2 * Block node draining tests
4 * Copyright (c) 2017 Kevin Wolf <kwolf@redhat.com>
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:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
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
25 #include "qemu/osdep.h"
26 #include "block/block_int.h"
27 #include "block/blockjob_int.h"
28 #include "sysemu/block-backend.h"
29 #include "qapi/error.h"
30 #include "qemu/main-loop.h"
33 static QemuEvent done_event
;
35 typedef struct BDRVTestState
{
37 AioContext
*bh_indirection_ctx
;
38 bool sleep_in_drain_begin
;
41 static void coroutine_fn
sleep_in_drain_begin(void *opaque
)
43 BlockDriverState
*bs
= opaque
;
45 qemu_co_sleep_ns(QEMU_CLOCK_REALTIME
, 100000);
46 bdrv_dec_in_flight(bs
);
49 static void bdrv_test_drain_begin(BlockDriverState
*bs
)
51 BDRVTestState
*s
= bs
->opaque
;
53 if (s
->sleep_in_drain_begin
) {
54 Coroutine
*co
= qemu_coroutine_create(sleep_in_drain_begin
, bs
);
55 bdrv_inc_in_flight(bs
);
56 aio_co_enter(bdrv_get_aio_context(bs
), co
);
60 static void bdrv_test_drain_end(BlockDriverState
*bs
)
62 BDRVTestState
*s
= bs
->opaque
;
66 static void bdrv_test_close(BlockDriverState
*bs
)
68 BDRVTestState
*s
= bs
->opaque
;
69 g_assert_cmpint(s
->drain_count
, >, 0);
72 static void co_reenter_bh(void *opaque
)
77 static int coroutine_fn
bdrv_test_co_preadv(BlockDriverState
*bs
,
78 int64_t offset
, int64_t bytes
,
80 BdrvRequestFlags flags
)
82 BDRVTestState
*s
= bs
->opaque
;
84 /* We want this request to stay until the polling loop in drain waits for
85 * it to complete. We need to sleep a while as bdrv_drain_invoke() comes
86 * first and polls its result, too, but it shouldn't accidentally complete
87 * this request yet. */
88 qemu_co_sleep_ns(QEMU_CLOCK_REALTIME
, 100000);
90 if (s
->bh_indirection_ctx
) {
91 aio_bh_schedule_oneshot(s
->bh_indirection_ctx
, co_reenter_bh
,
92 qemu_coroutine_self());
93 qemu_coroutine_yield();
99 static int bdrv_test_change_backing_file(BlockDriverState
*bs
,
100 const char *backing_file
,
101 const char *backing_fmt
)
106 static BlockDriver bdrv_test
= {
107 .format_name
= "test",
108 .instance_size
= sizeof(BDRVTestState
),
109 .supports_backing
= true,
111 .bdrv_close
= bdrv_test_close
,
112 .bdrv_co_preadv
= bdrv_test_co_preadv
,
114 .bdrv_drain_begin
= bdrv_test_drain_begin
,
115 .bdrv_drain_end
= bdrv_test_drain_end
,
117 .bdrv_child_perm
= bdrv_default_perms
,
119 .bdrv_change_backing_file
= bdrv_test_change_backing_file
,
122 static void aio_ret_cb(void *opaque
, int ret
)
124 int *aio_ret
= opaque
;
128 typedef struct CallInCoroutineData
{
131 } CallInCoroutineData
;
133 static coroutine_fn
void call_in_coroutine_entry(void *opaque
)
135 CallInCoroutineData
*data
= opaque
;
141 static void call_in_coroutine(void (*entry
)(void))
144 CallInCoroutineData data
= {
149 co
= qemu_coroutine_create(call_in_coroutine_entry
, &data
);
150 qemu_coroutine_enter(co
);
152 aio_poll(qemu_get_aio_context(), true);
162 static void do_drain_begin(enum drain_type drain_type
, BlockDriverState
*bs
)
164 switch (drain_type
) {
165 case BDRV_DRAIN_ALL
: bdrv_drain_all_begin(); break;
166 case BDRV_DRAIN
: bdrv_drained_begin(bs
); break;
167 default: g_assert_not_reached();
171 static void do_drain_end(enum drain_type drain_type
, BlockDriverState
*bs
)
173 switch (drain_type
) {
174 case BDRV_DRAIN_ALL
: bdrv_drain_all_end(); break;
175 case BDRV_DRAIN
: bdrv_drained_end(bs
); break;
176 default: g_assert_not_reached();
180 static void do_drain_begin_unlocked(enum drain_type drain_type
, BlockDriverState
*bs
)
182 if (drain_type
!= BDRV_DRAIN_ALL
) {
183 aio_context_acquire(bdrv_get_aio_context(bs
));
185 do_drain_begin(drain_type
, bs
);
186 if (drain_type
!= BDRV_DRAIN_ALL
) {
187 aio_context_release(bdrv_get_aio_context(bs
));
191 static BlockBackend
* no_coroutine_fn
test_setup(void)
194 BlockDriverState
*bs
, *backing
;
196 blk
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
197 bs
= bdrv_new_open_driver(&bdrv_test
, "test-node", BDRV_O_RDWR
,
199 blk_insert_bs(blk
, bs
, &error_abort
);
201 backing
= bdrv_new_open_driver(&bdrv_test
, "backing", 0, &error_abort
);
202 bdrv_set_backing_hd(bs
, backing
, &error_abort
);
210 static void do_drain_end_unlocked(enum drain_type drain_type
, BlockDriverState
*bs
)
212 if (drain_type
!= BDRV_DRAIN_ALL
) {
213 aio_context_acquire(bdrv_get_aio_context(bs
));
215 do_drain_end(drain_type
, bs
);
216 if (drain_type
!= BDRV_DRAIN_ALL
) {
217 aio_context_release(bdrv_get_aio_context(bs
));
221 static void test_drv_cb_common(BlockBackend
*blk
, enum drain_type drain_type
,
224 BlockDriverState
*bs
= blk_bs(blk
);
225 BlockDriverState
*backing
= bs
->backing
->bs
;
226 BDRVTestState
*s
, *backing_s
;
230 QEMUIOVector qiov
= QEMU_IOVEC_INIT_BUF(qiov
, NULL
, 0);
233 backing_s
= backing
->opaque
;
235 /* Simple bdrv_drain_all_begin/end pair, check that CBs are called */
236 g_assert_cmpint(s
->drain_count
, ==, 0);
237 g_assert_cmpint(backing_s
->drain_count
, ==, 0);
239 do_drain_begin(drain_type
, bs
);
241 g_assert_cmpint(s
->drain_count
, ==, 1);
242 g_assert_cmpint(backing_s
->drain_count
, ==, !!recursive
);
244 do_drain_end(drain_type
, bs
);
246 g_assert_cmpint(s
->drain_count
, ==, 0);
247 g_assert_cmpint(backing_s
->drain_count
, ==, 0);
249 /* Now do the same while a request is pending */
250 aio_ret
= -EINPROGRESS
;
251 acb
= blk_aio_preadv(blk
, 0, &qiov
, 0, aio_ret_cb
, &aio_ret
);
252 g_assert(acb
!= NULL
);
253 g_assert_cmpint(aio_ret
, ==, -EINPROGRESS
);
255 g_assert_cmpint(s
->drain_count
, ==, 0);
256 g_assert_cmpint(backing_s
->drain_count
, ==, 0);
258 do_drain_begin(drain_type
, bs
);
260 g_assert_cmpint(aio_ret
, ==, 0);
261 g_assert_cmpint(s
->drain_count
, ==, 1);
262 g_assert_cmpint(backing_s
->drain_count
, ==, !!recursive
);
264 do_drain_end(drain_type
, bs
);
266 g_assert_cmpint(s
->drain_count
, ==, 0);
267 g_assert_cmpint(backing_s
->drain_count
, ==, 0);
270 static void test_drv_cb_drain_all(void)
272 BlockBackend
*blk
= test_setup();
273 test_drv_cb_common(blk
, BDRV_DRAIN_ALL
, true);
277 static void test_drv_cb_drain(void)
279 BlockBackend
*blk
= test_setup();
280 test_drv_cb_common(blk
, BDRV_DRAIN
, false);
284 static void coroutine_fn
test_drv_cb_co_drain_all_entry(void)
286 BlockBackend
*blk
= blk_all_next(NULL
);
287 test_drv_cb_common(blk
, BDRV_DRAIN_ALL
, true);
290 static void test_drv_cb_co_drain_all(void)
292 BlockBackend
*blk
= test_setup();
293 call_in_coroutine(test_drv_cb_co_drain_all_entry
);
297 static void coroutine_fn
test_drv_cb_co_drain_entry(void)
299 BlockBackend
*blk
= blk_all_next(NULL
);
300 test_drv_cb_common(blk
, BDRV_DRAIN
, false);
303 static void test_drv_cb_co_drain(void)
305 BlockBackend
*blk
= test_setup();
306 call_in_coroutine(test_drv_cb_co_drain_entry
);
310 static void test_quiesce_common(BlockBackend
*blk
, enum drain_type drain_type
,
313 BlockDriverState
*bs
= blk_bs(blk
);
314 BlockDriverState
*backing
= bs
->backing
->bs
;
316 g_assert_cmpint(bs
->quiesce_counter
, ==, 0);
317 g_assert_cmpint(backing
->quiesce_counter
, ==, 0);
319 do_drain_begin(drain_type
, bs
);
321 if (drain_type
== BDRV_DRAIN_ALL
) {
322 g_assert_cmpint(bs
->quiesce_counter
, ==, 2);
324 g_assert_cmpint(bs
->quiesce_counter
, ==, 1);
326 g_assert_cmpint(backing
->quiesce_counter
, ==, !!recursive
);
328 do_drain_end(drain_type
, bs
);
330 g_assert_cmpint(bs
->quiesce_counter
, ==, 0);
331 g_assert_cmpint(backing
->quiesce_counter
, ==, 0);
334 static void test_quiesce_drain_all(void)
336 BlockBackend
*blk
= test_setup();
337 test_quiesce_common(blk
, BDRV_DRAIN_ALL
, true);
341 static void test_quiesce_drain(void)
343 BlockBackend
*blk
= test_setup();
344 test_quiesce_common(blk
, BDRV_DRAIN
, false);
348 static void coroutine_fn
test_quiesce_co_drain_all_entry(void)
350 BlockBackend
*blk
= blk_all_next(NULL
);
351 test_quiesce_common(blk
, BDRV_DRAIN_ALL
, true);
354 static void test_quiesce_co_drain_all(void)
356 BlockBackend
*blk
= test_setup();
357 call_in_coroutine(test_quiesce_co_drain_all_entry
);
361 static void coroutine_fn
test_quiesce_co_drain_entry(void)
363 BlockBackend
*blk
= blk_all_next(NULL
);
364 test_quiesce_common(blk
, BDRV_DRAIN
, false);
367 static void test_quiesce_co_drain(void)
369 BlockBackend
*blk
= test_setup();
370 call_in_coroutine(test_quiesce_co_drain_entry
);
374 static void test_nested(void)
377 BlockDriverState
*bs
, *backing
;
378 BDRVTestState
*s
, *backing_s
;
379 enum drain_type outer
, inner
;
381 blk
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
382 bs
= bdrv_new_open_driver(&bdrv_test
, "test-node", BDRV_O_RDWR
,
385 blk_insert_bs(blk
, bs
, &error_abort
);
387 backing
= bdrv_new_open_driver(&bdrv_test
, "backing", 0, &error_abort
);
388 backing_s
= backing
->opaque
;
389 bdrv_set_backing_hd(bs
, backing
, &error_abort
);
391 for (outer
= 0; outer
< DRAIN_TYPE_MAX
; outer
++) {
392 for (inner
= 0; inner
< DRAIN_TYPE_MAX
; inner
++) {
393 int backing_quiesce
= (outer
== BDRV_DRAIN_ALL
) +
394 (inner
== BDRV_DRAIN_ALL
);
396 g_assert_cmpint(bs
->quiesce_counter
, ==, 0);
397 g_assert_cmpint(backing
->quiesce_counter
, ==, 0);
398 g_assert_cmpint(s
->drain_count
, ==, 0);
399 g_assert_cmpint(backing_s
->drain_count
, ==, 0);
401 do_drain_begin(outer
, bs
);
402 do_drain_begin(inner
, bs
);
404 g_assert_cmpint(bs
->quiesce_counter
, ==, 2 + !!backing_quiesce
);
405 g_assert_cmpint(backing
->quiesce_counter
, ==, backing_quiesce
);
406 g_assert_cmpint(s
->drain_count
, ==, 1);
407 g_assert_cmpint(backing_s
->drain_count
, ==, !!backing_quiesce
);
409 do_drain_end(inner
, bs
);
410 do_drain_end(outer
, bs
);
412 g_assert_cmpint(bs
->quiesce_counter
, ==, 0);
413 g_assert_cmpint(backing
->quiesce_counter
, ==, 0);
414 g_assert_cmpint(s
->drain_count
, ==, 0);
415 g_assert_cmpint(backing_s
->drain_count
, ==, 0);
424 static void test_graph_change_drain_all(void)
426 BlockBackend
*blk_a
, *blk_b
;
427 BlockDriverState
*bs_a
, *bs_b
;
428 BDRVTestState
*a_s
, *b_s
;
430 /* Create node A with a BlockBackend */
431 blk_a
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
432 bs_a
= bdrv_new_open_driver(&bdrv_test
, "test-node-a", BDRV_O_RDWR
,
435 blk_insert_bs(blk_a
, bs_a
, &error_abort
);
437 g_assert_cmpint(bs_a
->quiesce_counter
, ==, 0);
438 g_assert_cmpint(a_s
->drain_count
, ==, 0);
440 /* Call bdrv_drain_all_begin() */
441 bdrv_drain_all_begin();
443 g_assert_cmpint(bs_a
->quiesce_counter
, ==, 1);
444 g_assert_cmpint(a_s
->drain_count
, ==, 1);
446 /* Create node B with a BlockBackend */
447 blk_b
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
448 bs_b
= bdrv_new_open_driver(&bdrv_test
, "test-node-b", BDRV_O_RDWR
,
451 blk_insert_bs(blk_b
, bs_b
, &error_abort
);
453 g_assert_cmpint(bs_a
->quiesce_counter
, ==, 1);
454 g_assert_cmpint(bs_b
->quiesce_counter
, ==, 1);
455 g_assert_cmpint(a_s
->drain_count
, ==, 1);
456 g_assert_cmpint(b_s
->drain_count
, ==, 1);
458 /* Unref and finally delete node A */
461 g_assert_cmpint(bs_a
->quiesce_counter
, ==, 1);
462 g_assert_cmpint(bs_b
->quiesce_counter
, ==, 1);
463 g_assert_cmpint(a_s
->drain_count
, ==, 1);
464 g_assert_cmpint(b_s
->drain_count
, ==, 1);
468 g_assert_cmpint(bs_b
->quiesce_counter
, ==, 1);
469 g_assert_cmpint(b_s
->drain_count
, ==, 1);
471 /* End the drained section */
472 bdrv_drain_all_end();
474 g_assert_cmpint(bs_b
->quiesce_counter
, ==, 0);
475 g_assert_cmpint(b_s
->drain_count
, ==, 0);
481 struct test_iothread_data
{
482 BlockDriverState
*bs
;
483 enum drain_type drain_type
;
488 static void coroutine_fn
test_iothread_drain_co_entry(void *opaque
)
490 struct test_iothread_data
*data
= opaque
;
492 do_drain_begin(data
->drain_type
, data
->bs
);
493 g_assert_cmpint(*data
->aio_ret
, ==, 0);
494 do_drain_end(data
->drain_type
, data
->bs
);
496 data
->co_done
= true;
500 static void test_iothread_aio_cb(void *opaque
, int ret
)
502 int *aio_ret
= opaque
;
504 qemu_event_set(&done_event
);
507 static void test_iothread_main_thread_bh(void *opaque
)
509 struct test_iothread_data
*data
= opaque
;
511 /* Test that the AioContext is not yet locked in a random BH that is
512 * executed during drain, otherwise this would deadlock. */
513 aio_context_acquire(bdrv_get_aio_context(data
->bs
));
514 bdrv_flush(data
->bs
);
515 aio_context_release(bdrv_get_aio_context(data
->bs
));
519 * Starts an AIO request on a BDS that runs in the AioContext of iothread 1.
520 * The request involves a BH on iothread 2 before it can complete.
522 * @drain_thread = 0 means that do_drain_begin/end are called from the main
523 * thread, @drain_thread = 1 means that they are called from iothread 1. Drain
524 * for this BDS cannot be called from iothread 2 because only the main thread
525 * may do cross-AioContext polling.
527 static void test_iothread_common(enum drain_type drain_type
, int drain_thread
)
530 BlockDriverState
*bs
;
535 struct test_iothread_data data
;
537 IOThread
*a
= iothread_new();
538 IOThread
*b
= iothread_new();
539 AioContext
*ctx_a
= iothread_get_aio_context(a
);
540 AioContext
*ctx_b
= iothread_get_aio_context(b
);
542 QEMUIOVector qiov
= QEMU_IOVEC_INIT_BUF(qiov
, NULL
, 0);
544 /* bdrv_drain_all() may only be called from the main loop thread */
545 if (drain_type
== BDRV_DRAIN_ALL
&& drain_thread
!= 0) {
549 blk
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
550 bs
= bdrv_new_open_driver(&bdrv_test
, "test-node", BDRV_O_RDWR
,
553 blk_insert_bs(blk
, bs
, &error_abort
);
554 blk_set_disable_request_queuing(blk
, true);
556 blk_set_aio_context(blk
, ctx_a
, &error_abort
);
557 aio_context_acquire(ctx_a
);
559 s
->bh_indirection_ctx
= ctx_b
;
561 aio_ret
= -EINPROGRESS
;
562 qemu_event_reset(&done_event
);
564 if (drain_thread
== 0) {
565 acb
= blk_aio_preadv(blk
, 0, &qiov
, 0, test_iothread_aio_cb
, &aio_ret
);
567 acb
= blk_aio_preadv(blk
, 0, &qiov
, 0, aio_ret_cb
, &aio_ret
);
569 g_assert(acb
!= NULL
);
570 g_assert_cmpint(aio_ret
, ==, -EINPROGRESS
);
572 aio_context_release(ctx_a
);
574 data
= (struct test_iothread_data
) {
576 .drain_type
= drain_type
,
580 switch (drain_thread
) {
582 if (drain_type
!= BDRV_DRAIN_ALL
) {
583 aio_context_acquire(ctx_a
);
586 aio_bh_schedule_oneshot(ctx_a
, test_iothread_main_thread_bh
, &data
);
588 /* The request is running on the IOThread a. Draining its block device
589 * will make sure that it has completed as far as the BDS is concerned,
590 * but the drain in this thread can continue immediately after
591 * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
593 do_drain_begin(drain_type
, bs
);
594 g_assert_cmpint(bs
->in_flight
, ==, 0);
596 if (drain_type
!= BDRV_DRAIN_ALL
) {
597 aio_context_release(ctx_a
);
599 qemu_event_wait(&done_event
);
600 if (drain_type
!= BDRV_DRAIN_ALL
) {
601 aio_context_acquire(ctx_a
);
604 g_assert_cmpint(aio_ret
, ==, 0);
605 do_drain_end(drain_type
, bs
);
607 if (drain_type
!= BDRV_DRAIN_ALL
) {
608 aio_context_release(ctx_a
);
612 co
= qemu_coroutine_create(test_iothread_drain_co_entry
, &data
);
613 aio_co_enter(ctx_a
, co
);
614 AIO_WAIT_WHILE_UNLOCKED(NULL
, !data
.co_done
);
617 g_assert_not_reached();
620 aio_context_acquire(ctx_a
);
621 blk_set_aio_context(blk
, qemu_get_aio_context(), &error_abort
);
622 aio_context_release(ctx_a
);
632 static void test_iothread_drain_all(void)
634 test_iothread_common(BDRV_DRAIN_ALL
, 0);
635 test_iothread_common(BDRV_DRAIN_ALL
, 1);
638 static void test_iothread_drain(void)
640 test_iothread_common(BDRV_DRAIN
, 0);
641 test_iothread_common(BDRV_DRAIN
, 1);
645 typedef struct TestBlockJob
{
647 BlockDriverState
*bs
;
651 bool should_complete
;
654 static int test_job_prepare(Job
*job
)
656 TestBlockJob
*s
= container_of(job
, TestBlockJob
, common
.job
);
658 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
660 return s
->prepare_ret
;
663 static void test_job_commit(Job
*job
)
665 TestBlockJob
*s
= container_of(job
, TestBlockJob
, common
.job
);
667 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
671 static void test_job_abort(Job
*job
)
673 TestBlockJob
*s
= container_of(job
, TestBlockJob
, common
.job
);
675 /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
679 static int coroutine_fn
test_job_run(Job
*job
, Error
**errp
)
681 TestBlockJob
*s
= container_of(job
, TestBlockJob
, common
.job
);
683 /* We are running the actual job code past the pause point in
687 job_transition_to_ready(&s
->common
.job
);
688 while (!s
->should_complete
) {
689 /* Avoid job_sleep_ns() because it marks the job as !busy. We want to
690 * emulate some actual activity (probably some I/O) here so that drain
691 * has to wait for this activity to stop. */
692 qemu_co_sleep_ns(QEMU_CLOCK_REALTIME
, 1000000);
694 job_pause_point(&s
->common
.job
);
700 static void test_job_complete(Job
*job
, Error
**errp
)
702 TestBlockJob
*s
= container_of(job
, TestBlockJob
, common
.job
);
703 s
->should_complete
= true;
706 BlockJobDriver test_job_driver
= {
708 .instance_size
= sizeof(TestBlockJob
),
709 .free
= block_job_free
,
710 .user_resume
= block_job_user_resume
,
712 .complete
= test_job_complete
,
713 .prepare
= test_job_prepare
,
714 .commit
= test_job_commit
,
715 .abort
= test_job_abort
,
719 enum test_job_result
{
722 TEST_JOB_FAIL_PREPARE
,
725 enum test_job_drain_node
{
727 TEST_JOB_DRAIN_SRC_CHILD
,
730 static void test_blockjob_common_drain_node(enum drain_type drain_type
,
732 enum test_job_result result
,
733 enum test_job_drain_node drain_node
)
735 BlockBackend
*blk_src
, *blk_target
;
736 BlockDriverState
*src
, *src_backing
, *src_overlay
, *target
, *drain_bs
;
739 IOThread
*iothread
= NULL
;
743 src
= bdrv_new_open_driver(&bdrv_test
, "source", BDRV_O_RDWR
,
745 src_backing
= bdrv_new_open_driver(&bdrv_test
, "source-backing",
746 BDRV_O_RDWR
, &error_abort
);
747 src_overlay
= bdrv_new_open_driver(&bdrv_test
, "source-overlay",
748 BDRV_O_RDWR
, &error_abort
);
750 bdrv_set_backing_hd(src_overlay
, src
, &error_abort
);
752 bdrv_set_backing_hd(src
, src_backing
, &error_abort
);
753 bdrv_unref(src_backing
);
755 blk_src
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
756 blk_insert_bs(blk_src
, src_overlay
, &error_abort
);
758 switch (drain_node
) {
759 case TEST_JOB_DRAIN_SRC
:
762 case TEST_JOB_DRAIN_SRC_CHILD
:
763 drain_bs
= src_backing
;
766 g_assert_not_reached();
770 iothread
= iothread_new();
771 ctx
= iothread_get_aio_context(iothread
);
772 blk_set_aio_context(blk_src
, ctx
, &error_abort
);
774 ctx
= qemu_get_aio_context();
777 target
= bdrv_new_open_driver(&bdrv_test
, "target", BDRV_O_RDWR
,
779 blk_target
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
780 blk_insert_bs(blk_target
, target
, &error_abort
);
781 blk_set_allow_aio_context_change(blk_target
, true);
783 aio_context_acquire(ctx
);
784 tjob
= block_job_create("job0", &test_job_driver
, NULL
, src
,
786 0, 0, NULL
, NULL
, &error_abort
);
789 block_job_add_bdrv(job
, "target", target
, 0, BLK_PERM_ALL
, &error_abort
);
792 case TEST_JOB_SUCCESS
:
794 case TEST_JOB_FAIL_RUN
:
795 tjob
->run_ret
= -EIO
;
797 case TEST_JOB_FAIL_PREPARE
:
798 tjob
->prepare_ret
= -EIO
;
801 aio_context_release(ctx
);
803 job_start(&job
->job
);
806 /* job_co_entry() is run in the I/O thread, wait for the actual job
807 * code to start (we don't want to catch the job in the pause point in
809 while (!tjob
->running
) {
810 aio_poll(qemu_get_aio_context(), false);
814 WITH_JOB_LOCK_GUARD() {
815 g_assert_cmpint(job
->job
.pause_count
, ==, 0);
816 g_assert_false(job
->job
.paused
);
817 g_assert_true(tjob
->running
);
818 g_assert_true(job
->job
.busy
); /* We're in qemu_co_sleep_ns() */
821 do_drain_begin_unlocked(drain_type
, drain_bs
);
823 WITH_JOB_LOCK_GUARD() {
824 if (drain_type
== BDRV_DRAIN_ALL
) {
825 /* bdrv_drain_all() drains both src and target */
826 g_assert_cmpint(job
->job
.pause_count
, ==, 2);
828 g_assert_cmpint(job
->job
.pause_count
, ==, 1);
830 g_assert_true(job
->job
.paused
);
831 g_assert_false(job
->job
.busy
); /* The job is paused */
834 do_drain_end_unlocked(drain_type
, drain_bs
);
838 * Here we are waiting for the paused status to change,
839 * so don't bother protecting the read every time.
841 * paused is reset in the I/O thread, wait for it
843 while (job
->job
.paused
) {
844 aio_poll(qemu_get_aio_context(), false);
848 WITH_JOB_LOCK_GUARD() {
849 g_assert_cmpint(job
->job
.pause_count
, ==, 0);
850 g_assert_false(job
->job
.paused
);
851 g_assert_true(job
->job
.busy
); /* We're in qemu_co_sleep_ns() */
854 do_drain_begin_unlocked(drain_type
, target
);
856 WITH_JOB_LOCK_GUARD() {
857 if (drain_type
== BDRV_DRAIN_ALL
) {
858 /* bdrv_drain_all() drains both src and target */
859 g_assert_cmpint(job
->job
.pause_count
, ==, 2);
861 g_assert_cmpint(job
->job
.pause_count
, ==, 1);
863 g_assert_true(job
->job
.paused
);
864 g_assert_false(job
->job
.busy
); /* The job is paused */
867 do_drain_end_unlocked(drain_type
, target
);
871 * Here we are waiting for the paused status to change,
872 * so don't bother protecting the read every time.
874 * paused is reset in the I/O thread, wait for it
876 while (job
->job
.paused
) {
877 aio_poll(qemu_get_aio_context(), false);
881 WITH_JOB_LOCK_GUARD() {
882 g_assert_cmpint(job
->job
.pause_count
, ==, 0);
883 g_assert_false(job
->job
.paused
);
884 g_assert_true(job
->job
.busy
); /* We're in qemu_co_sleep_ns() */
887 WITH_JOB_LOCK_GUARD() {
888 ret
= job_complete_sync_locked(&job
->job
, &error_abort
);
890 g_assert_cmpint(ret
, ==, (result
== TEST_JOB_SUCCESS
? 0 : -EIO
));
892 aio_context_acquire(ctx
);
894 blk_set_aio_context(blk_src
, qemu_get_aio_context(), &error_abort
);
895 assert(blk_get_aio_context(blk_target
) == qemu_get_aio_context());
897 aio_context_release(ctx
);
900 blk_unref(blk_target
);
901 bdrv_unref(src_overlay
);
905 iothread_join(iothread
);
909 static void test_blockjob_common(enum drain_type drain_type
, bool use_iothread
,
910 enum test_job_result result
)
912 test_blockjob_common_drain_node(drain_type
, use_iothread
, result
,
914 test_blockjob_common_drain_node(drain_type
, use_iothread
, result
,
915 TEST_JOB_DRAIN_SRC_CHILD
);
918 static void test_blockjob_drain_all(void)
920 test_blockjob_common(BDRV_DRAIN_ALL
, false, TEST_JOB_SUCCESS
);
923 static void test_blockjob_drain(void)
925 test_blockjob_common(BDRV_DRAIN
, false, TEST_JOB_SUCCESS
);
928 static void test_blockjob_error_drain_all(void)
930 test_blockjob_common(BDRV_DRAIN_ALL
, false, TEST_JOB_FAIL_RUN
);
931 test_blockjob_common(BDRV_DRAIN_ALL
, false, TEST_JOB_FAIL_PREPARE
);
934 static void test_blockjob_error_drain(void)
936 test_blockjob_common(BDRV_DRAIN
, false, TEST_JOB_FAIL_RUN
);
937 test_blockjob_common(BDRV_DRAIN
, false, TEST_JOB_FAIL_PREPARE
);
940 static void test_blockjob_iothread_drain_all(void)
942 test_blockjob_common(BDRV_DRAIN_ALL
, true, TEST_JOB_SUCCESS
);
945 static void test_blockjob_iothread_drain(void)
947 test_blockjob_common(BDRV_DRAIN
, true, TEST_JOB_SUCCESS
);
950 static void test_blockjob_iothread_error_drain_all(void)
952 test_blockjob_common(BDRV_DRAIN_ALL
, true, TEST_JOB_FAIL_RUN
);
953 test_blockjob_common(BDRV_DRAIN_ALL
, true, TEST_JOB_FAIL_PREPARE
);
956 static void test_blockjob_iothread_error_drain(void)
958 test_blockjob_common(BDRV_DRAIN
, true, TEST_JOB_FAIL_RUN
);
959 test_blockjob_common(BDRV_DRAIN
, true, TEST_JOB_FAIL_PREPARE
);
963 typedef struct BDRVTestTopState
{
964 BdrvChild
*wait_child
;
967 static void bdrv_test_top_close(BlockDriverState
*bs
)
969 BdrvChild
*c
, *next_c
;
971 bdrv_graph_wrlock(NULL
);
972 QLIST_FOREACH_SAFE(c
, &bs
->children
, next
, next_c
) {
973 bdrv_unref_child(bs
, c
);
975 bdrv_graph_wrunlock();
978 static int coroutine_fn GRAPH_RDLOCK
979 bdrv_test_top_co_preadv(BlockDriverState
*bs
, int64_t offset
, int64_t bytes
,
980 QEMUIOVector
*qiov
, BdrvRequestFlags flags
)
982 BDRVTestTopState
*tts
= bs
->opaque
;
983 return bdrv_co_preadv(tts
->wait_child
, offset
, bytes
, qiov
, flags
);
986 static BlockDriver bdrv_test_top_driver
= {
987 .format_name
= "test_top_driver",
988 .instance_size
= sizeof(BDRVTestTopState
),
990 .bdrv_close
= bdrv_test_top_close
,
991 .bdrv_co_preadv
= bdrv_test_top_co_preadv
,
993 .bdrv_child_perm
= bdrv_default_perms
,
996 typedef struct TestCoDeleteByDrainData
{
998 bool detach_instead_of_delete
;
1000 } TestCoDeleteByDrainData
;
1002 static void coroutine_fn
test_co_delete_by_drain(void *opaque
)
1004 TestCoDeleteByDrainData
*dbdd
= opaque
;
1005 BlockBackend
*blk
= dbdd
->blk
;
1006 BlockDriverState
*bs
= blk_bs(blk
);
1007 BDRVTestTopState
*tts
= bs
->opaque
;
1008 void *buffer
= g_malloc(65536);
1009 QEMUIOVector qiov
= QEMU_IOVEC_INIT_BUF(qiov
, buffer
, 65536);
1011 /* Pretend some internal write operation from parent to child.
1012 * Important: We have to read from the child, not from the parent!
1013 * Draining works by first propagating it all up the tree to the
1014 * root and then waiting for drainage from root to the leaves
1015 * (protocol nodes). If we have a request waiting on the root,
1016 * everything will be drained before we go back down the tree, but
1017 * we do not want that. We want to be in the middle of draining
1018 * when this following requests returns. */
1019 bdrv_graph_co_rdlock();
1020 bdrv_co_preadv(tts
->wait_child
, 0, 65536, &qiov
, 0);
1021 bdrv_graph_co_rdunlock();
1023 g_assert_cmpint(bs
->refcnt
, ==, 1);
1025 if (!dbdd
->detach_instead_of_delete
) {
1028 BdrvChild
*c
, *next_c
;
1029 QLIST_FOREACH_SAFE(c
, &bs
->children
, next
, next_c
) {
1030 bdrv_co_unref_child(bs
, c
);
1039 * Test what happens when some BDS has some children, you drain one of
1040 * them and this results in the BDS being deleted.
1042 * If @detach_instead_of_delete is set, the BDS is not going to be
1043 * deleted but will only detach all of its children.
1045 static void do_test_delete_by_drain(bool detach_instead_of_delete
,
1046 enum drain_type drain_type
)
1049 BlockDriverState
*bs
, *child_bs
, *null_bs
;
1050 BDRVTestTopState
*tts
;
1051 TestCoDeleteByDrainData dbdd
;
1054 bs
= bdrv_new_open_driver(&bdrv_test_top_driver
, "top", BDRV_O_RDWR
,
1056 bs
->total_sectors
= 65536 >> BDRV_SECTOR_BITS
;
1059 null_bs
= bdrv_open("null-co://", NULL
, NULL
, BDRV_O_RDWR
| BDRV_O_PROTOCOL
,
1061 bdrv_graph_wrlock(NULL
);
1062 bdrv_attach_child(bs
, null_bs
, "null-child", &child_of_bds
,
1063 BDRV_CHILD_DATA
, &error_abort
);
1064 bdrv_graph_wrunlock();
1066 /* This child will be the one to pass to requests through to, and
1067 * it will stall until a drain occurs */
1068 child_bs
= bdrv_new_open_driver(&bdrv_test
, "child", BDRV_O_RDWR
,
1070 child_bs
->total_sectors
= 65536 >> BDRV_SECTOR_BITS
;
1071 /* Takes our reference to child_bs */
1072 bdrv_graph_wrlock(NULL
);
1073 tts
->wait_child
= bdrv_attach_child(bs
, child_bs
, "wait-child",
1075 BDRV_CHILD_DATA
| BDRV_CHILD_PRIMARY
,
1077 bdrv_graph_wrunlock();
1079 /* This child is just there to be deleted
1080 * (for detach_instead_of_delete == true) */
1081 null_bs
= bdrv_open("null-co://", NULL
, NULL
, BDRV_O_RDWR
| BDRV_O_PROTOCOL
,
1083 bdrv_graph_wrlock(NULL
);
1084 bdrv_attach_child(bs
, null_bs
, "null-child", &child_of_bds
, BDRV_CHILD_DATA
,
1086 bdrv_graph_wrunlock();
1088 blk
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
1089 blk_insert_bs(blk
, bs
, &error_abort
);
1091 /* Referenced by blk now */
1094 g_assert_cmpint(bs
->refcnt
, ==, 1);
1095 g_assert_cmpint(child_bs
->refcnt
, ==, 1);
1096 g_assert_cmpint(null_bs
->refcnt
, ==, 1);
1099 dbdd
= (TestCoDeleteByDrainData
){
1101 .detach_instead_of_delete
= detach_instead_of_delete
,
1104 co
= qemu_coroutine_create(test_co_delete_by_drain
, &dbdd
);
1105 qemu_coroutine_enter(co
);
1107 /* Drain the child while the read operation is still pending.
1108 * This should result in the operation finishing and
1109 * test_co_delete_by_drain() resuming. Thus, @bs will be deleted
1110 * and the coroutine will exit while this drain operation is still
1112 switch (drain_type
) {
1115 bdrv_drain(child_bs
);
1116 bdrv_unref(child_bs
);
1118 case BDRV_DRAIN_ALL
:
1119 bdrv_drain_all_begin();
1120 bdrv_drain_all_end();
1123 g_assert_not_reached();
1126 while (!dbdd
.done
) {
1127 aio_poll(qemu_get_aio_context(), true);
1130 if (detach_instead_of_delete
) {
1131 /* Here, the reference has not passed over to the coroutine,
1132 * so we have to delete the BB ourselves */
1137 static void test_delete_by_drain(void)
1139 do_test_delete_by_drain(false, BDRV_DRAIN
);
1142 static void test_detach_by_drain_all(void)
1144 do_test_delete_by_drain(true, BDRV_DRAIN_ALL
);
1147 static void test_detach_by_drain(void)
1149 do_test_delete_by_drain(true, BDRV_DRAIN
);
1153 struct detach_by_parent_data
{
1154 BlockDriverState
*parent_b
;
1156 BlockDriverState
*c
;
1159 bool detach_on_drain
;
1161 static struct detach_by_parent_data detach_by_parent_data
;
1163 static void detach_indirect_bh(void *opaque
)
1165 struct detach_by_parent_data
*data
= opaque
;
1167 bdrv_dec_in_flight(data
->child_b
->bs
);
1169 bdrv_graph_wrlock(NULL
);
1170 bdrv_unref_child(data
->parent_b
, data
->child_b
);
1173 data
->child_c
= bdrv_attach_child(data
->parent_b
, data
->c
, "PB-C",
1174 &child_of_bds
, BDRV_CHILD_DATA
,
1176 bdrv_graph_wrunlock();
1179 static void detach_by_parent_aio_cb(void *opaque
, int ret
)
1181 struct detach_by_parent_data
*data
= &detach_by_parent_data
;
1183 g_assert_cmpint(ret
, ==, 0);
1184 if (data
->by_parent_cb
) {
1185 bdrv_inc_in_flight(data
->child_b
->bs
);
1186 detach_indirect_bh(data
);
1190 static void detach_by_driver_cb_drained_begin(BdrvChild
*child
)
1192 struct detach_by_parent_data
*data
= &detach_by_parent_data
;
1194 if (!data
->detach_on_drain
) {
1197 data
->detach_on_drain
= false;
1199 bdrv_inc_in_flight(data
->child_b
->bs
);
1200 aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1201 detach_indirect_bh
, &detach_by_parent_data
);
1202 child_of_bds
.drained_begin(child
);
1205 static BdrvChildClass detach_by_driver_cb_class
;
1214 * by_parent_cb == true: Test that parent callbacks don't poll
1216 * PA has a pending write request whose callback changes the child nodes of
1217 * PB: It removes B and adds C instead. The subtree of PB is drained, which
1218 * will indirectly drain the write request, too.
1220 * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1222 * PA's BdrvChildClass has a .drained_begin callback that schedules a BH
1223 * that does the same graph change. If bdrv_drain_invoke() calls it, the
1224 * state is messed up, but if it is only polled in the single
1225 * BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1227 static void test_detach_indirect(bool by_parent_cb
)
1230 BlockDriverState
*parent_a
, *parent_b
, *a
, *b
, *c
;
1231 BdrvChild
*child_a
, *child_b
;
1234 QEMUIOVector qiov
= QEMU_IOVEC_INIT_BUF(qiov
, NULL
, 0);
1236 if (!by_parent_cb
) {
1237 detach_by_driver_cb_class
= child_of_bds
;
1238 detach_by_driver_cb_class
.drained_begin
=
1239 detach_by_driver_cb_drained_begin
;
1240 detach_by_driver_cb_class
.drained_end
= NULL
;
1241 detach_by_driver_cb_class
.drained_poll
= NULL
;
1244 detach_by_parent_data
= (struct detach_by_parent_data
) {
1245 .detach_on_drain
= false,
1248 /* Create all involved nodes */
1249 parent_a
= bdrv_new_open_driver(&bdrv_test
, "parent-a", BDRV_O_RDWR
,
1251 parent_b
= bdrv_new_open_driver(&bdrv_test
, "parent-b", 0,
1254 a
= bdrv_new_open_driver(&bdrv_test
, "a", BDRV_O_RDWR
, &error_abort
);
1255 b
= bdrv_new_open_driver(&bdrv_test
, "b", BDRV_O_RDWR
, &error_abort
);
1256 c
= bdrv_new_open_driver(&bdrv_test
, "c", BDRV_O_RDWR
, &error_abort
);
1258 /* blk is a BB for parent-a */
1259 blk
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
1260 blk_insert_bs(blk
, parent_a
, &error_abort
);
1261 bdrv_unref(parent_a
);
1263 /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1264 * callback must not return immediately. */
1265 if (!by_parent_cb
) {
1266 BDRVTestState
*s
= parent_a
->opaque
;
1267 s
->sleep_in_drain_begin
= true;
1270 /* Set child relationships */
1273 bdrv_graph_wrlock(NULL
);
1274 child_b
= bdrv_attach_child(parent_b
, b
, "PB-B", &child_of_bds
,
1275 BDRV_CHILD_DATA
, &error_abort
);
1276 child_a
= bdrv_attach_child(parent_b
, a
, "PB-A", &child_of_bds
,
1277 BDRV_CHILD_COW
, &error_abort
);
1280 bdrv_attach_child(parent_a
, a
, "PA-A",
1281 by_parent_cb
? &child_of_bds
: &detach_by_driver_cb_class
,
1282 BDRV_CHILD_DATA
, &error_abort
);
1283 bdrv_graph_wrunlock();
1285 g_assert_cmpint(parent_a
->refcnt
, ==, 1);
1286 g_assert_cmpint(parent_b
->refcnt
, ==, 1);
1287 g_assert_cmpint(a
->refcnt
, ==, 3);
1288 g_assert_cmpint(b
->refcnt
, ==, 2);
1289 g_assert_cmpint(c
->refcnt
, ==, 1);
1291 g_assert(QLIST_FIRST(&parent_b
->children
) == child_a
);
1292 g_assert(QLIST_NEXT(child_a
, next
) == child_b
);
1293 g_assert(QLIST_NEXT(child_b
, next
) == NULL
);
1295 /* Start the evil write request */
1296 detach_by_parent_data
= (struct detach_by_parent_data
) {
1297 .parent_b
= parent_b
,
1300 .by_parent_cb
= by_parent_cb
,
1301 .detach_on_drain
= true,
1303 acb
= blk_aio_preadv(blk
, 0, &qiov
, 0, detach_by_parent_aio_cb
, NULL
);
1304 g_assert(acb
!= NULL
);
1306 /* Drain and check the expected result */
1307 bdrv_drained_begin(parent_b
);
1308 bdrv_drained_begin(a
);
1309 bdrv_drained_begin(b
);
1310 bdrv_drained_begin(c
);
1312 g_assert(detach_by_parent_data
.child_c
!= NULL
);
1314 g_assert_cmpint(parent_a
->refcnt
, ==, 1);
1315 g_assert_cmpint(parent_b
->refcnt
, ==, 1);
1316 g_assert_cmpint(a
->refcnt
, ==, 3);
1317 g_assert_cmpint(b
->refcnt
, ==, 1);
1318 g_assert_cmpint(c
->refcnt
, ==, 2);
1320 g_assert(QLIST_FIRST(&parent_b
->children
) == detach_by_parent_data
.child_c
);
1321 g_assert(QLIST_NEXT(detach_by_parent_data
.child_c
, next
) == child_a
);
1322 g_assert(QLIST_NEXT(child_a
, next
) == NULL
);
1324 g_assert_cmpint(parent_a
->quiesce_counter
, ==, 1);
1325 g_assert_cmpint(parent_b
->quiesce_counter
, ==, 3);
1326 g_assert_cmpint(a
->quiesce_counter
, ==, 1);
1327 g_assert_cmpint(b
->quiesce_counter
, ==, 1);
1328 g_assert_cmpint(c
->quiesce_counter
, ==, 1);
1330 bdrv_drained_end(parent_b
);
1331 bdrv_drained_end(a
);
1332 bdrv_drained_end(b
);
1333 bdrv_drained_end(c
);
1335 bdrv_unref(parent_b
);
1338 g_assert_cmpint(a
->refcnt
, ==, 1);
1339 g_assert_cmpint(b
->refcnt
, ==, 1);
1340 g_assert_cmpint(c
->refcnt
, ==, 1);
1346 static void test_detach_by_parent_cb(void)
1348 test_detach_indirect(true);
1351 static void test_detach_by_driver_cb(void)
1353 test_detach_indirect(false);
1356 static void test_append_to_drained(void)
1359 BlockDriverState
*base
, *overlay
;
1360 BDRVTestState
*base_s
, *overlay_s
;
1362 blk
= blk_new(qemu_get_aio_context(), BLK_PERM_ALL
, BLK_PERM_ALL
);
1363 base
= bdrv_new_open_driver(&bdrv_test
, "base", BDRV_O_RDWR
, &error_abort
);
1364 base_s
= base
->opaque
;
1365 blk_insert_bs(blk
, base
, &error_abort
);
1367 overlay
= bdrv_new_open_driver(&bdrv_test
, "overlay", BDRV_O_RDWR
,
1369 overlay_s
= overlay
->opaque
;
1371 do_drain_begin(BDRV_DRAIN
, base
);
1372 g_assert_cmpint(base
->quiesce_counter
, ==, 1);
1373 g_assert_cmpint(base_s
->drain_count
, ==, 1);
1374 g_assert_cmpint(base
->in_flight
, ==, 0);
1376 aio_context_acquire(qemu_get_aio_context());
1377 bdrv_append(overlay
, base
, &error_abort
);
1378 aio_context_release(qemu_get_aio_context());
1380 g_assert_cmpint(base
->in_flight
, ==, 0);
1381 g_assert_cmpint(overlay
->in_flight
, ==, 0);
1383 g_assert_cmpint(base
->quiesce_counter
, ==, 1);
1384 g_assert_cmpint(base_s
->drain_count
, ==, 1);
1385 g_assert_cmpint(overlay
->quiesce_counter
, ==, 1);
1386 g_assert_cmpint(overlay_s
->drain_count
, ==, 1);
1388 do_drain_end(BDRV_DRAIN
, base
);
1390 g_assert_cmpint(base
->quiesce_counter
, ==, 0);
1391 g_assert_cmpint(base_s
->drain_count
, ==, 0);
1392 g_assert_cmpint(overlay
->quiesce_counter
, ==, 0);
1393 g_assert_cmpint(overlay_s
->drain_count
, ==, 0);
1395 bdrv_unref(overlay
);
1400 static void test_set_aio_context(void)
1402 BlockDriverState
*bs
;
1403 IOThread
*a
= iothread_new();
1404 IOThread
*b
= iothread_new();
1405 AioContext
*ctx_a
= iothread_get_aio_context(a
);
1406 AioContext
*ctx_b
= iothread_get_aio_context(b
);
1408 bs
= bdrv_new_open_driver(&bdrv_test
, "test-node", BDRV_O_RDWR
,
1411 bdrv_drained_begin(bs
);
1412 bdrv_try_change_aio_context(bs
, ctx_a
, NULL
, &error_abort
);
1414 aio_context_acquire(ctx_a
);
1415 bdrv_drained_end(bs
);
1417 bdrv_drained_begin(bs
);
1418 bdrv_try_change_aio_context(bs
, ctx_b
, NULL
, &error_abort
);
1419 aio_context_release(ctx_a
);
1420 aio_context_acquire(ctx_b
);
1421 bdrv_try_change_aio_context(bs
, qemu_get_aio_context(), NULL
, &error_abort
);
1422 aio_context_release(ctx_b
);
1423 bdrv_drained_end(bs
);
1431 typedef struct TestDropBackingBlockJob
{
1433 bool should_complete
;
1435 BlockDriverState
*detach_also
;
1436 BlockDriverState
*bs
;
1437 } TestDropBackingBlockJob
;
1439 static int coroutine_fn
test_drop_backing_job_run(Job
*job
, Error
**errp
)
1441 TestDropBackingBlockJob
*s
=
1442 container_of(job
, TestDropBackingBlockJob
, common
.job
);
1444 while (!s
->should_complete
) {
1445 job_sleep_ns(job
, 0);
1451 static void test_drop_backing_job_commit(Job
*job
)
1453 TestDropBackingBlockJob
*s
=
1454 container_of(job
, TestDropBackingBlockJob
, common
.job
);
1456 bdrv_set_backing_hd(s
->bs
, NULL
, &error_abort
);
1457 bdrv_set_backing_hd(s
->detach_also
, NULL
, &error_abort
);
1459 *s
->did_complete
= true;
1462 static const BlockJobDriver test_drop_backing_job_driver
= {
1464 .instance_size
= sizeof(TestDropBackingBlockJob
),
1465 .free
= block_job_free
,
1466 .user_resume
= block_job_user_resume
,
1467 .run
= test_drop_backing_job_run
,
1468 .commit
= test_drop_backing_job_commit
,
1473 * Creates a child node with three parent nodes on it, and then runs a
1474 * block job on the final one, parent-node-2.
1476 * The job is then asked to complete before a section where the child
1479 * Ending this section will undrain the child's parents, first
1480 * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
1481 * list is in reverse order of how they were added. Ending the drain
1482 * on parent-node-2 will resume the job, thus completing it and
1483 * scheduling job_exit().
1485 * Ending the drain on parent-node-1 will poll the AioContext, which
1486 * lets job_exit() and thus test_drop_backing_job_commit() run. That
1487 * function first removes the child as parent-node-2's backing file.
1489 * In old (and buggy) implementations, there are two problems with
1491 * (A) bdrv_drain_invoke() polls for every node that leaves the
1492 * drained section. This means that job_exit() is scheduled
1493 * before the child has left the drained section. Its
1494 * quiesce_counter is therefore still 1 when it is removed from
1497 * (B) bdrv_replace_child_noperm() calls drained_end() on the old
1498 * child's parents as many times as the child is quiesced. This
1499 * means it will call drained_end() on parent-node-2 once.
1500 * Because parent-node-2 is no longer quiesced at this point, this
1503 * bdrv_replace_child_noperm() therefore must call drained_end() on
1504 * the parent only if it really is still drained because the child is
1507 * If removing child from parent-node-2 was successful (as it should
1508 * be), test_drop_backing_job_commit() will then also remove the child
1509 * from parent-node-0.
1511 * With an old version of our drain infrastructure ((A) above), that
1512 * resulted in the following flow:
1514 * 1. child attempts to leave its drained section. The call recurses
1517 * 2. parent-node-2 leaves the drained section. Polling in
1518 * bdrv_drain_invoke() will schedule job_exit().
1520 * 3. parent-node-1 leaves the drained section. Polling in
1521 * bdrv_drain_invoke() will run job_exit(), thus disconnecting
1522 * parent-node-0 from the child node.
1524 * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
1525 * iterate over the parents. Thus, it now accesses the BdrvChild
1526 * object that used to connect parent-node-0 and the child node.
1527 * However, that object no longer exists, so it accesses a dangling
1530 * The solution is to only poll once when running a bdrv_drained_end()
1531 * operation, specifically at the end when all drained_end()
1532 * operations for all involved nodes have been scheduled.
1533 * Note that this also solves (A) above, thus hiding (B).
1535 static void test_blockjob_commit_by_drained_end(void)
1537 BlockDriverState
*bs_child
, *bs_parents
[3];
1538 TestDropBackingBlockJob
*job
;
1539 bool job_has_completed
= false;
1542 bs_child
= bdrv_new_open_driver(&bdrv_test
, "child-node", BDRV_O_RDWR
,
1545 for (i
= 0; i
< 3; i
++) {
1547 snprintf(name
, sizeof(name
), "parent-node-%i", i
);
1548 bs_parents
[i
] = bdrv_new_open_driver(&bdrv_test
, name
, BDRV_O_RDWR
,
1550 bdrv_set_backing_hd(bs_parents
[i
], bs_child
, &error_abort
);
1553 job
= block_job_create("job", &test_drop_backing_job_driver
, NULL
,
1554 bs_parents
[2], 0, BLK_PERM_ALL
, 0, 0, NULL
, NULL
,
1556 job
->bs
= bs_parents
[2];
1558 job
->detach_also
= bs_parents
[0];
1559 job
->did_complete
= &job_has_completed
;
1561 job_start(&job
->common
.job
);
1563 job
->should_complete
= true;
1564 bdrv_drained_begin(bs_child
);
1565 g_assert(!job_has_completed
);
1566 bdrv_drained_end(bs_child
);
1567 aio_poll(qemu_get_aio_context(), false);
1568 g_assert(job_has_completed
);
1570 bdrv_unref(bs_parents
[0]);
1571 bdrv_unref(bs_parents
[1]);
1572 bdrv_unref(bs_parents
[2]);
1573 bdrv_unref(bs_child
);
1577 typedef struct TestSimpleBlockJob
{
1579 bool should_complete
;
1581 } TestSimpleBlockJob
;
1583 static int coroutine_fn
test_simple_job_run(Job
*job
, Error
**errp
)
1585 TestSimpleBlockJob
*s
= container_of(job
, TestSimpleBlockJob
, common
.job
);
1587 while (!s
->should_complete
) {
1588 job_sleep_ns(job
, 0);
1594 static void test_simple_job_clean(Job
*job
)
1596 TestSimpleBlockJob
*s
= container_of(job
, TestSimpleBlockJob
, common
.job
);
1597 *s
->did_complete
= true;
1600 static const BlockJobDriver test_simple_job_driver
= {
1602 .instance_size
= sizeof(TestSimpleBlockJob
),
1603 .free
= block_job_free
,
1604 .user_resume
= block_job_user_resume
,
1605 .run
= test_simple_job_run
,
1606 .clean
= test_simple_job_clean
,
1610 static int drop_intermediate_poll_update_filename(BdrvChild
*child
,
1611 BlockDriverState
*new_base
,
1612 const char *filename
,
1616 * We are free to poll here, which may change the block graph, if
1617 * it is not drained.
1620 /* If the job is not drained: Complete it, schedule job_exit() */
1621 aio_poll(qemu_get_current_aio_context(), false);
1622 /* If the job is not drained: Run job_exit(), finish the job */
1623 aio_poll(qemu_get_current_aio_context(), false);
1629 * Test a poll in the midst of bdrv_drop_intermediate().
1631 * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(),
1632 * which can yield or poll. This may lead to graph changes, unless
1633 * the whole subtree in question is drained.
1635 * We test this on the following graph:
1651 * node-2 --chain--> node-1 --chain--> node-0
1653 * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0).
1655 * This first updates node-2's backing filename by invoking
1656 * drop_intermediate_poll_update_filename(), which polls twice. This
1657 * causes the job to finish, which in turns causes the job-node to be
1660 * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it
1661 * already has a pointer to the BdrvChild edge between job-node and
1662 * node-1. When it tries to handle that edge, we probably get a
1663 * segmentation fault because the object no longer exists.
1666 * The solution is for bdrv_drop_intermediate() to drain top's
1667 * subtree. This prevents graph changes from happening just because
1668 * BdrvChildClass.update_filename() yields or polls. Thus, the block
1669 * job is paused during that drained section and must finish before or
1672 * (In addition, bdrv_replace_child() must keep the job paused.)
1674 static void test_drop_intermediate_poll(void)
1676 static BdrvChildClass chain_child_class
;
1677 BlockDriverState
*chain
[3];
1678 TestSimpleBlockJob
*job
;
1679 BlockDriverState
*job_node
;
1680 bool job_has_completed
= false;
1684 chain_child_class
= child_of_bds
;
1685 chain_child_class
.update_filename
= drop_intermediate_poll_update_filename
;
1687 for (i
= 0; i
< 3; i
++) {
1689 snprintf(name
, 32, "node-%i", i
);
1691 chain
[i
] = bdrv_new_open_driver(&bdrv_test
, name
, 0, &error_abort
);
1694 job_node
= bdrv_new_open_driver(&bdrv_test
, "job-node", BDRV_O_RDWR
,
1696 bdrv_set_backing_hd(job_node
, chain
[1], &error_abort
);
1699 * Establish the chain last, so the chain links are the first
1700 * elements in the BDS.parents lists
1702 bdrv_graph_wrlock(NULL
);
1703 for (i
= 0; i
< 3; i
++) {
1705 /* Takes the reference to chain[i - 1] */
1706 bdrv_attach_child(chain
[i
], chain
[i
- 1], "chain",
1707 &chain_child_class
, BDRV_CHILD_COW
, &error_abort
);
1710 bdrv_graph_wrunlock();
1712 job
= block_job_create("job", &test_simple_job_driver
, NULL
, job_node
,
1713 0, BLK_PERM_ALL
, 0, 0, NULL
, NULL
, &error_abort
);
1715 /* The job has a reference now */
1716 bdrv_unref(job_node
);
1718 job
->did_complete
= &job_has_completed
;
1720 job_start(&job
->common
.job
);
1721 job
->should_complete
= true;
1723 g_assert(!job_has_completed
);
1724 ret
= bdrv_drop_intermediate(chain
[1], chain
[0], NULL
);
1725 aio_poll(qemu_get_aio_context(), false);
1727 g_assert(job_has_completed
);
1729 bdrv_unref(chain
[2]);
1733 typedef struct BDRVReplaceTestState
{
1734 bool setup_completed
;
1741 bool yield_before_read
;
1743 Coroutine
*drain_co
;
1744 } BDRVReplaceTestState
;
1746 static void bdrv_replace_test_close(BlockDriverState
*bs
)
1751 * If @bs has a backing file:
1752 * Yield if .yield_before_read is true (and wait for drain_begin to
1754 * Forward the read to bs->backing. Set .has_read to true.
1755 * If drain_begin has woken us, wake it in turn.
1758 * Set .has_read to true and return success.
1760 static int coroutine_fn GRAPH_RDLOCK
1761 bdrv_replace_test_co_preadv(BlockDriverState
*bs
, int64_t offset
, int64_t bytes
,
1762 QEMUIOVector
*qiov
, BdrvRequestFlags flags
)
1764 BDRVReplaceTestState
*s
= bs
->opaque
;
1769 g_assert(!s
->drain_count
);
1771 s
->io_co
= qemu_coroutine_self();
1772 if (s
->yield_before_read
) {
1773 s
->yield_before_read
= false;
1774 qemu_coroutine_yield();
1778 ret
= bdrv_co_preadv(bs
->backing
, offset
, bytes
, qiov
, 0);
1781 /* Wake up drain_co if it runs */
1783 aio_co_wake(s
->drain_co
);
1793 static void coroutine_fn
bdrv_replace_test_drain_co(void *opaque
)
1795 BlockDriverState
*bs
= opaque
;
1796 BDRVReplaceTestState
*s
= bs
->opaque
;
1798 /* Keep waking io_co up until it is done */
1800 aio_co_wake(s
->io_co
);
1802 qemu_coroutine_yield();
1805 bdrv_dec_in_flight(bs
);
1809 * If .drain_count is 0, wake up .io_co if there is one; and set
1811 * Increment .drain_count.
1813 static void bdrv_replace_test_drain_begin(BlockDriverState
*bs
)
1815 BDRVReplaceTestState
*s
= bs
->opaque
;
1817 if (!s
->setup_completed
) {
1821 if (!s
->drain_count
) {
1822 s
->drain_co
= qemu_coroutine_create(bdrv_replace_test_drain_co
, bs
);
1823 bdrv_inc_in_flight(bs
);
1824 aio_co_enter(bdrv_get_aio_context(bs
), s
->drain_co
);
1825 s
->was_drained
= true;
1830 static void coroutine_fn
bdrv_replace_test_read_entry(void *opaque
)
1832 BlockDriverState
*bs
= opaque
;
1834 QEMUIOVector qiov
= QEMU_IOVEC_INIT_BUF(qiov
, &data
, 1);
1837 /* Queue a read request post-drain */
1838 bdrv_graph_co_rdlock();
1839 ret
= bdrv_replace_test_co_preadv(bs
, 0, 1, &qiov
, 0);
1840 bdrv_graph_co_rdunlock();
1843 bdrv_dec_in_flight(bs
);
1847 * Reduce .drain_count, set .was_undrained once it reaches 0.
1848 * If .drain_count reaches 0 and the node has a backing file, issue a
1851 static void bdrv_replace_test_drain_end(BlockDriverState
*bs
)
1853 BDRVReplaceTestState
*s
= bs
->opaque
;
1855 if (!s
->setup_completed
) {
1859 g_assert(s
->drain_count
> 0);
1860 if (!--s
->drain_count
) {
1861 s
->was_undrained
= true;
1864 Coroutine
*co
= qemu_coroutine_create(bdrv_replace_test_read_entry
,
1866 bdrv_inc_in_flight(bs
);
1867 aio_co_enter(bdrv_get_aio_context(bs
), co
);
1872 static BlockDriver bdrv_replace_test
= {
1873 .format_name
= "replace_test",
1874 .instance_size
= sizeof(BDRVReplaceTestState
),
1875 .supports_backing
= true,
1877 .bdrv_close
= bdrv_replace_test_close
,
1878 .bdrv_co_preadv
= bdrv_replace_test_co_preadv
,
1880 .bdrv_drain_begin
= bdrv_replace_test_drain_begin
,
1881 .bdrv_drain_end
= bdrv_replace_test_drain_end
,
1883 .bdrv_child_perm
= bdrv_default_perms
,
1886 static void coroutine_fn
test_replace_child_mid_drain_read_co(void *opaque
)
1891 ret
= blk_co_pread(opaque
, 0, 1, &data
, 0);
1896 * We test two things:
1897 * (1) bdrv_replace_child_noperm() must not undrain the parent if both
1898 * children are drained.
1899 * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
1900 * drained child. If the old child is drained, it must flush I/O
1901 * requests after the new one has been attached. If the new child
1902 * is drained, it must flush I/O requests before the old one is
1905 * To do so, we create one parent node and two child nodes; then
1906 * attach one of the children (old_child_bs) to the parent, then
1907 * drain both old_child_bs and new_child_bs according to
1908 * old_drain_count and new_drain_count, respectively, and finally
1909 * we invoke bdrv_replace_node() to replace old_child_bs by
1912 * The test block driver we use here (bdrv_replace_test) has a read
1914 * - For the parent node, can optionally yield, and then forwards the
1915 * read to bdrv_preadv(),
1916 * - For the child node, just returns immediately.
1918 * If the read yields, the drain_begin function will wake it up.
1920 * The drain_end function issues a read on the parent once it is fully
1921 * undrained (which simulates requests starting to come in again).
1923 static void do_test_replace_child_mid_drain(int old_drain_count
,
1924 int new_drain_count
)
1926 BlockBackend
*parent_blk
;
1927 BlockDriverState
*parent_bs
;
1928 BlockDriverState
*old_child_bs
, *new_child_bs
;
1929 BDRVReplaceTestState
*parent_s
;
1930 BDRVReplaceTestState
*old_child_s
, *new_child_s
;
1934 parent_bs
= bdrv_new_open_driver(&bdrv_replace_test
, "parent", 0,
1936 parent_s
= parent_bs
->opaque
;
1938 parent_blk
= blk_new(qemu_get_aio_context(),
1939 BLK_PERM_CONSISTENT_READ
, BLK_PERM_ALL
);
1940 blk_insert_bs(parent_blk
, parent_bs
, &error_abort
);
1942 old_child_bs
= bdrv_new_open_driver(&bdrv_replace_test
, "old-child", 0,
1944 new_child_bs
= bdrv_new_open_driver(&bdrv_replace_test
, "new-child", 0,
1946 old_child_s
= old_child_bs
->opaque
;
1947 new_child_s
= new_child_bs
->opaque
;
1949 /* So that we can read something */
1950 parent_bs
->total_sectors
= 1;
1951 old_child_bs
->total_sectors
= 1;
1952 new_child_bs
->total_sectors
= 1;
1954 bdrv_ref(old_child_bs
);
1955 bdrv_graph_wrlock(NULL
);
1956 bdrv_attach_child(parent_bs
, old_child_bs
, "child", &child_of_bds
,
1957 BDRV_CHILD_COW
, &error_abort
);
1958 bdrv_graph_wrunlock();
1959 parent_s
->setup_completed
= true;
1961 for (i
= 0; i
< old_drain_count
; i
++) {
1962 bdrv_drained_begin(old_child_bs
);
1964 for (i
= 0; i
< new_drain_count
; i
++) {
1965 bdrv_drained_begin(new_child_bs
);
1968 if (!old_drain_count
) {
1970 * Start a read operation that will yield, so it will not
1971 * complete before the node is drained.
1973 parent_s
->yield_before_read
= true;
1974 io_co
= qemu_coroutine_create(test_replace_child_mid_drain_read_co
,
1976 qemu_coroutine_enter(io_co
);
1979 /* If we have started a read operation, it should have yielded */
1980 g_assert(!parent_s
->has_read
);
1982 /* Reset drained status so we can see what bdrv_replace_node() does */
1983 parent_s
->was_drained
= false;
1984 parent_s
->was_undrained
= false;
1986 g_assert(parent_bs
->quiesce_counter
== old_drain_count
);
1987 bdrv_replace_node(old_child_bs
, new_child_bs
, &error_abort
);
1988 g_assert(parent_bs
->quiesce_counter
== new_drain_count
);
1990 if (!old_drain_count
&& !new_drain_count
) {
1992 * From undrained to undrained drains and undrains the parent,
1993 * because bdrv_replace_node() contains a drained section for
1996 g_assert(parent_s
->was_drained
&& parent_s
->was_undrained
);
1997 } else if (!old_drain_count
&& new_drain_count
) {
1999 * From undrained to drained should drain the parent and keep
2002 g_assert(parent_s
->was_drained
&& !parent_s
->was_undrained
);
2003 } else if (old_drain_count
&& !new_drain_count
) {
2005 * From drained to undrained should undrain the parent and
2008 g_assert(!parent_s
->was_drained
&& parent_s
->was_undrained
);
2009 } else /* if (old_drain_count && new_drain_count) */ {
2011 * From drained to drained must not undrain the parent at any
2014 g_assert(!parent_s
->was_drained
&& !parent_s
->was_undrained
);
2017 if (!old_drain_count
|| !new_drain_count
) {
2019 * If !old_drain_count, we have started a read request before
2020 * bdrv_replace_node(). If !new_drain_count, the parent must
2021 * have been undrained at some point, and
2022 * bdrv_replace_test_co_drain_end() starts a read request
2025 g_assert(parent_s
->has_read
);
2028 * If the parent was never undrained, there is no way to start
2031 g_assert(!parent_s
->has_read
);
2034 /* A drained child must have not received any request */
2035 g_assert(!(old_drain_count
&& old_child_s
->has_read
));
2036 g_assert(!(new_drain_count
&& new_child_s
->has_read
));
2038 for (i
= 0; i
< new_drain_count
; i
++) {
2039 bdrv_drained_end(new_child_bs
);
2041 for (i
= 0; i
< old_drain_count
; i
++) {
2042 bdrv_drained_end(old_child_bs
);
2046 * By now, bdrv_replace_test_co_drain_end() must have been called
2047 * at some point while the new child was attached to the parent.
2049 g_assert(parent_s
->has_read
);
2050 g_assert(new_child_s
->has_read
);
2052 blk_unref(parent_blk
);
2053 bdrv_unref(parent_bs
);
2054 bdrv_unref(old_child_bs
);
2055 bdrv_unref(new_child_bs
);
2058 static void test_replace_child_mid_drain(void)
2060 int old_drain_count
, new_drain_count
;
2062 for (old_drain_count
= 0; old_drain_count
< 2; old_drain_count
++) {
2063 for (new_drain_count
= 0; new_drain_count
< 2; new_drain_count
++) {
2064 do_test_replace_child_mid_drain(old_drain_count
, new_drain_count
);
2069 int main(int argc
, char **argv
)
2074 qemu_init_main_loop(&error_abort
);
2076 g_test_init(&argc
, &argv
, NULL
);
2077 qemu_event_init(&done_event
, false);
2079 g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all
);
2080 g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain
);
2082 g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
2083 test_drv_cb_co_drain_all
);
2084 g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain
);
2086 g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all
);
2087 g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain
);
2089 g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
2090 test_quiesce_co_drain_all
);
2091 g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain
);
2093 g_test_add_func("/bdrv-drain/nested", test_nested
);
2095 g_test_add_func("/bdrv-drain/graph-change/drain_all",
2096 test_graph_change_drain_all
);
2098 g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all
);
2099 g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain
);
2101 g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all
);
2102 g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain
);
2104 g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
2105 test_blockjob_error_drain_all
);
2106 g_test_add_func("/bdrv-drain/blockjob/error/drain",
2107 test_blockjob_error_drain
);
2109 g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
2110 test_blockjob_iothread_drain_all
);
2111 g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
2112 test_blockjob_iothread_drain
);
2114 g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
2115 test_blockjob_iothread_error_drain_all
);
2116 g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
2117 test_blockjob_iothread_error_drain
);
2119 g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain
);
2120 g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all
);
2121 g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain
);
2122 g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb
);
2123 g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb
);
2125 g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained
);
2127 g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context
);
2129 g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
2130 test_blockjob_commit_by_drained_end
);
2132 g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
2133 test_drop_intermediate_poll
);
2135 g_test_add_func("/bdrv-drain/replace_child/mid-drain",
2136 test_replace_child_mid_drain
);
2139 qemu_event_destroy(&done_event
);