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 }; 439