1 /* 2 * Multifd common code 3 * 4 * Copyright (c) 2019-2020 Red Hat Inc 5 * 6 * Authors: 7 * Juan Quintela <quintela@redhat.com> 8 * 9 * This work is licensed under the terms of the GNU GPL, version 2 or later. 10 * See the COPYING file in the top-level directory. 11 */ 12 13 #include "qemu/osdep.h" 14 #include "qemu/rcu.h" 15 #include "exec/target_page.h" 16 #include "sysemu/sysemu.h" 17 #include "exec/ramblock.h" 18 #include "qemu/error-report.h" 19 #include "qapi/error.h" 20 #include "ram.h" 21 #include "migration.h" 22 #include "socket.h" 23 #include "tls.h" 24 #include "qemu-file.h" 25 #include "trace.h" 26 #include "multifd.h" 27 28 #include "qemu/yank.h" 29 #include "io/channel-socket.h" 30 #include "yank_functions.h" 31 32 /* Multiple fd's */ 33 34 #define MULTIFD_MAGIC 0x11223344U 35 #define MULTIFD_VERSION 1 36 37 typedef struct { 38 uint32_t magic; 39 uint32_t version; 40 unsigned char uuid[16]; /* QemuUUID */ 41 uint8_t id; 42 uint8_t unused1[7]; /* Reserved for future use */ 43 uint64_t unused2[4]; /* Reserved for future use */ 44 } __attribute__((packed)) MultiFDInit_t; 45 46 /* Multifd without compression */ 47 48 /** 49 * nocomp_send_setup: setup send side 50 * 51 * For no compression this function does nothing. 52 * 53 * Returns 0 for success or -1 for error 54 * 55 * @p: Params for the channel that we are using 56 * @errp: pointer to an error 57 */ 58 static int nocomp_send_setup(MultiFDSendParams *p, Error **errp) 59 { 60 return 0; 61 } 62 63 /** 64 * nocomp_send_cleanup: cleanup send side 65 * 66 * For no compression this function does nothing. 67 * 68 * @p: Params for the channel that we are using 69 * @errp: pointer to an error 70 */ 71 static void nocomp_send_cleanup(MultiFDSendParams *p, Error **errp) 72 { 73 return; 74 } 75 76 /** 77 * nocomp_send_prepare: prepare date to be able to send 78 * 79 * For no compression we just have to calculate the size of the 80 * packet. 81 * 82 * Returns 0 for success or -1 for error 83 * 84 * @p: Params for the channel that we are using 85 * @errp: pointer to an error 86 */ 87 static int nocomp_send_prepare(MultiFDSendParams *p, Error **errp) 88 { 89 MultiFDPages_t *pages = p->pages; 90 size_t page_size = qemu_target_page_size(); 91 92 for (int i = 0; i < p->normal_num; i++) { 93 p->iov[p->iovs_num].iov_base = pages->block->host + p->normal[i]; 94 p->iov[p->iovs_num].iov_len = page_size; 95 p->iovs_num++; 96 } 97 98 p->next_packet_size = p->normal_num * page_size; 99 p->flags |= MULTIFD_FLAG_NOCOMP; 100 return 0; 101 } 102 103 /** 104 * nocomp_recv_setup: setup receive side 105 * 106 * For no compression this function does nothing. 107 * 108 * Returns 0 for success or -1 for error 109 * 110 * @p: Params for the channel that we are using 111 * @errp: pointer to an error 112 */ 113 static int nocomp_recv_setup(MultiFDRecvParams *p, Error **errp) 114 { 115 return 0; 116 } 117 118 /** 119 * nocomp_recv_cleanup: setup receive side 120 * 121 * For no compression this function does nothing. 122 * 123 * @p: Params for the channel that we are using 124 */ 125 static void nocomp_recv_cleanup(MultiFDRecvParams *p) 126 { 127 } 128 129 /** 130 * nocomp_recv_pages: read the data from the channel into actual pages 131 * 132 * For no compression we just need to read things into the correct place. 133 * 134 * Returns 0 for success or -1 for error 135 * 136 * @p: Params for the channel that we are using 137 * @errp: pointer to an error 138 */ 139 static int nocomp_recv_pages(MultiFDRecvParams *p, Error **errp) 140 { 141 uint32_t flags = p->flags & MULTIFD_FLAG_COMPRESSION_MASK; 142 size_t page_size = qemu_target_page_size(); 143 144 if (flags != MULTIFD_FLAG_NOCOMP) { 145 error_setg(errp, "multifd %u: flags received %x flags expected %x", 146 p->id, flags, MULTIFD_FLAG_NOCOMP); 147 return -1; 148 } 149 for (int i = 0; i < p->normal_num; i++) { 150 p->iov[i].iov_base = p->host + p->normal[i]; 151 p->iov[i].iov_len = page_size; 152 } 153 return qio_channel_readv_all(p->c, p->iov, p->normal_num, errp); 154 } 155 156 static MultiFDMethods multifd_nocomp_ops = { 157 .send_setup = nocomp_send_setup, 158 .send_cleanup = nocomp_send_cleanup, 159 .send_prepare = nocomp_send_prepare, 160 .recv_setup = nocomp_recv_setup, 161 .recv_cleanup = nocomp_recv_cleanup, 162 .recv_pages = nocomp_recv_pages 163 }; 164 165 static MultiFDMethods *multifd_ops[MULTIFD_COMPRESSION__MAX] = { 166 [MULTIFD_COMPRESSION_NONE] = &multifd_nocomp_ops, 167 }; 168 169 void multifd_register_ops(int method, MultiFDMethods *ops) 170 { 171 assert(0 < method && method < MULTIFD_COMPRESSION__MAX); 172 multifd_ops[method] = ops; 173 } 174 175 static int multifd_send_initial_packet(MultiFDSendParams *p, Error **errp) 176 { 177 MultiFDInit_t msg = {}; 178 int ret; 179 180 msg.magic = cpu_to_be32(MULTIFD_MAGIC); 181 msg.version = cpu_to_be32(MULTIFD_VERSION); 182 msg.id = p->id; 183 memcpy(msg.uuid, &qemu_uuid.data, sizeof(msg.uuid)); 184 185 ret = qio_channel_write_all(p->c, (char *)&msg, sizeof(msg), errp); 186 if (ret != 0) { 187 return -1; 188 } 189 return 0; 190 } 191 192 static int multifd_recv_initial_packet(QIOChannel *c, Error **errp) 193 { 194 MultiFDInit_t msg; 195 int ret; 196 197 ret = qio_channel_read_all(c, (char *)&msg, sizeof(msg), errp); 198 if (ret != 0) { 199 return -1; 200 } 201 202 msg.magic = be32_to_cpu(msg.magic); 203 msg.version = be32_to_cpu(msg.version); 204 205 if (msg.magic != MULTIFD_MAGIC) { 206 error_setg(errp, "multifd: received packet magic %x " 207 "expected %x", msg.magic, MULTIFD_MAGIC); 208 return -1; 209 } 210 211 if (msg.version != MULTIFD_VERSION) { 212 error_setg(errp, "multifd: received packet version %u " 213 "expected %u", msg.version, MULTIFD_VERSION); 214 return -1; 215 } 216 217 if (memcmp(msg.uuid, &qemu_uuid, sizeof(qemu_uuid))) { 218 char *uuid = qemu_uuid_unparse_strdup(&qemu_uuid); 219 char *msg_uuid = qemu_uuid_unparse_strdup((const QemuUUID *)msg.uuid); 220 221 error_setg(errp, "multifd: received uuid '%s' and expected " 222 "uuid '%s' for channel %hhd", msg_uuid, uuid, msg.id); 223 g_free(uuid); 224 g_free(msg_uuid); 225 return -1; 226 } 227 228 if (msg.id > migrate_multifd_channels()) { 229 error_setg(errp, "multifd: received channel version %u " 230 "expected %u", msg.version, MULTIFD_VERSION); 231 return -1; 232 } 233 234 return msg.id; 235 } 236 237 static MultiFDPages_t *multifd_pages_init(size_t size) 238 { 239 MultiFDPages_t *pages = g_new0(MultiFDPages_t, 1); 240 241 pages->allocated = size; 242 pages->offset = g_new0(ram_addr_t, size); 243 244 return pages; 245 } 246 247 static void multifd_pages_clear(MultiFDPages_t *pages) 248 { 249 pages->num = 0; 250 pages->allocated = 0; 251 pages->packet_num = 0; 252 pages->block = NULL; 253 g_free(pages->offset); 254 pages->offset = NULL; 255 g_free(pages); 256 } 257 258 static void multifd_send_fill_packet(MultiFDSendParams *p) 259 { 260 MultiFDPacket_t *packet = p->packet; 261 int i; 262 263 packet->flags = cpu_to_be32(p->flags); 264 packet->pages_alloc = cpu_to_be32(p->pages->allocated); 265 packet->normal_pages = cpu_to_be32(p->normal_num); 266 packet->next_packet_size = cpu_to_be32(p->next_packet_size); 267 packet->packet_num = cpu_to_be64(p->packet_num); 268 269 if (p->pages->block) { 270 strncpy(packet->ramblock, p->pages->block->idstr, 256); 271 } 272 273 for (i = 0; i < p->normal_num; i++) { 274 /* there are architectures where ram_addr_t is 32 bit */ 275 uint64_t temp = p->normal[i]; 276 277 packet->offset[i] = cpu_to_be64(temp); 278 } 279 } 280 281 static int multifd_recv_unfill_packet(MultiFDRecvParams *p, Error **errp) 282 { 283 MultiFDPacket_t *packet = p->packet; 284 size_t page_size = qemu_target_page_size(); 285 uint32_t page_count = MULTIFD_PACKET_SIZE / page_size; 286 RAMBlock *block; 287 int i; 288 289 packet->magic = be32_to_cpu(packet->magic); 290 if (packet->magic != MULTIFD_MAGIC) { 291 error_setg(errp, "multifd: received packet " 292 "magic %x and expected magic %x", 293 packet->magic, MULTIFD_MAGIC); 294 return -1; 295 } 296 297 packet->version = be32_to_cpu(packet->version); 298 if (packet->version != MULTIFD_VERSION) { 299 error_setg(errp, "multifd: received packet " 300 "version %u and expected version %u", 301 packet->version, MULTIFD_VERSION); 302 return -1; 303 } 304 305 p->flags = be32_to_cpu(packet->flags); 306 307 packet->pages_alloc = be32_to_cpu(packet->pages_alloc); 308 /* 309 * If we received a packet that is 100 times bigger than expected 310 * just stop migration. It is a magic number. 311 */ 312 if (packet->pages_alloc > page_count) { 313 error_setg(errp, "multifd: received packet " 314 "with size %u and expected a size of %u", 315 packet->pages_alloc, page_count) ; 316 return -1; 317 } 318 319 p->normal_num = be32_to_cpu(packet->normal_pages); 320 if (p->normal_num > packet->pages_alloc) { 321 error_setg(errp, "multifd: received packet " 322 "with %u pages and expected maximum pages are %u", 323 p->normal_num, packet->pages_alloc) ; 324 return -1; 325 } 326 327 p->next_packet_size = be32_to_cpu(packet->next_packet_size); 328 p->packet_num = be64_to_cpu(packet->packet_num); 329 330 if (p->normal_num == 0) { 331 return 0; 332 } 333 334 /* make sure that ramblock is 0 terminated */ 335 packet->ramblock[255] = 0; 336 block = qemu_ram_block_by_name(packet->ramblock); 337 if (!block) { 338 error_setg(errp, "multifd: unknown ram block %s", 339 packet->ramblock); 340 return -1; 341 } 342 343 p->host = block->host; 344 for (i = 0; i < p->normal_num; i++) { 345 uint64_t offset = be64_to_cpu(packet->offset[i]); 346 347 if (offset > (block->used_length - page_size)) { 348 error_setg(errp, "multifd: offset too long %" PRIu64 349 " (max " RAM_ADDR_FMT ")", 350 offset, block->used_length); 351 return -1; 352 } 353 p->normal[i] = offset; 354 } 355 356 return 0; 357 } 358 359 struct { 360 MultiFDSendParams *params; 361 /* array of pages to sent */ 362 MultiFDPages_t *pages; 363 /* global number of generated multifd packets */ 364 uint64_t packet_num; 365 /* send channels ready */ 366 QemuSemaphore channels_ready; 367 /* 368 * Have we already run terminate threads. There is a race when it 369 * happens that we got one error while we are exiting. 370 * We will use atomic operations. Only valid values are 0 and 1. 371 */ 372 int exiting; 373 /* multifd ops */ 374 MultiFDMethods *ops; 375 } *multifd_send_state; 376 377 /* 378 * How we use multifd_send_state->pages and channel->pages? 379 * 380 * We create a pages for each channel, and a main one. Each time that 381 * we need to send a batch of pages we interchange the ones between 382 * multifd_send_state and the channel that is sending it. There are 383 * two reasons for that: 384 * - to not have to do so many mallocs during migration 385 * - to make easier to know what to free at the end of migration 386 * 387 * This way we always know who is the owner of each "pages" struct, 388 * and we don't need any locking. It belongs to the migration thread 389 * or to the channel thread. Switching is safe because the migration 390 * thread is using the channel mutex when changing it, and the channel 391 * have to had finish with its own, otherwise pending_job can't be 392 * false. 393 */ 394 395 static int multifd_send_pages(QEMUFile *f) 396 { 397 int i; 398 static int next_channel; 399 MultiFDSendParams *p = NULL; /* make happy gcc */ 400 MultiFDPages_t *pages = multifd_send_state->pages; 401 uint64_t transferred; 402 403 if (qatomic_read(&multifd_send_state->exiting)) { 404 return -1; 405 } 406 407 qemu_sem_wait(&multifd_send_state->channels_ready); 408 /* 409 * next_channel can remain from a previous migration that was 410 * using more channels, so ensure it doesn't overflow if the 411 * limit is lower now. 412 */ 413 next_channel %= migrate_multifd_channels(); 414 for (i = next_channel;; i = (i + 1) % migrate_multifd_channels()) { 415 p = &multifd_send_state->params[i]; 416 417 qemu_mutex_lock(&p->mutex); 418 if (p->quit) { 419 error_report("%s: channel %d has already quit!", __func__, i); 420 qemu_mutex_unlock(&p->mutex); 421 return -1; 422 } 423 if (!p->pending_job) { 424 p->pending_job++; 425 next_channel = (i + 1) % migrate_multifd_channels(); 426 break; 427 } 428 qemu_mutex_unlock(&p->mutex); 429 } 430 assert(!p->pages->num); 431 assert(!p->pages->block); 432 433 p->packet_num = multifd_send_state->packet_num++; 434 multifd_send_state->pages = p->pages; 435 p->pages = pages; 436 transferred = ((uint64_t) pages->num) * qemu_target_page_size() 437 + p->packet_len; 438 qemu_file_update_transfer(f, transferred); 439 ram_counters.multifd_bytes += transferred; 440 ram_counters.transferred += transferred; 441 qemu_mutex_unlock(&p->mutex); 442 qemu_sem_post(&p->sem); 443 444 return 1; 445 } 446 447 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset) 448 { 449 MultiFDPages_t *pages = multifd_send_state->pages; 450 451 if (!pages->block) { 452 pages->block = block; 453 } 454 455 if (pages->block == block) { 456 pages->offset[pages->num] = offset; 457 pages->num++; 458 459 if (pages->num < pages->allocated) { 460 return 1; 461 } 462 } 463 464 if (multifd_send_pages(f) < 0) { 465 return -1; 466 } 467 468 if (pages->block != block) { 469 return multifd_queue_page(f, block, offset); 470 } 471 472 return 1; 473 } 474 475 static void multifd_send_terminate_threads(Error *err) 476 { 477 int i; 478 479 trace_multifd_send_terminate_threads(err != NULL); 480 481 if (err) { 482 MigrationState *s = migrate_get_current(); 483 migrate_set_error(s, err); 484 if (s->state == MIGRATION_STATUS_SETUP || 485 s->state == MIGRATION_STATUS_PRE_SWITCHOVER || 486 s->state == MIGRATION_STATUS_DEVICE || 487 s->state == MIGRATION_STATUS_ACTIVE) { 488 migrate_set_state(&s->state, s->state, 489 MIGRATION_STATUS_FAILED); 490 } 491 } 492 493 /* 494 * We don't want to exit each threads twice. Depending on where 495 * we get the error, or if there are two independent errors in two 496 * threads at the same time, we can end calling this function 497 * twice. 498 */ 499 if (qatomic_xchg(&multifd_send_state->exiting, 1)) { 500 return; 501 } 502 503 for (i = 0; i < migrate_multifd_channels(); i++) { 504 MultiFDSendParams *p = &multifd_send_state->params[i]; 505 506 qemu_mutex_lock(&p->mutex); 507 p->quit = true; 508 qemu_sem_post(&p->sem); 509 if (p->c) { 510 qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL); 511 } 512 qemu_mutex_unlock(&p->mutex); 513 } 514 } 515 516 void multifd_save_cleanup(void) 517 { 518 int i; 519 520 if (!migrate_use_multifd() || !migrate_multi_channels_is_allowed()) { 521 return; 522 } 523 multifd_send_terminate_threads(NULL); 524 for (i = 0; i < migrate_multifd_channels(); i++) { 525 MultiFDSendParams *p = &multifd_send_state->params[i]; 526 527 if (p->running) { 528 qemu_thread_join(&p->thread); 529 } 530 } 531 for (i = 0; i < migrate_multifd_channels(); i++) { 532 MultiFDSendParams *p = &multifd_send_state->params[i]; 533 Error *local_err = NULL; 534 535 if (p->registered_yank) { 536 migration_ioc_unregister_yank(p->c); 537 } 538 socket_send_channel_destroy(p->c); 539 p->c = NULL; 540 qemu_mutex_destroy(&p->mutex); 541 qemu_sem_destroy(&p->sem); 542 qemu_sem_destroy(&p->sem_sync); 543 g_free(p->name); 544 p->name = NULL; 545 multifd_pages_clear(p->pages); 546 p->pages = NULL; 547 p->packet_len = 0; 548 g_free(p->packet); 549 p->packet = NULL; 550 g_free(p->iov); 551 p->iov = NULL; 552 g_free(p->normal); 553 p->normal = NULL; 554 multifd_send_state->ops->send_cleanup(p, &local_err); 555 if (local_err) { 556 migrate_set_error(migrate_get_current(), local_err); 557 error_free(local_err); 558 } 559 } 560 qemu_sem_destroy(&multifd_send_state->channels_ready); 561 g_free(multifd_send_state->params); 562 multifd_send_state->params = NULL; 563 multifd_pages_clear(multifd_send_state->pages); 564 multifd_send_state->pages = NULL; 565 g_free(multifd_send_state); 566 multifd_send_state = NULL; 567 } 568 569 int multifd_send_sync_main(QEMUFile *f) 570 { 571 int i; 572 bool flush_zero_copy; 573 574 if (!migrate_use_multifd()) { 575 return 0; 576 } 577 if (multifd_send_state->pages->num) { 578 if (multifd_send_pages(f) < 0) { 579 error_report("%s: multifd_send_pages fail", __func__); 580 return -1; 581 } 582 } 583 584 /* 585 * When using zero-copy, it's necessary to flush the pages before any of 586 * the pages can be sent again, so we'll make sure the new version of the 587 * pages will always arrive _later_ than the old pages. 588 * 589 * Currently we achieve this by flushing the zero-page requested writes 590 * per ram iteration, but in the future we could potentially optimize it 591 * to be less frequent, e.g. only after we finished one whole scanning of 592 * all the dirty bitmaps. 593 */ 594 595 flush_zero_copy = migrate_use_zero_copy_send(); 596 597 for (i = 0; i < migrate_multifd_channels(); i++) { 598 MultiFDSendParams *p = &multifd_send_state->params[i]; 599 600 trace_multifd_send_sync_main_signal(p->id); 601 602 qemu_mutex_lock(&p->mutex); 603 604 if (p->quit) { 605 error_report("%s: channel %d has already quit", __func__, i); 606 qemu_mutex_unlock(&p->mutex); 607 return -1; 608 } 609 610 p->packet_num = multifd_send_state->packet_num++; 611 p->flags |= MULTIFD_FLAG_SYNC; 612 p->pending_job++; 613 qemu_file_update_transfer(f, p->packet_len); 614 ram_counters.multifd_bytes += p->packet_len; 615 ram_counters.transferred += p->packet_len; 616 qemu_mutex_unlock(&p->mutex); 617 qemu_sem_post(&p->sem); 618 619 if (flush_zero_copy && p->c) { 620 int ret; 621 Error *err = NULL; 622 623 ret = qio_channel_flush(p->c, &err); 624 if (ret < 0) { 625 error_report_err(err); 626 return -1; 627 } 628 } 629 } 630 for (i = 0; i < migrate_multifd_channels(); i++) { 631 MultiFDSendParams *p = &multifd_send_state->params[i]; 632 633 trace_multifd_send_sync_main_wait(p->id); 634 qemu_sem_wait(&p->sem_sync); 635 } 636 trace_multifd_send_sync_main(multifd_send_state->packet_num); 637 638 return 0; 639 } 640 641 static void *multifd_send_thread(void *opaque) 642 { 643 MultiFDSendParams *p = opaque; 644 Error *local_err = NULL; 645 int ret = 0; 646 bool use_zero_copy_send = migrate_use_zero_copy_send(); 647 648 trace_multifd_send_thread_start(p->id); 649 rcu_register_thread(); 650 651 if (multifd_send_initial_packet(p, &local_err) < 0) { 652 ret = -1; 653 goto out; 654 } 655 /* initial packet */ 656 p->num_packets = 1; 657 658 while (true) { 659 qemu_sem_wait(&p->sem); 660 661 if (qatomic_read(&multifd_send_state->exiting)) { 662 break; 663 } 664 qemu_mutex_lock(&p->mutex); 665 666 if (p->pending_job) { 667 uint64_t packet_num = p->packet_num; 668 uint32_t flags = p->flags; 669 p->normal_num = 0; 670 671 if (use_zero_copy_send) { 672 p->iovs_num = 0; 673 } else { 674 p->iovs_num = 1; 675 } 676 677 for (int i = 0; i < p->pages->num; i++) { 678 p->normal[p->normal_num] = p->pages->offset[i]; 679 p->normal_num++; 680 } 681 682 if (p->normal_num) { 683 ret = multifd_send_state->ops->send_prepare(p, &local_err); 684 if (ret != 0) { 685 qemu_mutex_unlock(&p->mutex); 686 break; 687 } 688 } 689 multifd_send_fill_packet(p); 690 p->flags = 0; 691 p->num_packets++; 692 p->total_normal_pages += p->normal_num; 693 p->pages->num = 0; 694 p->pages->block = NULL; 695 qemu_mutex_unlock(&p->mutex); 696 697 trace_multifd_send(p->id, packet_num, p->normal_num, flags, 698 p->next_packet_size); 699 700 if (use_zero_copy_send) { 701 /* Send header first, without zerocopy */ 702 ret = qio_channel_write_all(p->c, (void *)p->packet, 703 p->packet_len, &local_err); 704 if (ret != 0) { 705 break; 706 } 707 } else { 708 /* Send header using the same writev call */ 709 p->iov[0].iov_len = p->packet_len; 710 p->iov[0].iov_base = p->packet; 711 } 712 713 ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL, 714 0, p->write_flags, &local_err); 715 if (ret != 0) { 716 break; 717 } 718 719 qemu_mutex_lock(&p->mutex); 720 p->pending_job--; 721 qemu_mutex_unlock(&p->mutex); 722 723 if (flags & MULTIFD_FLAG_SYNC) { 724 qemu_sem_post(&p->sem_sync); 725 } 726 qemu_sem_post(&multifd_send_state->channels_ready); 727 } else if (p->quit) { 728 qemu_mutex_unlock(&p->mutex); 729 break; 730 } else { 731 qemu_mutex_unlock(&p->mutex); 732 /* sometimes there are spurious wakeups */ 733 } 734 } 735 736 out: 737 if (local_err) { 738 trace_multifd_send_error(p->id); 739 multifd_send_terminate_threads(local_err); 740 error_free(local_err); 741 } 742 743 /* 744 * Error happen, I will exit, but I can't just leave, tell 745 * who pay attention to me. 746 */ 747 if (ret != 0) { 748 qemu_sem_post(&p->sem_sync); 749 qemu_sem_post(&multifd_send_state->channels_ready); 750 } 751 752 qemu_mutex_lock(&p->mutex); 753 p->running = false; 754 qemu_mutex_unlock(&p->mutex); 755 756 rcu_unregister_thread(); 757 trace_multifd_send_thread_end(p->id, p->num_packets, p->total_normal_pages); 758 759 return NULL; 760 } 761 762 static bool multifd_channel_connect(MultiFDSendParams *p, 763 QIOChannel *ioc, 764 Error *error); 765 766 static void multifd_tls_outgoing_handshake(QIOTask *task, 767 gpointer opaque) 768 { 769 MultiFDSendParams *p = opaque; 770 QIOChannel *ioc = QIO_CHANNEL(qio_task_get_source(task)); 771 Error *err = NULL; 772 773 if (qio_task_propagate_error(task, &err)) { 774 trace_multifd_tls_outgoing_handshake_error(ioc, error_get_pretty(err)); 775 } else { 776 trace_multifd_tls_outgoing_handshake_complete(ioc); 777 } 778 779 if (!multifd_channel_connect(p, ioc, err)) { 780 /* 781 * Error happen, mark multifd_send_thread status as 'quit' although it 782 * is not created, and then tell who pay attention to me. 783 */ 784 p->quit = true; 785 qemu_sem_post(&multifd_send_state->channels_ready); 786 qemu_sem_post(&p->sem_sync); 787 } 788 } 789 790 static void *multifd_tls_handshake_thread(void *opaque) 791 { 792 MultiFDSendParams *p = opaque; 793 QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c); 794 795 qio_channel_tls_handshake(tioc, 796 multifd_tls_outgoing_handshake, 797 p, 798 NULL, 799 NULL); 800 return NULL; 801 } 802 803 static void multifd_tls_channel_connect(MultiFDSendParams *p, 804 QIOChannel *ioc, 805 Error **errp) 806 { 807 MigrationState *s = migrate_get_current(); 808 const char *hostname = s->hostname; 809 QIOChannelTLS *tioc; 810 811 tioc = migration_tls_client_create(s, ioc, hostname, errp); 812 if (!tioc) { 813 return; 814 } 815 816 object_unref(OBJECT(ioc)); 817 trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname); 818 qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing"); 819 p->c = QIO_CHANNEL(tioc); 820 qemu_thread_create(&p->thread, "multifd-tls-handshake-worker", 821 multifd_tls_handshake_thread, p, 822 QEMU_THREAD_JOINABLE); 823 } 824 825 static bool multifd_channel_connect(MultiFDSendParams *p, 826 QIOChannel *ioc, 827 Error *error) 828 { 829 trace_multifd_set_outgoing_channel( 830 ioc, object_get_typename(OBJECT(ioc)), 831 migrate_get_current()->hostname, error); 832 833 if (!error) { 834 if (migrate_use_tls() && 835 !object_dynamic_cast(OBJECT(ioc), 836 TYPE_QIO_CHANNEL_TLS)) { 837 multifd_tls_channel_connect(p, ioc, &error); 838 if (!error) { 839 /* 840 * tls_channel_connect will call back to this 841 * function after the TLS handshake, 842 * so we mustn't call multifd_send_thread until then 843 */ 844 return true; 845 } else { 846 return false; 847 } 848 } else { 849 migration_ioc_register_yank(ioc); 850 p->registered_yank = true; 851 p->c = ioc; 852 qemu_thread_create(&p->thread, p->name, multifd_send_thread, p, 853 QEMU_THREAD_JOINABLE); 854 } 855 return true; 856 } 857 858 return false; 859 } 860 861 static void multifd_new_send_channel_cleanup(MultiFDSendParams *p, 862 QIOChannel *ioc, Error *err) 863 { 864 migrate_set_error(migrate_get_current(), err); 865 /* Error happen, we need to tell who pay attention to me */ 866 qemu_sem_post(&multifd_send_state->channels_ready); 867 qemu_sem_post(&p->sem_sync); 868 /* 869 * Although multifd_send_thread is not created, but main migration 870 * thread neet to judge whether it is running, so we need to mark 871 * its status. 872 */ 873 p->quit = true; 874 object_unref(OBJECT(ioc)); 875 error_free(err); 876 } 877 878 static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque) 879 { 880 MultiFDSendParams *p = opaque; 881 QIOChannel *sioc = QIO_CHANNEL(qio_task_get_source(task)); 882 Error *local_err = NULL; 883 884 trace_multifd_new_send_channel_async(p->id); 885 if (qio_task_propagate_error(task, &local_err)) { 886 goto cleanup; 887 } else { 888 p->c = QIO_CHANNEL(sioc); 889 qio_channel_set_delay(p->c, false); 890 p->running = true; 891 if (!multifd_channel_connect(p, sioc, local_err)) { 892 goto cleanup; 893 } 894 return; 895 } 896 897 cleanup: 898 multifd_new_send_channel_cleanup(p, sioc, local_err); 899 } 900 901 int multifd_save_setup(Error **errp) 902 { 903 int thread_count; 904 uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size(); 905 uint8_t i; 906 907 if (!migrate_use_multifd()) { 908 return 0; 909 } 910 if (!migrate_multi_channels_is_allowed()) { 911 error_setg(errp, "multifd is not supported by current protocol"); 912 return -1; 913 } 914 915 thread_count = migrate_multifd_channels(); 916 multifd_send_state = g_malloc0(sizeof(*multifd_send_state)); 917 multifd_send_state->params = g_new0(MultiFDSendParams, thread_count); 918 multifd_send_state->pages = multifd_pages_init(page_count); 919 qemu_sem_init(&multifd_send_state->channels_ready, 0); 920 qatomic_set(&multifd_send_state->exiting, 0); 921 multifd_send_state->ops = multifd_ops[migrate_multifd_compression()]; 922 923 for (i = 0; i < thread_count; i++) { 924 MultiFDSendParams *p = &multifd_send_state->params[i]; 925 926 qemu_mutex_init(&p->mutex); 927 qemu_sem_init(&p->sem, 0); 928 qemu_sem_init(&p->sem_sync, 0); 929 p->quit = false; 930 p->pending_job = 0; 931 p->id = i; 932 p->pages = multifd_pages_init(page_count); 933 p->packet_len = sizeof(MultiFDPacket_t) 934 + sizeof(uint64_t) * page_count; 935 p->packet = g_malloc0(p->packet_len); 936 p->packet->magic = cpu_to_be32(MULTIFD_MAGIC); 937 p->packet->version = cpu_to_be32(MULTIFD_VERSION); 938 p->name = g_strdup_printf("multifdsend_%d", i); 939 /* We need one extra place for the packet header */ 940 p->iov = g_new0(struct iovec, page_count + 1); 941 p->normal = g_new0(ram_addr_t, page_count); 942 943 if (migrate_use_zero_copy_send()) { 944 p->write_flags = QIO_CHANNEL_WRITE_FLAG_ZERO_COPY; 945 } else { 946 p->write_flags = 0; 947 } 948 949 socket_send_channel_create(multifd_new_send_channel_async, p); 950 } 951 952 for (i = 0; i < thread_count; i++) { 953 MultiFDSendParams *p = &multifd_send_state->params[i]; 954 Error *local_err = NULL; 955 int ret; 956 957 ret = multifd_send_state->ops->send_setup(p, &local_err); 958 if (ret) { 959 error_propagate(errp, local_err); 960 return ret; 961 } 962 } 963 return 0; 964 } 965 966 struct { 967 MultiFDRecvParams *params; 968 /* number of created threads */ 969 int count; 970 /* syncs main thread and channels */ 971 QemuSemaphore sem_sync; 972 /* global number of generated multifd packets */ 973 uint64_t packet_num; 974 /* multifd ops */ 975 MultiFDMethods *ops; 976 } *multifd_recv_state; 977 978 static void multifd_recv_terminate_threads(Error *err) 979 { 980 int i; 981 982 trace_multifd_recv_terminate_threads(err != NULL); 983 984 if (err) { 985 MigrationState *s = migrate_get_current(); 986 migrate_set_error(s, err); 987 if (s->state == MIGRATION_STATUS_SETUP || 988 s->state == MIGRATION_STATUS_ACTIVE) { 989 migrate_set_state(&s->state, s->state, 990 MIGRATION_STATUS_FAILED); 991 } 992 } 993 994 for (i = 0; i < migrate_multifd_channels(); i++) { 995 MultiFDRecvParams *p = &multifd_recv_state->params[i]; 996 997 qemu_mutex_lock(&p->mutex); 998 p->quit = true; 999 /* 1000 * We could arrive here for two reasons: 1001 * - normal quit, i.e. everything went fine, just finished 1002 * - error quit: We close the channels so the channel threads 1003 * finish the qio_channel_read_all_eof() 1004 */ 1005 if (p->c) { 1006 qio_channel_shutdown(p->c, QIO_CHANNEL_SHUTDOWN_BOTH, NULL); 1007 } 1008 qemu_mutex_unlock(&p->mutex); 1009 } 1010 } 1011 1012 int multifd_load_cleanup(Error **errp) 1013 { 1014 int i; 1015 1016 if (!migrate_use_multifd() || !migrate_multi_channels_is_allowed()) { 1017 return 0; 1018 } 1019 multifd_recv_terminate_threads(NULL); 1020 for (i = 0; i < migrate_multifd_channels(); i++) { 1021 MultiFDRecvParams *p = &multifd_recv_state->params[i]; 1022 1023 if (p->running) { 1024 p->quit = true; 1025 /* 1026 * multifd_recv_thread may hung at MULTIFD_FLAG_SYNC handle code, 1027 * however try to wakeup it without harm in cleanup phase. 1028 */ 1029 qemu_sem_post(&p->sem_sync); 1030 qemu_thread_join(&p->thread); 1031 } 1032 } 1033 for (i = 0; i < migrate_multifd_channels(); i++) { 1034 MultiFDRecvParams *p = &multifd_recv_state->params[i]; 1035 1036 migration_ioc_unregister_yank(p->c); 1037 object_unref(OBJECT(p->c)); 1038 p->c = NULL; 1039 qemu_mutex_destroy(&p->mutex); 1040 qemu_sem_destroy(&p->sem_sync); 1041 g_free(p->name); 1042 p->name = NULL; 1043 p->packet_len = 0; 1044 g_free(p->packet); 1045 p->packet = NULL; 1046 g_free(p->iov); 1047 p->iov = NULL; 1048 g_free(p->normal); 1049 p->normal = NULL; 1050 multifd_recv_state->ops->recv_cleanup(p); 1051 } 1052 qemu_sem_destroy(&multifd_recv_state->sem_sync); 1053 g_free(multifd_recv_state->params); 1054 multifd_recv_state->params = NULL; 1055 g_free(multifd_recv_state); 1056 multifd_recv_state = NULL; 1057 1058 return 0; 1059 } 1060 1061 void multifd_recv_sync_main(void) 1062 { 1063 int i; 1064 1065 if (!migrate_use_multifd()) { 1066 return; 1067 } 1068 for (i = 0; i < migrate_multifd_channels(); i++) { 1069 MultiFDRecvParams *p = &multifd_recv_state->params[i]; 1070 1071 trace_multifd_recv_sync_main_wait(p->id); 1072 qemu_sem_wait(&multifd_recv_state->sem_sync); 1073 } 1074 for (i = 0; i < migrate_multifd_channels(); i++) { 1075 MultiFDRecvParams *p = &multifd_recv_state->params[i]; 1076 1077 WITH_QEMU_LOCK_GUARD(&p->mutex) { 1078 if (multifd_recv_state->packet_num < p->packet_num) { 1079 multifd_recv_state->packet_num = p->packet_num; 1080 } 1081 } 1082 trace_multifd_recv_sync_main_signal(p->id); 1083 qemu_sem_post(&p->sem_sync); 1084 } 1085 trace_multifd_recv_sync_main(multifd_recv_state->packet_num); 1086 } 1087 1088 static void *multifd_recv_thread(void *opaque) 1089 { 1090 MultiFDRecvParams *p = opaque; 1091 Error *local_err = NULL; 1092 int ret; 1093 1094 trace_multifd_recv_thread_start(p->id); 1095 rcu_register_thread(); 1096 1097 while (true) { 1098 uint32_t flags; 1099 1100 if (p->quit) { 1101 break; 1102 } 1103 1104 ret = qio_channel_read_all_eof(p->c, (void *)p->packet, 1105 p->packet_len, &local_err); 1106 if (ret == 0) { /* EOF */ 1107 break; 1108 } 1109 if (ret == -1) { /* Error */ 1110 break; 1111 } 1112 1113 qemu_mutex_lock(&p->mutex); 1114 ret = multifd_recv_unfill_packet(p, &local_err); 1115 if (ret) { 1116 qemu_mutex_unlock(&p->mutex); 1117 break; 1118 } 1119 1120 flags = p->flags; 1121 /* recv methods don't know how to handle the SYNC flag */ 1122 p->flags &= ~MULTIFD_FLAG_SYNC; 1123 trace_multifd_recv(p->id, p->packet_num, p->normal_num, flags, 1124 p->next_packet_size); 1125 p->num_packets++; 1126 p->total_normal_pages += p->normal_num; 1127 qemu_mutex_unlock(&p->mutex); 1128 1129 if (p->normal_num) { 1130 ret = multifd_recv_state->ops->recv_pages(p, &local_err); 1131 if (ret != 0) { 1132 break; 1133 } 1134 } 1135 1136 if (flags & MULTIFD_FLAG_SYNC) { 1137 qemu_sem_post(&multifd_recv_state->sem_sync); 1138 qemu_sem_wait(&p->sem_sync); 1139 } 1140 } 1141 1142 if (local_err) { 1143 multifd_recv_terminate_threads(local_err); 1144 error_free(local_err); 1145 } 1146 qemu_mutex_lock(&p->mutex); 1147 p->running = false; 1148 qemu_mutex_unlock(&p->mutex); 1149 1150 rcu_unregister_thread(); 1151 trace_multifd_recv_thread_end(p->id, p->num_packets, p->total_normal_pages); 1152 1153 return NULL; 1154 } 1155 1156 int multifd_load_setup(Error **errp) 1157 { 1158 int thread_count; 1159 uint32_t page_count = MULTIFD_PACKET_SIZE / qemu_target_page_size(); 1160 uint8_t i; 1161 1162 if (!migrate_use_multifd()) { 1163 return 0; 1164 } 1165 if (!migrate_multi_channels_is_allowed()) { 1166 error_setg(errp, "multifd is not supported by current protocol"); 1167 return -1; 1168 } 1169 thread_count = migrate_multifd_channels(); 1170 multifd_recv_state = g_malloc0(sizeof(*multifd_recv_state)); 1171 multifd_recv_state->params = g_new0(MultiFDRecvParams, thread_count); 1172 qatomic_set(&multifd_recv_state->count, 0); 1173 qemu_sem_init(&multifd_recv_state->sem_sync, 0); 1174 multifd_recv_state->ops = multifd_ops[migrate_multifd_compression()]; 1175 1176 for (i = 0; i < thread_count; i++) { 1177 MultiFDRecvParams *p = &multifd_recv_state->params[i]; 1178 1179 qemu_mutex_init(&p->mutex); 1180 qemu_sem_init(&p->sem_sync, 0); 1181 p->quit = false; 1182 p->id = i; 1183 p->packet_len = sizeof(MultiFDPacket_t) 1184 + sizeof(uint64_t) * page_count; 1185 p->packet = g_malloc0(p->packet_len); 1186 p->name = g_strdup_printf("multifdrecv_%d", i); 1187 p->iov = g_new0(struct iovec, page_count); 1188 p->normal = g_new0(ram_addr_t, page_count); 1189 } 1190 1191 for (i = 0; i < thread_count; i++) { 1192 MultiFDRecvParams *p = &multifd_recv_state->params[i]; 1193 Error *local_err = NULL; 1194 int ret; 1195 1196 ret = multifd_recv_state->ops->recv_setup(p, &local_err); 1197 if (ret) { 1198 error_propagate(errp, local_err); 1199 return ret; 1200 } 1201 } 1202 return 0; 1203 } 1204 1205 bool multifd_recv_all_channels_created(void) 1206 { 1207 int thread_count = migrate_multifd_channels(); 1208 1209 if (!migrate_use_multifd()) { 1210 return true; 1211 } 1212 1213 if (!multifd_recv_state) { 1214 /* Called before any connections created */ 1215 return false; 1216 } 1217 1218 return thread_count == qatomic_read(&multifd_recv_state->count); 1219 } 1220 1221 /* 1222 * Try to receive all multifd channels to get ready for the migration. 1223 * - Return true and do not set @errp when correctly receiving all channels; 1224 * - Return false and do not set @errp when correctly receiving the current one; 1225 * - Return false and set @errp when failing to receive the current channel. 1226 */ 1227 bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp) 1228 { 1229 MultiFDRecvParams *p; 1230 Error *local_err = NULL; 1231 int id; 1232 1233 id = multifd_recv_initial_packet(ioc, &local_err); 1234 if (id < 0) { 1235 multifd_recv_terminate_threads(local_err); 1236 error_propagate_prepend(errp, local_err, 1237 "failed to receive packet" 1238 " via multifd channel %d: ", 1239 qatomic_read(&multifd_recv_state->count)); 1240 return false; 1241 } 1242 trace_multifd_recv_new_channel(id); 1243 1244 p = &multifd_recv_state->params[id]; 1245 if (p->c != NULL) { 1246 error_setg(&local_err, "multifd: received id '%d' already setup'", 1247 id); 1248 multifd_recv_terminate_threads(local_err); 1249 error_propagate(errp, local_err); 1250 return false; 1251 } 1252 p->c = ioc; 1253 object_ref(OBJECT(ioc)); 1254 /* initial packet */ 1255 p->num_packets = 1; 1256 1257 p->running = true; 1258 qemu_thread_create(&p->thread, p->name, multifd_recv_thread, p, 1259 QEMU_THREAD_JOINABLE); 1260 qatomic_inc(&multifd_recv_state->count); 1261 return qatomic_read(&multifd_recv_state->count) == 1262 migrate_multifd_channels(); 1263 } 1264