1 /* 2 * Postcopy migration for RAM 3 * 4 * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates 5 * 6 * Authors: 7 * Dave Gilbert <dgilbert@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 14 /* 15 * Postcopy is a migration technique where the execution flips from the 16 * source to the destination before all the data has been copied. 17 */ 18 19 #include "qemu/osdep.h" 20 #include "qemu/rcu.h" 21 #include "qemu/madvise.h" 22 #include "exec/target_page.h" 23 #include "migration.h" 24 #include "qemu-file.h" 25 #include "savevm.h" 26 #include "postcopy-ram.h" 27 #include "ram.h" 28 #include "qapi/error.h" 29 #include "qemu/notify.h" 30 #include "qemu/rcu.h" 31 #include "sysemu/sysemu.h" 32 #include "qemu/error-report.h" 33 #include "trace.h" 34 #include "hw/boards.h" 35 #include "exec/ramblock.h" 36 #include "socket.h" 37 #include "qemu-file.h" 38 #include "yank_functions.h" 39 #include "tls.h" 40 41 /* Arbitrary limit on size of each discard command, 42 * keeps them around ~200 bytes 43 */ 44 #define MAX_DISCARDS_PER_COMMAND 12 45 46 struct PostcopyDiscardState { 47 const char *ramblock_name; 48 uint16_t cur_entry; 49 /* 50 * Start and length of a discard range (bytes) 51 */ 52 uint64_t start_list[MAX_DISCARDS_PER_COMMAND]; 53 uint64_t length_list[MAX_DISCARDS_PER_COMMAND]; 54 unsigned int nsentwords; 55 unsigned int nsentcmds; 56 }; 57 58 static NotifierWithReturnList postcopy_notifier_list; 59 60 void postcopy_infrastructure_init(void) 61 { 62 notifier_with_return_list_init(&postcopy_notifier_list); 63 } 64 65 void postcopy_add_notifier(NotifierWithReturn *nn) 66 { 67 notifier_with_return_list_add(&postcopy_notifier_list, nn); 68 } 69 70 void postcopy_remove_notifier(NotifierWithReturn *n) 71 { 72 notifier_with_return_remove(n); 73 } 74 75 int postcopy_notify(enum PostcopyNotifyReason reason, Error **errp) 76 { 77 struct PostcopyNotifyData pnd; 78 pnd.reason = reason; 79 pnd.errp = errp; 80 81 return notifier_with_return_list_notify(&postcopy_notifier_list, 82 &pnd); 83 } 84 85 /* 86 * NOTE: this routine is not thread safe, we can't call it concurrently. But it 87 * should be good enough for migration's purposes. 88 */ 89 void postcopy_thread_create(MigrationIncomingState *mis, 90 QemuThread *thread, const char *name, 91 void *(*fn)(void *), int joinable) 92 { 93 qemu_sem_init(&mis->thread_sync_sem, 0); 94 qemu_thread_create(thread, name, fn, mis, joinable); 95 qemu_sem_wait(&mis->thread_sync_sem); 96 qemu_sem_destroy(&mis->thread_sync_sem); 97 } 98 99 /* Postcopy needs to detect accesses to pages that haven't yet been copied 100 * across, and efficiently map new pages in, the techniques for doing this 101 * are target OS specific. 102 */ 103 #if defined(__linux__) 104 105 #include <poll.h> 106 #include <sys/ioctl.h> 107 #include <sys/syscall.h> 108 #include <asm/types.h> /* for __u64 */ 109 #endif 110 111 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD) 112 #include <sys/eventfd.h> 113 #include <linux/userfaultfd.h> 114 115 typedef struct PostcopyBlocktimeContext { 116 /* time when page fault initiated per vCPU */ 117 uint32_t *page_fault_vcpu_time; 118 /* page address per vCPU */ 119 uintptr_t *vcpu_addr; 120 uint32_t total_blocktime; 121 /* blocktime per vCPU */ 122 uint32_t *vcpu_blocktime; 123 /* point in time when last page fault was initiated */ 124 uint32_t last_begin; 125 /* number of vCPU are suspended */ 126 int smp_cpus_down; 127 uint64_t start_time; 128 129 /* 130 * Handler for exit event, necessary for 131 * releasing whole blocktime_ctx 132 */ 133 Notifier exit_notifier; 134 } PostcopyBlocktimeContext; 135 136 static void destroy_blocktime_context(struct PostcopyBlocktimeContext *ctx) 137 { 138 g_free(ctx->page_fault_vcpu_time); 139 g_free(ctx->vcpu_addr); 140 g_free(ctx->vcpu_blocktime); 141 g_free(ctx); 142 } 143 144 static void migration_exit_cb(Notifier *n, void *data) 145 { 146 PostcopyBlocktimeContext *ctx = container_of(n, PostcopyBlocktimeContext, 147 exit_notifier); 148 destroy_blocktime_context(ctx); 149 } 150 151 static struct PostcopyBlocktimeContext *blocktime_context_new(void) 152 { 153 MachineState *ms = MACHINE(qdev_get_machine()); 154 unsigned int smp_cpus = ms->smp.cpus; 155 PostcopyBlocktimeContext *ctx = g_new0(PostcopyBlocktimeContext, 1); 156 ctx->page_fault_vcpu_time = g_new0(uint32_t, smp_cpus); 157 ctx->vcpu_addr = g_new0(uintptr_t, smp_cpus); 158 ctx->vcpu_blocktime = g_new0(uint32_t, smp_cpus); 159 160 ctx->exit_notifier.notify = migration_exit_cb; 161 ctx->start_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); 162 qemu_add_exit_notifier(&ctx->exit_notifier); 163 return ctx; 164 } 165 166 static uint32List *get_vcpu_blocktime_list(PostcopyBlocktimeContext *ctx) 167 { 168 MachineState *ms = MACHINE(qdev_get_machine()); 169 uint32List *list = NULL; 170 int i; 171 172 for (i = ms->smp.cpus - 1; i >= 0; i--) { 173 QAPI_LIST_PREPEND(list, ctx->vcpu_blocktime[i]); 174 } 175 176 return list; 177 } 178 179 /* 180 * This function just populates MigrationInfo from postcopy's 181 * blocktime context. It will not populate MigrationInfo, 182 * unless postcopy-blocktime capability was set. 183 * 184 * @info: pointer to MigrationInfo to populate 185 */ 186 void fill_destination_postcopy_migration_info(MigrationInfo *info) 187 { 188 MigrationIncomingState *mis = migration_incoming_get_current(); 189 PostcopyBlocktimeContext *bc = mis->blocktime_ctx; 190 191 if (!bc) { 192 return; 193 } 194 195 info->has_postcopy_blocktime = true; 196 info->postcopy_blocktime = bc->total_blocktime; 197 info->has_postcopy_vcpu_blocktime = true; 198 info->postcopy_vcpu_blocktime = get_vcpu_blocktime_list(bc); 199 } 200 201 static uint32_t get_postcopy_total_blocktime(void) 202 { 203 MigrationIncomingState *mis = migration_incoming_get_current(); 204 PostcopyBlocktimeContext *bc = mis->blocktime_ctx; 205 206 if (!bc) { 207 return 0; 208 } 209 210 return bc->total_blocktime; 211 } 212 213 /** 214 * receive_ufd_features: check userfault fd features, to request only supported 215 * features in the future. 216 * 217 * Returns: true on success 218 * 219 * __NR_userfaultfd - should be checked before 220 * @features: out parameter will contain uffdio_api.features provided by kernel 221 * in case of success 222 */ 223 static bool receive_ufd_features(uint64_t *features) 224 { 225 struct uffdio_api api_struct = {0}; 226 int ufd; 227 bool ret = true; 228 229 /* if we are here __NR_userfaultfd should exists */ 230 ufd = syscall(__NR_userfaultfd, O_CLOEXEC); 231 if (ufd == -1) { 232 error_report("%s: syscall __NR_userfaultfd failed: %s", __func__, 233 strerror(errno)); 234 return false; 235 } 236 237 /* ask features */ 238 api_struct.api = UFFD_API; 239 api_struct.features = 0; 240 if (ioctl(ufd, UFFDIO_API, &api_struct)) { 241 error_report("%s: UFFDIO_API failed: %s", __func__, 242 strerror(errno)); 243 ret = false; 244 goto release_ufd; 245 } 246 247 *features = api_struct.features; 248 249 release_ufd: 250 close(ufd); 251 return ret; 252 } 253 254 /** 255 * request_ufd_features: this function should be called only once on a newly 256 * opened ufd, subsequent calls will lead to error. 257 * 258 * Returns: true on success 259 * 260 * @ufd: fd obtained from userfaultfd syscall 261 * @features: bit mask see UFFD_API_FEATURES 262 */ 263 static bool request_ufd_features(int ufd, uint64_t features) 264 { 265 struct uffdio_api api_struct = {0}; 266 uint64_t ioctl_mask; 267 268 api_struct.api = UFFD_API; 269 api_struct.features = features; 270 if (ioctl(ufd, UFFDIO_API, &api_struct)) { 271 error_report("%s failed: UFFDIO_API failed: %s", __func__, 272 strerror(errno)); 273 return false; 274 } 275 276 ioctl_mask = (__u64)1 << _UFFDIO_REGISTER | 277 (__u64)1 << _UFFDIO_UNREGISTER; 278 if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) { 279 error_report("Missing userfault features: %" PRIx64, 280 (uint64_t)(~api_struct.ioctls & ioctl_mask)); 281 return false; 282 } 283 284 return true; 285 } 286 287 static bool ufd_check_and_apply(int ufd, MigrationIncomingState *mis) 288 { 289 uint64_t asked_features = 0; 290 static uint64_t supported_features; 291 292 /* 293 * it's not possible to 294 * request UFFD_API twice per one fd 295 * userfault fd features is persistent 296 */ 297 if (!supported_features) { 298 if (!receive_ufd_features(&supported_features)) { 299 error_report("%s failed", __func__); 300 return false; 301 } 302 } 303 304 #ifdef UFFD_FEATURE_THREAD_ID 305 if (UFFD_FEATURE_THREAD_ID & supported_features) { 306 asked_features |= UFFD_FEATURE_THREAD_ID; 307 if (migrate_postcopy_blocktime()) { 308 if (!mis->blocktime_ctx) { 309 mis->blocktime_ctx = blocktime_context_new(); 310 } 311 } 312 } 313 #endif 314 315 /* 316 * request features, even if asked_features is 0, due to 317 * kernel expects UFFD_API before UFFDIO_REGISTER, per 318 * userfault file descriptor 319 */ 320 if (!request_ufd_features(ufd, asked_features)) { 321 error_report("%s failed: features %" PRIu64, __func__, 322 asked_features); 323 return false; 324 } 325 326 if (qemu_real_host_page_size() != ram_pagesize_summary()) { 327 bool have_hp = false; 328 /* We've got a huge page */ 329 #ifdef UFFD_FEATURE_MISSING_HUGETLBFS 330 have_hp = supported_features & UFFD_FEATURE_MISSING_HUGETLBFS; 331 #endif 332 if (!have_hp) { 333 error_report("Userfault on this host does not support huge pages"); 334 return false; 335 } 336 } 337 return true; 338 } 339 340 /* Callback from postcopy_ram_supported_by_host block iterator. 341 */ 342 static int test_ramblock_postcopiable(RAMBlock *rb, void *opaque) 343 { 344 const char *block_name = qemu_ram_get_idstr(rb); 345 ram_addr_t length = qemu_ram_get_used_length(rb); 346 size_t pagesize = qemu_ram_pagesize(rb); 347 348 if (length % pagesize) { 349 error_report("Postcopy requires RAM blocks to be a page size multiple," 350 " block %s is 0x" RAM_ADDR_FMT " bytes with a " 351 "page size of 0x%zx", block_name, length, pagesize); 352 return 1; 353 } 354 return 0; 355 } 356 357 /* 358 * Note: This has the side effect of munlock'ing all of RAM, that's 359 * normally fine since if the postcopy succeeds it gets turned back on at the 360 * end. 361 */ 362 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis) 363 { 364 long pagesize = qemu_real_host_page_size(); 365 int ufd = -1; 366 bool ret = false; /* Error unless we change it */ 367 void *testarea = NULL; 368 struct uffdio_register reg_struct; 369 struct uffdio_range range_struct; 370 uint64_t feature_mask; 371 Error *local_err = NULL; 372 373 if (qemu_target_page_size() > pagesize) { 374 error_report("Target page size bigger than host page size"); 375 goto out; 376 } 377 378 ufd = syscall(__NR_userfaultfd, O_CLOEXEC); 379 if (ufd == -1) { 380 error_report("%s: userfaultfd not available: %s", __func__, 381 strerror(errno)); 382 goto out; 383 } 384 385 /* Give devices a chance to object */ 386 if (postcopy_notify(POSTCOPY_NOTIFY_PROBE, &local_err)) { 387 error_report_err(local_err); 388 goto out; 389 } 390 391 /* Version and features check */ 392 if (!ufd_check_and_apply(ufd, mis)) { 393 goto out; 394 } 395 396 /* We don't support postcopy with shared RAM yet */ 397 if (foreach_not_ignored_block(test_ramblock_postcopiable, NULL)) { 398 goto out; 399 } 400 401 /* 402 * userfault and mlock don't go together; we'll put it back later if 403 * it was enabled. 404 */ 405 if (munlockall()) { 406 error_report("%s: munlockall: %s", __func__, strerror(errno)); 407 goto out; 408 } 409 410 /* 411 * We need to check that the ops we need are supported on anon memory 412 * To do that we need to register a chunk and see the flags that 413 * are returned. 414 */ 415 testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE | 416 MAP_ANONYMOUS, -1, 0); 417 if (testarea == MAP_FAILED) { 418 error_report("%s: Failed to map test area: %s", __func__, 419 strerror(errno)); 420 goto out; 421 } 422 g_assert(QEMU_PTR_IS_ALIGNED(testarea, pagesize)); 423 424 reg_struct.range.start = (uintptr_t)testarea; 425 reg_struct.range.len = pagesize; 426 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; 427 428 if (ioctl(ufd, UFFDIO_REGISTER, ®_struct)) { 429 error_report("%s userfault register: %s", __func__, strerror(errno)); 430 goto out; 431 } 432 433 range_struct.start = (uintptr_t)testarea; 434 range_struct.len = pagesize; 435 if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) { 436 error_report("%s userfault unregister: %s", __func__, strerror(errno)); 437 goto out; 438 } 439 440 feature_mask = (__u64)1 << _UFFDIO_WAKE | 441 (__u64)1 << _UFFDIO_COPY | 442 (__u64)1 << _UFFDIO_ZEROPAGE; 443 if ((reg_struct.ioctls & feature_mask) != feature_mask) { 444 error_report("Missing userfault map features: %" PRIx64, 445 (uint64_t)(~reg_struct.ioctls & feature_mask)); 446 goto out; 447 } 448 449 /* Success! */ 450 ret = true; 451 out: 452 if (testarea) { 453 munmap(testarea, pagesize); 454 } 455 if (ufd != -1) { 456 close(ufd); 457 } 458 return ret; 459 } 460 461 /* 462 * Setup an area of RAM so that it *can* be used for postcopy later; this 463 * must be done right at the start prior to pre-copy. 464 * opaque should be the MIS. 465 */ 466 static int init_range(RAMBlock *rb, void *opaque) 467 { 468 const char *block_name = qemu_ram_get_idstr(rb); 469 void *host_addr = qemu_ram_get_host_addr(rb); 470 ram_addr_t offset = qemu_ram_get_offset(rb); 471 ram_addr_t length = qemu_ram_get_used_length(rb); 472 trace_postcopy_init_range(block_name, host_addr, offset, length); 473 474 /* 475 * Save the used_length before running the guest. In case we have to 476 * resize RAM blocks when syncing RAM block sizes from the source during 477 * precopy, we'll update it manually via the ram block notifier. 478 */ 479 rb->postcopy_length = length; 480 481 /* 482 * We need the whole of RAM to be truly empty for postcopy, so things 483 * like ROMs and any data tables built during init must be zero'd 484 * - we're going to get the copy from the source anyway. 485 * (Precopy will just overwrite this data, so doesn't need the discard) 486 */ 487 if (ram_discard_range(block_name, 0, length)) { 488 return -1; 489 } 490 491 return 0; 492 } 493 494 /* 495 * At the end of migration, undo the effects of init_range 496 * opaque should be the MIS. 497 */ 498 static int cleanup_range(RAMBlock *rb, void *opaque) 499 { 500 const char *block_name = qemu_ram_get_idstr(rb); 501 void *host_addr = qemu_ram_get_host_addr(rb); 502 ram_addr_t offset = qemu_ram_get_offset(rb); 503 ram_addr_t length = rb->postcopy_length; 504 MigrationIncomingState *mis = opaque; 505 struct uffdio_range range_struct; 506 trace_postcopy_cleanup_range(block_name, host_addr, offset, length); 507 508 /* 509 * We turned off hugepage for the precopy stage with postcopy enabled 510 * we can turn it back on now. 511 */ 512 qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE); 513 514 /* 515 * We can also turn off userfault now since we should have all the 516 * pages. It can be useful to leave it on to debug postcopy 517 * if you're not sure it's always getting every page. 518 */ 519 range_struct.start = (uintptr_t)host_addr; 520 range_struct.len = length; 521 522 if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) { 523 error_report("%s: userfault unregister %s", __func__, strerror(errno)); 524 525 return -1; 526 } 527 528 return 0; 529 } 530 531 /* 532 * Initialise postcopy-ram, setting the RAM to a state where we can go into 533 * postcopy later; must be called prior to any precopy. 534 * called from arch_init's similarly named ram_postcopy_incoming_init 535 */ 536 int postcopy_ram_incoming_init(MigrationIncomingState *mis) 537 { 538 if (foreach_not_ignored_block(init_range, NULL)) { 539 return -1; 540 } 541 542 return 0; 543 } 544 545 static void postcopy_temp_pages_cleanup(MigrationIncomingState *mis) 546 { 547 int i; 548 549 if (mis->postcopy_tmp_pages) { 550 for (i = 0; i < mis->postcopy_channels; i++) { 551 if (mis->postcopy_tmp_pages[i].tmp_huge_page) { 552 munmap(mis->postcopy_tmp_pages[i].tmp_huge_page, 553 mis->largest_page_size); 554 mis->postcopy_tmp_pages[i].tmp_huge_page = NULL; 555 } 556 } 557 g_free(mis->postcopy_tmp_pages); 558 mis->postcopy_tmp_pages = NULL; 559 } 560 561 if (mis->postcopy_tmp_zero_page) { 562 munmap(mis->postcopy_tmp_zero_page, mis->largest_page_size); 563 mis->postcopy_tmp_zero_page = NULL; 564 } 565 } 566 567 /* 568 * At the end of a migration where postcopy_ram_incoming_init was called. 569 */ 570 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) 571 { 572 trace_postcopy_ram_incoming_cleanup_entry(); 573 574 if (mis->postcopy_prio_thread_created) { 575 qemu_thread_join(&mis->postcopy_prio_thread); 576 mis->postcopy_prio_thread_created = false; 577 } 578 579 if (mis->have_fault_thread) { 580 Error *local_err = NULL; 581 582 /* Let the fault thread quit */ 583 qatomic_set(&mis->fault_thread_quit, 1); 584 postcopy_fault_thread_notify(mis); 585 trace_postcopy_ram_incoming_cleanup_join(); 586 qemu_thread_join(&mis->fault_thread); 587 588 if (postcopy_notify(POSTCOPY_NOTIFY_INBOUND_END, &local_err)) { 589 error_report_err(local_err); 590 return -1; 591 } 592 593 if (foreach_not_ignored_block(cleanup_range, mis)) { 594 return -1; 595 } 596 597 trace_postcopy_ram_incoming_cleanup_closeuf(); 598 close(mis->userfault_fd); 599 close(mis->userfault_event_fd); 600 mis->have_fault_thread = false; 601 } 602 603 if (enable_mlock) { 604 if (os_mlock() < 0) { 605 error_report("mlock: %s", strerror(errno)); 606 /* 607 * It doesn't feel right to fail at this point, we have a valid 608 * VM state. 609 */ 610 } 611 } 612 613 postcopy_temp_pages_cleanup(mis); 614 615 trace_postcopy_ram_incoming_cleanup_blocktime( 616 get_postcopy_total_blocktime()); 617 618 trace_postcopy_ram_incoming_cleanup_exit(); 619 return 0; 620 } 621 622 /* 623 * Disable huge pages on an area 624 */ 625 static int nhp_range(RAMBlock *rb, void *opaque) 626 { 627 const char *block_name = qemu_ram_get_idstr(rb); 628 void *host_addr = qemu_ram_get_host_addr(rb); 629 ram_addr_t offset = qemu_ram_get_offset(rb); 630 ram_addr_t length = rb->postcopy_length; 631 trace_postcopy_nhp_range(block_name, host_addr, offset, length); 632 633 /* 634 * Before we do discards we need to ensure those discards really 635 * do delete areas of the page, even if THP thinks a hugepage would 636 * be a good idea, so force hugepages off. 637 */ 638 qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE); 639 640 return 0; 641 } 642 643 /* 644 * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard 645 * however leaving it until after precopy means that most of the precopy 646 * data is still THPd 647 */ 648 int postcopy_ram_prepare_discard(MigrationIncomingState *mis) 649 { 650 if (foreach_not_ignored_block(nhp_range, mis)) { 651 return -1; 652 } 653 654 postcopy_state_set(POSTCOPY_INCOMING_DISCARD); 655 656 return 0; 657 } 658 659 /* 660 * Mark the given area of RAM as requiring notification to unwritten areas 661 * Used as a callback on foreach_not_ignored_block. 662 * host_addr: Base of area to mark 663 * offset: Offset in the whole ram arena 664 * length: Length of the section 665 * opaque: MigrationIncomingState pointer 666 * Returns 0 on success 667 */ 668 static int ram_block_enable_notify(RAMBlock *rb, void *opaque) 669 { 670 MigrationIncomingState *mis = opaque; 671 struct uffdio_register reg_struct; 672 673 reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb); 674 reg_struct.range.len = rb->postcopy_length; 675 reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING; 676 677 /* Now tell our userfault_fd that it's responsible for this area */ 678 if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, ®_struct)) { 679 error_report("%s userfault register: %s", __func__, strerror(errno)); 680 return -1; 681 } 682 if (!(reg_struct.ioctls & ((__u64)1 << _UFFDIO_COPY))) { 683 error_report("%s userfault: Region doesn't support COPY", __func__); 684 return -1; 685 } 686 if (reg_struct.ioctls & ((__u64)1 << _UFFDIO_ZEROPAGE)) { 687 qemu_ram_set_uf_zeroable(rb); 688 } 689 690 return 0; 691 } 692 693 int postcopy_wake_shared(struct PostCopyFD *pcfd, 694 uint64_t client_addr, 695 RAMBlock *rb) 696 { 697 size_t pagesize = qemu_ram_pagesize(rb); 698 struct uffdio_range range; 699 int ret; 700 trace_postcopy_wake_shared(client_addr, qemu_ram_get_idstr(rb)); 701 range.start = ROUND_DOWN(client_addr, pagesize); 702 range.len = pagesize; 703 ret = ioctl(pcfd->fd, UFFDIO_WAKE, &range); 704 if (ret) { 705 error_report("%s: Failed to wake: %zx in %s (%s)", 706 __func__, (size_t)client_addr, qemu_ram_get_idstr(rb), 707 strerror(errno)); 708 } 709 return ret; 710 } 711 712 static int postcopy_request_page(MigrationIncomingState *mis, RAMBlock *rb, 713 ram_addr_t start, uint64_t haddr) 714 { 715 void *aligned = (void *)(uintptr_t)ROUND_DOWN(haddr, qemu_ram_pagesize(rb)); 716 717 /* 718 * Discarded pages (via RamDiscardManager) are never migrated. On unlikely 719 * access, place a zeropage, which will also set the relevant bits in the 720 * recv_bitmap accordingly, so we won't try placing a zeropage twice. 721 * 722 * Checking a single bit is sufficient to handle pagesize > TPS as either 723 * all relevant bits are set or not. 724 */ 725 assert(QEMU_IS_ALIGNED(start, qemu_ram_pagesize(rb))); 726 if (ramblock_page_is_discarded(rb, start)) { 727 bool received = ramblock_recv_bitmap_test_byte_offset(rb, start); 728 729 return received ? 0 : postcopy_place_page_zero(mis, aligned, rb); 730 } 731 732 return migrate_send_rp_req_pages(mis, rb, start, haddr); 733 } 734 735 /* 736 * Callback from shared fault handlers to ask for a page, 737 * the page must be specified by a RAMBlock and an offset in that rb 738 * Note: Only for use by shared fault handlers (in fault thread) 739 */ 740 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, 741 uint64_t client_addr, uint64_t rb_offset) 742 { 743 uint64_t aligned_rbo = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb)); 744 MigrationIncomingState *mis = migration_incoming_get_current(); 745 746 trace_postcopy_request_shared_page(pcfd->idstr, qemu_ram_get_idstr(rb), 747 rb_offset); 748 if (ramblock_recv_bitmap_test_byte_offset(rb, aligned_rbo)) { 749 trace_postcopy_request_shared_page_present(pcfd->idstr, 750 qemu_ram_get_idstr(rb), rb_offset); 751 return postcopy_wake_shared(pcfd, client_addr, rb); 752 } 753 postcopy_request_page(mis, rb, aligned_rbo, client_addr); 754 return 0; 755 } 756 757 static int get_mem_fault_cpu_index(uint32_t pid) 758 { 759 CPUState *cpu_iter; 760 761 CPU_FOREACH(cpu_iter) { 762 if (cpu_iter->thread_id == pid) { 763 trace_get_mem_fault_cpu_index(cpu_iter->cpu_index, pid); 764 return cpu_iter->cpu_index; 765 } 766 } 767 trace_get_mem_fault_cpu_index(-1, pid); 768 return -1; 769 } 770 771 static uint32_t get_low_time_offset(PostcopyBlocktimeContext *dc) 772 { 773 int64_t start_time_offset = qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - 774 dc->start_time; 775 return start_time_offset < 1 ? 1 : start_time_offset & UINT32_MAX; 776 } 777 778 /* 779 * This function is being called when pagefault occurs. It 780 * tracks down vCPU blocking time. 781 * 782 * @addr: faulted host virtual address 783 * @ptid: faulted process thread id 784 * @rb: ramblock appropriate to addr 785 */ 786 static void mark_postcopy_blocktime_begin(uintptr_t addr, uint32_t ptid, 787 RAMBlock *rb) 788 { 789 int cpu, already_received; 790 MigrationIncomingState *mis = migration_incoming_get_current(); 791 PostcopyBlocktimeContext *dc = mis->blocktime_ctx; 792 uint32_t low_time_offset; 793 794 if (!dc || ptid == 0) { 795 return; 796 } 797 cpu = get_mem_fault_cpu_index(ptid); 798 if (cpu < 0) { 799 return; 800 } 801 802 low_time_offset = get_low_time_offset(dc); 803 if (dc->vcpu_addr[cpu] == 0) { 804 qatomic_inc(&dc->smp_cpus_down); 805 } 806 807 qatomic_xchg(&dc->last_begin, low_time_offset); 808 qatomic_xchg(&dc->page_fault_vcpu_time[cpu], low_time_offset); 809 qatomic_xchg(&dc->vcpu_addr[cpu], addr); 810 811 /* 812 * check it here, not at the beginning of the function, 813 * due to, check could occur early than bitmap_set in 814 * qemu_ufd_copy_ioctl 815 */ 816 already_received = ramblock_recv_bitmap_test(rb, (void *)addr); 817 if (already_received) { 818 qatomic_xchg(&dc->vcpu_addr[cpu], 0); 819 qatomic_xchg(&dc->page_fault_vcpu_time[cpu], 0); 820 qatomic_dec(&dc->smp_cpus_down); 821 } 822 trace_mark_postcopy_blocktime_begin(addr, dc, dc->page_fault_vcpu_time[cpu], 823 cpu, already_received); 824 } 825 826 /* 827 * This function just provide calculated blocktime per cpu and trace it. 828 * Total blocktime is calculated in mark_postcopy_blocktime_end. 829 * 830 * 831 * Assume we have 3 CPU 832 * 833 * S1 E1 S1 E1 834 * -----***********------------xxx***************------------------------> CPU1 835 * 836 * S2 E2 837 * ------------****************xxx---------------------------------------> CPU2 838 * 839 * S3 E3 840 * ------------------------****xxx********-------------------------------> CPU3 841 * 842 * We have sequence S1,S2,E1,S3,S1,E2,E3,E1 843 * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include CPU3 844 * S3,S1,E2 - sequence includes all CPUs, in this case overlap will be S1,E2 - 845 * it's a part of total blocktime. 846 * S1 - here is last_begin 847 * Legend of the picture is following: 848 * * - means blocktime per vCPU 849 * x - means overlapped blocktime (total blocktime) 850 * 851 * @addr: host virtual address 852 */ 853 static void mark_postcopy_blocktime_end(uintptr_t addr) 854 { 855 MigrationIncomingState *mis = migration_incoming_get_current(); 856 PostcopyBlocktimeContext *dc = mis->blocktime_ctx; 857 MachineState *ms = MACHINE(qdev_get_machine()); 858 unsigned int smp_cpus = ms->smp.cpus; 859 int i, affected_cpu = 0; 860 bool vcpu_total_blocktime = false; 861 uint32_t read_vcpu_time, low_time_offset; 862 863 if (!dc) { 864 return; 865 } 866 867 low_time_offset = get_low_time_offset(dc); 868 /* lookup cpu, to clear it, 869 * that algorithm looks straightforward, but it's not 870 * optimal, more optimal algorithm is keeping tree or hash 871 * where key is address value is a list of */ 872 for (i = 0; i < smp_cpus; i++) { 873 uint32_t vcpu_blocktime = 0; 874 875 read_vcpu_time = qatomic_fetch_add(&dc->page_fault_vcpu_time[i], 0); 876 if (qatomic_fetch_add(&dc->vcpu_addr[i], 0) != addr || 877 read_vcpu_time == 0) { 878 continue; 879 } 880 qatomic_xchg(&dc->vcpu_addr[i], 0); 881 vcpu_blocktime = low_time_offset - read_vcpu_time; 882 affected_cpu += 1; 883 /* we need to know is that mark_postcopy_end was due to 884 * faulted page, another possible case it's prefetched 885 * page and in that case we shouldn't be here */ 886 if (!vcpu_total_blocktime && 887 qatomic_fetch_add(&dc->smp_cpus_down, 0) == smp_cpus) { 888 vcpu_total_blocktime = true; 889 } 890 /* continue cycle, due to one page could affect several vCPUs */ 891 dc->vcpu_blocktime[i] += vcpu_blocktime; 892 } 893 894 qatomic_sub(&dc->smp_cpus_down, affected_cpu); 895 if (vcpu_total_blocktime) { 896 dc->total_blocktime += low_time_offset - qatomic_fetch_add( 897 &dc->last_begin, 0); 898 } 899 trace_mark_postcopy_blocktime_end(addr, dc, dc->total_blocktime, 900 affected_cpu); 901 } 902 903 static void postcopy_pause_fault_thread(MigrationIncomingState *mis) 904 { 905 trace_postcopy_pause_fault_thread(); 906 qemu_sem_wait(&mis->postcopy_pause_sem_fault); 907 trace_postcopy_pause_fault_thread_continued(); 908 } 909 910 /* 911 * Handle faults detected by the USERFAULT markings 912 */ 913 static void *postcopy_ram_fault_thread(void *opaque) 914 { 915 MigrationIncomingState *mis = opaque; 916 struct uffd_msg msg; 917 int ret; 918 size_t index; 919 RAMBlock *rb = NULL; 920 921 trace_postcopy_ram_fault_thread_entry(); 922 rcu_register_thread(); 923 mis->last_rb = NULL; /* last RAMBlock we sent part of */ 924 qemu_sem_post(&mis->thread_sync_sem); 925 926 struct pollfd *pfd; 927 size_t pfd_len = 2 + mis->postcopy_remote_fds->len; 928 929 pfd = g_new0(struct pollfd, pfd_len); 930 931 pfd[0].fd = mis->userfault_fd; 932 pfd[0].events = POLLIN; 933 pfd[1].fd = mis->userfault_event_fd; 934 pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */ 935 trace_postcopy_ram_fault_thread_fds_core(pfd[0].fd, pfd[1].fd); 936 for (index = 0; index < mis->postcopy_remote_fds->len; index++) { 937 struct PostCopyFD *pcfd = &g_array_index(mis->postcopy_remote_fds, 938 struct PostCopyFD, index); 939 pfd[2 + index].fd = pcfd->fd; 940 pfd[2 + index].events = POLLIN; 941 trace_postcopy_ram_fault_thread_fds_extra(2 + index, pcfd->idstr, 942 pcfd->fd); 943 } 944 945 while (true) { 946 ram_addr_t rb_offset; 947 int poll_result; 948 949 /* 950 * We're mainly waiting for the kernel to give us a faulting HVA, 951 * however we can be told to quit via userfault_quit_fd which is 952 * an eventfd 953 */ 954 955 poll_result = poll(pfd, pfd_len, -1 /* Wait forever */); 956 if (poll_result == -1) { 957 error_report("%s: userfault poll: %s", __func__, strerror(errno)); 958 break; 959 } 960 961 if (!mis->to_src_file) { 962 /* 963 * Possibly someone tells us that the return path is 964 * broken already using the event. We should hold until 965 * the channel is rebuilt. 966 */ 967 postcopy_pause_fault_thread(mis); 968 } 969 970 if (pfd[1].revents) { 971 uint64_t tmp64 = 0; 972 973 /* Consume the signal */ 974 if (read(mis->userfault_event_fd, &tmp64, 8) != 8) { 975 /* Nothing obviously nicer than posting this error. */ 976 error_report("%s: read() failed", __func__); 977 } 978 979 if (qatomic_read(&mis->fault_thread_quit)) { 980 trace_postcopy_ram_fault_thread_quit(); 981 break; 982 } 983 } 984 985 if (pfd[0].revents) { 986 poll_result--; 987 ret = read(mis->userfault_fd, &msg, sizeof(msg)); 988 if (ret != sizeof(msg)) { 989 if (errno == EAGAIN) { 990 /* 991 * if a wake up happens on the other thread just after 992 * the poll, there is nothing to read. 993 */ 994 continue; 995 } 996 if (ret < 0) { 997 error_report("%s: Failed to read full userfault " 998 "message: %s", 999 __func__, strerror(errno)); 1000 break; 1001 } else { 1002 error_report("%s: Read %d bytes from userfaultfd " 1003 "expected %zd", 1004 __func__, ret, sizeof(msg)); 1005 break; /* Lost alignment, don't know what we'd read next */ 1006 } 1007 } 1008 if (msg.event != UFFD_EVENT_PAGEFAULT) { 1009 error_report("%s: Read unexpected event %ud from userfaultfd", 1010 __func__, msg.event); 1011 continue; /* It's not a page fault, shouldn't happen */ 1012 } 1013 1014 rb = qemu_ram_block_from_host( 1015 (void *)(uintptr_t)msg.arg.pagefault.address, 1016 true, &rb_offset); 1017 if (!rb) { 1018 error_report("postcopy_ram_fault_thread: Fault outside guest: %" 1019 PRIx64, (uint64_t)msg.arg.pagefault.address); 1020 break; 1021 } 1022 1023 rb_offset = ROUND_DOWN(rb_offset, qemu_ram_pagesize(rb)); 1024 trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address, 1025 qemu_ram_get_idstr(rb), 1026 rb_offset, 1027 msg.arg.pagefault.feat.ptid); 1028 mark_postcopy_blocktime_begin( 1029 (uintptr_t)(msg.arg.pagefault.address), 1030 msg.arg.pagefault.feat.ptid, rb); 1031 1032 retry: 1033 /* 1034 * Send the request to the source - we want to request one 1035 * of our host page sizes (which is >= TPS) 1036 */ 1037 ret = postcopy_request_page(mis, rb, rb_offset, 1038 msg.arg.pagefault.address); 1039 if (ret) { 1040 /* May be network failure, try to wait for recovery */ 1041 postcopy_pause_fault_thread(mis); 1042 goto retry; 1043 } 1044 } 1045 1046 /* Now handle any requests from external processes on shared memory */ 1047 /* TODO: May need to handle devices deregistering during postcopy */ 1048 for (index = 2; index < pfd_len && poll_result; index++) { 1049 if (pfd[index].revents) { 1050 struct PostCopyFD *pcfd = 1051 &g_array_index(mis->postcopy_remote_fds, 1052 struct PostCopyFD, index - 2); 1053 1054 poll_result--; 1055 if (pfd[index].revents & POLLERR) { 1056 error_report("%s: POLLERR on poll %zd fd=%d", 1057 __func__, index, pcfd->fd); 1058 pfd[index].events = 0; 1059 continue; 1060 } 1061 1062 ret = read(pcfd->fd, &msg, sizeof(msg)); 1063 if (ret != sizeof(msg)) { 1064 if (errno == EAGAIN) { 1065 /* 1066 * if a wake up happens on the other thread just after 1067 * the poll, there is nothing to read. 1068 */ 1069 continue; 1070 } 1071 if (ret < 0) { 1072 error_report("%s: Failed to read full userfault " 1073 "message: %s (shared) revents=%d", 1074 __func__, strerror(errno), 1075 pfd[index].revents); 1076 /*TODO: Could just disable this sharer */ 1077 break; 1078 } else { 1079 error_report("%s: Read %d bytes from userfaultfd " 1080 "expected %zd (shared)", 1081 __func__, ret, sizeof(msg)); 1082 /*TODO: Could just disable this sharer */ 1083 break; /*Lost alignment,don't know what we'd read next*/ 1084 } 1085 } 1086 if (msg.event != UFFD_EVENT_PAGEFAULT) { 1087 error_report("%s: Read unexpected event %ud " 1088 "from userfaultfd (shared)", 1089 __func__, msg.event); 1090 continue; /* It's not a page fault, shouldn't happen */ 1091 } 1092 /* Call the device handler registered with us */ 1093 ret = pcfd->handler(pcfd, &msg); 1094 if (ret) { 1095 error_report("%s: Failed to resolve shared fault on %zd/%s", 1096 __func__, index, pcfd->idstr); 1097 /* TODO: Fail? Disable this sharer? */ 1098 } 1099 } 1100 } 1101 } 1102 rcu_unregister_thread(); 1103 trace_postcopy_ram_fault_thread_exit(); 1104 g_free(pfd); 1105 return NULL; 1106 } 1107 1108 static int postcopy_temp_pages_setup(MigrationIncomingState *mis) 1109 { 1110 PostcopyTmpPage *tmp_page; 1111 int err, i, channels; 1112 void *temp_page; 1113 1114 if (migrate_postcopy_preempt()) { 1115 /* If preemption enabled, need extra channel for urgent requests */ 1116 mis->postcopy_channels = RAM_CHANNEL_MAX; 1117 } else { 1118 /* Both precopy/postcopy on the same channel */ 1119 mis->postcopy_channels = 1; 1120 } 1121 1122 channels = mis->postcopy_channels; 1123 mis->postcopy_tmp_pages = g_malloc0_n(sizeof(PostcopyTmpPage), channels); 1124 1125 for (i = 0; i < channels; i++) { 1126 tmp_page = &mis->postcopy_tmp_pages[i]; 1127 temp_page = mmap(NULL, mis->largest_page_size, PROT_READ | PROT_WRITE, 1128 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); 1129 if (temp_page == MAP_FAILED) { 1130 err = errno; 1131 error_report("%s: Failed to map postcopy_tmp_pages[%d]: %s", 1132 __func__, i, strerror(err)); 1133 /* Clean up will be done later */ 1134 return -err; 1135 } 1136 tmp_page->tmp_huge_page = temp_page; 1137 /* Initialize default states for each tmp page */ 1138 postcopy_temp_page_reset(tmp_page); 1139 } 1140 1141 /* 1142 * Map large zero page when kernel can't use UFFDIO_ZEROPAGE for hugepages 1143 */ 1144 mis->postcopy_tmp_zero_page = mmap(NULL, mis->largest_page_size, 1145 PROT_READ | PROT_WRITE, 1146 MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); 1147 if (mis->postcopy_tmp_zero_page == MAP_FAILED) { 1148 err = errno; 1149 mis->postcopy_tmp_zero_page = NULL; 1150 error_report("%s: Failed to map large zero page %s", 1151 __func__, strerror(err)); 1152 return -err; 1153 } 1154 1155 memset(mis->postcopy_tmp_zero_page, '\0', mis->largest_page_size); 1156 1157 return 0; 1158 } 1159 1160 int postcopy_ram_incoming_setup(MigrationIncomingState *mis) 1161 { 1162 /* Open the fd for the kernel to give us userfaults */ 1163 mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK); 1164 if (mis->userfault_fd == -1) { 1165 error_report("%s: Failed to open userfault fd: %s", __func__, 1166 strerror(errno)); 1167 return -1; 1168 } 1169 1170 /* 1171 * Although the host check already tested the API, we need to 1172 * do the check again as an ABI handshake on the new fd. 1173 */ 1174 if (!ufd_check_and_apply(mis->userfault_fd, mis)) { 1175 return -1; 1176 } 1177 1178 /* Now an eventfd we use to tell the fault-thread to quit */ 1179 mis->userfault_event_fd = eventfd(0, EFD_CLOEXEC); 1180 if (mis->userfault_event_fd == -1) { 1181 error_report("%s: Opening userfault_event_fd: %s", __func__, 1182 strerror(errno)); 1183 close(mis->userfault_fd); 1184 return -1; 1185 } 1186 1187 postcopy_thread_create(mis, &mis->fault_thread, "fault-default", 1188 postcopy_ram_fault_thread, QEMU_THREAD_JOINABLE); 1189 mis->have_fault_thread = true; 1190 1191 /* Mark so that we get notified of accesses to unwritten areas */ 1192 if (foreach_not_ignored_block(ram_block_enable_notify, mis)) { 1193 error_report("ram_block_enable_notify failed"); 1194 return -1; 1195 } 1196 1197 if (postcopy_temp_pages_setup(mis)) { 1198 /* Error dumped in the sub-function */ 1199 return -1; 1200 } 1201 1202 if (migrate_postcopy_preempt()) { 1203 /* 1204 * This thread needs to be created after the temp pages because 1205 * it'll fetch RAM_CHANNEL_POSTCOPY PostcopyTmpPage immediately. 1206 */ 1207 postcopy_thread_create(mis, &mis->postcopy_prio_thread, "fault-fast", 1208 postcopy_preempt_thread, QEMU_THREAD_JOINABLE); 1209 mis->postcopy_prio_thread_created = true; 1210 } 1211 1212 trace_postcopy_ram_enable_notify(); 1213 1214 return 0; 1215 } 1216 1217 static int qemu_ufd_copy_ioctl(MigrationIncomingState *mis, void *host_addr, 1218 void *from_addr, uint64_t pagesize, RAMBlock *rb) 1219 { 1220 int userfault_fd = mis->userfault_fd; 1221 int ret; 1222 1223 if (from_addr) { 1224 struct uffdio_copy copy_struct; 1225 copy_struct.dst = (uint64_t)(uintptr_t)host_addr; 1226 copy_struct.src = (uint64_t)(uintptr_t)from_addr; 1227 copy_struct.len = pagesize; 1228 copy_struct.mode = 0; 1229 ret = ioctl(userfault_fd, UFFDIO_COPY, ©_struct); 1230 } else { 1231 struct uffdio_zeropage zero_struct; 1232 zero_struct.range.start = (uint64_t)(uintptr_t)host_addr; 1233 zero_struct.range.len = pagesize; 1234 zero_struct.mode = 0; 1235 ret = ioctl(userfault_fd, UFFDIO_ZEROPAGE, &zero_struct); 1236 } 1237 if (!ret) { 1238 qemu_mutex_lock(&mis->page_request_mutex); 1239 ramblock_recv_bitmap_set_range(rb, host_addr, 1240 pagesize / qemu_target_page_size()); 1241 /* 1242 * If this page resolves a page fault for a previous recorded faulted 1243 * address, take a special note to maintain the requested page list. 1244 */ 1245 if (g_tree_lookup(mis->page_requested, host_addr)) { 1246 g_tree_remove(mis->page_requested, host_addr); 1247 mis->page_requested_count--; 1248 trace_postcopy_page_req_del(host_addr, mis->page_requested_count); 1249 } 1250 qemu_mutex_unlock(&mis->page_request_mutex); 1251 mark_postcopy_blocktime_end((uintptr_t)host_addr); 1252 } 1253 return ret; 1254 } 1255 1256 int postcopy_notify_shared_wake(RAMBlock *rb, uint64_t offset) 1257 { 1258 int i; 1259 MigrationIncomingState *mis = migration_incoming_get_current(); 1260 GArray *pcrfds = mis->postcopy_remote_fds; 1261 1262 for (i = 0; i < pcrfds->len; i++) { 1263 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); 1264 int ret = cur->waker(cur, rb, offset); 1265 if (ret) { 1266 return ret; 1267 } 1268 } 1269 return 0; 1270 } 1271 1272 /* 1273 * Place a host page (from) at (host) atomically 1274 * returns 0 on success 1275 */ 1276 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, 1277 RAMBlock *rb) 1278 { 1279 size_t pagesize = qemu_ram_pagesize(rb); 1280 1281 /* copy also acks to the kernel waking the stalled thread up 1282 * TODO: We can inhibit that ack and only do it if it was requested 1283 * which would be slightly cheaper, but we'd have to be careful 1284 * of the order of updating our page state. 1285 */ 1286 if (qemu_ufd_copy_ioctl(mis, host, from, pagesize, rb)) { 1287 int e = errno; 1288 error_report("%s: %s copy host: %p from: %p (size: %zd)", 1289 __func__, strerror(e), host, from, pagesize); 1290 1291 return -e; 1292 } 1293 1294 trace_postcopy_place_page(host); 1295 return postcopy_notify_shared_wake(rb, 1296 qemu_ram_block_host_offset(rb, host)); 1297 } 1298 1299 /* 1300 * Place a zero page at (host) atomically 1301 * returns 0 on success 1302 */ 1303 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, 1304 RAMBlock *rb) 1305 { 1306 size_t pagesize = qemu_ram_pagesize(rb); 1307 trace_postcopy_place_page_zero(host); 1308 1309 /* Normal RAMBlocks can zero a page using UFFDIO_ZEROPAGE 1310 * but it's not available for everything (e.g. hugetlbpages) 1311 */ 1312 if (qemu_ram_is_uf_zeroable(rb)) { 1313 if (qemu_ufd_copy_ioctl(mis, host, NULL, pagesize, rb)) { 1314 int e = errno; 1315 error_report("%s: %s zero host: %p", 1316 __func__, strerror(e), host); 1317 1318 return -e; 1319 } 1320 return postcopy_notify_shared_wake(rb, 1321 qemu_ram_block_host_offset(rb, 1322 host)); 1323 } else { 1324 return postcopy_place_page(mis, host, mis->postcopy_tmp_zero_page, rb); 1325 } 1326 } 1327 1328 #else 1329 /* No target OS support, stubs just fail */ 1330 void fill_destination_postcopy_migration_info(MigrationInfo *info) 1331 { 1332 } 1333 1334 bool postcopy_ram_supported_by_host(MigrationIncomingState *mis) 1335 { 1336 error_report("%s: No OS support", __func__); 1337 return false; 1338 } 1339 1340 int postcopy_ram_incoming_init(MigrationIncomingState *mis) 1341 { 1342 error_report("postcopy_ram_incoming_init: No OS support"); 1343 return -1; 1344 } 1345 1346 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis) 1347 { 1348 assert(0); 1349 return -1; 1350 } 1351 1352 int postcopy_ram_prepare_discard(MigrationIncomingState *mis) 1353 { 1354 assert(0); 1355 return -1; 1356 } 1357 1358 int postcopy_request_shared_page(struct PostCopyFD *pcfd, RAMBlock *rb, 1359 uint64_t client_addr, uint64_t rb_offset) 1360 { 1361 assert(0); 1362 return -1; 1363 } 1364 1365 int postcopy_ram_incoming_setup(MigrationIncomingState *mis) 1366 { 1367 assert(0); 1368 return -1; 1369 } 1370 1371 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from, 1372 RAMBlock *rb) 1373 { 1374 assert(0); 1375 return -1; 1376 } 1377 1378 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host, 1379 RAMBlock *rb) 1380 { 1381 assert(0); 1382 return -1; 1383 } 1384 1385 int postcopy_wake_shared(struct PostCopyFD *pcfd, 1386 uint64_t client_addr, 1387 RAMBlock *rb) 1388 { 1389 assert(0); 1390 return -1; 1391 } 1392 #endif 1393 1394 /* ------------------------------------------------------------------------- */ 1395 void postcopy_temp_page_reset(PostcopyTmpPage *tmp_page) 1396 { 1397 tmp_page->target_pages = 0; 1398 tmp_page->host_addr = NULL; 1399 /* 1400 * This is set to true when reset, and cleared as long as we received any 1401 * of the non-zero small page within this huge page. 1402 */ 1403 tmp_page->all_zero = true; 1404 } 1405 1406 void postcopy_fault_thread_notify(MigrationIncomingState *mis) 1407 { 1408 uint64_t tmp64 = 1; 1409 1410 /* 1411 * Wakeup the fault_thread. It's an eventfd that should currently 1412 * be at 0, we're going to increment it to 1 1413 */ 1414 if (write(mis->userfault_event_fd, &tmp64, 8) != 8) { 1415 /* Not much we can do here, but may as well report it */ 1416 error_report("%s: incrementing failed: %s", __func__, 1417 strerror(errno)); 1418 } 1419 } 1420 1421 /** 1422 * postcopy_discard_send_init: Called at the start of each RAMBlock before 1423 * asking to discard individual ranges. 1424 * 1425 * @ms: The current migration state. 1426 * @offset: the bitmap offset of the named RAMBlock in the migration bitmap. 1427 * @name: RAMBlock that discards will operate on. 1428 */ 1429 static PostcopyDiscardState pds = {0}; 1430 void postcopy_discard_send_init(MigrationState *ms, const char *name) 1431 { 1432 pds.ramblock_name = name; 1433 pds.cur_entry = 0; 1434 pds.nsentwords = 0; 1435 pds.nsentcmds = 0; 1436 } 1437 1438 /** 1439 * postcopy_discard_send_range: Called by the bitmap code for each chunk to 1440 * discard. May send a discard message, may just leave it queued to 1441 * be sent later. 1442 * 1443 * @ms: Current migration state. 1444 * @start,@length: a range of pages in the migration bitmap in the 1445 * RAM block passed to postcopy_discard_send_init() (length=1 is one page) 1446 */ 1447 void postcopy_discard_send_range(MigrationState *ms, unsigned long start, 1448 unsigned long length) 1449 { 1450 size_t tp_size = qemu_target_page_size(); 1451 /* Convert to byte offsets within the RAM block */ 1452 pds.start_list[pds.cur_entry] = start * tp_size; 1453 pds.length_list[pds.cur_entry] = length * tp_size; 1454 trace_postcopy_discard_send_range(pds.ramblock_name, start, length); 1455 pds.cur_entry++; 1456 pds.nsentwords++; 1457 1458 if (pds.cur_entry == MAX_DISCARDS_PER_COMMAND) { 1459 /* Full set, ship it! */ 1460 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, 1461 pds.ramblock_name, 1462 pds.cur_entry, 1463 pds.start_list, 1464 pds.length_list); 1465 pds.nsentcmds++; 1466 pds.cur_entry = 0; 1467 } 1468 } 1469 1470 /** 1471 * postcopy_discard_send_finish: Called at the end of each RAMBlock by the 1472 * bitmap code. Sends any outstanding discard messages, frees the PDS 1473 * 1474 * @ms: Current migration state. 1475 */ 1476 void postcopy_discard_send_finish(MigrationState *ms) 1477 { 1478 /* Anything unsent? */ 1479 if (pds.cur_entry) { 1480 qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file, 1481 pds.ramblock_name, 1482 pds.cur_entry, 1483 pds.start_list, 1484 pds.length_list); 1485 pds.nsentcmds++; 1486 } 1487 1488 trace_postcopy_discard_send_finish(pds.ramblock_name, pds.nsentwords, 1489 pds.nsentcmds); 1490 } 1491 1492 /* 1493 * Current state of incoming postcopy; note this is not part of 1494 * MigrationIncomingState since it's state is used during cleanup 1495 * at the end as MIS is being freed. 1496 */ 1497 static PostcopyState incoming_postcopy_state; 1498 1499 PostcopyState postcopy_state_get(void) 1500 { 1501 return qatomic_mb_read(&incoming_postcopy_state); 1502 } 1503 1504 /* Set the state and return the old state */ 1505 PostcopyState postcopy_state_set(PostcopyState new_state) 1506 { 1507 return qatomic_xchg(&incoming_postcopy_state, new_state); 1508 } 1509 1510 /* Register a handler for external shared memory postcopy 1511 * called on the destination. 1512 */ 1513 void postcopy_register_shared_ufd(struct PostCopyFD *pcfd) 1514 { 1515 MigrationIncomingState *mis = migration_incoming_get_current(); 1516 1517 mis->postcopy_remote_fds = g_array_append_val(mis->postcopy_remote_fds, 1518 *pcfd); 1519 } 1520 1521 /* Unregister a handler for external shared memory postcopy 1522 */ 1523 void postcopy_unregister_shared_ufd(struct PostCopyFD *pcfd) 1524 { 1525 guint i; 1526 MigrationIncomingState *mis = migration_incoming_get_current(); 1527 GArray *pcrfds = mis->postcopy_remote_fds; 1528 1529 if (!pcrfds) { 1530 /* migration has already finished and freed the array */ 1531 return; 1532 } 1533 for (i = 0; i < pcrfds->len; i++) { 1534 struct PostCopyFD *cur = &g_array_index(pcrfds, struct PostCopyFD, i); 1535 if (cur->fd == pcfd->fd) { 1536 mis->postcopy_remote_fds = g_array_remove_index(pcrfds, i); 1537 return; 1538 } 1539 } 1540 } 1541 1542 bool postcopy_preempt_new_channel(MigrationIncomingState *mis, QEMUFile *file) 1543 { 1544 /* 1545 * The new loading channel has its own threads, so it needs to be 1546 * blocked too. It's by default true, just be explicit. 1547 */ 1548 qemu_file_set_blocking(file, true); 1549 mis->postcopy_qemufile_dst = file; 1550 trace_postcopy_preempt_new_channel(); 1551 1552 /* Start the migration immediately */ 1553 return true; 1554 } 1555 1556 /* 1557 * Setup the postcopy preempt channel with the IOC. If ERROR is specified, 1558 * setup the error instead. This helper will free the ERROR if specified. 1559 */ 1560 static void 1561 postcopy_preempt_send_channel_done(MigrationState *s, 1562 QIOChannel *ioc, Error *local_err) 1563 { 1564 if (local_err) { 1565 migrate_set_error(s, local_err); 1566 error_free(local_err); 1567 } else { 1568 migration_ioc_register_yank(ioc); 1569 s->postcopy_qemufile_src = qemu_file_new_output(ioc); 1570 trace_postcopy_preempt_new_channel(); 1571 } 1572 1573 /* 1574 * Kick the waiter in all cases. The waiter should check upon 1575 * postcopy_qemufile_src to know whether it failed or not. 1576 */ 1577 qemu_sem_post(&s->postcopy_qemufile_src_sem); 1578 } 1579 1580 static void 1581 postcopy_preempt_tls_handshake(QIOTask *task, gpointer opaque) 1582 { 1583 g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task)); 1584 MigrationState *s = opaque; 1585 Error *local_err = NULL; 1586 1587 qio_task_propagate_error(task, &local_err); 1588 postcopy_preempt_send_channel_done(s, ioc, local_err); 1589 } 1590 1591 static void 1592 postcopy_preempt_send_channel_new(QIOTask *task, gpointer opaque) 1593 { 1594 g_autoptr(QIOChannel) ioc = QIO_CHANNEL(qio_task_get_source(task)); 1595 MigrationState *s = opaque; 1596 QIOChannelTLS *tioc; 1597 Error *local_err = NULL; 1598 1599 if (qio_task_propagate_error(task, &local_err)) { 1600 goto out; 1601 } 1602 1603 if (migrate_channel_requires_tls_upgrade(ioc)) { 1604 tioc = migration_tls_client_create(s, ioc, s->hostname, &local_err); 1605 if (!tioc) { 1606 goto out; 1607 } 1608 trace_postcopy_preempt_tls_handshake(); 1609 qio_channel_set_name(QIO_CHANNEL(tioc), "migration-tls-preempt"); 1610 qio_channel_tls_handshake(tioc, postcopy_preempt_tls_handshake, 1611 s, NULL, NULL); 1612 /* Setup the channel until TLS handshake finished */ 1613 return; 1614 } 1615 1616 out: 1617 /* This handles both good and error cases */ 1618 postcopy_preempt_send_channel_done(s, ioc, local_err); 1619 } 1620 1621 /* Returns 0 if channel established, -1 for error. */ 1622 int postcopy_preempt_wait_channel(MigrationState *s) 1623 { 1624 /* If preempt not enabled, no need to wait */ 1625 if (!migrate_postcopy_preempt()) { 1626 return 0; 1627 } 1628 1629 /* 1630 * We need the postcopy preempt channel to be established before 1631 * starting doing anything. 1632 */ 1633 qemu_sem_wait(&s->postcopy_qemufile_src_sem); 1634 1635 return s->postcopy_qemufile_src ? 0 : -1; 1636 } 1637 1638 int postcopy_preempt_setup(MigrationState *s, Error **errp) 1639 { 1640 if (!migrate_postcopy_preempt()) { 1641 return 0; 1642 } 1643 1644 if (!migrate_multi_channels_is_allowed()) { 1645 error_setg(errp, "Postcopy preempt is not supported as current " 1646 "migration stream does not support multi-channels."); 1647 return -1; 1648 } 1649 1650 /* Kick an async task to connect */ 1651 socket_send_channel_create(postcopy_preempt_send_channel_new, s); 1652 1653 return 0; 1654 } 1655 1656 static void postcopy_pause_ram_fast_load(MigrationIncomingState *mis) 1657 { 1658 trace_postcopy_pause_fast_load(); 1659 qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex); 1660 qemu_sem_wait(&mis->postcopy_pause_sem_fast_load); 1661 qemu_mutex_lock(&mis->postcopy_prio_thread_mutex); 1662 trace_postcopy_pause_fast_load_continued(); 1663 } 1664 1665 void *postcopy_preempt_thread(void *opaque) 1666 { 1667 MigrationIncomingState *mis = opaque; 1668 int ret; 1669 1670 trace_postcopy_preempt_thread_entry(); 1671 1672 rcu_register_thread(); 1673 1674 qemu_sem_post(&mis->thread_sync_sem); 1675 1676 /* Sending RAM_SAVE_FLAG_EOS to terminate this thread */ 1677 qemu_mutex_lock(&mis->postcopy_prio_thread_mutex); 1678 while (1) { 1679 ret = ram_load_postcopy(mis->postcopy_qemufile_dst, 1680 RAM_CHANNEL_POSTCOPY); 1681 /* If error happened, go into recovery routine */ 1682 if (ret) { 1683 postcopy_pause_ram_fast_load(mis); 1684 } else { 1685 /* We're done */ 1686 break; 1687 } 1688 } 1689 qemu_mutex_unlock(&mis->postcopy_prio_thread_mutex); 1690 1691 rcu_unregister_thread(); 1692 1693 trace_postcopy_preempt_thread_exit(); 1694 1695 return NULL; 1696 } 1697