root/net/ipv4/tcp_output.c

/* [<][>][^][v][top][bottom][index][help] */

DEFINITIONS

This source file includes following definitions.
  1. tcp_mstamp_refresh
  2. tcp_event_new_data_sent
  3. tcp_acceptable_seq
  4. tcp_advertise_mss
  5. tcp_cwnd_restart
  6. tcp_event_data_sent
  7. tcp_event_ack_sent
  8. tcp_select_initial_window
  9. tcp_select_window
  10. tcp_ecn_send_synack
  11. tcp_ecn_send_syn
  12. tcp_ecn_clear_syn
  13. tcp_ecn_make_synack
  14. tcp_ecn_send
  15. tcp_init_nondata_skb
  16. tcp_urg_mode
  17. smc_options_write
  18. tcp_options_write
  19. smc_set_option
  20. smc_set_option_cond
  21. tcp_syn_options
  22. tcp_synack_options
  23. tcp_established_options
  24. tcp_tsq_write
  25. tcp_tsq_handler
  26. tcp_tasklet_func
  27. tcp_release_cb
  28. tcp_tasklet_init
  29. tcp_wfree
  30. tcp_pace_kick
  31. tcp_update_skb_after_send
  32. __tcp_transmit_skb
  33. tcp_transmit_skb
  34. tcp_queue_skb
  35. tcp_set_skb_tso_segs
  36. tcp_adjust_pcount
  37. tcp_has_tx_tstamp
  38. tcp_fragment_tstamp
  39. tcp_skb_fragment_eor
  40. tcp_insert_write_queue_after
  41. tcp_fragment
  42. __pskb_trim_head
  43. tcp_trim_head
  44. __tcp_mtu_to_mss
  45. tcp_mtu_to_mss
  46. tcp_mss_to_mtu
  47. tcp_mtup_init
  48. tcp_sync_mss
  49. tcp_current_mss
  50. tcp_cwnd_application_limited
  51. tcp_cwnd_validate
  52. tcp_minshall_check
  53. tcp_minshall_update
  54. tcp_nagle_check
  55. tcp_tso_autosize
  56. tcp_tso_segs
  57. tcp_mss_split_point
  58. tcp_cwnd_test
  59. tcp_init_tso_segs
  60. tcp_nagle_test
  61. tcp_snd_wnd_test
  62. tso_fragment
  63. tcp_tso_should_defer
  64. tcp_mtu_check_reprobe
  65. tcp_can_coalesce_send_queue_head
  66. tcp_mtu_probe
  67. tcp_pacing_check
  68. tcp_small_queue_check
  69. tcp_chrono_set
  70. tcp_chrono_start
  71. tcp_chrono_stop
  72. tcp_write_xmit
  73. tcp_schedule_loss_probe
  74. skb_still_in_host_queue
  75. tcp_send_loss_probe
  76. __tcp_push_pending_frames
  77. tcp_push_one
  78. __tcp_select_window
  79. tcp_skb_collapse_tstamp
  80. tcp_collapse_retrans
  81. tcp_can_collapse
  82. tcp_retrans_try_collapse
  83. __tcp_retransmit_skb
  84. tcp_retransmit_skb
  85. tcp_xmit_retransmit_queue
  86. sk_forced_mem_schedule
  87. tcp_send_fin
  88. tcp_send_active_reset
  89. tcp_send_synack
  90. tcp_make_synack
  91. tcp_ca_dst_init
  92. tcp_connect_init
  93. tcp_connect_queue_skb
  94. tcp_send_syn_data
  95. tcp_connect
  96. tcp_send_delayed_ack
  97. __tcp_send_ack
  98. tcp_send_ack
  99. tcp_xmit_probe_skb
  100. tcp_send_window_probe
  101. tcp_write_wakeup
  102. tcp_send_probe0
  103. tcp_rtx_synack

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the  BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Implementation of the Transmission Control Protocol(TCP).
   8  *
   9  * Authors:     Ross Biro
  10  *              Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  11  *              Mark Evans, <evansmp@uhura.aston.ac.uk>
  12  *              Corey Minyard <wf-rch!minyard@relay.EU.net>
  13  *              Florian La Roche, <flla@stud.uni-sb.de>
  14  *              Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  15  *              Linus Torvalds, <torvalds@cs.helsinki.fi>
  16  *              Alan Cox, <gw4pts@gw4pts.ampr.org>
  17  *              Matthew Dillon, <dillon@apollo.west.oic.com>
  18  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  19  *              Jorge Cwik, <jorge@laser.satlink.net>
  20  */
  21 
  22 /*
  23  * Changes:     Pedro Roque     :       Retransmit queue handled by TCP.
  24  *                              :       Fragmentation on mtu decrease
  25  *                              :       Segment collapse on retransmit
  26  *                              :       AF independence
  27  *
  28  *              Linus Torvalds  :       send_delayed_ack
  29  *              David S. Miller :       Charge memory using the right skb
  30  *                                      during syn/ack processing.
  31  *              David S. Miller :       Output engine completely rewritten.
  32  *              Andrea Arcangeli:       SYNACK carry ts_recent in tsecr.
  33  *              Cacophonix Gaul :       draft-minshall-nagle-01
  34  *              J Hadi Salim    :       ECN support
  35  *
  36  */
  37 
  38 #define pr_fmt(fmt) "TCP: " fmt
  39 
  40 #include <net/tcp.h>
  41 
  42 #include <linux/compiler.h>
  43 #include <linux/gfp.h>
  44 #include <linux/module.h>
  45 #include <linux/static_key.h>
  46 
  47 #include <trace/events/tcp.h>
  48 
  49 /* Refresh clocks of a TCP socket,
  50  * ensuring monotically increasing values.
  51  */
  52 void tcp_mstamp_refresh(struct tcp_sock *tp)
  53 {
  54         u64 val = tcp_clock_ns();
  55 
  56         tp->tcp_clock_cache = val;
  57         tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
  58 }
  59 
  60 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  61                            int push_one, gfp_t gfp);
  62 
  63 /* Account for new data that has been sent to the network. */
  64 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  65 {
  66         struct inet_connection_sock *icsk = inet_csk(sk);
  67         struct tcp_sock *tp = tcp_sk(sk);
  68         unsigned int prior_packets = tp->packets_out;
  69 
  70         WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
  71 
  72         __skb_unlink(skb, &sk->sk_write_queue);
  73         tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
  74 
  75         if (tp->highest_sack == NULL)
  76                 tp->highest_sack = skb;
  77 
  78         tp->packets_out += tcp_skb_pcount(skb);
  79         if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  80                 tcp_rearm_rto(sk);
  81 
  82         NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  83                       tcp_skb_pcount(skb));
  84 }
  85 
  86 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  87  * window scaling factor due to loss of precision.
  88  * If window has been shrunk, what should we make? It is not clear at all.
  89  * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  90  * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  91  * invalid. OK, let's make this for now:
  92  */
  93 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  94 {
  95         const struct tcp_sock *tp = tcp_sk(sk);
  96 
  97         if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
  98             (tp->rx_opt.wscale_ok &&
  99              ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
 100                 return tp->snd_nxt;
 101         else
 102                 return tcp_wnd_end(tp);
 103 }
 104 
 105 /* Calculate mss to advertise in SYN segment.
 106  * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 107  *
 108  * 1. It is independent of path mtu.
 109  * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 110  * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 111  *    attached devices, because some buggy hosts are confused by
 112  *    large MSS.
 113  * 4. We do not make 3, we advertise MSS, calculated from first
 114  *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 115  *    This may be overridden via information stored in routing table.
 116  * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 117  *    probably even Jumbo".
 118  */
 119 static __u16 tcp_advertise_mss(struct sock *sk)
 120 {
 121         struct tcp_sock *tp = tcp_sk(sk);
 122         const struct dst_entry *dst = __sk_dst_get(sk);
 123         int mss = tp->advmss;
 124 
 125         if (dst) {
 126                 unsigned int metric = dst_metric_advmss(dst);
 127 
 128                 if (metric < mss) {
 129                         mss = metric;
 130                         tp->advmss = mss;
 131                 }
 132         }
 133 
 134         return (__u16)mss;
 135 }
 136 
 137 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 138  * This is the first part of cwnd validation mechanism.
 139  */
 140 void tcp_cwnd_restart(struct sock *sk, s32 delta)
 141 {
 142         struct tcp_sock *tp = tcp_sk(sk);
 143         u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
 144         u32 cwnd = tp->snd_cwnd;
 145 
 146         tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 147 
 148         tp->snd_ssthresh = tcp_current_ssthresh(sk);
 149         restart_cwnd = min(restart_cwnd, cwnd);
 150 
 151         while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 152                 cwnd >>= 1;
 153         tp->snd_cwnd = max(cwnd, restart_cwnd);
 154         tp->snd_cwnd_stamp = tcp_jiffies32;
 155         tp->snd_cwnd_used = 0;
 156 }
 157 
 158 /* Congestion state accounting after a packet has been sent. */
 159 static void tcp_event_data_sent(struct tcp_sock *tp,
 160                                 struct sock *sk)
 161 {
 162         struct inet_connection_sock *icsk = inet_csk(sk);
 163         const u32 now = tcp_jiffies32;
 164 
 165         if (tcp_packets_in_flight(tp) == 0)
 166                 tcp_ca_event(sk, CA_EVENT_TX_START);
 167 
 168         /* If this is the first data packet sent in response to the
 169          * previous received data,
 170          * and it is a reply for ato after last received packet,
 171          * increase pingpong count.
 172          */
 173         if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
 174             (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 175                 inet_csk_inc_pingpong_cnt(sk);
 176 
 177         tp->lsndtime = now;
 178 }
 179 
 180 /* Account for an ACK we sent. */
 181 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
 182                                       u32 rcv_nxt)
 183 {
 184         struct tcp_sock *tp = tcp_sk(sk);
 185 
 186         if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
 187                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
 188                               tp->compressed_ack - TCP_FASTRETRANS_THRESH);
 189                 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
 190                 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
 191                         __sock_put(sk);
 192         }
 193 
 194         if (unlikely(rcv_nxt != tp->rcv_nxt))
 195                 return;  /* Special ACK sent by DCTCP to reflect ECN */
 196         tcp_dec_quickack_mode(sk, pkts);
 197         inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 198 }
 199 
 200 /* Determine a window scaling and initial window to offer.
 201  * Based on the assumption that the given amount of space
 202  * will be offered. Store the results in the tp structure.
 203  * NOTE: for smooth operation initial space offering should
 204  * be a multiple of mss if possible. We assume here that mss >= 1.
 205  * This MUST be enforced by all callers.
 206  */
 207 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
 208                                __u32 *rcv_wnd, __u32 *window_clamp,
 209                                int wscale_ok, __u8 *rcv_wscale,
 210                                __u32 init_rcv_wnd)
 211 {
 212         unsigned int space = (__space < 0 ? 0 : __space);
 213 
 214         /* If no clamp set the clamp to the max possible scaled window */
 215         if (*window_clamp == 0)
 216                 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
 217         space = min(*window_clamp, space);
 218 
 219         /* Quantize space offering to a multiple of mss if possible. */
 220         if (space > mss)
 221                 space = rounddown(space, mss);
 222 
 223         /* NOTE: offering an initial window larger than 32767
 224          * will break some buggy TCP stacks. If the admin tells us
 225          * it is likely we could be speaking with such a buggy stack
 226          * we will truncate our initial window offering to 32K-1
 227          * unless the remote has sent us a window scaling option,
 228          * which we interpret as a sign the remote TCP is not
 229          * misinterpreting the window field as a signed quantity.
 230          */
 231         if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 232                 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 233         else
 234                 (*rcv_wnd) = min_t(u32, space, U16_MAX);
 235 
 236         if (init_rcv_wnd)
 237                 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 238 
 239         *rcv_wscale = 0;
 240         if (wscale_ok) {
 241                 /* Set window scaling on max possible window */
 242                 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
 243                 space = max_t(u32, space, sysctl_rmem_max);
 244                 space = min_t(u32, space, *window_clamp);
 245                 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
 246                                       0, TCP_MAX_WSCALE);
 247         }
 248         /* Set the clamp no higher than max representable value */
 249         (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
 250 }
 251 EXPORT_SYMBOL(tcp_select_initial_window);
 252 
 253 /* Chose a new window to advertise, update state in tcp_sock for the
 254  * socket, and return result with RFC1323 scaling applied.  The return
 255  * value can be stuffed directly into th->window for an outgoing
 256  * frame.
 257  */
 258 static u16 tcp_select_window(struct sock *sk)
 259 {
 260         struct tcp_sock *tp = tcp_sk(sk);
 261         u32 old_win = tp->rcv_wnd;
 262         u32 cur_win = tcp_receive_window(tp);
 263         u32 new_win = __tcp_select_window(sk);
 264 
 265         /* Never shrink the offered window */
 266         if (new_win < cur_win) {
 267                 /* Danger Will Robinson!
 268                  * Don't update rcv_wup/rcv_wnd here or else
 269                  * we will not be able to advertise a zero
 270                  * window in time.  --DaveM
 271                  *
 272                  * Relax Will Robinson.
 273                  */
 274                 if (new_win == 0)
 275                         NET_INC_STATS(sock_net(sk),
 276                                       LINUX_MIB_TCPWANTZEROWINDOWADV);
 277                 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 278         }
 279         tp->rcv_wnd = new_win;
 280         tp->rcv_wup = tp->rcv_nxt;
 281 
 282         /* Make sure we do not exceed the maximum possible
 283          * scaled window.
 284          */
 285         if (!tp->rx_opt.rcv_wscale &&
 286             sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 287                 new_win = min(new_win, MAX_TCP_WINDOW);
 288         else
 289                 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 290 
 291         /* RFC1323 scaling applied */
 292         new_win >>= tp->rx_opt.rcv_wscale;
 293 
 294         /* If we advertise zero window, disable fast path. */
 295         if (new_win == 0) {
 296                 tp->pred_flags = 0;
 297                 if (old_win)
 298                         NET_INC_STATS(sock_net(sk),
 299                                       LINUX_MIB_TCPTOZEROWINDOWADV);
 300         } else if (old_win == 0) {
 301                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
 302         }
 303 
 304         return new_win;
 305 }
 306 
 307 /* Packet ECN state for a SYN-ACK */
 308 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
 309 {
 310         const struct tcp_sock *tp = tcp_sk(sk);
 311 
 312         TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 313         if (!(tp->ecn_flags & TCP_ECN_OK))
 314                 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 315         else if (tcp_ca_needs_ecn(sk) ||
 316                  tcp_bpf_ca_needs_ecn(sk))
 317                 INET_ECN_xmit(sk);
 318 }
 319 
 320 /* Packet ECN state for a SYN.  */
 321 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
 322 {
 323         struct tcp_sock *tp = tcp_sk(sk);
 324         bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
 325         bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
 326                 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
 327 
 328         if (!use_ecn) {
 329                 const struct dst_entry *dst = __sk_dst_get(sk);
 330 
 331                 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
 332                         use_ecn = true;
 333         }
 334 
 335         tp->ecn_flags = 0;
 336 
 337         if (use_ecn) {
 338                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 339                 tp->ecn_flags = TCP_ECN_OK;
 340                 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
 341                         INET_ECN_xmit(sk);
 342         }
 343 }
 344 
 345 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
 346 {
 347         if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
 348                 /* tp->ecn_flags are cleared at a later point in time when
 349                  * SYN ACK is ultimatively being received.
 350                  */
 351                 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
 352 }
 353 
 354 static void
 355 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
 356 {
 357         if (inet_rsk(req)->ecn_ok)
 358                 th->ece = 1;
 359 }
 360 
 361 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 362  * be sent.
 363  */
 364 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
 365                          struct tcphdr *th, int tcp_header_len)
 366 {
 367         struct tcp_sock *tp = tcp_sk(sk);
 368 
 369         if (tp->ecn_flags & TCP_ECN_OK) {
 370                 /* Not-retransmitted data segment: set ECT and inject CWR. */
 371                 if (skb->len != tcp_header_len &&
 372                     !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 373                         INET_ECN_xmit(sk);
 374                         if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 375                                 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 376                                 th->cwr = 1;
 377                                 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 378                         }
 379                 } else if (!tcp_ca_needs_ecn(sk)) {
 380                         /* ACK or retransmitted segment: clear ECT|CE */
 381                         INET_ECN_dontxmit(sk);
 382                 }
 383                 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 384                         th->ece = 1;
 385         }
 386 }
 387 
 388 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
 389  * auto increment end seqno.
 390  */
 391 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 392 {
 393         skb->ip_summed = CHECKSUM_PARTIAL;
 394 
 395         TCP_SKB_CB(skb)->tcp_flags = flags;
 396         TCP_SKB_CB(skb)->sacked = 0;
 397 
 398         tcp_skb_pcount_set(skb, 1);
 399 
 400         TCP_SKB_CB(skb)->seq = seq;
 401         if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 402                 seq++;
 403         TCP_SKB_CB(skb)->end_seq = seq;
 404 }
 405 
 406 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 407 {
 408         return tp->snd_una != tp->snd_up;
 409 }
 410 
 411 #define OPTION_SACK_ADVERTISE   (1 << 0)
 412 #define OPTION_TS               (1 << 1)
 413 #define OPTION_MD5              (1 << 2)
 414 #define OPTION_WSCALE           (1 << 3)
 415 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
 416 #define OPTION_SMC              (1 << 9)
 417 
 418 static void smc_options_write(__be32 *ptr, u16 *options)
 419 {
 420 #if IS_ENABLED(CONFIG_SMC)
 421         if (static_branch_unlikely(&tcp_have_smc)) {
 422                 if (unlikely(OPTION_SMC & *options)) {
 423                         *ptr++ = htonl((TCPOPT_NOP  << 24) |
 424                                        (TCPOPT_NOP  << 16) |
 425                                        (TCPOPT_EXP <<  8) |
 426                                        (TCPOLEN_EXP_SMC_BASE));
 427                         *ptr++ = htonl(TCPOPT_SMC_MAGIC);
 428                 }
 429         }
 430 #endif
 431 }
 432 
 433 struct tcp_out_options {
 434         u16 options;            /* bit field of OPTION_* */
 435         u16 mss;                /* 0 to disable */
 436         u8 ws;                  /* window scale, 0 to disable */
 437         u8 num_sack_blocks;     /* number of SACK blocks to include */
 438         u8 hash_size;           /* bytes in hash_location */
 439         __u8 *hash_location;    /* temporary pointer, overloaded */
 440         __u32 tsval, tsecr;     /* need to include OPTION_TS */
 441         struct tcp_fastopen_cookie *fastopen_cookie;    /* Fast open cookie */
 442 };
 443 
 444 /* Write previously computed TCP options to the packet.
 445  *
 446  * Beware: Something in the Internet is very sensitive to the ordering of
 447  * TCP options, we learned this through the hard way, so be careful here.
 448  * Luckily we can at least blame others for their non-compliance but from
 449  * inter-operability perspective it seems that we're somewhat stuck with
 450  * the ordering which we have been using if we want to keep working with
 451  * those broken things (not that it currently hurts anybody as there isn't
 452  * particular reason why the ordering would need to be changed).
 453  *
 454  * At least SACK_PERM as the first option is known to lead to a disaster
 455  * (but it may well be that other scenarios fail similarly).
 456  */
 457 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 458                               struct tcp_out_options *opts)
 459 {
 460         u16 options = opts->options;    /* mungable copy */
 461 
 462         if (unlikely(OPTION_MD5 & options)) {
 463                 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 464                                (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
 465                 /* overload cookie hash location */
 466                 opts->hash_location = (__u8 *)ptr;
 467                 ptr += 4;
 468         }
 469 
 470         if (unlikely(opts->mss)) {
 471                 *ptr++ = htonl((TCPOPT_MSS << 24) |
 472                                (TCPOLEN_MSS << 16) |
 473                                opts->mss);
 474         }
 475 
 476         if (likely(OPTION_TS & options)) {
 477                 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 478                         *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 479                                        (TCPOLEN_SACK_PERM << 16) |
 480                                        (TCPOPT_TIMESTAMP << 8) |
 481                                        TCPOLEN_TIMESTAMP);
 482                         options &= ~OPTION_SACK_ADVERTISE;
 483                 } else {
 484                         *ptr++ = htonl((TCPOPT_NOP << 24) |
 485                                        (TCPOPT_NOP << 16) |
 486                                        (TCPOPT_TIMESTAMP << 8) |
 487                                        TCPOLEN_TIMESTAMP);
 488                 }
 489                 *ptr++ = htonl(opts->tsval);
 490                 *ptr++ = htonl(opts->tsecr);
 491         }
 492 
 493         if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 494                 *ptr++ = htonl((TCPOPT_NOP << 24) |
 495                                (TCPOPT_NOP << 16) |
 496                                (TCPOPT_SACK_PERM << 8) |
 497                                TCPOLEN_SACK_PERM);
 498         }
 499 
 500         if (unlikely(OPTION_WSCALE & options)) {
 501                 *ptr++ = htonl((TCPOPT_NOP << 24) |
 502                                (TCPOPT_WINDOW << 16) |
 503                                (TCPOLEN_WINDOW << 8) |
 504                                opts->ws);
 505         }
 506 
 507         if (unlikely(opts->num_sack_blocks)) {
 508                 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 509                         tp->duplicate_sack : tp->selective_acks;
 510                 int this_sack;
 511 
 512                 *ptr++ = htonl((TCPOPT_NOP  << 24) |
 513                                (TCPOPT_NOP  << 16) |
 514                                (TCPOPT_SACK <<  8) |
 515                                (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 516                                                      TCPOLEN_SACK_PERBLOCK)));
 517 
 518                 for (this_sack = 0; this_sack < opts->num_sack_blocks;
 519                      ++this_sack) {
 520                         *ptr++ = htonl(sp[this_sack].start_seq);
 521                         *ptr++ = htonl(sp[this_sack].end_seq);
 522                 }
 523 
 524                 tp->rx_opt.dsack = 0;
 525         }
 526 
 527         if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 528                 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 529                 u8 *p = (u8 *)ptr;
 530                 u32 len; /* Fast Open option length */
 531 
 532                 if (foc->exp) {
 533                         len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 534                         *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
 535                                      TCPOPT_FASTOPEN_MAGIC);
 536                         p += TCPOLEN_EXP_FASTOPEN_BASE;
 537                 } else {
 538                         len = TCPOLEN_FASTOPEN_BASE + foc->len;
 539                         *p++ = TCPOPT_FASTOPEN;
 540                         *p++ = len;
 541                 }
 542 
 543                 memcpy(p, foc->val, foc->len);
 544                 if ((len & 3) == 2) {
 545                         p[foc->len] = TCPOPT_NOP;
 546                         p[foc->len + 1] = TCPOPT_NOP;
 547                 }
 548                 ptr += (len + 3) >> 2;
 549         }
 550 
 551         smc_options_write(ptr, &options);
 552 }
 553 
 554 static void smc_set_option(const struct tcp_sock *tp,
 555                            struct tcp_out_options *opts,
 556                            unsigned int *remaining)
 557 {
 558 #if IS_ENABLED(CONFIG_SMC)
 559         if (static_branch_unlikely(&tcp_have_smc)) {
 560                 if (tp->syn_smc) {
 561                         if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 562                                 opts->options |= OPTION_SMC;
 563                                 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 564                         }
 565                 }
 566         }
 567 #endif
 568 }
 569 
 570 static void smc_set_option_cond(const struct tcp_sock *tp,
 571                                 const struct inet_request_sock *ireq,
 572                                 struct tcp_out_options *opts,
 573                                 unsigned int *remaining)
 574 {
 575 #if IS_ENABLED(CONFIG_SMC)
 576         if (static_branch_unlikely(&tcp_have_smc)) {
 577                 if (tp->syn_smc && ireq->smc_ok) {
 578                         if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 579                                 opts->options |= OPTION_SMC;
 580                                 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 581                         }
 582                 }
 583         }
 584 #endif
 585 }
 586 
 587 /* Compute TCP options for SYN packets. This is not the final
 588  * network wire format yet.
 589  */
 590 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 591                                 struct tcp_out_options *opts,
 592                                 struct tcp_md5sig_key **md5)
 593 {
 594         struct tcp_sock *tp = tcp_sk(sk);
 595         unsigned int remaining = MAX_TCP_OPTION_SPACE;
 596         struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 597 
 598         *md5 = NULL;
 599 #ifdef CONFIG_TCP_MD5SIG
 600         if (static_branch_unlikely(&tcp_md5_needed) &&
 601             rcu_access_pointer(tp->md5sig_info)) {
 602                 *md5 = tp->af_specific->md5_lookup(sk, sk);
 603                 if (*md5) {
 604                         opts->options |= OPTION_MD5;
 605                         remaining -= TCPOLEN_MD5SIG_ALIGNED;
 606                 }
 607         }
 608 #endif
 609 
 610         /* We always get an MSS option.  The option bytes which will be seen in
 611          * normal data packets should timestamps be used, must be in the MSS
 612          * advertised.  But we subtract them from tp->mss_cache so that
 613          * calculations in tcp_sendmsg are simpler etc.  So account for this
 614          * fact here if necessary.  If we don't do this correctly, as a
 615          * receiver we won't recognize data packets as being full sized when we
 616          * should, and thus we won't abide by the delayed ACK rules correctly.
 617          * SACKs don't matter, we never delay an ACK when we have any of those
 618          * going out.  */
 619         opts->mss = tcp_advertise_mss(sk);
 620         remaining -= TCPOLEN_MSS_ALIGNED;
 621 
 622         if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
 623                 opts->options |= OPTION_TS;
 624                 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
 625                 opts->tsecr = tp->rx_opt.ts_recent;
 626                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
 627         }
 628         if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
 629                 opts->ws = tp->rx_opt.rcv_wscale;
 630                 opts->options |= OPTION_WSCALE;
 631                 remaining -= TCPOLEN_WSCALE_ALIGNED;
 632         }
 633         if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
 634                 opts->options |= OPTION_SACK_ADVERTISE;
 635                 if (unlikely(!(OPTION_TS & opts->options)))
 636                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
 637         }
 638 
 639         if (fastopen && fastopen->cookie.len >= 0) {
 640                 u32 need = fastopen->cookie.len;
 641 
 642                 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 643                                                TCPOLEN_FASTOPEN_BASE;
 644                 need = (need + 3) & ~3U;  /* Align to 32 bits */
 645                 if (remaining >= need) {
 646                         opts->options |= OPTION_FAST_OPEN_COOKIE;
 647                         opts->fastopen_cookie = &fastopen->cookie;
 648                         remaining -= need;
 649                         tp->syn_fastopen = 1;
 650                         tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
 651                 }
 652         }
 653 
 654         smc_set_option(tp, opts, &remaining);
 655 
 656         return MAX_TCP_OPTION_SPACE - remaining;
 657 }
 658 
 659 /* Set up TCP options for SYN-ACKs. */
 660 static unsigned int tcp_synack_options(const struct sock *sk,
 661                                        struct request_sock *req,
 662                                        unsigned int mss, struct sk_buff *skb,
 663                                        struct tcp_out_options *opts,
 664                                        const struct tcp_md5sig_key *md5,
 665                                        struct tcp_fastopen_cookie *foc)
 666 {
 667         struct inet_request_sock *ireq = inet_rsk(req);
 668         unsigned int remaining = MAX_TCP_OPTION_SPACE;
 669 
 670 #ifdef CONFIG_TCP_MD5SIG
 671         if (md5) {
 672                 opts->options |= OPTION_MD5;
 673                 remaining -= TCPOLEN_MD5SIG_ALIGNED;
 674 
 675                 /* We can't fit any SACK blocks in a packet with MD5 + TS
 676                  * options. There was discussion about disabling SACK
 677                  * rather than TS in order to fit in better with old,
 678                  * buggy kernels, but that was deemed to be unnecessary.
 679                  */
 680                 ireq->tstamp_ok &= !ireq->sack_ok;
 681         }
 682 #endif
 683 
 684         /* We always send an MSS option. */
 685         opts->mss = mss;
 686         remaining -= TCPOLEN_MSS_ALIGNED;
 687 
 688         if (likely(ireq->wscale_ok)) {
 689                 opts->ws = ireq->rcv_wscale;
 690                 opts->options |= OPTION_WSCALE;
 691                 remaining -= TCPOLEN_WSCALE_ALIGNED;
 692         }
 693         if (likely(ireq->tstamp_ok)) {
 694                 opts->options |= OPTION_TS;
 695                 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
 696                 opts->tsecr = req->ts_recent;
 697                 remaining -= TCPOLEN_TSTAMP_ALIGNED;
 698         }
 699         if (likely(ireq->sack_ok)) {
 700                 opts->options |= OPTION_SACK_ADVERTISE;
 701                 if (unlikely(!ireq->tstamp_ok))
 702                         remaining -= TCPOLEN_SACKPERM_ALIGNED;
 703         }
 704         if (foc != NULL && foc->len >= 0) {
 705                 u32 need = foc->len;
 706 
 707                 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 708                                    TCPOLEN_FASTOPEN_BASE;
 709                 need = (need + 3) & ~3U;  /* Align to 32 bits */
 710                 if (remaining >= need) {
 711                         opts->options |= OPTION_FAST_OPEN_COOKIE;
 712                         opts->fastopen_cookie = foc;
 713                         remaining -= need;
 714                 }
 715         }
 716 
 717         smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
 718 
 719         return MAX_TCP_OPTION_SPACE - remaining;
 720 }
 721 
 722 /* Compute TCP options for ESTABLISHED sockets. This is not the
 723  * final wire format yet.
 724  */
 725 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 726                                         struct tcp_out_options *opts,
 727                                         struct tcp_md5sig_key **md5)
 728 {
 729         struct tcp_sock *tp = tcp_sk(sk);
 730         unsigned int size = 0;
 731         unsigned int eff_sacks;
 732 
 733         opts->options = 0;
 734 
 735         *md5 = NULL;
 736 #ifdef CONFIG_TCP_MD5SIG
 737         if (static_branch_unlikely(&tcp_md5_needed) &&
 738             rcu_access_pointer(tp->md5sig_info)) {
 739                 *md5 = tp->af_specific->md5_lookup(sk, sk);
 740                 if (*md5) {
 741                         opts->options |= OPTION_MD5;
 742                         size += TCPOLEN_MD5SIG_ALIGNED;
 743                 }
 744         }
 745 #endif
 746 
 747         if (likely(tp->rx_opt.tstamp_ok)) {
 748                 opts->options |= OPTION_TS;
 749                 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
 750                 opts->tsecr = tp->rx_opt.ts_recent;
 751                 size += TCPOLEN_TSTAMP_ALIGNED;
 752         }
 753 
 754         eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 755         if (unlikely(eff_sacks)) {
 756                 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 757                 opts->num_sack_blocks =
 758                         min_t(unsigned int, eff_sacks,
 759                               (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 760                               TCPOLEN_SACK_PERBLOCK);
 761                 if (likely(opts->num_sack_blocks))
 762                         size += TCPOLEN_SACK_BASE_ALIGNED +
 763                                 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 764         }
 765 
 766         return size;
 767 }
 768 
 769 
 770 /* TCP SMALL QUEUES (TSQ)
 771  *
 772  * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
 773  * to reduce RTT and bufferbloat.
 774  * We do this using a special skb destructor (tcp_wfree).
 775  *
 776  * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
 777  * needs to be reallocated in a driver.
 778  * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
 779  *
 780  * Since transmit from skb destructor is forbidden, we use a tasklet
 781  * to process all sockets that eventually need to send more skbs.
 782  * We use one tasklet per cpu, with its own queue of sockets.
 783  */
 784 struct tsq_tasklet {
 785         struct tasklet_struct   tasklet;
 786         struct list_head        head; /* queue of tcp sockets */
 787 };
 788 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
 789 
 790 static void tcp_tsq_write(struct sock *sk)
 791 {
 792         if ((1 << sk->sk_state) &
 793             (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
 794              TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
 795                 struct tcp_sock *tp = tcp_sk(sk);
 796 
 797                 if (tp->lost_out > tp->retrans_out &&
 798                     tp->snd_cwnd > tcp_packets_in_flight(tp)) {
 799                         tcp_mstamp_refresh(tp);
 800                         tcp_xmit_retransmit_queue(sk);
 801                 }
 802 
 803                 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
 804                                0, GFP_ATOMIC);
 805         }
 806 }
 807 
 808 static void tcp_tsq_handler(struct sock *sk)
 809 {
 810         bh_lock_sock(sk);
 811         if (!sock_owned_by_user(sk))
 812                 tcp_tsq_write(sk);
 813         else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
 814                 sock_hold(sk);
 815         bh_unlock_sock(sk);
 816 }
 817 /*
 818  * One tasklet per cpu tries to send more skbs.
 819  * We run in tasklet context but need to disable irqs when
 820  * transferring tsq->head because tcp_wfree() might
 821  * interrupt us (non NAPI drivers)
 822  */
 823 static void tcp_tasklet_func(unsigned long data)
 824 {
 825         struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
 826         LIST_HEAD(list);
 827         unsigned long flags;
 828         struct list_head *q, *n;
 829         struct tcp_sock *tp;
 830         struct sock *sk;
 831 
 832         local_irq_save(flags);
 833         list_splice_init(&tsq->head, &list);
 834         local_irq_restore(flags);
 835 
 836         list_for_each_safe(q, n, &list) {
 837                 tp = list_entry(q, struct tcp_sock, tsq_node);
 838                 list_del(&tp->tsq_node);
 839 
 840                 sk = (struct sock *)tp;
 841                 smp_mb__before_atomic();
 842                 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
 843 
 844                 tcp_tsq_handler(sk);
 845                 sk_free(sk);
 846         }
 847 }
 848 
 849 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |           \
 850                           TCPF_WRITE_TIMER_DEFERRED |   \
 851                           TCPF_DELACK_TIMER_DEFERRED |  \
 852                           TCPF_MTU_REDUCED_DEFERRED)
 853 /**
 854  * tcp_release_cb - tcp release_sock() callback
 855  * @sk: socket
 856  *
 857  * called from release_sock() to perform protocol dependent
 858  * actions before socket release.
 859  */
 860 void tcp_release_cb(struct sock *sk)
 861 {
 862         unsigned long flags, nflags;
 863 
 864         /* perform an atomic operation only if at least one flag is set */
 865         do {
 866                 flags = sk->sk_tsq_flags;
 867                 if (!(flags & TCP_DEFERRED_ALL))
 868                         return;
 869                 nflags = flags & ~TCP_DEFERRED_ALL;
 870         } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
 871 
 872         if (flags & TCPF_TSQ_DEFERRED) {
 873                 tcp_tsq_write(sk);
 874                 __sock_put(sk);
 875         }
 876         /* Here begins the tricky part :
 877          * We are called from release_sock() with :
 878          * 1) BH disabled
 879          * 2) sk_lock.slock spinlock held
 880          * 3) socket owned by us (sk->sk_lock.owned == 1)
 881          *
 882          * But following code is meant to be called from BH handlers,
 883          * so we should keep BH disabled, but early release socket ownership
 884          */
 885         sock_release_ownership(sk);
 886 
 887         if (flags & TCPF_WRITE_TIMER_DEFERRED) {
 888                 tcp_write_timer_handler(sk);
 889                 __sock_put(sk);
 890         }
 891         if (flags & TCPF_DELACK_TIMER_DEFERRED) {
 892                 tcp_delack_timer_handler(sk);
 893                 __sock_put(sk);
 894         }
 895         if (flags & TCPF_MTU_REDUCED_DEFERRED) {
 896                 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
 897                 __sock_put(sk);
 898         }
 899 }
 900 EXPORT_SYMBOL(tcp_release_cb);
 901 
 902 void __init tcp_tasklet_init(void)
 903 {
 904         int i;
 905 
 906         for_each_possible_cpu(i) {
 907                 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
 908 
 909                 INIT_LIST_HEAD(&tsq->head);
 910                 tasklet_init(&tsq->tasklet,
 911                              tcp_tasklet_func,
 912                              (unsigned long)tsq);
 913         }
 914 }
 915 
 916 /*
 917  * Write buffer destructor automatically called from kfree_skb.
 918  * We can't xmit new skbs from this context, as we might already
 919  * hold qdisc lock.
 920  */
 921 void tcp_wfree(struct sk_buff *skb)
 922 {
 923         struct sock *sk = skb->sk;
 924         struct tcp_sock *tp = tcp_sk(sk);
 925         unsigned long flags, nval, oval;
 926 
 927         /* Keep one reference on sk_wmem_alloc.
 928          * Will be released by sk_free() from here or tcp_tasklet_func()
 929          */
 930         WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
 931 
 932         /* If this softirq is serviced by ksoftirqd, we are likely under stress.
 933          * Wait until our queues (qdisc + devices) are drained.
 934          * This gives :
 935          * - less callbacks to tcp_write_xmit(), reducing stress (batches)
 936          * - chance for incoming ACK (processed by another cpu maybe)
 937          *   to migrate this flow (skb->ooo_okay will be eventually set)
 938          */
 939         if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
 940                 goto out;
 941 
 942         for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
 943                 struct tsq_tasklet *tsq;
 944                 bool empty;
 945 
 946                 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
 947                         goto out;
 948 
 949                 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
 950                 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
 951                 if (nval != oval)
 952                         continue;
 953 
 954                 /* queue this socket to tasklet queue */
 955                 local_irq_save(flags);
 956                 tsq = this_cpu_ptr(&tsq_tasklet);
 957                 empty = list_empty(&tsq->head);
 958                 list_add(&tp->tsq_node, &tsq->head);
 959                 if (empty)
 960                         tasklet_schedule(&tsq->tasklet);
 961                 local_irq_restore(flags);
 962                 return;
 963         }
 964 out:
 965         sk_free(sk);
 966 }
 967 
 968 /* Note: Called under soft irq.
 969  * We can call TCP stack right away, unless socket is owned by user.
 970  */
 971 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
 972 {
 973         struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
 974         struct sock *sk = (struct sock *)tp;
 975 
 976         tcp_tsq_handler(sk);
 977         sock_put(sk);
 978 
 979         return HRTIMER_NORESTART;
 980 }
 981 
 982 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
 983                                       u64 prior_wstamp)
 984 {
 985         struct tcp_sock *tp = tcp_sk(sk);
 986 
 987         if (sk->sk_pacing_status != SK_PACING_NONE) {
 988                 unsigned long rate = sk->sk_pacing_rate;
 989 
 990                 /* Original sch_fq does not pace first 10 MSS
 991                  * Note that tp->data_segs_out overflows after 2^32 packets,
 992                  * this is a minor annoyance.
 993                  */
 994                 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
 995                         u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
 996                         u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
 997 
 998                         /* take into account OS jitter */
 999                         len_ns -= min_t(u64, len_ns / 2, credit);
1000                         tp->tcp_wstamp_ns += len_ns;
1001                 }
1002         }
1003         list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1004 }
1005 
1006 /* This routine actually transmits TCP packets queued in by
1007  * tcp_do_sendmsg().  This is used by both the initial
1008  * transmission and possible later retransmissions.
1009  * All SKB's seen here are completely headerless.  It is our
1010  * job to build the TCP header, and pass the packet down to
1011  * IP so it can do the same plus pass the packet off to the
1012  * device.
1013  *
1014  * We are working here with either a clone of the original
1015  * SKB, or a fresh unique copy made by the retransmit engine.
1016  */
1017 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1018                               int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1019 {
1020         const struct inet_connection_sock *icsk = inet_csk(sk);
1021         struct inet_sock *inet;
1022         struct tcp_sock *tp;
1023         struct tcp_skb_cb *tcb;
1024         struct tcp_out_options opts;
1025         unsigned int tcp_options_size, tcp_header_size;
1026         struct sk_buff *oskb = NULL;
1027         struct tcp_md5sig_key *md5;
1028         struct tcphdr *th;
1029         u64 prior_wstamp;
1030         int err;
1031 
1032         BUG_ON(!skb || !tcp_skb_pcount(skb));
1033         tp = tcp_sk(sk);
1034         prior_wstamp = tp->tcp_wstamp_ns;
1035         tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1036         skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1037         if (clone_it) {
1038                 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1039                         - tp->snd_una;
1040                 oskb = skb;
1041 
1042                 tcp_skb_tsorted_save(oskb) {
1043                         if (unlikely(skb_cloned(oskb)))
1044                                 skb = pskb_copy(oskb, gfp_mask);
1045                         else
1046                                 skb = skb_clone(oskb, gfp_mask);
1047                 } tcp_skb_tsorted_restore(oskb);
1048 
1049                 if (unlikely(!skb))
1050                         return -ENOBUFS;
1051                 /* retransmit skbs might have a non zero value in skb->dev
1052                  * because skb->dev is aliased with skb->rbnode.rb_left
1053                  */
1054                 skb->dev = NULL;
1055         }
1056 
1057         inet = inet_sk(sk);
1058         tcb = TCP_SKB_CB(skb);
1059         memset(&opts, 0, sizeof(opts));
1060 
1061         if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1062                 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1063         } else {
1064                 tcp_options_size = tcp_established_options(sk, skb, &opts,
1065                                                            &md5);
1066                 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1067                  * at receiver : This slightly improve GRO performance.
1068                  * Note that we do not force the PSH flag for non GSO packets,
1069                  * because they might be sent under high congestion events,
1070                  * and in this case it is better to delay the delivery of 1-MSS
1071                  * packets and thus the corresponding ACK packet that would
1072                  * release the following packet.
1073                  */
1074                 if (tcp_skb_pcount(skb) > 1)
1075                         tcb->tcp_flags |= TCPHDR_PSH;
1076         }
1077         tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1078 
1079         /* if no packet is in qdisc/device queue, then allow XPS to select
1080          * another queue. We can be called from tcp_tsq_handler()
1081          * which holds one reference to sk.
1082          *
1083          * TODO: Ideally, in-flight pure ACK packets should not matter here.
1084          * One way to get this would be to set skb->truesize = 2 on them.
1085          */
1086         skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1087 
1088         /* If we had to use memory reserve to allocate this skb,
1089          * this might cause drops if packet is looped back :
1090          * Other socket might not have SOCK_MEMALLOC.
1091          * Packets not looped back do not care about pfmemalloc.
1092          */
1093         skb->pfmemalloc = 0;
1094 
1095         skb_push(skb, tcp_header_size);
1096         skb_reset_transport_header(skb);
1097 
1098         skb_orphan(skb);
1099         skb->sk = sk;
1100         skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1101         skb_set_hash_from_sk(skb, sk);
1102         refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1103 
1104         skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1105 
1106         /* Build TCP header and checksum it. */
1107         th = (struct tcphdr *)skb->data;
1108         th->source              = inet->inet_sport;
1109         th->dest                = inet->inet_dport;
1110         th->seq                 = htonl(tcb->seq);
1111         th->ack_seq             = htonl(rcv_nxt);
1112         *(((__be16 *)th) + 6)   = htons(((tcp_header_size >> 2) << 12) |
1113                                         tcb->tcp_flags);
1114 
1115         th->check               = 0;
1116         th->urg_ptr             = 0;
1117 
1118         /* The urg_mode check is necessary during a below snd_una win probe */
1119         if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1120                 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1121                         th->urg_ptr = htons(tp->snd_up - tcb->seq);
1122                         th->urg = 1;
1123                 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1124                         th->urg_ptr = htons(0xFFFF);
1125                         th->urg = 1;
1126                 }
1127         }
1128 
1129         tcp_options_write((__be32 *)(th + 1), tp, &opts);
1130         skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1131         if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1132                 th->window      = htons(tcp_select_window(sk));
1133                 tcp_ecn_send(sk, skb, th, tcp_header_size);
1134         } else {
1135                 /* RFC1323: The window in SYN & SYN/ACK segments
1136                  * is never scaled.
1137                  */
1138                 th->window      = htons(min(tp->rcv_wnd, 65535U));
1139         }
1140 #ifdef CONFIG_TCP_MD5SIG
1141         /* Calculate the MD5 hash, as we have all we need now */
1142         if (md5) {
1143                 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1144                 tp->af_specific->calc_md5_hash(opts.hash_location,
1145                                                md5, sk, skb);
1146         }
1147 #endif
1148 
1149         icsk->icsk_af_ops->send_check(sk, skb);
1150 
1151         if (likely(tcb->tcp_flags & TCPHDR_ACK))
1152                 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1153 
1154         if (skb->len != tcp_header_size) {
1155                 tcp_event_data_sent(tp, sk);
1156                 tp->data_segs_out += tcp_skb_pcount(skb);
1157                 tp->bytes_sent += skb->len - tcp_header_size;
1158         }
1159 
1160         if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1161                 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1162                               tcp_skb_pcount(skb));
1163 
1164         tp->segs_out += tcp_skb_pcount(skb);
1165         /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1166         skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1167         skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1168 
1169         /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1170 
1171         /* Cleanup our debris for IP stacks */
1172         memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1173                                sizeof(struct inet6_skb_parm)));
1174 
1175         tcp_add_tx_delay(skb, tp);
1176 
1177         err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1178 
1179         if (unlikely(err > 0)) {
1180                 tcp_enter_cwr(sk);
1181                 err = net_xmit_eval(err);
1182         }
1183         if (!err && oskb) {
1184                 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1185                 tcp_rate_skb_sent(sk, oskb);
1186         }
1187         return err;
1188 }
1189 
1190 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1191                             gfp_t gfp_mask)
1192 {
1193         return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1194                                   tcp_sk(sk)->rcv_nxt);
1195 }
1196 
1197 /* This routine just queues the buffer for sending.
1198  *
1199  * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1200  * otherwise socket can stall.
1201  */
1202 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1203 {
1204         struct tcp_sock *tp = tcp_sk(sk);
1205 
1206         /* Advance write_seq and place onto the write_queue. */
1207         WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1208         __skb_header_release(skb);
1209         tcp_add_write_queue_tail(sk, skb);
1210         sk_wmem_queued_add(sk, skb->truesize);
1211         sk_mem_charge(sk, skb->truesize);
1212 }
1213 
1214 /* Initialize TSO segments for a packet. */
1215 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1216 {
1217         if (skb->len <= mss_now) {
1218                 /* Avoid the costly divide in the normal
1219                  * non-TSO case.
1220                  */
1221                 tcp_skb_pcount_set(skb, 1);
1222                 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1223         } else {
1224                 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1225                 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1226         }
1227 }
1228 
1229 /* Pcount in the middle of the write queue got changed, we need to do various
1230  * tweaks to fix counters
1231  */
1232 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1233 {
1234         struct tcp_sock *tp = tcp_sk(sk);
1235 
1236         tp->packets_out -= decr;
1237 
1238         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1239                 tp->sacked_out -= decr;
1240         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1241                 tp->retrans_out -= decr;
1242         if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1243                 tp->lost_out -= decr;
1244 
1245         /* Reno case is special. Sigh... */
1246         if (tcp_is_reno(tp) && decr > 0)
1247                 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1248 
1249         if (tp->lost_skb_hint &&
1250             before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1251             (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1252                 tp->lost_cnt_hint -= decr;
1253 
1254         tcp_verify_left_out(tp);
1255 }
1256 
1257 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1258 {
1259         return TCP_SKB_CB(skb)->txstamp_ack ||
1260                 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1261 }
1262 
1263 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1264 {
1265         struct skb_shared_info *shinfo = skb_shinfo(skb);
1266 
1267         if (unlikely(tcp_has_tx_tstamp(skb)) &&
1268             !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1269                 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1270                 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1271 
1272                 shinfo->tx_flags &= ~tsflags;
1273                 shinfo2->tx_flags |= tsflags;
1274                 swap(shinfo->tskey, shinfo2->tskey);
1275                 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1276                 TCP_SKB_CB(skb)->txstamp_ack = 0;
1277         }
1278 }
1279 
1280 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1281 {
1282         TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1283         TCP_SKB_CB(skb)->eor = 0;
1284 }
1285 
1286 /* Insert buff after skb on the write or rtx queue of sk.  */
1287 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1288                                          struct sk_buff *buff,
1289                                          struct sock *sk,
1290                                          enum tcp_queue tcp_queue)
1291 {
1292         if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1293                 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1294         else
1295                 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1296 }
1297 
1298 /* Function to create two new TCP segments.  Shrinks the given segment
1299  * to the specified size and appends a new segment with the rest of the
1300  * packet to the list.  This won't be called frequently, I hope.
1301  * Remember, these are still headerless SKBs at this point.
1302  */
1303 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1304                  struct sk_buff *skb, u32 len,
1305                  unsigned int mss_now, gfp_t gfp)
1306 {
1307         struct tcp_sock *tp = tcp_sk(sk);
1308         struct sk_buff *buff;
1309         int nsize, old_factor;
1310         long limit;
1311         int nlen;
1312         u8 flags;
1313 
1314         if (WARN_ON(len > skb->len))
1315                 return -EINVAL;
1316 
1317         nsize = skb_headlen(skb) - len;
1318         if (nsize < 0)
1319                 nsize = 0;
1320 
1321         /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1322          * We need some allowance to not penalize applications setting small
1323          * SO_SNDBUF values.
1324          * Also allow first and last skb in retransmit queue to be split.
1325          */
1326         limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1327         if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1328                      tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1329                      skb != tcp_rtx_queue_head(sk) &&
1330                      skb != tcp_rtx_queue_tail(sk))) {
1331                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1332                 return -ENOMEM;
1333         }
1334 
1335         if (skb_unclone(skb, gfp))
1336                 return -ENOMEM;
1337 
1338         /* Get a new skb... force flag on. */
1339         buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1340         if (!buff)
1341                 return -ENOMEM; /* We'll just try again later. */
1342         skb_copy_decrypted(buff, skb);
1343 
1344         sk_wmem_queued_add(sk, buff->truesize);
1345         sk_mem_charge(sk, buff->truesize);
1346         nlen = skb->len - len - nsize;
1347         buff->truesize += nlen;
1348         skb->truesize -= nlen;
1349 
1350         /* Correct the sequence numbers. */
1351         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1352         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1353         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1354 
1355         /* PSH and FIN should only be set in the second packet. */
1356         flags = TCP_SKB_CB(skb)->tcp_flags;
1357         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1358         TCP_SKB_CB(buff)->tcp_flags = flags;
1359         TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1360         tcp_skb_fragment_eor(skb, buff);
1361 
1362         skb_split(skb, buff, len);
1363 
1364         buff->ip_summed = CHECKSUM_PARTIAL;
1365 
1366         buff->tstamp = skb->tstamp;
1367         tcp_fragment_tstamp(skb, buff);
1368 
1369         old_factor = tcp_skb_pcount(skb);
1370 
1371         /* Fix up tso_factor for both original and new SKB.  */
1372         tcp_set_skb_tso_segs(skb, mss_now);
1373         tcp_set_skb_tso_segs(buff, mss_now);
1374 
1375         /* Update delivered info for the new segment */
1376         TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1377 
1378         /* If this packet has been sent out already, we must
1379          * adjust the various packet counters.
1380          */
1381         if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1382                 int diff = old_factor - tcp_skb_pcount(skb) -
1383                         tcp_skb_pcount(buff);
1384 
1385                 if (diff)
1386                         tcp_adjust_pcount(sk, skb, diff);
1387         }
1388 
1389         /* Link BUFF into the send queue. */
1390         __skb_header_release(buff);
1391         tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1392         if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1393                 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1394 
1395         return 0;
1396 }
1397 
1398 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1399  * data is not copied, but immediately discarded.
1400  */
1401 static int __pskb_trim_head(struct sk_buff *skb, int len)
1402 {
1403         struct skb_shared_info *shinfo;
1404         int i, k, eat;
1405 
1406         eat = min_t(int, len, skb_headlen(skb));
1407         if (eat) {
1408                 __skb_pull(skb, eat);
1409                 len -= eat;
1410                 if (!len)
1411                         return 0;
1412         }
1413         eat = len;
1414         k = 0;
1415         shinfo = skb_shinfo(skb);
1416         for (i = 0; i < shinfo->nr_frags; i++) {
1417                 int size = skb_frag_size(&shinfo->frags[i]);
1418 
1419                 if (size <= eat) {
1420                         skb_frag_unref(skb, i);
1421                         eat -= size;
1422                 } else {
1423                         shinfo->frags[k] = shinfo->frags[i];
1424                         if (eat) {
1425                                 skb_frag_off_add(&shinfo->frags[k], eat);
1426                                 skb_frag_size_sub(&shinfo->frags[k], eat);
1427                                 eat = 0;
1428                         }
1429                         k++;
1430                 }
1431         }
1432         shinfo->nr_frags = k;
1433 
1434         skb->data_len -= len;
1435         skb->len = skb->data_len;
1436         return len;
1437 }
1438 
1439 /* Remove acked data from a packet in the transmit queue. */
1440 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1441 {
1442         u32 delta_truesize;
1443 
1444         if (skb_unclone(skb, GFP_ATOMIC))
1445                 return -ENOMEM;
1446 
1447         delta_truesize = __pskb_trim_head(skb, len);
1448 
1449         TCP_SKB_CB(skb)->seq += len;
1450         skb->ip_summed = CHECKSUM_PARTIAL;
1451 
1452         if (delta_truesize) {
1453                 skb->truesize      -= delta_truesize;
1454                 sk_wmem_queued_add(sk, -delta_truesize);
1455                 sk_mem_uncharge(sk, delta_truesize);
1456                 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1457         }
1458 
1459         /* Any change of skb->len requires recalculation of tso factor. */
1460         if (tcp_skb_pcount(skb) > 1)
1461                 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1462 
1463         return 0;
1464 }
1465 
1466 /* Calculate MSS not accounting any TCP options.  */
1467 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1468 {
1469         const struct tcp_sock *tp = tcp_sk(sk);
1470         const struct inet_connection_sock *icsk = inet_csk(sk);
1471         int mss_now;
1472 
1473         /* Calculate base mss without TCP options:
1474            It is MMS_S - sizeof(tcphdr) of rfc1122
1475          */
1476         mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1477 
1478         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1479         if (icsk->icsk_af_ops->net_frag_header_len) {
1480                 const struct dst_entry *dst = __sk_dst_get(sk);
1481 
1482                 if (dst && dst_allfrag(dst))
1483                         mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1484         }
1485 
1486         /* Clamp it (mss_clamp does not include tcp options) */
1487         if (mss_now > tp->rx_opt.mss_clamp)
1488                 mss_now = tp->rx_opt.mss_clamp;
1489 
1490         /* Now subtract optional transport overhead */
1491         mss_now -= icsk->icsk_ext_hdr_len;
1492 
1493         /* Then reserve room for full set of TCP options and 8 bytes of data */
1494         mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1495         return mss_now;
1496 }
1497 
1498 /* Calculate MSS. Not accounting for SACKs here.  */
1499 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1500 {
1501         /* Subtract TCP options size, not including SACKs */
1502         return __tcp_mtu_to_mss(sk, pmtu) -
1503                (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1504 }
1505 
1506 /* Inverse of above */
1507 int tcp_mss_to_mtu(struct sock *sk, int mss)
1508 {
1509         const struct tcp_sock *tp = tcp_sk(sk);
1510         const struct inet_connection_sock *icsk = inet_csk(sk);
1511         int mtu;
1512 
1513         mtu = mss +
1514               tp->tcp_header_len +
1515               icsk->icsk_ext_hdr_len +
1516               icsk->icsk_af_ops->net_header_len;
1517 
1518         /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1519         if (icsk->icsk_af_ops->net_frag_header_len) {
1520                 const struct dst_entry *dst = __sk_dst_get(sk);
1521 
1522                 if (dst && dst_allfrag(dst))
1523                         mtu += icsk->icsk_af_ops->net_frag_header_len;
1524         }
1525         return mtu;
1526 }
1527 EXPORT_SYMBOL(tcp_mss_to_mtu);
1528 
1529 /* MTU probing init per socket */
1530 void tcp_mtup_init(struct sock *sk)
1531 {
1532         struct tcp_sock *tp = tcp_sk(sk);
1533         struct inet_connection_sock *icsk = inet_csk(sk);
1534         struct net *net = sock_net(sk);
1535 
1536         icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1537         icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1538                                icsk->icsk_af_ops->net_header_len;
1539         icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1540         icsk->icsk_mtup.probe_size = 0;
1541         if (icsk->icsk_mtup.enabled)
1542                 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1543 }
1544 EXPORT_SYMBOL(tcp_mtup_init);
1545 
1546 /* This function synchronize snd mss to current pmtu/exthdr set.
1547 
1548    tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1549    for TCP options, but includes only bare TCP header.
1550 
1551    tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1552    It is minimum of user_mss and mss received with SYN.
1553    It also does not include TCP options.
1554 
1555    inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1556 
1557    tp->mss_cache is current effective sending mss, including
1558    all tcp options except for SACKs. It is evaluated,
1559    taking into account current pmtu, but never exceeds
1560    tp->rx_opt.mss_clamp.
1561 
1562    NOTE1. rfc1122 clearly states that advertised MSS
1563    DOES NOT include either tcp or ip options.
1564 
1565    NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1566    are READ ONLY outside this function.         --ANK (980731)
1567  */
1568 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1569 {
1570         struct tcp_sock *tp = tcp_sk(sk);
1571         struct inet_connection_sock *icsk = inet_csk(sk);
1572         int mss_now;
1573 
1574         if (icsk->icsk_mtup.search_high > pmtu)
1575                 icsk->icsk_mtup.search_high = pmtu;
1576 
1577         mss_now = tcp_mtu_to_mss(sk, pmtu);
1578         mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1579 
1580         /* And store cached results */
1581         icsk->icsk_pmtu_cookie = pmtu;
1582         if (icsk->icsk_mtup.enabled)
1583                 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1584         tp->mss_cache = mss_now;
1585 
1586         return mss_now;
1587 }
1588 EXPORT_SYMBOL(tcp_sync_mss);
1589 
1590 /* Compute the current effective MSS, taking SACKs and IP options,
1591  * and even PMTU discovery events into account.
1592  */
1593 unsigned int tcp_current_mss(struct sock *sk)
1594 {
1595         const struct tcp_sock *tp = tcp_sk(sk);
1596         const struct dst_entry *dst = __sk_dst_get(sk);
1597         u32 mss_now;
1598         unsigned int header_len;
1599         struct tcp_out_options opts;
1600         struct tcp_md5sig_key *md5;
1601 
1602         mss_now = tp->mss_cache;
1603 
1604         if (dst) {
1605                 u32 mtu = dst_mtu(dst);
1606                 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1607                         mss_now = tcp_sync_mss(sk, mtu);
1608         }
1609 
1610         header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1611                      sizeof(struct tcphdr);
1612         /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1613          * some common options. If this is an odd packet (because we have SACK
1614          * blocks etc) then our calculated header_len will be different, and
1615          * we have to adjust mss_now correspondingly */
1616         if (header_len != tp->tcp_header_len) {
1617                 int delta = (int) header_len - tp->tcp_header_len;
1618                 mss_now -= delta;
1619         }
1620 
1621         return mss_now;
1622 }
1623 
1624 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1625  * As additional protections, we do not touch cwnd in retransmission phases,
1626  * and if application hit its sndbuf limit recently.
1627  */
1628 static void tcp_cwnd_application_limited(struct sock *sk)
1629 {
1630         struct tcp_sock *tp = tcp_sk(sk);
1631 
1632         if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1633             sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1634                 /* Limited by application or receiver window. */
1635                 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1636                 u32 win_used = max(tp->snd_cwnd_used, init_win);
1637                 if (win_used < tp->snd_cwnd) {
1638                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
1639                         tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1640                 }
1641                 tp->snd_cwnd_used = 0;
1642         }
1643         tp->snd_cwnd_stamp = tcp_jiffies32;
1644 }
1645 
1646 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1647 {
1648         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1649         struct tcp_sock *tp = tcp_sk(sk);
1650 
1651         /* Track the maximum number of outstanding packets in each
1652          * window, and remember whether we were cwnd-limited then.
1653          */
1654         if (!before(tp->snd_una, tp->max_packets_seq) ||
1655             tp->packets_out > tp->max_packets_out) {
1656                 tp->max_packets_out = tp->packets_out;
1657                 tp->max_packets_seq = tp->snd_nxt;
1658                 tp->is_cwnd_limited = is_cwnd_limited;
1659         }
1660 
1661         if (tcp_is_cwnd_limited(sk)) {
1662                 /* Network is feed fully. */
1663                 tp->snd_cwnd_used = 0;
1664                 tp->snd_cwnd_stamp = tcp_jiffies32;
1665         } else {
1666                 /* Network starves. */
1667                 if (tp->packets_out > tp->snd_cwnd_used)
1668                         tp->snd_cwnd_used = tp->packets_out;
1669 
1670                 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1671                     (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1672                     !ca_ops->cong_control)
1673                         tcp_cwnd_application_limited(sk);
1674 
1675                 /* The following conditions together indicate the starvation
1676                  * is caused by insufficient sender buffer:
1677                  * 1) just sent some data (see tcp_write_xmit)
1678                  * 2) not cwnd limited (this else condition)
1679                  * 3) no more data to send (tcp_write_queue_empty())
1680                  * 4) application is hitting buffer limit (SOCK_NOSPACE)
1681                  */
1682                 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1683                     test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1684                     (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1685                         tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1686         }
1687 }
1688 
1689 /* Minshall's variant of the Nagle send check. */
1690 static bool tcp_minshall_check(const struct tcp_sock *tp)
1691 {
1692         return after(tp->snd_sml, tp->snd_una) &&
1693                 !after(tp->snd_sml, tp->snd_nxt);
1694 }
1695 
1696 /* Update snd_sml if this skb is under mss
1697  * Note that a TSO packet might end with a sub-mss segment
1698  * The test is really :
1699  * if ((skb->len % mss) != 0)
1700  *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1701  * But we can avoid doing the divide again given we already have
1702  *  skb_pcount = skb->len / mss_now
1703  */
1704 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1705                                 const struct sk_buff *skb)
1706 {
1707         if (skb->len < tcp_skb_pcount(skb) * mss_now)
1708                 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1709 }
1710 
1711 /* Return false, if packet can be sent now without violation Nagle's rules:
1712  * 1. It is full sized. (provided by caller in %partial bool)
1713  * 2. Or it contains FIN. (already checked by caller)
1714  * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1715  * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1716  *    With Minshall's modification: all sent small packets are ACKed.
1717  */
1718 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1719                             int nonagle)
1720 {
1721         return partial &&
1722                 ((nonagle & TCP_NAGLE_CORK) ||
1723                  (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1724 }
1725 
1726 /* Return how many segs we'd like on a TSO packet,
1727  * to send one TSO packet per ms
1728  */
1729 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1730                             int min_tso_segs)
1731 {
1732         u32 bytes, segs;
1733 
1734         bytes = min_t(unsigned long,
1735                       sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1736                       sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1737 
1738         /* Goal is to send at least one packet per ms,
1739          * not one big TSO packet every 100 ms.
1740          * This preserves ACK clocking and is consistent
1741          * with tcp_tso_should_defer() heuristic.
1742          */
1743         segs = max_t(u32, bytes / mss_now, min_tso_segs);
1744 
1745         return segs;
1746 }
1747 
1748 /* Return the number of segments we want in the skb we are transmitting.
1749  * See if congestion control module wants to decide; otherwise, autosize.
1750  */
1751 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1752 {
1753         const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1754         u32 min_tso, tso_segs;
1755 
1756         min_tso = ca_ops->min_tso_segs ?
1757                         ca_ops->min_tso_segs(sk) :
1758                         sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1759 
1760         tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1761         return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1762 }
1763 
1764 /* Returns the portion of skb which can be sent right away */
1765 static unsigned int tcp_mss_split_point(const struct sock *sk,
1766                                         const struct sk_buff *skb,
1767                                         unsigned int mss_now,
1768                                         unsigned int max_segs,
1769                                         int nonagle)
1770 {
1771         const struct tcp_sock *tp = tcp_sk(sk);
1772         u32 partial, needed, window, max_len;
1773 
1774         window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1775         max_len = mss_now * max_segs;
1776 
1777         if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1778                 return max_len;
1779 
1780         needed = min(skb->len, window);
1781 
1782         if (max_len <= needed)
1783                 return max_len;
1784 
1785         partial = needed % mss_now;
1786         /* If last segment is not a full MSS, check if Nagle rules allow us
1787          * to include this last segment in this skb.
1788          * Otherwise, we'll split the skb at last MSS boundary
1789          */
1790         if (tcp_nagle_check(partial != 0, tp, nonagle))
1791                 return needed - partial;
1792 
1793         return needed;
1794 }
1795 
1796 /* Can at least one segment of SKB be sent right now, according to the
1797  * congestion window rules?  If so, return how many segments are allowed.
1798  */
1799 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1800                                          const struct sk_buff *skb)
1801 {
1802         u32 in_flight, cwnd, halfcwnd;
1803 
1804         /* Don't be strict about the congestion window for the final FIN.  */
1805         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1806             tcp_skb_pcount(skb) == 1)
1807                 return 1;
1808 
1809         in_flight = tcp_packets_in_flight(tp);
1810         cwnd = tp->snd_cwnd;
1811         if (in_flight >= cwnd)
1812                 return 0;
1813 
1814         /* For better scheduling, ensure we have at least
1815          * 2 GSO packets in flight.
1816          */
1817         halfcwnd = max(cwnd >> 1, 1U);
1818         return min(halfcwnd, cwnd - in_flight);
1819 }
1820 
1821 /* Initialize TSO state of a skb.
1822  * This must be invoked the first time we consider transmitting
1823  * SKB onto the wire.
1824  */
1825 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1826 {
1827         int tso_segs = tcp_skb_pcount(skb);
1828 
1829         if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1830                 tcp_set_skb_tso_segs(skb, mss_now);
1831                 tso_segs = tcp_skb_pcount(skb);
1832         }
1833         return tso_segs;
1834 }
1835 
1836 
1837 /* Return true if the Nagle test allows this packet to be
1838  * sent now.
1839  */
1840 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1841                                   unsigned int cur_mss, int nonagle)
1842 {
1843         /* Nagle rule does not apply to frames, which sit in the middle of the
1844          * write_queue (they have no chances to get new data).
1845          *
1846          * This is implemented in the callers, where they modify the 'nonagle'
1847          * argument based upon the location of SKB in the send queue.
1848          */
1849         if (nonagle & TCP_NAGLE_PUSH)
1850                 return true;
1851 
1852         /* Don't use the nagle rule for urgent data (or for the final FIN). */
1853         if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1854                 return true;
1855 
1856         if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1857                 return true;
1858 
1859         return false;
1860 }
1861 
1862 /* Does at least the first segment of SKB fit into the send window? */
1863 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1864                              const struct sk_buff *skb,
1865                              unsigned int cur_mss)
1866 {
1867         u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1868 
1869         if (skb->len > cur_mss)
1870                 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1871 
1872         return !after(end_seq, tcp_wnd_end(tp));
1873 }
1874 
1875 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1876  * which is put after SKB on the list.  It is very much like
1877  * tcp_fragment() except that it may make several kinds of assumptions
1878  * in order to speed up the splitting operation.  In particular, we
1879  * know that all the data is in scatter-gather pages, and that the
1880  * packet has never been sent out before (and thus is not cloned).
1881  */
1882 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1883                         unsigned int mss_now, gfp_t gfp)
1884 {
1885         int nlen = skb->len - len;
1886         struct sk_buff *buff;
1887         u8 flags;
1888 
1889         /* All of a TSO frame must be composed of paged data.  */
1890         if (skb->len != skb->data_len)
1891                 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1892                                     skb, len, mss_now, gfp);
1893 
1894         buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1895         if (unlikely(!buff))
1896                 return -ENOMEM;
1897         skb_copy_decrypted(buff, skb);
1898 
1899         sk_wmem_queued_add(sk, buff->truesize);
1900         sk_mem_charge(sk, buff->truesize);
1901         buff->truesize += nlen;
1902         skb->truesize -= nlen;
1903 
1904         /* Correct the sequence numbers. */
1905         TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1906         TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1907         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1908 
1909         /* PSH and FIN should only be set in the second packet. */
1910         flags = TCP_SKB_CB(skb)->tcp_flags;
1911         TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1912         TCP_SKB_CB(buff)->tcp_flags = flags;
1913 
1914         /* This packet was never sent out yet, so no SACK bits. */
1915         TCP_SKB_CB(buff)->sacked = 0;
1916 
1917         tcp_skb_fragment_eor(skb, buff);
1918 
1919         buff->ip_summed = CHECKSUM_PARTIAL;
1920         skb_split(skb, buff, len);
1921         tcp_fragment_tstamp(skb, buff);
1922 
1923         /* Fix up tso_factor for both original and new SKB.  */
1924         tcp_set_skb_tso_segs(skb, mss_now);
1925         tcp_set_skb_tso_segs(buff, mss_now);
1926 
1927         /* Link BUFF into the send queue. */
1928         __skb_header_release(buff);
1929         tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
1930 
1931         return 0;
1932 }
1933 
1934 /* Try to defer sending, if possible, in order to minimize the amount
1935  * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1936  *
1937  * This algorithm is from John Heffner.
1938  */
1939 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1940                                  bool *is_cwnd_limited,
1941                                  bool *is_rwnd_limited,
1942                                  u32 max_segs)
1943 {
1944         const struct inet_connection_sock *icsk = inet_csk(sk);
1945         u32 send_win, cong_win, limit, in_flight;
1946         struct tcp_sock *tp = tcp_sk(sk);
1947         struct sk_buff *head;
1948         int win_divisor;
1949         s64 delta;
1950 
1951         if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1952                 goto send_now;
1953 
1954         /* Avoid bursty behavior by allowing defer
1955          * only if the last write was recent (1 ms).
1956          * Note that tp->tcp_wstamp_ns can be in the future if we have
1957          * packets waiting in a qdisc or device for EDT delivery.
1958          */
1959         delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
1960         if (delta > 0)
1961                 goto send_now;
1962 
1963         in_flight = tcp_packets_in_flight(tp);
1964 
1965         BUG_ON(tcp_skb_pcount(skb) <= 1);
1966         BUG_ON(tp->snd_cwnd <= in_flight);
1967 
1968         send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1969 
1970         /* From in_flight test above, we know that cwnd > in_flight.  */
1971         cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1972 
1973         limit = min(send_win, cong_win);
1974 
1975         /* If a full-sized TSO skb can be sent, do it. */
1976         if (limit >= max_segs * tp->mss_cache)
1977                 goto send_now;
1978 
1979         /* Middle in queue won't get any more data, full sendable already? */
1980         if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1981                 goto send_now;
1982 
1983         win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1984         if (win_divisor) {
1985                 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1986 
1987                 /* If at least some fraction of a window is available,
1988                  * just use it.
1989                  */
1990                 chunk /= win_divisor;
1991                 if (limit >= chunk)
1992                         goto send_now;
1993         } else {
1994                 /* Different approach, try not to defer past a single
1995                  * ACK.  Receiver should ACK every other full sized
1996                  * frame, so if we have space for more than 3 frames
1997                  * then send now.
1998                  */
1999                 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2000                         goto send_now;
2001         }
2002 
2003         /* TODO : use tsorted_sent_queue ? */
2004         head = tcp_rtx_queue_head(sk);
2005         if (!head)
2006                 goto send_now;
2007         delta = tp->tcp_clock_cache - head->tstamp;
2008         /* If next ACK is likely to come too late (half srtt), do not defer */
2009         if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2010                 goto send_now;
2011 
2012         /* Ok, it looks like it is advisable to defer.
2013          * Three cases are tracked :
2014          * 1) We are cwnd-limited
2015          * 2) We are rwnd-limited
2016          * 3) We are application limited.
2017          */
2018         if (cong_win < send_win) {
2019                 if (cong_win <= skb->len) {
2020                         *is_cwnd_limited = true;
2021                         return true;
2022                 }
2023         } else {
2024                 if (send_win <= skb->len) {
2025                         *is_rwnd_limited = true;
2026                         return true;
2027                 }
2028         }
2029 
2030         /* If this packet won't get more data, do not wait. */
2031         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2032             TCP_SKB_CB(skb)->eor)
2033                 goto send_now;
2034 
2035         return true;
2036 
2037 send_now:
2038         return false;
2039 }
2040 
2041 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2042 {
2043         struct inet_connection_sock *icsk = inet_csk(sk);
2044         struct tcp_sock *tp = tcp_sk(sk);
2045         struct net *net = sock_net(sk);
2046         u32 interval;
2047         s32 delta;
2048 
2049         interval = net->ipv4.sysctl_tcp_probe_interval;
2050         delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2051         if (unlikely(delta >= interval * HZ)) {
2052                 int mss = tcp_current_mss(sk);
2053 
2054                 /* Update current search range */
2055                 icsk->icsk_mtup.probe_size = 0;
2056                 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2057                         sizeof(struct tcphdr) +
2058                         icsk->icsk_af_ops->net_header_len;
2059                 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2060 
2061                 /* Update probe time stamp */
2062                 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2063         }
2064 }
2065 
2066 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2067 {
2068         struct sk_buff *skb, *next;
2069 
2070         skb = tcp_send_head(sk);
2071         tcp_for_write_queue_from_safe(skb, next, sk) {
2072                 if (len <= skb->len)
2073                         break;
2074 
2075                 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2076                         return false;
2077 
2078                 len -= skb->len;
2079         }
2080 
2081         return true;
2082 }
2083 
2084 /* Create a new MTU probe if we are ready.
2085  * MTU probe is regularly attempting to increase the path MTU by
2086  * deliberately sending larger packets.  This discovers routing
2087  * changes resulting in larger path MTUs.
2088  *
2089  * Returns 0 if we should wait to probe (no cwnd available),
2090  *         1 if a probe was sent,
2091  *         -1 otherwise
2092  */
2093 static int tcp_mtu_probe(struct sock *sk)
2094 {
2095         struct inet_connection_sock *icsk = inet_csk(sk);
2096         struct tcp_sock *tp = tcp_sk(sk);
2097         struct sk_buff *skb, *nskb, *next;
2098         struct net *net = sock_net(sk);
2099         int probe_size;
2100         int size_needed;
2101         int copy, len;
2102         int mss_now;
2103         int interval;
2104 
2105         /* Not currently probing/verifying,
2106          * not in recovery,
2107          * have enough cwnd, and
2108          * not SACKing (the variable headers throw things off)
2109          */
2110         if (likely(!icsk->icsk_mtup.enabled ||
2111                    icsk->icsk_mtup.probe_size ||
2112                    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2113                    tp->snd_cwnd < 11 ||
2114                    tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2115                 return -1;
2116 
2117         /* Use binary search for probe_size between tcp_mss_base,
2118          * and current mss_clamp. if (search_high - search_low)
2119          * smaller than a threshold, backoff from probing.
2120          */
2121         mss_now = tcp_current_mss(sk);
2122         probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2123                                     icsk->icsk_mtup.search_low) >> 1);
2124         size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2125         interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2126         /* When misfortune happens, we are reprobing actively,
2127          * and then reprobe timer has expired. We stick with current
2128          * probing process by not resetting search range to its orignal.
2129          */
2130         if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2131                 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2132                 /* Check whether enough time has elaplased for
2133                  * another round of probing.
2134                  */
2135                 tcp_mtu_check_reprobe(sk);
2136                 return -1;
2137         }
2138 
2139         /* Have enough data in the send queue to probe? */
2140         if (tp->write_seq - tp->snd_nxt < size_needed)
2141                 return -1;
2142 
2143         if (tp->snd_wnd < size_needed)
2144                 return -1;
2145         if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2146                 return 0;
2147 
2148         /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2149         if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2150                 if (!tcp_packets_in_flight(tp))
2151                         return -1;
2152                 else
2153                         return 0;
2154         }
2155 
2156         if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2157                 return -1;
2158 
2159         /* We're allowed to probe.  Build it now. */
2160         nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2161         if (!nskb)
2162                 return -1;
2163         sk_wmem_queued_add(sk, nskb->truesize);
2164         sk_mem_charge(sk, nskb->truesize);
2165 
2166         skb = tcp_send_head(sk);
2167         skb_copy_decrypted(nskb, skb);
2168 
2169         TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2170         TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2171         TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2172         TCP_SKB_CB(nskb)->sacked = 0;
2173         nskb->csum = 0;
2174         nskb->ip_summed = CHECKSUM_PARTIAL;
2175 
2176         tcp_insert_write_queue_before(nskb, skb, sk);
2177         tcp_highest_sack_replace(sk, skb, nskb);
2178 
2179         len = 0;
2180         tcp_for_write_queue_from_safe(skb, next, sk) {
2181                 copy = min_t(int, skb->len, probe_size - len);
2182                 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2183 
2184                 if (skb->len <= copy) {
2185                         /* We've eaten all the data from this skb.
2186                          * Throw it away. */
2187                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2188                         /* If this is the last SKB we copy and eor is set
2189                          * we need to propagate it to the new skb.
2190                          */
2191                         TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2192                         tcp_skb_collapse_tstamp(nskb, skb);
2193                         tcp_unlink_write_queue(skb, sk);
2194                         sk_wmem_free_skb(sk, skb);
2195                 } else {
2196                         TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2197                                                    ~(TCPHDR_FIN|TCPHDR_PSH);
2198                         if (!skb_shinfo(skb)->nr_frags) {
2199                                 skb_pull(skb, copy);
2200                         } else {
2201                                 __pskb_trim_head(skb, copy);
2202                                 tcp_set_skb_tso_segs(skb, mss_now);
2203                         }
2204                         TCP_SKB_CB(skb)->seq += copy;
2205                 }
2206 
2207                 len += copy;
2208 
2209                 if (len >= probe_size)
2210                         break;
2211         }
2212         tcp_init_tso_segs(nskb, nskb->len);
2213 
2214         /* We're ready to send.  If this fails, the probe will
2215          * be resegmented into mss-sized pieces by tcp_write_xmit().
2216          */
2217         if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2218                 /* Decrement cwnd here because we are sending
2219                  * effectively two packets. */
2220                 tp->snd_cwnd--;
2221                 tcp_event_new_data_sent(sk, nskb);
2222 
2223                 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2224                 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2225                 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2226 
2227                 return 1;
2228         }
2229 
2230         return -1;
2231 }
2232 
2233 static bool tcp_pacing_check(struct sock *sk)
2234 {
2235         struct tcp_sock *tp = tcp_sk(sk);
2236 
2237         if (!tcp_needs_internal_pacing(sk))
2238                 return false;
2239 
2240         if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2241                 return false;
2242 
2243         if (!hrtimer_is_queued(&tp->pacing_timer)) {
2244                 hrtimer_start(&tp->pacing_timer,
2245                               ns_to_ktime(tp->tcp_wstamp_ns),
2246                               HRTIMER_MODE_ABS_PINNED_SOFT);
2247                 sock_hold(sk);
2248         }
2249         return true;
2250 }
2251 
2252 /* TCP Small Queues :
2253  * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2254  * (These limits are doubled for retransmits)
2255  * This allows for :
2256  *  - better RTT estimation and ACK scheduling
2257  *  - faster recovery
2258  *  - high rates
2259  * Alas, some drivers / subsystems require a fair amount
2260  * of queued bytes to ensure line rate.
2261  * One example is wifi aggregation (802.11 AMPDU)
2262  */
2263 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2264                                   unsigned int factor)
2265 {
2266         unsigned long limit;
2267 
2268         limit = max_t(unsigned long,
2269                       2 * skb->truesize,
2270                       sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2271         if (sk->sk_pacing_status == SK_PACING_NONE)
2272                 limit = min_t(unsigned long, limit,
2273                               sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2274         limit <<= factor;
2275 
2276         if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2277             tcp_sk(sk)->tcp_tx_delay) {
2278                 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2279 
2280                 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2281                  * approximate our needs assuming an ~100% skb->truesize overhead.
2282                  * USEC_PER_SEC is approximated by 2^20.
2283                  * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2284                  */
2285                 extra_bytes >>= (20 - 1);
2286                 limit += extra_bytes;
2287         }
2288         if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2289                 /* Always send skb if rtx queue is empty.
2290                  * No need to wait for TX completion to call us back,
2291                  * after softirq/tasklet schedule.
2292                  * This helps when TX completions are delayed too much.
2293                  */
2294                 if (tcp_rtx_queue_empty(sk))
2295                         return false;
2296 
2297                 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2298                 /* It is possible TX completion already happened
2299                  * before we set TSQ_THROTTLED, so we must
2300                  * test again the condition.
2301                  */
2302                 smp_mb__after_atomic();
2303                 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2304                         return true;
2305         }
2306         return false;
2307 }
2308 
2309 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2310 {
2311         const u32 now = tcp_jiffies32;
2312         enum tcp_chrono old = tp->chrono_type;
2313 
2314         if (old > TCP_CHRONO_UNSPEC)
2315                 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2316         tp->chrono_start = now;
2317         tp->chrono_type = new;
2318 }
2319 
2320 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2321 {
2322         struct tcp_sock *tp = tcp_sk(sk);
2323 
2324         /* If there are multiple conditions worthy of tracking in a
2325          * chronograph then the highest priority enum takes precedence
2326          * over the other conditions. So that if something "more interesting"
2327          * starts happening, stop the previous chrono and start a new one.
2328          */
2329         if (type > tp->chrono_type)
2330                 tcp_chrono_set(tp, type);
2331 }
2332 
2333 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2334 {
2335         struct tcp_sock *tp = tcp_sk(sk);
2336 
2337 
2338         /* There are multiple conditions worthy of tracking in a
2339          * chronograph, so that the highest priority enum takes
2340          * precedence over the other conditions (see tcp_chrono_start).
2341          * If a condition stops, we only stop chrono tracking if
2342          * it's the "most interesting" or current chrono we are
2343          * tracking and starts busy chrono if we have pending data.
2344          */
2345         if (tcp_rtx_and_write_queues_empty(sk))
2346                 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2347         else if (type == tp->chrono_type)
2348                 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2349 }
2350 
2351 /* This routine writes packets to the network.  It advances the
2352  * send_head.  This happens as incoming acks open up the remote
2353  * window for us.
2354  *
2355  * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2356  * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2357  * account rare use of URG, this is not a big flaw.
2358  *
2359  * Send at most one packet when push_one > 0. Temporarily ignore
2360  * cwnd limit to force at most one packet out when push_one == 2.
2361 
2362  * Returns true, if no segments are in flight and we have queued segments,
2363  * but cannot send anything now because of SWS or another problem.
2364  */
2365 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2366                            int push_one, gfp_t gfp)
2367 {
2368         struct tcp_sock *tp = tcp_sk(sk);
2369         struct sk_buff *skb;
2370         unsigned int tso_segs, sent_pkts;
2371         int cwnd_quota;
2372         int result;
2373         bool is_cwnd_limited = false, is_rwnd_limited = false;
2374         u32 max_segs;
2375 
2376         sent_pkts = 0;
2377 
2378         tcp_mstamp_refresh(tp);
2379         if (!push_one) {
2380                 /* Do MTU probing. */
2381                 result = tcp_mtu_probe(sk);
2382                 if (!result) {
2383                         return false;
2384                 } else if (result > 0) {
2385                         sent_pkts = 1;
2386                 }
2387         }
2388 
2389         max_segs = tcp_tso_segs(sk, mss_now);
2390         while ((skb = tcp_send_head(sk))) {
2391                 unsigned int limit;
2392 
2393                 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2394                         /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2395                         skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2396                         list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2397                         tcp_init_tso_segs(skb, mss_now);
2398                         goto repair; /* Skip network transmission */
2399                 }
2400 
2401                 if (tcp_pacing_check(sk))
2402                         break;
2403 
2404                 tso_segs = tcp_init_tso_segs(skb, mss_now);
2405                 BUG_ON(!tso_segs);
2406 
2407                 cwnd_quota = tcp_cwnd_test(tp, skb);
2408                 if (!cwnd_quota) {
2409                         if (push_one == 2)
2410                                 /* Force out a loss probe pkt. */
2411                                 cwnd_quota = 1;
2412                         else
2413                                 break;
2414                 }
2415 
2416                 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2417                         is_rwnd_limited = true;
2418                         break;
2419                 }
2420 
2421                 if (tso_segs == 1) {
2422                         if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2423                                                      (tcp_skb_is_last(sk, skb) ?
2424                                                       nonagle : TCP_NAGLE_PUSH))))
2425                                 break;
2426                 } else {
2427                         if (!push_one &&
2428                             tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2429                                                  &is_rwnd_limited, max_segs))
2430                                 break;
2431                 }
2432 
2433                 limit = mss_now;
2434                 if (tso_segs > 1 && !tcp_urg_mode(tp))
2435                         limit = tcp_mss_split_point(sk, skb, mss_now,
2436                                                     min_t(unsigned int,
2437                                                           cwnd_quota,
2438                                                           max_segs),
2439                                                     nonagle);
2440 
2441                 if (skb->len > limit &&
2442                     unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2443                         break;
2444 
2445                 if (tcp_small_queue_check(sk, skb, 0))
2446                         break;
2447 
2448                 /* Argh, we hit an empty skb(), presumably a thread
2449                  * is sleeping in sendmsg()/sk_stream_wait_memory().
2450                  * We do not want to send a pure-ack packet and have
2451                  * a strange looking rtx queue with empty packet(s).
2452                  */
2453                 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2454                         break;
2455 
2456                 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2457                         break;
2458 
2459 repair:
2460                 /* Advance the send_head.  This one is sent out.
2461                  * This call will increment packets_out.
2462                  */
2463                 tcp_event_new_data_sent(sk, skb);
2464 
2465                 tcp_minshall_update(tp, mss_now, skb);
2466                 sent_pkts += tcp_skb_pcount(skb);
2467 
2468                 if (push_one)
2469                         break;
2470         }
2471 
2472         if (is_rwnd_limited)
2473                 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2474         else
2475                 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2476 
2477         if (likely(sent_pkts)) {
2478                 if (tcp_in_cwnd_reduction(sk))
2479                         tp->prr_out += sent_pkts;
2480 
2481                 /* Send one loss probe per tail loss episode. */
2482                 if (push_one != 2)
2483                         tcp_schedule_loss_probe(sk, false);
2484                 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2485                 tcp_cwnd_validate(sk, is_cwnd_limited);
2486                 return false;
2487         }
2488         return !tp->packets_out && !tcp_write_queue_empty(sk);
2489 }
2490 
2491 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2492 {
2493         struct inet_connection_sock *icsk = inet_csk(sk);
2494         struct tcp_sock *tp = tcp_sk(sk);
2495         u32 timeout, rto_delta_us;
2496         int early_retrans;
2497 
2498         /* Don't do any loss probe on a Fast Open connection before 3WHS
2499          * finishes.
2500          */
2501         if (rcu_access_pointer(tp->fastopen_rsk))
2502                 return false;
2503 
2504         early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2505         /* Schedule a loss probe in 2*RTT for SACK capable connections
2506          * not in loss recovery, that are either limited by cwnd or application.
2507          */
2508         if ((early_retrans != 3 && early_retrans != 4) ||
2509             !tp->packets_out || !tcp_is_sack(tp) ||
2510             (icsk->icsk_ca_state != TCP_CA_Open &&
2511              icsk->icsk_ca_state != TCP_CA_CWR))
2512                 return false;
2513 
2514         /* Probe timeout is 2*rtt. Add minimum RTO to account
2515          * for delayed ack when there's one outstanding packet. If no RTT
2516          * sample is available then probe after TCP_TIMEOUT_INIT.
2517          */
2518         if (tp->srtt_us) {
2519                 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2520                 if (tp->packets_out == 1)
2521                         timeout += TCP_RTO_MIN;
2522                 else
2523                         timeout += TCP_TIMEOUT_MIN;
2524         } else {
2525                 timeout = TCP_TIMEOUT_INIT;
2526         }
2527 
2528         /* If the RTO formula yields an earlier time, then use that time. */
2529         rto_delta_us = advancing_rto ?
2530                         jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2531                         tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2532         if (rto_delta_us > 0)
2533                 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2534 
2535         tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2536                              TCP_RTO_MAX, NULL);
2537         return true;
2538 }
2539 
2540 /* Thanks to skb fast clones, we can detect if a prior transmit of
2541  * a packet is still in a qdisc or driver queue.
2542  * In this case, there is very little point doing a retransmit !
2543  */
2544 static bool skb_still_in_host_queue(const struct sock *sk,
2545                                     const struct sk_buff *skb)
2546 {
2547         if (unlikely(skb_fclone_busy(sk, skb))) {
2548                 NET_INC_STATS(sock_net(sk),
2549                               LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2550                 return true;
2551         }
2552         return false;
2553 }
2554 
2555 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2556  * retransmit the last segment.
2557  */
2558 void tcp_send_loss_probe(struct sock *sk)
2559 {
2560         struct tcp_sock *tp = tcp_sk(sk);
2561         struct sk_buff *skb;
2562         int pcount;
2563         int mss = tcp_current_mss(sk);
2564 
2565         skb = tcp_send_head(sk);
2566         if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2567                 pcount = tp->packets_out;
2568                 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2569                 if (tp->packets_out > pcount)
2570                         goto probe_sent;
2571                 goto rearm_timer;
2572         }
2573         skb = skb_rb_last(&sk->tcp_rtx_queue);
2574         if (unlikely(!skb)) {
2575                 WARN_ONCE(tp->packets_out,
2576                           "invalid inflight: %u state %u cwnd %u mss %d\n",
2577                           tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2578                 inet_csk(sk)->icsk_pending = 0;
2579                 return;
2580         }
2581 
2582         /* At most one outstanding TLP retransmission. */
2583         if (tp->tlp_high_seq)
2584                 goto rearm_timer;
2585 
2586         if (skb_still_in_host_queue(sk, skb))
2587                 goto rearm_timer;
2588 
2589         pcount = tcp_skb_pcount(skb);
2590         if (WARN_ON(!pcount))
2591                 goto rearm_timer;
2592 
2593         if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2594                 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2595                                           (pcount - 1) * mss, mss,
2596                                           GFP_ATOMIC)))
2597                         goto rearm_timer;
2598                 skb = skb_rb_next(skb);
2599         }
2600 
2601         if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2602                 goto rearm_timer;
2603 
2604         if (__tcp_retransmit_skb(sk, skb, 1))
2605                 goto rearm_timer;
2606 
2607         /* Record snd_nxt for loss detection. */
2608         tp->tlp_high_seq = tp->snd_nxt;
2609 
2610 probe_sent:
2611         NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2612         /* Reset s.t. tcp_rearm_rto will restart timer from now */
2613         inet_csk(sk)->icsk_pending = 0;
2614 rearm_timer:
2615         tcp_rearm_rto(sk);
2616 }
2617 
2618 /* Push out any pending frames which were held back due to
2619  * TCP_CORK or attempt at coalescing tiny packets.
2620  * The socket must be locked by the caller.
2621  */
2622 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2623                                int nonagle)
2624 {
2625         /* If we are closed, the bytes will have to remain here.
2626          * In time closedown will finish, we empty the write queue and
2627          * all will be happy.
2628          */
2629         if (unlikely(sk->sk_state == TCP_CLOSE))
2630                 return;
2631 
2632         if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2633                            sk_gfp_mask(sk, GFP_ATOMIC)))
2634                 tcp_check_probe_timer(sk);
2635 }
2636 
2637 /* Send _single_ skb sitting at the send head. This function requires
2638  * true push pending frames to setup probe timer etc.
2639  */
2640 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2641 {
2642         struct sk_buff *skb = tcp_send_head(sk);
2643 
2644         BUG_ON(!skb || skb->len < mss_now);
2645 
2646         tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2647 }
2648 
2649 /* This function returns the amount that we can raise the
2650  * usable window based on the following constraints
2651  *
2652  * 1. The window can never be shrunk once it is offered (RFC 793)
2653  * 2. We limit memory per socket
2654  *
2655  * RFC 1122:
2656  * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2657  *  RECV.NEXT + RCV.WIN fixed until:
2658  *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2659  *
2660  * i.e. don't raise the right edge of the window until you can raise
2661  * it at least MSS bytes.
2662  *
2663  * Unfortunately, the recommended algorithm breaks header prediction,
2664  * since header prediction assumes th->window stays fixed.
2665  *
2666  * Strictly speaking, keeping th->window fixed violates the receiver
2667  * side SWS prevention criteria. The problem is that under this rule
2668  * a stream of single byte packets will cause the right side of the
2669  * window to always advance by a single byte.
2670  *
2671  * Of course, if the sender implements sender side SWS prevention
2672  * then this will not be a problem.
2673  *
2674  * BSD seems to make the following compromise:
2675  *
2676  *      If the free space is less than the 1/4 of the maximum
2677  *      space available and the free space is less than 1/2 mss,
2678  *      then set the window to 0.
2679  *      [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2680  *      Otherwise, just prevent the window from shrinking
2681  *      and from being larger than the largest representable value.
2682  *
2683  * This prevents incremental opening of the window in the regime
2684  * where TCP is limited by the speed of the reader side taking
2685  * data out of the TCP receive queue. It does nothing about
2686  * those cases where the window is constrained on the sender side
2687  * because the pipeline is full.
2688  *
2689  * BSD also seems to "accidentally" limit itself to windows that are a
2690  * multiple of MSS, at least until the free space gets quite small.
2691  * This would appear to be a side effect of the mbuf implementation.
2692  * Combining these two algorithms results in the observed behavior
2693  * of having a fixed window size at almost all times.
2694  *
2695  * Below we obtain similar behavior by forcing the offered window to
2696  * a multiple of the mss when it is feasible to do so.
2697  *
2698  * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2699  * Regular options like TIMESTAMP are taken into account.
2700  */
2701 u32 __tcp_select_window(struct sock *sk)
2702 {
2703         struct inet_connection_sock *icsk = inet_csk(sk);
2704         struct tcp_sock *tp = tcp_sk(sk);
2705         /* MSS for the peer's data.  Previous versions used mss_clamp
2706          * here.  I don't know if the value based on our guesses
2707          * of peer's MSS is better for the performance.  It's more correct
2708          * but may be worse for the performance because of rcv_mss
2709          * fluctuations.  --SAW  1998/11/1
2710          */
2711         int mss = icsk->icsk_ack.rcv_mss;
2712         int free_space = tcp_space(sk);
2713         int allowed_space = tcp_full_space(sk);
2714         int full_space = min_t(int, tp->window_clamp, allowed_space);
2715         int window;
2716 
2717         if (unlikely(mss > full_space)) {
2718                 mss = full_space;
2719                 if (mss <= 0)
2720                         return 0;
2721         }
2722         if (free_space < (full_space >> 1)) {
2723                 icsk->icsk_ack.quick = 0;
2724 
2725                 if (tcp_under_memory_pressure(sk))
2726                         tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2727                                                4U * tp->advmss);
2728 
2729                 /* free_space might become our new window, make sure we don't
2730                  * increase it due to wscale.
2731                  */
2732                 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2733 
2734                 /* if free space is less than mss estimate, or is below 1/16th
2735                  * of the maximum allowed, try to move to zero-window, else
2736                  * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2737                  * new incoming data is dropped due to memory limits.
2738                  * With large window, mss test triggers way too late in order
2739                  * to announce zero window in time before rmem limit kicks in.
2740                  */
2741                 if (free_space < (allowed_space >> 4) || free_space < mss)
2742                         return 0;
2743         }
2744 
2745         if (free_space > tp->rcv_ssthresh)
2746                 free_space = tp->rcv_ssthresh;
2747 
2748         /* Don't do rounding if we are using window scaling, since the
2749          * scaled window will not line up with the MSS boundary anyway.
2750          */
2751         if (tp->rx_opt.rcv_wscale) {
2752                 window = free_space;
2753 
2754                 /* Advertise enough space so that it won't get scaled away.
2755                  * Import case: prevent zero window announcement if
2756                  * 1<<rcv_wscale > mss.
2757                  */
2758                 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2759         } else {
2760                 window = tp->rcv_wnd;
2761                 /* Get the largest window that is a nice multiple of mss.
2762                  * Window clamp already applied above.
2763                  * If our current window offering is within 1 mss of the
2764                  * free space we just keep it. This prevents the divide
2765                  * and multiply from happening most of the time.
2766                  * We also don't do any window rounding when the free space
2767                  * is too small.
2768                  */
2769                 if (window <= free_space - mss || window > free_space)
2770                         window = rounddown(free_space, mss);
2771                 else if (mss == full_space &&
2772                          free_space > window + (full_space >> 1))
2773                         window = free_space;
2774         }
2775 
2776         return window;
2777 }
2778 
2779 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2780                              const struct sk_buff *next_skb)
2781 {
2782         if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2783                 const struct skb_shared_info *next_shinfo =
2784                         skb_shinfo(next_skb);
2785                 struct skb_shared_info *shinfo = skb_shinfo(skb);
2786 
2787                 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2788                 shinfo->tskey = next_shinfo->tskey;
2789                 TCP_SKB_CB(skb)->txstamp_ack |=
2790                         TCP_SKB_CB(next_skb)->txstamp_ack;
2791         }
2792 }
2793 
2794 /* Collapses two adjacent SKB's during retransmission. */
2795 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2796 {
2797         struct tcp_sock *tp = tcp_sk(sk);
2798         struct sk_buff *next_skb = skb_rb_next(skb);
2799         int next_skb_size;
2800 
2801         next_skb_size = next_skb->len;
2802 
2803         BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2804 
2805         if (next_skb_size) {
2806                 if (next_skb_size <= skb_availroom(skb))
2807                         skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2808                                       next_skb_size);
2809                 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2810                         return false;
2811         }
2812         tcp_highest_sack_replace(sk, next_skb, skb);
2813 
2814         /* Update sequence range on original skb. */
2815         TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2816 
2817         /* Merge over control information. This moves PSH/FIN etc. over */
2818         TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2819 
2820         /* All done, get rid of second SKB and account for it so
2821          * packet counting does not break.
2822          */
2823         TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2824         TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2825 
2826         /* changed transmit queue under us so clear hints */
2827         tcp_clear_retrans_hints_partial(tp);
2828         if (next_skb == tp->retransmit_skb_hint)
2829                 tp->retransmit_skb_hint = skb;
2830 
2831         tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2832 
2833         tcp_skb_collapse_tstamp(skb, next_skb);
2834 
2835         tcp_rtx_queue_unlink_and_free(next_skb, sk);
2836         return true;
2837 }
2838 
2839 /* Check if coalescing SKBs is legal. */
2840 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2841 {
2842         if (tcp_skb_pcount(skb) > 1)
2843                 return false;
2844         if (skb_cloned(skb))
2845                 return false;
2846         /* Some heuristics for collapsing over SACK'd could be invented */
2847         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2848                 return false;
2849 
2850         return true;
2851 }
2852 
2853 /* Collapse packets in the retransmit queue to make to create
2854  * less packets on the wire. This is only done on retransmission.
2855  */
2856 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2857                                      int space)
2858 {
2859         struct tcp_sock *tp = tcp_sk(sk);
2860         struct sk_buff *skb = to, *tmp;
2861         bool first = true;
2862 
2863         if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2864                 return;
2865         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2866                 return;
2867 
2868         skb_rbtree_walk_from_safe(skb, tmp) {
2869                 if (!tcp_can_collapse(sk, skb))
2870                         break;
2871 
2872                 if (!tcp_skb_can_collapse_to(to))
2873                         break;
2874 
2875                 space -= skb->len;
2876 
2877                 if (first) {
2878                         first = false;
2879                         continue;
2880                 }
2881 
2882                 if (space < 0)
2883                         break;
2884 
2885                 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2886                         break;
2887 
2888                 if (!tcp_collapse_retrans(sk, to))
2889                         break;
2890         }
2891 }
2892 
2893 /* This retransmits one SKB.  Policy decisions and retransmit queue
2894  * state updates are done by the caller.  Returns non-zero if an
2895  * error occurred which prevented the send.
2896  */
2897 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2898 {
2899         struct inet_connection_sock *icsk = inet_csk(sk);
2900         struct tcp_sock *tp = tcp_sk(sk);
2901         unsigned int cur_mss;
2902         int diff, len, err;
2903 
2904 
2905         /* Inconclusive MTU probe */
2906         if (icsk->icsk_mtup.probe_size)
2907                 icsk->icsk_mtup.probe_size = 0;
2908 
2909         /* Do not sent more than we queued. 1/4 is reserved for possible
2910          * copying overhead: fragmentation, tunneling, mangling etc.
2911          */
2912         if (refcount_read(&sk->sk_wmem_alloc) >
2913             min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2914                   sk->sk_sndbuf))
2915                 return -EAGAIN;
2916 
2917         if (skb_still_in_host_queue(sk, skb))
2918                 return -EBUSY;
2919 
2920         if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2921                 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2922                         WARN_ON_ONCE(1);
2923                         return -EINVAL;
2924                 }
2925                 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2926                         return -ENOMEM;
2927         }
2928 
2929         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2930                 return -EHOSTUNREACH; /* Routing failure or similar. */
2931 
2932         cur_mss = tcp_current_mss(sk);
2933 
2934         /* If receiver has shrunk his window, and skb is out of
2935          * new window, do not retransmit it. The exception is the
2936          * case, when window is shrunk to zero. In this case
2937          * our retransmit serves as a zero window probe.
2938          */
2939         if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2940             TCP_SKB_CB(skb)->seq != tp->snd_una)
2941                 return -EAGAIN;
2942 
2943         len = cur_mss * segs;
2944         if (skb->len > len) {
2945                 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2946                                  cur_mss, GFP_ATOMIC))
2947                         return -ENOMEM; /* We'll try again later. */
2948         } else {
2949                 if (skb_unclone(skb, GFP_ATOMIC))
2950                         return -ENOMEM;
2951 
2952                 diff = tcp_skb_pcount(skb);
2953                 tcp_set_skb_tso_segs(skb, cur_mss);
2954                 diff -= tcp_skb_pcount(skb);
2955                 if (diff)
2956                         tcp_adjust_pcount(sk, skb, diff);
2957                 if (skb->len < cur_mss)
2958                         tcp_retrans_try_collapse(sk, skb, cur_mss);
2959         }
2960 
2961         /* RFC3168, section 6.1.1.1. ECN fallback */
2962         if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2963                 tcp_ecn_clear_syn(sk, skb);
2964 
2965         /* Update global and local TCP statistics. */
2966         segs = tcp_skb_pcount(skb);
2967         TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2968         if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2969                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2970         tp->total_retrans += segs;
2971         tp->bytes_retrans += skb->len;
2972 
2973         /* make sure skb->data is aligned on arches that require it
2974          * and check if ack-trimming & collapsing extended the headroom
2975          * beyond what csum_start can cover.
2976          */
2977         if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2978                      skb_headroom(skb) >= 0xFFFF)) {
2979                 struct sk_buff *nskb;
2980 
2981                 tcp_skb_tsorted_save(skb) {
2982                         nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2983                         if (nskb) {
2984                                 nskb->dev = NULL;
2985                                 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
2986                         } else {
2987                                 err = -ENOBUFS;
2988                         }
2989                 } tcp_skb_tsorted_restore(skb);
2990 
2991                 if (!err) {
2992                         tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
2993                         tcp_rate_skb_sent(sk, skb);
2994                 }
2995         } else {
2996                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2997         }
2998 
2999         /* To avoid taking spuriously low RTT samples based on a timestamp
3000          * for a transmit that never happened, always mark EVER_RETRANS
3001          */
3002         TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3003 
3004         if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3005                 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3006                                   TCP_SKB_CB(skb)->seq, segs, err);
3007 
3008         if (likely(!err)) {
3009                 trace_tcp_retransmit_skb(sk, skb);
3010         } else if (err != -EBUSY) {
3011                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3012         }
3013         return err;
3014 }
3015 
3016 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3017 {
3018         struct tcp_sock *tp = tcp_sk(sk);
3019         int err = __tcp_retransmit_skb(sk, skb, segs);
3020 
3021         if (err == 0) {
3022 #if FASTRETRANS_DEBUG > 0
3023                 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3024                         net_dbg_ratelimited("retrans_out leaked\n");
3025                 }
3026 #endif
3027                 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3028                 tp->retrans_out += tcp_skb_pcount(skb);
3029         }
3030 
3031         /* Save stamp of the first (attempted) retransmit. */
3032         if (!tp->retrans_stamp)
3033                 tp->retrans_stamp = tcp_skb_timestamp(skb);
3034 
3035         if (tp->undo_retrans < 0)
3036                 tp->undo_retrans = 0;
3037         tp->undo_retrans += tcp_skb_pcount(skb);
3038         return err;
3039 }
3040 
3041 /* This gets called after a retransmit timeout, and the initially
3042  * retransmitted data is acknowledged.  It tries to continue
3043  * resending the rest of the retransmit queue, until either
3044  * we've sent it all or the congestion window limit is reached.
3045  */
3046 void tcp_xmit_retransmit_queue(struct sock *sk)
3047 {
3048         const struct inet_connection_sock *icsk = inet_csk(sk);
3049         struct sk_buff *skb, *rtx_head, *hole = NULL;
3050         struct tcp_sock *tp = tcp_sk(sk);
3051         u32 max_segs;
3052         int mib_idx;
3053 
3054         if (!tp->packets_out)
3055                 return;
3056 
3057         rtx_head = tcp_rtx_queue_head(sk);
3058         skb = tp->retransmit_skb_hint ?: rtx_head;
3059         max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3060         skb_rbtree_walk_from(skb) {
3061                 __u8 sacked;
3062                 int segs;
3063 
3064                 if (tcp_pacing_check(sk))
3065                         break;
3066 
3067                 /* we could do better than to assign each time */
3068                 if (!hole)
3069                         tp->retransmit_skb_hint = skb;
3070 
3071                 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3072                 if (segs <= 0)
3073                         return;
3074                 sacked = TCP_SKB_CB(skb)->sacked;
3075                 /* In case tcp_shift_skb_data() have aggregated large skbs,
3076                  * we need to make sure not sending too bigs TSO packets
3077                  */
3078                 segs = min_t(int, segs, max_segs);
3079 
3080                 if (tp->retrans_out >= tp->lost_out) {
3081                         break;
3082                 } else if (!(sacked & TCPCB_LOST)) {
3083                         if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3084                                 hole = skb;
3085                         continue;
3086 
3087                 } else {
3088                         if (icsk->icsk_ca_state != TCP_CA_Loss)
3089                                 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3090                         else
3091                                 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3092                 }
3093 
3094                 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3095                         continue;
3096 
3097                 if (tcp_small_queue_check(sk, skb, 1))
3098                         return;
3099 
3100                 if (tcp_retransmit_skb(sk, skb, segs))
3101                         return;
3102 
3103                 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3104 
3105                 if (tcp_in_cwnd_reduction(sk))
3106                         tp->prr_out += tcp_skb_pcount(skb);
3107 
3108                 if (skb == rtx_head &&
3109                     icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3110                         tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3111                                              inet_csk(sk)->icsk_rto,
3112                                              TCP_RTO_MAX,
3113                                              skb);
3114         }
3115 }
3116 
3117 /* We allow to exceed memory limits for FIN packets to expedite
3118  * connection tear down and (memory) recovery.
3119  * Otherwise tcp_send_fin() could be tempted to either delay FIN
3120  * or even be forced to close flow without any FIN.
3121  * In general, we want to allow one skb per socket to avoid hangs
3122  * with edge trigger epoll()
3123  */
3124 void sk_forced_mem_schedule(struct sock *sk, int size)
3125 {
3126         int amt;
3127 
3128         if (size <= sk->sk_forward_alloc)
3129                 return;
3130         amt = sk_mem_pages(size);
3131         sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3132         sk_memory_allocated_add(sk, amt);
3133 
3134         if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3135                 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3136 }
3137 
3138 /* Send a FIN. The caller locks the socket for us.
3139  * We should try to send a FIN packet really hard, but eventually give up.
3140  */
3141 void tcp_send_fin(struct sock *sk)
3142 {
3143         struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3144         struct tcp_sock *tp = tcp_sk(sk);
3145 
3146         /* Optimization, tack on the FIN if we have one skb in write queue and
3147          * this skb was not yet sent, or we are under memory pressure.
3148          * Note: in the latter case, FIN packet will be sent after a timeout,
3149          * as TCP stack thinks it has already been transmitted.
3150          */
3151         if (!tskb && tcp_under_memory_pressure(sk))
3152                 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3153 
3154         if (tskb) {
3155                 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3156                 TCP_SKB_CB(tskb)->end_seq++;
3157                 tp->write_seq++;
3158                 if (tcp_write_queue_empty(sk)) {
3159                         /* This means tskb was already sent.
3160                          * Pretend we included the FIN on previous transmit.
3161                          * We need to set tp->snd_nxt to the value it would have
3162                          * if FIN had been sent. This is because retransmit path
3163                          * does not change tp->snd_nxt.
3164                          */
3165                         WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3166                         return;
3167                 }
3168         } else {
3169                 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3170                 if (unlikely(!skb))
3171                         return;
3172 
3173                 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3174                 skb_reserve(skb, MAX_TCP_HEADER);
3175                 sk_forced_mem_schedule(sk, skb->truesize);
3176                 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3177                 tcp_init_nondata_skb(skb, tp->write_seq,
3178                                      TCPHDR_ACK | TCPHDR_FIN);
3179                 tcp_queue_skb(sk, skb);
3180         }
3181         __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3182 }
3183 
3184 /* We get here when a process closes a file descriptor (either due to
3185  * an explicit close() or as a byproduct of exit()'ing) and there
3186  * was unread data in the receive queue.  This behavior is recommended
3187  * by RFC 2525, section 2.17.  -DaveM
3188  */
3189 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3190 {
3191         struct sk_buff *skb;
3192 
3193         TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3194 
3195         /* NOTE: No TCP options attached and we never retransmit this. */
3196         skb = alloc_skb(MAX_TCP_HEADER, priority);
3197         if (!skb) {
3198                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3199                 return;
3200         }
3201 
3202         /* Reserve space for headers and prepare control bits. */
3203         skb_reserve(skb, MAX_TCP_HEADER);
3204         tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3205                              TCPHDR_ACK | TCPHDR_RST);
3206         tcp_mstamp_refresh(tcp_sk(sk));
3207         /* Send it off. */
3208         if (tcp_transmit_skb(sk, skb, 0, priority))
3209                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3210 
3211         /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3212          * skb here is different to the troublesome skb, so use NULL
3213          */
3214         trace_tcp_send_reset(sk, NULL);
3215 }
3216 
3217 /* Send a crossed SYN-ACK during socket establishment.
3218  * WARNING: This routine must only be called when we have already sent
3219  * a SYN packet that crossed the incoming SYN that caused this routine
3220  * to get called. If this assumption fails then the initial rcv_wnd
3221  * and rcv_wscale values will not be correct.
3222  */
3223 int tcp_send_synack(struct sock *sk)
3224 {
3225         struct sk_buff *skb;
3226 
3227         skb = tcp_rtx_queue_head(sk);
3228         if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3229                 pr_err("%s: wrong queue state\n", __func__);
3230                 return -EFAULT;
3231         }
3232         if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3233                 if (skb_cloned(skb)) {
3234                         struct sk_buff *nskb;
3235 
3236                         tcp_skb_tsorted_save(skb) {
3237                                 nskb = skb_copy(skb, GFP_ATOMIC);
3238                         } tcp_skb_tsorted_restore(skb);
3239                         if (!nskb)
3240                                 return -ENOMEM;
3241                         INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3242                         tcp_highest_sack_replace(sk, skb, nskb);
3243                         tcp_rtx_queue_unlink_and_free(skb, sk);
3244                         __skb_header_release(nskb);
3245                         tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3246                         sk_wmem_queued_add(sk, nskb->truesize);
3247                         sk_mem_charge(sk, nskb->truesize);
3248                         skb = nskb;
3249                 }
3250 
3251                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3252                 tcp_ecn_send_synack(sk, skb);
3253         }
3254         return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3255 }
3256 
3257 /**
3258  * tcp_make_synack - Prepare a SYN-ACK.
3259  * sk: listener socket
3260  * dst: dst entry attached to the SYNACK
3261  * req: request_sock pointer
3262  *
3263  * Allocate one skb and build a SYNACK packet.
3264  * @dst is consumed : Caller should not use it again.
3265  */
3266 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3267                                 struct request_sock *req,
3268                                 struct tcp_fastopen_cookie *foc,
3269                                 enum tcp_synack_type synack_type)
3270 {
3271         struct inet_request_sock *ireq = inet_rsk(req);
3272         const struct tcp_sock *tp = tcp_sk(sk);
3273         struct tcp_md5sig_key *md5 = NULL;
3274         struct tcp_out_options opts;
3275         struct sk_buff *skb;
3276         int tcp_header_size;
3277         struct tcphdr *th;
3278         int mss;
3279         u64 now;
3280 
3281         skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3282         if (unlikely(!skb)) {
3283                 dst_release(dst);
3284                 return NULL;
3285         }
3286         /* Reserve space for headers. */
3287         skb_reserve(skb, MAX_TCP_HEADER);
3288 
3289         switch (synack_type) {
3290         case TCP_SYNACK_NORMAL:
3291                 skb_set_owner_w(skb, req_to_sk(req));
3292                 break;
3293         case TCP_SYNACK_COOKIE:
3294                 /* Under synflood, we do not attach skb to a socket,
3295                  * to avoid false sharing.
3296                  */
3297                 break;
3298         case TCP_SYNACK_FASTOPEN:
3299                 /* sk is a const pointer, because we want to express multiple
3300                  * cpu might call us concurrently.
3301                  * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3302                  */
3303                 skb_set_owner_w(skb, (struct sock *)sk);
3304                 break;
3305         }
3306         skb_dst_set(skb, dst);
3307 
3308         mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3309 
3310         memset(&opts, 0, sizeof(opts));
3311         now = tcp_clock_ns();
3312 #ifdef CONFIG_SYN_COOKIES
3313         if (unlikely(req->cookie_ts))
3314                 skb->skb_mstamp_ns = cookie_init_timestamp(req);
3315         else
3316 #endif
3317         {
3318                 skb->skb_mstamp_ns = now;
3319                 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3320                         tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3321         }
3322 
3323 #ifdef CONFIG_TCP_MD5SIG
3324         rcu_read_lock();
3325         md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3326 #endif
3327         skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3328         tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3329                                              foc) + sizeof(*th);
3330 
3331         skb_push(skb, tcp_header_size);
3332         skb_reset_transport_header(skb);
3333 
3334         th = (struct tcphdr *)skb->data;
3335         memset(th, 0, sizeof(struct tcphdr));
3336         th->syn = 1;
3337         th->ack = 1;
3338         tcp_ecn_make_synack(req, th);
3339         th->source = htons(ireq->ir_num);
3340         th->dest = ireq->ir_rmt_port;
3341         skb->mark = ireq->ir_mark;
3342         skb->ip_summed = CHECKSUM_PARTIAL;
3343         th->seq = htonl(tcp_rsk(req)->snt_isn);
3344         /* XXX data is queued and acked as is. No buffer/window check */
3345         th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3346 
3347         /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3348         th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3349         tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3350         th->doff = (tcp_header_size >> 2);
3351         __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3352 
3353 #ifdef CONFIG_TCP_MD5SIG
3354         /* Okay, we have all we need - do the md5 hash if needed */
3355         if (md5)
3356                 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3357                                                md5, req_to_sk(req), skb);
3358         rcu_read_unlock();
3359 #endif
3360 
3361         skb->skb_mstamp_ns = now;
3362         tcp_add_tx_delay(skb, tp);
3363 
3364         return skb;
3365 }
3366 EXPORT_SYMBOL(tcp_make_synack);
3367 
3368 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3369 {
3370         struct inet_connection_sock *icsk = inet_csk(sk);
3371         const struct tcp_congestion_ops *ca;
3372         u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3373 
3374         if (ca_key == TCP_CA_UNSPEC)
3375                 return;
3376 
3377         rcu_read_lock();
3378         ca = tcp_ca_find_key(ca_key);
3379         if (likely(ca && try_module_get(ca->owner))) {
3380                 module_put(icsk->icsk_ca_ops->owner);
3381                 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3382                 icsk->icsk_ca_ops = ca;
3383         }
3384         rcu_read_unlock();
3385 }
3386 
3387 /* Do all connect socket setups that can be done AF independent. */
3388 static void tcp_connect_init(struct sock *sk)
3389 {
3390         const struct dst_entry *dst = __sk_dst_get(sk);
3391         struct tcp_sock *tp = tcp_sk(sk);
3392         __u8 rcv_wscale;
3393         u32 rcv_wnd;
3394 
3395         /* We'll fix this up when we get a response from the other end.
3396          * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3397          */
3398         tp->tcp_header_len = sizeof(struct tcphdr);
3399         if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3400                 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3401 
3402 #ifdef CONFIG_TCP_MD5SIG
3403         if (tp->af_specific->md5_lookup(sk, sk))
3404                 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3405 #endif
3406 
3407         /* If user gave his TCP_MAXSEG, record it to clamp */
3408         if (tp->rx_opt.user_mss)
3409                 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3410         tp->max_window = 0;
3411         tcp_mtup_init(sk);
3412         tcp_sync_mss(sk, dst_mtu(dst));
3413 
3414         tcp_ca_dst_init(sk, dst);
3415 
3416         if (!tp->window_clamp)
3417                 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3418         tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3419 
3420         tcp_initialize_rcv_mss(sk);
3421 
3422         /* limit the window selection if the user enforce a smaller rx buffer */
3423         if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3424             (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3425                 tp->window_clamp = tcp_full_space(sk);
3426 
3427         rcv_wnd = tcp_rwnd_init_bpf(sk);
3428         if (rcv_wnd == 0)
3429                 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3430 
3431         tcp_select_initial_window(sk, tcp_full_space(sk),
3432                                   tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3433                                   &tp->rcv_wnd,
3434                                   &tp->window_clamp,
3435                                   sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3436                                   &rcv_wscale,
3437                                   rcv_wnd);
3438 
3439         tp->rx_opt.rcv_wscale = rcv_wscale;
3440         tp->rcv_ssthresh = tp->rcv_wnd;
3441 
3442         sk->sk_err = 0;
3443         sock_reset_flag(sk, SOCK_DONE);
3444         tp->snd_wnd = 0;
3445         tcp_init_wl(tp, 0);
3446         tcp_write_queue_purge(sk);
3447         tp->snd_una = tp->write_seq;
3448         tp->snd_sml = tp->write_seq;
3449         tp->snd_up = tp->write_seq;
3450         WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3451 
3452         if (likely(!tp->repair))
3453                 tp->rcv_nxt = 0;
3454         else
3455                 tp->rcv_tstamp = tcp_jiffies32;
3456         tp->rcv_wup = tp->rcv_nxt;
3457         WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3458 
3459         inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3460         inet_csk(sk)->icsk_retransmits = 0;
3461         tcp_clear_retrans(tp);
3462 }
3463 
3464 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3465 {
3466         struct tcp_sock *tp = tcp_sk(sk);
3467         struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3468 
3469         tcb->end_seq += skb->len;
3470         __skb_header_release(skb);
3471         sk_wmem_queued_add(sk, skb->truesize);
3472         sk_mem_charge(sk, skb->truesize);
3473         WRITE_ONCE(tp->write_seq, tcb->end_seq);
3474         tp->packets_out += tcp_skb_pcount(skb);
3475 }
3476 
3477 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3478  * queue a data-only packet after the regular SYN, such that regular SYNs
3479  * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3480  * only the SYN sequence, the data are retransmitted in the first ACK.
3481  * If cookie is not cached or other error occurs, falls back to send a
3482  * regular SYN with Fast Open cookie request option.
3483  */
3484 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3485 {
3486         struct tcp_sock *tp = tcp_sk(sk);
3487         struct tcp_fastopen_request *fo = tp->fastopen_req;
3488         int space, err = 0;
3489         struct sk_buff *syn_data;
3490 
3491         tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3492         if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3493                 goto fallback;
3494 
3495         /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3496          * user-MSS. Reserve maximum option space for middleboxes that add
3497          * private TCP options. The cost is reduced data space in SYN :(
3498          */
3499         tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3500 
3501         space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3502                 MAX_TCP_OPTION_SPACE;
3503 
3504         space = min_t(size_t, space, fo->size);
3505 
3506         /* limit to order-0 allocations */
3507         space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3508 
3509         syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3510         if (!syn_data)
3511                 goto fallback;
3512         syn_data->ip_summed = CHECKSUM_PARTIAL;
3513         memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3514         if (space) {
3515                 int copied = copy_from_iter(skb_put(syn_data, space), space,
3516                                             &fo->data->msg_iter);
3517                 if (unlikely(!copied)) {
3518                         tcp_skb_tsorted_anchor_cleanup(syn_data);
3519                         kfree_skb(syn_data);
3520                         goto fallback;
3521                 }
3522                 if (copied != space) {
3523                         skb_trim(syn_data, copied);
3524                         space = copied;
3525                 }
3526                 skb_zcopy_set(syn_data, fo->uarg, NULL);
3527         }
3528         /* No more data pending in inet_wait_for_connect() */
3529         if (space == fo->size)
3530                 fo->data = NULL;
3531         fo->copied = space;
3532 
3533         tcp_connect_queue_skb(sk, syn_data);
3534         if (syn_data->len)
3535                 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3536 
3537         err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3538 
3539         syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3540 
3541         /* Now full SYN+DATA was cloned and sent (or not),
3542          * remove the SYN from the original skb (syn_data)
3543          * we keep in write queue in case of a retransmit, as we
3544          * also have the SYN packet (with no data) in the same queue.
3545          */
3546         TCP_SKB_CB(syn_data)->seq++;
3547         TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3548         if (!err) {
3549                 tp->syn_data = (fo->copied > 0);
3550                 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3551                 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3552                 goto done;
3553         }
3554 
3555         /* data was not sent, put it in write_queue */
3556         __skb_queue_tail(&sk->sk_write_queue, syn_data);
3557         tp->packets_out -= tcp_skb_pcount(syn_data);
3558 
3559 fallback:
3560         /* Send a regular SYN with Fast Open cookie request option */
3561         if (fo->cookie.len > 0)
3562                 fo->cookie.len = 0;
3563         err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3564         if (err)
3565                 tp->syn_fastopen = 0;
3566 done:
3567         fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3568         return err;
3569 }
3570 
3571 /* Build a SYN and send it off. */
3572 int tcp_connect(struct sock *sk)
3573 {
3574         struct tcp_sock *tp = tcp_sk(sk);
3575         struct sk_buff *buff;
3576         int err;
3577 
3578         tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3579 
3580         if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3581                 return -EHOSTUNREACH; /* Routing failure or similar. */
3582 
3583         tcp_connect_init(sk);
3584 
3585         if (unlikely(tp->repair)) {
3586                 tcp_finish_connect(sk, NULL);
3587                 return 0;
3588         }
3589 
3590         buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3591         if (unlikely(!buff))
3592                 return -ENOBUFS;
3593 
3594         tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3595         tcp_mstamp_refresh(tp);
3596         tp->retrans_stamp = tcp_time_stamp(tp);
3597         tcp_connect_queue_skb(sk, buff);
3598         tcp_ecn_send_syn(sk, buff);
3599         tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3600 
3601         /* Send off SYN; include data in Fast Open. */
3602         err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3603               tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3604         if (err == -ECONNREFUSED)
3605                 return err;
3606 
3607         /* We change tp->snd_nxt after the tcp_transmit_skb() call
3608          * in order to make this packet get counted in tcpOutSegs.
3609          */
3610         WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3611         tp->pushed_seq = tp->write_seq;
3612         buff = tcp_send_head(sk);
3613         if (unlikely(buff)) {
3614                 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3615                 tp->pushed_seq  = TCP_SKB_CB(buff)->seq;
3616         }
3617         TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3618 
3619         /* Timer for repeating the SYN until an answer. */
3620         inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3621                                   inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3622         return 0;
3623 }
3624 EXPORT_SYMBOL(tcp_connect);
3625 
3626 /* Send out a delayed ack, the caller does the policy checking
3627  * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3628  * for details.
3629  */
3630 void tcp_send_delayed_ack(struct sock *sk)
3631 {
3632         struct inet_connection_sock *icsk = inet_csk(sk);
3633         int ato = icsk->icsk_ack.ato;
3634         unsigned long timeout;
3635 
3636         if (ato > TCP_DELACK_MIN) {
3637                 const struct tcp_sock *tp = tcp_sk(sk);
3638                 int max_ato = HZ / 2;
3639 
3640                 if (inet_csk_in_pingpong_mode(sk) ||
3641                     (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3642                         max_ato = TCP_DELACK_MAX;
3643 
3644                 /* Slow path, intersegment interval is "high". */
3645 
3646                 /* If some rtt estimate is known, use it to bound delayed ack.
3647                  * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3648                  * directly.
3649                  */
3650                 if (tp->srtt_us) {
3651                         int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3652                                         TCP_DELACK_MIN);
3653 
3654                         if (rtt < max_ato)
3655                                 max_ato = rtt;
3656                 }
3657 
3658                 ato = min(ato, max_ato);
3659         }
3660 
3661         /* Stay within the limit we were given */
3662         timeout = jiffies + ato;
3663 
3664         /* Use new timeout only if there wasn't a older one earlier. */
3665         if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3666                 /* If delack timer was blocked or is about to expire,
3667                  * send ACK now.
3668                  */
3669                 if (icsk->icsk_ack.blocked ||
3670                     time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3671                         tcp_send_ack(sk);
3672                         return;
3673                 }
3674 
3675                 if (!time_before(timeout, icsk->icsk_ack.timeout))
3676                         timeout = icsk->icsk_ack.timeout;
3677         }
3678         icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3679         icsk->icsk_ack.timeout = timeout;
3680         sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3681 }
3682 
3683 /* This routine sends an ack and also updates the window. */
3684 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3685 {
3686         struct sk_buff *buff;
3687 
3688         /* If we have been reset, we may not send again. */
3689         if (sk->sk_state == TCP_CLOSE)
3690                 return;
3691 
3692         /* We are not putting this on the write queue, so
3693          * tcp_transmit_skb() will set the ownership to this
3694          * sock.
3695          */
3696         buff = alloc_skb(MAX_TCP_HEADER,
3697                          sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3698         if (unlikely(!buff)) {
3699                 inet_csk_schedule_ack(sk);
3700                 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3701                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3702                                           TCP_DELACK_MAX, TCP_RTO_MAX);
3703                 return;
3704         }
3705 
3706         /* Reserve space for headers and prepare control bits. */
3707         skb_reserve(buff, MAX_TCP_HEADER);
3708         tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3709 
3710         /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3711          * too much.
3712          * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3713          */
3714         skb_set_tcp_pure_ack(buff);
3715 
3716         /* Send it off, this clears delayed acks for us. */
3717         __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3718 }
3719 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3720 
3721 void tcp_send_ack(struct sock *sk)
3722 {
3723         __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3724 }
3725 
3726 /* This routine sends a packet with an out of date sequence
3727  * number. It assumes the other end will try to ack it.
3728  *
3729  * Question: what should we make while urgent mode?
3730  * 4.4BSD forces sending single byte of data. We cannot send
3731  * out of window data, because we have SND.NXT==SND.MAX...
3732  *
3733  * Current solution: to send TWO zero-length segments in urgent mode:
3734  * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3735  * out-of-date with SND.UNA-1 to probe window.
3736  */
3737 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3738 {
3739         struct tcp_sock *tp = tcp_sk(sk);
3740         struct sk_buff *skb;
3741 
3742         /* We don't queue it, tcp_transmit_skb() sets ownership. */
3743         skb = alloc_skb(MAX_TCP_HEADER,
3744                         sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3745         if (!skb)
3746                 return -1;
3747 
3748         /* Reserve space for headers and set control bits. */
3749         skb_reserve(skb, MAX_TCP_HEADER);
3750         /* Use a previous sequence.  This should cause the other
3751          * end to send an ack.  Don't queue or clone SKB, just
3752          * send it.
3753          */
3754         tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3755         NET_INC_STATS(sock_net(sk), mib);
3756         return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3757 }
3758 
3759 /* Called from setsockopt( ... TCP_REPAIR ) */
3760 void tcp_send_window_probe(struct sock *sk)
3761 {
3762         if (sk->sk_state == TCP_ESTABLISHED) {
3763                 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3764                 tcp_mstamp_refresh(tcp_sk(sk));
3765                 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3766         }
3767 }
3768 
3769 /* Initiate keepalive or window probe from timer. */
3770 int tcp_write_wakeup(struct sock *sk, int mib)
3771 {
3772         struct tcp_sock *tp = tcp_sk(sk);
3773         struct sk_buff *skb;
3774 
3775         if (sk->sk_state == TCP_CLOSE)
3776                 return -1;
3777 
3778         skb = tcp_send_head(sk);
3779         if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3780                 int err;
3781                 unsigned int mss = tcp_current_mss(sk);
3782                 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3783 
3784                 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3785                         tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3786 
3787                 /* We are probing the opening of a window
3788                  * but the window size is != 0
3789                  * must have been a result SWS avoidance ( sender )
3790                  */
3791                 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3792                     skb->len > mss) {
3793                         seg_size = min(seg_size, mss);
3794                         TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3795                         if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3796                                          skb, seg_size, mss, GFP_ATOMIC))
3797                                 return -1;
3798                 } else if (!tcp_skb_pcount(skb))
3799                         tcp_set_skb_tso_segs(skb, mss);
3800 
3801                 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3802                 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3803                 if (!err)
3804                         tcp_event_new_data_sent(sk, skb);
3805                 return err;
3806         } else {
3807                 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3808                         tcp_xmit_probe_skb(sk, 1, mib);
3809                 return tcp_xmit_probe_skb(sk, 0, mib);
3810         }
3811 }
3812 
3813 /* A window probe timeout has occurred.  If window is not closed send
3814  * a partial packet else a zero probe.
3815  */
3816 void tcp_send_probe0(struct sock *sk)
3817 {
3818         struct inet_connection_sock *icsk = inet_csk(sk);
3819         struct tcp_sock *tp = tcp_sk(sk);
3820         struct net *net = sock_net(sk);
3821         unsigned long timeout;
3822         int err;
3823 
3824         err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3825 
3826         if (tp->packets_out || tcp_write_queue_empty(sk)) {
3827                 /* Cancel probe timer, if it is not required. */
3828                 icsk->icsk_probes_out = 0;
3829                 icsk->icsk_backoff = 0;
3830                 return;
3831         }
3832 
3833         icsk->icsk_probes_out++;
3834         if (err <= 0) {
3835                 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3836                         icsk->icsk_backoff++;
3837                 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
3838         } else {
3839                 /* If packet was not sent due to local congestion,
3840                  * Let senders fight for local resources conservatively.
3841                  */
3842                 timeout = TCP_RESOURCE_PROBE_INTERVAL;
3843         }
3844         tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
3845 }
3846 
3847 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3848 {
3849         const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3850         struct flowi fl;
3851         int res;
3852 
3853         tcp_rsk(req)->txhash = net_tx_rndhash();
3854         res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3855         if (!res) {
3856                 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3857                 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3858                 if (unlikely(tcp_passive_fastopen(sk)))
3859                         tcp_sk(sk)->total_retrans++;
3860                 trace_tcp_retransmit_synack(sk, req);
3861         }
3862         return res;
3863 }
3864 EXPORT_SYMBOL(tcp_rtx_synack);

/* [<][>][^][v][top][bottom][index][help] */