net/vmw_vsock/vmci_transport_notify_qstate.c

   1 /*
   2  * VMware vSockets Driver
   3  *
   4  * Copyright (C) 2009-2013 VMware, Inc. All rights reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License as published by the Free
   8  * Software Foundation version 2 and no later version.
   9  *
  10  * This program is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  13  * more details.
  14  */
  15
  16 #include <linux/types.h>
  17 #include <linux/socket.h>
  18 #include <linux/stddef.h>
  19 #include <net/sock.h>
  20
  21 #include "vmci_transport_notify.h"
  22
  23 #define PKT_FIELD(vsk, field_name) \
  24         (vmci_trans(vsk)->notify.pkt_q_state.field_name)
  25
  26 static bool vmci_transport_notify_waiting_write(struct vsock_sock *vsk)
  27 {
  28         bool retval;
  29         u64 notify_limit;
  30
  31         if (!PKT_FIELD(vsk, peer_waiting_write))
  32                 return false;
  33
  34         /* When the sender blocks, we take that as a sign that the sender is
  35          * faster than the receiver. To reduce the transmit rate of the sender,
  36          * we delay the sending of the read notification by decreasing the
  37          * write_notify_window. The notification is delayed until the number of
  38          * bytes used in the queue drops below the write_notify_window.
  39          */
  40
  41         if (!PKT_FIELD(vsk, peer_waiting_write_detected)) {
  42                 PKT_FIELD(vsk, peer_waiting_write_detected) = true;
  43                 if (PKT_FIELD(vsk, write_notify_window) < PAGE_SIZE) {
  44                         PKT_FIELD(vsk, write_notify_window) =
  45                             PKT_FIELD(vsk, write_notify_min_window);
  46                 } else {
  47                         PKT_FIELD(vsk, write_notify_window) -= PAGE_SIZE;
  48                         if (PKT_FIELD(vsk, write_notify_window) <
  49                             PKT_FIELD(vsk, write_notify_min_window))
  50                                 PKT_FIELD(vsk, write_notify_window) =
  51                                     PKT_FIELD(vsk, write_notify_min_window);
  52
  53                 }
  54         }
  55         notify_limit = vmci_trans(vsk)->consume_size -
  56                 PKT_FIELD(vsk, write_notify_window);
  57
  58         /* The notify_limit is used to delay notifications in the case where
  59          * flow control is enabled. Below the test is expressed in terms of
  60          * free space in the queue: if free_space > ConsumeSize -
  61          * write_notify_window then notify An alternate way of expressing this
  62          * is to rewrite the expression to use the data ready in the receive
  63          * queue: if write_notify_window > bufferReady then notify as
  64          * free_space == ConsumeSize - bufferReady.
  65          */
  66
  67         retval = vmci_qpair_consume_free_space(vmci_trans(vsk)->qpair) >
  68                 notify_limit;
  69
  70         if (retval) {
  71                 /* Once we notify the peer, we reset the detected flag so the
  72                  * next wait will again cause a decrease in the window size.
  73                  */
  74
  75                 PKT_FIELD(vsk, peer_waiting_write_detected) = false;
  76         }
  77         return retval;
  78 }
  79
  80 static void
  81 vmci_transport_handle_read(struct sock *sk,
  82                            struct vmci_transport_packet *pkt,
  83                            bool bottom_half,
  84                            struct sockaddr_vm *dst, struct sockaddr_vm *src)
  85 {
  86         sk->sk_write_space(sk);
  87 }
  88
  89 static void
  90 vmci_transport_handle_wrote(struct sock *sk,
  91                             struct vmci_transport_packet *pkt,
  92                             bool bottom_half,
  93                             struct sockaddr_vm *dst, struct sockaddr_vm *src)
  94 {
  95         sk->sk_data_ready(sk);
  96 }
  97
  98 static void vsock_block_update_write_window(struct sock *sk)
  99 {
 100         struct vsock_sock *vsk = vsock_sk(sk);
 101
 102         if (PKT_FIELD(vsk, write_notify_window) < vmci_trans(vsk)->consume_size)
 103                 PKT_FIELD(vsk, write_notify_window) =
 104                     min(PKT_FIELD(vsk, write_notify_window) + PAGE_SIZE,
 105                         vmci_trans(vsk)->consume_size);
 106 }
 107
 108 static int vmci_transport_send_read_notification(struct sock *sk)
 109 {
 110         struct vsock_sock *vsk;
 111         bool sent_read;
 112         unsigned int retries;
 113         int err;
 114
 115         vsk = vsock_sk(sk);
 116         sent_read = false;
 117         retries = 0;
 118         err = 0;
 119
 120         if (vmci_transport_notify_waiting_write(vsk)) {
 121                 /* Notify the peer that we have read, retrying the send on
 122                  * failure up to our maximum value.  XXX For now we just log
 123                  * the failure, but later we should schedule a work item to
 124                  * handle the resend until it succeeds.  That would require
 125                  * keeping track of work items in the vsk and cleaning them up
 126                  * upon socket close.
 127                  */
 128                 while (!(vsk->peer_shutdown & RCV_SHUTDOWN) &&
 129                        !sent_read &&
 130                        retries < VMCI_TRANSPORT_MAX_DGRAM_RESENDS) {
 131                         err = vmci_transport_send_read(sk);
 132                         if (err >= 0)
 133                                 sent_read = true;
 134
 135                         retries++;
 136                 }
 137
 138                 if (retries >= VMCI_TRANSPORT_MAX_DGRAM_RESENDS && !sent_read)
 139                         pr_err("%p unable to send read notification to peer\n",
 140                                sk);
 141                 else
 142                         PKT_FIELD(vsk, peer_waiting_write) = false;
 143
 144         }
 145         return err;
 146 }
 147
 148 static void vmci_transport_notify_pkt_socket_init(struct sock *sk)
 149 {
 150         struct vsock_sock *vsk = vsock_sk(sk);
 151
 152         PKT_FIELD(vsk, write_notify_window) = PAGE_SIZE;
 153         PKT_FIELD(vsk, write_notify_min_window) = PAGE_SIZE;
 154         PKT_FIELD(vsk, peer_waiting_write) = false;
 155         PKT_FIELD(vsk, peer_waiting_write_detected) = false;
 156 }
 157
 158 static void vmci_transport_notify_pkt_socket_destruct(struct vsock_sock *vsk)
 159 {
 160         PKT_FIELD(vsk, write_notify_window) = PAGE_SIZE;
 161         PKT_FIELD(vsk, write_notify_min_window) = PAGE_SIZE;
 162         PKT_FIELD(vsk, peer_waiting_write) = false;
 163         PKT_FIELD(vsk, peer_waiting_write_detected) = false;
 164 }
 165
 166 static int
 167 vmci_transport_notify_pkt_poll_in(struct sock *sk,
 168                                   size_t target, bool *data_ready_now)
 169 {
 170         struct vsock_sock *vsk = vsock_sk(sk);
 171
 172         if (vsock_stream_has_data(vsk)) {
 173                 *data_ready_now = true;
 174         } else {
 175                 /* We can't read right now because there is nothing in the
 176                  * queue. Ask for notifications when there is something to
 177                  * read.
 178                  */
 179                 if (sk->sk_state == SS_CONNECTED)
 180                         vsock_block_update_write_window(sk);
 181                 *data_ready_now = false;
 182         }
 183
 184         return 0;
 185 }
 186
 187 static int
 188 vmci_transport_notify_pkt_poll_out(struct sock *sk,
 189                                    size_t target, bool *space_avail_now)
 190 {
 191         s64 produce_q_free_space;
 192         struct vsock_sock *vsk = vsock_sk(sk);
 193
 194         produce_q_free_space = vsock_stream_has_space(vsk);
 195         if (produce_q_free_space > 0) {
 196                 *space_avail_now = true;
 197                 return 0;
 198         } else if (produce_q_free_space == 0) {
 199                 /* This is a connected socket but we can't currently send data.
 200                  * Nothing else to do.
 201                  */
 202                 *space_avail_now = false;
 203         }
 204
 205         return 0;
 206 }
 207
 208 static int
 209 vmci_transport_notify_pkt_recv_init(
 210                                 struct sock *sk,
 211                                 size_t target,
 212                                 struct vmci_transport_recv_notify_data *data)
 213 {
 214         struct vsock_sock *vsk = vsock_sk(sk);
 215
 216         data->consume_head = 0;
 217         data->produce_tail = 0;
 218         data->notify_on_block = false;
 219
 220         if (PKT_FIELD(vsk, write_notify_min_window) < target + 1) {
 221                 PKT_FIELD(vsk, write_notify_min_window) = target + 1;
 222                 if (PKT_FIELD(vsk, write_notify_window) <
 223                     PKT_FIELD(vsk, write_notify_min_window)) {
 224                         /* If the current window is smaller than the new
 225                          * minimal window size, we need to reevaluate whether
 226                          * we need to notify the sender. If the number of ready
 227                          * bytes are smaller than the new window, we need to
 228                          * send a notification to the sender before we block.
 229                          */
 230
 231                         PKT_FIELD(vsk, write_notify_window) =
 232                             PKT_FIELD(vsk, write_notify_min_window);
 233                         data->notify_on_block = true;
 234                 }
 235         }
 236
 237         return 0;
 238 }
 239
 240 static int
 241 vmci_transport_notify_pkt_recv_pre_block(
 242                                 struct sock *sk,
 243                                 size_t target,
 244                                 struct vmci_transport_recv_notify_data *data)
 245 {
 246         int err = 0;
 247
 248         vsock_block_update_write_window(sk);
 249
 250         if (data->notify_on_block) {
 251                 err = vmci_transport_send_read_notification(sk);
 252                 if (err < 0)
 253                         return err;
 254                 data->notify_on_block = false;
 255         }
 256
 257         return err;
 258 }
 259
 260 static int
 261 vmci_transport_notify_pkt_recv_post_dequeue(
 262                                 struct sock *sk,
 263                                 size_t target,
 264                                 ssize_t copied,
 265                                 bool data_read,
 266                                 struct vmci_transport_recv_notify_data *data)
 267 {
 268         struct vsock_sock *vsk;
 269         int err;
 270         bool was_full = false;
 271         u64 free_space;
 272
 273         vsk = vsock_sk(sk);
 274         err = 0;
 275
 276         if (data_read) {
 277                 smp_mb();
 278
 279                 free_space =
 280                         vmci_qpair_consume_free_space(vmci_trans(vsk)->qpair);
 281                 was_full = free_space == copied;
 282
 283                 if (was_full)
 284                         PKT_FIELD(vsk, peer_waiting_write) = true;
 285
 286                 err = vmci_transport_send_read_notification(sk);
 287                 if (err < 0)
 288                         return err;
 289
 290                 /* See the comment in
 291                  * vmci_transport_notify_pkt_send_post_enqueue().
 292                  */
 293                 sk->sk_data_ready(sk);
 294         }
 295
 296         return err;
 297 }
 298
 299 static int
 300 vmci_transport_notify_pkt_send_init(
 301                                 struct sock *sk,
 302                                 struct vmci_transport_send_notify_data *data)
 303 {
 304         data->consume_head = 0;
 305         data->produce_tail = 0;
 306
 307         return 0;
 308 }
 309
 310 static int
 311 vmci_transport_notify_pkt_send_post_enqueue(
 312                                 struct sock *sk,
 313                                 ssize_t written,
 314                                 struct vmci_transport_send_notify_data *data)
 315 {
 316         int err = 0;
 317         struct vsock_sock *vsk;
 318         bool sent_wrote = false;
 319         bool was_empty;
 320         int retries = 0;
 321
 322         vsk = vsock_sk(sk);
 323
 324         smp_mb();
 325
 326         was_empty =
 327                 vmci_qpair_produce_buf_ready(vmci_trans(vsk)->qpair) == written;
 328         if (was_empty) {
 329                 while (!(vsk->peer_shutdown & RCV_SHUTDOWN) &&
 330                        !sent_wrote &&
 331                        retries < VMCI_TRANSPORT_MAX_DGRAM_RESENDS) {
 332                         err = vmci_transport_send_wrote(sk);
 333                         if (err >= 0)
 334                                 sent_wrote = true;
 335
 336                         retries++;
 337                 }
 338         }
 339
 340         if (retries >= VMCI_TRANSPORT_MAX_DGRAM_RESENDS && !sent_wrote) {
 341                 pr_err("%p unable to send wrote notification to peer\n",
 342                        sk);
 343                 return err;
 344         }
 345
 346         return err;
 347 }
 348
 349 static void
 350 vmci_transport_notify_pkt_handle_pkt(
 351                                 struct sock *sk,
 352                                 struct vmci_transport_packet *pkt,
 353                                 bool bottom_half,
 354                                 struct sockaddr_vm *dst,
 355                                 struct sockaddr_vm *src, bool *pkt_processed)
 356 {
 357         bool processed = false;
 358
 359         switch (pkt->type) {
 360         case VMCI_TRANSPORT_PACKET_TYPE_WROTE:
 361                 vmci_transport_handle_wrote(sk, pkt, bottom_half, dst, src);
 362                 processed = true;
 363                 break;
 364         case VMCI_TRANSPORT_PACKET_TYPE_READ:
 365                 vmci_transport_handle_read(sk, pkt, bottom_half, dst, src);
 366                 processed = true;
 367                 break;
 368         }
 369
 370         if (pkt_processed)
 371                 *pkt_processed = processed;
 372 }
 373
 374 static void vmci_transport_notify_pkt_process_request(struct sock *sk)
 375 {
 376         struct vsock_sock *vsk = vsock_sk(sk);
 377
 378         PKT_FIELD(vsk, write_notify_window) = vmci_trans(vsk)->consume_size;
 379         if (vmci_trans(vsk)->consume_size <
 380                 PKT_FIELD(vsk, write_notify_min_window))
 381                 PKT_FIELD(vsk, write_notify_min_window) =
 382                         vmci_trans(vsk)->consume_size;
 383 }
 384
 385 static void vmci_transport_notify_pkt_process_negotiate(struct sock *sk)
 386 {
 387         struct vsock_sock *vsk = vsock_sk(sk);
 388
 389         PKT_FIELD(vsk, write_notify_window) = vmci_trans(vsk)->consume_size;
 390         if (vmci_trans(vsk)->consume_size <
 391                 PKT_FIELD(vsk, write_notify_min_window))
 392                 PKT_FIELD(vsk, write_notify_min_window) =
 393                         vmci_trans(vsk)->consume_size;
 394 }
 395
 396 static int
 397 vmci_transport_notify_pkt_recv_pre_dequeue(
 398                                 struct sock *sk,
 399                                 size_t target,
 400                                 struct vmci_transport_recv_notify_data *data)
 401 {
 402         return 0; /* NOP for QState. */
 403 }
 404
 405 static int
 406 vmci_transport_notify_pkt_send_pre_block(
 407                                 struct sock *sk,
 408                                 struct vmci_transport_send_notify_data *data)
 409 {
 410         return 0; /* NOP for QState. */
 411 }
 412
 413 static int
 414 vmci_transport_notify_pkt_send_pre_enqueue(
 415                                 struct sock *sk,
 416                                 struct vmci_transport_send_notify_data *data)
 417 {
 418         return 0; /* NOP for QState. */
 419 }
 420
 421 /* Socket always on control packet based operations. */
 422 struct vmci_transport_notify_ops vmci_transport_notify_pkt_q_state_ops = {
 423         vmci_transport_notify_pkt_socket_init,
 424         vmci_transport_notify_pkt_socket_destruct,
 425         vmci_transport_notify_pkt_poll_in,
 426         vmci_transport_notify_pkt_poll_out,
 427         vmci_transport_notify_pkt_handle_pkt,
 428         vmci_transport_notify_pkt_recv_init,
 429         vmci_transport_notify_pkt_recv_pre_block,
 430         vmci_transport_notify_pkt_recv_pre_dequeue,
 431         vmci_transport_notify_pkt_recv_post_dequeue,
 432         vmci_transport_notify_pkt_send_init,
 433         vmci_transport_notify_pkt_send_pre_block,
 434         vmci_transport_notify_pkt_send_pre_enqueue,
 435         vmci_transport_notify_pkt_send_post_enqueue,
 436         vmci_transport_notify_pkt_process_request,
 437         vmci_transport_notify_pkt_process_negotiate,
 438 };