tests/test-bdrv-drain.c

   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.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 bdrv_test_co_drain_begin(BlockDriverState *bs)
  42 {
  43     BDRVTestState *s = bs->opaque;
  44     s->drain_count++;
  45     if (s->sleep_in_drain_begin) {
  46         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
  47     }
  48 }
  49
  50 static void coroutine_fn bdrv_test_co_drain_end(BlockDriverState *bs)
  51 {
  52     BDRVTestState *s = bs->opaque;
  53     s->drain_count--;
  54 }
  55
  56 static void bdrv_test_close(BlockDriverState *bs)
  57 {
  58     BDRVTestState *s = bs->opaque;
  59     g_assert_cmpint(s->drain_count, >, 0);
  60 }
  61
  62 static void co_reenter_bh(void *opaque)
  63 {
  64     aio_co_wake(opaque);
  65 }
  66
  67 static int coroutine_fn bdrv_test_co_preadv(BlockDriverState *bs,
  68                                             uint64_t offset, uint64_t bytes,
  69                                             QEMUIOVector *qiov, int flags)
  70 {
  71     BDRVTestState *s = bs->opaque;
  72
  73     /* We want this request to stay until the polling loop in drain waits for
  74      * it to complete. We need to sleep a while as bdrv_drain_invoke() comes
  75      * first and polls its result, too, but it shouldn't accidentally complete
  76      * this request yet. */
  77     qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
  78
  79     if (s->bh_indirection_ctx) {
  80         aio_bh_schedule_oneshot(s->bh_indirection_ctx, co_reenter_bh,
  81                                 qemu_coroutine_self());
  82         qemu_coroutine_yield();
  83     }
  84
  85     return 0;
  86 }
  87
  88 static void bdrv_test_child_perm(BlockDriverState *bs, BdrvChild *c,
  89                                  const BdrvChildRole *role,
  90                                  BlockReopenQueue *reopen_queue,
  91                                  uint64_t perm, uint64_t shared,
  92                                  uint64_t *nperm, uint64_t *nshared)
  93 {
  94     /* bdrv_format_default_perms() accepts only these two, so disguise
  95      * detach_by_driver_cb_role as one of them. */
  96     if (role != &child_file && role != &child_backing) {
  97         role = &child_file;
  98     }
  99
 100     bdrv_format_default_perms(bs, c, role, reopen_queue, perm, shared,
 101                               nperm, nshared);
 102 }
 103
 104 static int bdrv_test_change_backing_file(BlockDriverState *bs,
 105                                          const char *backing_file,
 106                                          const char *backing_fmt)
 107 {
 108     return 0;
 109 }
 110
 111 static BlockDriver bdrv_test = {
 112     .format_name            = "test",
 113     .instance_size          = sizeof(BDRVTestState),
 114
 115     .bdrv_close             = bdrv_test_close,
 116     .bdrv_co_preadv         = bdrv_test_co_preadv,
 117
 118     .bdrv_co_drain_begin    = bdrv_test_co_drain_begin,
 119     .bdrv_co_drain_end      = bdrv_test_co_drain_end,
 120
 121     .bdrv_child_perm        = bdrv_test_child_perm,
 122
 123     .bdrv_change_backing_file = bdrv_test_change_backing_file,
 124 };
 125
 126 static void aio_ret_cb(void *opaque, int ret)
 127 {
 128     int *aio_ret = opaque;
 129     *aio_ret = ret;
 130 }
 131
 132 typedef struct CallInCoroutineData {
 133     void (*entry)(void);
 134     bool done;
 135 } CallInCoroutineData;
 136
 137 static coroutine_fn void call_in_coroutine_entry(void *opaque)
 138 {
 139     CallInCoroutineData *data = opaque;
 140
 141     data->entry();
 142     data->done = true;
 143 }
 144
 145 static void call_in_coroutine(void (*entry)(void))
 146 {
 147     Coroutine *co;
 148     CallInCoroutineData data = {
 149         .entry  = entry,
 150         .done   = false,
 151     };
 152
 153     co = qemu_coroutine_create(call_in_coroutine_entry, &data);
 154     qemu_coroutine_enter(co);
 155     while (!data.done) {
 156         aio_poll(qemu_get_aio_context(), true);
 157     }
 158 }
 159
 160 enum drain_type {
 161     BDRV_DRAIN_ALL,
 162     BDRV_DRAIN,
 163     BDRV_SUBTREE_DRAIN,
 164     DRAIN_TYPE_MAX,
 165 };
 166
 167 static void do_drain_begin(enum drain_type drain_type, BlockDriverState *bs)
 168 {
 169     switch (drain_type) {
 170     case BDRV_DRAIN_ALL:        bdrv_drain_all_begin(); break;
 171     case BDRV_DRAIN:            bdrv_drained_begin(bs); break;
 172     case BDRV_SUBTREE_DRAIN:    bdrv_subtree_drained_begin(bs); break;
 173     default:                    g_assert_not_reached();
 174     }
 175 }
 176
 177 static void do_drain_end(enum drain_type drain_type, BlockDriverState *bs)
 178 {
 179     switch (drain_type) {
 180     case BDRV_DRAIN_ALL:        bdrv_drain_all_end(); break;
 181     case BDRV_DRAIN:            bdrv_drained_end(bs); break;
 182     case BDRV_SUBTREE_DRAIN:    bdrv_subtree_drained_end(bs); break;
 183     default:                    g_assert_not_reached();
 184     }
 185 }
 186
 187 static void do_drain_begin_unlocked(enum drain_type drain_type, BlockDriverState *bs)
 188 {
 189     if (drain_type != BDRV_DRAIN_ALL) {
 190         aio_context_acquire(bdrv_get_aio_context(bs));
 191     }
 192     do_drain_begin(drain_type, bs);
 193     if (drain_type != BDRV_DRAIN_ALL) {
 194         aio_context_release(bdrv_get_aio_context(bs));
 195     }
 196 }
 197
 198 static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs)
 199 {
 200     if (drain_type != BDRV_DRAIN_ALL) {
 201         aio_context_acquire(bdrv_get_aio_context(bs));
 202     }
 203     do_drain_end(drain_type, bs);
 204     if (drain_type != BDRV_DRAIN_ALL) {
 205         aio_context_release(bdrv_get_aio_context(bs));
 206     }
 207 }
 208
 209 static void test_drv_cb_common(enum drain_type drain_type, bool recursive)
 210 {
 211     BlockBackend *blk;
 212     BlockDriverState *bs, *backing;
 213     BDRVTestState *s, *backing_s;
 214     BlockAIOCB *acb;
 215     int aio_ret;
 216
 217     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
 218
 219     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 220     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 221                               &error_abort);
 222     s = bs->opaque;
 223     blk_insert_bs(blk, bs, &error_abort);
 224
 225     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 226     backing_s = backing->opaque;
 227     bdrv_set_backing_hd(bs, backing, &error_abort);
 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     bdrv_unref(backing);
 264     bdrv_unref(bs);
 265     blk_unref(blk);
 266 }
 267
 268 static void test_drv_cb_drain_all(void)
 269 {
 270     test_drv_cb_common(BDRV_DRAIN_ALL, true);
 271 }
 272
 273 static void test_drv_cb_drain(void)
 274 {
 275     test_drv_cb_common(BDRV_DRAIN, false);
 276 }
 277
 278 static void test_drv_cb_drain_subtree(void)
 279 {
 280     test_drv_cb_common(BDRV_SUBTREE_DRAIN, true);
 281 }
 282
 283 static void test_drv_cb_co_drain_all(void)
 284 {
 285     call_in_coroutine(test_drv_cb_drain_all);
 286 }
 287
 288 static void test_drv_cb_co_drain(void)
 289 {
 290     call_in_coroutine(test_drv_cb_drain);
 291 }
 292
 293 static void test_drv_cb_co_drain_subtree(void)
 294 {
 295     call_in_coroutine(test_drv_cb_drain_subtree);
 296 }
 297
 298 static void test_quiesce_common(enum drain_type drain_type, bool recursive)
 299 {
 300     BlockBackend *blk;
 301     BlockDriverState *bs, *backing;
 302
 303     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 304     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 305                               &error_abort);
 306     blk_insert_bs(blk, bs, &error_abort);
 307
 308     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 309     bdrv_set_backing_hd(bs, backing, &error_abort);
 310
 311     g_assert_cmpint(bs->quiesce_counter, ==, 0);
 312     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 313
 314     do_drain_begin(drain_type, bs);
 315
 316     g_assert_cmpint(bs->quiesce_counter, ==, 1);
 317     g_assert_cmpint(backing->quiesce_counter, ==, !!recursive);
 318
 319     do_drain_end(drain_type, bs);
 320
 321     g_assert_cmpint(bs->quiesce_counter, ==, 0);
 322     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 323
 324     bdrv_unref(backing);
 325     bdrv_unref(bs);
 326     blk_unref(blk);
 327 }
 328
 329 static void test_quiesce_drain_all(void)
 330 {
 331     test_quiesce_common(BDRV_DRAIN_ALL, true);
 332 }
 333
 334 static void test_quiesce_drain(void)
 335 {
 336     test_quiesce_common(BDRV_DRAIN, false);
 337 }
 338
 339 static void test_quiesce_drain_subtree(void)
 340 {
 341     test_quiesce_common(BDRV_SUBTREE_DRAIN, true);
 342 }
 343
 344 static void test_quiesce_co_drain_all(void)
 345 {
 346     call_in_coroutine(test_quiesce_drain_all);
 347 }
 348
 349 static void test_quiesce_co_drain(void)
 350 {
 351     call_in_coroutine(test_quiesce_drain);
 352 }
 353
 354 static void test_quiesce_co_drain_subtree(void)
 355 {
 356     call_in_coroutine(test_quiesce_drain_subtree);
 357 }
 358
 359 static void test_nested(void)
 360 {
 361     BlockBackend *blk;
 362     BlockDriverState *bs, *backing;
 363     BDRVTestState *s, *backing_s;
 364     enum drain_type outer, inner;
 365
 366     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 367     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 368                               &error_abort);
 369     s = bs->opaque;
 370     blk_insert_bs(blk, bs, &error_abort);
 371
 372     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 373     backing_s = backing->opaque;
 374     bdrv_set_backing_hd(bs, backing, &error_abort);
 375
 376     for (outer = 0; outer < DRAIN_TYPE_MAX; outer++) {
 377         for (inner = 0; inner < DRAIN_TYPE_MAX; inner++) {
 378             int backing_quiesce = (outer != BDRV_DRAIN) +
 379                                   (inner != BDRV_DRAIN);
 380
 381             g_assert_cmpint(bs->quiesce_counter, ==, 0);
 382             g_assert_cmpint(backing->quiesce_counter, ==, 0);
 383             g_assert_cmpint(s->drain_count, ==, 0);
 384             g_assert_cmpint(backing_s->drain_count, ==, 0);
 385
 386             do_drain_begin(outer, bs);
 387             do_drain_begin(inner, bs);
 388
 389             g_assert_cmpint(bs->quiesce_counter, ==, 2);
 390             g_assert_cmpint(backing->quiesce_counter, ==, backing_quiesce);
 391             g_assert_cmpint(s->drain_count, ==, 2);
 392             g_assert_cmpint(backing_s->drain_count, ==, backing_quiesce);
 393
 394             do_drain_end(inner, bs);
 395             do_drain_end(outer, bs);
 396
 397             g_assert_cmpint(bs->quiesce_counter, ==, 0);
 398             g_assert_cmpint(backing->quiesce_counter, ==, 0);
 399             g_assert_cmpint(s->drain_count, ==, 0);
 400             g_assert_cmpint(backing_s->drain_count, ==, 0);
 401         }
 402     }
 403
 404     bdrv_unref(backing);
 405     bdrv_unref(bs);
 406     blk_unref(blk);
 407 }
 408
 409 static void test_multiparent(void)
 410 {
 411     BlockBackend *blk_a, *blk_b;
 412     BlockDriverState *bs_a, *bs_b, *backing;
 413     BDRVTestState *a_s, *b_s, *backing_s;
 414
 415     blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 416     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
 417                                 &error_abort);
 418     a_s = bs_a->opaque;
 419     blk_insert_bs(blk_a, bs_a, &error_abort);
 420
 421     blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 422     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
 423                                 &error_abort);
 424     b_s = bs_b->opaque;
 425     blk_insert_bs(blk_b, bs_b, &error_abort);
 426
 427     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 428     backing_s = backing->opaque;
 429     bdrv_set_backing_hd(bs_a, backing, &error_abort);
 430     bdrv_set_backing_hd(bs_b, backing, &error_abort);
 431
 432     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 433     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 434     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 435     g_assert_cmpint(a_s->drain_count, ==, 0);
 436     g_assert_cmpint(b_s->drain_count, ==, 0);
 437     g_assert_cmpint(backing_s->drain_count, ==, 0);
 438
 439     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 440
 441     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 442     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 443     g_assert_cmpint(backing->quiesce_counter, ==, 1);
 444     g_assert_cmpint(a_s->drain_count, ==, 1);
 445     g_assert_cmpint(b_s->drain_count, ==, 1);
 446     g_assert_cmpint(backing_s->drain_count, ==, 1);
 447
 448     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
 449
 450     g_assert_cmpint(bs_a->quiesce_counter, ==, 2);
 451     g_assert_cmpint(bs_b->quiesce_counter, ==, 2);
 452     g_assert_cmpint(backing->quiesce_counter, ==, 2);
 453     g_assert_cmpint(a_s->drain_count, ==, 2);
 454     g_assert_cmpint(b_s->drain_count, ==, 2);
 455     g_assert_cmpint(backing_s->drain_count, ==, 2);
 456
 457     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
 458
 459     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 460     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 461     g_assert_cmpint(backing->quiesce_counter, ==, 1);
 462     g_assert_cmpint(a_s->drain_count, ==, 1);
 463     g_assert_cmpint(b_s->drain_count, ==, 1);
 464     g_assert_cmpint(backing_s->drain_count, ==, 1);
 465
 466     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 467
 468     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 469     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 470     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 471     g_assert_cmpint(a_s->drain_count, ==, 0);
 472     g_assert_cmpint(b_s->drain_count, ==, 0);
 473     g_assert_cmpint(backing_s->drain_count, ==, 0);
 474
 475     bdrv_unref(backing);
 476     bdrv_unref(bs_a);
 477     bdrv_unref(bs_b);
 478     blk_unref(blk_a);
 479     blk_unref(blk_b);
 480 }
 481
 482 static void test_graph_change_drain_subtree(void)
 483 {
 484     BlockBackend *blk_a, *blk_b;
 485     BlockDriverState *bs_a, *bs_b, *backing;
 486     BDRVTestState *a_s, *b_s, *backing_s;
 487
 488     blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 489     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
 490                                 &error_abort);
 491     a_s = bs_a->opaque;
 492     blk_insert_bs(blk_a, bs_a, &error_abort);
 493
 494     blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 495     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
 496                                 &error_abort);
 497     b_s = bs_b->opaque;
 498     blk_insert_bs(blk_b, bs_b, &error_abort);
 499
 500     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 501     backing_s = backing->opaque;
 502     bdrv_set_backing_hd(bs_a, backing, &error_abort);
 503
 504     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 505     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 506     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 507     g_assert_cmpint(a_s->drain_count, ==, 0);
 508     g_assert_cmpint(b_s->drain_count, ==, 0);
 509     g_assert_cmpint(backing_s->drain_count, ==, 0);
 510
 511     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 512     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 513     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 514     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
 515     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
 516
 517     bdrv_set_backing_hd(bs_b, backing, &error_abort);
 518     g_assert_cmpint(bs_a->quiesce_counter, ==, 5);
 519     g_assert_cmpint(bs_b->quiesce_counter, ==, 5);
 520     g_assert_cmpint(backing->quiesce_counter, ==, 5);
 521     g_assert_cmpint(a_s->drain_count, ==, 5);
 522     g_assert_cmpint(b_s->drain_count, ==, 5);
 523     g_assert_cmpint(backing_s->drain_count, ==, 5);
 524
 525     bdrv_set_backing_hd(bs_b, NULL, &error_abort);
 526     g_assert_cmpint(bs_a->quiesce_counter, ==, 3);
 527     g_assert_cmpint(bs_b->quiesce_counter, ==, 2);
 528     g_assert_cmpint(backing->quiesce_counter, ==, 3);
 529     g_assert_cmpint(a_s->drain_count, ==, 3);
 530     g_assert_cmpint(b_s->drain_count, ==, 2);
 531     g_assert_cmpint(backing_s->drain_count, ==, 3);
 532
 533     bdrv_set_backing_hd(bs_b, backing, &error_abort);
 534     g_assert_cmpint(bs_a->quiesce_counter, ==, 5);
 535     g_assert_cmpint(bs_b->quiesce_counter, ==, 5);
 536     g_assert_cmpint(backing->quiesce_counter, ==, 5);
 537     g_assert_cmpint(a_s->drain_count, ==, 5);
 538     g_assert_cmpint(b_s->drain_count, ==, 5);
 539     g_assert_cmpint(backing_s->drain_count, ==, 5);
 540
 541     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
 542     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
 543     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 544     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 545     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 546
 547     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 548     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 549     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 550     g_assert_cmpint(a_s->drain_count, ==, 0);
 551     g_assert_cmpint(b_s->drain_count, ==, 0);
 552     g_assert_cmpint(backing_s->drain_count, ==, 0);
 553
 554     bdrv_unref(backing);
 555     bdrv_unref(bs_a);
 556     bdrv_unref(bs_b);
 557     blk_unref(blk_a);
 558     blk_unref(blk_b);
 559 }
 560
 561 static void test_graph_change_drain_all(void)
 562 {
 563     BlockBackend *blk_a, *blk_b;
 564     BlockDriverState *bs_a, *bs_b;
 565     BDRVTestState *a_s, *b_s;
 566
 567     /* Create node A with a BlockBackend */
 568     blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 569     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
 570                                 &error_abort);
 571     a_s = bs_a->opaque;
 572     blk_insert_bs(blk_a, bs_a, &error_abort);
 573
 574     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 575     g_assert_cmpint(a_s->drain_count, ==, 0);
 576
 577     /* Call bdrv_drain_all_begin() */
 578     bdrv_drain_all_begin();
 579
 580     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 581     g_assert_cmpint(a_s->drain_count, ==, 1);
 582
 583     /* Create node B with a BlockBackend */
 584     blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 585     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
 586                                 &error_abort);
 587     b_s = bs_b->opaque;
 588     blk_insert_bs(blk_b, bs_b, &error_abort);
 589
 590     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 591     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 592     g_assert_cmpint(a_s->drain_count, ==, 1);
 593     g_assert_cmpint(b_s->drain_count, ==, 1);
 594
 595     /* Unref and finally delete node A */
 596     blk_unref(blk_a);
 597
 598     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 599     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 600     g_assert_cmpint(a_s->drain_count, ==, 1);
 601     g_assert_cmpint(b_s->drain_count, ==, 1);
 602
 603     bdrv_unref(bs_a);
 604
 605     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 606     g_assert_cmpint(b_s->drain_count, ==, 1);
 607
 608     /* End the drained section */
 609     bdrv_drain_all_end();
 610
 611     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 612     g_assert_cmpint(b_s->drain_count, ==, 0);
 613
 614     bdrv_unref(bs_b);
 615     blk_unref(blk_b);
 616 }
 617
 618 struct test_iothread_data {
 619     BlockDriverState *bs;
 620     enum drain_type drain_type;
 621     int *aio_ret;
 622 };
 623
 624 static void test_iothread_drain_entry(void *opaque)
 625 {
 626     struct test_iothread_data *data = opaque;
 627
 628     aio_context_acquire(bdrv_get_aio_context(data->bs));
 629     do_drain_begin(data->drain_type, data->bs);
 630     g_assert_cmpint(*data->aio_ret, ==, 0);
 631     do_drain_end(data->drain_type, data->bs);
 632     aio_context_release(bdrv_get_aio_context(data->bs));
 633
 634     qemu_event_set(&done_event);
 635 }
 636
 637 static void test_iothread_aio_cb(void *opaque, int ret)
 638 {
 639     int *aio_ret = opaque;
 640     *aio_ret = ret;
 641     qemu_event_set(&done_event);
 642 }
 643
 644 static void test_iothread_main_thread_bh(void *opaque)
 645 {
 646     struct test_iothread_data *data = opaque;
 647
 648     /* Test that the AioContext is not yet locked in a random BH that is
 649      * executed during drain, otherwise this would deadlock. */
 650     aio_context_acquire(bdrv_get_aio_context(data->bs));
 651     bdrv_flush(data->bs);
 652     aio_context_release(bdrv_get_aio_context(data->bs));
 653 }
 654
 655 /*
 656  * Starts an AIO request on a BDS that runs in the AioContext of iothread 1.
 657  * The request involves a BH on iothread 2 before it can complete.
 658  *
 659  * @drain_thread = 0 means that do_drain_begin/end are called from the main
 660  * thread, @drain_thread = 1 means that they are called from iothread 1. Drain
 661  * for this BDS cannot be called from iothread 2 because only the main thread
 662  * may do cross-AioContext polling.
 663  */
 664 static void test_iothread_common(enum drain_type drain_type, int drain_thread)
 665 {
 666     BlockBackend *blk;
 667     BlockDriverState *bs;
 668     BDRVTestState *s;
 669     BlockAIOCB *acb;
 670     int aio_ret;
 671     struct test_iothread_data data;
 672
 673     IOThread *a = iothread_new();
 674     IOThread *b = iothread_new();
 675     AioContext *ctx_a = iothread_get_aio_context(a);
 676     AioContext *ctx_b = iothread_get_aio_context(b);
 677
 678     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
 679
 680     /* bdrv_drain_all() may only be called from the main loop thread */
 681     if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) {
 682         goto out;
 683     }
 684
 685     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 686     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 687                               &error_abort);
 688     s = bs->opaque;
 689     blk_insert_bs(blk, bs, &error_abort);
 690     blk_set_disable_request_queuing(blk, true);
 691
 692     blk_set_aio_context(blk, ctx_a, &error_abort);
 693     aio_context_acquire(ctx_a);
 694
 695     s->bh_indirection_ctx = ctx_b;
 696
 697     aio_ret = -EINPROGRESS;
 698     qemu_event_reset(&done_event);
 699
 700     if (drain_thread == 0) {
 701         acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret);
 702     } else {
 703         acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
 704     }
 705     g_assert(acb != NULL);
 706     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
 707
 708     aio_context_release(ctx_a);
 709
 710     data = (struct test_iothread_data) {
 711         .bs         = bs,
 712         .drain_type = drain_type,
 713         .aio_ret    = &aio_ret,
 714     };
 715
 716     switch (drain_thread) {
 717     case 0:
 718         if (drain_type != BDRV_DRAIN_ALL) {
 719             aio_context_acquire(ctx_a);
 720         }
 721
 722         aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data);
 723
 724         /* The request is running on the IOThread a. Draining its block device
 725          * will make sure that it has completed as far as the BDS is concerned,
 726          * but the drain in this thread can continue immediately after
 727          * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
 728          * later. */
 729         do_drain_begin(drain_type, bs);
 730         g_assert_cmpint(bs->in_flight, ==, 0);
 731
 732         if (drain_type != BDRV_DRAIN_ALL) {
 733             aio_context_release(ctx_a);
 734         }
 735         qemu_event_wait(&done_event);
 736         if (drain_type != BDRV_DRAIN_ALL) {
 737             aio_context_acquire(ctx_a);
 738         }
 739
 740         g_assert_cmpint(aio_ret, ==, 0);
 741         do_drain_end(drain_type, bs);
 742
 743         if (drain_type != BDRV_DRAIN_ALL) {
 744             aio_context_release(ctx_a);
 745         }
 746         break;
 747     case 1:
 748         aio_bh_schedule_oneshot(ctx_a, test_iothread_drain_entry, &data);
 749         qemu_event_wait(&done_event);
 750         break;
 751     default:
 752         g_assert_not_reached();
 753     }
 754
 755     aio_context_acquire(ctx_a);
 756     blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort);
 757     aio_context_release(ctx_a);
 758
 759     bdrv_unref(bs);
 760     blk_unref(blk);
 761
 762 out:
 763     iothread_join(a);
 764     iothread_join(b);
 765 }
 766
 767 static void test_iothread_drain_all(void)
 768 {
 769     test_iothread_common(BDRV_DRAIN_ALL, 0);
 770     test_iothread_common(BDRV_DRAIN_ALL, 1);
 771 }
 772
 773 static void test_iothread_drain(void)
 774 {
 775     test_iothread_common(BDRV_DRAIN, 0);
 776     test_iothread_common(BDRV_DRAIN, 1);
 777 }
 778
 779 static void test_iothread_drain_subtree(void)
 780 {
 781     test_iothread_common(BDRV_SUBTREE_DRAIN, 0);
 782     test_iothread_common(BDRV_SUBTREE_DRAIN, 1);
 783 }
 784
 785
 786 typedef struct TestBlockJob {
 787     BlockJob common;
 788     int run_ret;
 789     int prepare_ret;
 790     bool running;
 791     bool should_complete;
 792 } TestBlockJob;
 793
 794 static int test_job_prepare(Job *job)
 795 {
 796     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 797
 798     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
 799     blk_flush(s->common.blk);
 800     return s->prepare_ret;
 801 }
 802
 803 static void test_job_commit(Job *job)
 804 {
 805     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 806
 807     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
 808     blk_flush(s->common.blk);
 809 }
 810
 811 static void test_job_abort(Job *job)
 812 {
 813     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 814
 815     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
 816     blk_flush(s->common.blk);
 817 }
 818
 819 static int coroutine_fn test_job_run(Job *job, Error **errp)
 820 {
 821     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 822
 823     /* We are running the actual job code past the pause point in
 824      * job_co_entry(). */
 825     s->running = true;
 826
 827     job_transition_to_ready(&s->common.job);
 828     while (!s->should_complete) {
 829         /* Avoid job_sleep_ns() because it marks the job as !busy. We want to
 830          * emulate some actual activity (probably some I/O) here so that drain
 831          * has to wait for this activity to stop. */
 832         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000);
 833
 834         job_pause_point(&s->common.job);
 835     }
 836
 837     return s->run_ret;
 838 }
 839
 840 static void test_job_complete(Job *job, Error **errp)
 841 {
 842     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 843     s->should_complete = true;
 844 }
 845
 846 BlockJobDriver test_job_driver = {
 847     .job_driver = {
 848         .instance_size  = sizeof(TestBlockJob),
 849         .free           = block_job_free,
 850         .user_resume    = block_job_user_resume,
 851         .run            = test_job_run,
 852         .complete       = test_job_complete,
 853         .prepare        = test_job_prepare,
 854         .commit         = test_job_commit,
 855         .abort          = test_job_abort,
 856     },
 857 };
 858
 859 enum test_job_result {
 860     TEST_JOB_SUCCESS,
 861     TEST_JOB_FAIL_RUN,
 862     TEST_JOB_FAIL_PREPARE,
 863 };
 864
 865 enum test_job_drain_node {
 866     TEST_JOB_DRAIN_SRC,
 867     TEST_JOB_DRAIN_SRC_CHILD,
 868     TEST_JOB_DRAIN_SRC_PARENT,
 869 };
 870
 871 static void test_blockjob_common_drain_node(enum drain_type drain_type,
 872                                             bool use_iothread,
 873                                             enum test_job_result result,
 874                                             enum test_job_drain_node drain_node)
 875 {
 876     BlockBackend *blk_src, *blk_target;
 877     BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs;
 878     BlockJob *job;
 879     TestBlockJob *tjob;
 880     IOThread *iothread = NULL;
 881     AioContext *ctx;
 882     int ret;
 883
 884     src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR,
 885                                &error_abort);
 886     src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing",
 887                                        BDRV_O_RDWR, &error_abort);
 888     src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay",
 889                                        BDRV_O_RDWR, &error_abort);
 890
 891     bdrv_set_backing_hd(src_overlay, src, &error_abort);
 892     bdrv_unref(src);
 893     bdrv_set_backing_hd(src, src_backing, &error_abort);
 894     bdrv_unref(src_backing);
 895
 896     blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 897     blk_insert_bs(blk_src, src_overlay, &error_abort);
 898
 899     switch (drain_node) {
 900     case TEST_JOB_DRAIN_SRC:
 901         drain_bs = src;
 902         break;
 903     case TEST_JOB_DRAIN_SRC_CHILD:
 904         drain_bs = src_backing;
 905         break;
 906     case TEST_JOB_DRAIN_SRC_PARENT:
 907         drain_bs = src_overlay;
 908         break;
 909     default:
 910         g_assert_not_reached();
 911     }
 912
 913     if (use_iothread) {
 914         iothread = iothread_new();
 915         ctx = iothread_get_aio_context(iothread);
 916         blk_set_aio_context(blk_src, ctx, &error_abort);
 917     } else {
 918         ctx = qemu_get_aio_context();
 919     }
 920
 921     target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR,
 922                                   &error_abort);
 923     blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
 924     blk_insert_bs(blk_target, target, &error_abort);
 925     blk_set_allow_aio_context_change(blk_target, true);
 926
 927     aio_context_acquire(ctx);
 928     tjob = block_job_create("job0", &test_job_driver, NULL, src,
 929                             0, BLK_PERM_ALL,
 930                             0, 0, NULL, NULL, &error_abort);
 931     job = &tjob->common;
 932     block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort);
 933
 934     switch (result) {
 935     case TEST_JOB_SUCCESS:
 936         break;
 937     case TEST_JOB_FAIL_RUN:
 938         tjob->run_ret = -EIO;
 939         break;
 940     case TEST_JOB_FAIL_PREPARE:
 941         tjob->prepare_ret = -EIO;
 942         break;
 943     }
 944
 945     job_start(&job->job);
 946     aio_context_release(ctx);
 947
 948     if (use_iothread) {
 949         /* job_co_entry() is run in the I/O thread, wait for the actual job
 950          * code to start (we don't want to catch the job in the pause point in
 951          * job_co_entry(). */
 952         while (!tjob->running) {
 953             aio_poll(qemu_get_aio_context(), false);
 954         }
 955     }
 956
 957     g_assert_cmpint(job->job.pause_count, ==, 0);
 958     g_assert_false(job->job.paused);
 959     g_assert_true(tjob->running);
 960     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
 961
 962     do_drain_begin_unlocked(drain_type, drain_bs);
 963
 964     if (drain_type == BDRV_DRAIN_ALL) {
 965         /* bdrv_drain_all() drains both src and target */
 966         g_assert_cmpint(job->job.pause_count, ==, 2);
 967     } else {
 968         g_assert_cmpint(job->job.pause_count, ==, 1);
 969     }
 970     g_assert_true(job->job.paused);
 971     g_assert_false(job->job.busy); /* The job is paused */
 972
 973     do_drain_end_unlocked(drain_type, drain_bs);
 974
 975     if (use_iothread) {
 976         /* paused is reset in the I/O thread, wait for it */
 977         while (job->job.paused) {
 978             aio_poll(qemu_get_aio_context(), false);
 979         }
 980     }
 981
 982     g_assert_cmpint(job->job.pause_count, ==, 0);
 983     g_assert_false(job->job.paused);
 984     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
 985
 986     do_drain_begin_unlocked(drain_type, target);
 987
 988     if (drain_type == BDRV_DRAIN_ALL) {
 989         /* bdrv_drain_all() drains both src and target */
 990         g_assert_cmpint(job->job.pause_count, ==, 2);
 991     } else {
 992         g_assert_cmpint(job->job.pause_count, ==, 1);
 993     }
 994     g_assert_true(job->job.paused);
 995     g_assert_false(job->job.busy); /* The job is paused */
 996
 997     do_drain_end_unlocked(drain_type, target);
 998
 999     if (use_iothread) {
1000         /* paused is reset in the I/O thread, wait for it */
1001         while (job->job.paused) {
1002             aio_poll(qemu_get_aio_context(), false);
1003         }
1004     }
1005
1006     g_assert_cmpint(job->job.pause_count, ==, 0);
1007     g_assert_false(job->job.paused);
1008     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
1009
1010     aio_context_acquire(ctx);
1011     ret = job_complete_sync(&job->job, &error_abort);
1012     g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO));
1013
1014     if (use_iothread) {
1015         blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort);
1016         assert(blk_get_aio_context(blk_target) == qemu_get_aio_context());
1017     }
1018     aio_context_release(ctx);
1019
1020     blk_unref(blk_src);
1021     blk_unref(blk_target);
1022     bdrv_unref(src_overlay);
1023     bdrv_unref(target);
1024
1025     if (iothread) {
1026         iothread_join(iothread);
1027     }
1028 }
1029
1030 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread,
1031                                  enum test_job_result result)
1032 {
1033     test_blockjob_common_drain_node(drain_type, use_iothread, result,
1034                                     TEST_JOB_DRAIN_SRC);
1035     test_blockjob_common_drain_node(drain_type, use_iothread, result,
1036                                     TEST_JOB_DRAIN_SRC_CHILD);
1037     if (drain_type == BDRV_SUBTREE_DRAIN) {
1038         test_blockjob_common_drain_node(drain_type, use_iothread, result,
1039                                         TEST_JOB_DRAIN_SRC_PARENT);
1040     }
1041 }
1042
1043 static void test_blockjob_drain_all(void)
1044 {
1045     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS);
1046 }
1047
1048 static void test_blockjob_drain(void)
1049 {
1050     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS);
1051 }
1052
1053 static void test_blockjob_drain_subtree(void)
1054 {
1055     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_SUCCESS);
1056 }
1057
1058 static void test_blockjob_error_drain_all(void)
1059 {
1060     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN);
1061     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE);
1062 }
1063
1064 static void test_blockjob_error_drain(void)
1065 {
1066     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN);
1067     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE);
1068 }
1069
1070 static void test_blockjob_error_drain_subtree(void)
1071 {
1072     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_RUN);
1073     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_PREPARE);
1074 }
1075
1076 static void test_blockjob_iothread_drain_all(void)
1077 {
1078     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS);
1079 }
1080
1081 static void test_blockjob_iothread_drain(void)
1082 {
1083     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS);
1084 }
1085
1086 static void test_blockjob_iothread_drain_subtree(void)
1087 {
1088     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_SUCCESS);
1089 }
1090
1091 static void test_blockjob_iothread_error_drain_all(void)
1092 {
1093     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN);
1094     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE);
1095 }
1096
1097 static void test_blockjob_iothread_error_drain(void)
1098 {
1099     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN);
1100     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE);
1101 }
1102
1103 static void test_blockjob_iothread_error_drain_subtree(void)
1104 {
1105     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_RUN);
1106     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_PREPARE);
1107 }
1108
1109
1110 typedef struct BDRVTestTopState {
1111     BdrvChild *wait_child;
1112 } BDRVTestTopState;
1113
1114 static void bdrv_test_top_close(BlockDriverState *bs)
1115 {
1116     BdrvChild *c, *next_c;
1117     QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1118         bdrv_unref_child(bs, c);
1119     }
1120 }
1121
1122 static int coroutine_fn bdrv_test_top_co_preadv(BlockDriverState *bs,
1123                                                 uint64_t offset, uint64_t bytes,
1124                                                 QEMUIOVector *qiov, int flags)
1125 {
1126     BDRVTestTopState *tts = bs->opaque;
1127     return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags);
1128 }
1129
1130 static BlockDriver bdrv_test_top_driver = {
1131     .format_name            = "test_top_driver",
1132     .instance_size          = sizeof(BDRVTestTopState),
1133
1134     .bdrv_close             = bdrv_test_top_close,
1135     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
1136
1137     .bdrv_child_perm        = bdrv_format_default_perms,
1138 };
1139
1140 typedef struct TestCoDeleteByDrainData {
1141     BlockBackend *blk;
1142     bool detach_instead_of_delete;
1143     bool done;
1144 } TestCoDeleteByDrainData;
1145
1146 static void coroutine_fn test_co_delete_by_drain(void *opaque)
1147 {
1148     TestCoDeleteByDrainData *dbdd = opaque;
1149     BlockBackend *blk = dbdd->blk;
1150     BlockDriverState *bs = blk_bs(blk);
1151     BDRVTestTopState *tts = bs->opaque;
1152     void *buffer = g_malloc(65536);
1153     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536);
1154
1155     /* Pretend some internal write operation from parent to child.
1156      * Important: We have to read from the child, not from the parent!
1157      * Draining works by first propagating it all up the tree to the
1158      * root and then waiting for drainage from root to the leaves
1159      * (protocol nodes).  If we have a request waiting on the root,
1160      * everything will be drained before we go back down the tree, but
1161      * we do not want that.  We want to be in the middle of draining
1162      * when this following requests returns. */
1163     bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0);
1164
1165     g_assert_cmpint(bs->refcnt, ==, 1);
1166
1167     if (!dbdd->detach_instead_of_delete) {
1168         blk_unref(blk);
1169     } else {
1170         BdrvChild *c, *next_c;
1171         QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1172             bdrv_unref_child(bs, c);
1173         }
1174     }
1175
1176     dbdd->done = true;
1177     g_free(buffer);
1178 }
1179
1180 /**
1181  * Test what happens when some BDS has some children, you drain one of
1182  * them and this results in the BDS being deleted.
1183  *
1184  * If @detach_instead_of_delete is set, the BDS is not going to be
1185  * deleted but will only detach all of its children.
1186  */
1187 static void do_test_delete_by_drain(bool detach_instead_of_delete,
1188                                     enum drain_type drain_type)
1189 {
1190     BlockBackend *blk;
1191     BlockDriverState *bs, *child_bs, *null_bs;
1192     BDRVTestTopState *tts;
1193     TestCoDeleteByDrainData dbdd;
1194     Coroutine *co;
1195
1196     bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR,
1197                               &error_abort);
1198     bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1199     tts = bs->opaque;
1200
1201     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1202                         &error_abort);
1203     bdrv_attach_child(bs, null_bs, "null-child", &child_file, &error_abort);
1204
1205     /* This child will be the one to pass to requests through to, and
1206      * it will stall until a drain occurs */
1207     child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR,
1208                                     &error_abort);
1209     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1210     /* Takes our reference to child_bs */
1211     tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child", &child_file,
1212                                         &error_abort);
1213
1214     /* This child is just there to be deleted
1215      * (for detach_instead_of_delete == true) */
1216     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1217                         &error_abort);
1218     bdrv_attach_child(bs, null_bs, "null-child", &child_file, &error_abort);
1219
1220     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1221     blk_insert_bs(blk, bs, &error_abort);
1222
1223     /* Referenced by blk now */
1224     bdrv_unref(bs);
1225
1226     g_assert_cmpint(bs->refcnt, ==, 1);
1227     g_assert_cmpint(child_bs->refcnt, ==, 1);
1228     g_assert_cmpint(null_bs->refcnt, ==, 1);
1229
1230
1231     dbdd = (TestCoDeleteByDrainData){
1232         .blk = blk,
1233         .detach_instead_of_delete = detach_instead_of_delete,
1234         .done = false,
1235     };
1236     co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd);
1237     qemu_coroutine_enter(co);
1238
1239     /* Drain the child while the read operation is still pending.
1240      * This should result in the operation finishing and
1241      * test_co_delete_by_drain() resuming.  Thus, @bs will be deleted
1242      * and the coroutine will exit while this drain operation is still
1243      * in progress. */
1244     switch (drain_type) {
1245     case BDRV_DRAIN:
1246         bdrv_ref(child_bs);
1247         bdrv_drain(child_bs);
1248         bdrv_unref(child_bs);
1249         break;
1250     case BDRV_SUBTREE_DRAIN:
1251         /* Would have to ref/unref bs here for !detach_instead_of_delete, but
1252          * then the whole test becomes pointless because the graph changes
1253          * don't occur during the drain any more. */
1254         assert(detach_instead_of_delete);
1255         bdrv_subtree_drained_begin(bs);
1256         bdrv_subtree_drained_end(bs);
1257         break;
1258     case BDRV_DRAIN_ALL:
1259         bdrv_drain_all_begin();
1260         bdrv_drain_all_end();
1261         break;
1262     default:
1263         g_assert_not_reached();
1264     }
1265
1266     while (!dbdd.done) {
1267         aio_poll(qemu_get_aio_context(), true);
1268     }
1269
1270     if (detach_instead_of_delete) {
1271         /* Here, the reference has not passed over to the coroutine,
1272          * so we have to delete the BB ourselves */
1273         blk_unref(blk);
1274     }
1275 }
1276
1277 static void test_delete_by_drain(void)
1278 {
1279     do_test_delete_by_drain(false, BDRV_DRAIN);
1280 }
1281
1282 static void test_detach_by_drain_all(void)
1283 {
1284     do_test_delete_by_drain(true, BDRV_DRAIN_ALL);
1285 }
1286
1287 static void test_detach_by_drain(void)
1288 {
1289     do_test_delete_by_drain(true, BDRV_DRAIN);
1290 }
1291
1292 static void test_detach_by_drain_subtree(void)
1293 {
1294     do_test_delete_by_drain(true, BDRV_SUBTREE_DRAIN);
1295 }
1296
1297
1298 struct detach_by_parent_data {
1299     BlockDriverState *parent_b;
1300     BdrvChild *child_b;
1301     BlockDriverState *c;
1302     BdrvChild *child_c;
1303     bool by_parent_cb;
1304 };
1305 static struct detach_by_parent_data detach_by_parent_data;
1306
1307 static void detach_indirect_bh(void *opaque)
1308 {
1309     struct detach_by_parent_data *data = opaque;
1310
1311     bdrv_unref_child(data->parent_b, data->child_b);
1312
1313     bdrv_ref(data->c);
1314     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
1315                                       &child_file, &error_abort);
1316 }
1317
1318 static void detach_by_parent_aio_cb(void *opaque, int ret)
1319 {
1320     struct detach_by_parent_data *data = &detach_by_parent_data;
1321
1322     g_assert_cmpint(ret, ==, 0);
1323     if (data->by_parent_cb) {
1324         detach_indirect_bh(data);
1325     }
1326 }
1327
1328 static void detach_by_driver_cb_drained_begin(BdrvChild *child)
1329 {
1330     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1331                             detach_indirect_bh, &detach_by_parent_data);
1332     child_file.drained_begin(child);
1333 }
1334
1335 static BdrvChildRole detach_by_driver_cb_role;
1336
1337 /*
1338  * Initial graph:
1339  *
1340  * PA     PB
1341  *    \ /   \
1342  *     A     B     C
1343  *
1344  * by_parent_cb == true:  Test that parent callbacks don't poll
1345  *
1346  *     PA has a pending write request whose callback changes the child nodes of
1347  *     PB: It removes B and adds C instead. The subtree of PB is drained, which
1348  *     will indirectly drain the write request, too.
1349  *
1350  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1351  *
1352  *     PA's BdrvChildRole has a .drained_begin callback that schedules a BH
1353  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
1354  *     state is messed up, but if it is only polled in the single
1355  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1356  */
1357 static void test_detach_indirect(bool by_parent_cb)
1358 {
1359     BlockBackend *blk;
1360     BlockDriverState *parent_a, *parent_b, *a, *b, *c;
1361     BdrvChild *child_a, *child_b;
1362     BlockAIOCB *acb;
1363
1364     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
1365
1366     if (!by_parent_cb) {
1367         detach_by_driver_cb_role = child_file;
1368         detach_by_driver_cb_role.drained_begin =
1369             detach_by_driver_cb_drained_begin;
1370     }
1371
1372     /* Create all involved nodes */
1373     parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR,
1374                                     &error_abort);
1375     parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0,
1376                                     &error_abort);
1377
1378     a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort);
1379     b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort);
1380     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
1381
1382     /* blk is a BB for parent-a */
1383     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1384     blk_insert_bs(blk, parent_a, &error_abort);
1385     bdrv_unref(parent_a);
1386
1387     /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1388      * callback must not return immediately. */
1389     if (!by_parent_cb) {
1390         BDRVTestState *s = parent_a->opaque;
1391         s->sleep_in_drain_begin = true;
1392     }
1393
1394     /* Set child relationships */
1395     bdrv_ref(b);
1396     bdrv_ref(a);
1397     child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_file, &error_abort);
1398     child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_backing, &error_abort);
1399
1400     bdrv_ref(a);
1401     bdrv_attach_child(parent_a, a, "PA-A",
1402                       by_parent_cb ? &child_file : &detach_by_driver_cb_role,
1403                       &error_abort);
1404
1405     g_assert_cmpint(parent_a->refcnt, ==, 1);
1406     g_assert_cmpint(parent_b->refcnt, ==, 1);
1407     g_assert_cmpint(a->refcnt, ==, 3);
1408     g_assert_cmpint(b->refcnt, ==, 2);
1409     g_assert_cmpint(c->refcnt, ==, 1);
1410
1411     g_assert(QLIST_FIRST(&parent_b->children) == child_a);
1412     g_assert(QLIST_NEXT(child_a, next) == child_b);
1413     g_assert(QLIST_NEXT(child_b, next) == NULL);
1414
1415     /* Start the evil write request */
1416     detach_by_parent_data = (struct detach_by_parent_data) {
1417         .parent_b = parent_b,
1418         .child_b = child_b,
1419         .c = c,
1420         .by_parent_cb = by_parent_cb,
1421     };
1422     acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL);
1423     g_assert(acb != NULL);
1424
1425     /* Drain and check the expected result */
1426     bdrv_subtree_drained_begin(parent_b);
1427
1428     g_assert(detach_by_parent_data.child_c != NULL);
1429
1430     g_assert_cmpint(parent_a->refcnt, ==, 1);
1431     g_assert_cmpint(parent_b->refcnt, ==, 1);
1432     g_assert_cmpint(a->refcnt, ==, 3);
1433     g_assert_cmpint(b->refcnt, ==, 1);
1434     g_assert_cmpint(c->refcnt, ==, 2);
1435
1436     g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c);
1437     g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a);
1438     g_assert(QLIST_NEXT(child_a, next) == NULL);
1439
1440     g_assert_cmpint(parent_a->quiesce_counter, ==, 1);
1441     g_assert_cmpint(parent_b->quiesce_counter, ==, 1);
1442     g_assert_cmpint(a->quiesce_counter, ==, 1);
1443     g_assert_cmpint(b->quiesce_counter, ==, 0);
1444     g_assert_cmpint(c->quiesce_counter, ==, 1);
1445
1446     bdrv_subtree_drained_end(parent_b);
1447
1448     bdrv_unref(parent_b);
1449     blk_unref(blk);
1450
1451     g_assert_cmpint(a->refcnt, ==, 1);
1452     g_assert_cmpint(b->refcnt, ==, 1);
1453     g_assert_cmpint(c->refcnt, ==, 1);
1454     bdrv_unref(a);
1455     bdrv_unref(b);
1456     bdrv_unref(c);
1457 }
1458
1459 static void test_detach_by_parent_cb(void)
1460 {
1461     test_detach_indirect(true);
1462 }
1463
1464 static void test_detach_by_driver_cb(void)
1465 {
1466     test_detach_indirect(false);
1467 }
1468
1469 static void test_append_to_drained(void)
1470 {
1471     BlockBackend *blk;
1472     BlockDriverState *base, *overlay;
1473     BDRVTestState *base_s, *overlay_s;
1474
1475     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1476     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
1477     base_s = base->opaque;
1478     blk_insert_bs(blk, base, &error_abort);
1479
1480     overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR,
1481                                    &error_abort);
1482     overlay_s = overlay->opaque;
1483
1484     do_drain_begin(BDRV_DRAIN, base);
1485     g_assert_cmpint(base->quiesce_counter, ==, 1);
1486     g_assert_cmpint(base_s->drain_count, ==, 1);
1487     g_assert_cmpint(base->in_flight, ==, 0);
1488
1489     /* Takes ownership of overlay, so we don't have to unref it later */
1490     bdrv_append(overlay, base, &error_abort);
1491     g_assert_cmpint(base->in_flight, ==, 0);
1492     g_assert_cmpint(overlay->in_flight, ==, 0);
1493
1494     g_assert_cmpint(base->quiesce_counter, ==, 1);
1495     g_assert_cmpint(base_s->drain_count, ==, 1);
1496     g_assert_cmpint(overlay->quiesce_counter, ==, 1);
1497     g_assert_cmpint(overlay_s->drain_count, ==, 1);
1498
1499     do_drain_end(BDRV_DRAIN, base);
1500
1501     g_assert_cmpint(base->quiesce_counter, ==, 0);
1502     g_assert_cmpint(base_s->drain_count, ==, 0);
1503     g_assert_cmpint(overlay->quiesce_counter, ==, 0);
1504     g_assert_cmpint(overlay_s->drain_count, ==, 0);
1505
1506     bdrv_unref(base);
1507     blk_unref(blk);
1508 }
1509
1510 static void test_set_aio_context(void)
1511 {
1512     BlockDriverState *bs;
1513     IOThread *a = iothread_new();
1514     IOThread *b = iothread_new();
1515     AioContext *ctx_a = iothread_get_aio_context(a);
1516     AioContext *ctx_b = iothread_get_aio_context(b);
1517
1518     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
1519                               &error_abort);
1520
1521     bdrv_drained_begin(bs);
1522     bdrv_try_set_aio_context(bs, ctx_a, &error_abort);
1523
1524     aio_context_acquire(ctx_a);
1525     bdrv_drained_end(bs);
1526
1527     bdrv_drained_begin(bs);
1528     bdrv_try_set_aio_context(bs, ctx_b, &error_abort);
1529     aio_context_release(ctx_a);
1530     aio_context_acquire(ctx_b);
1531     bdrv_try_set_aio_context(bs, qemu_get_aio_context(), &error_abort);
1532     aio_context_release(ctx_b);
1533     bdrv_drained_end(bs);
1534
1535     bdrv_unref(bs);
1536     iothread_join(a);
1537     iothread_join(b);
1538 }
1539
1540
1541 typedef struct TestDropBackingBlockJob {
1542     BlockJob common;
1543     bool should_complete;
1544     bool *did_complete;
1545     BlockDriverState *detach_also;
1546 } TestDropBackingBlockJob;
1547
1548 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp)
1549 {
1550     TestDropBackingBlockJob *s =
1551         container_of(job, TestDropBackingBlockJob, common.job);
1552
1553     while (!s->should_complete) {
1554         job_sleep_ns(job, 0);
1555     }
1556
1557     return 0;
1558 }
1559
1560 static void test_drop_backing_job_commit(Job *job)
1561 {
1562     TestDropBackingBlockJob *s =
1563         container_of(job, TestDropBackingBlockJob, common.job);
1564
1565     bdrv_set_backing_hd(blk_bs(s->common.blk), NULL, &error_abort);
1566     bdrv_set_backing_hd(s->detach_also, NULL, &error_abort);
1567
1568     *s->did_complete = true;
1569 }
1570
1571 static const BlockJobDriver test_drop_backing_job_driver = {
1572     .job_driver = {
1573         .instance_size  = sizeof(TestDropBackingBlockJob),
1574         .free           = block_job_free,
1575         .user_resume    = block_job_user_resume,
1576         .run            = test_drop_backing_job_run,
1577         .commit         = test_drop_backing_job_commit,
1578     }
1579 };
1580
1581 /**
1582  * Creates a child node with three parent nodes on it, and then runs a
1583  * block job on the final one, parent-node-2.
1584  *
1585  * The job is then asked to complete before a section where the child
1586  * is drained.
1587  *
1588  * Ending this section will undrain the child's parents, first
1589  * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
1590  * list is in reverse order of how they were added.  Ending the drain
1591  * on parent-node-2 will resume the job, thus completing it and
1592  * scheduling job_exit().
1593  *
1594  * Ending the drain on parent-node-1 will poll the AioContext, which
1595  * lets job_exit() and thus test_drop_backing_job_commit() run.  That
1596  * function first removes the child as parent-node-2's backing file.
1597  *
1598  * In old (and buggy) implementations, there are two problems with
1599  * that:
1600  * (A) bdrv_drain_invoke() polls for every node that leaves the
1601  *     drained section.  This means that job_exit() is scheduled
1602  *     before the child has left the drained section.  Its
1603  *     quiesce_counter is therefore still 1 when it is removed from
1604  *     parent-node-2.
1605  *
1606  * (B) bdrv_replace_child_noperm() calls drained_end() on the old
1607  *     child's parents as many times as the child is quiesced.  This
1608  *     means it will call drained_end() on parent-node-2 once.
1609  *     Because parent-node-2 is no longer quiesced at this point, this
1610  *     will fail.
1611  *
1612  * bdrv_replace_child_noperm() therefore must call drained_end() on
1613  * the parent only if it really is still drained because the child is
1614  * drained.
1615  *
1616  * If removing child from parent-node-2 was successful (as it should
1617  * be), test_drop_backing_job_commit() will then also remove the child
1618  * from parent-node-0.
1619  *
1620  * With an old version of our drain infrastructure ((A) above), that
1621  * resulted in the following flow:
1622  *
1623  * 1. child attempts to leave its drained section.  The call recurses
1624  *    to its parents.
1625  *
1626  * 2. parent-node-2 leaves the drained section.  Polling in
1627  *    bdrv_drain_invoke() will schedule job_exit().
1628  *
1629  * 3. parent-node-1 leaves the drained section.  Polling in
1630  *    bdrv_drain_invoke() will run job_exit(), thus disconnecting
1631  *    parent-node-0 from the child node.
1632  *
1633  * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
1634  *    iterate over the parents.  Thus, it now accesses the BdrvChild
1635  *    object that used to connect parent-node-0 and the child node.
1636  *    However, that object no longer exists, so it accesses a dangling
1637  *    pointer.
1638  *
1639  * The solution is to only poll once when running a bdrv_drained_end()
1640  * operation, specifically at the end when all drained_end()
1641  * operations for all involved nodes have been scheduled.
1642  * Note that this also solves (A) above, thus hiding (B).
1643  */
1644 static void test_blockjob_commit_by_drained_end(void)
1645 {
1646     BlockDriverState *bs_child, *bs_parents[3];
1647     TestDropBackingBlockJob *job;
1648     bool job_has_completed = false;
1649     int i;
1650
1651     bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR,
1652                                     &error_abort);
1653
1654     for (i = 0; i < 3; i++) {
1655         char name[32];
1656         snprintf(name, sizeof(name), "parent-node-%i", i);
1657         bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR,
1658                                              &error_abort);
1659         bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort);
1660     }
1661
1662     job = block_job_create("job", &test_drop_backing_job_driver, NULL,
1663                            bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL,
1664                            &error_abort);
1665
1666     job->detach_also = bs_parents[0];
1667     job->did_complete = &job_has_completed;
1668
1669     job_start(&job->common.job);
1670
1671     job->should_complete = true;
1672     bdrv_drained_begin(bs_child);
1673     g_assert(!job_has_completed);
1674     bdrv_drained_end(bs_child);
1675     g_assert(job_has_completed);
1676
1677     bdrv_unref(bs_parents[0]);
1678     bdrv_unref(bs_parents[1]);
1679     bdrv_unref(bs_parents[2]);
1680     bdrv_unref(bs_child);
1681 }
1682
1683
1684 typedef struct TestSimpleBlockJob {
1685     BlockJob common;
1686     bool should_complete;
1687     bool *did_complete;
1688 } TestSimpleBlockJob;
1689
1690 static int coroutine_fn test_simple_job_run(Job *job, Error **errp)
1691 {
1692     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1693
1694     while (!s->should_complete) {
1695         job_sleep_ns(job, 0);
1696     }
1697
1698     return 0;
1699 }
1700
1701 static void test_simple_job_clean(Job *job)
1702 {
1703     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1704     *s->did_complete = true;
1705 }
1706
1707 static const BlockJobDriver test_simple_job_driver = {
1708     .job_driver = {
1709         .instance_size  = sizeof(TestSimpleBlockJob),
1710         .free           = block_job_free,
1711         .user_resume    = block_job_user_resume,
1712         .run            = test_simple_job_run,
1713         .clean          = test_simple_job_clean,
1714     },
1715 };
1716
1717 static int drop_intermediate_poll_update_filename(BdrvChild *child,
1718                                                   BlockDriverState *new_base,
1719                                                   const char *filename,
1720                                                   Error **errp)
1721 {
1722     /*
1723      * We are free to poll here, which may change the block graph, if
1724      * it is not drained.
1725      */
1726
1727     /* If the job is not drained: Complete it, schedule job_exit() */
1728     aio_poll(qemu_get_current_aio_context(), false);
1729     /* If the job is not drained: Run job_exit(), finish the job */
1730     aio_poll(qemu_get_current_aio_context(), false);
1731
1732     return 0;
1733 }
1734
1735 /**
1736  * Test a poll in the midst of bdrv_drop_intermediate().
1737  *
1738  * bdrv_drop_intermediate() calls BdrvChildRole.update_filename(),
1739  * which can yield or poll.  This may lead to graph changes, unless
1740  * the whole subtree in question is drained.
1741  *
1742  * We test this on the following graph:
1743  *
1744  *                    Job
1745  *
1746  *                     |
1747  *                  job-node
1748  *                     |
1749  *                     v
1750  *
1751  *                  job-node
1752  *
1753  *                     |
1754  *                  backing
1755  *                     |
1756  *                     v
1757  *
1758  * node-2 --chain--> node-1 --chain--> node-0
1759  *
1760  * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0).
1761  *
1762  * This first updates node-2's backing filename by invoking
1763  * drop_intermediate_poll_update_filename(), which polls twice.  This
1764  * causes the job to finish, which in turns causes the job-node to be
1765  * deleted.
1766  *
1767  * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it
1768  * already has a pointer to the BdrvChild edge between job-node and
1769  * node-1.  When it tries to handle that edge, we probably get a
1770  * segmentation fault because the object no longer exists.
1771  *
1772  *
1773  * The solution is for bdrv_drop_intermediate() to drain top's
1774  * subtree.  This prevents graph changes from happening just because
1775  * BdrvChildRole.update_filename() yields or polls.  Thus, the block
1776  * job is paused during that drained section and must finish before or
1777  * after.
1778  *
1779  * (In addition, bdrv_replace_child() must keep the job paused.)
1780  */
1781 static void test_drop_intermediate_poll(void)
1782 {
1783     static BdrvChildRole chain_child_role;
1784     BlockDriverState *chain[3];
1785     TestSimpleBlockJob *job;
1786     BlockDriverState *job_node;
1787     bool job_has_completed = false;
1788     int i;
1789     int ret;
1790
1791     chain_child_role = child_backing;
1792     chain_child_role.update_filename = drop_intermediate_poll_update_filename;
1793
1794     for (i = 0; i < 3; i++) {
1795         char name[32];
1796         snprintf(name, 32, "node-%i", i);
1797
1798         chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort);
1799     }
1800
1801     job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR,
1802                                     &error_abort);
1803     bdrv_set_backing_hd(job_node, chain[1], &error_abort);
1804
1805     /*
1806      * Establish the chain last, so the chain links are the first
1807      * elements in the BDS.parents lists
1808      */
1809     for (i = 0; i < 3; i++) {
1810         if (i) {
1811             /* Takes the reference to chain[i - 1] */
1812             chain[i]->backing = bdrv_attach_child(chain[i], chain[i - 1],
1813                                                   "chain", &chain_child_role,
1814                                                   &error_abort);
1815         }
1816     }
1817
1818     job = block_job_create("job", &test_simple_job_driver, NULL, job_node,
1819                            0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort);
1820
1821     /* The job has a reference now */
1822     bdrv_unref(job_node);
1823
1824     job->did_complete = &job_has_completed;
1825
1826     job_start(&job->common.job);
1827     job->should_complete = true;
1828
1829     g_assert(!job_has_completed);
1830     ret = bdrv_drop_intermediate(chain[1], chain[0], NULL);
1831     g_assert(ret == 0);
1832     g_assert(job_has_completed);
1833
1834     bdrv_unref(chain[2]);
1835 }
1836
1837
1838 typedef struct BDRVReplaceTestState {
1839     bool was_drained;
1840     bool was_undrained;
1841     bool has_read;
1842
1843     int drain_count;
1844
1845     bool yield_before_read;
1846     Coroutine *io_co;
1847     Coroutine *drain_co;
1848 } BDRVReplaceTestState;
1849
1850 static void bdrv_replace_test_close(BlockDriverState *bs)
1851 {
1852 }
1853
1854 /**
1855  * If @bs has a backing file:
1856  *   Yield if .yield_before_read is true (and wait for drain_begin to
1857  *   wake us up).
1858  *   Forward the read to bs->backing.  Set .has_read to true.
1859  *   If drain_begin has woken us, wake it in turn.
1860  *
1861  * Otherwise:
1862  *   Set .has_read to true and return success.
1863  */
1864 static int coroutine_fn bdrv_replace_test_co_preadv(BlockDriverState *bs,
1865                                                     uint64_t offset,
1866                                                     uint64_t bytes,
1867                                                     QEMUIOVector *qiov,
1868                                                     int flags)
1869 {
1870     BDRVReplaceTestState *s = bs->opaque;
1871
1872     if (bs->backing) {
1873         int ret;
1874
1875         g_assert(!s->drain_count);
1876
1877         s->io_co = qemu_coroutine_self();
1878         if (s->yield_before_read) {
1879             s->yield_before_read = false;
1880             qemu_coroutine_yield();
1881         }
1882         s->io_co = NULL;
1883
1884         ret = bdrv_preadv(bs->backing, offset, qiov);
1885         s->has_read = true;
1886
1887         /* Wake up drain_co if it runs */
1888         if (s->drain_co) {
1889             aio_co_wake(s->drain_co);
1890         }
1891
1892         return ret;
1893     }
1894
1895     s->has_read = true;
1896     return 0;
1897 }
1898
1899 /**
1900  * If .drain_count is 0, wake up .io_co if there is one; and set
1901  * .was_drained.
1902  * Increment .drain_count.
1903  */
1904 static void coroutine_fn bdrv_replace_test_co_drain_begin(BlockDriverState *bs)
1905 {
1906     BDRVReplaceTestState *s = bs->opaque;
1907
1908     if (!s->drain_count) {
1909         /* Keep waking io_co up until it is done */
1910         s->drain_co = qemu_coroutine_self();
1911         while (s->io_co) {
1912             aio_co_wake(s->io_co);
1913             s->io_co = NULL;
1914             qemu_coroutine_yield();
1915         }
1916         s->drain_co = NULL;
1917
1918         s->was_drained = true;
1919     }
1920     s->drain_count++;
1921 }
1922
1923 /**
1924  * Reduce .drain_count, set .was_undrained once it reaches 0.
1925  * If .drain_count reaches 0 and the node has a backing file, issue a
1926  * read request.
1927  */
1928 static void coroutine_fn bdrv_replace_test_co_drain_end(BlockDriverState *bs)
1929 {
1930     BDRVReplaceTestState *s = bs->opaque;
1931
1932     g_assert(s->drain_count > 0);
1933     if (!--s->drain_count) {
1934         int ret;
1935
1936         s->was_undrained = true;
1937
1938         if (bs->backing) {
1939             char data;
1940             QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
1941
1942             /* Queue a read request post-drain */
1943             ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
1944             g_assert(ret >= 0);
1945         }
1946     }
1947 }
1948
1949 static BlockDriver bdrv_replace_test = {
1950     .format_name            = "replace_test",
1951     .instance_size          = sizeof(BDRVReplaceTestState),
1952
1953     .bdrv_close             = bdrv_replace_test_close,
1954     .bdrv_co_preadv         = bdrv_replace_test_co_preadv,
1955
1956     .bdrv_co_drain_begin    = bdrv_replace_test_co_drain_begin,
1957     .bdrv_co_drain_end      = bdrv_replace_test_co_drain_end,
1958
1959     .bdrv_child_perm        = bdrv_format_default_perms,
1960 };
1961
1962 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque)
1963 {
1964     int ret;
1965     char data;
1966
1967     ret = blk_co_pread(opaque, 0, 1, &data, 0);
1968     g_assert(ret >= 0);
1969 }
1970
1971 /**
1972  * We test two things:
1973  * (1) bdrv_replace_child_noperm() must not undrain the parent if both
1974  *     children are drained.
1975  * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
1976  *     drained child.  If the old child is drained, it must flush I/O
1977  *     requests after the new one has been attached.  If the new child
1978  *     is drained, it must flush I/O requests before the old one is
1979  *     detached.
1980  *
1981  * To do so, we create one parent node and two child nodes; then
1982  * attach one of the children (old_child_bs) to the parent, then
1983  * drain both old_child_bs and new_child_bs according to
1984  * old_drain_count and new_drain_count, respectively, and finally
1985  * we invoke bdrv_replace_node() to replace old_child_bs by
1986  * new_child_bs.
1987  *
1988  * The test block driver we use here (bdrv_replace_test) has a read
1989  * function that:
1990  * - For the parent node, can optionally yield, and then forwards the
1991  *   read to bdrv_preadv(),
1992  * - For the child node, just returns immediately.
1993  *
1994  * If the read yields, the drain_begin function will wake it up.
1995  *
1996  * The drain_end function issues a read on the parent once it is fully
1997  * undrained (which simulates requests starting to come in again).
1998  */
1999 static void do_test_replace_child_mid_drain(int old_drain_count,
2000                                             int new_drain_count)
2001 {
2002     BlockBackend *parent_blk;
2003     BlockDriverState *parent_bs;
2004     BlockDriverState *old_child_bs, *new_child_bs;
2005     BDRVReplaceTestState *parent_s;
2006     BDRVReplaceTestState *old_child_s, *new_child_s;
2007     Coroutine *io_co;
2008     int i;
2009
2010     parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0,
2011                                      &error_abort);
2012     parent_s = parent_bs->opaque;
2013
2014     parent_blk = blk_new(qemu_get_aio_context(),
2015                          BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
2016     blk_insert_bs(parent_blk, parent_bs, &error_abort);
2017
2018     old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0,
2019                                         &error_abort);
2020     new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0,
2021                                         &error_abort);
2022     old_child_s = old_child_bs->opaque;
2023     new_child_s = new_child_bs->opaque;
2024
2025     /* So that we can read something */
2026     parent_bs->total_sectors = 1;
2027     old_child_bs->total_sectors = 1;
2028     new_child_bs->total_sectors = 1;
2029
2030     bdrv_ref(old_child_bs);
2031     parent_bs->backing = bdrv_attach_child(parent_bs, old_child_bs, "child",
2032                                            &child_backing, &error_abort);
2033
2034     for (i = 0; i < old_drain_count; i++) {
2035         bdrv_drained_begin(old_child_bs);
2036     }
2037     for (i = 0; i < new_drain_count; i++) {
2038         bdrv_drained_begin(new_child_bs);
2039     }
2040
2041     if (!old_drain_count) {
2042         /*
2043          * Start a read operation that will yield, so it will not
2044          * complete before the node is drained.
2045          */
2046         parent_s->yield_before_read = true;
2047         io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co,
2048                                       parent_blk);
2049         qemu_coroutine_enter(io_co);
2050     }
2051
2052     /* If we have started a read operation, it should have yielded */
2053     g_assert(!parent_s->has_read);
2054
2055     /* Reset drained status so we can see what bdrv_replace_node() does */
2056     parent_s->was_drained = false;
2057     parent_s->was_undrained = false;
2058
2059     g_assert(parent_bs->quiesce_counter == old_drain_count);
2060     bdrv_replace_node(old_child_bs, new_child_bs, &error_abort);
2061     g_assert(parent_bs->quiesce_counter == new_drain_count);
2062
2063     if (!old_drain_count && !new_drain_count) {
2064         /*
2065          * From undrained to undrained drains and undrains the parent,
2066          * because bdrv_replace_node() contains a drained section for
2067          * @old_child_bs.
2068          */
2069         g_assert(parent_s->was_drained && parent_s->was_undrained);
2070     } else if (!old_drain_count && new_drain_count) {
2071         /*
2072          * From undrained to drained should drain the parent and keep
2073          * it that way.
2074          */
2075         g_assert(parent_s->was_drained && !parent_s->was_undrained);
2076     } else if (old_drain_count && !new_drain_count) {
2077         /*
2078          * From drained to undrained should undrain the parent and
2079          * keep it that way.
2080          */
2081         g_assert(!parent_s->was_drained && parent_s->was_undrained);
2082     } else /* if (old_drain_count && new_drain_count) */ {
2083         /*
2084          * From drained to drained must not undrain the parent at any
2085          * point
2086          */
2087         g_assert(!parent_s->was_drained && !parent_s->was_undrained);
2088     }
2089
2090     if (!old_drain_count || !new_drain_count) {
2091         /*
2092          * If !old_drain_count, we have started a read request before
2093          * bdrv_replace_node().  If !new_drain_count, the parent must
2094          * have been undrained at some point, and
2095          * bdrv_replace_test_co_drain_end() starts a read request
2096          * then.
2097          */
2098         g_assert(parent_s->has_read);
2099     } else {
2100         /*
2101          * If the parent was never undrained, there is no way to start
2102          * a read request.
2103          */
2104         g_assert(!parent_s->has_read);
2105     }
2106
2107     /* A drained child must have not received any request */
2108     g_assert(!(old_drain_count && old_child_s->has_read));
2109     g_assert(!(new_drain_count && new_child_s->has_read));
2110
2111     for (i = 0; i < new_drain_count; i++) {
2112         bdrv_drained_end(new_child_bs);
2113     }
2114     for (i = 0; i < old_drain_count; i++) {
2115         bdrv_drained_end(old_child_bs);
2116     }
2117
2118     /*
2119      * By now, bdrv_replace_test_co_drain_end() must have been called
2120      * at some point while the new child was attached to the parent.
2121      */
2122     g_assert(parent_s->has_read);
2123     g_assert(new_child_s->has_read);
2124
2125     blk_unref(parent_blk);
2126     bdrv_unref(parent_bs);
2127     bdrv_unref(old_child_bs);
2128     bdrv_unref(new_child_bs);
2129 }
2130
2131 static void test_replace_child_mid_drain(void)
2132 {
2133     int old_drain_count, new_drain_count;
2134
2135     for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) {
2136         for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) {
2137             do_test_replace_child_mid_drain(old_drain_count, new_drain_count);
2138         }
2139     }
2140 }
2141
2142 int main(int argc, char **argv)
2143 {
2144     int ret;
2145
2146     bdrv_init();
2147     qemu_init_main_loop(&error_abort);
2148
2149     g_test_init(&argc, &argv, NULL);
2150     qemu_event_init(&done_event, false);
2151
2152     g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all);
2153     g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain);
2154     g_test_add_func("/bdrv-drain/driver-cb/drain_subtree",
2155                     test_drv_cb_drain_subtree);
2156
2157     g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
2158                     test_drv_cb_co_drain_all);
2159     g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain);
2160     g_test_add_func("/bdrv-drain/driver-cb/co/drain_subtree",
2161                     test_drv_cb_co_drain_subtree);
2162
2163
2164     g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all);
2165     g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain);
2166     g_test_add_func("/bdrv-drain/quiesce/drain_subtree",
2167                     test_quiesce_drain_subtree);
2168
2169     g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
2170                     test_quiesce_co_drain_all);
2171     g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain);
2172     g_test_add_func("/bdrv-drain/quiesce/co/drain_subtree",
2173                     test_quiesce_co_drain_subtree);
2174
2175     g_test_add_func("/bdrv-drain/nested", test_nested);
2176     g_test_add_func("/bdrv-drain/multiparent", test_multiparent);
2177
2178     g_test_add_func("/bdrv-drain/graph-change/drain_subtree",
2179                     test_graph_change_drain_subtree);
2180     g_test_add_func("/bdrv-drain/graph-change/drain_all",
2181                     test_graph_change_drain_all);
2182
2183     g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all);
2184     g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain);
2185     g_test_add_func("/bdrv-drain/iothread/drain_subtree",
2186                     test_iothread_drain_subtree);
2187
2188     g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all);
2189     g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain);
2190     g_test_add_func("/bdrv-drain/blockjob/drain_subtree",
2191                     test_blockjob_drain_subtree);
2192
2193     g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
2194                     test_blockjob_error_drain_all);
2195     g_test_add_func("/bdrv-drain/blockjob/error/drain",
2196                     test_blockjob_error_drain);
2197     g_test_add_func("/bdrv-drain/blockjob/error/drain_subtree",
2198                     test_blockjob_error_drain_subtree);
2199
2200     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
2201                     test_blockjob_iothread_drain_all);
2202     g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
2203                     test_blockjob_iothread_drain);
2204     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_subtree",
2205                     test_blockjob_iothread_drain_subtree);
2206
2207     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
2208                     test_blockjob_iothread_error_drain_all);
2209     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
2210                     test_blockjob_iothread_error_drain);
2211     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_subtree",
2212                     test_blockjob_iothread_error_drain_subtree);
2213
2214     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
2215     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
2216     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
2217     g_test_add_func("/bdrv-drain/detach/drain_subtree", test_detach_by_drain_subtree);
2218     g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb);
2219     g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb);
2220
2221     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
2222
2223     g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context);
2224
2225     g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
2226                     test_blockjob_commit_by_drained_end);
2227
2228     g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
2229                     test_drop_intermediate_poll);
2230
2231     g_test_add_func("/bdrv-drain/replace_child/mid-drain",
2232                     test_replace_child_mid_drain);
2233
2234     ret = g_test_run();
2235     qemu_event_destroy(&done_event);
2236     return ret;
2237 }