tests/test-bdrv-drain.c

   1 /*
   2  * Block node draining tests
   3  *
   4  * Copyright (c) 2017 Kevin Wolf <[email protected]>
   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 "iothread.h"
  31
  32 static QemuEvent done_event;
  33
  34 typedef struct BDRVTestState {
  35     int drain_count;
  36     AioContext *bh_indirection_ctx;
  37     bool sleep_in_drain_begin;
  38 } BDRVTestState;
  39
  40 static void coroutine_fn bdrv_test_co_drain_begin(BlockDriverState *bs)
  41 {
  42     BDRVTestState *s = bs->opaque;
  43     s->drain_count++;
  44     if (s->sleep_in_drain_begin) {
  45         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
  46     }
  47 }
  48
  49 static void coroutine_fn bdrv_test_co_drain_end(BlockDriverState *bs)
  50 {
  51     BDRVTestState *s = bs->opaque;
  52     s->drain_count--;
  53 }
  54
  55 static void bdrv_test_close(BlockDriverState *bs)
  56 {
  57     BDRVTestState *s = bs->opaque;
  58     g_assert_cmpint(s->drain_count, >, 0);
  59 }
  60
  61 static void co_reenter_bh(void *opaque)
  62 {
  63     aio_co_wake(opaque);
  64 }
  65
  66 static int coroutine_fn bdrv_test_co_preadv(BlockDriverState *bs,
  67                                             uint64_t offset, uint64_t bytes,
  68                                             QEMUIOVector *qiov, int flags)
  69 {
  70     BDRVTestState *s = bs->opaque;
  71
  72     /* We want this request to stay until the polling loop in drain waits for
  73      * it to complete. We need to sleep a while as bdrv_drain_invoke() comes
  74      * first and polls its result, too, but it shouldn't accidentally complete
  75      * this request yet. */
  76     qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
  77
  78     if (s->bh_indirection_ctx) {
  79         aio_bh_schedule_oneshot(s->bh_indirection_ctx, co_reenter_bh,
  80                                 qemu_coroutine_self());
  81         qemu_coroutine_yield();
  82     }
  83
  84     return 0;
  85 }
  86
  87 static void bdrv_test_child_perm(BlockDriverState *bs, BdrvChild *c,
  88                                  const BdrvChildRole *role,
  89                                  BlockReopenQueue *reopen_queue,
  90                                  uint64_t perm, uint64_t shared,
  91                                  uint64_t *nperm, uint64_t *nshared)
  92 {
  93     /* bdrv_format_default_perms() accepts only these two, so disguise
  94      * detach_by_driver_cb_role as one of them. */
  95     if (role != &child_file && role != &child_backing) {
  96         role = &child_file;
  97     }
  98
  99     bdrv_format_default_perms(bs, c, role, reopen_queue, perm, shared,
 100                               nperm, nshared);
 101 }
 102
 103 static BlockDriver bdrv_test = {
 104     .format_name            = "test",
 105     .instance_size          = sizeof(BDRVTestState),
 106
 107     .bdrv_close             = bdrv_test_close,
 108     .bdrv_co_preadv         = bdrv_test_co_preadv,
 109
 110     .bdrv_co_drain_begin    = bdrv_test_co_drain_begin,
 111     .bdrv_co_drain_end      = bdrv_test_co_drain_end,
 112
 113     .bdrv_child_perm        = bdrv_test_child_perm,
 114 };
 115
 116 static void aio_ret_cb(void *opaque, int ret)
 117 {
 118     int *aio_ret = opaque;
 119     *aio_ret = ret;
 120 }
 121
 122 typedef struct CallInCoroutineData {
 123     void (*entry)(void);
 124     bool done;
 125 } CallInCoroutineData;
 126
 127 static coroutine_fn void call_in_coroutine_entry(void *opaque)
 128 {
 129     CallInCoroutineData *data = opaque;
 130
 131     data->entry();
 132     data->done = true;
 133 }
 134
 135 static void call_in_coroutine(void (*entry)(void))
 136 {
 137     Coroutine *co;
 138     CallInCoroutineData data = {
 139         .entry  = entry,
 140         .done   = false,
 141     };
 142
 143     co = qemu_coroutine_create(call_in_coroutine_entry, &data);
 144     qemu_coroutine_enter(co);
 145     while (!data.done) {
 146         aio_poll(qemu_get_aio_context(), true);
 147     }
 148 }
 149
 150 enum drain_type {
 151     BDRV_DRAIN_ALL,
 152     BDRV_DRAIN,
 153     BDRV_SUBTREE_DRAIN,
 154     DRAIN_TYPE_MAX,
 155 };
 156
 157 static void do_drain_begin(enum drain_type drain_type, BlockDriverState *bs)
 158 {
 159     switch (drain_type) {
 160     case BDRV_DRAIN_ALL:        bdrv_drain_all_begin(); break;
 161     case BDRV_DRAIN:            bdrv_drained_begin(bs); break;
 162     case BDRV_SUBTREE_DRAIN:    bdrv_subtree_drained_begin(bs); break;
 163     default:                    g_assert_not_reached();
 164     }
 165 }
 166
 167 static void do_drain_end(enum drain_type drain_type, BlockDriverState *bs)
 168 {
 169     switch (drain_type) {
 170     case BDRV_DRAIN_ALL:        bdrv_drain_all_end(); break;
 171     case BDRV_DRAIN:            bdrv_drained_end(bs); break;
 172     case BDRV_SUBTREE_DRAIN:    bdrv_subtree_drained_end(bs); break;
 173     default:                    g_assert_not_reached();
 174     }
 175 }
 176
 177 static void do_drain_begin_unlocked(enum drain_type drain_type, BlockDriverState *bs)
 178 {
 179     if (drain_type != BDRV_DRAIN_ALL) {
 180         aio_context_acquire(bdrv_get_aio_context(bs));
 181     }
 182     do_drain_begin(drain_type, bs);
 183     if (drain_type != BDRV_DRAIN_ALL) {
 184         aio_context_release(bdrv_get_aio_context(bs));
 185     }
 186 }
 187
 188 static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs)
 189 {
 190     if (drain_type != BDRV_DRAIN_ALL) {
 191         aio_context_acquire(bdrv_get_aio_context(bs));
 192     }
 193     do_drain_end(drain_type, bs);
 194     if (drain_type != BDRV_DRAIN_ALL) {
 195         aio_context_release(bdrv_get_aio_context(bs));
 196     }
 197 }
 198
 199 static void test_drv_cb_common(enum drain_type drain_type, bool recursive)
 200 {
 201     BlockBackend *blk;
 202     BlockDriverState *bs, *backing;
 203     BDRVTestState *s, *backing_s;
 204     BlockAIOCB *acb;
 205     int aio_ret;
 206
 207     QEMUIOVector qiov;
 208     struct iovec iov = {
 209         .iov_base = NULL,
 210         .iov_len = 0,
 211     };
 212     qemu_iovec_init_external(&qiov, &iov, 1);
 213
 214     blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 215     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 216                               &error_abort);
 217     s = bs->opaque;
 218     blk_insert_bs(blk, bs, &error_abort);
 219
 220     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 221     backing_s = backing->opaque;
 222     bdrv_set_backing_hd(bs, backing, &error_abort);
 223
 224     /* Simple bdrv_drain_all_begin/end pair, check that CBs are called */
 225     g_assert_cmpint(s->drain_count, ==, 0);
 226     g_assert_cmpint(backing_s->drain_count, ==, 0);
 227
 228     do_drain_begin(drain_type, bs);
 229
 230     g_assert_cmpint(s->drain_count, ==, 1);
 231     g_assert_cmpint(backing_s->drain_count, ==, !!recursive);
 232
 233     do_drain_end(drain_type, bs);
 234
 235     g_assert_cmpint(s->drain_count, ==, 0);
 236     g_assert_cmpint(backing_s->drain_count, ==, 0);
 237
 238     /* Now do the same while a request is pending */
 239     aio_ret = -EINPROGRESS;
 240     acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
 241     g_assert(acb != NULL);
 242     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
 243
 244     g_assert_cmpint(s->drain_count, ==, 0);
 245     g_assert_cmpint(backing_s->drain_count, ==, 0);
 246
 247     do_drain_begin(drain_type, bs);
 248
 249     g_assert_cmpint(aio_ret, ==, 0);
 250     g_assert_cmpint(s->drain_count, ==, 1);
 251     g_assert_cmpint(backing_s->drain_count, ==, !!recursive);
 252
 253     do_drain_end(drain_type, bs);
 254
 255     g_assert_cmpint(s->drain_count, ==, 0);
 256     g_assert_cmpint(backing_s->drain_count, ==, 0);
 257
 258     bdrv_unref(backing);
 259     bdrv_unref(bs);
 260     blk_unref(blk);
 261 }
 262
 263 static void test_drv_cb_drain_all(void)
 264 {
 265     test_drv_cb_common(BDRV_DRAIN_ALL, true);
 266 }
 267
 268 static void test_drv_cb_drain(void)
 269 {
 270     test_drv_cb_common(BDRV_DRAIN, false);
 271 }
 272
 273 static void test_drv_cb_drain_subtree(void)
 274 {
 275     test_drv_cb_common(BDRV_SUBTREE_DRAIN, true);
 276 }
 277
 278 static void test_drv_cb_co_drain_all(void)
 279 {
 280     call_in_coroutine(test_drv_cb_drain_all);
 281 }
 282
 283 static void test_drv_cb_co_drain(void)
 284 {
 285     call_in_coroutine(test_drv_cb_drain);
 286 }
 287
 288 static void test_drv_cb_co_drain_subtree(void)
 289 {
 290     call_in_coroutine(test_drv_cb_drain_subtree);
 291 }
 292
 293 static void test_quiesce_common(enum drain_type drain_type, bool recursive)
 294 {
 295     BlockBackend *blk;
 296     BlockDriverState *bs, *backing;
 297
 298     blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 299     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 300                               &error_abort);
 301     blk_insert_bs(blk, bs, &error_abort);
 302
 303     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 304     bdrv_set_backing_hd(bs, backing, &error_abort);
 305
 306     g_assert_cmpint(bs->quiesce_counter, ==, 0);
 307     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 308
 309     do_drain_begin(drain_type, bs);
 310
 311     g_assert_cmpint(bs->quiesce_counter, ==, 1);
 312     g_assert_cmpint(backing->quiesce_counter, ==, !!recursive);
 313
 314     do_drain_end(drain_type, bs);
 315
 316     g_assert_cmpint(bs->quiesce_counter, ==, 0);
 317     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 318
 319     bdrv_unref(backing);
 320     bdrv_unref(bs);
 321     blk_unref(blk);
 322 }
 323
 324 static void test_quiesce_drain_all(void)
 325 {
 326     test_quiesce_common(BDRV_DRAIN_ALL, true);
 327 }
 328
 329 static void test_quiesce_drain(void)
 330 {
 331     test_quiesce_common(BDRV_DRAIN, false);
 332 }
 333
 334 static void test_quiesce_drain_subtree(void)
 335 {
 336     test_quiesce_common(BDRV_SUBTREE_DRAIN, true);
 337 }
 338
 339 static void test_quiesce_co_drain_all(void)
 340 {
 341     call_in_coroutine(test_quiesce_drain_all);
 342 }
 343
 344 static void test_quiesce_co_drain(void)
 345 {
 346     call_in_coroutine(test_quiesce_drain);
 347 }
 348
 349 static void test_quiesce_co_drain_subtree(void)
 350 {
 351     call_in_coroutine(test_quiesce_drain_subtree);
 352 }
 353
 354 static void test_nested(void)
 355 {
 356     BlockBackend *blk;
 357     BlockDriverState *bs, *backing;
 358     BDRVTestState *s, *backing_s;
 359     enum drain_type outer, inner;
 360
 361     blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 362     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 363                               &error_abort);
 364     s = bs->opaque;
 365     blk_insert_bs(blk, bs, &error_abort);
 366
 367     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 368     backing_s = backing->opaque;
 369     bdrv_set_backing_hd(bs, backing, &error_abort);
 370
 371     for (outer = 0; outer < DRAIN_TYPE_MAX; outer++) {
 372         for (inner = 0; inner < DRAIN_TYPE_MAX; inner++) {
 373             int backing_quiesce = (outer != BDRV_DRAIN) +
 374                                   (inner != BDRV_DRAIN);
 375
 376             g_assert_cmpint(bs->quiesce_counter, ==, 0);
 377             g_assert_cmpint(backing->quiesce_counter, ==, 0);
 378             g_assert_cmpint(s->drain_count, ==, 0);
 379             g_assert_cmpint(backing_s->drain_count, ==, 0);
 380
 381             do_drain_begin(outer, bs);
 382             do_drain_begin(inner, bs);
 383
 384             g_assert_cmpint(bs->quiesce_counter, ==, 2);
 385             g_assert_cmpint(backing->quiesce_counter, ==, backing_quiesce);
 386             g_assert_cmpint(s->drain_count, ==, 2);
 387             g_assert_cmpint(backing_s->drain_count, ==, backing_quiesce);
 388
 389             do_drain_end(inner, bs);
 390             do_drain_end(outer, bs);
 391
 392             g_assert_cmpint(bs->quiesce_counter, ==, 0);
 393             g_assert_cmpint(backing->quiesce_counter, ==, 0);
 394             g_assert_cmpint(s->drain_count, ==, 0);
 395             g_assert_cmpint(backing_s->drain_count, ==, 0);
 396         }
 397     }
 398
 399     bdrv_unref(backing);
 400     bdrv_unref(bs);
 401     blk_unref(blk);
 402 }
 403
 404 static void test_multiparent(void)
 405 {
 406     BlockBackend *blk_a, *blk_b;
 407     BlockDriverState *bs_a, *bs_b, *backing;
 408     BDRVTestState *a_s, *b_s, *backing_s;
 409
 410     blk_a = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 411     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
 412                                 &error_abort);
 413     a_s = bs_a->opaque;
 414     blk_insert_bs(blk_a, bs_a, &error_abort);
 415
 416     blk_b = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 417     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
 418                                 &error_abort);
 419     b_s = bs_b->opaque;
 420     blk_insert_bs(blk_b, bs_b, &error_abort);
 421
 422     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 423     backing_s = backing->opaque;
 424     bdrv_set_backing_hd(bs_a, backing, &error_abort);
 425     bdrv_set_backing_hd(bs_b, backing, &error_abort);
 426
 427     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 428     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 429     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 430     g_assert_cmpint(a_s->drain_count, ==, 0);
 431     g_assert_cmpint(b_s->drain_count, ==, 0);
 432     g_assert_cmpint(backing_s->drain_count, ==, 0);
 433
 434     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 435
 436     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 437     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 438     g_assert_cmpint(backing->quiesce_counter, ==, 1);
 439     g_assert_cmpint(a_s->drain_count, ==, 1);
 440     g_assert_cmpint(b_s->drain_count, ==, 1);
 441     g_assert_cmpint(backing_s->drain_count, ==, 1);
 442
 443     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
 444
 445     g_assert_cmpint(bs_a->quiesce_counter, ==, 2);
 446     g_assert_cmpint(bs_b->quiesce_counter, ==, 2);
 447     g_assert_cmpint(backing->quiesce_counter, ==, 2);
 448     g_assert_cmpint(a_s->drain_count, ==, 2);
 449     g_assert_cmpint(b_s->drain_count, ==, 2);
 450     g_assert_cmpint(backing_s->drain_count, ==, 2);
 451
 452     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
 453
 454     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 455     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 456     g_assert_cmpint(backing->quiesce_counter, ==, 1);
 457     g_assert_cmpint(a_s->drain_count, ==, 1);
 458     g_assert_cmpint(b_s->drain_count, ==, 1);
 459     g_assert_cmpint(backing_s->drain_count, ==, 1);
 460
 461     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 462
 463     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 464     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 465     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 466     g_assert_cmpint(a_s->drain_count, ==, 0);
 467     g_assert_cmpint(b_s->drain_count, ==, 0);
 468     g_assert_cmpint(backing_s->drain_count, ==, 0);
 469
 470     bdrv_unref(backing);
 471     bdrv_unref(bs_a);
 472     bdrv_unref(bs_b);
 473     blk_unref(blk_a);
 474     blk_unref(blk_b);
 475 }
 476
 477 static void test_graph_change_drain_subtree(void)
 478 {
 479     BlockBackend *blk_a, *blk_b;
 480     BlockDriverState *bs_a, *bs_b, *backing;
 481     BDRVTestState *a_s, *b_s, *backing_s;
 482
 483     blk_a = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 484     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
 485                                 &error_abort);
 486     a_s = bs_a->opaque;
 487     blk_insert_bs(blk_a, bs_a, &error_abort);
 488
 489     blk_b = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 490     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
 491                                 &error_abort);
 492     b_s = bs_b->opaque;
 493     blk_insert_bs(blk_b, bs_b, &error_abort);
 494
 495     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
 496     backing_s = backing->opaque;
 497     bdrv_set_backing_hd(bs_a, backing, &error_abort);
 498
 499     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 500     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 501     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 502     g_assert_cmpint(a_s->drain_count, ==, 0);
 503     g_assert_cmpint(b_s->drain_count, ==, 0);
 504     g_assert_cmpint(backing_s->drain_count, ==, 0);
 505
 506     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 507     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 508     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
 509     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
 510     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
 511
 512     bdrv_set_backing_hd(bs_b, backing, &error_abort);
 513     g_assert_cmpint(bs_a->quiesce_counter, ==, 5);
 514     g_assert_cmpint(bs_b->quiesce_counter, ==, 5);
 515     g_assert_cmpint(backing->quiesce_counter, ==, 5);
 516     g_assert_cmpint(a_s->drain_count, ==, 5);
 517     g_assert_cmpint(b_s->drain_count, ==, 5);
 518     g_assert_cmpint(backing_s->drain_count, ==, 5);
 519
 520     bdrv_set_backing_hd(bs_b, NULL, &error_abort);
 521     g_assert_cmpint(bs_a->quiesce_counter, ==, 3);
 522     g_assert_cmpint(bs_b->quiesce_counter, ==, 2);
 523     g_assert_cmpint(backing->quiesce_counter, ==, 3);
 524     g_assert_cmpint(a_s->drain_count, ==, 3);
 525     g_assert_cmpint(b_s->drain_count, ==, 2);
 526     g_assert_cmpint(backing_s->drain_count, ==, 3);
 527
 528     bdrv_set_backing_hd(bs_b, backing, &error_abort);
 529     g_assert_cmpint(bs_a->quiesce_counter, ==, 5);
 530     g_assert_cmpint(bs_b->quiesce_counter, ==, 5);
 531     g_assert_cmpint(backing->quiesce_counter, ==, 5);
 532     g_assert_cmpint(a_s->drain_count, ==, 5);
 533     g_assert_cmpint(b_s->drain_count, ==, 5);
 534     g_assert_cmpint(backing_s->drain_count, ==, 5);
 535
 536     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
 537     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
 538     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 539     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 540     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
 541
 542     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 543     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 544     g_assert_cmpint(backing->quiesce_counter, ==, 0);
 545     g_assert_cmpint(a_s->drain_count, ==, 0);
 546     g_assert_cmpint(b_s->drain_count, ==, 0);
 547     g_assert_cmpint(backing_s->drain_count, ==, 0);
 548
 549     bdrv_unref(backing);
 550     bdrv_unref(bs_a);
 551     bdrv_unref(bs_b);
 552     blk_unref(blk_a);
 553     blk_unref(blk_b);
 554 }
 555
 556 static void test_graph_change_drain_all(void)
 557 {
 558     BlockBackend *blk_a, *blk_b;
 559     BlockDriverState *bs_a, *bs_b;
 560     BDRVTestState *a_s, *b_s;
 561
 562     /* Create node A with a BlockBackend */
 563     blk_a = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 564     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
 565                                 &error_abort);
 566     a_s = bs_a->opaque;
 567     blk_insert_bs(blk_a, bs_a, &error_abort);
 568
 569     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
 570     g_assert_cmpint(a_s->drain_count, ==, 0);
 571
 572     /* Call bdrv_drain_all_begin() */
 573     bdrv_drain_all_begin();
 574
 575     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 576     g_assert_cmpint(a_s->drain_count, ==, 1);
 577
 578     /* Create node B with a BlockBackend */
 579     blk_b = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 580     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
 581                                 &error_abort);
 582     b_s = bs_b->opaque;
 583     blk_insert_bs(blk_b, bs_b, &error_abort);
 584
 585     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 586     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 587     g_assert_cmpint(a_s->drain_count, ==, 1);
 588     g_assert_cmpint(b_s->drain_count, ==, 1);
 589
 590     /* Unref and finally delete node A */
 591     blk_unref(blk_a);
 592
 593     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
 594     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 595     g_assert_cmpint(a_s->drain_count, ==, 1);
 596     g_assert_cmpint(b_s->drain_count, ==, 1);
 597
 598     bdrv_unref(bs_a);
 599
 600     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
 601     g_assert_cmpint(b_s->drain_count, ==, 1);
 602
 603     /* End the drained section */
 604     bdrv_drain_all_end();
 605
 606     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
 607     g_assert_cmpint(b_s->drain_count, ==, 0);
 608
 609     bdrv_unref(bs_b);
 610     blk_unref(blk_b);
 611 }
 612
 613 struct test_iothread_data {
 614     BlockDriverState *bs;
 615     enum drain_type drain_type;
 616     int *aio_ret;
 617 };
 618
 619 static void test_iothread_drain_entry(void *opaque)
 620 {
 621     struct test_iothread_data *data = opaque;
 622
 623     aio_context_acquire(bdrv_get_aio_context(data->bs));
 624     do_drain_begin(data->drain_type, data->bs);
 625     g_assert_cmpint(*data->aio_ret, ==, 0);
 626     do_drain_end(data->drain_type, data->bs);
 627     aio_context_release(bdrv_get_aio_context(data->bs));
 628
 629     qemu_event_set(&done_event);
 630 }
 631
 632 static void test_iothread_aio_cb(void *opaque, int ret)
 633 {
 634     int *aio_ret = opaque;
 635     *aio_ret = ret;
 636     qemu_event_set(&done_event);
 637 }
 638
 639 /*
 640  * Starts an AIO request on a BDS that runs in the AioContext of iothread 1.
 641  * The request involves a BH on iothread 2 before it can complete.
 642  *
 643  * @drain_thread = 0 means that do_drain_begin/end are called from the main
 644  * thread, @drain_thread = 1 means that they are called from iothread 1. Drain
 645  * for this BDS cannot be called from iothread 2 because only the main thread
 646  * may do cross-AioContext polling.
 647  */
 648 static void test_iothread_common(enum drain_type drain_type, int drain_thread)
 649 {
 650     BlockBackend *blk;
 651     BlockDriverState *bs;
 652     BDRVTestState *s;
 653     BlockAIOCB *acb;
 654     int aio_ret;
 655     struct test_iothread_data data;
 656
 657     IOThread *a = iothread_new();
 658     IOThread *b = iothread_new();
 659     AioContext *ctx_a = iothread_get_aio_context(a);
 660     AioContext *ctx_b = iothread_get_aio_context(b);
 661
 662     QEMUIOVector qiov;
 663     struct iovec iov = {
 664         .iov_base = NULL,
 665         .iov_len = 0,
 666     };
 667     qemu_iovec_init_external(&qiov, &iov, 1);
 668
 669     /* bdrv_drain_all() may only be called from the main loop thread */
 670     if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) {
 671         goto out;
 672     }
 673
 674     blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 675     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
 676                               &error_abort);
 677     s = bs->opaque;
 678     blk_insert_bs(blk, bs, &error_abort);
 679
 680     blk_set_aio_context(blk, ctx_a);
 681     aio_context_acquire(ctx_a);
 682
 683     s->bh_indirection_ctx = ctx_b;
 684
 685     aio_ret = -EINPROGRESS;
 686     if (drain_thread == 0) {
 687         acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret);
 688     } else {
 689         acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
 690     }
 691     g_assert(acb != NULL);
 692     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
 693
 694     aio_context_release(ctx_a);
 695
 696     data = (struct test_iothread_data) {
 697         .bs         = bs,
 698         .drain_type = drain_type,
 699         .aio_ret    = &aio_ret,
 700     };
 701
 702     switch (drain_thread) {
 703     case 0:
 704         if (drain_type != BDRV_DRAIN_ALL) {
 705             aio_context_acquire(ctx_a);
 706         }
 707
 708         /* The request is running on the IOThread a. Draining its block device
 709          * will make sure that it has completed as far as the BDS is concerned,
 710          * but the drain in this thread can continue immediately after
 711          * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
 712          * later. */
 713         qemu_event_reset(&done_event);
 714         do_drain_begin(drain_type, bs);
 715         g_assert_cmpint(bs->in_flight, ==, 0);
 716
 717         if (drain_type != BDRV_DRAIN_ALL) {
 718             aio_context_release(ctx_a);
 719         }
 720         qemu_event_wait(&done_event);
 721         if (drain_type != BDRV_DRAIN_ALL) {
 722             aio_context_acquire(ctx_a);
 723         }
 724
 725         g_assert_cmpint(aio_ret, ==, 0);
 726         do_drain_end(drain_type, bs);
 727
 728         if (drain_type != BDRV_DRAIN_ALL) {
 729             aio_context_release(ctx_a);
 730         }
 731         break;
 732     case 1:
 733         qemu_event_reset(&done_event);
 734         aio_bh_schedule_oneshot(ctx_a, test_iothread_drain_entry, &data);
 735         qemu_event_wait(&done_event);
 736         break;
 737     default:
 738         g_assert_not_reached();
 739     }
 740
 741     aio_context_acquire(ctx_a);
 742     blk_set_aio_context(blk, qemu_get_aio_context());
 743     aio_context_release(ctx_a);
 744
 745     bdrv_unref(bs);
 746     blk_unref(blk);
 747
 748 out:
 749     iothread_join(a);
 750     iothread_join(b);
 751 }
 752
 753 static void test_iothread_drain_all(void)
 754 {
 755     test_iothread_common(BDRV_DRAIN_ALL, 0);
 756     test_iothread_common(BDRV_DRAIN_ALL, 1);
 757 }
 758
 759 static void test_iothread_drain(void)
 760 {
 761     test_iothread_common(BDRV_DRAIN, 0);
 762     test_iothread_common(BDRV_DRAIN, 1);
 763 }
 764
 765 static void test_iothread_drain_subtree(void)
 766 {
 767     test_iothread_common(BDRV_SUBTREE_DRAIN, 0);
 768     test_iothread_common(BDRV_SUBTREE_DRAIN, 1);
 769 }
 770
 771
 772 typedef struct TestBlockJob {
 773     BlockJob common;
 774     bool should_complete;
 775 } TestBlockJob;
 776
 777 static int coroutine_fn test_job_run(Job *job, Error **errp)
 778 {
 779     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 780
 781     job_transition_to_ready(&s->common.job);
 782     while (!s->should_complete) {
 783         /* Avoid block_job_sleep_ns() because it marks the job as !busy. We
 784          * want to emulate some actual activity (probably some I/O) here so
 785          * that drain has to wait for this acitivity to stop. */
 786         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
 787         job_pause_point(&s->common.job);
 788     }
 789
 790     return 0;
 791 }
 792
 793 static void test_job_complete(Job *job, Error **errp)
 794 {
 795     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 796     s->should_complete = true;
 797 }
 798
 799 BlockJobDriver test_job_driver = {
 800     .job_driver = {
 801         .instance_size  = sizeof(TestBlockJob),
 802         .free           = block_job_free,
 803         .user_resume    = block_job_user_resume,
 804         .drain          = block_job_drain,
 805         .run            = test_job_run,
 806         .complete       = test_job_complete,
 807     },
 808 };
 809
 810 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread)
 811 {
 812     BlockBackend *blk_src, *blk_target;
 813     BlockDriverState *src, *target;
 814     BlockJob *job;
 815     IOThread *iothread = NULL;
 816     AioContext *ctx;
 817     int ret;
 818
 819     src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR,
 820                                &error_abort);
 821     blk_src = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 822     blk_insert_bs(blk_src, src, &error_abort);
 823
 824     if (use_iothread) {
 825         iothread = iothread_new();
 826         ctx = iothread_get_aio_context(iothread);
 827         blk_set_aio_context(blk_src, ctx);
 828     } else {
 829         ctx = qemu_get_aio_context();
 830     }
 831
 832     target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR,
 833                                   &error_abort);
 834     blk_target = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
 835     blk_insert_bs(blk_target, target, &error_abort);
 836
 837     aio_context_acquire(ctx);
 838     job = block_job_create("job0", &test_job_driver, NULL, src, 0, BLK_PERM_ALL,
 839                            0, 0, NULL, NULL, &error_abort);
 840     block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort);
 841     job_start(&job->job);
 842     aio_context_release(ctx);
 843
 844     g_assert_cmpint(job->job.pause_count, ==, 0);
 845     g_assert_false(job->job.paused);
 846     g_assert_true(job->job.busy); /* We're in job_sleep_ns() */
 847
 848     do_drain_begin_unlocked(drain_type, src);
 849
 850     if (drain_type == BDRV_DRAIN_ALL) {
 851         /* bdrv_drain_all() drains both src and target */
 852         g_assert_cmpint(job->job.pause_count, ==, 2);
 853     } else {
 854         g_assert_cmpint(job->job.pause_count, ==, 1);
 855     }
 856     g_assert_true(job->job.paused);
 857     g_assert_false(job->job.busy); /* The job is paused */
 858
 859     do_drain_end_unlocked(drain_type, src);
 860
 861     if (use_iothread) {
 862         /* paused is reset in the I/O thread, wait for it */
 863         while (job->job.paused) {
 864             aio_poll(qemu_get_aio_context(), false);
 865         }
 866     }
 867
 868     g_assert_cmpint(job->job.pause_count, ==, 0);
 869     g_assert_false(job->job.paused);
 870     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
 871
 872     do_drain_begin(drain_type, target);
 873
 874     if (drain_type == BDRV_DRAIN_ALL) {
 875         /* bdrv_drain_all() drains both src and target */
 876         g_assert_cmpint(job->job.pause_count, ==, 2);
 877     } else {
 878         g_assert_cmpint(job->job.pause_count, ==, 1);
 879     }
 880     g_assert_true(job->job.paused);
 881     g_assert_false(job->job.busy); /* The job is paused */
 882
 883     do_drain_end(drain_type, target);
 884
 885     if (use_iothread) {
 886         /* paused is reset in the I/O thread, wait for it */
 887         while (job->job.paused) {
 888             aio_poll(qemu_get_aio_context(), false);
 889         }
 890     }
 891
 892     g_assert_cmpint(job->job.pause_count, ==, 0);
 893     g_assert_false(job->job.paused);
 894     g_assert_true(job->job.busy); /* We're in job_sleep_ns() */
 895
 896     aio_context_acquire(ctx);
 897     ret = job_complete_sync(&job->job, &error_abort);
 898     g_assert_cmpint(ret, ==, 0);
 899
 900     if (use_iothread) {
 901         blk_set_aio_context(blk_src, qemu_get_aio_context());
 902     }
 903     aio_context_release(ctx);
 904
 905     blk_unref(blk_src);
 906     blk_unref(blk_target);
 907     bdrv_unref(src);
 908     bdrv_unref(target);
 909
 910     if (iothread) {
 911         iothread_join(iothread);
 912     }
 913 }
 914
 915 static void test_blockjob_drain_all(void)
 916 {
 917     test_blockjob_common(BDRV_DRAIN_ALL, false);
 918 }
 919
 920 static void test_blockjob_drain(void)
 921 {
 922     test_blockjob_common(BDRV_DRAIN, false);
 923 }
 924
 925 static void test_blockjob_drain_subtree(void)
 926 {
 927     test_blockjob_common(BDRV_SUBTREE_DRAIN, false);
 928 }
 929
 930 static void test_blockjob_iothread_drain_all(void)
 931 {
 932     test_blockjob_common(BDRV_DRAIN_ALL, true);
 933 }
 934
 935 static void test_blockjob_iothread_drain(void)
 936 {
 937     test_blockjob_common(BDRV_DRAIN, true);
 938 }
 939
 940 static void test_blockjob_iothread_drain_subtree(void)
 941 {
 942     test_blockjob_common(BDRV_SUBTREE_DRAIN, true);
 943 }
 944
 945
 946 typedef struct BDRVTestTopState {
 947     BdrvChild *wait_child;
 948 } BDRVTestTopState;
 949
 950 static void bdrv_test_top_close(BlockDriverState *bs)
 951 {
 952     BdrvChild *c, *next_c;
 953     QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
 954         bdrv_unref_child(bs, c);
 955     }
 956 }
 957
 958 static int coroutine_fn bdrv_test_top_co_preadv(BlockDriverState *bs,
 959                                                 uint64_t offset, uint64_t bytes,
 960                                                 QEMUIOVector *qiov, int flags)
 961 {
 962     BDRVTestTopState *tts = bs->opaque;
 963     return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags);
 964 }
 965
 966 static BlockDriver bdrv_test_top_driver = {
 967     .format_name            = "test_top_driver",
 968     .instance_size          = sizeof(BDRVTestTopState),
 969
 970     .bdrv_close             = bdrv_test_top_close,
 971     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
 972
 973     .bdrv_child_perm        = bdrv_format_default_perms,
 974 };
 975
 976 typedef struct TestCoDeleteByDrainData {
 977     BlockBackend *blk;
 978     bool detach_instead_of_delete;
 979     bool done;
 980 } TestCoDeleteByDrainData;
 981
 982 static void coroutine_fn test_co_delete_by_drain(void *opaque)
 983 {
 984     TestCoDeleteByDrainData *dbdd = opaque;
 985     BlockBackend *blk = dbdd->blk;
 986     BlockDriverState *bs = blk_bs(blk);
 987     BDRVTestTopState *tts = bs->opaque;
 988     void *buffer = g_malloc(65536);
 989     QEMUIOVector qiov;
 990     struct iovec iov = {
 991         .iov_base = buffer,
 992         .iov_len  = 65536,
 993     };
 994
 995     qemu_iovec_init_external(&qiov, &iov, 1);
 996
 997     /* Pretend some internal write operation from parent to child.
 998      * Important: We have to read from the child, not from the parent!
 999      * Draining works by first propagating it all up the tree to the
1000      * root and then waiting for drainage from root to the leaves
1001      * (protocol nodes).  If we have a request waiting on the root,
1002      * everything will be drained before we go back down the tree, but
1003      * we do not want that.  We want to be in the middle of draining
1004      * when this following requests returns. */
1005     bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0);
1006
1007     g_assert_cmpint(bs->refcnt, ==, 1);
1008
1009     if (!dbdd->detach_instead_of_delete) {
1010         blk_unref(blk);
1011     } else {
1012         BdrvChild *c, *next_c;
1013         QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1014             bdrv_unref_child(bs, c);
1015         }
1016     }
1017
1018     dbdd->done = true;
1019     g_free(buffer);
1020 }
1021
1022 /**
1023  * Test what happens when some BDS has some children, you drain one of
1024  * them and this results in the BDS being deleted.
1025  *
1026  * If @detach_instead_of_delete is set, the BDS is not going to be
1027  * deleted but will only detach all of its children.
1028  */
1029 static void do_test_delete_by_drain(bool detach_instead_of_delete,
1030                                     enum drain_type drain_type)
1031 {
1032     BlockBackend *blk;
1033     BlockDriverState *bs, *child_bs, *null_bs;
1034     BDRVTestTopState *tts;
1035     TestCoDeleteByDrainData dbdd;
1036     Coroutine *co;
1037
1038     bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR,
1039                               &error_abort);
1040     bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1041     tts = bs->opaque;
1042
1043     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1044                         &error_abort);
1045     bdrv_attach_child(bs, null_bs, "null-child", &child_file, &error_abort);
1046
1047     /* This child will be the one to pass to requests through to, and
1048      * it will stall until a drain occurs */
1049     child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR,
1050                                     &error_abort);
1051     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1052     /* Takes our reference to child_bs */
1053     tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child", &child_file,
1054                                         &error_abort);
1055
1056     /* This child is just there to be deleted
1057      * (for detach_instead_of_delete == true) */
1058     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1059                         &error_abort);
1060     bdrv_attach_child(bs, null_bs, "null-child", &child_file, &error_abort);
1061
1062     blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
1063     blk_insert_bs(blk, bs, &error_abort);
1064
1065     /* Referenced by blk now */
1066     bdrv_unref(bs);
1067
1068     g_assert_cmpint(bs->refcnt, ==, 1);
1069     g_assert_cmpint(child_bs->refcnt, ==, 1);
1070     g_assert_cmpint(null_bs->refcnt, ==, 1);
1071
1072
1073     dbdd = (TestCoDeleteByDrainData){
1074         .blk = blk,
1075         .detach_instead_of_delete = detach_instead_of_delete,
1076         .done = false,
1077     };
1078     co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd);
1079     qemu_coroutine_enter(co);
1080
1081     /* Drain the child while the read operation is still pending.
1082      * This should result in the operation finishing and
1083      * test_co_delete_by_drain() resuming.  Thus, @bs will be deleted
1084      * and the coroutine will exit while this drain operation is still
1085      * in progress. */
1086     switch (drain_type) {
1087     case BDRV_DRAIN:
1088         bdrv_ref(child_bs);
1089         bdrv_drain(child_bs);
1090         bdrv_unref(child_bs);
1091         break;
1092     case BDRV_SUBTREE_DRAIN:
1093         /* Would have to ref/unref bs here for !detach_instead_of_delete, but
1094          * then the whole test becomes pointless because the graph changes
1095          * don't occur during the drain any more. */
1096         assert(detach_instead_of_delete);
1097         bdrv_subtree_drained_begin(bs);
1098         bdrv_subtree_drained_end(bs);
1099         break;
1100     case BDRV_DRAIN_ALL:
1101         bdrv_drain_all_begin();
1102         bdrv_drain_all_end();
1103         break;
1104     default:
1105         g_assert_not_reached();
1106     }
1107
1108     while (!dbdd.done) {
1109         aio_poll(qemu_get_aio_context(), true);
1110     }
1111
1112     if (detach_instead_of_delete) {
1113         /* Here, the reference has not passed over to the coroutine,
1114          * so we have to delete the BB ourselves */
1115         blk_unref(blk);
1116     }
1117 }
1118
1119 static void test_delete_by_drain(void)
1120 {
1121     do_test_delete_by_drain(false, BDRV_DRAIN);
1122 }
1123
1124 static void test_detach_by_drain_all(void)
1125 {
1126     do_test_delete_by_drain(true, BDRV_DRAIN_ALL);
1127 }
1128
1129 static void test_detach_by_drain(void)
1130 {
1131     do_test_delete_by_drain(true, BDRV_DRAIN);
1132 }
1133
1134 static void test_detach_by_drain_subtree(void)
1135 {
1136     do_test_delete_by_drain(true, BDRV_SUBTREE_DRAIN);
1137 }
1138
1139
1140 struct detach_by_parent_data {
1141     BlockDriverState *parent_b;
1142     BdrvChild *child_b;
1143     BlockDriverState *c;
1144     BdrvChild *child_c;
1145     bool by_parent_cb;
1146 };
1147 static struct detach_by_parent_data detach_by_parent_data;
1148
1149 static void detach_indirect_bh(void *opaque)
1150 {
1151     struct detach_by_parent_data *data = opaque;
1152
1153     bdrv_unref_child(data->parent_b, data->child_b);
1154
1155     bdrv_ref(data->c);
1156     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
1157                                       &child_file, &error_abort);
1158 }
1159
1160 static void detach_by_parent_aio_cb(void *opaque, int ret)
1161 {
1162     struct detach_by_parent_data *data = &detach_by_parent_data;
1163
1164     g_assert_cmpint(ret, ==, 0);
1165     if (data->by_parent_cb) {
1166         detach_indirect_bh(data);
1167     }
1168 }
1169
1170 static void detach_by_driver_cb_drained_begin(BdrvChild *child)
1171 {
1172     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1173                             detach_indirect_bh, &detach_by_parent_data);
1174     child_file.drained_begin(child);
1175 }
1176
1177 static BdrvChildRole detach_by_driver_cb_role;
1178
1179 /*
1180  * Initial graph:
1181  *
1182  * PA     PB
1183  *    \ /   \
1184  *     A     B     C
1185  *
1186  * by_parent_cb == true:  Test that parent callbacks don't poll
1187  *
1188  *     PA has a pending write request whose callback changes the child nodes of
1189  *     PB: It removes B and adds C instead. The subtree of PB is drained, which
1190  *     will indirectly drain the write request, too.
1191  *
1192  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1193  *
1194  *     PA's BdrvChildRole has a .drained_begin callback that schedules a BH
1195  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
1196  *     state is messed up, but if it is only polled in the single
1197  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1198  */
1199 static void test_detach_indirect(bool by_parent_cb)
1200 {
1201     BlockBackend *blk;
1202     BlockDriverState *parent_a, *parent_b, *a, *b, *c;
1203     BdrvChild *child_a, *child_b;
1204     BlockAIOCB *acb;
1205
1206     QEMUIOVector qiov;
1207     struct iovec iov = {
1208         .iov_base = NULL,
1209         .iov_len = 0,
1210     };
1211     qemu_iovec_init_external(&qiov, &iov, 1);
1212
1213     if (!by_parent_cb) {
1214         detach_by_driver_cb_role = child_file;
1215         detach_by_driver_cb_role.drained_begin =
1216             detach_by_driver_cb_drained_begin;
1217     }
1218
1219     /* Create all involved nodes */
1220     parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR,
1221                                     &error_abort);
1222     parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0,
1223                                     &error_abort);
1224
1225     a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort);
1226     b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort);
1227     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
1228
1229     /* blk is a BB for parent-a */
1230     blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
1231     blk_insert_bs(blk, parent_a, &error_abort);
1232     bdrv_unref(parent_a);
1233
1234     /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1235      * callback must not return immediately. */
1236     if (!by_parent_cb) {
1237         BDRVTestState *s = parent_a->opaque;
1238         s->sleep_in_drain_begin = true;
1239     }
1240
1241     /* Set child relationships */
1242     bdrv_ref(b);
1243     bdrv_ref(a);
1244     child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_file, &error_abort);
1245     child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_backing, &error_abort);
1246
1247     bdrv_ref(a);
1248     bdrv_attach_child(parent_a, a, "PA-A",
1249                       by_parent_cb ? &child_file : &detach_by_driver_cb_role,
1250                       &error_abort);
1251
1252     g_assert_cmpint(parent_a->refcnt, ==, 1);
1253     g_assert_cmpint(parent_b->refcnt, ==, 1);
1254     g_assert_cmpint(a->refcnt, ==, 3);
1255     g_assert_cmpint(b->refcnt, ==, 2);
1256     g_assert_cmpint(c->refcnt, ==, 1);
1257
1258     g_assert(QLIST_FIRST(&parent_b->children) == child_a);
1259     g_assert(QLIST_NEXT(child_a, next) == child_b);
1260     g_assert(QLIST_NEXT(child_b, next) == NULL);
1261
1262     /* Start the evil write request */
1263     detach_by_parent_data = (struct detach_by_parent_data) {
1264         .parent_b = parent_b,
1265         .child_b = child_b,
1266         .c = c,
1267         .by_parent_cb = by_parent_cb,
1268     };
1269     acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL);
1270     g_assert(acb != NULL);
1271
1272     /* Drain and check the expected result */
1273     bdrv_subtree_drained_begin(parent_b);
1274
1275     g_assert(detach_by_parent_data.child_c != NULL);
1276
1277     g_assert_cmpint(parent_a->refcnt, ==, 1);
1278     g_assert_cmpint(parent_b->refcnt, ==, 1);
1279     g_assert_cmpint(a->refcnt, ==, 3);
1280     g_assert_cmpint(b->refcnt, ==, 1);
1281     g_assert_cmpint(c->refcnt, ==, 2);
1282
1283     g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c);
1284     g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a);
1285     g_assert(QLIST_NEXT(child_a, next) == NULL);
1286
1287     g_assert_cmpint(parent_a->quiesce_counter, ==, 1);
1288     g_assert_cmpint(parent_b->quiesce_counter, ==, 1);
1289     g_assert_cmpint(a->quiesce_counter, ==, 1);
1290     g_assert_cmpint(b->quiesce_counter, ==, 0);
1291     g_assert_cmpint(c->quiesce_counter, ==, 1);
1292
1293     bdrv_subtree_drained_end(parent_b);
1294
1295     bdrv_unref(parent_b);
1296     blk_unref(blk);
1297
1298     /* XXX Once bdrv_close() unref's children instead of just detaching them,
1299      * this won't be necessary any more. */
1300     bdrv_unref(a);
1301     bdrv_unref(a);
1302     bdrv_unref(c);
1303
1304     g_assert_cmpint(a->refcnt, ==, 1);
1305     g_assert_cmpint(b->refcnt, ==, 1);
1306     g_assert_cmpint(c->refcnt, ==, 1);
1307     bdrv_unref(a);
1308     bdrv_unref(b);
1309     bdrv_unref(c);
1310 }
1311
1312 static void test_detach_by_parent_cb(void)
1313 {
1314     test_detach_indirect(true);
1315 }
1316
1317 static void test_detach_by_driver_cb(void)
1318 {
1319     test_detach_indirect(false);
1320 }
1321
1322 static void test_append_to_drained(void)
1323 {
1324     BlockBackend *blk;
1325     BlockDriverState *base, *overlay;
1326     BDRVTestState *base_s, *overlay_s;
1327
1328     blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
1329     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
1330     base_s = base->opaque;
1331     blk_insert_bs(blk, base, &error_abort);
1332
1333     overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR,
1334                                    &error_abort);
1335     overlay_s = overlay->opaque;
1336
1337     do_drain_begin(BDRV_DRAIN, base);
1338     g_assert_cmpint(base->quiesce_counter, ==, 1);
1339     g_assert_cmpint(base_s->drain_count, ==, 1);
1340     g_assert_cmpint(base->in_flight, ==, 0);
1341
1342     /* Takes ownership of overlay, so we don't have to unref it later */
1343     bdrv_append(overlay, base, &error_abort);
1344     g_assert_cmpint(base->in_flight, ==, 0);
1345     g_assert_cmpint(overlay->in_flight, ==, 0);
1346
1347     g_assert_cmpint(base->quiesce_counter, ==, 1);
1348     g_assert_cmpint(base_s->drain_count, ==, 1);
1349     g_assert_cmpint(overlay->quiesce_counter, ==, 1);
1350     g_assert_cmpint(overlay_s->drain_count, ==, 1);
1351
1352     do_drain_end(BDRV_DRAIN, base);
1353
1354     g_assert_cmpint(base->quiesce_counter, ==, 0);
1355     g_assert_cmpint(base_s->drain_count, ==, 0);
1356     g_assert_cmpint(overlay->quiesce_counter, ==, 0);
1357     g_assert_cmpint(overlay_s->drain_count, ==, 0);
1358
1359     bdrv_unref(base);
1360     blk_unref(blk);
1361 }
1362
1363 int main(int argc, char **argv)
1364 {
1365     int ret;
1366
1367     bdrv_init();
1368     qemu_init_main_loop(&error_abort);
1369
1370     g_test_init(&argc, &argv, NULL);
1371     qemu_event_init(&done_event, false);
1372
1373     g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all);
1374     g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain);
1375     g_test_add_func("/bdrv-drain/driver-cb/drain_subtree",
1376                     test_drv_cb_drain_subtree);
1377
1378     g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
1379                     test_drv_cb_co_drain_all);
1380     g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain);
1381     g_test_add_func("/bdrv-drain/driver-cb/co/drain_subtree",
1382                     test_drv_cb_co_drain_subtree);
1383
1384
1385     g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all);
1386     g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain);
1387     g_test_add_func("/bdrv-drain/quiesce/drain_subtree",
1388                     test_quiesce_drain_subtree);
1389
1390     g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
1391                     test_quiesce_co_drain_all);
1392     g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain);
1393     g_test_add_func("/bdrv-drain/quiesce/co/drain_subtree",
1394                     test_quiesce_co_drain_subtree);
1395
1396     g_test_add_func("/bdrv-drain/nested", test_nested);
1397     g_test_add_func("/bdrv-drain/multiparent", test_multiparent);
1398
1399     g_test_add_func("/bdrv-drain/graph-change/drain_subtree",
1400                     test_graph_change_drain_subtree);
1401     g_test_add_func("/bdrv-drain/graph-change/drain_all",
1402                     test_graph_change_drain_all);
1403
1404     g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all);
1405     g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain);
1406     g_test_add_func("/bdrv-drain/iothread/drain_subtree",
1407                     test_iothread_drain_subtree);
1408
1409     g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all);
1410     g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain);
1411     g_test_add_func("/bdrv-drain/blockjob/drain_subtree",
1412                     test_blockjob_drain_subtree);
1413
1414     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
1415                     test_blockjob_iothread_drain_all);
1416     g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
1417                     test_blockjob_iothread_drain);
1418     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_subtree",
1419                     test_blockjob_iothread_drain_subtree);
1420
1421     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
1422     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
1423     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
1424     g_test_add_func("/bdrv-drain/detach/drain_subtree", test_detach_by_drain_subtree);
1425     g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb);
1426     g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb);
1427
1428     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
1429
1430     ret = g_test_run();
1431     qemu_event_destroy(&done_event);
1432     return ret;
1433 }