++ = htonl(sp[this_sack].start_seq);
478                         *ptr++ = htonl(sp[this_sack].end_seq);
479                 }
480 
481                 tp->rx_opt.dsack = 0;
482         }
483 
484         if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
485                 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
486 
487                 *ptr++ = htonl((TCPOPT_EXP << 24) |
488                                ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
489                                TCPOPT_FASTOPEN_MAGIC);
490 
491                 memcpy(ptr, foc->val, foc->len);
492                 if ((foc->len & 3) == 2) {
493                         u8 *align = ((u8 *)ptr) + foc->len;
494                         align[0] = align[1] = TCPOPT_NOP;
495                 }
496                 ptr += (foc->len + 3) >> 2;
497         }
498 }
499 
500 /* Compute TCP options for SYN packets. This is not the final
501  * network wire format yet.
502  */
503 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
504                                 struct tcp_out_options *opts,
505                                 struct tcp_md5sig_key **md5)
506 {
507         struct tcp_sock *tp = tcp_sk(sk);
508         unsigned int remaining = MAX_TCP_OPTION_SPACE;
509         struct tcp_fastopen_request *fastopen = tp->fastopen_req;
510 
511 #ifdef CONFIG_TCP_MD5SIG
512         *md5 = tp->af_specific->md5_lookup(sk, sk);
513         if (*md5) {
514                 opts->options |= OPTION_MD5;
515                 remaining -= TCPOLEN_MD5SIG_ALIGNED;
516         }
517 #else
518         *md5 = NULL;
519 #endif
520 
521         /* We always get an MSS option.  The option bytes which will be seen in
522          * normal data packets should timestamps be used, must be in the MSS
523          * advertised.  But we subtract them from tp->mss_cache so that
524          * calculations in tcp_sendmsg are simpler etc.  So account for this
525          * fact here if necessary.  If we don't do this correctly, as a
526          * receiver we won't recognize data packets as being full sized when we
527          * should, and thus we won't abide by the delayed ACK rules correctly.
528          * SACKs don't matter, we never delay an ACK when we have any of those
529          * going out.  */
530         opts->mss = tcp_advertise_mss(sk);
531         remaining -= TCPOLEN_MSS_ALIGNED;
532 
533         if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
534                 opts->options |= OPTION_TS;
535                 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
536                 opts->tsecr = tp->rx_opt.ts_recent;
537                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
538         }
539         if (likely(sysctl_tcp_window_scaling)) {
540                 opts->ws = tp->rx_opt.rcv_wscale;
541                 opts->options |= OPTION_WSCALE;
542                 remaining -= TCPOLEN_WSCALE_ALIGNED;
543         }
544         if (likely(sysctl_tcp_sack)) {
545                 opts->options |= OPTION_SACK_ADVERTISE;
546                 if (unlikely(!(OPTION_TS & opts->options)))
547                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
548         }
549 
550         if (fastopen && fastopen->cookie.len >= 0) {
551                 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
552                 need = (need + 3) & ~3U;  /* Align to 32 bits */
553                 if (remaining >= need) {
554                         opts->options |= OPTION_FAST_OPEN_COOKIE;
555                         opts->fastopen_cookie = &fastopen->cookie;
556                         remaining -= need;
557                         tp->syn_fastopen = 1;
558                 }
559         }
560 
561         return MAX_TCP_OPTION_SPACE - remaining;
562 }
563 
564 /* Set up TCP options for SYN-ACKs. */
565 static unsigned int tcp_synack_options(struct sock *sk,
566                                    struct request_sock *req,
567                                    unsigned int mss, struct sk_buff *skb,
568                                    struct tcp_out_options *opts,
569                                    struct tcp_md5sig_key **md5,
570                                    struct tcp_fastopen_cookie *foc)
571 {
572         struct inet_request_sock *ireq = inet_rsk(req);
573         unsigned int remaining = MAX_TCP_OPTION_SPACE;
574 
575 #ifdef CONFIG_TCP_MD5SIG
576         *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
577         if (*md5) {
578                 opts->options |= OPTION_MD5;
579                 remaining -= TCPOLEN_MD5SIG_ALIGNED;
580 
581                 /* We can't fit any SACK blocks in a packet with MD5 + TS
582                  * options. There was discussion about disabling SACK
583                  * rather than TS in order to fit in better with old,
584                  * buggy kernels, but that was deemed to be unnecessary.
585                  */
586                 ireq->tstamp_ok &= !ireq->sack_ok;
587         }
588 #else
589         *md5 = NULL;
590 #endif
591 
592         /* We always send an MSS option. */
593         opts->mss = mss;
594         remaining -= TCPOLEN_MSS_ALIGNED;
595 
596         if (likely(ireq->wscale_ok)) {
597                 opts->ws = ireq->rcv_wscale;
598                 opts->options |= OPTION_WSCALE;
599                 remaining -= TCPOLEN_WSCALE_ALIGNED;
600         }
601         if (likely(ireq->tstamp_ok)) {
602                 opts->options |= OPTION_TS;
603                 opts->tsval = TCP_SKB_CB(skb)->when;
604                 opts->tsecr = req->ts_recent;
605                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
606         }
607         if (likely(ireq->sack_ok)) {
608                 opts->options |= OPTION_SACK_ADVERTISE;
609                 if (unlikely(!ireq->tstamp_ok))
610                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
611         }
612         if (foc != NULL) {
613                 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
614                 need = (need + 3) & ~3U;  /* Align to 32 bits */
615                 if (remaining >= need) {
616                         opts->options |= OPTION_FAST_OPEN_COOKIE;
617                         opts->fastopen_cookie = foc;
618                         remaining -= need;
619                 }
620         }
621 
622         return MAX_TCP_OPTION_SPACE - remaining;
623 }
624 
625 /* Compute TCP options for ESTABLISHED sockets. This is not the
626  * final wire format yet.
627  */
628 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
629                                         struct tcp_out_options *opts,
630                                         struct tcp_md5sig_key **md5)
631 {
632         struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
633         struct tcp_sock *tp = tcp_sk(sk);
634         unsigned int size = 0;
635         unsigned int eff_sacks;
636 
637 #ifdef CONFIG_TCP_MD5SIG
638         *md5 = tp->af_specific->md5_lookup(sk, sk);
639         if (unlikely(*md5)) {
640                 opts->options |= OPTION_MD5;
641                 size += TCPOLEN_MD5SIG_ALIGNED;
642         }
643 #else
644         *md5 = NULL;
645 #endif
646 
647         if (likely(tp->rx_opt.tstamp_ok)) {
648                 opts->options |= OPTION_TS;
649                 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
650                 opts->tsecr = tp->rx_opt.ts_recent;
651                 size += TCPOLEN_TSTAMP_ALIGNED;
652         }
653 
654         eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
655         if (unlikely(eff_sacks)) {
656                 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
657                 opts->num_sack_blocks =
658                         min_t(unsigned int, eff_sacks,
659                               (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
660                               TCPOLEN_SACK_PERBLOCK);
661                 size += TCPOLEN_SACK_BASE_ALIGNED +
662                         opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
663         }
664 
665         return size;
666 }
667 
668 
669 /* TCP SMALL QUEUES (TSQ)
670  *
671  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
672  * to reduce RTT and bufferbloat.
673  * We do this using a special skb destructor (tcp_wfree).
674  *
675  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
676  * needs to be reallocated in a driver.
677  * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
678  *
679  * Since transmit from skb destructor is forbidden, we use a tasklet
680  * to process all sockets that eventually need to send more skbs.
681  * We use one tasklet per cpu, with its own queue of sockets.
682  */
683 struct tsq_tasklet {
684         struct tasklet_struct   tasklet;
685         struct list_head        head; /* queue of tcp sockets */
686 };
687 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
688 
689 static void tcp_tsq_handler(struct sock *sk)
690 {
691         if ((1 << sk->sk_state) &
692             (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
693              TCPF_CLOSE_WAIT  | TCPF_LAST_ACK))
694                 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
695 }
696 /*
697  * One tasklest per cpu tries to send more skbs.
698  * We run in tasklet context but need to disable irqs when
699  * transfering tsq->head because tcp_wfree() might
700  * interrupt us (non NAPI drivers)
701  */
702 static void tcp_tasklet_func(unsigned long data)
703 {
704         struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
705         LIST_HEAD(list);
706         unsigned long flags;
707         struct list_head *q, *n;
708         struct tcp_sock *tp;
709         struct sock *sk;
710 
711         local_irq_save(flags);
712         list_splice_init(&tsq->head, &list);
713         local_irq_restore(flags);
714 
715         list_for_each_safe(q, n, &list) {
716                 tp = list_entry(q, struct tcp_sock, tsq_node);
717                 list_del(&tp->tsq_node);
718 
719                 sk = (struct sock *)tp;
720                 bh_lock_sock(sk);
721 
722                 if (!sock_owned_by_user(sk)) {
723                         tcp_tsq_handler(sk);
724                 } else {
725                         /* defer the work to tcp_release_cb() */
726                         set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
727                 }
728                 bh_unlock_sock(sk);
729 
730                 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
731                 sk_free(sk);
732         }
733 }
734 
735 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) |           \
736                           (1UL << TCP_WRITE_TIMER_DEFERRED) |   \
737                           (1UL << TCP_DELACK_TIMER_DEFERRED) |  \
738                           (1UL << TCP_MTU_REDUCED_DEFERRED))
739 /**
740  * tcp_release_cb - tcp release_sock() callback
741  * @sk: socket
742  *
743  * called from release_sock() to perform protocol dependent
744  * actions before socket release.
745  */
746 void tcp_release_cb(struct sock *sk)
747 {
748         struct tcp_sock *tp = tcp_sk(sk);
749         unsigned long flags, nflags;
750 
751         /* perform an atomic operation only if at least one flag is set */
752         do {
753                 flags = tp->tsq_flags;
754                 if (!(flags & TCP_DEFERRED_ALL))
755                         return;
756                 nflags = flags & ~TCP_DEFERRED_ALL;
757         } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
758 
759         if (flags & (1UL << TCP_TSQ_DEFERRED))
760                 tcp_tsq_handler(sk);
761 
762         if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
763                 tcp_write_timer_handler(sk);
764                 __sock_put(sk);
765         }
766         if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
767                 tcp_delack_timer_handler(sk);
768                 __sock_put(sk);
769         }
770         if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
771                 sk->sk_prot->mtu_reduced(sk);
772                 __sock_put(sk);
773         }
774 }
775 EXPORT_SYMBOL(tcp_release_cb);
776 
777 void __init tcp_tasklet_init(void)
778 {
779         int i;
780 
781         for_each_possible_cpu(i) {
782                 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
783 
784                 INIT_LIST_HEAD(&tsq->head);
785                 tasklet_init(&tsq->tasklet,
786                              tcp_tasklet_func,
787                              (unsigned long)tsq);
788         }
789 }
790 
791 /*
792  * Write buffer destructor automatically called from kfree_skb.
793  * We cant xmit new skbs from this context, as we might already
794  * hold qdisc lock.
795  */
796 void tcp_wfree(struct sk_buff *skb)
797 {
798         struct sock *sk = skb->sk;
799         struct tcp_sock *tp = tcp_sk(sk);
800 
801         if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
802             !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
803                 unsigned long flags;
804                 struct tsq_tasklet *tsq;
805 
806                 /* Keep a ref on socket.
807                  * This last ref will be released in tcp_tasklet_func()
808                  */
809                 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
810 
811                 /* queue this socket to tasklet queue */
812                 local_irq_save(flags);
813                 tsq = &__get_cpu_var(tsq_tasklet);
814                 list_add(&tp->tsq_node, &tsq->head);
815                 tasklet_schedule(&tsq->tasklet);
816                 local_irq_restore(flags);
817         } else {
818                 sock_wfree(skb);
819         }
820 }
821 
822 /* This routine actually transmits TCP packets queued in by
823  * tcp_do_sendmsg().  This is used by both the initial
824  * transmission and possible later retransmissions.
825  * All SKB's seen here are completely headerless.  It is our
826  * job to build the TCP header, and pass the packet down to
827  * IP so it can do the same plus pass the packet off to the
828  * device.
829  *
830  * We are working here with either a clone of the original
831  * SKB, or a fresh unique copy made by the retransmit engine.
832  */
833 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
834                             gfp_t gfp_mask)
835 {
836         const struct inet_connection_sock *icsk = inet_csk(sk);
837         struct inet_sock *inet;
838         struct tcp_sock *tp;
839         struct tcp_skb_cb *tcb;
840         struct tcp_out_options opts;
841         unsigned int tcp_options_size, tcp_header_size;
842         struct tcp_md5sig_key *md5;
843         struct tcphdr *th;
844         int err;
845 
846         BUG_ON(!skb || !tcp_skb_pcount(skb));
847 
848         /* If congestion control is doing timestamping, we must
849          * take such a timestamp before we potentially clone/copy.
850          */
851         if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
852                 __net_timestamp(skb);
853 
854         if (likely(clone_it)) {
855                 const struct sk_buff *fclone = skb + 1;
856 
857                 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
858                              fclone->fclone == SKB_FCLONE_CLONE))
859                         NET_INC_STATS_BH(sock_net(sk),
860                                          LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
861 
862                 if (unlikely(skb_cloned(skb)))
863                         skb = pskb_copy(skb, gfp_mask);
864                 else
865                         skb = skb_clone(skb, gfp_mask);
866                 if (unlikely(!skb))
867                         return -ENOBUFS;
868         }
869 
870         inet = inet_sk(sk);
871         tp = tcp_sk(sk);
872         tcb = TCP_SKB_CB(skb);
873         memset(&opts, 0, sizeof(opts));
874 
875         if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
876                 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
877         else
878                 tcp_options_size = tcp_established_options(sk, skb, &opts,
879                                                            &md5);
880         tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
881 
882         if (tcp_packets_in_flight(tp) == 0)
883                 tcp_ca_event(sk, CA_EVENT_TX_START);
884 
885         /* if no packet is in qdisc/device queue, then allow XPS to select
886          * another queue.
887          */
888         skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
889 
890         skb_push(skb, tcp_header_size);
891         skb_reset_transport_header(skb);
892 
893         skb_orphan(skb);
894         skb->sk = sk;
895         skb->destructor = tcp_wfree;
896         atomic_add(skb->truesize, &sk->sk_wmem_alloc);
897 
898         /* Build TCP header and checksum it. */
899         th = tcp_hdr(skb);
900         th->source              = inet->inet_sport;
901         th->dest                = inet->inet_dport;
902         th->seq                 = htonl(tcb->seq);
903         th->ack_seq             = htonl(tp->rcv_nxt);
904         *(((__be16 *)th) + 6)   = htons(((tcp_header_size >> 2) << 12) |
905                                         tcb->tcp_flags);
906 
907         if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
908                 /* RFC1323: The window in SYN & SYN/ACK segments
909                  * is never scaled.
910                  */
911                 th->window      = htons(min(tp->rcv_wnd, 65535U));
912         } else {
913                 th->window      = htons(tcp_select_window(sk));
914         }
915         th->check               = 0;
916         th->urg_ptr             = 0;
917 
918         /* The urg_mode check is necessary during a below snd_una win probe */
919         if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
920                 if (before(tp->snd_up, tcb->seq + 0x10000)) {
921                         th->urg_ptr = htons(tp->snd_up - tcb->seq);
922                         th->urg = 1;
923                 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
924                         th->urg_ptr = htons(0xFFFF);
925                         th->urg = 1;
926                 }
927         }
928 
929         tcp_options_write((__be32 *)(th + 1), tp, &opts);
930         if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
931                 TCP_ECN_send(sk, skb, tcp_header_size);
932 
933 #ifdef CONFIG_TCP_MD5SIG
934         /* Calculate the MD5 hash, as we have all we need now */
935         if (md5) {
936                 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
937                 tp->af_specific->calc_md5_hash(opts.hash_location,
938                                                md5, sk, NULL, skb);
939         }
940 #endif
941 
942         icsk->icsk_af_ops->send_check(sk, skb);
943 
944         if (likely(tcb->tcp_flags & TCPHDR_ACK))
945                 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
946 
947         if (skb->len != tcp_header_size)
948                 tcp_event_data_sent(tp, sk);
949 
950         if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
951                 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
952                               tcp_skb_pcount(skb));
953 
954         err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
955         if (likely(err <= 0))
956                 return err;
957 
958         tcp_enter_cwr(sk, 1);
959 
960         return net_xmit_eval(err);
961 }
962 
963 /* This routine just queues the buffer for sending.
964  *
965  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
966  * otherwise socket can stall.
967  */
968 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
969 {
970         struct tcp_sock *tp = tcp_sk(sk);
971 
972         /* Advance write_seq and place onto the write_queue. */
973         tp->write_seq = TCP_SKB_CB(skb)->end_seq;
974         skb_header_release(skb);
975         tcp_add_write_queue_tail(sk, skb);
976         sk->sk_wmem_queued += skb->truesize;
977         sk_mem_charge(sk, skb->truesize);
978 }
979 
980 /* Initialize TSO segments for a packet. */
981 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
982                                  unsigned int mss_now)
983 {
984         /* Make sure we own this skb before messing gso_size/gso_segs */
985         WARN_ON_ONCE(skb_cloned(skb));
986 
987         if (skb->len <= mss_now || !sk_can_gso(sk) ||
988             skb->ip_summed == CHECKSUM_NONE) {
989                 /* Avoid the costly divide in the normal
990                  * non-TSO case.
991                  */
992                 skb_shinfo(skb)->gso_segs = 1;
993                 skb_shinfo(skb)->gso_size = 0;
994                 skb_shinfo(skb)->gso_type = 0;
995         } else {
996                 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
997                 skb_shinfo(skb)->gso_size = mss_now;
998                 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
999         }
1000 }
1001 
1002 /* When a modification to fackets out becomes necessary, we need to check
1003  * skb is counted to fackets_out or not.
1004  */
1005 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1006                                    int decr)
1007 {
1008         struct tcp_sock *tp = tcp_sk(sk);
1009 
1010         if (!tp->sacked_out || tcp_is_reno(tp))
1011                 return;
1012 
1013         if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1014                 tp->fackets_out -= decr;
1015 }
1016 
1017 /* Pcount in the middle of the write queue got changed, we need to do various
1018  * tweaks to fix counters
1019  */
1020 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1021 {
1022         struct tcp_sock *tp = tcp_sk(sk);
1023 
1024         tp->packets_out -= decr;
1025 
1026         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1027                 tp->sacked_out -= decr;
1028         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1029                 tp->retrans_out -= decr;
1030         if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1031                 tp->lost_out -= decr;
1032 
1033         /* Reno case is special. Sigh... */
1034         if (tcp_is_reno(tp) && decr > 0)
1035                 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1036 
1037         tcp_adjust_fackets_out(sk, skb, decr);
1038 
1039         if (tp->lost_skb_hint &&
1040             before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1041             (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1042                 tp->lost_cnt_hint -= decr;
1043 
1044         tcp_verify_left_out(tp);
1045 }
1046 
1047 /* Function to create two new TCP segments.  Shrinks the given segment
1048  * to the specified size and appends a new segment with the rest of the
1049  * packet to the list.  This won't be called frequently, I hope.
1050  * Remember, these are still headerless SKBs at this point.
1051  */
1052 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1053                  unsigned int mss_now)
1054 {
1055         struct tcp_sock *tp = tcp_sk(sk);
1056         struct sk_buff *buff;
1057         int nsize, old_factor;
1058         int nlen;
1059         u8 flags;
1060 
1061         if (WARN_ON(len > skb->len))
1062                 return -EINVAL;
1063 
1064         nsize = skb_headlen(skb) - len;
1065         if (nsize < 0)
1066                 nsize = 0;
1067 
1068         if (skb_unclone(skb, GFP_ATOMIC))
1069                 return -ENOMEM;
1070 
1071         /* Get a new skb... force flag on. */
1072         buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1073         if (buff == NULL)
1074                 return -ENOMEM; /* We'll just try again later. */
1075 
1076         sk->sk_wmem_queued += buff->truesize;
1077         sk_mem_charge(sk, buff->truesize);
1078         nlen = skb->len - len - nsize;
1079         buff->truesize += nlen;
1080         skb->truesize -= nlen;
1081 
1082         /* Correct the sequence numbers. */
1083         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1084         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1085         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1086 
1087         /* PSH and FIN should only be set in the second packet. */
1088         flags = TCP_SKB_CB(skb)->tcp_flags;
1089         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1090         TCP_SKB_CB(buff)->tcp_flags = flags;
1091         TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1092 
1093         if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1094                 /* Copy and checksum data tail into the new buffer. */
1095                 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1096                                                        skb_put(buff, nsize),
1097                                                        nsize, 0);
1098 
1099                 skb_trim(skb, len);
1100 
1101                 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1102         } else {
1103                 skb->ip_summed = CHECKSUM_PARTIAL;
1104                 skb_split(skb, buff, len);
1105         }
1106 
1107         buff->ip_summed = skb->ip_summed;
1108 
1109         /* Looks stupid, but our code really uses when of
1110          * skbs, which it never sent before. --ANK
1111          */
1112         TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1113         buff->tstamp = skb->tstamp;
1114 
1115         old_factor = tcp_skb_pcount(skb);
1116 
1117         /* Fix up tso_factor for both original and new SKB.  */
1118         tcp_set_skb_tso_segs(sk, skb, mss_now);
1119         tcp_set_skb_tso_segs(sk, buff, mss_now);
1120 
1121         /* If this packet has been sent out already, we must
1122          * adjust the various packet counters.
1123          */
1124         if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1125                 int diff = old_factor - tcp_skb_pcount(skb) -
1126                         tcp_skb_pcount(buff);
1127 
1128                 if (diff)
1129                         tcp_adjust_pcount(sk, skb, diff);
1130         }
1131 
1132         /* Link BUFF into the send queue. */
1133         skb_header_release(buff);
1134         tcp_insert_write_queue_after(skb, buff, sk);
1135 
1136         return 0;
1137 }
1138 
1139 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1140  * eventually). The difference is that pulled data not copied, but
1141  * immediately discarded.
1142  */
1143 static void __pskb_trim_head(struct sk_buff *skb, int len)
1144 {
1145         int i, k, eat;
1146 
1147         eat = min_t(int, len, skb_headlen(skb));
1148         if (eat) {
1149                 __skb_pull(skb, eat);
1150                 len -= eat;
1151                 if (!len)
1152                         return;
1153         }
1154         eat = len;
1155         k = 0;
1156         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1157                 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1158 
1159                 if (size <= eat) {
1160                         skb_frag_unref(skb, i);
1161                         eat -= size;
1162                 } else {
1163                         skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1164                         if (eat) {
1165                                 skb_shinfo(skb)->frags[k].page_offset += eat;
1166                                 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1167                                 eat = 0;
1168                         }
1169                         k++;
1170                 }
1171         }
1172         skb_shinfo(skb)->nr_frags = k;
1173 
1174         skb_reset_tail_pointer(skb);
1175         skb->data_len -= len;
1176         skb->len = skb->data_len;
1177 }
1178 
1179 /* Remove acked data from a packet in the transmit queue. */
1180 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1181 {
1182         if (skb_unclone(skb, GFP_ATOMIC))
1183                 return -ENOMEM;
1184 
1185         __pskb_trim_head(skb, len);
1186 
1187         TCP_SKB_CB(skb)->seq += len;
1188         skb->ip_summed = CHECKSUM_PARTIAL;
1189 
1190         skb->truesize        -= len;
1191         sk->sk_wmem_queued   -= len;
1192         sk_mem_uncharge(sk, len);
1193         sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1194 
1195         /* Any change of skb->len requires recalculation of tso factor. */
1196         if (tcp_skb_pcount(skb) > 1)
1197                 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1198 
1199         return 0;
1200 }
1201 
1202 /* Calculate MSS not accounting any TCP options.  */
1203 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1204 {
1205         const struct tcp_sock *tp = tcp_sk(sk);
1206         const struct inet_connection_sock *icsk = inet_csk(sk);
1207         int mss_now;
1208 
1209         /* Calculate base mss without TCP options:
1210            It is MMS_S - sizeof(tcphdr) of rfc1122
1211          */
1212         mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1213 
1214         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1215         if (icsk->icsk_af_ops->net_frag_header_len) {
1216                 const struct dst_entry *dst = __sk_dst_get(sk);
1217 
1218                 if (dst && dst_allfrag(dst))
1219                         mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1220         }
1221 
1222         /* Clamp it (mss_clamp does not include tcp options) */
1223         if (mss_now > tp->rx_opt.mss_clamp)
1224                 mss_now = tp->rx_opt.mss_clamp;
1225 
1226         /* Now subtract optional transport overhead */
1227         mss_now -= icsk->icsk_ext_hdr_len;
1228 
1229         /* Then reserve room for full set of TCP options and 8 bytes of data */
1230         if (mss_now < 48)
1231                 mss_now = 48;
1232         return mss_now;
1233 }
1234 
1235 /* Calculate MSS. Not accounting for SACKs here.  */
1236 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1237 {
1238         /* Subtract TCP options size, not including SACKs */
1239         return __tcp_mtu_to_mss(sk, pmtu) -
1240                (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1241 }
1242 
1243 /* Inverse of above */
1244 int tcp_mss_to_mtu(struct sock *sk, int mss)
1245 {
1246         const struct tcp_sock *tp = tcp_sk(sk);
1247         const struct inet_connection_sock *icsk = inet_csk(sk);
1248         int mtu;
1249 
1250         mtu = mss +
1251               tp->tcp_header_len +
1252               icsk->icsk_ext_hdr_len +
1253               icsk->icsk_af_ops->net_header_len;
1254 
1255         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1256         if (icsk->icsk_af_ops->net_frag_header_len) {
1257                 const struct dst_entry *dst = __sk_dst_get(sk);
1258 
1259                 if (dst && dst_allfrag(dst))
1260                         mtu += icsk->icsk_af_ops->net_frag_header_len;
1261         }
1262         return mtu;
1263 }
1264 
1265 /* MTU probing init per socket */
1266 void tcp_mtup_init(struct sock *sk)
1267 {
1268         struct tcp_sock *tp = tcp_sk(sk);
1269         struct inet_connection_sock *icsk = inet_csk(sk);
1270 
1271         icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1272         icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1273                                icsk->icsk_af_ops->net_header_len;
1274         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1275         icsk->icsk_mtup.probe_size = 0;
1276 }
1277 EXPORT_SYMBOL(tcp_mtup_init);
1278 
1279 /* This function synchronize snd mss to current pmtu/exthdr set.
1280 
1281    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1282    for TCP options, but includes only bare TCP header.
1283 
1284    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1285    It is minimum of user_mss and mss received with SYN.
1286    It also does not include TCP options.
1287 
1288    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1289 
1290    tp->mss_cache is current effective sending mss, including
1291    all tcp options except for SACKs. It is evaluated,
1292    taking into account current pmtu, but never exceeds
1293    tp->rx_opt.mss_clamp.
1294 
1295    NOTE1. rfc1122 clearly states that advertised MSS
1296    DOES NOT include either tcp or ip options.
1297 
1298    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1299    are READ ONLY outside this function.         --ANK (980731)
1300  */
1301 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1302 {
1303         struct tcp_sock *tp = tcp_sk(sk);
1304         struct inet_connection_sock *icsk = inet_csk(sk);
1305         int mss_now;
1306 
1307         if (icsk->icsk_mtup.search_high > pmtu)
1308                 icsk->icsk_mtup.search_high = pmtu;
1309 
1310         mss_now = tcp_mtu_to_mss(sk, pmtu);
1311         mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1312 
1313         /* And store cached results */
1314         icsk->icsk_pmtu_cookie = pmtu;
1315         if (icsk->icsk_mtup.enabled)
1316                 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1317         tp->mss_cache = mss_now;
1318 
1319         return mss_now;
1320 }
1321 EXPORT_SYMBOL(tcp_sync_mss);
1322 
1323 /* Compute the current effective MSS, taking SACKs and IP options,
1324  * and even PMTU discovery events into account.
1325  */
1326 unsigned int tcp_current_mss(struct sock *sk)
1327 {
1328         const struct tcp_sock *tp = tcp_sk(sk);
1329         const struct dst_entry *dst = __sk_dst_get(sk);
1330         u32 mss_now;
1331         unsigned int header_len;
1332         struct tcp_out_options opts;
1333         struct tcp_md5sig_key *md5;
1334 
1335         mss_now = tp->mss_cache;
1336 
1337         if (dst) {
1338                 u32 mtu = dst_mtu(dst);
1339                 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1340                         mss_now = tcp_sync_mss(sk, mtu);
1341         }
1342 
1343         header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1344                      sizeof(struct tcphdr);
1345         /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1346          * some common options. If this is an odd packet (because we have SACK
1347          * blocks etc) then our calculated header_len will be different, and
1348          * we have to adjust mss_now correspondingly */
1349         if (header_len != tp->tcp_header_len) {
1350                 int delta = (int) header_len - tp->tcp_header_len;
1351                 mss_now -= delta;
1352         }
1353 
1354         return mss_now;
1355 }
1356 
1357 /* Congestion window validation. (RFC2861) */
1358 static void tcp_cwnd_validate(struct sock *sk)
1359 {
1360         struct tcp_sock *tp = tcp_sk(sk);
1361 
1362         if (tp->packets_out >= tp->snd_cwnd) {
1363                 /* Network is feed fully. */
1364                 tp->snd_cwnd_used = 0;
1365                 tp->snd_cwnd_stamp = tcp_time_stamp;
1366         } else {
1367                 /* Network starves. */
1368                 if (tp->packets_out > tp->snd_cwnd_used)
1369                         tp->snd_cwnd_used = tp->packets_out;
1370 
1371                 if (sysctl_tcp_slow_start_after_idle &&
1372                     (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1373                         tcp_cwnd_application_limited(sk);
1374         }
1375 }
1376 
1377 /* Returns the portion of skb which can be sent right away without
1378  * introducing MSS oddities to segment boundaries. In rare cases where
1379  * mss_now != mss_cache, we will request caller to create a small skb
1380  * per input skb which could be mostly avoided here (if desired).
1381  *
1382  * We explicitly want to create a request for splitting write queue tail
1383  * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1384  * thus all the complexity (cwnd_len is always MSS multiple which we
1385  * return whenever allowed by the other factors). Basically we need the
1386  * modulo only when the receiver window alone is the limiting factor or
1387  * when we would be allowed to send the split-due-to-Nagle skb fully.
1388  */
1389 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1390                                         unsigned int mss_now, unsigned int max_segs)
1391 {
1392         const struct tcp_sock *tp = tcp_sk(sk);
1393         u32 needed, window, max_len;
1394 
1395         window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1396         max_len = mss_now * max_segs;
1397 
1398         if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1399                 return max_len;
1400 
1401         needed = min(skb->len, window);
1402 
1403         if (max_len <= needed)
1404                 return max_len;
1405 
1406         return needed - needed % mss_now;
1407 }
1408 
1409 /* Can at least one segment of SKB be sent right now, according to the
1410  * congestion window rules?  If so, return how many segments are allowed.
1411  */
1412 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1413                                          const struct sk_buff *skb)
1414 {
1415         u32 in_flight, cwnd;
1416 
1417         /* Don't be strict about the congestion window for the final FIN.  */
1418         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1419             tcp_skb_pcount(skb) == 1)
1420                 return 1;
1421 
1422         in_flight = tcp_packets_in_flight(tp);
1423         cwnd = tp->snd_cwnd;
1424         if (in_flight < cwnd)
1425                 return (cwnd - in_flight);
1426 
1427         return 0;
1428 }
1429 
1430 /* Initialize TSO state of a skb.
1431  * This must be invoked the first time we consider transmitting
1432  * SKB onto the wire.
1433  */
1434 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1435                              unsigned int mss_now)
1436 {
1437         int tso_segs = tcp_skb_pcount(skb);
1438 
1439         if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1440                 tcp_set_skb_tso_segs(sk, skb, mss_now);
1441                 tso_segs = tcp_skb_pcount(skb);
1442         }
1443         return tso_segs;
1444 }
1445 
1446 /* Minshall's variant of the Nagle send check. */
1447 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1448 {
1449         return after(tp->snd_sml, tp->snd_una) &&
1450                 !after(tp->snd_sml, tp->snd_nxt);
1451 }
1452 
1453 /* Return false, if packet can be sent now without violation Nagle's rules:
1454  * 1. It is full sized.
1455  * 2. Or it contains FIN. (already checked by caller)
1456  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1457  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1458  *    With Minshall's modification: all sent small packets are ACKed.
1459  */
1460 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1461                                   const struct sk_buff *skb,
1462                                   unsigned int mss_now, int nonagle)
1463 {
1464         return skb->len < mss_now &&
1465                 ((nonagle & TCP_NAGLE_CORK) ||
1466                  (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1467 }
1468 
1469 /* Return true if the Nagle test allows this packet to be
1470  * sent now.
1471  */
1472 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1473                                   unsigned int cur_mss, int nonagle)
1474 {
1475         /* Nagle rule does not apply to frames, which sit in the middle of the
1476          * write_queue (they have no chances to get new data).
1477          *
1478          * This is implemented in the callers, where they modify the 'nonagle'
1479          * argument based upon the location of SKB in the send queue.
1480          */
1481         if (nonagle & TCP_NAGLE_PUSH)
1482                 return true;
1483 
1484         /* Don't use the nagle rule for urgent data (or for the final FIN). */
1485         if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1486                 return true;
1487 
1488         if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1489                 return true;
1490 
1491         return false;
1492 }
1493 
1494 /* Does at least the first segment of SKB fit into the send window? */
1495 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1496                              const struct sk_buff *skb,
1497                              unsigned int cur_mss)
1498 {
1499         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1500 
1501         if (skb->len > cur_mss)
1502                 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1503 
1504         return !after(end_seq, tcp_wnd_end(tp));
1505 }
1506 
1507 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1508  * should be put on the wire right now.  If so, it returns the number of
1509  * packets allowed by the congestion window.
1510  */
1511 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1512                                  unsigned int cur_mss, int nonagle)
1513 {
1514         const struct tcp_sock *tp = tcp_sk(sk);
1515         unsigned int cwnd_quota;
1516 
1517         tcp_init_tso_segs(sk, skb, cur_mss);
1518 
1519         if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1520                 return 0;
1521 
1522         cwnd_quota = tcp_cwnd_test(tp, skb);
1523         if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1524                 cwnd_quota = 0;
1525 
1526         return cwnd_quota;
1527 }
1528 
1529 /* Test if sending is allowed right now. */
1530 bool tcp_may_send_now(struct sock *sk)
1531 {
1532         const struct tcp_sock *tp = tcp_sk(sk);
1533         struct sk_buff *skb = tcp_send_head(sk);
1534 
1535         return skb &&
1536                 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1537                              (tcp_skb_is_last(sk, skb) ?
1538                               tp->nonagle : TCP_NAGLE_PUSH));
1539 }
1540 
1541 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1542  * which is put after SKB on the list.  It is very much like
1543  * tcp_fragment() except that it may make several kinds of assumptions
1544  * in order to speed up the splitting operation.  In particular, we
1545  * know that all the data is in scatter-gather pages, and that the
1546  * packet has never been sent out before (and thus is not cloned).
1547  */
1548 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1549                         unsigned int mss_now, gfp_t gfp)
1550 {
1551         struct sk_buff *buff;
1552         int nlen = skb->len - len;
1553         u8 flags;
1554 
1555         /* All of a TSO frame must be composed of paged data.  */
1556         if (skb->len != skb->data_len)
1557                 return tcp_fragment(sk, skb, len, mss_now);
1558 
1559         buff = sk_stream_alloc_skb(sk, 0, gfp);
1560         if (unlikely(buff == NULL))
1561                 return -ENOMEM;
1562 
1563         sk->sk_wmem_queued += buff->truesize;
1564         sk_mem_charge(sk, buff->truesize);
1565         buff->truesize += nlen;
1566         skb->truesize -= nlen;
1567 
1568         /* Correct the sequence numbers. */
1569         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1570         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1571         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1572 
1573         /* PSH and FIN should only be set in the second packet. */
1574         flags = TCP_SKB_CB(skb)->tcp_flags;
1575         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1576         TCP_SKB_CB(buff)->tcp_flags = flags;
1577 
1578         /* This packet was never sent out yet, so no SACK bits. */
1579         TCP_SKB_CB(buff)->sacked = 0;
1580 
1581         buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1582         skb_split(skb, buff, len);
1583 
1584         /* Fix up tso_factor for both original and new SKB.  */
1585         tcp_set_skb_tso_segs(sk, skb, mss_now);
1586         tcp_set_skb_tso_segs(sk, buff, mss_now);
1587 
1588         /* Link BUFF into the send queue. */
1589         skb_header_release(buff);
1590         tcp_insert_write_queue_after(skb, buff, sk);
1591 
1592         return 0;
1593 }
1594 
1595 /* Try to defer sending, if possible, in order to minimize the amount
1596  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1597  *
1598  * This algorithm is from John Heffner.
1599  */
1600 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1601 {
1602         struct tcp_sock *tp = tcp_sk(sk);
1603         const struct inet_connection_sock *icsk = inet_csk(sk);
1604         u32 send_win, cong_win, limit, in_flight;
1605         int win_divisor;
1606 
1607         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1608                 goto send_now;
1609 
1610         if (icsk->icsk_ca_state != TCP_CA_Open)
1611                 goto send_now;
1612 
1613         /* Defer for less than two clock ticks. */
1614         if (tp->tso_deferred &&
1615             (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1616                 goto send_now;
1617 
1618         in_flight = tcp_packets_in_flight(tp);
1619 
1620         BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1621 
1622         send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1623 
1624         /* From in_flight test above, we know that cwnd > in_flight.  */
1625         cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1626 
1627         limit = min(send_win, cong_win);
1628 
1629         /* If a full-sized TSO skb can be sent, do it. */
1630         if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1631                            tp->xmit_size_goal_segs * tp->mss_cache))
1632                 goto send_now;
1633 
1634         /* Middle in queue won't get any more data, full sendable already? */
1635         if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1636                 goto send_now;
1637 
1638         win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1639         if (win_divisor) {
1640                 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1641 
1642                 /* If at least some fraction of a window is available,
1643                  * just use it.
1644                  */
1645                 chunk /= win_divisor;
1646                 if (limit >= chunk)
1647                         goto send_now;
1648         } else {
1649                 /* Different approach, try not to defer past a single
1650                  * ACK.  Receiver should ACK every other full sized
1651                  * frame, so if we have space for more than 3 frames
1652                  * then send now.
1653                  */
1654                 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1655                         goto send_now;
1656         }
1657 
1658         /* Ok, it looks like it is advisable to defer.
1659          * Do not rearm the timer if already set to not break TCP ACK clocking.
1660          */
1661         if (!tp->tso_deferred)
1662                 tp->tso_deferred = 1 | (jiffies << 1);
1663 
1664         return true;
1665 
1666 send_now:
1667         tp->tso_deferred = 0;
1668         return false;
1669 }
1670 
1671 /* Create a new MTU probe if we are ready.
1672  * MTU probe is regularly attempting to increase the path MTU by
1673  * deliberately sending larger packets.  This discovers routing
1674  * changes resulting in larger path MTUs.
1675  *
1676  * Returns 0 if we should wait to probe (no cwnd available),
1677  *         1 if a probe was sent,
1678  *         -1 otherwise
1679  */
1680 static int tcp_mtu_probe(struct sock *sk)
1681 {
1682         struct tcp_sock *tp = tcp_sk(sk);
1683         struct inet_connection_sock *icsk = inet_csk(sk);
1684         struct sk_buff *skb, *nskb, *next;
1685         int len;
1686         int probe_size;
1687         int size_needed;
1688         int copy;
1689         int mss_now;
1690 
1691         /* Not currently probing/verifying,
1692          * not in recovery,
1693          * have enough cwnd, and
1694          * not SACKing (the variable headers throw things off) */
1695         if (!icsk->icsk_mtup.enabled ||
1696             icsk->icsk_mtup.probe_size ||
1697             inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1698             tp->snd_cwnd < 11 ||
1699             tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1700                 return -1;
1701 
1702         /* Very simple search strategy: just double the MSS. */
1703         mss_now = tcp_current_mss(sk);
1704         probe_size = 2 * tp->mss_cache;
1705         size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1706         if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1707                 /* TODO: set timer for probe_converge_event */
1708                 return -1;
1709         }
1710 
1711         /* Have enough data in the send queue to probe? */
1712         if (tp->write_seq - tp->snd_nxt < size_needed)
1713                 return -1;
1714 
1715         if (tp->snd_wnd < size_needed)
1716                 return -1;
1717         if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1718                 return 0;
1719 
1720         /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1721         if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1722                 if (!tcp_packets_in_flight(tp))
1723                         return -1;
1724                 else
1725                         return 0;
1726         }
1727 
1728         /* We're allowed to probe.  Build it now. */
1729         if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1730                 return -1;
1731         sk->sk_wmem_queued += nskb->truesize;
1732         sk_mem_charge(sk, nskb->truesize);
1733 
1734         skb = tcp_send_head(sk);
1735 
1736         TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1737         TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1738         TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1739         TCP_SKB_CB(nskb)->sacked = 0;
1740         nskb->csum = 0;
1741         nskb->ip_summed = skb->ip_summed;
1742 
1743         tcp_insert_write_queue_before(nskb, skb, sk);
1744 
1745         len = 0;
1746         tcp_for_write_queue_from_safe(skb, next, sk) {
1747                 copy = min_t(int, skb->len, probe_size - len);
1748                 if (nskb->ip_summed)
1749                         skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1750                 else
1751                         nskb->csum = skb_copy_and_csum_bits(skb, 0,
1752                                                             skb_put(nskb, copy),
1753                                                             copy, nskb->csum);
1754 
1755                 if (skb->len <= copy) {
1756                         /* We've eaten all the data from this skb.
1757                          * Throw it away. */
1758                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1759                         tcp_unlink_write_queue(skb, sk);
1760                         sk_wmem_free_skb(sk, skb);
1761                 } else {
1762                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1763                                                    ~(TCPHDR_FIN|TCPHDR_PSH);
1764                         if (!skb_shinfo(skb)->nr_frags) {
1765                                 skb_pull(skb, copy);
1766                                 if (skb->ip_summed != CHECKSUM_PARTIAL)
1767                                         skb->csum = csum_partial(skb->data,
1768                                                                  skb->len, 0);
1769                         } else {
1770                                 __pskb_trim_head(skb, copy);
1771                                 tcp_set_skb_tso_segs(sk, skb, mss_now);
1772                         }
1773                         TCP_SKB_CB(skb)->seq += copy;
1774                 }
1775 
1776                 len += copy;
1777 
1778                 if (len >= probe_size)
1779                         break;
1780         }
1781         tcp_init_tso_segs(sk, nskb, nskb->len);
1782 
1783         /* We're ready to send.  If this fails, the probe will
1784          * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1785         TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1786         if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1787                 /* Decrement cwnd here because we are sending
1788                  * effectively two packets. */
1789                 tp->snd_cwnd--;
1790                 tcp_event_new_data_sent(sk, nskb);
1791 
1792                 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1793                 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1794                 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1795 
1796                 return 1;
1797         }
1798 
1799         return -1;
1800 }
1801 
1802 /* This routine writes packets to the network.  It advances the
1803  * send_head.  This happens as incoming acks open up the remote
1804  * window for us.
1805  *
1806  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1807  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1808  * account rare use of URG, this is not a big flaw.
1809  *
1810  * Send at most one packet when push_one > 0. Temporarily ignore
1811  * cwnd limit to force at most one packet out when push_one == 2.
1812 
1813  * Returns true, if no segments are in flight and we have queued segments,
1814  * but cannot send anything now because of SWS or another problem.
1815  */
1816 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1817                            int push_one, gfp_t gfp)
1818 {
1819         struct tcp_sock *tp = tcp_sk(sk);
1820         struct sk_buff *skb;
1821         unsigned int tso_segs, sent_pkts;
1822         int cwnd_quota;
1823         int result;
1824 
1825         sent_pkts = 0;
1826 
1827         if (!push_one) {
1828                 /* Do MTU probing. */
1829                 result = tcp_mtu_probe(sk);
1830                 if (!result) {
1831                         return false;
1832                 } else if (result > 0) {
1833                         sent_pkts = 1;
1834                 }
1835         }
1836 
1837         while ((skb = tcp_send_head(sk))) {
1838                 unsigned int limit;
1839 
1840                 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1841                 BUG_ON(!tso_segs);
1842 
1843                 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1844                         goto repair; /* Skip network transmission */
1845 
1846                 cwnd_quota = tcp_cwnd_test(tp, skb);
1847                 if (!cwnd_quota) {
1848                         if (push_one == 2)
1849                                 /* Force out a loss probe pkt. */
1850                                 cwnd_quota = 1;
1851                         else
1852                                 break;
1853                 }
1854 
1855                 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1856                         break;
1857 
1858                 if (tso_segs == 1) {
1859                         if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1860                                                      (tcp_skb_is_last(sk, skb) ?
1861                                                       nonagle : TCP_NAGLE_PUSH))))
1862                                 break;
1863                 } else {
1864                         if (!push_one && tcp_tso_should_defer(sk, skb))
1865                                 break;
1866                 }
1867 
1868                 /* TCP Small Queues :
1869                  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1870                  * This allows for :
1871                  *  - better RTT estimation and ACK scheduling
1872                  *  - faster recovery
1873                  *  - high rates
1874                  */
1875                 limit = max(skb->truesize, sk->sk_pacing_rate >> 10);
1876 
1877                 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1878                         set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1879                         break;
1880                 }
1881 
1882                 limit = mss_now;
1883                 if (tso_segs > 1 && !tcp_urg_mode(tp))
1884                         limit = tcp_mss_split_point(sk, skb, mss_now,
1885                                                     min_t(unsigned int,
1886                                                           cwnd_quota,
1887                                                           sk->sk_gso_max_segs));
1888 
1889                 if (skb->len > limit &&
1890                     unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1891                         break;
1892 
1893                 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1894 
1895                 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1896                         break;
1897 
1898 repair:
1899                 /* Advance the send_head.  This one is sent out.
1900                  * This call will increment packets_out.
1901                  */
1902                 tcp_event_new_data_sent(sk, skb);
1903 
1904                 tcp_minshall_update(tp, mss_now, skb);
1905                 sent_pkts += tcp_skb_pcount(skb);
1906 
1907                 if (push_one)
1908                         break;
1909         }
1910 
1911         if (likely(sent_pkts)) {
1912                 if (tcp_in_cwnd_reduction(sk))
1913                         tp->prr_out += sent_pkts;
1914 
1915                 /* Send one loss probe per tail loss episode. */
1916                 if (push_one != 2)
1917                         tcp_schedule_loss_probe(sk);
1918                 tcp_cwnd_validate(sk);
1919                 return false;
1920         }
1921         return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1922 }
1923 
1924 bool tcp_schedule_loss_probe(struct sock *sk)
1925 {
1926         struct inet_connection_sock *icsk = inet_csk(sk);
1927         struct tcp_sock *tp = tcp_sk(sk);
1928         u32 timeout, tlp_time_stamp, rto_time_stamp;
1929         u32 rtt = tp->srtt >> 3;
1930 
1931         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1932                 return false;
1933         /* No consecutive loss probes. */
1934         if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1935                 tcp_rearm_rto(sk);
1936                 return false;
1937         }
1938         /* Don't do any loss probe on a Fast Open connection before 3WHS
1939          * finishes.
1940          */
1941         if (sk->sk_state == TCP_SYN_RECV)
1942                 return false;
1943 
1944         /* TLP is only scheduled when next timer event is RTO. */
1945         if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1946                 return false;
1947 
1948         /* Schedule a loss probe in 2*RTT for SACK capable connections
1949          * in Open state, that are either limited by cwnd or application.
1950          */
1951         if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1952             !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1953                 return false;
1954 
1955         if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1956              tcp_send_head(sk))
1957                 return false;
1958 
1959         /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1960          * for delayed ack when there's one outstanding packet.
1961          */
1962         timeout = rtt << 1;
1963         if (tp->packets_out == 1)
1964                 timeout = max_t(u32, timeout,
1965                                 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1966         timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1967 
1968         /* If RTO is shorter, just schedule TLP in its place. */
1969         tlp_time_stamp = tcp_time_stamp + timeout;
1970         rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1971         if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1972                 s32 delta = rto_time_stamp - tcp_time_stamp;
1973                 if (delta > 0)
1974                         timeout = delta;
1975         }
1976 
1977         inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1978                                   TCP_RTO_MAX);
1979         return true;
1980 }
1981 
1982 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1983  * retransmit the last segment.
1984  */
1985 void tcp_send_loss_probe(struct sock *sk)
1986 {
1987         struct tcp_sock *tp = tcp_sk(sk);
1988         struct sk_buff *skb;
1989         int pcount;
1990         int mss = tcp_current_mss(sk);
1991         int err = -1;
1992 
1993         if (tcp_send_head(sk) != NULL) {
1994                 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
1995                 goto rearm_timer;
1996         }
1997 
1998         /* At most one outstanding TLP retransmission. */
1999         if (tp->tlp_high_seq)
2000                 goto rearm_timer;
2001 
2002         /* Retransmit last segment. */
2003         skb = tcp_write_queue_tail(sk);
2004         if (WARN_ON(!skb))
2005                 goto rearm_timer;
2006 
2007         pcount = tcp_skb_pcount(skb);
2008         if (WARN_ON(!pcount))
2009                 goto rearm_timer;
2010 
2011         if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2012                 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2013                         goto rearm_timer;
2014                 skb = tcp_write_queue_tail(sk);
2015         }
2016 
2017         if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2018                 goto rearm_timer;
2019 
2020         /* Probe with zero data doesn't trigger fast recovery. */
2021         if (skb->len > 0)
2022                 err = __tcp_retransmit_skb(sk, skb);
2023 
2024         /* Record snd_nxt for loss detection. */
2025         if (likely(!err))
2026                 tp->tlp_high_seq = tp->snd_nxt;
2027 
2028 rearm_timer:
2029         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2030                                   inet_csk(sk)->icsk_rto,
2031                                   TCP_RTO_MAX);
2032 
2033         if (likely(!err))
2034                 NET_INC_STATS_BH(sock_net(sk),
2035                                  LINUX_MIB_TCPLOSSPROBES);
2036         return;
2037 }
2038 
2039 /* Push out any pending frames which were held back due to
2040  * TCP_CORK or attempt at coalescing tiny packets.
2041  * The socket must be locked by the caller.
2042  */
2043 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2044                                int nonagle)
2045 {
2046         /* If we are closed, the bytes will have to remain here.
2047          * In time closedown will finish, we empty the write queue and
2048          * all will be happy.
2049          */
2050         if (unlikely(sk->sk_state == TCP_CLOSE))
2051                 return;
2052 
2053         if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2054                            sk_gfp_atomic(sk, GFP_ATOMIC)))
2055                 tcp_check_probe_timer(sk);
2056 }
2057 
2058 /* Send _single_ skb sitting at the send head. This function requires
2059  * true push pending frames to setup probe timer etc.
2060  */
2061 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2062 {
2063         struct sk_buff *skb = tcp_send_head(sk);
2064 
2065         BUG_ON(!skb || skb->len < mss_now);
2066 
2067         tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2068 }
2069 
2070 /* This function returns the amount that we can raise the
2071  * usable window based on the following constraints
2072  *
2073  * 1. The window can never be shrunk once it is offered (RFC 793)
2074  * 2. We limit memory per socket
2075  *
2076  * RFC 1122:
2077  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2078  *  RECV.NEXT + RCV.WIN fixed until:
2079  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2080  *
2081  * i.e. don't raise the right edge of the window until you can raise
2082  * it at least MSS bytes.
2083  *
2084  * Unfortunately, the recommended algorithm breaks header prediction,
2085  * since header prediction assumes th->window stays fixed.
2086  *
2087  * Strictly speaking, keeping th->window fixed violates the receiver
2088  * side SWS prevention criteria. The problem is that under this rule
2089  * a stream of single byte packets will cause the right side of the
2090  * window to always advance by a single byte.
2091  *
2092  * Of course, if the sender implements sender side SWS prevention
2093  * then this will not be a problem.
2094  *
2095  * BSD seems to make the following compromise:
2096  *
2097  *      If the free space is less than the 1/4 of the maximum
2098  *      space available and the free space is less than 1/2 mss,
2099  *      then set the window to 0.
2100  *      [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2101  *      Otherwise, just prevent the window from shrinking
2102  *      and from being larger than the largest representable value.
2103  *
2104  * This prevents incremental opening of the window in the regime
2105  * where TCP is limited by the speed of the reader side taking
2106  * data out of the TCP receive queue. It does nothing about
2107  * those cases where the window is constrained on the sender side
2108  * because the pipeline is full.
2109  *
2110  * BSD also seems to "accidentally" limit itself to windows that are a
2111  * multiple of MSS, at least until the free space gets quite small.
2112  * This would appear to be a side effect of the mbuf implementation.
2113  * Combining these two algorithms results in the observed behavior
2114  * of having a fixed window size at almost all times.
2115  *
2116  * Below we obtain similar behavior by forcing the offered window to
2117  * a multiple of the mss when it is feasible to do so.
2118  *
2119  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2120  * Regular options like TIMESTAMP are taken into account.
2121  */
2122 u32 __tcp_select_window(struct sock *sk)
2123 {
2124         struct inet_connection_sock *icsk = inet_csk(sk);
2125         struct tcp_sock *tp = tcp_sk(sk);
2126         /* MSS for the peer's data.  Previous versions used mss_clamp
2127          * here.  I don't know if the value based on our guesses
2128          * of peer's MSS is better for the performance.  It's more correct
2129          * but may be worse for the performance because of rcv_mss
2130          * fluctuations.  --SAW  1998/11/1
2131          */
2132         int mss = icsk->icsk_ack.rcv_mss;
2133         int free_space = tcp_space(sk);
2134         int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2135         int window;
2136 
2137         if (mss > full_space)
2138                 mss = full_space;
2139 
2140         if (free_space < (full_space >> 1)) {
2141                 icsk->icsk_ack.quick = 0;
2142 
2143                 if (sk_under_memory_pressure(sk))
2144                         tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2145                                                4U * tp->advmss);
2146 
2147                 if (free_space < mss)
2148                         return 0;
2149         }
2150 
2151         if (free_space > tp->rcv_ssthresh)
2152                 free_space = tp->rcv_ssthresh;
2153 
2154         /* Don't do rounding if we are using window scaling, since the
2155          * scaled window will not line up with the MSS boundary anyway.
2156          */
2157         window = tp->rcv_wnd;
2158         if (tp->rx_opt.rcv_wscale) {
2159                 window = free_space;
2160 
2161                 /* Advertise enough space so that it won't get scaled away.
2162                  * Import case: prevent zero window announcement if
2163                  * 1<<rcv_wscale > mss.
2164                  */
2165                 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2166                         window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2167                                   << tp->rx_opt.rcv_wscale);
2168         } else {
2169                 /* Get the largest window that is a nice multiple of mss.
2170                  * Window clamp already applied above.
2171                  * If our current window offering is within 1 mss of the
2172                  * free space we just keep it. This prevents the divide
2173                  * and multiply from happening most of the time.
2174                  * We also don't do any window rounding when the free space
2175                  * is too small.
2176                  */
2177                 if (window <= free_space - mss || window > free_space)
2178                         window = (free_space / mss) * mss;
2179                 else if (mss == full_space &&
2180                          free_space > window + (full_space >> 1))
2181                         window = free_space;
2182         }
2183 
2184         return window;
2185 }
2186 
2187 /* Collapses two adjacent SKB's during retransmission. */
2188 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2189 {
2190         struct tcp_sock *tp = tcp_sk(sk);
2191         struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2192         int skb_size, next_skb_size;
2193 
2194         skb_size = skb->len;
2195         next_skb_size = next_skb->len;
2196 
2197         BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2198 
2199         tcp_highest_sack_combine(sk, next_skb, skb);
2200 
2201         tcp_unlink_write_queue(next_skb, sk);
2202 
2203         skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2204                                   next_skb_size);
2205 
2206         if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2207                 skb->ip_summed = CHECKSUM_PARTIAL;
2208 
2209         if (skb->ip_summed != CHECKSUM_PARTIAL)
2210                 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2211 
2212         /* Update sequence range on original skb. */
2213         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2214 
2215         /* Merge over control information. This moves PSH/FIN etc. over */
2216         TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2217 
2218         /* All done, get rid of second SKB and account for it so
2219          * packet counting does not break.
2220          */
2221         TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2222 
2223         /* changed transmit queue under us so clear hints */
2224         tcp_clear_retrans_hints_partial(tp);
2225         if (next_skb == tp->retransmit_skb_hint)
2226                 tp->retransmit_skb_hint = skb;
2227 
2228         tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2229 
2230         sk_wmem_free_skb(sk, next_skb);
2231 }
2232 
2233 /* Check if coalescing SKBs is legal. */
2234 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2235 {
2236         if (tcp_skb_pcount(skb) > 1)
2237                 return false;
2238         /* TODO: SACK collapsing could be used to remove this condition */
2239         if (skb_shinfo(skb)->nr_frags != 0)
2240                 return false;
2241         if (skb_cloned(skb))
2242                 return false;
2243         if (skb == tcp_send_head(sk))
2244                 return false;
2245         /* Some heurestics for collapsing over SACK'd could be invented */
2246         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2247                 return false;
2248 
2249         return true;
2250 }
2251 
2252 /* Collapse packets in the retransmit queue to make to create
2253  * less packets on the wire. This is only done on retransmission.
2254  */
2255 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2256                                      int space)
2257 {
2258         struct tcp_sock *tp = tcp_sk(sk);
2259         struct sk_buff *skb = to, *tmp;
2260         bool first = true;
2261 
2262         if (!sysctl_tcp_retrans_collapse)
2263                 return;
2264         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2265                 return;
2266 
2267         tcp_for_write_queue_from_safe(skb, tmp, sk) {
2268                 if (!tcp_can_collapse(sk, skb))
2269                         break;
2270 
2271                 space -= skb->len;
2272 
2273                 if (first) {
2274                         first = false;
2275                         continue;
2276                 }
2277 
2278                 if (space < 0)
2279                         break;
2280                 /* Punt if not enough space exists in the first SKB for
2281                  * the data in the second
2282                  */
2283                 if (skb->len > skb_availroom(to))
2284                         break;
2285 
2286                 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2287                         break;
2288 
2289                 tcp_collapse_retrans(sk, to);
2290         }
2291 }
2292 
2293 /* This retransmits one SKB.  Policy decisions and retransmit queue
2294  * state updates are done by the caller.  Returns non-zero if an
2295  * error occurred which prevented the send.
2296  */
2297 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2298 {
2299         struct tcp_sock *tp = tcp_sk(sk);
2300         struct inet_connection_sock *icsk = inet_csk(sk);
2301         unsigned int cur_mss;
2302 
2303         /* Inconslusive MTU probe */
2304         if (icsk->icsk_mtup.probe_size) {
2305                 icsk->icsk_mtup.probe_size = 0;
2306         }
2307 
2308         /* Do not sent more than we queued. 1/4 is reserved for possible
2309          * copying overhead: fragmentation, tunneling, mangling etc.
2310          */
2311         if (atomic_read(&sk->sk_wmem_alloc) >
2312             min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2313                 return -EAGAIN;
2314 
2315         if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2316                 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2317                         BUG();
2318                 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2319                         return -ENOMEM;
2320         }
2321 
2322         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2323                 return -EHOSTUNREACH; /* Routing failure or similar. */
2324 
2325         cur_mss = tcp_current_mss(sk);
2326 
2327         /* If receiver has shrunk his window, and skb is out of
2328          * new window, do not retransmit it. The exception is the
2329          * case, when window is shrunk to zero. In this case
2330          * our retransmit serves as a zero window probe.
2331          */
2332         if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2333             TCP_SKB_CB(skb)->seq != tp->snd_una)
2334                 return -EAGAIN;
2335 
2336         if (skb->len > cur_mss) {
2337                 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2338                         return -ENOMEM; /* We'll try again later. */
2339         } else {
2340                 int oldpcount = tcp_skb_pcount(skb);
2341 
2342                 if (unlikely(oldpcount > 1)) {
2343                         if (skb_unclone(skb, GFP_ATOMIC))
2344                                 return -ENOMEM;
2345                         tcp_init_tso_segs(sk, skb, cur_mss);
2346                         tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2347                 }
2348         }
2349 
2350         tcp_retrans_try_collapse(sk, skb, cur_mss);
2351 
2352         /* Some Solaris stacks overoptimize and ignore the FIN on a
2353          * retransmit when old data is attached.  So strip it off
2354          * since it is cheap to do so and saves bytes on the network.
2355          */
2356         if (skb->len > 0 &&
2357             (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2358             tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2359                 if (!pskb_trim(skb, 0)) {
2360                         /* Reuse, even though it does some unnecessary work */
2361                         tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2362                                              TCP_SKB_CB(skb)->tcp_flags);
2363                         skb->ip_summed = CHECKSUM_NONE;
2364                 }
2365         }
2366 
2367         /* Make a copy, if the first transmission SKB clone we made
2368          * is still in somebody's hands, else make a clone.
2369          */
2370         TCP_SKB_CB(skb)->when = tcp_time_stamp;
2371 
2372         /* make sure skb->data is aligned on arches that require it
2373          * and check if ack-trimming & collapsing extended the headroom
2374          * beyond what csum_start can cover.
2375          */
2376         if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2377                      skb_headroom(skb) >= 0xFFFF)) {
2378                 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2379                                                    GFP_ATOMIC);
2380                 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2381                               -ENOBUFS;
2382         } else {
2383                 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2384         }
2385 }
2386 
2387 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2388 {
2389         struct tcp_sock *tp = tcp_sk(sk);
2390         int err = __tcp_retransmit_skb(sk, skb);
2391 
2392         if (err == 0) {
2393                 /* Update global TCP statistics. */
2394                 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2395 
2396                 tp->total_retrans++;
2397 
2398 #if FASTRETRANS_DEBUG > 0
2399                 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2400                         net_dbg_ratelimited("retrans_out leaked\n");
2401                 }
2402 #endif
2403                 if (!tp->retrans_out)
2404                         tp->lost_retrans_low = tp->snd_nxt;
2405                 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2406                 tp->retrans_out += tcp_skb_pcount(skb);
2407 
2408                 /* Save stamp of the first retransmit. */
2409                 if (!tp->retrans_stamp)
2410                         tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2411 
2412                 tp->undo_retrans += tcp_skb_pcount(skb);
2413 
2414                 /* snd_nxt is stored to detect loss of retransmitted segment,
2415                  * see tcp_input.c tcp_sacktag_write_queue().
2416                  */
2417                 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2418         } else {
2419                 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2420         }
2421         return err;
2422 }
2423 
2424 /* Check if we forward retransmits are possible in the current
2425  * window/congestion state.
2426  */
2427 static bool tcp_can_forward_retransmit(struct sock *sk)
2428 {
2429         const struct inet_connection_sock *icsk = inet_csk(sk);
2430         const struct tcp_sock *tp = tcp_sk(sk);
2431 
2432         /* Forward retransmissions are possible only during Recovery. */
2433         if (icsk->icsk_ca_state != TCP_CA_Recovery)
2434                 return false;
2435 
2436         /* No forward retransmissions in Reno are possible. */
2437         if (tcp_is_reno(tp))
2438                 return false;
2439 
2440         /* Yeah, we have to make difficult choice between forward transmission
2441          * and retransmission... Both ways have their merits...
2442          *
2443          * For now we do not retransmit anything, while we have some new
2444          * segments to send. In the other cases, follow rule 3 for
2445          * NextSeg() specified in RFC3517.
2446          */
2447 
2448         if (tcp_may_send_now(sk))
2449                 return false;
2450 
2451         return true;
2452 }
2453 
2454 /* This gets called after a retransmit timeout, and the initially
2455  * retransmitted data is acknowledged.  It tries to continue
2456  * resending the rest of the retransmit queue, until either
2457  * we've sent it all or the congestion window limit is reached.
2458  * If doing SACK, the first ACK which comes back for a timeout
2459  * based retransmit packet might feed us FACK information again.
2460  * If so, we use it to avoid unnecessarily retransmissions.
2461  */
2462 void tcp_xmit_retransmit_queue(struct sock *sk)
2463 {
2464         const struct inet_connection_sock *icsk = inet_csk(sk);
2465         struct tcp_sock *tp = tcp_sk(sk);
2466         struct sk_buff *skb;
2467         struct sk_buff *hole = NULL;
2468         u32 last_lost;
2469         int mib_idx;
2470         int fwd_rexmitting = 0;
2471 
2472         if (!tp->packets_out)
2473                 return;
2474 
2475         if (!tp->lost_out)
2476                 tp->retransmit_high = tp->snd_una;
2477 
2478         if (tp->retransmit_skb_hint) {
2479                 skb = tp->retransmit_skb_hint;
2480                 last_lost = TCP_SKB_CB(skb)->end_seq;
2481                 if (after(last_lost, tp->retransmit_high))
2482                         last_lost = tp->retransmit_high;
2483         } else {
2484                 skb = tcp_write_queue_head(sk);
2485                 last_lost = tp->snd_una;
2486         }
2487 
2488         tcp_for_write_queue_from(skb, sk) {
2489                 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2490 
2491                 if (skb == tcp_send_head(sk))
2492                         break;
2493                 /* we could do better than to assign each time */
2494                 if (hole == NULL)
2495                         tp->retransmit_skb_hint = skb;
2496 
2497                 /* Assume this retransmit will generate
2498                  * only one packet for congestion window
2499                  * calculation purposes.  This works because
2500                  * tcp_retransmit_skb() will chop up the
2501                  * packet to be MSS sized and all the
2502                  * packet counting works out.
2503                  */
2504                 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2505                         return;
2506 
2507                 if (fwd_rexmitting) {
2508 begin_fwd:
2509                         if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2510                                 break;
2511                         mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2512 
2513                 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2514                         tp->retransmit_high = last_lost;
2515                         if (!tcp_can_forward_retransmit(sk))
2516                                 break;
2517                         /* Backtrack if necessary to non-L'ed skb */
2518                         if (hole != NULL) {
2519                                 skb = hole;
2520                                 hole = NULL;
2521                         }
2522                         fwd_rexmitting = 1;
2523                         goto begin_fwd;
2524 
2525                 } else if (!(sacked & TCPCB_LOST)) {
2526                         if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2527                                 hole = skb;
2528                         continue;
2529 
2530                 } else {
2531                         last_lost = TCP_SKB_CB(skb)->end_seq;
2532                         if (icsk->icsk_ca_state != TCP_CA_Loss)
2533                                 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2534                         else
2535                                 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2536                 }
2537 
2538                 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2539                         continue;
2540 
2541                 if (tcp_retransmit_skb(sk, skb))
2542                         return;
2543 
2544                 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2545 
2546                 if (tcp_in_cwnd_reduction(sk))
2547                         tp->prr_out += tcp_skb_pcount(skb);
2548 
2549                 if (skb == tcp_write_queue_head(sk))
2550                         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2551                                                   inet_csk(sk)->icsk_rto,
2552                                                   TCP_RTO_MAX);
2553         }
2554 }
2555 
2556 /* Send a fin.  The caller locks the socket for us.  This cannot be
2557  * allowed to fail queueing a FIN frame under any circumstances.
2558  */
2559 void tcp_send_fin(struct sock *sk)
2560 {
2561         struct tcp_sock *tp = tcp_sk(sk);
2562         struct sk_buff *skb = tcp_write_queue_tail(sk);
2563         int mss_now;
2564 
2565         /* Optimization, tack on the FIN if we have a queue of
2566          * unsent frames.  But be careful about outgoing SACKS
2567          * and IP options.
2568          */
2569         mss_now = tcp_current_mss(sk);
2570 
2571         if (tcp_send_head(sk) != NULL) {
2572                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2573                 TCP_SKB_CB(skb)->end_seq++;
2574                 tp->write_seq++;
2575         } else {
2576                 /* Socket is locked, keep trying until memory is available. */
2577                 for (;;) {
2578                         skb = alloc_skb_fclone(MAX_TCP_HEADER,
2579                                                sk->sk_allocation);
2580                         if (skb)
2581                                 break;
2582                         yield();
2583                 }
2584 
2585                 /* Reserve space for headers and prepare control bits. */
2586                 skb_reserve(skb, MAX_TCP_HEADER);
2587                 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2588                 tcp_init_nondata_skb(skb, tp->write_seq,
2589                                      TCPHDR_ACK | TCPHDR_FIN);
2590                 tcp_queue_skb(sk, skb);
2591         }
2592         __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2593 }
2594 
2595 /* We get here when a process closes a file descriptor (either due to
2596  * an explicit close() or as a byproduct of exit()'ing) and there
2597  * was unread data in the receive queue.  This behavior is recommended
2598  * by RFC 2525, section 2.17.  -DaveM
2599  */
2600 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2601 {
2602         struct sk_buff *skb;
2603 
2604         /* NOTE: No TCP options attached and we never retransmit this. */
2605         skb = alloc_skb(MAX_TCP_HEADER, priority);
2606         if (!skb) {
2607                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2608                 return;
2609         }
2610 
2611         /* Reserve space for headers and prepare control bits. */
2612         skb_reserve(skb, MAX_TCP_HEADER);
2613         tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2614                              TCPHDR_ACK | TCPHDR_RST);
2615         /* Send it off. */
2616         TCP_SKB_CB(skb)->when = tcp_time_stamp;
2617         if (tcp_transmit_skb(sk, skb, 0, priority))
2618                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2619 
2620         TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2621 }
2622 
2623 /* Send a crossed SYN-ACK during socket establishment.
2624  * WARNING: This routine must only be called when we have already sent
2625  * a SYN packet that crossed the incoming SYN that caused this routine
2626  * to get called. If this assumption fails then the initial rcv_wnd
2627  * and rcv_wscale values will not be correct.
2628  */
2629 int tcp_send_synack(struct sock *sk)
2630 {
2631         struct sk_buff *skb;
2632 
2633         skb = tcp_write_queue_head(sk);
2634         if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2635                 pr_debug("%s: wrong queue state\n", __func__);
2636                 return -EFAULT;
2637         }
2638         if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2639                 if (skb_cloned(skb)) {
2640                         struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2641                         if (nskb == NULL)
2642                                 return -ENOMEM;
2643                         tcp_unlink_write_queue(skb, sk);
2644                         skb_header_release(nskb);
2645                         __tcp_add_write_queue_head(sk, nskb);
2646                         sk_wmem_free_skb(sk, skb);
2647                         sk->sk_wmem_queued += nskb->truesize;
2648                         sk_mem_charge(sk, nskb->truesize);
2649                         skb = nskb;
2650                 }
2651 
2652                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2653                 TCP_ECN_send_synack(tcp_sk(sk), skb);
2654         }
2655         TCP_SKB_CB(skb)->when = tcp_time_stamp;
2656         return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2657 }
2658 
2659 /**
2660  * tcp_make_synack - Prepare a SYN-ACK.
2661  * sk: listener socket
2662  * dst: dst entry attached to the SYNACK
2663  * req: request_sock pointer
2664  *
2665  * Allocate one skb and build a SYNACK packet.
2666  * @dst is consumed : Caller should not use it again.
2667  */
2668 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2669                                 struct request_sock *req,
2670                                 struct tcp_fastopen_cookie *foc)
2671 {
2672         struct tcp_out_options opts;
2673         struct inet_request_sock *ireq = inet_rsk(req);
2674         struct tcp_sock *tp = tcp_sk(sk);
2675         struct tcphdr *th;
2676         struct sk_buff *skb;
2677         struct tcp_md5sig_key *md5;
2678         int tcp_header_size;
2679         int mss;
2680 
2681         skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2682         if (unlikely(!skb)) {
2683                 dst_release(dst);
2684                 return NULL;
2685         }
2686         /* Reserve space for headers. */
2687         skb_reserve(skb, MAX_TCP_HEADER);
2688 
2689         skb_dst_set(skb, dst);
2690         security_skb_owned_by(skb, sk);
2691 
2692         mss = dst_metric_advmss(dst);
2693         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2694                 mss = tp->rx_opt.user_mss;
2695 
2696         if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2697                 __u8 rcv_wscale;
2698                 /* Set this up on the first call only */
2699                 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2700 
2701                 /* limit the window selection if the user enforce a smaller rx buffer */
2702                 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2703                     (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2704                         req->window_clamp = tcp_full_space(sk);
2705 
2706                 /* tcp_full_space because it is guaranteed to be the first packet */
2707                 tcp_select_initial_window(tcp_full_space(sk),
2708                         mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2709                         &req->rcv_wnd,
2710                         &req->window_clamp,
2711                         ireq->wscale_ok,
2712                         &rcv_wscale,
2713                         dst_metric(dst, RTAX_INITRWND));
2714                 ireq->rcv_wscale = rcv_wscale;
2715         }
2716 
2717         memset(&opts, 0, sizeof(opts));
2718 #ifdef CONFIG_SYN_COOKIES
2719         if (unlikely(req->cookie_ts))
2720                 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2721         else
2722 #endif
2723         TCP_SKB_CB(skb)->when = tcp_time_stamp;
2724         tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2725                                              foc) + sizeof(*th);
2726 
2727         skb_push(skb, tcp_header_size);
2728         skb_reset_transport_header(skb);
2729 
2730         th = tcp_hdr(skb);
2731         memset(th, 0, sizeof(struct tcphdr));
2732         th->syn = 1;
2733         th->ack = 1;
2734         TCP_ECN_make_synack(req, th);
2735         th->source = ireq->loc_port;
2736         th->dest = ireq->rmt_port;
2737         /* Setting of flags are superfluous here for callers (and ECE is
2738          * not even correctly set)
2739          */
2740         tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2741                              TCPHDR_SYN | TCPHDR_ACK);
2742 
2743         th->seq = htonl(TCP_SKB_CB(skb)->seq);
2744         /* XXX data is queued and acked as is. No buffer/window check */
2745         th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2746 
2747         /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2748         th->window = htons(min(req->rcv_wnd, 65535U));
2749         tcp_options_write((__be32 *)(th + 1), tp, &opts);
2750         th->doff = (tcp_header_size >> 2);
2751         TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2752 
2753 #ifdef CONFIG_TCP_MD5SIG
2754         /* Okay, we have all we need - do the md5 hash if needed */
2755         if (md5) {
2756                 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2757                                                md5, NULL, req, skb);
2758         }
2759 #endif
2760 
2761         return skb;
2762 }
2763 EXPORT_SYMBOL(tcp_make_synack);
2764 
2765 /* Do all connect socket setups that can be done AF independent. */
2766 void tcp_connect_init(struct sock *sk)
2767 {
2768         const struct dst_entry *dst = __sk_dst_get(sk);
2769         struct tcp_sock *tp = tcp_sk(sk);
2770         __u8 rcv_wscale;
2771 
2772         /* We'll fix this up when we get a response from the other end.
2773          * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2774          */
2775         tp->tcp_header_len = sizeof(struct tcphdr) +
2776                 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2777 
2778 #ifdef CONFIG_TCP_MD5SIG
2779         if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2780                 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2781 #endif
2782 
2783         /* If user gave his TCP_MAXSEG, record it to clamp */
2784         if (tp->rx_opt.user_mss)
2785                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2786         tp->max_window = 0;
2787         tcp_mtup_init(sk);
2788         tcp_sync_mss(sk, dst_mtu(dst));
2789 
2790         if (!tp->window_clamp)
2791                 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2792         tp->advmss = dst_metric_advmss(dst);
2793         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2794                 tp->advmss = tp->rx_opt.user_mss;
2795 
2796         tcp_initialize_rcv_mss(sk);
2797 
2798         /* limit the window selection if the user enforce a smaller rx buffer */
2799         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2800             (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2801                 tp->window_clamp = tcp_full_space(sk);
2802 
2803         tcp_select_initial_window(tcp_full_space(sk),
2804                                   tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2805                                   &tp->rcv_wnd,
2806                                   &tp->window_clamp,
2807                                   sysctl_tcp_window_scaling,
2808                                   &rcv_wscale,
2809                                   dst_metric(dst, RTAX_INITRWND));
2810 
2811         tp->rx_opt.rcv_wscale = rcv_wscale;
2812         tp->rcv_ssthresh = tp->rcv_wnd;
2813 
2814         sk->sk_err = 0;
2815         sock_reset_flag(sk, SOCK_DONE);
2816         tp->snd_wnd = 0;
2817         tcp_init_wl(tp, 0);
2818         tp->snd_una = tp->write_seq;
2819         tp->snd_sml = tp->write_seq;
2820         tp->snd_up = tp->write_seq;
2821         tp->snd_nxt = tp->write_seq;
2822 
2823         if (likely(!tp->repair))
2824                 tp->rcv_nxt = 0;
2825         else
2826                 tp->rcv_tstamp = tcp_time_stamp;
2827         tp->rcv_wup = tp->rcv_nxt;
2828         tp->copied_seq = tp->rcv_nxt;
2829 
2830         inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2831         inet_csk(sk)->icsk_retransmits = 0;
2832         tcp_clear_retrans(tp);
2833 }
2834 
2835 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2836 {
2837         struct tcp_sock *tp = tcp_sk(sk);
2838         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2839 
2840         tcb->end_seq += skb->len;
2841         skb_header_release(skb);
2842         __tcp_add_write_queue_tail(sk, skb);
2843         sk->sk_wmem_queued += skb->truesize;
2844         sk_mem_charge(sk, skb->truesize);
2845         tp->write_seq = tcb->end_seq;
2846         tp->packets_out += tcp_skb_pcount(skb);
2847 }
2848 
2849 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2850  * queue a data-only packet after the regular SYN, such that regular SYNs
2851  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2852  * only the SYN sequence, the data are retransmitted in the first ACK.
2853  * If cookie is not cached or other error occurs, falls back to send a
2854  * regular SYN with Fast Open cookie request option.
2855  */
2856 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2857 {
2858         struct tcp_sock *tp = tcp_sk(sk);
2859         struct tcp_fastopen_request *fo = tp->fastopen_req;
2860         int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2861         struct sk_buff *syn_data = NULL, *data;
2862         unsigned long last_syn_loss = 0;
2863 
2864         tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
2865         tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2866                                &syn_loss, &last_syn_loss);
2867         /* Recurring FO SYN losses: revert to regular handshake temporarily */
2868         if (syn_loss > 1 &&
2869             time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2870                 fo->cookie.len = -1;
2871                 goto fallback;
2872         }
2873 
2874         if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2875                 fo->cookie.len = -1;
2876         else if (fo->cookie.len <= 0)
2877                 goto fallback;
2878 
2879         /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2880          * user-MSS. Reserve maximum option space for middleboxes that add
2881          * private TCP options. The cost is reduced data space in SYN :(
2882          */
2883         if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2884                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2885         space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2886                 MAX_TCP_OPTION_SPACE;
2887 
2888         syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2889                                    sk->sk_allocation);
2890         if (syn_data == NULL)
2891                 goto fallback;
2892 
2893         for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2894                 struct iovec *iov = &fo->data->msg_iov[i];
2895                 unsigned char __user *from = iov->iov_base;
2896                 int len = iov->iov_len;
2897 
2898                 if (syn_data->len + len > space)
2899                         len = space - syn_data->len;
2900                 else if (i + 1 == iovlen)
2901                         /* No more data pending in inet_wait_for_connect() */
2902                         fo->data = NULL;
2903 
2904                 if (skb_add_data(syn_data, from, len))
2905                         goto fallback;
2906         }
2907 
2908         /* Queue a data-only packet after the regular SYN for retransmission */
2909         data = pskb_copy(syn_data, sk->sk_allocation);
2910         if (data == NULL)
2911                 goto fallback;
2912         TCP_SKB_CB(data)->seq++;
2913         TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2914         TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2915         tcp_connect_queue_skb(sk, data);
2916         fo->copied = data->len;
2917 
2918         if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2919                 tp->syn_data = (fo->copied > 0);
2920                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2921                 goto done;
2922         }
2923         syn_data = NULL;
2924 
2925 fallback:
2926         /* Send a regular SYN with Fast Open cookie request option */
2927         if (fo->cookie.len > 0)
2928                 fo->cookie.len = 0;
2929         err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2930         if (err)
2931                 tp->syn_fastopen = 0;
2932         kfree_skb(syn_data);
2933 done:
2934         fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
2935         return err;
2936 }
2937 
2938 /* Build a SYN and send it off. */
2939 int tcp_connect(struct sock *sk)
2940 {
2941         struct tcp_sock *tp = tcp_sk(sk);
2942         struct sk_buff *buff;
2943         int err;
2944 
2945         tcp_connect_init(sk);
2946 
2947         if (unlikely(tp->repair)) {
2948                 tcp_finish_connect(sk, NULL);
2949                 return 0;
2950         }
2951 
2952         buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2953         if (unlikely(buff == NULL))
2954                 return -ENOBUFS;
2955 
2956         /* Reserve space for headers. */
2957         skb_reserve(buff, MAX_TCP_HEADER);
2958 
2959         tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2960         tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2961         tcp_connect_queue_skb(sk, buff);
2962         TCP_ECN_send_syn(sk, buff);
2963 
2964         /* Send off SYN; include data in Fast Open. */
2965         err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2966               tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2967         if (err == -ECONNREFUSED)
2968                 return err;
2969 
2970         /* We change tp->snd_nxt after the tcp_transmit_skb() call
2971          * in order to make this packet get counted in tcpOutSegs.
2972          */
2973         tp->snd_nxt = tp->write_seq;
2974         tp->pushed_seq = tp->write_seq;
2975         TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2976 
2977         /* Timer for repeating the SYN until an answer. */
2978         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2979                                   inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2980         return 0;
2981 }
2982 EXPORT_SYMBOL(tcp_connect);
2983 
2984 /* Send out a delayed ack, the caller does the policy checking
2985  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2986  * for details.
2987  */
2988 void tcp_send_delayed_ack(struct sock *sk)
2989 {
2990         struct inet_connection_sock *icsk = inet_csk(sk);
2991         int ato = icsk->icsk_ack.ato;
2992         unsigned long timeout;
2993 
2994         if (ato > TCP_DELACK_MIN) {
2995                 const struct tcp_sock *tp = tcp_sk(sk);
2996                 int max_ato = HZ / 2;
2997 
2998                 if (icsk->icsk_ack.pingpong ||
2999                     (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3000                         max_ato = TCP_DELACK_MAX;
3001 
3002                 /* Slow path, intersegment interval is "high". */
3003 
3004                 /* If some rtt estimate is known, use it to bound delayed ack.
3005                  * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3006                  * directly.
3007                  */
3008                 if (tp->srtt) {
3009                         int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3010 
3011                         if (rtt < max_ato)
3012                                 max_ato = rtt;
3013                 }
3014 
3015                 ato = min(ato, max_ato);
3016         }
3017 
3018         /* Stay within the limit we were given */
3019         timeout = jiffies + ato;
3020 
3021         /* Use new timeout only if there wasn't a older one earlier. */
3022         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3023                 /* If delack timer was blocked or is about to expire,
3024                  * send ACK now.
3025                  */
3026                 if (icsk->icsk_ack.blocked ||
3027                     time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3028                         tcp_send_ack(sk);
3029                         return;
3030                 }
3031 
3032                 if (!time_before(timeout, icsk->icsk_ack.timeout))
3033                         timeout = icsk->icsk_ack.timeout;
3034         }
3035         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3036         icsk->icsk_ack.timeout = timeout;
3037         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3038 }
3039 
3040 /* This routine sends an ack and also updates the window. */
3041 void tcp_send_ack(struct sock *sk)
3042 {
3043         struct sk_buff *buff;
3044 
3045         /* If we have been reset, we may not send again. */
3046         if (sk->sk_state == TCP_CLOSE)
3047                 return;
3048 
3049         /* We are not putting this on the write queue, so
3050          * tcp_transmit_skb() will set the ownership to this
3051          * sock.
3052          */
3053         buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3054         if (buff == NULL) {
3055                 inet_csk_schedule_ack(sk);
3056                 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3057                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3058                                           TCP_DELACK_MAX, TCP_RTO_MAX);
3059                 return;
3060         }
3061 
3062         /* Reserve space for headers and prepare control bits. */
3063         skb_reserve(buff, MAX_TCP_HEADER);
3064         tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3065 
3066         /* Send it off, this clears delayed acks for us. */
3067         TCP_SKB_CB(buff)->when = tcp_time_stamp;
3068         tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3069 }
3070 
3071 /* This routine sends a packet with an out of date sequence
3072  * number. It assumes the other end will try to ack it.
3073  *
3074  * Question: what should we make while urgent mode?
3075  * 4.4BSD forces sending single byte of data. We cannot send
3076  * out of window data, because we have SND.NXT==SND.MAX...
3077  *
3078  * Current solution: to send TWO zero-length segments in urgent mode:
3079  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3080  * out-of-date with SND.UNA-1 to probe window.
3081  */
3082 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3083 {
3084         struct tcp_sock *tp = tcp_sk(sk);
3085         struct sk_buff *skb;
3086 
3087         /* We don't queue it, tcp_transmit_skb() sets ownership. */
3088         skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3089         if (skb == NULL)
3090                 return -1;
3091 
3092         /* Reserve space for headers and set control bits. */
3093         skb_reserve(skb, MAX_TCP_HEADER);
3094         /* Use a previous sequence.  This should cause the other
3095          * end to send an ack.  Don't queue or clone SKB, just
3096          * send it.
3097          */
3098         tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3099         TCP_SKB_CB(skb)->when = tcp_time_stamp;
3100         return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3101 }
3102 
3103 void tcp_send_window_probe(struct sock *sk)
3104 {
3105         if (sk->sk_state == TCP_ESTABLISHED) {
3106                 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3107                 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3108                 tcp_xmit_probe_skb(sk, 0);
3109         }
3110 }
3111 
3112 /* Initiate keepalive or window probe from timer. */
3113 int tcp_write_wakeup(struct sock *sk)
3114 {
3115         struct tcp_sock *tp = tcp_sk(sk);
3116         struct sk_buff *skb;
3117 
3118         if (sk->sk_state == TCP_CLOSE)
3119                 return -1;
3120 
3121         if ((skb = tcp_send_head(sk)) != NULL &&
3122             before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3123                 int err;
3124                 unsigned int mss = tcp_current_mss(sk);
3125                 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3126 
3127                 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3128                         tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3129 
3130                 /* We are probing the opening of a window
3131                  * but the window size is != 0
3132                  * must have been a result SWS avoidance ( sender )
3133                  */
3134                 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3135                     skb->len > mss) {
3136                         seg_size = min(seg_size, mss);
3137                         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3138                         if (tcp_fragment(sk, skb, seg_size, mss))
3139                                 return -1;
3140                 } else if (!tcp_skb_pcount(skb))
3141                         tcp_set_skb_tso_segs(sk, skb, mss);
3142 
3143                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3144                 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3145                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3146                 if (!err)
3147                         tcp_event_new_data_sent(sk, skb);
3148                 return err;
3149         } else {
3150                 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3151                         tcp_xmit_probe_skb(sk, 1);
3152                 return tcp_xmit_probe_skb(sk, 0);
3153         }
3154 }
3155 
3156 /* A window probe timeout has occurred.  If window is not closed send
3157  * a partial packet else a zero probe.
3158  */
3159 void tcp_send_probe0(struct sock *sk)
3160 {
3161         struct inet_connection_sock *icsk = inet_csk(sk);
3162         struct tcp_sock *tp = tcp_sk(sk);
3163         int err;
3164 
3165         err = tcp_write_wakeup(sk);
3166 
3167         if (tp->packets_out || !tcp_send_head(sk)) {
3168                 /* Cancel probe timer, if it is not required. */
3169                 icsk->icsk_probes_out = 0;
3170                 icsk->icsk_backoff = 0;
3171                 return;
3172         }
3173 
3174         if (err <= 0) {
3175                 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3176                         icsk->icsk_backoff++;
3177                 icsk->icsk_probes_out++;
3178                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3179                                           min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3180                                           TCP_RTO_MAX);
3181         } else {
3182                 /* If packet was not sent due to local congestion,
3183                  * do not backoff and do not remember icsk_probes_out.
3184                  * Let local senders to fight for local resources.
3185                  *
3186                  * Use accumulated backoff yet.
3187                  */
3188                 if (!icsk->icsk_probes_out)
3189                         icsk->icsk_probes_out = 1;
3190                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3191                                           min(icsk->icsk_rto << icsk->icsk_backoff,
3192                                               TCP_RESOURCE_PROBE_INTERVAL),
3193                                           TCP_RTO_MAX);
3194         }
3195 }
3196 

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

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp