1 /* 2 * QTest testcase for precopy migration 3 * 4 * Copyright (c) 2016-2018 Red Hat, Inc. and/or its affiliates 5 * based on the vhost-user-test.c that is: 6 * Copyright (c) 2014 Virtual Open Systems Sarl. 7 * 8 * This work is licensed under the terms of the GNU GPL, version 2 or later. 9 * See the COPYING file in the top-level directory. 10 * 11 */ 12 13 #include "qemu/osdep.h" 14 #include "chardev/char.h" 15 #include "crypto/tlscredspsk.h" 16 #include "libqtest.h" 17 #include "migration/bootfile.h" 18 #include "migration/framework.h" 19 #include "migration/migration-qmp.h" 20 #include "migration/migration-util.h" 21 #include "ppc-util.h" 22 #include "qapi/qmp/qlist.h" 23 #include "qemu/module.h" 24 #include "qemu/option.h" 25 #include "qemu/range.h" 26 #include "qemu/sockets.h" 27 28 29 /* 30 * Dirtylimit stop working if dirty page rate error 31 * value less than DIRTYLIMIT_TOLERANCE_RANGE 32 */ 33 #define DIRTYLIMIT_TOLERANCE_RANGE 25 /* MB/s */ 34 35 static char *tmpfs; 36 37 static void test_precopy_unix_plain(void) 38 { 39 g_autofree char *uri = g_strdup_printf("unix:%s/migsocket", tmpfs); 40 MigrateCommon args = { 41 .listen_uri = uri, 42 .connect_uri = uri, 43 /* 44 * The simplest use case of precopy, covering smoke tests of 45 * get-dirty-log dirty tracking. 46 */ 47 .live = true, 48 }; 49 50 test_precopy_common(&args); 51 } 52 53 static void test_precopy_unix_suspend_live(void) 54 { 55 g_autofree char *uri = g_strdup_printf("unix:%s/migsocket", tmpfs); 56 MigrateCommon args = { 57 .listen_uri = uri, 58 .connect_uri = uri, 59 /* 60 * despite being live, the test is fast because the src 61 * suspends immediately. 62 */ 63 .live = true, 64 .start.suspend_me = true, 65 }; 66 67 test_precopy_common(&args); 68 } 69 70 static void test_precopy_unix_suspend_notlive(void) 71 { 72 g_autofree char *uri = g_strdup_printf("unix:%s/migsocket", tmpfs); 73 MigrateCommon args = { 74 .listen_uri = uri, 75 .connect_uri = uri, 76 .start.suspend_me = true, 77 }; 78 79 test_precopy_common(&args); 80 } 81 82 static void test_precopy_unix_dirty_ring(void) 83 { 84 g_autofree char *uri = g_strdup_printf("unix:%s/migsocket", tmpfs); 85 MigrateCommon args = { 86 .start = { 87 .use_dirty_ring = true, 88 }, 89 .listen_uri = uri, 90 .connect_uri = uri, 91 /* 92 * Besides the precopy/unix basic test, cover dirty ring interface 93 * rather than get-dirty-log. 94 */ 95 .live = true, 96 }; 97 98 test_precopy_common(&args); 99 } 100 101 static void test_precopy_tcp_plain(void) 102 { 103 MigrateCommon args = { 104 .listen_uri = "tcp:127.0.0.1:0", 105 }; 106 107 test_precopy_common(&args); 108 } 109 110 static void *migrate_hook_start_switchover_ack(QTestState *from, QTestState *to) 111 { 112 113 migrate_set_capability(from, "return-path", true); 114 migrate_set_capability(to, "return-path", true); 115 116 migrate_set_capability(from, "switchover-ack", true); 117 migrate_set_capability(to, "switchover-ack", true); 118 119 return NULL; 120 } 121 122 static void test_precopy_tcp_switchover_ack(void) 123 { 124 MigrateCommon args = { 125 .listen_uri = "tcp:127.0.0.1:0", 126 .start_hook = migrate_hook_start_switchover_ack, 127 /* 128 * Source VM must be running in order to consider the switchover ACK 129 * when deciding to do switchover or not. 130 */ 131 .live = true, 132 }; 133 134 test_precopy_common(&args); 135 } 136 137 #ifndef _WIN32 138 static void *migrate_hook_start_fd(QTestState *from, 139 QTestState *to) 140 { 141 int ret; 142 int pair[2]; 143 144 /* Create two connected sockets for migration */ 145 ret = qemu_socketpair(PF_LOCAL, SOCK_STREAM, 0, pair); 146 g_assert_cmpint(ret, ==, 0); 147 148 /* Send the 1st socket to the target */ 149 qtest_qmp_fds_assert_success(to, &pair[0], 1, 150 "{ 'execute': 'getfd'," 151 " 'arguments': { 'fdname': 'fd-mig' }}"); 152 close(pair[0]); 153 154 /* Start incoming migration from the 1st socket */ 155 migrate_incoming_qmp(to, "fd:fd-mig", "{}"); 156 157 /* Send the 2nd socket to the target */ 158 qtest_qmp_fds_assert_success(from, &pair[1], 1, 159 "{ 'execute': 'getfd'," 160 " 'arguments': { 'fdname': 'fd-mig' }}"); 161 close(pair[1]); 162 163 return NULL; 164 } 165 166 static void migrate_hook_end_fd(QTestState *from, 167 QTestState *to, 168 void *opaque) 169 { 170 QDict *rsp; 171 const char *error_desc; 172 173 /* Test closing fds */ 174 /* 175 * We assume, that QEMU removes named fd from its list, 176 * so this should fail. 177 */ 178 rsp = qtest_qmp(from, 179 "{ 'execute': 'closefd'," 180 " 'arguments': { 'fdname': 'fd-mig' }}"); 181 g_assert_true(qdict_haskey(rsp, "error")); 182 error_desc = qdict_get_str(qdict_get_qdict(rsp, "error"), "desc"); 183 g_assert_cmpstr(error_desc, ==, "File descriptor named 'fd-mig' not found"); 184 qobject_unref(rsp); 185 186 rsp = qtest_qmp(to, 187 "{ 'execute': 'closefd'," 188 " 'arguments': { 'fdname': 'fd-mig' }}"); 189 g_assert_true(qdict_haskey(rsp, "error")); 190 error_desc = qdict_get_str(qdict_get_qdict(rsp, "error"), "desc"); 191 g_assert_cmpstr(error_desc, ==, "File descriptor named 'fd-mig' not found"); 192 qobject_unref(rsp); 193 } 194 195 static void test_precopy_fd_socket(void) 196 { 197 MigrateCommon args = { 198 .listen_uri = "defer", 199 .connect_uri = "fd:fd-mig", 200 .start_hook = migrate_hook_start_fd, 201 .end_hook = migrate_hook_end_fd, 202 }; 203 test_precopy_common(&args); 204 } 205 206 static void *migrate_hook_start_precopy_fd_file(QTestState *from, 207 QTestState *to) 208 { 209 g_autofree char *file = g_strdup_printf("%s/%s", tmpfs, FILE_TEST_FILENAME); 210 int src_flags = O_CREAT | O_RDWR; 211 int dst_flags = O_CREAT | O_RDWR; 212 int fds[2]; 213 214 fds[0] = open(file, src_flags, 0660); 215 assert(fds[0] != -1); 216 217 fds[1] = open(file, dst_flags, 0660); 218 assert(fds[1] != -1); 219 220 221 qtest_qmp_fds_assert_success(to, &fds[0], 1, 222 "{ 'execute': 'getfd'," 223 " 'arguments': { 'fdname': 'fd-mig' }}"); 224 225 qtest_qmp_fds_assert_success(from, &fds[1], 1, 226 "{ 'execute': 'getfd'," 227 " 'arguments': { 'fdname': 'fd-mig' }}"); 228 229 close(fds[0]); 230 close(fds[1]); 231 232 return NULL; 233 } 234 235 static void test_precopy_fd_file(void) 236 { 237 MigrateCommon args = { 238 .listen_uri = "defer", 239 .connect_uri = "fd:fd-mig", 240 .start_hook = migrate_hook_start_precopy_fd_file, 241 .end_hook = migrate_hook_end_fd, 242 }; 243 test_file_common(&args, true); 244 } 245 #endif /* _WIN32 */ 246 247 /* 248 * The way auto_converge works, we need to do too many passes to 249 * run this test. Auto_converge logic is only run once every 250 * three iterations, so: 251 * 252 * - 3 iterations without auto_converge enabled 253 * - 3 iterations with pct = 5 254 * - 3 iterations with pct = 30 255 * - 3 iterations with pct = 55 256 * - 3 iterations with pct = 80 257 * - 3 iterations with pct = 95 (max(95, 80 + 25)) 258 * 259 * To make things even worse, we need to run the initial stage at 260 * 3MB/s so we enter autoconverge even when host is (over)loaded. 261 */ 262 static void test_auto_converge(void) 263 { 264 g_autofree char *uri = g_strdup_printf("unix:%s/migsocket", tmpfs); 265 MigrateStart args = {}; 266 QTestState *from, *to; 267 int64_t percentage; 268 269 /* 270 * We want the test to be stable and as fast as possible. 271 * E.g., with 1Gb/s bandwidth migration may pass without throttling, 272 * so we need to decrease a bandwidth. 273 */ 274 const int64_t init_pct = 5, inc_pct = 25, max_pct = 95; 275 uint64_t prev_dirty_sync_cnt, dirty_sync_cnt; 276 int max_try_count, hit = 0; 277 278 if (migrate_start(&from, &to, uri, &args)) { 279 return; 280 } 281 282 migrate_set_capability(from, "auto-converge", true); 283 migrate_set_parameter_int(from, "cpu-throttle-initial", init_pct); 284 migrate_set_parameter_int(from, "cpu-throttle-increment", inc_pct); 285 migrate_set_parameter_int(from, "max-cpu-throttle", max_pct); 286 287 /* 288 * Set the initial parameters so that the migration could not converge 289 * without throttling. 290 */ 291 migrate_ensure_non_converge(from); 292 293 /* To check remaining size after precopy */ 294 migrate_set_capability(from, "pause-before-switchover", true); 295 296 /* Wait for the first serial output from the source */ 297 wait_for_serial("src_serial"); 298 299 migrate_qmp(from, to, uri, NULL, "{}"); 300 301 /* Wait for throttling begins */ 302 percentage = 0; 303 do { 304 percentage = read_migrate_property_int(from, "cpu-throttle-percentage"); 305 if (percentage != 0) { 306 break; 307 } 308 usleep(20); 309 g_assert_false(get_src()->stop_seen); 310 } while (true); 311 /* The first percentage of throttling should be at least init_pct */ 312 g_assert_cmpint(percentage, >=, init_pct); 313 314 /* 315 * End the loop when the dirty sync count greater than 1. 316 */ 317 while ((dirty_sync_cnt = get_migration_pass(from)) < 2) { 318 usleep(1000 * 1000); 319 } 320 321 prev_dirty_sync_cnt = dirty_sync_cnt; 322 323 /* 324 * The RAMBlock dirty sync count must changes in 5 seconds, here we set 325 * the timeout to 10 seconds to ensure it changes. 326 * 327 * Note that migrate_ensure_non_converge set the max-bandwidth to 3MB/s, 328 * while the qtest mem is >= 100MB, one iteration takes at least 33s (100/3) 329 * to complete; this ensures that the RAMBlock dirty sync occurs. 330 */ 331 max_try_count = 10; 332 while (--max_try_count) { 333 dirty_sync_cnt = get_migration_pass(from); 334 if (dirty_sync_cnt != prev_dirty_sync_cnt) { 335 hit = 1; 336 break; 337 } 338 prev_dirty_sync_cnt = dirty_sync_cnt; 339 sleep(1); 340 } 341 g_assert_cmpint(hit, ==, 1); 342 343 /* Now, when we tested that throttling works, let it converge */ 344 migrate_ensure_converge(from); 345 346 /* 347 * Wait for pre-switchover status to check last throttle percentage 348 * and remaining. These values will be zeroed later 349 */ 350 wait_for_migration_status(from, "pre-switchover", NULL); 351 352 /* The final percentage of throttling shouldn't be greater than max_pct */ 353 percentage = read_migrate_property_int(from, "cpu-throttle-percentage"); 354 g_assert_cmpint(percentage, <=, max_pct); 355 migrate_continue(from, "pre-switchover"); 356 357 qtest_qmp_eventwait(to, "RESUME"); 358 359 wait_for_serial("dest_serial"); 360 wait_for_migration_complete(from); 361 362 migrate_end(from, to, true); 363 } 364 365 static void * 366 migrate_hook_start_precopy_tcp_multifd(QTestState *from, 367 QTestState *to) 368 { 369 return migrate_hook_start_precopy_tcp_multifd_common(from, to, "none"); 370 } 371 372 static void * 373 migrate_hook_start_precopy_tcp_multifd_zero_page_legacy(QTestState *from, 374 QTestState *to) 375 { 376 migrate_hook_start_precopy_tcp_multifd_common(from, to, "none"); 377 migrate_set_parameter_str(from, "zero-page-detection", "legacy"); 378 return NULL; 379 } 380 381 static void * 382 migrate_hook_start_precopy_tcp_multifd_no_zero_page(QTestState *from, 383 QTestState *to) 384 { 385 migrate_hook_start_precopy_tcp_multifd_common(from, to, "none"); 386 migrate_set_parameter_str(from, "zero-page-detection", "none"); 387 return NULL; 388 } 389 390 static void test_multifd_tcp_uri_none(void) 391 { 392 MigrateCommon args = { 393 .listen_uri = "defer", 394 .start_hook = migrate_hook_start_precopy_tcp_multifd, 395 /* 396 * Multifd is more complicated than most of the features, it 397 * directly takes guest page buffers when sending, make sure 398 * everything will work alright even if guest page is changing. 399 */ 400 .live = true, 401 }; 402 test_precopy_common(&args); 403 } 404 405 static void test_multifd_tcp_zero_page_legacy(void) 406 { 407 MigrateCommon args = { 408 .listen_uri = "defer", 409 .start_hook = migrate_hook_start_precopy_tcp_multifd_zero_page_legacy, 410 /* 411 * Multifd is more complicated than most of the features, it 412 * directly takes guest page buffers when sending, make sure 413 * everything will work alright even if guest page is changing. 414 */ 415 .live = true, 416 }; 417 test_precopy_common(&args); 418 } 419 420 static void test_multifd_tcp_no_zero_page(void) 421 { 422 MigrateCommon args = { 423 .listen_uri = "defer", 424 .start_hook = migrate_hook_start_precopy_tcp_multifd_no_zero_page, 425 /* 426 * Multifd is more complicated than most of the features, it 427 * directly takes guest page buffers when sending, make sure 428 * everything will work alright even if guest page is changing. 429 */ 430 .live = true, 431 }; 432 test_precopy_common(&args); 433 } 434 435 static void test_multifd_tcp_channels_none(void) 436 { 437 MigrateCommon args = { 438 .listen_uri = "defer", 439 .start_hook = migrate_hook_start_precopy_tcp_multifd, 440 .live = true, 441 .connect_channels = ("[ { 'channel-type': 'main'," 442 " 'addr': { 'transport': 'socket'," 443 " 'type': 'inet'," 444 " 'host': '127.0.0.1'," 445 " 'port': '0' } } ]"), 446 }; 447 test_precopy_common(&args); 448 } 449 450 /* 451 * This test does: 452 * source target 453 * migrate_incoming 454 * migrate 455 * migrate_cancel 456 * launch another target 457 * migrate 458 * 459 * And see that it works 460 */ 461 static void test_multifd_tcp_cancel(void) 462 { 463 MigrateStart args = { 464 .hide_stderr = true, 465 }; 466 QTestState *from, *to, *to2; 467 468 if (migrate_start(&from, &to, "defer", &args)) { 469 return; 470 } 471 472 migrate_ensure_non_converge(from); 473 migrate_prepare_for_dirty_mem(from); 474 475 migrate_set_parameter_int(from, "multifd-channels", 16); 476 migrate_set_parameter_int(to, "multifd-channels", 16); 477 478 migrate_set_capability(from, "multifd", true); 479 migrate_set_capability(to, "multifd", true); 480 481 /* Start incoming migration from the 1st socket */ 482 migrate_incoming_qmp(to, "tcp:127.0.0.1:0", "{}"); 483 484 /* Wait for the first serial output from the source */ 485 wait_for_serial("src_serial"); 486 487 migrate_qmp(from, to, NULL, NULL, "{}"); 488 489 migrate_wait_for_dirty_mem(from, to); 490 491 migrate_cancel(from); 492 493 /* Make sure QEMU process "to" exited */ 494 qtest_set_expected_status(to, EXIT_FAILURE); 495 qtest_wait_qemu(to); 496 qtest_quit(to); 497 498 /* 499 * Ensure the source QEMU finishes its cancellation process before we 500 * proceed with the setup of the next migration. The migrate_start() 501 * function and others might want to interact with the source in a way that 502 * is not possible while the migration is not canceled properly. For 503 * example, setting migration capabilities when the migration is still 504 * running leads to an error. 505 */ 506 wait_for_migration_status(from, "cancelled", NULL); 507 508 args = (MigrateStart){ 509 .only_target = true, 510 }; 511 512 if (migrate_start(&from, &to2, "defer", &args)) { 513 return; 514 } 515 516 migrate_set_parameter_int(to2, "multifd-channels", 16); 517 518 migrate_set_capability(to2, "multifd", true); 519 520 /* Start incoming migration from the 1st socket */ 521 migrate_incoming_qmp(to2, "tcp:127.0.0.1:0", "{}"); 522 523 migrate_ensure_non_converge(from); 524 525 migrate_qmp(from, to2, NULL, NULL, "{}"); 526 527 migrate_wait_for_dirty_mem(from, to2); 528 529 migrate_ensure_converge(from); 530 531 wait_for_stop(from, get_src()); 532 qtest_qmp_eventwait(to2, "RESUME"); 533 534 wait_for_serial("dest_serial"); 535 wait_for_migration_complete(from); 536 migrate_end(from, to2, true); 537 } 538 539 static void calc_dirty_rate(QTestState *who, uint64_t calc_time) 540 { 541 qtest_qmp_assert_success(who, 542 "{ 'execute': 'calc-dirty-rate'," 543 "'arguments': { " 544 "'calc-time': %" PRIu64 "," 545 "'mode': 'dirty-ring' }}", 546 calc_time); 547 } 548 549 static QDict *query_dirty_rate(QTestState *who) 550 { 551 return qtest_qmp_assert_success_ref(who, 552 "{ 'execute': 'query-dirty-rate' }"); 553 } 554 555 static void dirtylimit_set_all(QTestState *who, uint64_t dirtyrate) 556 { 557 qtest_qmp_assert_success(who, 558 "{ 'execute': 'set-vcpu-dirty-limit'," 559 "'arguments': { " 560 "'dirty-rate': %" PRIu64 " } }", 561 dirtyrate); 562 } 563 564 static void cancel_vcpu_dirty_limit(QTestState *who) 565 { 566 qtest_qmp_assert_success(who, 567 "{ 'execute': 'cancel-vcpu-dirty-limit' }"); 568 } 569 570 static QDict *query_vcpu_dirty_limit(QTestState *who) 571 { 572 QDict *rsp; 573 574 rsp = qtest_qmp(who, "{ 'execute': 'query-vcpu-dirty-limit' }"); 575 g_assert(!qdict_haskey(rsp, "error")); 576 g_assert(qdict_haskey(rsp, "return")); 577 578 return rsp; 579 } 580 581 static bool calc_dirtyrate_ready(QTestState *who) 582 { 583 QDict *rsp_return; 584 const char *status; 585 bool ready; 586 587 rsp_return = query_dirty_rate(who); 588 g_assert(rsp_return); 589 590 status = qdict_get_str(rsp_return, "status"); 591 g_assert(status); 592 ready = g_strcmp0(status, "measuring"); 593 qobject_unref(rsp_return); 594 595 return ready; 596 } 597 598 static void wait_for_calc_dirtyrate_complete(QTestState *who, 599 int64_t time_s) 600 { 601 int max_try_count = 10000; 602 usleep(time_s * 1000000); 603 604 while (!calc_dirtyrate_ready(who) && max_try_count--) { 605 usleep(1000); 606 } 607 608 /* 609 * Set the timeout with 10 s(max_try_count * 1000us), 610 * if dirtyrate measurement not complete, fail test. 611 */ 612 g_assert_cmpint(max_try_count, !=, 0); 613 } 614 615 static int64_t get_dirty_rate(QTestState *who) 616 { 617 QDict *rsp_return; 618 const char *status; 619 QList *rates; 620 const QListEntry *entry; 621 QDict *rate; 622 int64_t dirtyrate; 623 624 rsp_return = query_dirty_rate(who); 625 g_assert(rsp_return); 626 627 status = qdict_get_str(rsp_return, "status"); 628 g_assert(status); 629 g_assert_cmpstr(status, ==, "measured"); 630 631 rates = qdict_get_qlist(rsp_return, "vcpu-dirty-rate"); 632 g_assert(rates && !qlist_empty(rates)); 633 634 entry = qlist_first(rates); 635 g_assert(entry); 636 637 rate = qobject_to(QDict, qlist_entry_obj(entry)); 638 g_assert(rate); 639 640 dirtyrate = qdict_get_try_int(rate, "dirty-rate", -1); 641 642 qobject_unref(rsp_return); 643 return dirtyrate; 644 } 645 646 static int64_t get_limit_rate(QTestState *who) 647 { 648 QDict *rsp_return; 649 QList *rates; 650 const QListEntry *entry; 651 QDict *rate; 652 int64_t dirtyrate; 653 654 rsp_return = query_vcpu_dirty_limit(who); 655 g_assert(rsp_return); 656 657 rates = qdict_get_qlist(rsp_return, "return"); 658 g_assert(rates && !qlist_empty(rates)); 659 660 entry = qlist_first(rates); 661 g_assert(entry); 662 663 rate = qobject_to(QDict, qlist_entry_obj(entry)); 664 g_assert(rate); 665 666 dirtyrate = qdict_get_try_int(rate, "limit-rate", -1); 667 668 qobject_unref(rsp_return); 669 return dirtyrate; 670 } 671 672 static QTestState *dirtylimit_start_vm(void) 673 { 674 QTestState *vm = NULL; 675 g_autofree gchar *cmd = NULL; 676 const char *bootpath; 677 678 bootpath = bootfile_create(qtest_get_arch(), tmpfs, false); 679 cmd = g_strdup_printf("-accel kvm,dirty-ring-size=4096 " 680 "-name dirtylimit-test,debug-threads=on " 681 "-m 150M -smp 1 " 682 "-serial file:%s/vm_serial " 683 "-drive file=%s,format=raw ", 684 tmpfs, bootpath); 685 686 vm = qtest_init(cmd); 687 return vm; 688 } 689 690 static void dirtylimit_stop_vm(QTestState *vm) 691 { 692 g_autofree char *path = g_strdup_printf("%s/%s", tmpfs, "vm_serial"); 693 694 qtest_quit(vm); 695 unlink(path); 696 } 697 698 static void test_vcpu_dirty_limit(void) 699 { 700 QTestState *vm; 701 int64_t origin_rate; 702 int64_t quota_rate; 703 int64_t rate ; 704 int max_try_count = 20; 705 int hit = 0; 706 707 /* Start vm for vcpu dirtylimit test */ 708 vm = dirtylimit_start_vm(); 709 710 /* Wait for the first serial output from the vm*/ 711 wait_for_serial("vm_serial"); 712 713 /* Do dirtyrate measurement with calc time equals 1s */ 714 calc_dirty_rate(vm, 1); 715 716 /* Sleep calc time and wait for calc dirtyrate complete */ 717 wait_for_calc_dirtyrate_complete(vm, 1); 718 719 /* Query original dirty page rate */ 720 origin_rate = get_dirty_rate(vm); 721 722 /* VM booted from bootsect should dirty memory steadily */ 723 assert(origin_rate != 0); 724 725 /* Setup quota dirty page rate at half of origin */ 726 quota_rate = origin_rate / 2; 727 728 /* Set dirtylimit */ 729 dirtylimit_set_all(vm, quota_rate); 730 731 /* 732 * Check if set-vcpu-dirty-limit and query-vcpu-dirty-limit 733 * works literally 734 */ 735 g_assert_cmpint(quota_rate, ==, get_limit_rate(vm)); 736 737 /* Sleep a bit to check if it take effect */ 738 usleep(2000000); 739 740 /* 741 * Check if dirtylimit take effect realistically, set the 742 * timeout with 20 s(max_try_count * 1s), if dirtylimit 743 * doesn't take effect, fail test. 744 */ 745 while (--max_try_count) { 746 calc_dirty_rate(vm, 1); 747 wait_for_calc_dirtyrate_complete(vm, 1); 748 rate = get_dirty_rate(vm); 749 750 /* 751 * Assume hitting if current rate is less 752 * than quota rate (within accepting error) 753 */ 754 if (rate < (quota_rate + DIRTYLIMIT_TOLERANCE_RANGE)) { 755 hit = 1; 756 break; 757 } 758 } 759 760 g_assert_cmpint(hit, ==, 1); 761 762 hit = 0; 763 max_try_count = 20; 764 765 /* Check if dirtylimit cancellation take effect */ 766 cancel_vcpu_dirty_limit(vm); 767 while (--max_try_count) { 768 calc_dirty_rate(vm, 1); 769 wait_for_calc_dirtyrate_complete(vm, 1); 770 rate = get_dirty_rate(vm); 771 772 /* 773 * Assume dirtylimit be canceled if current rate is 774 * greater than quota rate (within accepting error) 775 */ 776 if (rate > (quota_rate + DIRTYLIMIT_TOLERANCE_RANGE)) { 777 hit = 1; 778 break; 779 } 780 } 781 782 g_assert_cmpint(hit, ==, 1); 783 dirtylimit_stop_vm(vm); 784 } 785 786 static void migrate_dirty_limit_wait_showup(QTestState *from, 787 const int64_t period, 788 const int64_t value) 789 { 790 /* Enable dirty limit capability */ 791 migrate_set_capability(from, "dirty-limit", true); 792 793 /* Set dirty limit parameters */ 794 migrate_set_parameter_int(from, "x-vcpu-dirty-limit-period", period); 795 migrate_set_parameter_int(from, "vcpu-dirty-limit", value); 796 797 /* Make sure migrate can't converge */ 798 migrate_ensure_non_converge(from); 799 800 /* To check limit rate after precopy */ 801 migrate_set_capability(from, "pause-before-switchover", true); 802 803 /* Wait for the serial output from the source */ 804 wait_for_serial("src_serial"); 805 } 806 807 /* 808 * This test does: 809 * source destination 810 * start vm 811 * start incoming vm 812 * migrate 813 * wait dirty limit to begin 814 * cancel migrate 815 * cancellation check 816 * restart incoming vm 817 * migrate 818 * wait dirty limit to begin 819 * wait pre-switchover event 820 * convergence condition check 821 * 822 * And see if dirty limit migration works correctly. 823 * This test case involves many passes, so it runs in slow mode only. 824 */ 825 static void test_dirty_limit(void) 826 { 827 g_autofree char *uri = g_strdup_printf("unix:%s/migsocket", tmpfs); 828 QTestState *from, *to; 829 int64_t remaining; 830 uint64_t throttle_us_per_full; 831 /* 832 * We want the test to be stable and as fast as possible. 833 * E.g., with 1Gb/s bandwidth migration may pass without dirty limit, 834 * so we need to decrease a bandwidth. 835 */ 836 const int64_t dirtylimit_period = 1000, dirtylimit_value = 50; 837 const int64_t max_bandwidth = 400000000; /* ~400Mb/s */ 838 const int64_t downtime_limit = 250; /* 250ms */ 839 /* 840 * We migrate through unix-socket (> 500Mb/s). 841 * Thus, expected migration speed ~= bandwidth limit (< 500Mb/s). 842 * So, we can predict expected_threshold 843 */ 844 const int64_t expected_threshold = max_bandwidth * downtime_limit / 1000; 845 int max_try_count = 10; 846 MigrateCommon args = { 847 .start = { 848 .hide_stderr = true, 849 .use_dirty_ring = true, 850 }, 851 .listen_uri = uri, 852 .connect_uri = uri, 853 }; 854 855 /* Start src, dst vm */ 856 if (migrate_start(&from, &to, args.listen_uri, &args.start)) { 857 return; 858 } 859 860 /* Prepare for dirty limit migration and wait src vm show up */ 861 migrate_dirty_limit_wait_showup(from, dirtylimit_period, dirtylimit_value); 862 863 /* Start migrate */ 864 migrate_qmp(from, to, args.connect_uri, NULL, "{}"); 865 866 /* Wait for dirty limit throttle begin */ 867 throttle_us_per_full = 0; 868 while (throttle_us_per_full == 0) { 869 throttle_us_per_full = 870 read_migrate_property_int(from, 871 "dirty-limit-throttle-time-per-round"); 872 usleep(100); 873 g_assert_false(get_src()->stop_seen); 874 } 875 876 /* Now cancel migrate and wait for dirty limit throttle switch off */ 877 migrate_cancel(from); 878 wait_for_migration_status(from, "cancelled", NULL); 879 880 /* destination always fails after cancel */ 881 migration_event_wait(to, "failed"); 882 qtest_set_expected_status(to, EXIT_FAILURE); 883 qtest_quit(to); 884 885 /* Check if dirty limit throttle switched off, set timeout 1ms */ 886 do { 887 throttle_us_per_full = 888 read_migrate_property_int(from, 889 "dirty-limit-throttle-time-per-round"); 890 usleep(100); 891 g_assert_false(get_src()->stop_seen); 892 } while (throttle_us_per_full != 0 && --max_try_count); 893 894 /* Assert dirty limit is not in service */ 895 g_assert_cmpint(throttle_us_per_full, ==, 0); 896 897 args = (MigrateCommon) { 898 .start = { 899 .only_target = true, 900 .use_dirty_ring = true, 901 }, 902 .listen_uri = uri, 903 .connect_uri = uri, 904 }; 905 906 /* Restart dst vm, src vm already show up so we needn't wait anymore */ 907 if (migrate_start(&from, &to, args.listen_uri, &args.start)) { 908 return; 909 } 910 911 /* Start migrate */ 912 migrate_qmp(from, to, args.connect_uri, NULL, "{}"); 913 914 /* Wait for dirty limit throttle begin */ 915 throttle_us_per_full = 0; 916 while (throttle_us_per_full == 0) { 917 throttle_us_per_full = 918 read_migrate_property_int(from, 919 "dirty-limit-throttle-time-per-round"); 920 usleep(100); 921 g_assert_false(get_src()->stop_seen); 922 } 923 924 /* 925 * The dirty limit rate should equals the return value of 926 * query-vcpu-dirty-limit if dirty limit cap set 927 */ 928 g_assert_cmpint(dirtylimit_value, ==, get_limit_rate(from)); 929 930 /* Now, we have tested if dirty limit works, let it converge */ 931 migrate_set_parameter_int(from, "downtime-limit", downtime_limit); 932 migrate_set_parameter_int(from, "max-bandwidth", max_bandwidth); 933 934 /* 935 * Wait for pre-switchover status to check if migration 936 * satisfy the convergence condition 937 */ 938 wait_for_migration_status(from, "pre-switchover", NULL); 939 940 remaining = read_ram_property_int(from, "remaining"); 941 g_assert_cmpint(remaining, <, 942 (expected_threshold + expected_threshold / 100)); 943 944 migrate_continue(from, "pre-switchover"); 945 946 qtest_qmp_eventwait(to, "RESUME"); 947 948 wait_for_serial("dest_serial"); 949 wait_for_migration_complete(from); 950 951 migrate_end(from, to, true); 952 } 953 954 void migration_test_add_precopy(MigrationTestEnv *env) 955 { 956 tmpfs = env->tmpfs; 957 958 if (env->is_x86) { 959 migration_test_add("/migration/precopy/unix/suspend/live", 960 test_precopy_unix_suspend_live); 961 migration_test_add("/migration/precopy/unix/suspend/notlive", 962 test_precopy_unix_suspend_notlive); 963 } 964 965 migration_test_add("/migration/precopy/unix/plain", 966 test_precopy_unix_plain); 967 968 migration_test_add("/migration/precopy/tcp/plain", test_precopy_tcp_plain); 969 970 migration_test_add("/migration/precopy/tcp/plain/switchover-ack", 971 test_precopy_tcp_switchover_ack); 972 973 #ifndef _WIN32 974 migration_test_add("/migration/precopy/fd/tcp", 975 test_precopy_fd_socket); 976 migration_test_add("/migration/precopy/fd/file", 977 test_precopy_fd_file); 978 #endif 979 980 /* 981 * See explanation why this test is slow on function definition 982 */ 983 if (g_test_slow()) { 984 migration_test_add("/migration/auto_converge", 985 test_auto_converge); 986 if (g_str_equal(env->arch, "x86_64") && 987 env->has_kvm && env->has_dirty_ring) { 988 migration_test_add("/dirty_limit", 989 test_dirty_limit); 990 } 991 } 992 migration_test_add("/migration/multifd/tcp/uri/plain/none", 993 test_multifd_tcp_uri_none); 994 migration_test_add("/migration/multifd/tcp/channels/plain/none", 995 test_multifd_tcp_channels_none); 996 migration_test_add("/migration/multifd/tcp/plain/zero-page/legacy", 997 test_multifd_tcp_zero_page_legacy); 998 migration_test_add("/migration/multifd/tcp/plain/zero-page/none", 999 test_multifd_tcp_no_zero_page); 1000 migration_test_add("/migration/multifd/tcp/plain/cancel", 1001 test_multifd_tcp_cancel); 1002 if (g_str_equal(env->arch, "x86_64") 1003 && env->has_kvm && env->has_dirty_ring) { 1004 1005 migration_test_add("/migration/dirty_ring", 1006 test_precopy_unix_dirty_ring); 1007 if (qtest_has_machine("pc") && g_test_slow()) { 1008 migration_test_add("/migration/vcpu_dirty_limit", 1009 test_vcpu_dirty_limit); 1010 } 1011 } 1012 } 1013