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 "system/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_co_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_co_change_backing_file = bdrv_test_co_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 do_drain_begin(drain_type, bs); 183 } 184 185 static BlockBackend * no_coroutine_fn test_setup(void) 186 { 187 BlockBackend *blk; 188 BlockDriverState *bs, *backing; 189 190 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 191 bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR, 192 &error_abort); 193 blk_insert_bs(blk, bs, &error_abort); 194 195 backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort); 196 bdrv_set_backing_hd(bs, backing, &error_abort); 197 198 bdrv_unref(backing); 199 bdrv_unref(bs); 200 201 return blk; 202 } 203 204 static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs) 205 { 206 do_drain_end(drain_type, bs); 207 } 208 209 /* 210 * Locking the block graph would be a bit cumbersome here because this function 211 * is called both in coroutine and non-coroutine context. We know this is a test 212 * and nothing else is running, so don't bother with TSA. 213 */ 214 static void coroutine_mixed_fn TSA_NO_TSA 215 test_drv_cb_common(BlockBackend *blk, enum drain_type drain_type, 216 bool recursive) 217 { 218 BlockDriverState *bs = blk_bs(blk); 219 BlockDriverState *backing = bs->backing->bs; 220 BDRVTestState *s, *backing_s; 221 BlockAIOCB *acb; 222 int aio_ret; 223 224 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0); 225 226 s = bs->opaque; 227 backing_s = backing->opaque; 228 229 /* Simple bdrv_drain_all_begin/end pair, check that CBs are called */ 230 g_assert_cmpint(s->drain_count, ==, 0); 231 g_assert_cmpint(backing_s->drain_count, ==, 0); 232 233 do_drain_begin(drain_type, bs); 234 235 g_assert_cmpint(s->drain_count, ==, 1); 236 g_assert_cmpint(backing_s->drain_count, ==, !!recursive); 237 238 do_drain_end(drain_type, bs); 239 240 g_assert_cmpint(s->drain_count, ==, 0); 241 g_assert_cmpint(backing_s->drain_count, ==, 0); 242 243 /* Now do the same while a request is pending */ 244 aio_ret = -EINPROGRESS; 245 acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret); 246 g_assert(acb != NULL); 247 g_assert_cmpint(aio_ret, ==, -EINPROGRESS); 248 249 g_assert_cmpint(s->drain_count, ==, 0); 250 g_assert_cmpint(backing_s->drain_count, ==, 0); 251 252 do_drain_begin(drain_type, bs); 253 254 g_assert_cmpint(aio_ret, ==, 0); 255 g_assert_cmpint(s->drain_count, ==, 1); 256 g_assert_cmpint(backing_s->drain_count, ==, !!recursive); 257 258 do_drain_end(drain_type, bs); 259 260 g_assert_cmpint(s->drain_count, ==, 0); 261 g_assert_cmpint(backing_s->drain_count, ==, 0); 262 } 263 264 static void test_drv_cb_drain_all(void) 265 { 266 BlockBackend *blk = test_setup(); 267 test_drv_cb_common(blk, BDRV_DRAIN_ALL, true); 268 blk_unref(blk); 269 } 270 271 static void test_drv_cb_drain(void) 272 { 273 BlockBackend *blk = test_setup(); 274 test_drv_cb_common(blk, BDRV_DRAIN, false); 275 blk_unref(blk); 276 } 277 278 static void coroutine_fn test_drv_cb_co_drain_all_entry(void) 279 { 280 BlockBackend *blk = blk_all_next(NULL); 281 test_drv_cb_common(blk, BDRV_DRAIN_ALL, true); 282 } 283 284 static void test_drv_cb_co_drain_all(void) 285 { 286 BlockBackend *blk = test_setup(); 287 call_in_coroutine(test_drv_cb_co_drain_all_entry); 288 blk_unref(blk); 289 } 290 291 static void coroutine_fn test_drv_cb_co_drain_entry(void) 292 { 293 BlockBackend *blk = blk_all_next(NULL); 294 test_drv_cb_common(blk, BDRV_DRAIN, false); 295 } 296 297 static void test_drv_cb_co_drain(void) 298 { 299 BlockBackend *blk = test_setup(); 300 call_in_coroutine(test_drv_cb_co_drain_entry); 301 blk_unref(blk); 302 } 303 304 /* 305 * Locking the block graph would be a bit cumbersome here because this function 306 * is called both in coroutine and non-coroutine context. We know this is a test 307 * and nothing else is running, so don't bother with TSA. 308 */ 309 static void coroutine_mixed_fn TSA_NO_TSA 310 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 bdrv_flush(data->bs); 512 bdrv_dec_in_flight(data->bs); /* incremented by test_iothread_common() */ 513 } 514 515 /* 516 * Starts an AIO request on a BDS that runs in the AioContext of iothread 1. 517 * The request involves a BH on iothread 2 before it can complete. 518 * 519 * @drain_thread = 0 means that do_drain_begin/end are called from the main 520 * thread, @drain_thread = 1 means that they are called from iothread 1. Drain 521 * for this BDS cannot be called from iothread 2 because only the main thread 522 * may do cross-AioContext polling. 523 */ 524 static void test_iothread_common(enum drain_type drain_type, int drain_thread) 525 { 526 BlockBackend *blk; 527 BlockDriverState *bs; 528 BDRVTestState *s; 529 BlockAIOCB *acb; 530 Coroutine *co; 531 int aio_ret; 532 struct test_iothread_data data; 533 534 IOThread *a = iothread_new(); 535 IOThread *b = iothread_new(); 536 AioContext *ctx_a = iothread_get_aio_context(a); 537 AioContext *ctx_b = iothread_get_aio_context(b); 538 539 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0); 540 541 /* bdrv_drain_all() may only be called from the main loop thread */ 542 if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) { 543 goto out; 544 } 545 546 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 547 bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR, 548 &error_abort); 549 s = bs->opaque; 550 blk_insert_bs(blk, bs, &error_abort); 551 blk_set_disable_request_queuing(blk, true); 552 553 blk_set_aio_context(blk, ctx_a, &error_abort); 554 555 s->bh_indirection_ctx = ctx_b; 556 557 aio_ret = -EINPROGRESS; 558 qemu_event_reset(&done_event); 559 560 if (drain_thread == 0) { 561 acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret); 562 } else { 563 acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret); 564 } 565 g_assert(acb != NULL); 566 g_assert_cmpint(aio_ret, ==, -EINPROGRESS); 567 568 data = (struct test_iothread_data) { 569 .bs = bs, 570 .drain_type = drain_type, 571 .aio_ret = &aio_ret, 572 }; 573 574 switch (drain_thread) { 575 case 0: 576 /* 577 * Increment in_flight so that do_drain_begin() waits for 578 * test_iothread_main_thread_bh(). This prevents the race between 579 * test_iothread_main_thread_bh() in IOThread a and do_drain_begin() in 580 * this thread. test_iothread_main_thread_bh() decrements in_flight. 581 */ 582 bdrv_inc_in_flight(bs); 583 aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data); 584 585 /* The request is running on the IOThread a. Draining its block device 586 * will make sure that it has completed as far as the BDS is concerned, 587 * but the drain in this thread can continue immediately after 588 * bdrv_dec_in_flight() and aio_ret might be assigned only slightly 589 * later. */ 590 do_drain_begin(drain_type, bs); 591 g_assert_cmpint(bs->in_flight, ==, 0); 592 593 qemu_event_wait(&done_event); 594 595 g_assert_cmpint(aio_ret, ==, 0); 596 do_drain_end(drain_type, bs); 597 break; 598 case 1: 599 co = qemu_coroutine_create(test_iothread_drain_co_entry, &data); 600 aio_co_enter(ctx_a, co); 601 AIO_WAIT_WHILE_UNLOCKED(NULL, !data.co_done); 602 break; 603 default: 604 g_assert_not_reached(); 605 } 606 607 blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort); 608 609 bdrv_unref(bs); 610 blk_unref(blk); 611 612 out: 613 iothread_join(a); 614 iothread_join(b); 615 } 616 617 static void test_iothread_drain_all(void) 618 { 619 test_iothread_common(BDRV_DRAIN_ALL, 0); 620 test_iothread_common(BDRV_DRAIN_ALL, 1); 621 } 622 623 static void test_iothread_drain(void) 624 { 625 test_iothread_common(BDRV_DRAIN, 0); 626 test_iothread_common(BDRV_DRAIN, 1); 627 } 628 629 630 typedef struct TestBlockJob { 631 BlockJob common; 632 BlockDriverState *bs; 633 int run_ret; 634 int prepare_ret; 635 636 /* Accessed with atomics */ 637 bool running; 638 bool should_complete; 639 } TestBlockJob; 640 641 static int test_job_prepare(Job *job) 642 { 643 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 644 645 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */ 646 bdrv_flush(s->bs); 647 return s->prepare_ret; 648 } 649 650 static void test_job_commit(Job *job) 651 { 652 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 653 654 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */ 655 bdrv_flush(s->bs); 656 } 657 658 static void test_job_abort(Job *job) 659 { 660 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 661 662 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */ 663 bdrv_flush(s->bs); 664 } 665 666 static int coroutine_fn test_job_run(Job *job, Error **errp) 667 { 668 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 669 670 /* We are running the actual job code past the pause point in 671 * job_co_entry(). */ 672 qatomic_set(&s->running, true); 673 674 job_transition_to_ready(&s->common.job); 675 while (!qatomic_read(&s->should_complete)) { 676 /* Avoid job_sleep_ns() because it marks the job as !busy. We want to 677 * emulate some actual activity (probably some I/O) here so that drain 678 * has to wait for this activity to stop. */ 679 qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000); 680 681 job_pause_point(&s->common.job); 682 } 683 684 return s->run_ret; 685 } 686 687 static void test_job_complete(Job *job, Error **errp) 688 { 689 TestBlockJob *s = container_of(job, TestBlockJob, common.job); 690 qatomic_set(&s->should_complete, true); 691 } 692 693 BlockJobDriver test_job_driver = { 694 .job_driver = { 695 .instance_size = sizeof(TestBlockJob), 696 .free = block_job_free, 697 .user_resume = block_job_user_resume, 698 .run = test_job_run, 699 .complete = test_job_complete, 700 .prepare = test_job_prepare, 701 .commit = test_job_commit, 702 .abort = test_job_abort, 703 }, 704 }; 705 706 enum test_job_result { 707 TEST_JOB_SUCCESS, 708 TEST_JOB_FAIL_RUN, 709 TEST_JOB_FAIL_PREPARE, 710 }; 711 712 enum test_job_drain_node { 713 TEST_JOB_DRAIN_SRC, 714 TEST_JOB_DRAIN_SRC_CHILD, 715 }; 716 717 static void test_blockjob_common_drain_node(enum drain_type drain_type, 718 bool use_iothread, 719 enum test_job_result result, 720 enum test_job_drain_node drain_node) 721 { 722 BlockBackend *blk_src, *blk_target; 723 BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs; 724 BlockJob *job; 725 TestBlockJob *tjob; 726 IOThread *iothread = NULL; 727 int ret = -1; 728 729 src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR, 730 &error_abort); 731 src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing", 732 BDRV_O_RDWR, &error_abort); 733 src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay", 734 BDRV_O_RDWR, &error_abort); 735 736 bdrv_set_backing_hd(src_overlay, src, &error_abort); 737 bdrv_unref(src); 738 bdrv_set_backing_hd(src, src_backing, &error_abort); 739 bdrv_unref(src_backing); 740 741 blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 742 blk_insert_bs(blk_src, src_overlay, &error_abort); 743 744 switch (drain_node) { 745 case TEST_JOB_DRAIN_SRC: 746 drain_bs = src; 747 break; 748 case TEST_JOB_DRAIN_SRC_CHILD: 749 drain_bs = src_backing; 750 break; 751 default: 752 g_assert_not_reached(); 753 } 754 755 if (use_iothread) { 756 AioContext *ctx; 757 758 iothread = iothread_new(); 759 ctx = iothread_get_aio_context(iothread); 760 blk_set_aio_context(blk_src, ctx, &error_abort); 761 } 762 763 target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR, 764 &error_abort); 765 blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 766 blk_insert_bs(blk_target, target, &error_abort); 767 blk_set_allow_aio_context_change(blk_target, true); 768 769 tjob = block_job_create("job0", &test_job_driver, NULL, src, 770 0, BLK_PERM_ALL, 771 0, 0, NULL, NULL, &error_abort); 772 tjob->bs = src; 773 job = &tjob->common; 774 775 bdrv_drain_all_begin(); 776 bdrv_graph_wrlock(); 777 block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort); 778 bdrv_graph_wrunlock(); 779 bdrv_drain_all_end(); 780 781 switch (result) { 782 case TEST_JOB_SUCCESS: 783 break; 784 case TEST_JOB_FAIL_RUN: 785 tjob->run_ret = -EIO; 786 break; 787 case TEST_JOB_FAIL_PREPARE: 788 tjob->prepare_ret = -EIO; 789 break; 790 } 791 792 job_start(&job->job); 793 794 if (use_iothread) { 795 /* job_co_entry() is run in the I/O thread, wait for the actual job 796 * code to start (we don't want to catch the job in the pause point in 797 * job_co_entry(). */ 798 while (!qatomic_read(&tjob->running)) { 799 aio_poll(qemu_get_aio_context(), false); 800 } 801 } 802 803 WITH_JOB_LOCK_GUARD() { 804 g_assert_cmpint(job->job.pause_count, ==, 0); 805 g_assert_false(job->job.paused); 806 g_assert_true(qatomic_read(&tjob->running)); 807 g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */ 808 } 809 810 do_drain_begin_unlocked(drain_type, drain_bs); 811 812 WITH_JOB_LOCK_GUARD() { 813 if (drain_type == BDRV_DRAIN_ALL) { 814 /* bdrv_drain_all() drains both src and target */ 815 g_assert_cmpint(job->job.pause_count, ==, 2); 816 } else { 817 g_assert_cmpint(job->job.pause_count, ==, 1); 818 } 819 g_assert_true(job->job.paused); 820 g_assert_false(job->job.busy); /* The job is paused */ 821 } 822 823 do_drain_end_unlocked(drain_type, drain_bs); 824 825 if (use_iothread) { 826 /* 827 * Here we are waiting for the paused status to change, 828 * so don't bother protecting the read every time. 829 * 830 * paused is reset in the I/O thread, wait for it 831 */ 832 while (job_is_paused(&job->job)) { 833 aio_poll(qemu_get_aio_context(), false); 834 } 835 } 836 837 WITH_JOB_LOCK_GUARD() { 838 g_assert_cmpint(job->job.pause_count, ==, 0); 839 g_assert_false(job->job.paused); 840 g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */ 841 } 842 843 do_drain_begin_unlocked(drain_type, target); 844 845 WITH_JOB_LOCK_GUARD() { 846 if (drain_type == BDRV_DRAIN_ALL) { 847 /* bdrv_drain_all() drains both src and target */ 848 g_assert_cmpint(job->job.pause_count, ==, 2); 849 } else { 850 g_assert_cmpint(job->job.pause_count, ==, 1); 851 } 852 g_assert_true(job->job.paused); 853 g_assert_false(job->job.busy); /* The job is paused */ 854 } 855 856 do_drain_end_unlocked(drain_type, target); 857 858 if (use_iothread) { 859 /* 860 * Here we are waiting for the paused status to change, 861 * so don't bother protecting the read every time. 862 * 863 * paused is reset in the I/O thread, wait for it 864 */ 865 while (job_is_paused(&job->job)) { 866 aio_poll(qemu_get_aio_context(), false); 867 } 868 } 869 870 WITH_JOB_LOCK_GUARD() { 871 g_assert_cmpint(job->job.pause_count, ==, 0); 872 g_assert_false(job->job.paused); 873 g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */ 874 } 875 876 WITH_JOB_LOCK_GUARD() { 877 ret = job_complete_sync_locked(&job->job, &error_abort); 878 } 879 g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO)); 880 881 if (use_iothread) { 882 blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort); 883 assert(blk_get_aio_context(blk_target) == qemu_get_aio_context()); 884 } 885 886 blk_unref(blk_src); 887 blk_unref(blk_target); 888 bdrv_unref(src_overlay); 889 bdrv_unref(target); 890 891 if (iothread) { 892 iothread_join(iothread); 893 } 894 } 895 896 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread, 897 enum test_job_result result) 898 { 899 test_blockjob_common_drain_node(drain_type, use_iothread, result, 900 TEST_JOB_DRAIN_SRC); 901 test_blockjob_common_drain_node(drain_type, use_iothread, result, 902 TEST_JOB_DRAIN_SRC_CHILD); 903 } 904 905 static void test_blockjob_drain_all(void) 906 { 907 test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS); 908 } 909 910 static void test_blockjob_drain(void) 911 { 912 test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS); 913 } 914 915 static void test_blockjob_error_drain_all(void) 916 { 917 test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN); 918 test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE); 919 } 920 921 static void test_blockjob_error_drain(void) 922 { 923 test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN); 924 test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE); 925 } 926 927 static void test_blockjob_iothread_drain_all(void) 928 { 929 test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS); 930 } 931 932 static void test_blockjob_iothread_drain(void) 933 { 934 test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS); 935 } 936 937 static void test_blockjob_iothread_error_drain_all(void) 938 { 939 test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN); 940 test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE); 941 } 942 943 static void test_blockjob_iothread_error_drain(void) 944 { 945 test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN); 946 test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE); 947 } 948 949 950 typedef struct BDRVTestTopState { 951 BdrvChild *wait_child; 952 } BDRVTestTopState; 953 954 static void bdrv_test_top_close(BlockDriverState *bs) 955 { 956 BdrvChild *c, *next_c; 957 958 bdrv_drain_all_begin(); 959 bdrv_graph_wrlock(); 960 QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) { 961 bdrv_unref_child(bs, c); 962 } 963 bdrv_graph_wrunlock(); 964 bdrv_drain_all_end(); 965 } 966 967 static int coroutine_fn GRAPH_RDLOCK 968 bdrv_test_top_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes, 969 QEMUIOVector *qiov, BdrvRequestFlags flags) 970 { 971 BDRVTestTopState *tts = bs->opaque; 972 return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags); 973 } 974 975 static BlockDriver bdrv_test_top_driver = { 976 .format_name = "test_top_driver", 977 .instance_size = sizeof(BDRVTestTopState), 978 979 .bdrv_close = bdrv_test_top_close, 980 .bdrv_co_preadv = bdrv_test_top_co_preadv, 981 982 .bdrv_child_perm = bdrv_default_perms, 983 }; 984 985 typedef struct TestCoDeleteByDrainData { 986 BlockBackend *blk; 987 bool detach_instead_of_delete; 988 bool done; 989 } TestCoDeleteByDrainData; 990 991 static void coroutine_fn test_co_delete_by_drain(void *opaque) 992 { 993 TestCoDeleteByDrainData *dbdd = opaque; 994 BlockBackend *blk = dbdd->blk; 995 BlockDriverState *bs = blk_bs(blk); 996 BDRVTestTopState *tts = bs->opaque; 997 void *buffer = g_malloc(65536); 998 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536); 999 1000 /* Pretend some internal write operation from parent to child. 1001 * Important: We have to read from the child, not from the parent! 1002 * Draining works by first propagating it all up the tree to the 1003 * root and then waiting for drainage from root to the leaves 1004 * (protocol nodes). If we have a request waiting on the root, 1005 * everything will be drained before we go back down the tree, but 1006 * we do not want that. We want to be in the middle of draining 1007 * when this following requests returns. */ 1008 bdrv_graph_co_rdlock(); 1009 bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0); 1010 bdrv_graph_co_rdunlock(); 1011 1012 g_assert_cmpint(bs->refcnt, ==, 1); 1013 1014 if (!dbdd->detach_instead_of_delete) { 1015 blk_co_unref(blk); 1016 } else { 1017 BdrvChild *c, *next_c; 1018 bdrv_graph_co_rdlock(); 1019 QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) { 1020 bdrv_graph_co_rdunlock(); 1021 bdrv_drain_all_begin(); 1022 bdrv_co_unref_child(bs, c); 1023 bdrv_drain_all_end(); 1024 bdrv_graph_co_rdlock(); 1025 } 1026 bdrv_graph_co_rdunlock(); 1027 } 1028 1029 dbdd->done = true; 1030 g_free(buffer); 1031 } 1032 1033 /** 1034 * Test what happens when some BDS has some children, you drain one of 1035 * them and this results in the BDS being deleted. 1036 * 1037 * If @detach_instead_of_delete is set, the BDS is not going to be 1038 * deleted but will only detach all of its children. 1039 */ 1040 static void do_test_delete_by_drain(bool detach_instead_of_delete, 1041 enum drain_type drain_type) 1042 { 1043 BlockBackend *blk; 1044 BlockDriverState *bs, *child_bs, *null_bs; 1045 BDRVTestTopState *tts; 1046 TestCoDeleteByDrainData dbdd; 1047 Coroutine *co; 1048 1049 bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR, 1050 &error_abort); 1051 bs->total_sectors = 65536 >> BDRV_SECTOR_BITS; 1052 tts = bs->opaque; 1053 1054 null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL, 1055 &error_abort); 1056 bdrv_drain_all_begin(); 1057 bdrv_graph_wrlock(); 1058 bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, 1059 BDRV_CHILD_DATA, &error_abort); 1060 bdrv_graph_wrunlock(); 1061 bdrv_drain_all_end(); 1062 1063 /* This child will be the one to pass to requests through to, and 1064 * it will stall until a drain occurs */ 1065 child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR, 1066 &error_abort); 1067 child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS; 1068 /* Takes our reference to child_bs */ 1069 bdrv_drain_all_begin(); 1070 bdrv_graph_wrlock(); 1071 tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child", 1072 &child_of_bds, 1073 BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY, 1074 &error_abort); 1075 bdrv_graph_wrunlock(); 1076 bdrv_drain_all_end(); 1077 1078 /* This child is just there to be deleted 1079 * (for detach_instead_of_delete == true) */ 1080 null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL, 1081 &error_abort); 1082 bdrv_drain_all_begin(); 1083 bdrv_graph_wrlock(); 1084 bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA, 1085 &error_abort); 1086 bdrv_graph_wrunlock(); 1087 bdrv_drain_all_end(); 1088 1089 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 1090 blk_insert_bs(blk, bs, &error_abort); 1091 1092 /* Referenced by blk now */ 1093 bdrv_unref(bs); 1094 1095 g_assert_cmpint(bs->refcnt, ==, 1); 1096 g_assert_cmpint(child_bs->refcnt, ==, 1); 1097 g_assert_cmpint(null_bs->refcnt, ==, 1); 1098 1099 1100 dbdd = (TestCoDeleteByDrainData){ 1101 .blk = blk, 1102 .detach_instead_of_delete = detach_instead_of_delete, 1103 .done = false, 1104 }; 1105 co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd); 1106 qemu_coroutine_enter(co); 1107 1108 /* Drain the child while the read operation is still pending. 1109 * This should result in the operation finishing and 1110 * test_co_delete_by_drain() resuming. Thus, @bs will be deleted 1111 * and the coroutine will exit while this drain operation is still 1112 * in progress. */ 1113 switch (drain_type) { 1114 case BDRV_DRAIN: 1115 bdrv_ref(child_bs); 1116 bdrv_drain(child_bs); 1117 bdrv_unref(child_bs); 1118 break; 1119 case BDRV_DRAIN_ALL: 1120 bdrv_drain_all_begin(); 1121 bdrv_drain_all_end(); 1122 break; 1123 default: 1124 g_assert_not_reached(); 1125 } 1126 1127 while (!dbdd.done) { 1128 aio_poll(qemu_get_aio_context(), true); 1129 } 1130 1131 if (detach_instead_of_delete) { 1132 /* Here, the reference has not passed over to the coroutine, 1133 * so we have to delete the BB ourselves */ 1134 blk_unref(blk); 1135 } 1136 } 1137 1138 static void test_delete_by_drain(void) 1139 { 1140 do_test_delete_by_drain(false, BDRV_DRAIN); 1141 } 1142 1143 static void test_detach_by_drain_all(void) 1144 { 1145 do_test_delete_by_drain(true, BDRV_DRAIN_ALL); 1146 } 1147 1148 static void test_detach_by_drain(void) 1149 { 1150 do_test_delete_by_drain(true, BDRV_DRAIN); 1151 } 1152 1153 1154 struct detach_by_parent_data { 1155 BlockDriverState *parent_b; 1156 BdrvChild *child_b; 1157 BlockDriverState *c; 1158 BdrvChild *child_c; 1159 bool by_parent_cb; 1160 bool detach_on_drain; 1161 }; 1162 static struct detach_by_parent_data detach_by_parent_data; 1163 1164 static void no_coroutine_fn detach_indirect_bh(void *opaque) 1165 { 1166 struct detach_by_parent_data *data = opaque; 1167 1168 bdrv_dec_in_flight(data->child_b->bs); 1169 1170 bdrv_drain_all_begin(); 1171 bdrv_graph_wrlock(); 1172 bdrv_unref_child(data->parent_b, data->child_b); 1173 1174 bdrv_ref(data->c); 1175 data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C", 1176 &child_of_bds, BDRV_CHILD_DATA, 1177 &error_abort); 1178 bdrv_graph_wrunlock(); 1179 bdrv_drain_all_end(); 1180 } 1181 1182 static void coroutine_mixed_fn detach_by_parent_aio_cb(void *opaque, int ret) 1183 { 1184 struct detach_by_parent_data *data = &detach_by_parent_data; 1185 1186 g_assert_cmpint(ret, ==, 0); 1187 if (data->by_parent_cb) { 1188 bdrv_inc_in_flight(data->child_b->bs); 1189 aio_bh_schedule_oneshot(qemu_get_current_aio_context(), 1190 detach_indirect_bh, &detach_by_parent_data); 1191 } 1192 } 1193 1194 static void GRAPH_RDLOCK detach_by_driver_cb_drained_begin(BdrvChild *child) 1195 { 1196 struct detach_by_parent_data *data = &detach_by_parent_data; 1197 1198 if (!data->detach_on_drain) { 1199 return; 1200 } 1201 data->detach_on_drain = false; 1202 1203 bdrv_inc_in_flight(data->child_b->bs); 1204 aio_bh_schedule_oneshot(qemu_get_current_aio_context(), 1205 detach_indirect_bh, &detach_by_parent_data); 1206 child_of_bds.drained_begin(child); 1207 } 1208 1209 static BdrvChildClass detach_by_driver_cb_class; 1210 1211 /* 1212 * Initial graph: 1213 * 1214 * PA PB 1215 * \ / \ 1216 * A B C 1217 * 1218 * by_parent_cb == true: Test that parent callbacks don't poll 1219 * 1220 * PA has a pending write request whose callback changes the child nodes of 1221 * PB: It removes B and adds C instead. The subtree of PB is drained, which 1222 * will indirectly drain the write request, too. 1223 * 1224 * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll 1225 * 1226 * PA's BdrvChildClass has a .drained_begin callback that schedules a BH 1227 * that does the same graph change. If bdrv_drain_invoke() calls it, the 1228 * state is messed up, but if it is only polled in the single 1229 * BDRV_POLL_WHILE() at the end of the drain, this should work fine. 1230 */ 1231 static void TSA_NO_TSA test_detach_indirect(bool by_parent_cb) 1232 { 1233 BlockBackend *blk; 1234 BlockDriverState *parent_a, *parent_b, *a, *b, *c; 1235 BdrvChild *child_a, *child_b; 1236 BlockAIOCB *acb; 1237 1238 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0); 1239 1240 if (!by_parent_cb) { 1241 detach_by_driver_cb_class = child_of_bds; 1242 detach_by_driver_cb_class.drained_begin = 1243 detach_by_driver_cb_drained_begin; 1244 detach_by_driver_cb_class.drained_end = NULL; 1245 detach_by_driver_cb_class.drained_poll = NULL; 1246 } 1247 1248 detach_by_parent_data = (struct detach_by_parent_data) { 1249 .detach_on_drain = false, 1250 }; 1251 1252 /* Create all involved nodes */ 1253 parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR, 1254 &error_abort); 1255 parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0, 1256 &error_abort); 1257 1258 a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort); 1259 b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort); 1260 c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort); 1261 1262 /* blk is a BB for parent-a */ 1263 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 1264 blk_insert_bs(blk, parent_a, &error_abort); 1265 bdrv_unref(parent_a); 1266 1267 /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver 1268 * callback must not return immediately. */ 1269 if (!by_parent_cb) { 1270 BDRVTestState *s = parent_a->opaque; 1271 s->sleep_in_drain_begin = true; 1272 } 1273 1274 /* Set child relationships */ 1275 bdrv_ref(b); 1276 bdrv_ref(a); 1277 bdrv_drain_all_begin(); 1278 bdrv_graph_wrlock(); 1279 child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds, 1280 BDRV_CHILD_DATA, &error_abort); 1281 child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds, 1282 BDRV_CHILD_COW, &error_abort); 1283 1284 bdrv_ref(a); 1285 bdrv_attach_child(parent_a, a, "PA-A", 1286 by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class, 1287 BDRV_CHILD_DATA, &error_abort); 1288 bdrv_graph_wrunlock(); 1289 bdrv_drain_all_end(); 1290 1291 g_assert_cmpint(parent_a->refcnt, ==, 1); 1292 g_assert_cmpint(parent_b->refcnt, ==, 1); 1293 g_assert_cmpint(a->refcnt, ==, 3); 1294 g_assert_cmpint(b->refcnt, ==, 2); 1295 g_assert_cmpint(c->refcnt, ==, 1); 1296 1297 g_assert(QLIST_FIRST(&parent_b->children) == child_a); 1298 g_assert(QLIST_NEXT(child_a, next) == child_b); 1299 g_assert(QLIST_NEXT(child_b, next) == NULL); 1300 1301 /* Start the evil write request */ 1302 detach_by_parent_data = (struct detach_by_parent_data) { 1303 .parent_b = parent_b, 1304 .child_b = child_b, 1305 .c = c, 1306 .by_parent_cb = by_parent_cb, 1307 .detach_on_drain = true, 1308 }; 1309 acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL); 1310 g_assert(acb != NULL); 1311 1312 /* Drain and check the expected result */ 1313 bdrv_drained_begin(parent_b); 1314 bdrv_drained_begin(a); 1315 bdrv_drained_begin(b); 1316 bdrv_drained_begin(c); 1317 1318 g_assert(detach_by_parent_data.child_c != NULL); 1319 1320 g_assert_cmpint(parent_a->refcnt, ==, 1); 1321 g_assert_cmpint(parent_b->refcnt, ==, 1); 1322 g_assert_cmpint(a->refcnt, ==, 3); 1323 g_assert_cmpint(b->refcnt, ==, 1); 1324 g_assert_cmpint(c->refcnt, ==, 2); 1325 1326 g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c); 1327 g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a); 1328 g_assert(QLIST_NEXT(child_a, next) == NULL); 1329 1330 g_assert_cmpint(parent_a->quiesce_counter, ==, 1); 1331 g_assert_cmpint(parent_b->quiesce_counter, ==, 3); 1332 g_assert_cmpint(a->quiesce_counter, ==, 1); 1333 g_assert_cmpint(b->quiesce_counter, ==, 1); 1334 g_assert_cmpint(c->quiesce_counter, ==, 1); 1335 1336 bdrv_drained_end(parent_b); 1337 bdrv_drained_end(a); 1338 bdrv_drained_end(b); 1339 bdrv_drained_end(c); 1340 1341 bdrv_unref(parent_b); 1342 blk_unref(blk); 1343 1344 g_assert_cmpint(a->refcnt, ==, 1); 1345 g_assert_cmpint(b->refcnt, ==, 1); 1346 g_assert_cmpint(c->refcnt, ==, 1); 1347 bdrv_unref(a); 1348 bdrv_unref(b); 1349 bdrv_unref(c); 1350 } 1351 1352 static void test_detach_by_parent_cb(void) 1353 { 1354 test_detach_indirect(true); 1355 } 1356 1357 static void test_detach_by_driver_cb(void) 1358 { 1359 test_detach_indirect(false); 1360 } 1361 1362 static void test_append_to_drained(void) 1363 { 1364 BlockBackend *blk; 1365 BlockDriverState *base, *overlay; 1366 BDRVTestState *base_s, *overlay_s; 1367 1368 blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL); 1369 base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort); 1370 base_s = base->opaque; 1371 blk_insert_bs(blk, base, &error_abort); 1372 1373 overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR, 1374 &error_abort); 1375 overlay_s = overlay->opaque; 1376 1377 do_drain_begin(BDRV_DRAIN, base); 1378 g_assert_cmpint(base->quiesce_counter, ==, 1); 1379 g_assert_cmpint(base_s->drain_count, ==, 1); 1380 g_assert_cmpint(base->in_flight, ==, 0); 1381 1382 bdrv_append(overlay, base, &error_abort); 1383 1384 g_assert_cmpint(base->in_flight, ==, 0); 1385 g_assert_cmpint(overlay->in_flight, ==, 0); 1386 1387 g_assert_cmpint(base->quiesce_counter, ==, 1); 1388 g_assert_cmpint(base_s->drain_count, ==, 1); 1389 g_assert_cmpint(overlay->quiesce_counter, ==, 1); 1390 g_assert_cmpint(overlay_s->drain_count, ==, 1); 1391 1392 do_drain_end(BDRV_DRAIN, base); 1393 1394 g_assert_cmpint(base->quiesce_counter, ==, 0); 1395 g_assert_cmpint(base_s->drain_count, ==, 0); 1396 g_assert_cmpint(overlay->quiesce_counter, ==, 0); 1397 g_assert_cmpint(overlay_s->drain_count, ==, 0); 1398 1399 bdrv_unref(overlay); 1400 bdrv_unref(base); 1401 blk_unref(blk); 1402 } 1403 1404 static void test_set_aio_context(void) 1405 { 1406 BlockDriverState *bs; 1407 IOThread *a = iothread_new(); 1408 IOThread *b = iothread_new(); 1409 AioContext *ctx_a = iothread_get_aio_context(a); 1410 AioContext *ctx_b = iothread_get_aio_context(b); 1411 1412 bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR, 1413 &error_abort); 1414 1415 bdrv_try_change_aio_context(bs, ctx_a, NULL, &error_abort); 1416 1417 bdrv_try_change_aio_context(bs, ctx_b, NULL, &error_abort); 1418 bdrv_try_change_aio_context(bs, qemu_get_aio_context(), NULL, &error_abort); 1419 1420 bdrv_unref(bs); 1421 iothread_join(a); 1422 iothread_join(b); 1423 } 1424 1425 1426 typedef struct TestDropBackingBlockJob { 1427 BlockJob common; 1428 bool *did_complete; 1429 BlockDriverState *detach_also; 1430 BlockDriverState *bs; 1431 1432 /* Accessed with atomics */ 1433 bool should_complete; 1434 } TestDropBackingBlockJob; 1435 1436 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp) 1437 { 1438 TestDropBackingBlockJob *s = 1439 container_of(job, TestDropBackingBlockJob, common.job); 1440 1441 while (!qatomic_read(&s->should_complete)) { 1442 job_sleep_ns(job, 0); 1443 } 1444 1445 return 0; 1446 } 1447 1448 static void test_drop_backing_job_commit(Job *job) 1449 { 1450 TestDropBackingBlockJob *s = 1451 container_of(job, TestDropBackingBlockJob, common.job); 1452 1453 bdrv_set_backing_hd(s->bs, NULL, &error_abort); 1454 bdrv_set_backing_hd(s->detach_also, NULL, &error_abort); 1455 1456 *s->did_complete = true; 1457 } 1458 1459 static const BlockJobDriver test_drop_backing_job_driver = { 1460 .job_driver = { 1461 .instance_size = sizeof(TestDropBackingBlockJob), 1462 .free = block_job_free, 1463 .user_resume = block_job_user_resume, 1464 .run = test_drop_backing_job_run, 1465 .commit = test_drop_backing_job_commit, 1466 } 1467 }; 1468 1469 /** 1470 * Creates a child node with three parent nodes on it, and then runs a 1471 * block job on the final one, parent-node-2. 1472 * 1473 * The job is then asked to complete before a section where the child 1474 * is drained. 1475 * 1476 * Ending this section will undrain the child's parents, first 1477 * parent-node-2, then parent-node-1, then parent-node-0 -- the parent 1478 * list is in reverse order of how they were added. Ending the drain 1479 * on parent-node-2 will resume the job, thus completing it and 1480 * scheduling job_exit(). 1481 * 1482 * Ending the drain on parent-node-1 will poll the AioContext, which 1483 * lets job_exit() and thus test_drop_backing_job_commit() run. That 1484 * function first removes the child as parent-node-2's backing file. 1485 * 1486 * In old (and buggy) implementations, there are two problems with 1487 * that: 1488 * (A) bdrv_drain_invoke() polls for every node that leaves the 1489 * drained section. This means that job_exit() is scheduled 1490 * before the child has left the drained section. Its 1491 * quiesce_counter is therefore still 1 when it is removed from 1492 * parent-node-2. 1493 * 1494 * (B) bdrv_replace_child_noperm() calls drained_end() on the old 1495 * child's parents as many times as the child is quiesced. This 1496 * means it will call drained_end() on parent-node-2 once. 1497 * Because parent-node-2 is no longer quiesced at this point, this 1498 * will fail. 1499 * 1500 * bdrv_replace_child_noperm() therefore must call drained_end() on 1501 * the parent only if it really is still drained because the child is 1502 * drained. 1503 * 1504 * If removing child from parent-node-2 was successful (as it should 1505 * be), test_drop_backing_job_commit() will then also remove the child 1506 * from parent-node-0. 1507 * 1508 * With an old version of our drain infrastructure ((A) above), that 1509 * resulted in the following flow: 1510 * 1511 * 1. child attempts to leave its drained section. The call recurses 1512 * to its parents. 1513 * 1514 * 2. parent-node-2 leaves the drained section. Polling in 1515 * bdrv_drain_invoke() will schedule job_exit(). 1516 * 1517 * 3. parent-node-1 leaves the drained section. Polling in 1518 * bdrv_drain_invoke() will run job_exit(), thus disconnecting 1519 * parent-node-0 from the child node. 1520 * 1521 * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to 1522 * iterate over the parents. Thus, it now accesses the BdrvChild 1523 * object that used to connect parent-node-0 and the child node. 1524 * However, that object no longer exists, so it accesses a dangling 1525 * pointer. 1526 * 1527 * The solution is to only poll once when running a bdrv_drained_end() 1528 * operation, specifically at the end when all drained_end() 1529 * operations for all involved nodes have been scheduled. 1530 * Note that this also solves (A) above, thus hiding (B). 1531 */ 1532 static void test_blockjob_commit_by_drained_end(void) 1533 { 1534 BlockDriverState *bs_child, *bs_parents[3]; 1535 TestDropBackingBlockJob *job; 1536 bool job_has_completed = false; 1537 int i; 1538 1539 bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR, 1540 &error_abort); 1541 1542 for (i = 0; i < 3; i++) { 1543 char name[32]; 1544 snprintf(name, sizeof(name), "parent-node-%i", i); 1545 bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR, 1546 &error_abort); 1547 bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort); 1548 } 1549 1550 job = block_job_create("job", &test_drop_backing_job_driver, NULL, 1551 bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL, 1552 &error_abort); 1553 job->bs = bs_parents[2]; 1554 1555 job->detach_also = bs_parents[0]; 1556 job->did_complete = &job_has_completed; 1557 1558 job_start(&job->common.job); 1559 1560 qatomic_set(&job->should_complete, true); 1561 bdrv_drained_begin(bs_child); 1562 g_assert(!job_has_completed); 1563 bdrv_drained_end(bs_child); 1564 aio_poll(qemu_get_aio_context(), false); 1565 g_assert(job_has_completed); 1566 1567 bdrv_unref(bs_parents[0]); 1568 bdrv_unref(bs_parents[1]); 1569 bdrv_unref(bs_parents[2]); 1570 bdrv_unref(bs_child); 1571 } 1572 1573 1574 typedef struct TestSimpleBlockJob { 1575 BlockJob common; 1576 bool *did_complete; 1577 1578 /* Accessed with atomics */ 1579 bool should_complete; 1580 } TestSimpleBlockJob; 1581 1582 static int coroutine_fn test_simple_job_run(Job *job, Error **errp) 1583 { 1584 TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job); 1585 1586 while (!qatomic_read(&s->should_complete)) { 1587 job_sleep_ns(job, 0); 1588 } 1589 1590 return 0; 1591 } 1592 1593 static void test_simple_job_clean(Job *job) 1594 { 1595 TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job); 1596 *s->did_complete = true; 1597 } 1598 1599 static const BlockJobDriver test_simple_job_driver = { 1600 .job_driver = { 1601 .instance_size = sizeof(TestSimpleBlockJob), 1602 .free = block_job_free, 1603 .user_resume = block_job_user_resume, 1604 .run = test_simple_job_run, 1605 .clean = test_simple_job_clean, 1606 }, 1607 }; 1608 1609 static int drop_intermediate_poll_update_filename(BdrvChild *child, 1610 BlockDriverState *new_base, 1611 const char *filename, 1612 bool backing_mask_protocol, 1613 Error **errp) 1614 { 1615 /* 1616 * We are free to poll here, which may change the block graph, if 1617 * it is not drained. 1618 */ 1619 1620 /* If the job is not drained: Complete it, schedule job_exit() */ 1621 aio_poll(qemu_get_current_aio_context(), false); 1622 /* If the job is not drained: Run job_exit(), finish the job */ 1623 aio_poll(qemu_get_current_aio_context(), false); 1624 1625 return 0; 1626 } 1627 1628 /** 1629 * Test a poll in the midst of bdrv_drop_intermediate(). 1630 * 1631 * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(), 1632 * which can yield or poll. This may lead to graph changes, unless 1633 * the whole subtree in question is drained. 1634 * 1635 * We test this on the following graph: 1636 * 1637 * Job 1638 * 1639 * | 1640 * job-node 1641 * | 1642 * v 1643 * 1644 * job-node 1645 * 1646 * | 1647 * backing 1648 * | 1649 * v 1650 * 1651 * node-2 --chain--> node-1 --chain--> node-0 1652 * 1653 * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0). 1654 * 1655 * This first updates node-2's backing filename by invoking 1656 * drop_intermediate_poll_update_filename(), which polls twice. This 1657 * causes the job to finish, which in turns causes the job-node to be 1658 * deleted. 1659 * 1660 * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it 1661 * already has a pointer to the BdrvChild edge between job-node and 1662 * node-1. When it tries to handle that edge, we probably get a 1663 * segmentation fault because the object no longer exists. 1664 * 1665 * 1666 * The solution is for bdrv_drop_intermediate() to drain top's 1667 * subtree. This prevents graph changes from happening just because 1668 * BdrvChildClass.update_filename() yields or polls. Thus, the block 1669 * job is paused during that drained section and must finish before or 1670 * after. 1671 * 1672 * (In addition, bdrv_replace_child() must keep the job paused.) 1673 */ 1674 static void test_drop_intermediate_poll(void) 1675 { 1676 static BdrvChildClass chain_child_class; 1677 BlockDriverState *chain[3]; 1678 TestSimpleBlockJob *job; 1679 BlockDriverState *job_node; 1680 bool job_has_completed = false; 1681 int i; 1682 int ret; 1683 1684 chain_child_class = child_of_bds; 1685 chain_child_class.update_filename = drop_intermediate_poll_update_filename; 1686 1687 for (i = 0; i < 3; i++) { 1688 char name[32]; 1689 snprintf(name, 32, "node-%i", i); 1690 1691 chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort); 1692 } 1693 1694 job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR, 1695 &error_abort); 1696 bdrv_set_backing_hd(job_node, chain[1], &error_abort); 1697 1698 /* 1699 * Establish the chain last, so the chain links are the first 1700 * elements in the BDS.parents lists 1701 */ 1702 bdrv_drain_all_begin(); 1703 bdrv_graph_wrlock(); 1704 for (i = 0; i < 3; i++) { 1705 if (i) { 1706 /* Takes the reference to chain[i - 1] */ 1707 bdrv_attach_child(chain[i], chain[i - 1], "chain", 1708 &chain_child_class, BDRV_CHILD_COW, &error_abort); 1709 } 1710 } 1711 bdrv_graph_wrunlock(); 1712 bdrv_drain_all_end(); 1713 1714 job = block_job_create("job", &test_simple_job_driver, NULL, job_node, 1715 0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort); 1716 1717 /* The job has a reference now */ 1718 bdrv_unref(job_node); 1719 1720 job->did_complete = &job_has_completed; 1721 1722 job_start(&job->common.job); 1723 qatomic_set(&job->should_complete, true); 1724 1725 g_assert(!job_has_completed); 1726 ret = bdrv_drop_intermediate(chain[1], chain[0], NULL, false); 1727 aio_poll(qemu_get_aio_context(), false); 1728 g_assert(ret == 0); 1729 g_assert(job_has_completed); 1730 1731 bdrv_unref(chain[2]); 1732 } 1733 1734 1735 typedef struct BDRVReplaceTestState { 1736 bool setup_completed; 1737 bool was_drained; 1738 bool was_undrained; 1739 bool has_read; 1740 1741 int drain_count; 1742 1743 bool yield_before_read; 1744 Coroutine *io_co; 1745 Coroutine *drain_co; 1746 } BDRVReplaceTestState; 1747 1748 static void bdrv_replace_test_close(BlockDriverState *bs) 1749 { 1750 } 1751 1752 /** 1753 * If @bs has a backing file: 1754 * Yield if .yield_before_read is true (and wait for drain_begin to 1755 * wake us up). 1756 * Forward the read to bs->backing. Set .has_read to true. 1757 * If drain_begin has woken us, wake it in turn. 1758 * 1759 * Otherwise: 1760 * Set .has_read to true and return success. 1761 */ 1762 static int coroutine_fn GRAPH_RDLOCK 1763 bdrv_replace_test_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes, 1764 QEMUIOVector *qiov, BdrvRequestFlags flags) 1765 { 1766 BDRVReplaceTestState *s = bs->opaque; 1767 1768 if (bs->backing) { 1769 int ret; 1770 1771 g_assert(!s->drain_count); 1772 1773 s->io_co = qemu_coroutine_self(); 1774 if (s->yield_before_read) { 1775 s->yield_before_read = false; 1776 qemu_coroutine_yield(); 1777 } 1778 s->io_co = NULL; 1779 1780 ret = bdrv_co_preadv(bs->backing, offset, bytes, qiov, 0); 1781 s->has_read = true; 1782 1783 /* Wake up drain_co if it runs */ 1784 if (s->drain_co) { 1785 aio_co_wake(s->drain_co); 1786 } 1787 1788 return ret; 1789 } 1790 1791 s->has_read = true; 1792 return 0; 1793 } 1794 1795 static void coroutine_fn bdrv_replace_test_drain_co(void *opaque) 1796 { 1797 BlockDriverState *bs = opaque; 1798 BDRVReplaceTestState *s = bs->opaque; 1799 1800 /* Keep waking io_co up until it is done */ 1801 while (s->io_co) { 1802 aio_co_wake(s->io_co); 1803 s->io_co = NULL; 1804 qemu_coroutine_yield(); 1805 } 1806 s->drain_co = NULL; 1807 bdrv_dec_in_flight(bs); 1808 } 1809 1810 /** 1811 * If .drain_count is 0, wake up .io_co if there is one; and set 1812 * .was_drained. 1813 * Increment .drain_count. 1814 */ 1815 static void bdrv_replace_test_drain_begin(BlockDriverState *bs) 1816 { 1817 BDRVReplaceTestState *s = bs->opaque; 1818 1819 if (!s->setup_completed) { 1820 return; 1821 } 1822 1823 if (!s->drain_count) { 1824 s->drain_co = qemu_coroutine_create(bdrv_replace_test_drain_co, bs); 1825 bdrv_inc_in_flight(bs); 1826 aio_co_enter(bdrv_get_aio_context(bs), s->drain_co); 1827 s->was_drained = true; 1828 } 1829 s->drain_count++; 1830 } 1831 1832 static void coroutine_fn bdrv_replace_test_read_entry(void *opaque) 1833 { 1834 BlockDriverState *bs = opaque; 1835 char data; 1836 QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1); 1837 int ret; 1838 1839 /* Queue a read request post-drain */ 1840 bdrv_graph_co_rdlock(); 1841 ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0); 1842 bdrv_graph_co_rdunlock(); 1843 1844 g_assert(ret >= 0); 1845 bdrv_dec_in_flight(bs); 1846 } 1847 1848 /** 1849 * Reduce .drain_count, set .was_undrained once it reaches 0. 1850 * If .drain_count reaches 0 and the node has a backing file, issue a 1851 * read request. 1852 */ 1853 static void bdrv_replace_test_drain_end(BlockDriverState *bs) 1854 { 1855 BDRVReplaceTestState *s = bs->opaque; 1856 1857 GRAPH_RDLOCK_GUARD_MAINLOOP(); 1858 1859 if (!s->setup_completed) { 1860 return; 1861 } 1862 1863 g_assert(s->drain_count > 0); 1864 if (!--s->drain_count) { 1865 s->was_undrained = true; 1866 1867 if (bs->backing) { 1868 Coroutine *co = qemu_coroutine_create(bdrv_replace_test_read_entry, 1869 bs); 1870 bdrv_inc_in_flight(bs); 1871 aio_co_enter(bdrv_get_aio_context(bs), co); 1872 } 1873 } 1874 } 1875 1876 static BlockDriver bdrv_replace_test = { 1877 .format_name = "replace_test", 1878 .instance_size = sizeof(BDRVReplaceTestState), 1879 .supports_backing = true, 1880 1881 .bdrv_close = bdrv_replace_test_close, 1882 .bdrv_co_preadv = bdrv_replace_test_co_preadv, 1883 1884 .bdrv_drain_begin = bdrv_replace_test_drain_begin, 1885 .bdrv_drain_end = bdrv_replace_test_drain_end, 1886 1887 .bdrv_child_perm = bdrv_default_perms, 1888 }; 1889 1890 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque) 1891 { 1892 int ret; 1893 char data; 1894 1895 ret = blk_co_pread(opaque, 0, 1, &data, 0); 1896 g_assert(ret >= 0); 1897 } 1898 1899 /** 1900 * We test two things: 1901 * (1) bdrv_replace_child_noperm() must not undrain the parent if both 1902 * children are drained. 1903 * (2) bdrv_replace_child_noperm() must never flush I/O requests to a 1904 * drained child. If the old child is drained, it must flush I/O 1905 * requests after the new one has been attached. If the new child 1906 * is drained, it must flush I/O requests before the old one is 1907 * detached. 1908 * 1909 * To do so, we create one parent node and two child nodes; then 1910 * attach one of the children (old_child_bs) to the parent, then 1911 * drain both old_child_bs and new_child_bs according to 1912 * old_drain_count and new_drain_count, respectively, and finally 1913 * we invoke bdrv_replace_node() to replace old_child_bs by 1914 * new_child_bs. 1915 * 1916 * The test block driver we use here (bdrv_replace_test) has a read 1917 * function that: 1918 * - For the parent node, can optionally yield, and then forwards the 1919 * read to bdrv_preadv(), 1920 * - For the child node, just returns immediately. 1921 * 1922 * If the read yields, the drain_begin function will wake it up. 1923 * 1924 * The drain_end function issues a read on the parent once it is fully 1925 * undrained (which simulates requests starting to come in again). 1926 */ 1927 static void do_test_replace_child_mid_drain(int old_drain_count, 1928 int new_drain_count) 1929 { 1930 BlockBackend *parent_blk; 1931 BlockDriverState *parent_bs; 1932 BlockDriverState *old_child_bs, *new_child_bs; 1933 BDRVReplaceTestState *parent_s; 1934 BDRVReplaceTestState *old_child_s, *new_child_s; 1935 Coroutine *io_co; 1936 int i; 1937 1938 parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0, 1939 &error_abort); 1940 parent_s = parent_bs->opaque; 1941 1942 parent_blk = blk_new(qemu_get_aio_context(), 1943 BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL); 1944 blk_insert_bs(parent_blk, parent_bs, &error_abort); 1945 1946 old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0, 1947 &error_abort); 1948 new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0, 1949 &error_abort); 1950 old_child_s = old_child_bs->opaque; 1951 new_child_s = new_child_bs->opaque; 1952 1953 /* So that we can read something */ 1954 parent_bs->total_sectors = 1; 1955 old_child_bs->total_sectors = 1; 1956 new_child_bs->total_sectors = 1; 1957 1958 bdrv_ref(old_child_bs); 1959 bdrv_drain_all_begin(); 1960 bdrv_graph_wrlock(); 1961 bdrv_attach_child(parent_bs, old_child_bs, "child", &child_of_bds, 1962 BDRV_CHILD_COW, &error_abort); 1963 bdrv_graph_wrunlock(); 1964 bdrv_drain_all_end(); 1965 parent_s->setup_completed = true; 1966 1967 for (i = 0; i < old_drain_count; i++) { 1968 bdrv_drained_begin(old_child_bs); 1969 } 1970 for (i = 0; i < new_drain_count; i++) { 1971 bdrv_drained_begin(new_child_bs); 1972 } 1973 1974 if (!old_drain_count) { 1975 /* 1976 * Start a read operation that will yield, so it will not 1977 * complete before the node is drained. 1978 */ 1979 parent_s->yield_before_read = true; 1980 io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co, 1981 parent_blk); 1982 qemu_coroutine_enter(io_co); 1983 } 1984 1985 /* If we have started a read operation, it should have yielded */ 1986 g_assert(!parent_s->has_read); 1987 1988 /* Reset drained status so we can see what bdrv_replace_node() does */ 1989 parent_s->was_drained = false; 1990 parent_s->was_undrained = false; 1991 1992 g_assert(parent_bs->quiesce_counter == old_drain_count); 1993 bdrv_drained_begin(old_child_bs); 1994 bdrv_drained_begin(new_child_bs); 1995 bdrv_graph_wrlock(); 1996 bdrv_replace_node(old_child_bs, new_child_bs, &error_abort); 1997 bdrv_graph_wrunlock(); 1998 bdrv_drained_end(new_child_bs); 1999 bdrv_drained_end(old_child_bs); 2000 g_assert(parent_bs->quiesce_counter == new_drain_count); 2001 2002 if (!old_drain_count && !new_drain_count) { 2003 /* 2004 * From undrained to undrained drains and undrains the parent, 2005 * because bdrv_replace_node() contains a drained section for 2006 * @old_child_bs. 2007 */ 2008 g_assert(parent_s->was_drained && parent_s->was_undrained); 2009 } else if (!old_drain_count && new_drain_count) { 2010 /* 2011 * From undrained to drained should drain the parent and keep 2012 * it that way. 2013 */ 2014 g_assert(parent_s->was_drained && !parent_s->was_undrained); 2015 } else if (old_drain_count && !new_drain_count) { 2016 /* 2017 * From drained to undrained should undrain the parent and 2018 * keep it that way. 2019 */ 2020 g_assert(!parent_s->was_drained && parent_s->was_undrained); 2021 } else /* if (old_drain_count && new_drain_count) */ { 2022 /* 2023 * From drained to drained must not undrain the parent at any 2024 * point 2025 */ 2026 g_assert(!parent_s->was_drained && !parent_s->was_undrained); 2027 } 2028 2029 if (!old_drain_count || !new_drain_count) { 2030 /* 2031 * If !old_drain_count, we have started a read request before 2032 * bdrv_replace_node(). If !new_drain_count, the parent must 2033 * have been undrained at some point, and 2034 * bdrv_replace_test_co_drain_end() starts a read request 2035 * then. 2036 */ 2037 g_assert(parent_s->has_read); 2038 } else { 2039 /* 2040 * If the parent was never undrained, there is no way to start 2041 * a read request. 2042 */ 2043 g_assert(!parent_s->has_read); 2044 } 2045 2046 /* A drained child must have not received any request */ 2047 g_assert(!(old_drain_count && old_child_s->has_read)); 2048 g_assert(!(new_drain_count && new_child_s->has_read)); 2049 2050 for (i = 0; i < new_drain_count; i++) { 2051 bdrv_drained_end(new_child_bs); 2052 } 2053 for (i = 0; i < old_drain_count; i++) { 2054 bdrv_drained_end(old_child_bs); 2055 } 2056 2057 /* 2058 * By now, bdrv_replace_test_co_drain_end() must have been called 2059 * at some point while the new child was attached to the parent. 2060 */ 2061 g_assert(parent_s->has_read); 2062 g_assert(new_child_s->has_read); 2063 2064 blk_unref(parent_blk); 2065 bdrv_unref(parent_bs); 2066 bdrv_unref(old_child_bs); 2067 bdrv_unref(new_child_bs); 2068 } 2069 2070 static void test_replace_child_mid_drain(void) 2071 { 2072 int old_drain_count, new_drain_count; 2073 2074 for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) { 2075 for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) { 2076 do_test_replace_child_mid_drain(old_drain_count, new_drain_count); 2077 } 2078 } 2079 } 2080 2081 int main(int argc, char **argv) 2082 { 2083 int ret; 2084 2085 bdrv_init(); 2086 qemu_init_main_loop(&error_abort); 2087 2088 g_test_init(&argc, &argv, NULL); 2089 qemu_event_init(&done_event, false); 2090 2091 g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all); 2092 g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain); 2093 2094 g_test_add_func("/bdrv-drain/driver-cb/co/drain_all", 2095 test_drv_cb_co_drain_all); 2096 g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain); 2097 2098 g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all); 2099 g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain); 2100 2101 g_test_add_func("/bdrv-drain/quiesce/co/drain_all", 2102 test_quiesce_co_drain_all); 2103 g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain); 2104 2105 g_test_add_func("/bdrv-drain/nested", test_nested); 2106 2107 g_test_add_func("/bdrv-drain/graph-change/drain_all", 2108 test_graph_change_drain_all); 2109 2110 g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all); 2111 g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain); 2112 2113 g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all); 2114 g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain); 2115 2116 g_test_add_func("/bdrv-drain/blockjob/error/drain_all", 2117 test_blockjob_error_drain_all); 2118 g_test_add_func("/bdrv-drain/blockjob/error/drain", 2119 test_blockjob_error_drain); 2120 2121 g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all", 2122 test_blockjob_iothread_drain_all); 2123 g_test_add_func("/bdrv-drain/blockjob/iothread/drain", 2124 test_blockjob_iothread_drain); 2125 2126 g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all", 2127 test_blockjob_iothread_error_drain_all); 2128 g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain", 2129 test_blockjob_iothread_error_drain); 2130 2131 g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain); 2132 g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all); 2133 g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain); 2134 g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb); 2135 g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb); 2136 2137 g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained); 2138 2139 g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context); 2140 2141 g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end", 2142 test_blockjob_commit_by_drained_end); 2143 2144 g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll", 2145 test_drop_intermediate_poll); 2146 2147 g_test_add_func("/bdrv-drain/replace_child/mid-drain", 2148 test_replace_child_mid_drain); 2149 2150 ret = g_test_run(); 2151 qemu_event_destroy(&done_event); 2152 return ret; 2153 } 2154