1 /* 2 * Block node draining tests 3 * 4 * Copyright (c) 2017 Kevin Wolf <kwolf@redhat.com> 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a copy 7 * of this software and associated documentation files (the "Software"), to deal 8 * in the Software without restriction, including without limitation the rights 9 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 10 * copies of the Software, and to permit persons to whom the Software is 11 * furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 22 * THE SOFTWARE. 23 */ 24 25 #include "qemu/osdep.h" 26 #include "block/block_int.h" 27 #include "block/blockjob_int.h" 28 #include "sysemu/block-backend.h" 29 #include "qapi/error.h" 30 #include "qemu/main-loop.h" 31 #include "iothread.h" 32 33 static QemuEvent done_event; 34 35 typedef struct BDRVTestState { 36 int drain_count; 37 AioContext *bh_indirection_ctx; 38 bool sleep_in_drain_begin; 39 } BDRVTestState; 40 41 static void coroutine_fn sleep_in_drain_begin(void *opaque) 42 { 43 BlockDriverState *bs = opaque; 44 45 qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000); 46 bdrv_dec_in_flight(bs); 47 } 48 49 static void bdrv_test_drain_begin(BlockDriverState *bs) 50 { 51 BDRVTestState *s = bs->opaque; 52 s->drain_count++; 53 if (s->sleep_in_drain_begin) { 54 Coroutine *co = qemu_coroutine_create(sleep_in_drain_begin, bs); 55 bdrv_inc_in_flight(bs); 56 aio_co_enter(bdrv_get_aio_context(bs), co); 57 } 58 } 59 60 static void bdrv_test_drain_end(BlockDriverState *bs) 61 { 62 BDRVTestState *s = bs->opaque; 63 s->drain_count--; 64 } 65 66 static void bdrv_test_close(BlockDriverState *bs) 67 { 68 BDRVTestState *s = bs->opaque; 69 g_assert_cmpint(s->drain_count, >, 0); 70 } 71 72 static void co_reenter_bh(void *opaque) 73 { 74 aio_co_wake(opaque); 75 } 76 77 static int coroutine_fn bdrv_test_co_preadv(BlockDriverState *bs, 78 int64_t offset, int64_t bytes, 79 QEMUIOVector *qiov, 80 BdrvRequestFlags flags) 81 { 82 BDRVTestState *s = bs->opaque; 83 84 /* We want this request to stay until the polling loop in drain waits for 85 * it to complete. We need to sleep a while as bdrv_drain_invoke() comes 86 * first and polls its result, too, but it shouldn't accidentally complete 87 * this request yet. */ 88 qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000); 89 90 if (s->bh_indirection_ctx) { 91 aio_bh_schedule_oneshot(s->bh_indirection_ctx, co_reenter_bh, 92 qemu_coroutine_self()); 93 qemu_coroutine_yield(); 94 } 95 96 return 0; 97 } 98 99 static int bdrv_test_change_backing_file(BlockDriverState *bs, 100 const char *backing_file, 101 const char *backing_fmt) 102 { 103 return 0; 104 } 105 106 static BlockDriver bdrv_test = { 107 .format_name = "test", 108 .instance_size = sizeof(BDRVTestState), 109 .supports_backing = true, 110 111 .bdrv_close = bdrv_test_close, 112 .bdrv_co_preadv = bdrv_test_co_preadv, 113 114 .bdrv_drain_begin = bdrv_test_drain_begin, 115 .bdrv_drain_end = bdrv_test_drain_end, 116 117 .bdrv_child_perm = bdrv_default_perms, 118 119 .bdrv_change_backing_file = bdrv_test_change_backing_file, 120 }; 121 122 static void aio_ret_cb(void *opaque, int ret) 123 { 124 int *aio_ret = opaque; 125 *aio_ret = ret; 126 } 127 128 typedef struct CallInCoroutineData { 129 void (*entry)(void); 130 bool done; 131 } CallInCoroutineData; 132 133 static coroutine_fn void call_in_coroutine_entry(void *opaque) 134 { 135 CallInCoroutineData *data = opaque; 136 137 data->entry(); 138 data->done = true; 139 } 140 141 static void call_in_coroutine(void (*entry)(void)) 142 { 143 Coroutine *co; 144 CallInCoroutineData data = { 145 .entry = entry, 146 .done = false, 147 }; 148 149 co = qemu_coroutine_create(call_in_coroutine_entry, &data); 150 qemu_coroutine_enter(co); 151 while (!data.done) { 152 aio_poll(qemu_get_aio_context(), true); 153 } 154 } 155 156 enum drain_type { 157 BDRV_DRAIN_ALL, 158 BDRV_DRAIN, 159 DRAIN_TYPE_MAX, 160 }; 161 162 static void do_drain_begin(enum drain_type drain_type, BlockDriverState *bs) 163 { 164 switch (drain_type) { 165 case BDRV_DRAIN_ALL: bdrv_drain_all_begin(); break; 166 case BDRV_DRAIN: bdrv_drained_begin(bs); break; 167 default: g_assert_not_reached(); 168 } 169 } 170 171 static void do_drain_end(enum drain_type drain_type, BlockDriverState *bs) 172 { 173 switch (drain_type) { 174 case BDRV_DRAIN_ALL: bdrv_drain_all_end(); break; 175 case BDRV_DRAIN: bdrv_drained_end(bs); break; 176 default: g_assert_not_reached(); 177 } 178 } 179 180 static void do_drain_begin_unlocked(enum drain_type drain_type, BlockDriverState *bs) 181 { 182 if (drain_type != BDRV_DRAIN_ALL) { 183 aio_context_acquire(bdrv_get_aio_context(bs)); 184 } 185 do_drain_begin(drain_type, bs); 186 if (drain_type != BDRV_DRAIN_ALL) { 187 aio_context_release(bdrv_get_aio_context(bs)); 188 } 189 } 190 191 static BlockBackend * no_coroutine_fn test_setup(void) 192 { 193 BlockBackend *blk; 194 BlockDriverState *bs, *backing; 195 196 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 197 bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR, 198 &error_abort); 199 blk_insert_bs(blk, bs, &error_abort); 200 201 backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort); 202 bdrv_set_backing_hd(bs, backing, &error_abort); 203 204 bdrv_unref(backing); 205 bdrv_unref(bs); 206 207 return blk; 208 } 209 210 static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs) 211 { 212 if (drain_type != BDRV_DRAIN_ALL) { 213 aio_context_acquire(bdrv_get_aio_context(bs)); 214 } 215 do_drain_end(drain_type, bs); 216 if (drain_type != BDRV_DRAIN_ALL) { 217 aio_context_release(bdrv_get_aio_context(bs)); 218 } 219 } 220 221 static void test_drv_cb_common(BlockBackend *blk, enum drain_type drain_type, 222 bool recursive) 223 { 224 BlockDriverState *bs = blk_bs(blk); 225 BlockDriverState *backing = bs->backing->bs; 226 BDRVTestState *s, *backing_s; 227 BlockAIOCB *acb; 228 int aio_ret; 229 230 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0); 231 232 s = bs->opaque; 233 backing_s = backing->opaque; 234 235 /* Simple bdrv_drain_all_begin/end pair, check that CBs are called */ 236 g_assert_cmpint(s->drain_count, ==, 0); 237 g_assert_cmpint(backing_s->drain_count, ==, 0); 238 239 do_drain_begin(drain_type, bs); 240 241 g_assert_cmpint(s->drain_count, ==, 1); 242 g_assert_cmpint(backing_s->drain_count, ==, !!recursive); 243 244 do_drain_end(drain_type, bs); 245 246 g_assert_cmpint(s->drain_count, ==, 0); 247 g_assert_cmpint(backing_s->drain_count, ==, 0); 248 249 /* Now do the same while a request is pending */ 250 aio_ret = -EINPROGRESS; 251 acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret); 252 g_assert(acb != NULL); 253 g_assert_cmpint(aio_ret, ==, -EINPROGRESS); 254 255 g_assert_cmpint(s->drain_count, ==, 0); 256 g_assert_cmpint(backing_s->drain_count, ==, 0); 257 258 do_drain_begin(drain_type, bs); 259 260 g_assert_cmpint(aio_ret, ==, 0); 261 g_assert_cmpint(s->drain_count, ==, 1); 262 g_assert_cmpint(backing_s->drain_count, ==, !!recursive); 263 264 do_drain_end(drain_type, bs); 265 266 g_assert_cmpint(s->drain_count, ==, 0); 267 g_assert_cmpint(backing_s->drain_count, ==, 0); 268 } 269 270 static void test_drv_cb_drain_all(void) 271 { 272 BlockBackend *blk = test_setup(); 273 test_drv_cb_common(blk, BDRV_DRAIN_ALL, true); 274 blk_unref(blk); 275 } 276 277 static void test_drv_cb_drain(void) 278 { 279 BlockBackend *blk = test_setup(); 280 test_drv_cb_common(blk, BDRV_DRAIN, false); 281 blk_unref(blk); 282 } 283 284 static void coroutine_fn test_drv_cb_co_drain_all_entry(void) 285 { 286 BlockBackend *blk = blk_all_next(NULL); 287 test_drv_cb_common(blk, BDRV_DRAIN_ALL, true); 288 } 289 290 static void test_drv_cb_co_drain_all(void) 291 { 292 BlockBackend *blk = test_setup(); 293 call_in_coroutine(test_drv_cb_co_drain_all_entry); 294 blk_unref(blk); 295 } 296 297 static void coroutine_fn test_drv_cb_co_drain_entry(void) 298 { 299 BlockBackend *blk = blk_all_next(NULL); 300 test_drv_cb_common(blk, BDRV_DRAIN, false); 301 } 302 303 static void test_drv_cb_co_drain(void) 304 { 305 BlockBackend *blk = test_setup(); 306 call_in_coroutine(test_drv_cb_co_drain_entry); 307 blk_unref(blk); 308 } 309 310 static void test_quiesce_common(BlockBackend *blk, enum drain_type drain_type, 311 bool recursive) 312 { 313 BlockDriverState *bs = blk_bs(blk); 314 BlockDriverState *backing = bs->backing->bs; 315 316 g_assert_cmpint(bs->quiesce_counter, ==, 0); 317 g_assert_cmpint(backing->quiesce_counter, ==, 0); 318 319 do_drain_begin(drain_type, bs); 320 321 if (drain_type == BDRV_DRAIN_ALL) { 322 g_assert_cmpint(bs->quiesce_counter, ==, 2); 323 } else { 324 g_assert_cmpint(bs->quiesce_counter, ==, 1); 325 } 326 g_assert_cmpint(backing->quiesce_counter, ==, !!recursive); 327 328 do_drain_end(drain_type, bs); 329 330 g_assert_cmpint(bs->quiesce_counter, ==, 0); 331 g_assert_cmpint(backing->quiesce_counter, ==, 0); 332 } 333 334 static void test_quiesce_drain_all(void) 335 { 336 BlockBackend *blk = test_setup(); 337 test_quiesce_common(blk, BDRV_DRAIN_ALL, true); 338 blk_unref(blk); 339 } 340 341 static void test_quiesce_drain(void) 342 { 343 BlockBackend *blk = test_setup(); 344 test_quiesce_common(blk, BDRV_DRAIN, false); 345 blk_unref(blk); 346 } 347 348 static void coroutine_fn test_quiesce_co_drain_all_entry(void) 349 { 350 BlockBackend *blk = blk_all_next(NULL); 351 test_quiesce_common(blk, BDRV_DRAIN_ALL, true); 352 } 353 354 static void test_quiesce_co_drain_all(void) 355 { 356 BlockBackend *blk = test_setup(); 357 call_in_coroutine(test_quiesce_co_drain_all_entry); 358 blk_unref(blk); 359 } 360 361 static void coroutine_fn test_quiesce_co_drain_entry(void) 362 { 363 BlockBackend *blk = blk_all_next(NULL); 364 test_quiesce_common(blk, BDRV_DRAIN, false); 365 } 366 367 static void test_quiesce_co_drain(void) 368 { 369 BlockBackend *blk = test_setup(); 370 call_in_coroutine(test_quiesce_co_drain_entry); 371 blk_unref(blk); 372 } 373 374 static void test_nested(void) 375 { 376 BlockBackend *blk; 377 BlockDriverState *bs, *backing; 378 BDRVTestState *s, *backing_s; 379 enum drain_type outer, inner; 380 381 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 382 bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR, 383 &error_abort); 384 s = bs->opaque; 385 blk_insert_bs(blk, bs, &error_abort); 386 387 backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort); 388 backing_s = backing->opaque; 389 bdrv_set_backing_hd(bs, backing, &error_abort); 390 391 for (outer = 0; outer < DRAIN_TYPE_MAX; outer++) { 392 for (inner = 0; inner < DRAIN_TYPE_MAX; inner++) { 393 int backing_quiesce = (outer == BDRV_DRAIN_ALL) + 394 (inner == BDRV_DRAIN_ALL); 395 396 g_assert_cmpint(bs->quiesce_counter, ==, 0); 397 g_assert_cmpint(backing->quiesce_counter, ==, 0); 398 g_assert_cmpint(s->drain_count, ==, 0); 399 g_assert_cmpint(backing_s->drain_count, ==, 0); 400 401 do_drain_begin(outer, bs); 402 do_drain_begin(inner, bs); 403 404 g_assert_cmpint(bs->quiesce_counter, ==, 2 + !!backing_quiesce); 405 g_assert_cmpint(backing->quiesce_counter, ==, backing_quiesce); 406 g_assert_cmpint(s->drain_count, ==, 1); 407 g_assert_cmpint(backing_s->drain_count, ==, !!backing_quiesce); 408 409 do_drain_end(inner, bs); 410 do_drain_end(outer, bs); 411 412 g_assert_cmpint(bs->quiesce_counter, ==, 0); 413 g_assert_cmpint(backing->quiesce_counter, ==, 0); 414 g_assert_cmpint(s->drain_count, ==, 0); 415 g_assert_cmpint(backing_s->drain_count, ==, 0); 416 } 417 } 418 419 bdrv_unref(backing); 420 bdrv_unref(bs); 421 blk_unref(blk); 422 } 423 424 static void test_graph_change_drain_all(void) 425 { 426 BlockBackend *blk_a, *blk_b; 427 BlockDriverState *bs_a, *bs_b; 428 BDRVTestState *a_s, *b_s; 429 430 /* Create node A with a BlockBackend */ 431 blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 432 bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR, 433 &error_abort); 434 a_s = bs_a->opaque; 435 blk_insert_bs(blk_a, bs_a, &error_abort); 436 437 g_assert_cmpint(bs_a->quiesce_counter, ==, 0); 438 g_assert_cmpint(a_s->drain_count, ==, 0); 439 440 /* Call bdrv_drain_all_begin() */ 441 bdrv_drain_all_begin(); 442 443 g_assert_cmpint(bs_a->quiesce_counter, ==, 1); 444 g_assert_cmpint(a_s->drain_count, ==, 1); 445 446 /* Create node B with a BlockBackend */ 447 blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 448 bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR, 449 &error_abort); 450 b_s = bs_b->opaque; 451 blk_insert_bs(blk_b, bs_b, &error_abort); 452 453 g_assert_cmpint(bs_a->quiesce_counter, ==, 1); 454 g_assert_cmpint(bs_b->quiesce_counter, ==, 1); 455 g_assert_cmpint(a_s->drain_count, ==, 1); 456 g_assert_cmpint(b_s->drain_count, ==, 1); 457 458 /* Unref and finally delete node A */ 459 blk_unref(blk_a); 460 461 g_assert_cmpint(bs_a->quiesce_counter, ==, 1); 462 g_assert_cmpint(bs_b->quiesce_counter, ==, 1); 463 g_assert_cmpint(a_s->drain_count, ==, 1); 464 g_assert_cmpint(b_s->drain_count, ==, 1); 465 466 bdrv_unref(bs_a); 467 468 g_assert_cmpint(bs_b->quiesce_counter, ==, 1); 469 g_assert_cmpint(b_s->drain_count, ==, 1); 470 471 /* End the drained section */ 472 bdrv_drain_all_end(); 473 474 g_assert_cmpint(bs_b->quiesce_counter, ==, 0); 475 g_assert_cmpint(b_s->drain_count, ==, 0); 476 477 bdrv_unref(bs_b); 478 blk_unref(blk_b); 479 } 480 481 struct test_iothread_data { 482 BlockDriverState *bs; 483 enum drain_type drain_type; 484 int *aio_ret; 485 bool co_done; 486 }; 487 488 static void coroutine_fn test_iothread_drain_co_entry(void *opaque) 489 { 490 struct test_iothread_data *data = opaque; 491 492 do_drain_begin(data->drain_type, data->bs); 493 g_assert_cmpint(*data->aio_ret, ==, 0); 494 do_drain_end(data->drain_type, data->bs); 495 496 data->co_done = true; 497 aio_wait_kick(); 498 } 499 500 static void test_iothread_aio_cb(void *opaque, int ret) 501 { 502 int *aio_ret = opaque; 503 *aio_ret = ret; 504 qemu_event_set(&done_event); 505 } 506 507 static void test_iothread_main_thread_bh(void *opaque) 508 { 509 struct test_iothread_data *data = opaque; 510 511 /* Test that the AioContext is not yet locked in a random BH that is 512 * executed during drain, otherwise this would deadlock. */ 513 aio_context_acquire(bdrv_get_aio_context(data->bs)); 514 bdrv_flush(data->bs); 515 aio_context_release(bdrv_get_aio_context(data->bs)); 516 } 517 518 /* 519 * Starts an AIO request on a BDS that runs in the AioContext of iothread 1. 520 * The request involves a BH on iothread 2 before it can complete. 521 * 522 * @drain_thread = 0 means that do_drain_begin/end are called from the main 523 * thread, @drain_thread = 1 means that they are called from iothread 1. Drain 524 * for this BDS cannot be called from iothread 2 because only the main thread 525 * may do cross-AioContext polling. 526 */ 527 static void test_iothread_common(enum drain_type drain_type, int drain_thread) 528 { 529 BlockBackend *blk; 530 BlockDriverState *bs; 531 BDRVTestState *s; 532 BlockAIOCB *acb; 533 Coroutine *co; 534 int aio_ret; 535 struct test_iothread_data data; 536 537 IOThread *a = iothread_new(); 538 IOThread *b = iothread_new(); 539 AioContext *ctx_a = iothread_get_aio_context(a); 540 AioContext *ctx_b = iothread_get_aio_context(b); 541 542 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0); 543 544 /* bdrv_drain_all() may only be called from the main loop thread */ 545 if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) { 546 goto out; 547 } 548 549 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 550 bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR, 551 &error_abort); 552 s = bs->opaque; 553 blk_insert_bs(blk, bs, &error_abort); 554 blk_set_disable_request_queuing(blk, true); 555 556 blk_set_aio_context(blk, ctx_a, &error_abort); 557 aio_context_acquire(ctx_a); 558 559 s->bh_indirection_ctx = ctx_b; 560 561 aio_ret = -EINPROGRESS; 562 qemu_event_reset(&done_event); 563 564 if (drain_thread == 0) { 565 acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret); 566 } else { 567 acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret); 568 } 569 g_assert(acb != NULL); 570 g_assert_cmpint(aio_ret, ==, -EINPROGRESS); 571 572 aio_context_release(ctx_a); 573 574 data = (struct test_iothread_data) { 575 .bs = bs, 576 .drain_type = drain_type, 577 .aio_ret = &aio_ret, 578 }; 579 580 switch (drain_thread) { 581 case 0: 582 if (drain_type != BDRV_DRAIN_ALL) { 583 aio_context_acquire(ctx_a); 584 } 585 586 aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data); 587 588 /* The request is running on the IOThread a. Draining its block device 589 * will make sure that it has completed as far as the BDS is concerned, 590 * but the drain in this thread can continue immediately after 591 * bdrv_dec_in_flight() and aio_ret might be assigned only slightly 592 * later. */ 593 do_drain_begin(drain_type, bs); 594 g_assert_cmpint(bs->in_flight, ==, 0); 595 596 if (drain_type != BDRV_DRAIN_ALL) { 597 aio_context_release(ctx_a); 598 } 599 qemu_event_wait(&done_event); 600 if (drain_type != BDRV_DRAIN_ALL) { 601 aio_context_acquire(ctx_a); 602 } 603 604 g_assert_cmpint(aio_ret, ==, 0); 605 do_drain_end(drain_type, bs); 606 607 if (drain_type != BDRV_DRAIN_ALL) { 608 aio_context_release(ctx_a); 609 } 610 break; 611 case 1: 612 co = qemu_coroutine_create(test_iothread_drain_co_entry, &data); 613 aio_co_enter(ctx_a, co); 614 AIO_WAIT_WHILE_UNLOCKED(NULL, !data.co_done); 615 break; 616 default: 617 g_assert_not_reached(); 618 } 619 620 aio_context_acquire(ctx_a); 621 blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort); 622 aio_context_release(ctx_a); 623 624 bdrv_unref(bs); 625 blk_unref(blk); 626 627 out: 628 iothread_join(a); 629 iothread_join(b); 630 } 631 632 static void test_iothread_drain_all(void) 633 { 634 test_iothread_common(BDRV_DRAIN_ALL, 0); 635 test_iothread_common(BDRV_DRAIN_ALL, 1); 636 } 637 638 static void test_iothread_drain(void) 639 { 640 test_iothread_common(BDRV_DRAIN, 0); 641 test_iothread_common(BDRV_DRAIN, 1); 642 } 643 644 645 typedef struct TestBlockJob { 646 BlockJob common; 647 BlockDriverState *bs; 648 int run_ret; 649 int prepare_ret; 650 bool running; 651 bool should_complete; 652 } TestBlockJob; 653 654 static int test_job_prepare(Job *job) 655 { 656 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 657 658 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */ 659 bdrv_flush(s->bs); 660 return s->prepare_ret; 661 } 662 663 static void test_job_commit(Job *job) 664 { 665 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 666 667 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */ 668 bdrv_flush(s->bs); 669 } 670 671 static void test_job_abort(Job *job) 672 { 673 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 674 675 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */ 676 bdrv_flush(s->bs); 677 } 678 679 static int coroutine_fn test_job_run(Job *job, Error **errp) 680 { 681 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 682 683 /* We are running the actual job code past the pause point in 684 * job_co_entry(). */ 685 s->running = true; 686 687 job_transition_to_ready(&s->common.job); 688 while (!s->should_complete) { 689 /* Avoid job_sleep_ns() because it marks the job as !busy. We want to 690 * emulate some actual activity (probably some I/O) here so that drain 691 * has to wait for this activity to stop. */ 692 qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000); 693 694 job_pause_point(&s->common.job); 695 } 696 697 return s->run_ret; 698 } 699 700 static void test_job_complete(Job *job, Error **errp) 701 { 702 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 703 s->should_complete = true; 704 } 705 706 BlockJobDriver test_job_driver = { 707 .job_driver = { 708 .instance_size = sizeof(TestBlockJob), 709 .free = block_job_free, 710 .user_resume = block_job_user_resume, 711 .run = test_job_run, 712 .complete = test_job_complete, 713 .prepare = test_job_prepare, 714 .commit = test_job_commit, 715 .abort = test_job_abort, 716 }, 717 }; 718 719 enum test_job_result { 720 TEST_JOB_SUCCESS, 721 TEST_JOB_FAIL_RUN, 722 TEST_JOB_FAIL_PREPARE, 723 }; 724 725 enum test_job_drain_node { 726 TEST_JOB_DRAIN_SRC, 727 TEST_JOB_DRAIN_SRC_CHILD, 728 }; 729 730 static void test_blockjob_common_drain_node(enum drain_type drain_type, 731 bool use_iothread, 732 enum test_job_result result, 733 enum test_job_drain_node drain_node) 734 { 735 BlockBackend *blk_src, *blk_target; 736 BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs; 737 BlockJob *job; 738 TestBlockJob *tjob; 739 IOThread *iothread = NULL; 740 AioContext *ctx; 741 int ret; 742 743 src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR, 744 &error_abort); 745 src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing", 746 BDRV_O_RDWR, &error_abort); 747 src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay", 748 BDRV_O_RDWR, &error_abort); 749 750 bdrv_set_backing_hd(src_overlay, src, &error_abort); 751 bdrv_unref(src); 752 bdrv_set_backing_hd(src, src_backing, &error_abort); 753 bdrv_unref(src_backing); 754 755 blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 756 blk_insert_bs(blk_src, src_overlay, &error_abort); 757 758 switch (drain_node) { 759 case TEST_JOB_DRAIN_SRC: 760 drain_bs = src; 761 break; 762 case TEST_JOB_DRAIN_SRC_CHILD: 763 drain_bs = src_backing; 764 break; 765 default: 766 g_assert_not_reached(); 767 } 768 769 if (use_iothread) { 770 iothread = iothread_new(); 771 ctx = iothread_get_aio_context(iothread); 772 blk_set_aio_context(blk_src, ctx, &error_abort); 773 } else { 774 ctx = qemu_get_aio_context(); 775 } 776 777 target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR, 778 &error_abort); 779 blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 780 blk_insert_bs(blk_target, target, &error_abort); 781 blk_set_allow_aio_context_change(blk_target, true); 782 783 aio_context_acquire(ctx); 784 tjob = block_job_create("job0", &test_job_driver, NULL, src, 785 0, BLK_PERM_ALL, 786 0, 0, NULL, NULL, &error_abort); 787 tjob->bs = src; 788 job = &tjob->common; 789 block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort); 790 791 switch (result) { 792 case TEST_JOB_SUCCESS: 793 break; 794 case TEST_JOB_FAIL_RUN: 795 tjob->run_ret = -EIO; 796 break; 797 case TEST_JOB_FAIL_PREPARE: 798 tjob->prepare_ret = -EIO; 799 break; 800 } 801 aio_context_release(ctx); 802 803 job_start(&job->job); 804 805 if (use_iothread) { 806 /* job_co_entry() is run in the I/O thread, wait for the actual job 807 * code to start (we don't want to catch the job in the pause point in 808 * job_co_entry(). */ 809 while (!tjob->running) { 810 aio_poll(qemu_get_aio_context(), false); 811 } 812 } 813 814 WITH_JOB_LOCK_GUARD() { 815 g_assert_cmpint(job->job.pause_count, ==, 0); 816 g_assert_false(job->job.paused); 817 g_assert_true(tjob->running); 818 g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */ 819 } 820 821 do_drain_begin_unlocked(drain_type, drain_bs); 822 823 WITH_JOB_LOCK_GUARD() { 824 if (drain_type == BDRV_DRAIN_ALL) { 825 /* bdrv_drain_all() drains both src and target */ 826 g_assert_cmpint(job->job.pause_count, ==, 2); 827 } else { 828 g_assert_cmpint(job->job.pause_count, ==, 1); 829 } 830 g_assert_true(job->job.paused); 831 g_assert_false(job->job.busy); /* The job is paused */ 832 } 833 834 do_drain_end_unlocked(drain_type, drain_bs); 835 836 if (use_iothread) { 837 /* 838 * Here we are waiting for the paused status to change, 839 * so don't bother protecting the read every time. 840 * 841 * paused is reset in the I/O thread, wait for it 842 */ 843 while (job->job.paused) { 844 aio_poll(qemu_get_aio_context(), false); 845 } 846 } 847 848 WITH_JOB_LOCK_GUARD() { 849 g_assert_cmpint(job->job.pause_count, ==, 0); 850 g_assert_false(job->job.paused); 851 g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */ 852 } 853 854 do_drain_begin_unlocked(drain_type, target); 855 856 WITH_JOB_LOCK_GUARD() { 857 if (drain_type == BDRV_DRAIN_ALL) { 858 /* bdrv_drain_all() drains both src and target */ 859 g_assert_cmpint(job->job.pause_count, ==, 2); 860 } else { 861 g_assert_cmpint(job->job.pause_count, ==, 1); 862 } 863 g_assert_true(job->job.paused); 864 g_assert_false(job->job.busy); /* The job is paused */ 865 } 866 867 do_drain_end_unlocked(drain_type, target); 868 869 if (use_iothread) { 870 /* 871 * Here we are waiting for the paused status to change, 872 * so don't bother protecting the read every time. 873 * 874 * paused is reset in the I/O thread, wait for it 875 */ 876 while (job->job.paused) { 877 aio_poll(qemu_get_aio_context(), false); 878 } 879 } 880 881 WITH_JOB_LOCK_GUARD() { 882 g_assert_cmpint(job->job.pause_count, ==, 0); 883 g_assert_false(job->job.paused); 884 g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */ 885 } 886 887 WITH_JOB_LOCK_GUARD() { 888 ret = job_complete_sync_locked(&job->job, &error_abort); 889 } 890 g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO)); 891 892 aio_context_acquire(ctx); 893 if (use_iothread) { 894 blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort); 895 assert(blk_get_aio_context(blk_target) == qemu_get_aio_context()); 896 } 897 aio_context_release(ctx); 898 899 blk_unref(blk_src); 900 blk_unref(blk_target); 901 bdrv_unref(src_overlay); 902 bdrv_unref(target); 903 904 if (iothread) { 905 iothread_join(iothread); 906 } 907 } 908 909 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread, 910 enum test_job_result result) 911 { 912 test_blockjob_common_drain_node(drain_type, use_iothread, result, 913 TEST_JOB_DRAIN_SRC); 914 test_blockjob_common_drain_node(drain_type, use_iothread, result, 915 TEST_JOB_DRAIN_SRC_CHILD); 916 } 917 918 static void test_blockjob_drain_all(void) 919 { 920 test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS); 921 } 922 923 static void test_blockjob_drain(void) 924 { 925 test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS); 926 } 927 928 static void test_blockjob_error_drain_all(void) 929 { 930 test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN); 931 test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE); 932 } 933 934 static void test_blockjob_error_drain(void) 935 { 936 test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN); 937 test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE); 938 } 939 940 static void test_blockjob_iothread_drain_all(void) 941 { 942 test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS); 943 } 944 945 static void test_blockjob_iothread_drain(void) 946 { 947 test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS); 948 } 949 950 static void test_blockjob_iothread_error_drain_all(void) 951 { 952 test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN); 953 test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE); 954 } 955 956 static void test_blockjob_iothread_error_drain(void) 957 { 958 test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN); 959 test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE); 960 } 961 962 963 typedef struct BDRVTestTopState { 964 BdrvChild *wait_child; 965 } BDRVTestTopState; 966 967 static void bdrv_test_top_close(BlockDriverState *bs) 968 { 969 BdrvChild *c, *next_c; 970 QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) { 971 bdrv_unref_child(bs, c); 972 } 973 } 974 975 static int coroutine_fn GRAPH_RDLOCK 976 bdrv_test_top_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes, 977 QEMUIOVector *qiov, BdrvRequestFlags flags) 978 { 979 BDRVTestTopState *tts = bs->opaque; 980 return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags); 981 } 982 983 static BlockDriver bdrv_test_top_driver = { 984 .format_name = "test_top_driver", 985 .instance_size = sizeof(BDRVTestTopState), 986 987 .bdrv_close = bdrv_test_top_close, 988 .bdrv_co_preadv = bdrv_test_top_co_preadv, 989 990 .bdrv_child_perm = bdrv_default_perms, 991 }; 992 993 typedef struct TestCoDeleteByDrainData { 994 BlockBackend *blk; 995 bool detach_instead_of_delete; 996 bool done; 997 } TestCoDeleteByDrainData; 998 999 static void coroutine_fn test_co_delete_by_drain(void *opaque) 1000 { 1001 TestCoDeleteByDrainData *dbdd = opaque; 1002 BlockBackend *blk = dbdd->blk; 1003 BlockDriverState *bs = blk_bs(blk); 1004 BDRVTestTopState *tts = bs->opaque; 1005 void *buffer = g_malloc(65536); 1006 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536); 1007 1008 /* Pretend some internal write operation from parent to child. 1009 * Important: We have to read from the child, not from the parent! 1010 * Draining works by first propagating it all up the tree to the 1011 * root and then waiting for drainage from root to the leaves 1012 * (protocol nodes). If we have a request waiting on the root, 1013 * everything will be drained before we go back down the tree, but 1014 * we do not want that. We want to be in the middle of draining 1015 * when this following requests returns. */ 1016 bdrv_graph_co_rdlock(); 1017 bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0); 1018 bdrv_graph_co_rdunlock(); 1019 1020 g_assert_cmpint(bs->refcnt, ==, 1); 1021 1022 if (!dbdd->detach_instead_of_delete) { 1023 blk_co_unref(blk); 1024 } else { 1025 BdrvChild *c, *next_c; 1026 QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) { 1027 bdrv_unref_child(bs, c); 1028 } 1029 } 1030 1031 dbdd->done = true; 1032 g_free(buffer); 1033 } 1034 1035 /** 1036 * Test what happens when some BDS has some children, you drain one of 1037 * them and this results in the BDS being deleted. 1038 * 1039 * If @detach_instead_of_delete is set, the BDS is not going to be 1040 * deleted but will only detach all of its children. 1041 */ 1042 static void do_test_delete_by_drain(bool detach_instead_of_delete, 1043 enum drain_type drain_type) 1044 { 1045 BlockBackend *blk; 1046 BlockDriverState *bs, *child_bs, *null_bs; 1047 BDRVTestTopState *tts; 1048 TestCoDeleteByDrainData dbdd; 1049 Coroutine *co; 1050 1051 bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR, 1052 &error_abort); 1053 bs->total_sectors = 65536 >> BDRV_SECTOR_BITS; 1054 tts = bs->opaque; 1055 1056 null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL, 1057 &error_abort); 1058 bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, 1059 BDRV_CHILD_DATA, &error_abort); 1060 1061 /* This child will be the one to pass to requests through to, and 1062 * it will stall until a drain occurs */ 1063 child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR, 1064 &error_abort); 1065 child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS; 1066 /* Takes our reference to child_bs */ 1067 tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child", 1068 &child_of_bds, 1069 BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY, 1070 &error_abort); 1071 1072 /* This child is just there to be deleted 1073 * (for detach_instead_of_delete == true) */ 1074 null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL, 1075 &error_abort); 1076 bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA, 1077 &error_abort); 1078 1079 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 1080 blk_insert_bs(blk, bs, &error_abort); 1081 1082 /* Referenced by blk now */ 1083 bdrv_unref(bs); 1084 1085 g_assert_cmpint(bs->refcnt, ==, 1); 1086 g_assert_cmpint(child_bs->refcnt, ==, 1); 1087 g_assert_cmpint(null_bs->refcnt, ==, 1); 1088 1089 1090 dbdd = (TestCoDeleteByDrainData){ 1091 .blk = blk, 1092 .detach_instead_of_delete = detach_instead_of_delete, 1093 .done = false, 1094 }; 1095 co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd); 1096 qemu_coroutine_enter(co); 1097 1098 /* Drain the child while the read operation is still pending. 1099 * This should result in the operation finishing and 1100 * test_co_delete_by_drain() resuming. Thus, @bs will be deleted 1101 * and the coroutine will exit while this drain operation is still 1102 * in progress. */ 1103 switch (drain_type) { 1104 case BDRV_DRAIN: 1105 bdrv_ref(child_bs); 1106 bdrv_drain(child_bs); 1107 bdrv_unref(child_bs); 1108 break; 1109 case BDRV_DRAIN_ALL: 1110 bdrv_drain_all_begin(); 1111 bdrv_drain_all_end(); 1112 break; 1113 default: 1114 g_assert_not_reached(); 1115 } 1116 1117 while (!dbdd.done) { 1118 aio_poll(qemu_get_aio_context(), true); 1119 } 1120 1121 if (detach_instead_of_delete) { 1122 /* Here, the reference has not passed over to the coroutine, 1123 * so we have to delete the BB ourselves */ 1124 blk_unref(blk); 1125 } 1126 } 1127 1128 static void test_delete_by_drain(void) 1129 { 1130 do_test_delete_by_drain(false, BDRV_DRAIN); 1131 } 1132 1133 static void test_detach_by_drain_all(void) 1134 { 1135 do_test_delete_by_drain(true, BDRV_DRAIN_ALL); 1136 } 1137 1138 static void test_detach_by_drain(void) 1139 { 1140 do_test_delete_by_drain(true, BDRV_DRAIN); 1141 } 1142 1143 1144 struct detach_by_parent_data { 1145 BlockDriverState *parent_b; 1146 BdrvChild *child_b; 1147 BlockDriverState *c; 1148 BdrvChild *child_c; 1149 bool by_parent_cb; 1150 bool detach_on_drain; 1151 }; 1152 static struct detach_by_parent_data detach_by_parent_data; 1153 1154 static void detach_indirect_bh(void *opaque) 1155 { 1156 struct detach_by_parent_data *data = opaque; 1157 1158 bdrv_dec_in_flight(data->child_b->bs); 1159 bdrv_unref_child(data->parent_b, data->child_b); 1160 1161 bdrv_ref(data->c); 1162 data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C", 1163 &child_of_bds, BDRV_CHILD_DATA, 1164 &error_abort); 1165 } 1166 1167 static void detach_by_parent_aio_cb(void *opaque, int ret) 1168 { 1169 struct detach_by_parent_data *data = &detach_by_parent_data; 1170 1171 g_assert_cmpint(ret, ==, 0); 1172 if (data->by_parent_cb) { 1173 bdrv_inc_in_flight(data->child_b->bs); 1174 detach_indirect_bh(data); 1175 } 1176 } 1177 1178 static void detach_by_driver_cb_drained_begin(BdrvChild *child) 1179 { 1180 struct detach_by_parent_data *data = &detach_by_parent_data; 1181 1182 if (!data->detach_on_drain) { 1183 return; 1184 } 1185 data->detach_on_drain = false; 1186 1187 bdrv_inc_in_flight(data->child_b->bs); 1188 aio_bh_schedule_oneshot(qemu_get_current_aio_context(), 1189 detach_indirect_bh, &detach_by_parent_data); 1190 child_of_bds.drained_begin(child); 1191 } 1192 1193 static BdrvChildClass detach_by_driver_cb_class; 1194 1195 /* 1196 * Initial graph: 1197 * 1198 * PA PB 1199 * \ / \ 1200 * A B C 1201 * 1202 * by_parent_cb == true: Test that parent callbacks don't poll 1203 * 1204 * PA has a pending write request whose callback changes the child nodes of 1205 * PB: It removes B and adds C instead. The subtree of PB is drained, which 1206 * will indirectly drain the write request, too. 1207 * 1208 * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll 1209 * 1210 * PA's BdrvChildClass has a .drained_begin callback that schedules a BH 1211 * that does the same graph change. If bdrv_drain_invoke() calls it, the 1212 * state is messed up, but if it is only polled in the single 1213 * BDRV_POLL_WHILE() at the end of the drain, this should work fine. 1214 */ 1215 static void test_detach_indirect(bool by_parent_cb) 1216 { 1217 BlockBackend *blk; 1218 BlockDriverState *parent_a, *parent_b, *a, *b, *c; 1219 BdrvChild *child_a, *child_b; 1220 BlockAIOCB *acb; 1221 1222 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0); 1223 1224 if (!by_parent_cb) { 1225 detach_by_driver_cb_class = child_of_bds; 1226 detach_by_driver_cb_class.drained_begin = 1227 detach_by_driver_cb_drained_begin; 1228 detach_by_driver_cb_class.drained_end = NULL; 1229 detach_by_driver_cb_class.drained_poll = NULL; 1230 } 1231 1232 detach_by_parent_data = (struct detach_by_parent_data) { 1233 .detach_on_drain = false, 1234 }; 1235 1236 /* Create all involved nodes */ 1237 parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR, 1238 &error_abort); 1239 parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0, 1240 &error_abort); 1241 1242 a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort); 1243 b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort); 1244 c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort); 1245 1246 /* blk is a BB for parent-a */ 1247 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 1248 blk_insert_bs(blk, parent_a, &error_abort); 1249 bdrv_unref(parent_a); 1250 1251 /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver 1252 * callback must not return immediately. */ 1253 if (!by_parent_cb) { 1254 BDRVTestState *s = parent_a->opaque; 1255 s->sleep_in_drain_begin = true; 1256 } 1257 1258 /* Set child relationships */ 1259 bdrv_ref(b); 1260 bdrv_ref(a); 1261 child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds, 1262 BDRV_CHILD_DATA, &error_abort); 1263 child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds, 1264 BDRV_CHILD_COW, &error_abort); 1265 1266 bdrv_ref(a); 1267 bdrv_attach_child(parent_a, a, "PA-A", 1268 by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class, 1269 BDRV_CHILD_DATA, &error_abort); 1270 1271 g_assert_cmpint(parent_a->refcnt, ==, 1); 1272 g_assert_cmpint(parent_b->refcnt, ==, 1); 1273 g_assert_cmpint(a->refcnt, ==, 3); 1274 g_assert_cmpint(b->refcnt, ==, 2); 1275 g_assert_cmpint(c->refcnt, ==, 1); 1276 1277 g_assert(QLIST_FIRST(&parent_b->children) == child_a); 1278 g_assert(QLIST_NEXT(child_a, next) == child_b); 1279 g_assert(QLIST_NEXT(child_b, next) == NULL); 1280 1281 /* Start the evil write request */ 1282 detach_by_parent_data = (struct detach_by_parent_data) { 1283 .parent_b = parent_b, 1284 .child_b = child_b, 1285 .c = c, 1286 .by_parent_cb = by_parent_cb, 1287 .detach_on_drain = true, 1288 }; 1289 acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL); 1290 g_assert(acb != NULL); 1291 1292 /* Drain and check the expected result */ 1293 bdrv_drained_begin(parent_b); 1294 bdrv_drained_begin(a); 1295 bdrv_drained_begin(b); 1296 bdrv_drained_begin(c); 1297 1298 g_assert(detach_by_parent_data.child_c != NULL); 1299 1300 g_assert_cmpint(parent_a->refcnt, ==, 1); 1301 g_assert_cmpint(parent_b->refcnt, ==, 1); 1302 g_assert_cmpint(a->refcnt, ==, 3); 1303 g_assert_cmpint(b->refcnt, ==, 1); 1304 g_assert_cmpint(c->refcnt, ==, 2); 1305 1306 g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c); 1307 g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a); 1308 g_assert(QLIST_NEXT(child_a, next) == NULL); 1309 1310 g_assert_cmpint(parent_a->quiesce_counter, ==, 1); 1311 g_assert_cmpint(parent_b->quiesce_counter, ==, 3); 1312 g_assert_cmpint(a->quiesce_counter, ==, 1); 1313 g_assert_cmpint(b->quiesce_counter, ==, 1); 1314 g_assert_cmpint(c->quiesce_counter, ==, 1); 1315 1316 bdrv_drained_end(parent_b); 1317 bdrv_drained_end(a); 1318 bdrv_drained_end(b); 1319 bdrv_drained_end(c); 1320 1321 bdrv_unref(parent_b); 1322 blk_unref(blk); 1323 1324 g_assert_cmpint(a->refcnt, ==, 1); 1325 g_assert_cmpint(b->refcnt, ==, 1); 1326 g_assert_cmpint(c->refcnt, ==, 1); 1327 bdrv_unref(a); 1328 bdrv_unref(b); 1329 bdrv_unref(c); 1330 } 1331 1332 static void test_detach_by_parent_cb(void) 1333 { 1334 test_detach_indirect(true); 1335 } 1336 1337 static void test_detach_by_driver_cb(void) 1338 { 1339 test_detach_indirect(false); 1340 } 1341 1342 static void test_append_to_drained(void) 1343 { 1344 BlockBackend *blk; 1345 BlockDriverState *base, *overlay; 1346 BDRVTestState *base_s, *overlay_s; 1347 1348 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 1349 base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort); 1350 base_s = base->opaque; 1351 blk_insert_bs(blk, base, &error_abort); 1352 1353 overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR, 1354 &error_abort); 1355 overlay_s = overlay->opaque; 1356 1357 do_drain_begin(BDRV_DRAIN, base); 1358 g_assert_cmpint(base->quiesce_counter, ==, 1); 1359 g_assert_cmpint(base_s->drain_count, ==, 1); 1360 g_assert_cmpint(base->in_flight, ==, 0); 1361 1362 bdrv_append(overlay, base, &error_abort); 1363 g_assert_cmpint(base->in_flight, ==, 0); 1364 g_assert_cmpint(overlay->in_flight, ==, 0); 1365 1366 g_assert_cmpint(base->quiesce_counter, ==, 1); 1367 g_assert_cmpint(base_s->drain_count, ==, 1); 1368 g_assert_cmpint(overlay->quiesce_counter, ==, 1); 1369 g_assert_cmpint(overlay_s->drain_count, ==, 1); 1370 1371 do_drain_end(BDRV_DRAIN, base); 1372 1373 g_assert_cmpint(base->quiesce_counter, ==, 0); 1374 g_assert_cmpint(base_s->drain_count, ==, 0); 1375 g_assert_cmpint(overlay->quiesce_counter, ==, 0); 1376 g_assert_cmpint(overlay_s->drain_count, ==, 0); 1377 1378 bdrv_unref(overlay); 1379 bdrv_unref(base); 1380 blk_unref(blk); 1381 } 1382 1383 static void test_set_aio_context(void) 1384 { 1385 BlockDriverState *bs; 1386 IOThread *a = iothread_new(); 1387 IOThread *b = iothread_new(); 1388 AioContext *ctx_a = iothread_get_aio_context(a); 1389 AioContext *ctx_b = iothread_get_aio_context(b); 1390 1391 bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR, 1392 &error_abort); 1393 1394 bdrv_drained_begin(bs); 1395 bdrv_try_change_aio_context(bs, ctx_a, NULL, &error_abort); 1396 1397 aio_context_acquire(ctx_a); 1398 bdrv_drained_end(bs); 1399 1400 bdrv_drained_begin(bs); 1401 bdrv_try_change_aio_context(bs, ctx_b, NULL, &error_abort); 1402 aio_context_release(ctx_a); 1403 aio_context_acquire(ctx_b); 1404 bdrv_try_change_aio_context(bs, qemu_get_aio_context(), NULL, &error_abort); 1405 aio_context_release(ctx_b); 1406 bdrv_drained_end(bs); 1407 1408 bdrv_unref(bs); 1409 iothread_join(a); 1410 iothread_join(b); 1411 } 1412 1413 1414 typedef struct TestDropBackingBlockJob { 1415 BlockJob common; 1416 bool should_complete; 1417 bool *did_complete; 1418 BlockDriverState *detach_also; 1419 BlockDriverState *bs; 1420 } TestDropBackingBlockJob; 1421 1422 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp) 1423 { 1424 TestDropBackingBlockJob *s = 1425 container_of(job, TestDropBackingBlockJob, common.job); 1426 1427 while (!s->should_complete) { 1428 job_sleep_ns(job, 0); 1429 } 1430 1431 return 0; 1432 } 1433 1434 static void test_drop_backing_job_commit(Job *job) 1435 { 1436 TestDropBackingBlockJob *s = 1437 container_of(job, TestDropBackingBlockJob, common.job); 1438 1439 bdrv_set_backing_hd(s->bs, NULL, &error_abort); 1440 bdrv_set_backing_hd(s->detach_also, NULL, &error_abort); 1441 1442 *s->did_complete = true; 1443 } 1444 1445 static const BlockJobDriver test_drop_backing_job_driver = { 1446 .job_driver = { 1447 .instance_size = sizeof(TestDropBackingBlockJob), 1448 .free = block_job_free, 1449 .user_resume = block_job_user_resume, 1450 .run = test_drop_backing_job_run, 1451 .commit = test_drop_backing_job_commit, 1452 } 1453 }; 1454 1455 /** 1456 * Creates a child node with three parent nodes on it, and then runs a 1457 * block job on the final one, parent-node-2. 1458 * 1459 * The job is then asked to complete before a section where the child 1460 * is drained. 1461 * 1462 * Ending this section will undrain the child's parents, first 1463 * parent-node-2, then parent-node-1, then parent-node-0 -- the parent 1464 * list is in reverse order of how they were added. Ending the drain 1465 * on parent-node-2 will resume the job, thus completing it and 1466 * scheduling job_exit(). 1467 * 1468 * Ending the drain on parent-node-1 will poll the AioContext, which 1469 * lets job_exit() and thus test_drop_backing_job_commit() run. That 1470 * function first removes the child as parent-node-2's backing file. 1471 * 1472 * In old (and buggy) implementations, there are two problems with 1473 * that: 1474 * (A) bdrv_drain_invoke() polls for every node that leaves the 1475 * drained section. This means that job_exit() is scheduled 1476 * before the child has left the drained section. Its 1477 * quiesce_counter is therefore still 1 when it is removed from 1478 * parent-node-2. 1479 * 1480 * (B) bdrv_replace_child_noperm() calls drained_end() on the old 1481 * child's parents as many times as the child is quiesced. This 1482 * means it will call drained_end() on parent-node-2 once. 1483 * Because parent-node-2 is no longer quiesced at this point, this 1484 * will fail. 1485 * 1486 * bdrv_replace_child_noperm() therefore must call drained_end() on 1487 * the parent only if it really is still drained because the child is 1488 * drained. 1489 * 1490 * If removing child from parent-node-2 was successful (as it should 1491 * be), test_drop_backing_job_commit() will then also remove the child 1492 * from parent-node-0. 1493 * 1494 * With an old version of our drain infrastructure ((A) above), that 1495 * resulted in the following flow: 1496 * 1497 * 1. child attempts to leave its drained section. The call recurses 1498 * to its parents. 1499 * 1500 * 2. parent-node-2 leaves the drained section. Polling in 1501 * bdrv_drain_invoke() will schedule job_exit(). 1502 * 1503 * 3. parent-node-1 leaves the drained section. Polling in 1504 * bdrv_drain_invoke() will run job_exit(), thus disconnecting 1505 * parent-node-0 from the child node. 1506 * 1507 * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to 1508 * iterate over the parents. Thus, it now accesses the BdrvChild 1509 * object that used to connect parent-node-0 and the child node. 1510 * However, that object no longer exists, so it accesses a dangling 1511 * pointer. 1512 * 1513 * The solution is to only poll once when running a bdrv_drained_end() 1514 * operation, specifically at the end when all drained_end() 1515 * operations for all involved nodes have been scheduled. 1516 * Note that this also solves (A) above, thus hiding (B). 1517 */ 1518 static void test_blockjob_commit_by_drained_end(void) 1519 { 1520 BlockDriverState *bs_child, *bs_parents[3]; 1521 TestDropBackingBlockJob *job; 1522 bool job_has_completed = false; 1523 int i; 1524 1525 bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR, 1526 &error_abort); 1527 1528 for (i = 0; i < 3; i++) { 1529 char name[32]; 1530 snprintf(name, sizeof(name), "parent-node-%i", i); 1531 bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR, 1532 &error_abort); 1533 bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort); 1534 } 1535 1536 job = block_job_create("job", &test_drop_backing_job_driver, NULL, 1537 bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL, 1538 &error_abort); 1539 job->bs = bs_parents[2]; 1540 1541 job->detach_also = bs_parents[0]; 1542 job->did_complete = &job_has_completed; 1543 1544 job_start(&job->common.job); 1545 1546 job->should_complete = true; 1547 bdrv_drained_begin(bs_child); 1548 g_assert(!job_has_completed); 1549 bdrv_drained_end(bs_child); 1550 aio_poll(qemu_get_aio_context(), false); 1551 g_assert(job_has_completed); 1552 1553 bdrv_unref(bs_parents[0]); 1554 bdrv_unref(bs_parents[1]); 1555 bdrv_unref(bs_parents[2]); 1556 bdrv_unref(bs_child); 1557 } 1558 1559 1560 typedef struct TestSimpleBlockJob { 1561 BlockJob common; 1562 bool should_complete; 1563 bool *did_complete; 1564 } TestSimpleBlockJob; 1565 1566 static int coroutine_fn test_simple_job_run(Job *job, Error **errp) 1567 { 1568 TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job); 1569 1570 while (!s->should_complete) { 1571 job_sleep_ns(job, 0); 1572 } 1573 1574 return 0; 1575 } 1576 1577 static void test_simple_job_clean(Job *job) 1578 { 1579 TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job); 1580 *s->did_complete = true; 1581 } 1582 1583 static const BlockJobDriver test_simple_job_driver = { 1584 .job_driver = { 1585 .instance_size = sizeof(TestSimpleBlockJob), 1586 .free = block_job_free, 1587 .user_resume = block_job_user_resume, 1588 .run = test_simple_job_run, 1589 .clean = test_simple_job_clean, 1590 }, 1591 }; 1592 1593 static int drop_intermediate_poll_update_filename(BdrvChild *child, 1594 BlockDriverState *new_base, 1595 const char *filename, 1596 Error **errp) 1597 { 1598 /* 1599 * We are free to poll here, which may change the block graph, if 1600 * it is not drained. 1601 */ 1602 1603 /* If the job is not drained: Complete it, schedule job_exit() */ 1604 aio_poll(qemu_get_current_aio_context(), false); 1605 /* If the job is not drained: Run job_exit(), finish the job */ 1606 aio_poll(qemu_get_current_aio_context(), false); 1607 1608 return 0; 1609 } 1610 1611 /** 1612 * Test a poll in the midst of bdrv_drop_intermediate(). 1613 * 1614 * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(), 1615 * which can yield or poll. This may lead to graph changes, unless 1616 * the whole subtree in question is drained. 1617 * 1618 * We test this on the following graph: 1619 * 1620 * Job 1621 * 1622 * | 1623 * job-node 1624 * | 1625 * v 1626 * 1627 * job-node 1628 * 1629 * | 1630 * backing 1631 * | 1632 * v 1633 * 1634 * node-2 --chain--> node-1 --chain--> node-0 1635 * 1636 * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0). 1637 * 1638 * This first updates node-2's backing filename by invoking 1639 * drop_intermediate_poll_update_filename(), which polls twice. This 1640 * causes the job to finish, which in turns causes the job-node to be 1641 * deleted. 1642 * 1643 * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it 1644 * already has a pointer to the BdrvChild edge between job-node and 1645 * node-1. When it tries to handle that edge, we probably get a 1646 * segmentation fault because the object no longer exists. 1647 * 1648 * 1649 * The solution is for bdrv_drop_intermediate() to drain top's 1650 * subtree. This prevents graph changes from happening just because 1651 * BdrvChildClass.update_filename() yields or polls. Thus, the block 1652 * job is paused during that drained section and must finish before or 1653 * after. 1654 * 1655 * (In addition, bdrv_replace_child() must keep the job paused.) 1656 */ 1657 static void test_drop_intermediate_poll(void) 1658 { 1659 static BdrvChildClass chain_child_class; 1660 BlockDriverState *chain[3]; 1661 TestSimpleBlockJob *job; 1662 BlockDriverState *job_node; 1663 bool job_has_completed = false; 1664 int i; 1665 int ret; 1666 1667 chain_child_class = child_of_bds; 1668 chain_child_class.update_filename = drop_intermediate_poll_update_filename; 1669 1670 for (i = 0; i < 3; i++) { 1671 char name[32]; 1672 snprintf(name, 32, "node-%i", i); 1673 1674 chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort); 1675 } 1676 1677 job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR, 1678 &error_abort); 1679 bdrv_set_backing_hd(job_node, chain[1], &error_abort); 1680 1681 /* 1682 * Establish the chain last, so the chain links are the first 1683 * elements in the BDS.parents lists 1684 */ 1685 for (i = 0; i < 3; i++) { 1686 if (i) { 1687 /* Takes the reference to chain[i - 1] */ 1688 bdrv_attach_child(chain[i], chain[i - 1], "chain", 1689 &chain_child_class, BDRV_CHILD_COW, &error_abort); 1690 } 1691 } 1692 1693 job = block_job_create("job", &test_simple_job_driver, NULL, job_node, 1694 0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort); 1695 1696 /* The job has a reference now */ 1697 bdrv_unref(job_node); 1698 1699 job->did_complete = &job_has_completed; 1700 1701 job_start(&job->common.job); 1702 job->should_complete = true; 1703 1704 g_assert(!job_has_completed); 1705 ret = bdrv_drop_intermediate(chain[1], chain[0], NULL); 1706 aio_poll(qemu_get_aio_context(), false); 1707 g_assert(ret == 0); 1708 g_assert(job_has_completed); 1709 1710 bdrv_unref(chain[2]); 1711 } 1712 1713 1714 typedef struct BDRVReplaceTestState { 1715 bool setup_completed; 1716 bool was_drained; 1717 bool was_undrained; 1718 bool has_read; 1719 1720 int drain_count; 1721 1722 bool yield_before_read; 1723 Coroutine *io_co; 1724 Coroutine *drain_co; 1725 } BDRVReplaceTestState; 1726 1727 static void bdrv_replace_test_close(BlockDriverState *bs) 1728 { 1729 } 1730 1731 /** 1732 * If @bs has a backing file: 1733 * Yield if .yield_before_read is true (and wait for drain_begin to 1734 * wake us up). 1735 * Forward the read to bs->backing. Set .has_read to true. 1736 * If drain_begin has woken us, wake it in turn. 1737 * 1738 * Otherwise: 1739 * Set .has_read to true and return success. 1740 */ 1741 static int coroutine_fn GRAPH_RDLOCK 1742 bdrv_replace_test_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes, 1743 QEMUIOVector *qiov, BdrvRequestFlags flags) 1744 { 1745 BDRVReplaceTestState *s = bs->opaque; 1746 1747 if (bs->backing) { 1748 int ret; 1749 1750 g_assert(!s->drain_count); 1751 1752 s->io_co = qemu_coroutine_self(); 1753 if (s->yield_before_read) { 1754 s->yield_before_read = false; 1755 qemu_coroutine_yield(); 1756 } 1757 s->io_co = NULL; 1758 1759 ret = bdrv_co_preadv(bs->backing, offset, bytes, qiov, 0); 1760 s->has_read = true; 1761 1762 /* Wake up drain_co if it runs */ 1763 if (s->drain_co) { 1764 aio_co_wake(s->drain_co); 1765 } 1766 1767 return ret; 1768 } 1769 1770 s->has_read = true; 1771 return 0; 1772 } 1773 1774 static void coroutine_fn bdrv_replace_test_drain_co(void *opaque) 1775 { 1776 BlockDriverState *bs = opaque; 1777 BDRVReplaceTestState *s = bs->opaque; 1778 1779 /* Keep waking io_co up until it is done */ 1780 while (s->io_co) { 1781 aio_co_wake(s->io_co); 1782 s->io_co = NULL; 1783 qemu_coroutine_yield(); 1784 } 1785 s->drain_co = NULL; 1786 bdrv_dec_in_flight(bs); 1787 } 1788 1789 /** 1790 * If .drain_count is 0, wake up .io_co if there is one; and set 1791 * .was_drained. 1792 * Increment .drain_count. 1793 */ 1794 static void bdrv_replace_test_drain_begin(BlockDriverState *bs) 1795 { 1796 BDRVReplaceTestState *s = bs->opaque; 1797 1798 if (!s->setup_completed) { 1799 return; 1800 } 1801 1802 if (!s->drain_count) { 1803 s->drain_co = qemu_coroutine_create(bdrv_replace_test_drain_co, bs); 1804 bdrv_inc_in_flight(bs); 1805 aio_co_enter(bdrv_get_aio_context(bs), s->drain_co); 1806 s->was_drained = true; 1807 } 1808 s->drain_count++; 1809 } 1810 1811 static void coroutine_fn bdrv_replace_test_read_entry(void *opaque) 1812 { 1813 BlockDriverState *bs = opaque; 1814 char data; 1815 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1); 1816 int ret; 1817 1818 /* Queue a read request post-drain */ 1819 bdrv_graph_co_rdlock(); 1820 ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0); 1821 bdrv_graph_co_rdunlock(); 1822 1823 g_assert(ret >= 0); 1824 bdrv_dec_in_flight(bs); 1825 } 1826 1827 /** 1828 * Reduce .drain_count, set .was_undrained once it reaches 0. 1829 * If .drain_count reaches 0 and the node has a backing file, issue a 1830 * read request. 1831 */ 1832 static void bdrv_replace_test_drain_end(BlockDriverState *bs) 1833 { 1834 BDRVReplaceTestState *s = bs->opaque; 1835 1836 if (!s->setup_completed) { 1837 return; 1838 } 1839 1840 g_assert(s->drain_count > 0); 1841 if (!--s->drain_count) { 1842 s->was_undrained = true; 1843 1844 if (bs->backing) { 1845 Coroutine *co = qemu_coroutine_create(bdrv_replace_test_read_entry, 1846 bs); 1847 bdrv_inc_in_flight(bs); 1848 aio_co_enter(bdrv_get_aio_context(bs), co); 1849 } 1850 } 1851 } 1852 1853 static BlockDriver bdrv_replace_test = { 1854 .format_name = "replace_test", 1855 .instance_size = sizeof(BDRVReplaceTestState), 1856 .supports_backing = true, 1857 1858 .bdrv_close = bdrv_replace_test_close, 1859 .bdrv_co_preadv = bdrv_replace_test_co_preadv, 1860 1861 .bdrv_drain_begin = bdrv_replace_test_drain_begin, 1862 .bdrv_drain_end = bdrv_replace_test_drain_end, 1863 1864 .bdrv_child_perm = bdrv_default_perms, 1865 }; 1866 1867 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque) 1868 { 1869 int ret; 1870 char data; 1871 1872 ret = blk_co_pread(opaque, 0, 1, &data, 0); 1873 g_assert(ret >= 0); 1874 } 1875 1876 /** 1877 * We test two things: 1878 * (1) bdrv_replace_child_noperm() must not undrain the parent if both 1879 * children are drained. 1880 * (2) bdrv_replace_child_noperm() must never flush I/O requests to a 1881 * drained child. If the old child is drained, it must flush I/O 1882 * requests after the new one has been attached. If the new child 1883 * is drained, it must flush I/O requests before the old one is 1884 * detached. 1885 * 1886 * To do so, we create one parent node and two child nodes; then 1887 * attach one of the children (old_child_bs) to the parent, then 1888 * drain both old_child_bs and new_child_bs according to 1889 * old_drain_count and new_drain_count, respectively, and finally 1890 * we invoke bdrv_replace_node() to replace old_child_bs by 1891 * new_child_bs. 1892 * 1893 * The test block driver we use here (bdrv_replace_test) has a read 1894 * function that: 1895 * - For the parent node, can optionally yield, and then forwards the 1896 * read to bdrv_preadv(), 1897 * - For the child node, just returns immediately. 1898 * 1899 * If the read yields, the drain_begin function will wake it up. 1900 * 1901 * The drain_end function issues a read on the parent once it is fully 1902 * undrained (which simulates requests starting to come in again). 1903 */ 1904 static void do_test_replace_child_mid_drain(int old_drain_count, 1905 int new_drain_count) 1906 { 1907 BlockBackend *parent_blk; 1908 BlockDriverState *parent_bs; 1909 BlockDriverState *old_child_bs, *new_child_bs; 1910 BDRVReplaceTestState *parent_s; 1911 BDRVReplaceTestState *old_child_s, *new_child_s; 1912 Coroutine *io_co; 1913 int i; 1914 1915 parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0, 1916 &error_abort); 1917 parent_s = parent_bs->opaque; 1918 1919 parent_blk = blk_new(qemu_get_aio_context(), 1920 BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL); 1921 blk_insert_bs(parent_blk, parent_bs, &error_abort); 1922 1923 old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0, 1924 &error_abort); 1925 new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0, 1926 &error_abort); 1927 old_child_s = old_child_bs->opaque; 1928 new_child_s = new_child_bs->opaque; 1929 1930 /* So that we can read something */ 1931 parent_bs->total_sectors = 1; 1932 old_child_bs->total_sectors = 1; 1933 new_child_bs->total_sectors = 1; 1934 1935 bdrv_ref(old_child_bs); 1936 bdrv_attach_child(parent_bs, old_child_bs, "child", &child_of_bds, 1937 BDRV_CHILD_COW, &error_abort); 1938 parent_s->setup_completed = true; 1939 1940 for (i = 0; i < old_drain_count; i++) { 1941 bdrv_drained_begin(old_child_bs); 1942 } 1943 for (i = 0; i < new_drain_count; i++) { 1944 bdrv_drained_begin(new_child_bs); 1945 } 1946 1947 if (!old_drain_count) { 1948 /* 1949 * Start a read operation that will yield, so it will not 1950 * complete before the node is drained. 1951 */ 1952 parent_s->yield_before_read = true; 1953 io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co, 1954 parent_blk); 1955 qemu_coroutine_enter(io_co); 1956 } 1957 1958 /* If we have started a read operation, it should have yielded */ 1959 g_assert(!parent_s->has_read); 1960 1961 /* Reset drained status so we can see what bdrv_replace_node() does */ 1962 parent_s->was_drained = false; 1963 parent_s->was_undrained = false; 1964 1965 g_assert(parent_bs->quiesce_counter == old_drain_count); 1966 bdrv_replace_node(old_child_bs, new_child_bs, &error_abort); 1967 g_assert(parent_bs->quiesce_counter == new_drain_count); 1968 1969 if (!old_drain_count && !new_drain_count) { 1970 /* 1971 * From undrained to undrained drains and undrains the parent, 1972 * because bdrv_replace_node() contains a drained section for 1973 * @old_child_bs. 1974 */ 1975 g_assert(parent_s->was_drained && parent_s->was_undrained); 1976 } else if (!old_drain_count && new_drain_count) { 1977 /* 1978 * From undrained to drained should drain the parent and keep 1979 * it that way. 1980 */ 1981 g_assert(parent_s->was_drained && !parent_s->was_undrained); 1982 } else if (old_drain_count && !new_drain_count) { 1983 /* 1984 * From drained to undrained should undrain the parent and 1985 * keep it that way. 1986 */ 1987 g_assert(!parent_s->was_drained && parent_s->was_undrained); 1988 } else /* if (old_drain_count && new_drain_count) */ { 1989 /* 1990 * From drained to drained must not undrain the parent at any 1991 * point 1992 */ 1993 g_assert(!parent_s->was_drained && !parent_s->was_undrained); 1994 } 1995 1996 if (!old_drain_count || !new_drain_count) { 1997 /* 1998 * If !old_drain_count, we have started a read request before 1999 * bdrv_replace_node(). If !new_drain_count, the parent must 2000 * have been undrained at some point, and 2001 * bdrv_replace_test_co_drain_end() starts a read request 2002 * then. 2003 */ 2004 g_assert(parent_s->has_read); 2005 } else { 2006 /* 2007 * If the parent was never undrained, there is no way to start 2008 * a read request. 2009 */ 2010 g_assert(!parent_s->has_read); 2011 } 2012 2013 /* A drained child must have not received any request */ 2014 g_assert(!(old_drain_count && old_child_s->has_read)); 2015 g_assert(!(new_drain_count && new_child_s->has_read)); 2016 2017 for (i = 0; i < new_drain_count; i++) { 2018 bdrv_drained_end(new_child_bs); 2019 } 2020 for (i = 0; i < old_drain_count; i++) { 2021 bdrv_drained_end(old_child_bs); 2022 } 2023 2024 /* 2025 * By now, bdrv_replace_test_co_drain_end() must have been called 2026 * at some point while the new child was attached to the parent. 2027 */ 2028 g_assert(parent_s->has_read); 2029 g_assert(new_child_s->has_read); 2030 2031 blk_unref(parent_blk); 2032 bdrv_unref(parent_bs); 2033 bdrv_unref(old_child_bs); 2034 bdrv_unref(new_child_bs); 2035 } 2036 2037 static void test_replace_child_mid_drain(void) 2038 { 2039 int old_drain_count, new_drain_count; 2040 2041 for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) { 2042 for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) { 2043 do_test_replace_child_mid_drain(old_drain_count, new_drain_count); 2044 } 2045 } 2046 } 2047 2048 int main(int argc, char **argv) 2049 { 2050 int ret; 2051 2052 bdrv_init(); 2053 qemu_init_main_loop(&error_abort); 2054 2055 g_test_init(&argc, &argv, NULL); 2056 qemu_event_init(&done_event, false); 2057 2058 g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all); 2059 g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain); 2060 2061 g_test_add_func("/bdrv-drain/driver-cb/co/drain_all", 2062 test_drv_cb_co_drain_all); 2063 g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain); 2064 2065 g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all); 2066 g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain); 2067 2068 g_test_add_func("/bdrv-drain/quiesce/co/drain_all", 2069 test_quiesce_co_drain_all); 2070 g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain); 2071 2072 g_test_add_func("/bdrv-drain/nested", test_nested); 2073 2074 g_test_add_func("/bdrv-drain/graph-change/drain_all", 2075 test_graph_change_drain_all); 2076 2077 g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all); 2078 g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain); 2079 2080 g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all); 2081 g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain); 2082 2083 g_test_add_func("/bdrv-drain/blockjob/error/drain_all", 2084 test_blockjob_error_drain_all); 2085 g_test_add_func("/bdrv-drain/blockjob/error/drain", 2086 test_blockjob_error_drain); 2087 2088 g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all", 2089 test_blockjob_iothread_drain_all); 2090 g_test_add_func("/bdrv-drain/blockjob/iothread/drain", 2091 test_blockjob_iothread_drain); 2092 2093 g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all", 2094 test_blockjob_iothread_error_drain_all); 2095 g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain", 2096 test_blockjob_iothread_error_drain); 2097 2098 g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain); 2099 g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all); 2100 g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain); 2101 g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb); 2102 g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb); 2103 2104 g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained); 2105 2106 g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context); 2107 2108 g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end", 2109 test_blockjob_commit_by_drained_end); 2110 2111 g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll", 2112 test_drop_intermediate_poll); 2113 2114 g_test_add_func("/bdrv-drain/replace_child/mid-drain", 2115 test_replace_child_mid_drain); 2116 2117 ret = g_test_run(); 2118 qemu_event_destroy(&done_event); 2119 return ret; 2120 } 2121