xref: /openbmc/qemu/tests/unit/test-bdrv-drain.c (revision 4921d0a7)
1 /*
2  * Block node draining tests
3  *
4  * Copyright (c) 2017 Kevin Wolf <kwolf@redhat.com>
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  */
24 
25 #include "qemu/osdep.h"
26 #include "block/block_int.h"
27 #include "block/blockjob_int.h"
28 #include "sysemu/block-backend.h"
29 #include "qapi/error.h"
30 #include "qemu/main-loop.h"
31 #include "iothread.h"
32 
33 static QemuEvent done_event;
34 
35 typedef struct BDRVTestState {
36     int drain_count;
37     AioContext *bh_indirection_ctx;
38     bool sleep_in_drain_begin;
39 } BDRVTestState;
40 
41 static void coroutine_fn sleep_in_drain_begin(void *opaque)
42 {
43     BlockDriverState *bs = opaque;
44 
45     qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
46     bdrv_dec_in_flight(bs);
47 }
48 
49 static void bdrv_test_drain_begin(BlockDriverState *bs)
50 {
51     BDRVTestState *s = bs->opaque;
52     s->drain_count++;
53     if (s->sleep_in_drain_begin) {
54         Coroutine *co = qemu_coroutine_create(sleep_in_drain_begin, bs);
55         bdrv_inc_in_flight(bs);
56         aio_co_enter(bdrv_get_aio_context(bs), co);
57     }
58 }
59 
60 static void bdrv_test_drain_end(BlockDriverState *bs)
61 {
62     BDRVTestState *s = bs->opaque;
63     s->drain_count--;
64 }
65 
66 static void bdrv_test_close(BlockDriverState *bs)
67 {
68     BDRVTestState *s = bs->opaque;
69     g_assert_cmpint(s->drain_count, >, 0);
70 }
71 
72 static void co_reenter_bh(void *opaque)
73 {
74     aio_co_wake(opaque);
75 }
76 
77 static int coroutine_fn bdrv_test_co_preadv(BlockDriverState *bs,
78                                             int64_t offset, int64_t bytes,
79                                             QEMUIOVector *qiov,
80                                             BdrvRequestFlags flags)
81 {
82     BDRVTestState *s = bs->opaque;
83 
84     /* We want this request to stay until the polling loop in drain waits for
85      * it to complete. We need to sleep a while as bdrv_drain_invoke() comes
86      * first and polls its result, too, but it shouldn't accidentally complete
87      * this request yet. */
88     qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
89 
90     if (s->bh_indirection_ctx) {
91         aio_bh_schedule_oneshot(s->bh_indirection_ctx, co_reenter_bh,
92                                 qemu_coroutine_self());
93         qemu_coroutine_yield();
94     }
95 
96     return 0;
97 }
98 
99 static int bdrv_test_co_change_backing_file(BlockDriverState *bs,
100                                             const char *backing_file,
101                                             const char *backing_fmt)
102 {
103     return 0;
104 }
105 
106 static BlockDriver bdrv_test = {
107     .format_name            = "test",
108     .instance_size          = sizeof(BDRVTestState),
109     .supports_backing       = true,
110 
111     .bdrv_close             = bdrv_test_close,
112     .bdrv_co_preadv         = bdrv_test_co_preadv,
113 
114     .bdrv_drain_begin       = bdrv_test_drain_begin,
115     .bdrv_drain_end         = bdrv_test_drain_end,
116 
117     .bdrv_child_perm        = bdrv_default_perms,
118 
119     .bdrv_co_change_backing_file = bdrv_test_co_change_backing_file,
120 };
121 
122 static void aio_ret_cb(void *opaque, int ret)
123 {
124     int *aio_ret = opaque;
125     *aio_ret = ret;
126 }
127 
128 typedef struct CallInCoroutineData {
129     void (*entry)(void);
130     bool done;
131 } CallInCoroutineData;
132 
133 static coroutine_fn void call_in_coroutine_entry(void *opaque)
134 {
135     CallInCoroutineData *data = opaque;
136 
137     data->entry();
138     data->done = true;
139 }
140 
141 static void call_in_coroutine(void (*entry)(void))
142 {
143     Coroutine *co;
144     CallInCoroutineData data = {
145         .entry  = entry,
146         .done   = false,
147     };
148 
149     co = qemu_coroutine_create(call_in_coroutine_entry, &data);
150     qemu_coroutine_enter(co);
151     while (!data.done) {
152         aio_poll(qemu_get_aio_context(), true);
153     }
154 }
155 
156 enum drain_type {
157     BDRV_DRAIN_ALL,
158     BDRV_DRAIN,
159     DRAIN_TYPE_MAX,
160 };
161 
162 static void do_drain_begin(enum drain_type drain_type, BlockDriverState *bs)
163 {
164     switch (drain_type) {
165     case BDRV_DRAIN_ALL:        bdrv_drain_all_begin(); break;
166     case BDRV_DRAIN:            bdrv_drained_begin(bs); break;
167     default:                    g_assert_not_reached();
168     }
169 }
170 
171 static void do_drain_end(enum drain_type drain_type, BlockDriverState *bs)
172 {
173     switch (drain_type) {
174     case BDRV_DRAIN_ALL:        bdrv_drain_all_end(); break;
175     case BDRV_DRAIN:            bdrv_drained_end(bs); break;
176     default:                    g_assert_not_reached();
177     }
178 }
179 
180 static void do_drain_begin_unlocked(enum drain_type drain_type, BlockDriverState *bs)
181 {
182     do_drain_begin(drain_type, bs);
183 }
184 
185 static BlockBackend * no_coroutine_fn test_setup(void)
186 {
187     BlockBackend *blk;
188     BlockDriverState *bs, *backing;
189 
190     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
191     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
192                               &error_abort);
193     blk_insert_bs(blk, bs, &error_abort);
194 
195     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
196     bdrv_set_backing_hd(bs, backing, &error_abort);
197 
198     bdrv_unref(backing);
199     bdrv_unref(bs);
200 
201     return blk;
202 }
203 
204 static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs)
205 {
206     do_drain_end(drain_type, bs);
207 }
208 
209 /*
210  * Locking the block graph would be a bit cumbersome here because this function
211  * is called both in coroutine and non-coroutine context. We know this is a test
212  * and nothing else is running, so don't bother with TSA.
213  */
214 static void coroutine_mixed_fn TSA_NO_TSA
215 test_drv_cb_common(BlockBackend *blk, enum drain_type drain_type,
216                    bool recursive)
217 {
218     BlockDriverState *bs = blk_bs(blk);
219     BlockDriverState *backing = bs->backing->bs;
220     BDRVTestState *s, *backing_s;
221     BlockAIOCB *acb;
222     int aio_ret;
223 
224     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
225 
226     s = bs->opaque;
227     backing_s = backing->opaque;
228 
229     /* Simple bdrv_drain_all_begin/end pair, check that CBs are called */
230     g_assert_cmpint(s->drain_count, ==, 0);
231     g_assert_cmpint(backing_s->drain_count, ==, 0);
232 
233     do_drain_begin(drain_type, bs);
234 
235     g_assert_cmpint(s->drain_count, ==, 1);
236     g_assert_cmpint(backing_s->drain_count, ==, !!recursive);
237 
238     do_drain_end(drain_type, bs);
239 
240     g_assert_cmpint(s->drain_count, ==, 0);
241     g_assert_cmpint(backing_s->drain_count, ==, 0);
242 
243     /* Now do the same while a request is pending */
244     aio_ret = -EINPROGRESS;
245     acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
246     g_assert(acb != NULL);
247     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
248 
249     g_assert_cmpint(s->drain_count, ==, 0);
250     g_assert_cmpint(backing_s->drain_count, ==, 0);
251 
252     do_drain_begin(drain_type, bs);
253 
254     g_assert_cmpint(aio_ret, ==, 0);
255     g_assert_cmpint(s->drain_count, ==, 1);
256     g_assert_cmpint(backing_s->drain_count, ==, !!recursive);
257 
258     do_drain_end(drain_type, bs);
259 
260     g_assert_cmpint(s->drain_count, ==, 0);
261     g_assert_cmpint(backing_s->drain_count, ==, 0);
262 }
263 
264 static void test_drv_cb_drain_all(void)
265 {
266     BlockBackend *blk = test_setup();
267     test_drv_cb_common(blk, BDRV_DRAIN_ALL, true);
268     blk_unref(blk);
269 }
270 
271 static void test_drv_cb_drain(void)
272 {
273     BlockBackend *blk = test_setup();
274     test_drv_cb_common(blk, BDRV_DRAIN, false);
275     blk_unref(blk);
276 }
277 
278 static void coroutine_fn test_drv_cb_co_drain_all_entry(void)
279 {
280     BlockBackend *blk = blk_all_next(NULL);
281     test_drv_cb_common(blk, BDRV_DRAIN_ALL, true);
282 }
283 
284 static void test_drv_cb_co_drain_all(void)
285 {
286     BlockBackend *blk = test_setup();
287     call_in_coroutine(test_drv_cb_co_drain_all_entry);
288     blk_unref(blk);
289 }
290 
291 static void coroutine_fn test_drv_cb_co_drain_entry(void)
292 {
293     BlockBackend *blk = blk_all_next(NULL);
294     test_drv_cb_common(blk, BDRV_DRAIN, false);
295 }
296 
297 static void test_drv_cb_co_drain(void)
298 {
299     BlockBackend *blk = test_setup();
300     call_in_coroutine(test_drv_cb_co_drain_entry);
301     blk_unref(blk);
302 }
303 
304 /*
305  * Locking the block graph would be a bit cumbersome here because this function
306  * is called both in coroutine and non-coroutine context. We know this is a test
307  * and nothing else is running, so don't bother with TSA.
308  */
309 static void coroutine_mixed_fn TSA_NO_TSA
310 test_quiesce_common(BlockBackend *blk, enum drain_type drain_type,
311                     bool recursive)
312 {
313     BlockDriverState *bs = blk_bs(blk);
314     BlockDriverState *backing = bs->backing->bs;
315 
316     g_assert_cmpint(bs->quiesce_counter, ==, 0);
317     g_assert_cmpint(backing->quiesce_counter, ==, 0);
318 
319     do_drain_begin(drain_type, bs);
320 
321     if (drain_type == BDRV_DRAIN_ALL) {
322         g_assert_cmpint(bs->quiesce_counter, ==, 2);
323     } else {
324         g_assert_cmpint(bs->quiesce_counter, ==, 1);
325     }
326     g_assert_cmpint(backing->quiesce_counter, ==, !!recursive);
327 
328     do_drain_end(drain_type, bs);
329 
330     g_assert_cmpint(bs->quiesce_counter, ==, 0);
331     g_assert_cmpint(backing->quiesce_counter, ==, 0);
332 }
333 
334 static void test_quiesce_drain_all(void)
335 {
336     BlockBackend *blk = test_setup();
337     test_quiesce_common(blk, BDRV_DRAIN_ALL, true);
338     blk_unref(blk);
339 }
340 
341 static void test_quiesce_drain(void)
342 {
343     BlockBackend *blk = test_setup();
344     test_quiesce_common(blk, BDRV_DRAIN, false);
345     blk_unref(blk);
346 }
347 
348 static void coroutine_fn test_quiesce_co_drain_all_entry(void)
349 {
350     BlockBackend *blk = blk_all_next(NULL);
351     test_quiesce_common(blk, BDRV_DRAIN_ALL, true);
352 }
353 
354 static void test_quiesce_co_drain_all(void)
355 {
356     BlockBackend *blk = test_setup();
357     call_in_coroutine(test_quiesce_co_drain_all_entry);
358     blk_unref(blk);
359 }
360 
361 static void coroutine_fn test_quiesce_co_drain_entry(void)
362 {
363     BlockBackend *blk = blk_all_next(NULL);
364     test_quiesce_common(blk, BDRV_DRAIN, false);
365 }
366 
367 static void test_quiesce_co_drain(void)
368 {
369     BlockBackend *blk = test_setup();
370     call_in_coroutine(test_quiesce_co_drain_entry);
371     blk_unref(blk);
372 }
373 
374 static void test_nested(void)
375 {
376     BlockBackend *blk;
377     BlockDriverState *bs, *backing;
378     BDRVTestState *s, *backing_s;
379     enum drain_type outer, inner;
380 
381     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
382     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
383                               &error_abort);
384     s = bs->opaque;
385     blk_insert_bs(blk, bs, &error_abort);
386 
387     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
388     backing_s = backing->opaque;
389     bdrv_set_backing_hd(bs, backing, &error_abort);
390 
391     for (outer = 0; outer < DRAIN_TYPE_MAX; outer++) {
392         for (inner = 0; inner < DRAIN_TYPE_MAX; inner++) {
393             int backing_quiesce = (outer == BDRV_DRAIN_ALL) +
394                                   (inner == BDRV_DRAIN_ALL);
395 
396             g_assert_cmpint(bs->quiesce_counter, ==, 0);
397             g_assert_cmpint(backing->quiesce_counter, ==, 0);
398             g_assert_cmpint(s->drain_count, ==, 0);
399             g_assert_cmpint(backing_s->drain_count, ==, 0);
400 
401             do_drain_begin(outer, bs);
402             do_drain_begin(inner, bs);
403 
404             g_assert_cmpint(bs->quiesce_counter, ==, 2 + !!backing_quiesce);
405             g_assert_cmpint(backing->quiesce_counter, ==, backing_quiesce);
406             g_assert_cmpint(s->drain_count, ==, 1);
407             g_assert_cmpint(backing_s->drain_count, ==, !!backing_quiesce);
408 
409             do_drain_end(inner, bs);
410             do_drain_end(outer, bs);
411 
412             g_assert_cmpint(bs->quiesce_counter, ==, 0);
413             g_assert_cmpint(backing->quiesce_counter, ==, 0);
414             g_assert_cmpint(s->drain_count, ==, 0);
415             g_assert_cmpint(backing_s->drain_count, ==, 0);
416         }
417     }
418 
419     bdrv_unref(backing);
420     bdrv_unref(bs);
421     blk_unref(blk);
422 }
423 
424 static void test_graph_change_drain_all(void)
425 {
426     BlockBackend *blk_a, *blk_b;
427     BlockDriverState *bs_a, *bs_b;
428     BDRVTestState *a_s, *b_s;
429 
430     /* Create node A with a BlockBackend */
431     blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
432     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
433                                 &error_abort);
434     a_s = bs_a->opaque;
435     blk_insert_bs(blk_a, bs_a, &error_abort);
436 
437     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
438     g_assert_cmpint(a_s->drain_count, ==, 0);
439 
440     /* Call bdrv_drain_all_begin() */
441     bdrv_drain_all_begin();
442 
443     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
444     g_assert_cmpint(a_s->drain_count, ==, 1);
445 
446     /* Create node B with a BlockBackend */
447     blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
448     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
449                                 &error_abort);
450     b_s = bs_b->opaque;
451     blk_insert_bs(blk_b, bs_b, &error_abort);
452 
453     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
454     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
455     g_assert_cmpint(a_s->drain_count, ==, 1);
456     g_assert_cmpint(b_s->drain_count, ==, 1);
457 
458     /* Unref and finally delete node A */
459     blk_unref(blk_a);
460 
461     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
462     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
463     g_assert_cmpint(a_s->drain_count, ==, 1);
464     g_assert_cmpint(b_s->drain_count, ==, 1);
465 
466     bdrv_unref(bs_a);
467 
468     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
469     g_assert_cmpint(b_s->drain_count, ==, 1);
470 
471     /* End the drained section */
472     bdrv_drain_all_end();
473 
474     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
475     g_assert_cmpint(b_s->drain_count, ==, 0);
476 
477     bdrv_unref(bs_b);
478     blk_unref(blk_b);
479 }
480 
481 struct test_iothread_data {
482     BlockDriverState *bs;
483     enum drain_type drain_type;
484     int *aio_ret;
485     bool co_done;
486 };
487 
488 static void coroutine_fn test_iothread_drain_co_entry(void *opaque)
489 {
490     struct test_iothread_data *data = opaque;
491 
492     do_drain_begin(data->drain_type, data->bs);
493     g_assert_cmpint(*data->aio_ret, ==, 0);
494     do_drain_end(data->drain_type, data->bs);
495 
496     data->co_done = true;
497     aio_wait_kick();
498 }
499 
500 static void test_iothread_aio_cb(void *opaque, int ret)
501 {
502     int *aio_ret = opaque;
503     *aio_ret = ret;
504     qemu_event_set(&done_event);
505 }
506 
507 static void test_iothread_main_thread_bh(void *opaque)
508 {
509     struct test_iothread_data *data = opaque;
510 
511     bdrv_flush(data->bs);
512     bdrv_dec_in_flight(data->bs); /* incremented by test_iothread_common() */
513 }
514 
515 /*
516  * Starts an AIO request on a BDS that runs in the AioContext of iothread 1.
517  * The request involves a BH on iothread 2 before it can complete.
518  *
519  * @drain_thread = 0 means that do_drain_begin/end are called from the main
520  * thread, @drain_thread = 1 means that they are called from iothread 1. Drain
521  * for this BDS cannot be called from iothread 2 because only the main thread
522  * may do cross-AioContext polling.
523  */
524 static void test_iothread_common(enum drain_type drain_type, int drain_thread)
525 {
526     BlockBackend *blk;
527     BlockDriverState *bs;
528     BDRVTestState *s;
529     BlockAIOCB *acb;
530     Coroutine *co;
531     int aio_ret;
532     struct test_iothread_data data;
533 
534     IOThread *a = iothread_new();
535     IOThread *b = iothread_new();
536     AioContext *ctx_a = iothread_get_aio_context(a);
537     AioContext *ctx_b = iothread_get_aio_context(b);
538 
539     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
540 
541     /* bdrv_drain_all() may only be called from the main loop thread */
542     if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) {
543         goto out;
544     }
545 
546     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
547     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
548                               &error_abort);
549     s = bs->opaque;
550     blk_insert_bs(blk, bs, &error_abort);
551     blk_set_disable_request_queuing(blk, true);
552 
553     blk_set_aio_context(blk, ctx_a, &error_abort);
554 
555     s->bh_indirection_ctx = ctx_b;
556 
557     aio_ret = -EINPROGRESS;
558     qemu_event_reset(&done_event);
559 
560     if (drain_thread == 0) {
561         acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret);
562     } else {
563         acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
564     }
565     g_assert(acb != NULL);
566     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
567 
568     data = (struct test_iothread_data) {
569         .bs         = bs,
570         .drain_type = drain_type,
571         .aio_ret    = &aio_ret,
572     };
573 
574     switch (drain_thread) {
575     case 0:
576         /*
577          * Increment in_flight so that do_drain_begin() waits for
578          * test_iothread_main_thread_bh(). This prevents the race between
579          * test_iothread_main_thread_bh() in IOThread a and do_drain_begin() in
580          * this thread. test_iothread_main_thread_bh() decrements in_flight.
581          */
582         bdrv_inc_in_flight(bs);
583         aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data);
584 
585         /* The request is running on the IOThread a. Draining its block device
586          * will make sure that it has completed as far as the BDS is concerned,
587          * but the drain in this thread can continue immediately after
588          * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
589          * later. */
590         do_drain_begin(drain_type, bs);
591         g_assert_cmpint(bs->in_flight, ==, 0);
592 
593         qemu_event_wait(&done_event);
594 
595         g_assert_cmpint(aio_ret, ==, 0);
596         do_drain_end(drain_type, bs);
597         break;
598     case 1:
599         co = qemu_coroutine_create(test_iothread_drain_co_entry, &data);
600         aio_co_enter(ctx_a, co);
601         AIO_WAIT_WHILE_UNLOCKED(NULL, !data.co_done);
602         break;
603     default:
604         g_assert_not_reached();
605     }
606 
607     blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort);
608 
609     bdrv_unref(bs);
610     blk_unref(blk);
611 
612 out:
613     iothread_join(a);
614     iothread_join(b);
615 }
616 
617 static void test_iothread_drain_all(void)
618 {
619     test_iothread_common(BDRV_DRAIN_ALL, 0);
620     test_iothread_common(BDRV_DRAIN_ALL, 1);
621 }
622 
623 static void test_iothread_drain(void)
624 {
625     test_iothread_common(BDRV_DRAIN, 0);
626     test_iothread_common(BDRV_DRAIN, 1);
627 }
628 
629 
630 typedef struct TestBlockJob {
631     BlockJob common;
632     BlockDriverState *bs;
633     int run_ret;
634     int prepare_ret;
635     bool running;
636     bool should_complete;
637 } TestBlockJob;
638 
639 static int test_job_prepare(Job *job)
640 {
641     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
642 
643     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
644     bdrv_flush(s->bs);
645     return s->prepare_ret;
646 }
647 
648 static void test_job_commit(Job *job)
649 {
650     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
651 
652     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
653     bdrv_flush(s->bs);
654 }
655 
656 static void test_job_abort(Job *job)
657 {
658     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
659 
660     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
661     bdrv_flush(s->bs);
662 }
663 
664 static int coroutine_fn test_job_run(Job *job, Error **errp)
665 {
666     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
667 
668     /* We are running the actual job code past the pause point in
669      * job_co_entry(). */
670     s->running = true;
671 
672     job_transition_to_ready(&s->common.job);
673     while (!s->should_complete) {
674         /* Avoid job_sleep_ns() because it marks the job as !busy. We want to
675          * emulate some actual activity (probably some I/O) here so that drain
676          * has to wait for this activity to stop. */
677         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000);
678 
679         job_pause_point(&s->common.job);
680     }
681 
682     return s->run_ret;
683 }
684 
685 static void test_job_complete(Job *job, Error **errp)
686 {
687     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
688     s->should_complete = true;
689 }
690 
691 BlockJobDriver test_job_driver = {
692     .job_driver = {
693         .instance_size  = sizeof(TestBlockJob),
694         .free           = block_job_free,
695         .user_resume    = block_job_user_resume,
696         .run            = test_job_run,
697         .complete       = test_job_complete,
698         .prepare        = test_job_prepare,
699         .commit         = test_job_commit,
700         .abort          = test_job_abort,
701     },
702 };
703 
704 enum test_job_result {
705     TEST_JOB_SUCCESS,
706     TEST_JOB_FAIL_RUN,
707     TEST_JOB_FAIL_PREPARE,
708 };
709 
710 enum test_job_drain_node {
711     TEST_JOB_DRAIN_SRC,
712     TEST_JOB_DRAIN_SRC_CHILD,
713 };
714 
715 static void test_blockjob_common_drain_node(enum drain_type drain_type,
716                                             bool use_iothread,
717                                             enum test_job_result result,
718                                             enum test_job_drain_node drain_node)
719 {
720     BlockBackend *blk_src, *blk_target;
721     BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs;
722     BlockJob *job;
723     TestBlockJob *tjob;
724     IOThread *iothread = NULL;
725     int ret;
726 
727     src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR,
728                                &error_abort);
729     src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing",
730                                        BDRV_O_RDWR, &error_abort);
731     src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay",
732                                        BDRV_O_RDWR, &error_abort);
733 
734     bdrv_set_backing_hd(src_overlay, src, &error_abort);
735     bdrv_unref(src);
736     bdrv_set_backing_hd(src, src_backing, &error_abort);
737     bdrv_unref(src_backing);
738 
739     blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
740     blk_insert_bs(blk_src, src_overlay, &error_abort);
741 
742     switch (drain_node) {
743     case TEST_JOB_DRAIN_SRC:
744         drain_bs = src;
745         break;
746     case TEST_JOB_DRAIN_SRC_CHILD:
747         drain_bs = src_backing;
748         break;
749     default:
750         g_assert_not_reached();
751     }
752 
753     if (use_iothread) {
754         AioContext *ctx;
755 
756         iothread = iothread_new();
757         ctx = iothread_get_aio_context(iothread);
758         blk_set_aio_context(blk_src, ctx, &error_abort);
759     }
760 
761     target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR,
762                                   &error_abort);
763     blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
764     blk_insert_bs(blk_target, target, &error_abort);
765     blk_set_allow_aio_context_change(blk_target, true);
766 
767     tjob = block_job_create("job0", &test_job_driver, NULL, src,
768                             0, BLK_PERM_ALL,
769                             0, 0, NULL, NULL, &error_abort);
770     tjob->bs = src;
771     job = &tjob->common;
772 
773     bdrv_graph_wrlock();
774     block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort);
775     bdrv_graph_wrunlock();
776 
777     switch (result) {
778     case TEST_JOB_SUCCESS:
779         break;
780     case TEST_JOB_FAIL_RUN:
781         tjob->run_ret = -EIO;
782         break;
783     case TEST_JOB_FAIL_PREPARE:
784         tjob->prepare_ret = -EIO;
785         break;
786     }
787 
788     job_start(&job->job);
789 
790     if (use_iothread) {
791         /* job_co_entry() is run in the I/O thread, wait for the actual job
792          * code to start (we don't want to catch the job in the pause point in
793          * job_co_entry(). */
794         while (!tjob->running) {
795             aio_poll(qemu_get_aio_context(), false);
796         }
797     }
798 
799     WITH_JOB_LOCK_GUARD() {
800         g_assert_cmpint(job->job.pause_count, ==, 0);
801         g_assert_false(job->job.paused);
802         g_assert_true(tjob->running);
803         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
804     }
805 
806     do_drain_begin_unlocked(drain_type, drain_bs);
807 
808     WITH_JOB_LOCK_GUARD() {
809         if (drain_type == BDRV_DRAIN_ALL) {
810             /* bdrv_drain_all() drains both src and target */
811             g_assert_cmpint(job->job.pause_count, ==, 2);
812         } else {
813             g_assert_cmpint(job->job.pause_count, ==, 1);
814         }
815         g_assert_true(job->job.paused);
816         g_assert_false(job->job.busy); /* The job is paused */
817     }
818 
819     do_drain_end_unlocked(drain_type, drain_bs);
820 
821     if (use_iothread) {
822         /*
823          * Here we are waiting for the paused status to change,
824          * so don't bother protecting the read every time.
825          *
826          * paused is reset in the I/O thread, wait for it
827          */
828         while (job->job.paused) {
829             aio_poll(qemu_get_aio_context(), false);
830         }
831     }
832 
833     WITH_JOB_LOCK_GUARD() {
834         g_assert_cmpint(job->job.pause_count, ==, 0);
835         g_assert_false(job->job.paused);
836         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
837     }
838 
839     do_drain_begin_unlocked(drain_type, target);
840 
841     WITH_JOB_LOCK_GUARD() {
842         if (drain_type == BDRV_DRAIN_ALL) {
843             /* bdrv_drain_all() drains both src and target */
844             g_assert_cmpint(job->job.pause_count, ==, 2);
845         } else {
846             g_assert_cmpint(job->job.pause_count, ==, 1);
847         }
848         g_assert_true(job->job.paused);
849         g_assert_false(job->job.busy); /* The job is paused */
850     }
851 
852     do_drain_end_unlocked(drain_type, target);
853 
854     if (use_iothread) {
855         /*
856          * Here we are waiting for the paused status to change,
857          * so don't bother protecting the read every time.
858          *
859          * paused is reset in the I/O thread, wait for it
860          */
861         while (job->job.paused) {
862             aio_poll(qemu_get_aio_context(), false);
863         }
864     }
865 
866     WITH_JOB_LOCK_GUARD() {
867         g_assert_cmpint(job->job.pause_count, ==, 0);
868         g_assert_false(job->job.paused);
869         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
870     }
871 
872     WITH_JOB_LOCK_GUARD() {
873         ret = job_complete_sync_locked(&job->job, &error_abort);
874     }
875     g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO));
876 
877     if (use_iothread) {
878         blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort);
879         assert(blk_get_aio_context(blk_target) == qemu_get_aio_context());
880     }
881 
882     blk_unref(blk_src);
883     blk_unref(blk_target);
884     bdrv_unref(src_overlay);
885     bdrv_unref(target);
886 
887     if (iothread) {
888         iothread_join(iothread);
889     }
890 }
891 
892 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread,
893                                  enum test_job_result result)
894 {
895     test_blockjob_common_drain_node(drain_type, use_iothread, result,
896                                     TEST_JOB_DRAIN_SRC);
897     test_blockjob_common_drain_node(drain_type, use_iothread, result,
898                                     TEST_JOB_DRAIN_SRC_CHILD);
899 }
900 
901 static void test_blockjob_drain_all(void)
902 {
903     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS);
904 }
905 
906 static void test_blockjob_drain(void)
907 {
908     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS);
909 }
910 
911 static void test_blockjob_error_drain_all(void)
912 {
913     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN);
914     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE);
915 }
916 
917 static void test_blockjob_error_drain(void)
918 {
919     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN);
920     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE);
921 }
922 
923 static void test_blockjob_iothread_drain_all(void)
924 {
925     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS);
926 }
927 
928 static void test_blockjob_iothread_drain(void)
929 {
930     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS);
931 }
932 
933 static void test_blockjob_iothread_error_drain_all(void)
934 {
935     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN);
936     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE);
937 }
938 
939 static void test_blockjob_iothread_error_drain(void)
940 {
941     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN);
942     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE);
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 
954     bdrv_graph_wrlock();
955     QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
956         bdrv_unref_child(bs, c);
957     }
958     bdrv_graph_wrunlock();
959 }
960 
961 static int coroutine_fn GRAPH_RDLOCK
962 bdrv_test_top_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
963                         QEMUIOVector *qiov, BdrvRequestFlags flags)
964 {
965     BDRVTestTopState *tts = bs->opaque;
966     return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags);
967 }
968 
969 static BlockDriver bdrv_test_top_driver = {
970     .format_name            = "test_top_driver",
971     .instance_size          = sizeof(BDRVTestTopState),
972 
973     .bdrv_close             = bdrv_test_top_close,
974     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
975 
976     .bdrv_child_perm        = bdrv_default_perms,
977 };
978 
979 typedef struct TestCoDeleteByDrainData {
980     BlockBackend *blk;
981     bool detach_instead_of_delete;
982     bool done;
983 } TestCoDeleteByDrainData;
984 
985 static void coroutine_fn test_co_delete_by_drain(void *opaque)
986 {
987     TestCoDeleteByDrainData *dbdd = opaque;
988     BlockBackend *blk = dbdd->blk;
989     BlockDriverState *bs = blk_bs(blk);
990     BDRVTestTopState *tts = bs->opaque;
991     void *buffer = g_malloc(65536);
992     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536);
993 
994     /* Pretend some internal write operation from parent to child.
995      * Important: We have to read from the child, not from the parent!
996      * Draining works by first propagating it all up the tree to the
997      * root and then waiting for drainage from root to the leaves
998      * (protocol nodes).  If we have a request waiting on the root,
999      * everything will be drained before we go back down the tree, but
1000      * we do not want that.  We want to be in the middle of draining
1001      * when this following requests returns. */
1002     bdrv_graph_co_rdlock();
1003     bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0);
1004     bdrv_graph_co_rdunlock();
1005 
1006     g_assert_cmpint(bs->refcnt, ==, 1);
1007 
1008     if (!dbdd->detach_instead_of_delete) {
1009         blk_co_unref(blk);
1010     } else {
1011         BdrvChild *c, *next_c;
1012         bdrv_graph_co_rdlock();
1013         QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1014             bdrv_graph_co_rdunlock();
1015             bdrv_co_unref_child(bs, c);
1016             bdrv_graph_co_rdlock();
1017         }
1018         bdrv_graph_co_rdunlock();
1019     }
1020 
1021     dbdd->done = true;
1022     g_free(buffer);
1023 }
1024 
1025 /**
1026  * Test what happens when some BDS has some children, you drain one of
1027  * them and this results in the BDS being deleted.
1028  *
1029  * If @detach_instead_of_delete is set, the BDS is not going to be
1030  * deleted but will only detach all of its children.
1031  */
1032 static void do_test_delete_by_drain(bool detach_instead_of_delete,
1033                                     enum drain_type drain_type)
1034 {
1035     BlockBackend *blk;
1036     BlockDriverState *bs, *child_bs, *null_bs;
1037     BDRVTestTopState *tts;
1038     TestCoDeleteByDrainData dbdd;
1039     Coroutine *co;
1040 
1041     bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR,
1042                               &error_abort);
1043     bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1044     tts = bs->opaque;
1045 
1046     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1047                         &error_abort);
1048     bdrv_graph_wrlock();
1049     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds,
1050                       BDRV_CHILD_DATA, &error_abort);
1051     bdrv_graph_wrunlock();
1052 
1053     /* This child will be the one to pass to requests through to, and
1054      * it will stall until a drain occurs */
1055     child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR,
1056                                     &error_abort);
1057     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1058     /* Takes our reference to child_bs */
1059     bdrv_graph_wrlock();
1060     tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child",
1061                                         &child_of_bds,
1062                                         BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY,
1063                                         &error_abort);
1064     bdrv_graph_wrunlock();
1065 
1066     /* This child is just there to be deleted
1067      * (for detach_instead_of_delete == true) */
1068     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1069                         &error_abort);
1070     bdrv_graph_wrlock();
1071     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA,
1072                       &error_abort);
1073     bdrv_graph_wrunlock();
1074 
1075     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1076     blk_insert_bs(blk, bs, &error_abort);
1077 
1078     /* Referenced by blk now */
1079     bdrv_unref(bs);
1080 
1081     g_assert_cmpint(bs->refcnt, ==, 1);
1082     g_assert_cmpint(child_bs->refcnt, ==, 1);
1083     g_assert_cmpint(null_bs->refcnt, ==, 1);
1084 
1085 
1086     dbdd = (TestCoDeleteByDrainData){
1087         .blk = blk,
1088         .detach_instead_of_delete = detach_instead_of_delete,
1089         .done = false,
1090     };
1091     co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd);
1092     qemu_coroutine_enter(co);
1093 
1094     /* Drain the child while the read operation is still pending.
1095      * This should result in the operation finishing and
1096      * test_co_delete_by_drain() resuming.  Thus, @bs will be deleted
1097      * and the coroutine will exit while this drain operation is still
1098      * in progress. */
1099     switch (drain_type) {
1100     case BDRV_DRAIN:
1101         bdrv_ref(child_bs);
1102         bdrv_drain(child_bs);
1103         bdrv_unref(child_bs);
1104         break;
1105     case BDRV_DRAIN_ALL:
1106         bdrv_drain_all_begin();
1107         bdrv_drain_all_end();
1108         break;
1109     default:
1110         g_assert_not_reached();
1111     }
1112 
1113     while (!dbdd.done) {
1114         aio_poll(qemu_get_aio_context(), true);
1115     }
1116 
1117     if (detach_instead_of_delete) {
1118         /* Here, the reference has not passed over to the coroutine,
1119          * so we have to delete the BB ourselves */
1120         blk_unref(blk);
1121     }
1122 }
1123 
1124 static void test_delete_by_drain(void)
1125 {
1126     do_test_delete_by_drain(false, BDRV_DRAIN);
1127 }
1128 
1129 static void test_detach_by_drain_all(void)
1130 {
1131     do_test_delete_by_drain(true, BDRV_DRAIN_ALL);
1132 }
1133 
1134 static void test_detach_by_drain(void)
1135 {
1136     do_test_delete_by_drain(true, BDRV_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     bool detach_on_drain;
1147 };
1148 static struct detach_by_parent_data detach_by_parent_data;
1149 
1150 static void no_coroutine_fn detach_indirect_bh(void *opaque)
1151 {
1152     struct detach_by_parent_data *data = opaque;
1153 
1154     bdrv_dec_in_flight(data->child_b->bs);
1155 
1156     bdrv_graph_wrlock();
1157     bdrv_unref_child(data->parent_b, data->child_b);
1158 
1159     bdrv_ref(data->c);
1160     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
1161                                       &child_of_bds, BDRV_CHILD_DATA,
1162                                       &error_abort);
1163     bdrv_graph_wrunlock();
1164 }
1165 
1166 static void coroutine_mixed_fn detach_by_parent_aio_cb(void *opaque, int ret)
1167 {
1168     struct detach_by_parent_data *data = &detach_by_parent_data;
1169 
1170     g_assert_cmpint(ret, ==, 0);
1171     if (data->by_parent_cb) {
1172         bdrv_inc_in_flight(data->child_b->bs);
1173         aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1174                                 detach_indirect_bh, &detach_by_parent_data);
1175     }
1176 }
1177 
1178 static void GRAPH_RDLOCK detach_by_driver_cb_drained_begin(BdrvChild *child)
1179 {
1180     struct detach_by_parent_data *data = &detach_by_parent_data;
1181 
1182     if (!data->detach_on_drain) {
1183         return;
1184     }
1185     data->detach_on_drain = false;
1186 
1187     bdrv_inc_in_flight(data->child_b->bs);
1188     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1189                             detach_indirect_bh, &detach_by_parent_data);
1190     child_of_bds.drained_begin(child);
1191 }
1192 
1193 static BdrvChildClass detach_by_driver_cb_class;
1194 
1195 /*
1196  * Initial graph:
1197  *
1198  * PA     PB
1199  *    \ /   \
1200  *     A     B     C
1201  *
1202  * by_parent_cb == true:  Test that parent callbacks don't poll
1203  *
1204  *     PA has a pending write request whose callback changes the child nodes of
1205  *     PB: It removes B and adds C instead. The subtree of PB is drained, which
1206  *     will indirectly drain the write request, too.
1207  *
1208  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1209  *
1210  *     PA's BdrvChildClass has a .drained_begin callback that schedules a BH
1211  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
1212  *     state is messed up, but if it is only polled in the single
1213  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1214  */
1215 static void TSA_NO_TSA test_detach_indirect(bool by_parent_cb)
1216 {
1217     BlockBackend *blk;
1218     BlockDriverState *parent_a, *parent_b, *a, *b, *c;
1219     BdrvChild *child_a, *child_b;
1220     BlockAIOCB *acb;
1221 
1222     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
1223 
1224     if (!by_parent_cb) {
1225         detach_by_driver_cb_class = child_of_bds;
1226         detach_by_driver_cb_class.drained_begin =
1227             detach_by_driver_cb_drained_begin;
1228         detach_by_driver_cb_class.drained_end = NULL;
1229         detach_by_driver_cb_class.drained_poll = NULL;
1230     }
1231 
1232     detach_by_parent_data = (struct detach_by_parent_data) {
1233         .detach_on_drain = false,
1234     };
1235 
1236     /* Create all involved nodes */
1237     parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR,
1238                                     &error_abort);
1239     parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0,
1240                                     &error_abort);
1241 
1242     a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort);
1243     b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort);
1244     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
1245 
1246     /* blk is a BB for parent-a */
1247     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1248     blk_insert_bs(blk, parent_a, &error_abort);
1249     bdrv_unref(parent_a);
1250 
1251     /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1252      * callback must not return immediately. */
1253     if (!by_parent_cb) {
1254         BDRVTestState *s = parent_a->opaque;
1255         s->sleep_in_drain_begin = true;
1256     }
1257 
1258     /* Set child relationships */
1259     bdrv_ref(b);
1260     bdrv_ref(a);
1261     bdrv_graph_wrlock();
1262     child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds,
1263                                 BDRV_CHILD_DATA, &error_abort);
1264     child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds,
1265                                 BDRV_CHILD_COW, &error_abort);
1266 
1267     bdrv_ref(a);
1268     bdrv_attach_child(parent_a, a, "PA-A",
1269                       by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class,
1270                       BDRV_CHILD_DATA, &error_abort);
1271     bdrv_graph_wrunlock();
1272 
1273     g_assert_cmpint(parent_a->refcnt, ==, 1);
1274     g_assert_cmpint(parent_b->refcnt, ==, 1);
1275     g_assert_cmpint(a->refcnt, ==, 3);
1276     g_assert_cmpint(b->refcnt, ==, 2);
1277     g_assert_cmpint(c->refcnt, ==, 1);
1278 
1279     g_assert(QLIST_FIRST(&parent_b->children) == child_a);
1280     g_assert(QLIST_NEXT(child_a, next) == child_b);
1281     g_assert(QLIST_NEXT(child_b, next) == NULL);
1282 
1283     /* Start the evil write request */
1284     detach_by_parent_data = (struct detach_by_parent_data) {
1285         .parent_b = parent_b,
1286         .child_b = child_b,
1287         .c = c,
1288         .by_parent_cb = by_parent_cb,
1289         .detach_on_drain = true,
1290     };
1291     acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL);
1292     g_assert(acb != NULL);
1293 
1294     /* Drain and check the expected result */
1295     bdrv_drained_begin(parent_b);
1296     bdrv_drained_begin(a);
1297     bdrv_drained_begin(b);
1298     bdrv_drained_begin(c);
1299 
1300     g_assert(detach_by_parent_data.child_c != NULL);
1301 
1302     g_assert_cmpint(parent_a->refcnt, ==, 1);
1303     g_assert_cmpint(parent_b->refcnt, ==, 1);
1304     g_assert_cmpint(a->refcnt, ==, 3);
1305     g_assert_cmpint(b->refcnt, ==, 1);
1306     g_assert_cmpint(c->refcnt, ==, 2);
1307 
1308     g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c);
1309     g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a);
1310     g_assert(QLIST_NEXT(child_a, next) == NULL);
1311 
1312     g_assert_cmpint(parent_a->quiesce_counter, ==, 1);
1313     g_assert_cmpint(parent_b->quiesce_counter, ==, 3);
1314     g_assert_cmpint(a->quiesce_counter, ==, 1);
1315     g_assert_cmpint(b->quiesce_counter, ==, 1);
1316     g_assert_cmpint(c->quiesce_counter, ==, 1);
1317 
1318     bdrv_drained_end(parent_b);
1319     bdrv_drained_end(a);
1320     bdrv_drained_end(b);
1321     bdrv_drained_end(c);
1322 
1323     bdrv_unref(parent_b);
1324     blk_unref(blk);
1325 
1326     g_assert_cmpint(a->refcnt, ==, 1);
1327     g_assert_cmpint(b->refcnt, ==, 1);
1328     g_assert_cmpint(c->refcnt, ==, 1);
1329     bdrv_unref(a);
1330     bdrv_unref(b);
1331     bdrv_unref(c);
1332 }
1333 
1334 static void test_detach_by_parent_cb(void)
1335 {
1336     test_detach_indirect(true);
1337 }
1338 
1339 static void test_detach_by_driver_cb(void)
1340 {
1341     test_detach_indirect(false);
1342 }
1343 
1344 static void test_append_to_drained(void)
1345 {
1346     BlockBackend *blk;
1347     BlockDriverState *base, *overlay;
1348     BDRVTestState *base_s, *overlay_s;
1349 
1350     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1351     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
1352     base_s = base->opaque;
1353     blk_insert_bs(blk, base, &error_abort);
1354 
1355     overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR,
1356                                    &error_abort);
1357     overlay_s = overlay->opaque;
1358 
1359     do_drain_begin(BDRV_DRAIN, base);
1360     g_assert_cmpint(base->quiesce_counter, ==, 1);
1361     g_assert_cmpint(base_s->drain_count, ==, 1);
1362     g_assert_cmpint(base->in_flight, ==, 0);
1363 
1364     bdrv_append(overlay, base, &error_abort);
1365 
1366     g_assert_cmpint(base->in_flight, ==, 0);
1367     g_assert_cmpint(overlay->in_flight, ==, 0);
1368 
1369     g_assert_cmpint(base->quiesce_counter, ==, 1);
1370     g_assert_cmpint(base_s->drain_count, ==, 1);
1371     g_assert_cmpint(overlay->quiesce_counter, ==, 1);
1372     g_assert_cmpint(overlay_s->drain_count, ==, 1);
1373 
1374     do_drain_end(BDRV_DRAIN, base);
1375 
1376     g_assert_cmpint(base->quiesce_counter, ==, 0);
1377     g_assert_cmpint(base_s->drain_count, ==, 0);
1378     g_assert_cmpint(overlay->quiesce_counter, ==, 0);
1379     g_assert_cmpint(overlay_s->drain_count, ==, 0);
1380 
1381     bdrv_unref(overlay);
1382     bdrv_unref(base);
1383     blk_unref(blk);
1384 }
1385 
1386 static void test_set_aio_context(void)
1387 {
1388     BlockDriverState *bs;
1389     IOThread *a = iothread_new();
1390     IOThread *b = iothread_new();
1391     AioContext *ctx_a = iothread_get_aio_context(a);
1392     AioContext *ctx_b = iothread_get_aio_context(b);
1393 
1394     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
1395                               &error_abort);
1396 
1397     bdrv_drained_begin(bs);
1398     bdrv_try_change_aio_context(bs, ctx_a, NULL, &error_abort);
1399     bdrv_drained_end(bs);
1400 
1401     bdrv_drained_begin(bs);
1402     bdrv_try_change_aio_context(bs, ctx_b, NULL, &error_abort);
1403     bdrv_try_change_aio_context(bs, qemu_get_aio_context(), NULL, &error_abort);
1404     bdrv_drained_end(bs);
1405 
1406     bdrv_unref(bs);
1407     iothread_join(a);
1408     iothread_join(b);
1409 }
1410 
1411 
1412 typedef struct TestDropBackingBlockJob {
1413     BlockJob common;
1414     bool should_complete;
1415     bool *did_complete;
1416     BlockDriverState *detach_also;
1417     BlockDriverState *bs;
1418 } TestDropBackingBlockJob;
1419 
1420 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp)
1421 {
1422     TestDropBackingBlockJob *s =
1423         container_of(job, TestDropBackingBlockJob, common.job);
1424 
1425     while (!s->should_complete) {
1426         job_sleep_ns(job, 0);
1427     }
1428 
1429     return 0;
1430 }
1431 
1432 static void test_drop_backing_job_commit(Job *job)
1433 {
1434     TestDropBackingBlockJob *s =
1435         container_of(job, TestDropBackingBlockJob, common.job);
1436 
1437     bdrv_set_backing_hd(s->bs, NULL, &error_abort);
1438     bdrv_set_backing_hd(s->detach_also, NULL, &error_abort);
1439 
1440     *s->did_complete = true;
1441 }
1442 
1443 static const BlockJobDriver test_drop_backing_job_driver = {
1444     .job_driver = {
1445         .instance_size  = sizeof(TestDropBackingBlockJob),
1446         .free           = block_job_free,
1447         .user_resume    = block_job_user_resume,
1448         .run            = test_drop_backing_job_run,
1449         .commit         = test_drop_backing_job_commit,
1450     }
1451 };
1452 
1453 /**
1454  * Creates a child node with three parent nodes on it, and then runs a
1455  * block job on the final one, parent-node-2.
1456  *
1457  * The job is then asked to complete before a section where the child
1458  * is drained.
1459  *
1460  * Ending this section will undrain the child's parents, first
1461  * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
1462  * list is in reverse order of how they were added.  Ending the drain
1463  * on parent-node-2 will resume the job, thus completing it and
1464  * scheduling job_exit().
1465  *
1466  * Ending the drain on parent-node-1 will poll the AioContext, which
1467  * lets job_exit() and thus test_drop_backing_job_commit() run.  That
1468  * function first removes the child as parent-node-2's backing file.
1469  *
1470  * In old (and buggy) implementations, there are two problems with
1471  * that:
1472  * (A) bdrv_drain_invoke() polls for every node that leaves the
1473  *     drained section.  This means that job_exit() is scheduled
1474  *     before the child has left the drained section.  Its
1475  *     quiesce_counter is therefore still 1 when it is removed from
1476  *     parent-node-2.
1477  *
1478  * (B) bdrv_replace_child_noperm() calls drained_end() on the old
1479  *     child's parents as many times as the child is quiesced.  This
1480  *     means it will call drained_end() on parent-node-2 once.
1481  *     Because parent-node-2 is no longer quiesced at this point, this
1482  *     will fail.
1483  *
1484  * bdrv_replace_child_noperm() therefore must call drained_end() on
1485  * the parent only if it really is still drained because the child is
1486  * drained.
1487  *
1488  * If removing child from parent-node-2 was successful (as it should
1489  * be), test_drop_backing_job_commit() will then also remove the child
1490  * from parent-node-0.
1491  *
1492  * With an old version of our drain infrastructure ((A) above), that
1493  * resulted in the following flow:
1494  *
1495  * 1. child attempts to leave its drained section.  The call recurses
1496  *    to its parents.
1497  *
1498  * 2. parent-node-2 leaves the drained section.  Polling in
1499  *    bdrv_drain_invoke() will schedule job_exit().
1500  *
1501  * 3. parent-node-1 leaves the drained section.  Polling in
1502  *    bdrv_drain_invoke() will run job_exit(), thus disconnecting
1503  *    parent-node-0 from the child node.
1504  *
1505  * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
1506  *    iterate over the parents.  Thus, it now accesses the BdrvChild
1507  *    object that used to connect parent-node-0 and the child node.
1508  *    However, that object no longer exists, so it accesses a dangling
1509  *    pointer.
1510  *
1511  * The solution is to only poll once when running a bdrv_drained_end()
1512  * operation, specifically at the end when all drained_end()
1513  * operations for all involved nodes have been scheduled.
1514  * Note that this also solves (A) above, thus hiding (B).
1515  */
1516 static void test_blockjob_commit_by_drained_end(void)
1517 {
1518     BlockDriverState *bs_child, *bs_parents[3];
1519     TestDropBackingBlockJob *job;
1520     bool job_has_completed = false;
1521     int i;
1522 
1523     bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR,
1524                                     &error_abort);
1525 
1526     for (i = 0; i < 3; i++) {
1527         char name[32];
1528         snprintf(name, sizeof(name), "parent-node-%i", i);
1529         bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR,
1530                                              &error_abort);
1531         bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort);
1532     }
1533 
1534     job = block_job_create("job", &test_drop_backing_job_driver, NULL,
1535                            bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL,
1536                            &error_abort);
1537     job->bs = bs_parents[2];
1538 
1539     job->detach_also = bs_parents[0];
1540     job->did_complete = &job_has_completed;
1541 
1542     job_start(&job->common.job);
1543 
1544     job->should_complete = true;
1545     bdrv_drained_begin(bs_child);
1546     g_assert(!job_has_completed);
1547     bdrv_drained_end(bs_child);
1548     aio_poll(qemu_get_aio_context(), false);
1549     g_assert(job_has_completed);
1550 
1551     bdrv_unref(bs_parents[0]);
1552     bdrv_unref(bs_parents[1]);
1553     bdrv_unref(bs_parents[2]);
1554     bdrv_unref(bs_child);
1555 }
1556 
1557 
1558 typedef struct TestSimpleBlockJob {
1559     BlockJob common;
1560     bool should_complete;
1561     bool *did_complete;
1562 } TestSimpleBlockJob;
1563 
1564 static int coroutine_fn test_simple_job_run(Job *job, Error **errp)
1565 {
1566     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1567 
1568     while (!s->should_complete) {
1569         job_sleep_ns(job, 0);
1570     }
1571 
1572     return 0;
1573 }
1574 
1575 static void test_simple_job_clean(Job *job)
1576 {
1577     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1578     *s->did_complete = true;
1579 }
1580 
1581 static const BlockJobDriver test_simple_job_driver = {
1582     .job_driver = {
1583         .instance_size  = sizeof(TestSimpleBlockJob),
1584         .free           = block_job_free,
1585         .user_resume    = block_job_user_resume,
1586         .run            = test_simple_job_run,
1587         .clean          = test_simple_job_clean,
1588     },
1589 };
1590 
1591 static int drop_intermediate_poll_update_filename(BdrvChild *child,
1592                                                   BlockDriverState *new_base,
1593                                                   const char *filename,
1594                                                   bool backing_mask_protocol,
1595                                                   Error **errp)
1596 {
1597     /*
1598      * We are free to poll here, which may change the block graph, if
1599      * it is not drained.
1600      */
1601 
1602     /* If the job is not drained: Complete it, schedule job_exit() */
1603     aio_poll(qemu_get_current_aio_context(), false);
1604     /* If the job is not drained: Run job_exit(), finish the job */
1605     aio_poll(qemu_get_current_aio_context(), false);
1606 
1607     return 0;
1608 }
1609 
1610 /**
1611  * Test a poll in the midst of bdrv_drop_intermediate().
1612  *
1613  * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(),
1614  * which can yield or poll.  This may lead to graph changes, unless
1615  * the whole subtree in question is drained.
1616  *
1617  * We test this on the following graph:
1618  *
1619  *                    Job
1620  *
1621  *                     |
1622  *                  job-node
1623  *                     |
1624  *                     v
1625  *
1626  *                  job-node
1627  *
1628  *                     |
1629  *                  backing
1630  *                     |
1631  *                     v
1632  *
1633  * node-2 --chain--> node-1 --chain--> node-0
1634  *
1635  * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0).
1636  *
1637  * This first updates node-2's backing filename by invoking
1638  * drop_intermediate_poll_update_filename(), which polls twice.  This
1639  * causes the job to finish, which in turns causes the job-node to be
1640  * deleted.
1641  *
1642  * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it
1643  * already has a pointer to the BdrvChild edge between job-node and
1644  * node-1.  When it tries to handle that edge, we probably get a
1645  * segmentation fault because the object no longer exists.
1646  *
1647  *
1648  * The solution is for bdrv_drop_intermediate() to drain top's
1649  * subtree.  This prevents graph changes from happening just because
1650  * BdrvChildClass.update_filename() yields or polls.  Thus, the block
1651  * job is paused during that drained section and must finish before or
1652  * after.
1653  *
1654  * (In addition, bdrv_replace_child() must keep the job paused.)
1655  */
1656 static void test_drop_intermediate_poll(void)
1657 {
1658     static BdrvChildClass chain_child_class;
1659     BlockDriverState *chain[3];
1660     TestSimpleBlockJob *job;
1661     BlockDriverState *job_node;
1662     bool job_has_completed = false;
1663     int i;
1664     int ret;
1665 
1666     chain_child_class = child_of_bds;
1667     chain_child_class.update_filename = drop_intermediate_poll_update_filename;
1668 
1669     for (i = 0; i < 3; i++) {
1670         char name[32];
1671         snprintf(name, 32, "node-%i", i);
1672 
1673         chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort);
1674     }
1675 
1676     job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR,
1677                                     &error_abort);
1678     bdrv_set_backing_hd(job_node, chain[1], &error_abort);
1679 
1680     /*
1681      * Establish the chain last, so the chain links are the first
1682      * elements in the BDS.parents lists
1683      */
1684     bdrv_graph_wrlock();
1685     for (i = 0; i < 3; i++) {
1686         if (i) {
1687             /* Takes the reference to chain[i - 1] */
1688             bdrv_attach_child(chain[i], chain[i - 1], "chain",
1689                               &chain_child_class, BDRV_CHILD_COW, &error_abort);
1690         }
1691     }
1692     bdrv_graph_wrunlock();
1693 
1694     job = block_job_create("job", &test_simple_job_driver, NULL, job_node,
1695                            0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort);
1696 
1697     /* The job has a reference now */
1698     bdrv_unref(job_node);
1699 
1700     job->did_complete = &job_has_completed;
1701 
1702     job_start(&job->common.job);
1703     job->should_complete = true;
1704 
1705     g_assert(!job_has_completed);
1706     ret = bdrv_drop_intermediate(chain[1], chain[0], NULL, false);
1707     aio_poll(qemu_get_aio_context(), false);
1708     g_assert(ret == 0);
1709     g_assert(job_has_completed);
1710 
1711     bdrv_unref(chain[2]);
1712 }
1713 
1714 
1715 typedef struct BDRVReplaceTestState {
1716     bool setup_completed;
1717     bool was_drained;
1718     bool was_undrained;
1719     bool has_read;
1720 
1721     int drain_count;
1722 
1723     bool yield_before_read;
1724     Coroutine *io_co;
1725     Coroutine *drain_co;
1726 } BDRVReplaceTestState;
1727 
1728 static void bdrv_replace_test_close(BlockDriverState *bs)
1729 {
1730 }
1731 
1732 /**
1733  * If @bs has a backing file:
1734  *   Yield if .yield_before_read is true (and wait for drain_begin to
1735  *   wake us up).
1736  *   Forward the read to bs->backing.  Set .has_read to true.
1737  *   If drain_begin has woken us, wake it in turn.
1738  *
1739  * Otherwise:
1740  *   Set .has_read to true and return success.
1741  */
1742 static int coroutine_fn GRAPH_RDLOCK
1743 bdrv_replace_test_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1744                             QEMUIOVector *qiov, BdrvRequestFlags flags)
1745 {
1746     BDRVReplaceTestState *s = bs->opaque;
1747 
1748     if (bs->backing) {
1749         int ret;
1750 
1751         g_assert(!s->drain_count);
1752 
1753         s->io_co = qemu_coroutine_self();
1754         if (s->yield_before_read) {
1755             s->yield_before_read = false;
1756             qemu_coroutine_yield();
1757         }
1758         s->io_co = NULL;
1759 
1760         ret = bdrv_co_preadv(bs->backing, offset, bytes, qiov, 0);
1761         s->has_read = true;
1762 
1763         /* Wake up drain_co if it runs */
1764         if (s->drain_co) {
1765             aio_co_wake(s->drain_co);
1766         }
1767 
1768         return ret;
1769     }
1770 
1771     s->has_read = true;
1772     return 0;
1773 }
1774 
1775 static void coroutine_fn bdrv_replace_test_drain_co(void *opaque)
1776 {
1777     BlockDriverState *bs = opaque;
1778     BDRVReplaceTestState *s = bs->opaque;
1779 
1780     /* Keep waking io_co up until it is done */
1781     while (s->io_co) {
1782         aio_co_wake(s->io_co);
1783         s->io_co = NULL;
1784         qemu_coroutine_yield();
1785     }
1786     s->drain_co = NULL;
1787     bdrv_dec_in_flight(bs);
1788 }
1789 
1790 /**
1791  * If .drain_count is 0, wake up .io_co if there is one; and set
1792  * .was_drained.
1793  * Increment .drain_count.
1794  */
1795 static void bdrv_replace_test_drain_begin(BlockDriverState *bs)
1796 {
1797     BDRVReplaceTestState *s = bs->opaque;
1798 
1799     if (!s->setup_completed) {
1800         return;
1801     }
1802 
1803     if (!s->drain_count) {
1804         s->drain_co = qemu_coroutine_create(bdrv_replace_test_drain_co, bs);
1805         bdrv_inc_in_flight(bs);
1806         aio_co_enter(bdrv_get_aio_context(bs), s->drain_co);
1807         s->was_drained = true;
1808     }
1809     s->drain_count++;
1810 }
1811 
1812 static void coroutine_fn bdrv_replace_test_read_entry(void *opaque)
1813 {
1814     BlockDriverState *bs = opaque;
1815     char data;
1816     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
1817     int ret;
1818 
1819     /* Queue a read request post-drain */
1820     bdrv_graph_co_rdlock();
1821     ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
1822     bdrv_graph_co_rdunlock();
1823 
1824     g_assert(ret >= 0);
1825     bdrv_dec_in_flight(bs);
1826 }
1827 
1828 /**
1829  * Reduce .drain_count, set .was_undrained once it reaches 0.
1830  * If .drain_count reaches 0 and the node has a backing file, issue a
1831  * read request.
1832  */
1833 static void bdrv_replace_test_drain_end(BlockDriverState *bs)
1834 {
1835     BDRVReplaceTestState *s = bs->opaque;
1836 
1837     GRAPH_RDLOCK_GUARD_MAINLOOP();
1838 
1839     if (!s->setup_completed) {
1840         return;
1841     }
1842 
1843     g_assert(s->drain_count > 0);
1844     if (!--s->drain_count) {
1845         s->was_undrained = true;
1846 
1847         if (bs->backing) {
1848             Coroutine *co = qemu_coroutine_create(bdrv_replace_test_read_entry,
1849                                                   bs);
1850             bdrv_inc_in_flight(bs);
1851             aio_co_enter(bdrv_get_aio_context(bs), co);
1852         }
1853     }
1854 }
1855 
1856 static BlockDriver bdrv_replace_test = {
1857     .format_name            = "replace_test",
1858     .instance_size          = sizeof(BDRVReplaceTestState),
1859     .supports_backing       = true,
1860 
1861     .bdrv_close             = bdrv_replace_test_close,
1862     .bdrv_co_preadv         = bdrv_replace_test_co_preadv,
1863 
1864     .bdrv_drain_begin       = bdrv_replace_test_drain_begin,
1865     .bdrv_drain_end         = bdrv_replace_test_drain_end,
1866 
1867     .bdrv_child_perm        = bdrv_default_perms,
1868 };
1869 
1870 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque)
1871 {
1872     int ret;
1873     char data;
1874 
1875     ret = blk_co_pread(opaque, 0, 1, &data, 0);
1876     g_assert(ret >= 0);
1877 }
1878 
1879 /**
1880  * We test two things:
1881  * (1) bdrv_replace_child_noperm() must not undrain the parent if both
1882  *     children are drained.
1883  * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
1884  *     drained child.  If the old child is drained, it must flush I/O
1885  *     requests after the new one has been attached.  If the new child
1886  *     is drained, it must flush I/O requests before the old one is
1887  *     detached.
1888  *
1889  * To do so, we create one parent node and two child nodes; then
1890  * attach one of the children (old_child_bs) to the parent, then
1891  * drain both old_child_bs and new_child_bs according to
1892  * old_drain_count and new_drain_count, respectively, and finally
1893  * we invoke bdrv_replace_node() to replace old_child_bs by
1894  * new_child_bs.
1895  *
1896  * The test block driver we use here (bdrv_replace_test) has a read
1897  * function that:
1898  * - For the parent node, can optionally yield, and then forwards the
1899  *   read to bdrv_preadv(),
1900  * - For the child node, just returns immediately.
1901  *
1902  * If the read yields, the drain_begin function will wake it up.
1903  *
1904  * The drain_end function issues a read on the parent once it is fully
1905  * undrained (which simulates requests starting to come in again).
1906  */
1907 static void do_test_replace_child_mid_drain(int old_drain_count,
1908                                             int new_drain_count)
1909 {
1910     BlockBackend *parent_blk;
1911     BlockDriverState *parent_bs;
1912     BlockDriverState *old_child_bs, *new_child_bs;
1913     BDRVReplaceTestState *parent_s;
1914     BDRVReplaceTestState *old_child_s, *new_child_s;
1915     Coroutine *io_co;
1916     int i;
1917 
1918     parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0,
1919                                      &error_abort);
1920     parent_s = parent_bs->opaque;
1921 
1922     parent_blk = blk_new(qemu_get_aio_context(),
1923                          BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
1924     blk_insert_bs(parent_blk, parent_bs, &error_abort);
1925 
1926     old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0,
1927                                         &error_abort);
1928     new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0,
1929                                         &error_abort);
1930     old_child_s = old_child_bs->opaque;
1931     new_child_s = new_child_bs->opaque;
1932 
1933     /* So that we can read something */
1934     parent_bs->total_sectors = 1;
1935     old_child_bs->total_sectors = 1;
1936     new_child_bs->total_sectors = 1;
1937 
1938     bdrv_ref(old_child_bs);
1939     bdrv_graph_wrlock();
1940     bdrv_attach_child(parent_bs, old_child_bs, "child", &child_of_bds,
1941                       BDRV_CHILD_COW, &error_abort);
1942     bdrv_graph_wrunlock();
1943     parent_s->setup_completed = true;
1944 
1945     for (i = 0; i < old_drain_count; i++) {
1946         bdrv_drained_begin(old_child_bs);
1947     }
1948     for (i = 0; i < new_drain_count; i++) {
1949         bdrv_drained_begin(new_child_bs);
1950     }
1951 
1952     if (!old_drain_count) {
1953         /*
1954          * Start a read operation that will yield, so it will not
1955          * complete before the node is drained.
1956          */
1957         parent_s->yield_before_read = true;
1958         io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co,
1959                                       parent_blk);
1960         qemu_coroutine_enter(io_co);
1961     }
1962 
1963     /* If we have started a read operation, it should have yielded */
1964     g_assert(!parent_s->has_read);
1965 
1966     /* Reset drained status so we can see what bdrv_replace_node() does */
1967     parent_s->was_drained = false;
1968     parent_s->was_undrained = false;
1969 
1970     g_assert(parent_bs->quiesce_counter == old_drain_count);
1971     bdrv_drained_begin(old_child_bs);
1972     bdrv_drained_begin(new_child_bs);
1973     bdrv_graph_wrlock();
1974     bdrv_replace_node(old_child_bs, new_child_bs, &error_abort);
1975     bdrv_graph_wrunlock();
1976     bdrv_drained_end(new_child_bs);
1977     bdrv_drained_end(old_child_bs);
1978     g_assert(parent_bs->quiesce_counter == new_drain_count);
1979 
1980     if (!old_drain_count && !new_drain_count) {
1981         /*
1982          * From undrained to undrained drains and undrains the parent,
1983          * because bdrv_replace_node() contains a drained section for
1984          * @old_child_bs.
1985          */
1986         g_assert(parent_s->was_drained && parent_s->was_undrained);
1987     } else if (!old_drain_count && new_drain_count) {
1988         /*
1989          * From undrained to drained should drain the parent and keep
1990          * it that way.
1991          */
1992         g_assert(parent_s->was_drained && !parent_s->was_undrained);
1993     } else if (old_drain_count && !new_drain_count) {
1994         /*
1995          * From drained to undrained should undrain the parent and
1996          * keep it that way.
1997          */
1998         g_assert(!parent_s->was_drained && parent_s->was_undrained);
1999     } else /* if (old_drain_count && new_drain_count) */ {
2000         /*
2001          * From drained to drained must not undrain the parent at any
2002          * point
2003          */
2004         g_assert(!parent_s->was_drained && !parent_s->was_undrained);
2005     }
2006 
2007     if (!old_drain_count || !new_drain_count) {
2008         /*
2009          * If !old_drain_count, we have started a read request before
2010          * bdrv_replace_node().  If !new_drain_count, the parent must
2011          * have been undrained at some point, and
2012          * bdrv_replace_test_co_drain_end() starts a read request
2013          * then.
2014          */
2015         g_assert(parent_s->has_read);
2016     } else {
2017         /*
2018          * If the parent was never undrained, there is no way to start
2019          * a read request.
2020          */
2021         g_assert(!parent_s->has_read);
2022     }
2023 
2024     /* A drained child must have not received any request */
2025     g_assert(!(old_drain_count && old_child_s->has_read));
2026     g_assert(!(new_drain_count && new_child_s->has_read));
2027 
2028     for (i = 0; i < new_drain_count; i++) {
2029         bdrv_drained_end(new_child_bs);
2030     }
2031     for (i = 0; i < old_drain_count; i++) {
2032         bdrv_drained_end(old_child_bs);
2033     }
2034 
2035     /*
2036      * By now, bdrv_replace_test_co_drain_end() must have been called
2037      * at some point while the new child was attached to the parent.
2038      */
2039     g_assert(parent_s->has_read);
2040     g_assert(new_child_s->has_read);
2041 
2042     blk_unref(parent_blk);
2043     bdrv_unref(parent_bs);
2044     bdrv_unref(old_child_bs);
2045     bdrv_unref(new_child_bs);
2046 }
2047 
2048 static void test_replace_child_mid_drain(void)
2049 {
2050     int old_drain_count, new_drain_count;
2051 
2052     for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) {
2053         for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) {
2054             do_test_replace_child_mid_drain(old_drain_count, new_drain_count);
2055         }
2056     }
2057 }
2058 
2059 int main(int argc, char **argv)
2060 {
2061     int ret;
2062 
2063     bdrv_init();
2064     qemu_init_main_loop(&error_abort);
2065 
2066     g_test_init(&argc, &argv, NULL);
2067     qemu_event_init(&done_event, false);
2068 
2069     g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all);
2070     g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain);
2071 
2072     g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
2073                     test_drv_cb_co_drain_all);
2074     g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain);
2075 
2076     g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all);
2077     g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain);
2078 
2079     g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
2080                     test_quiesce_co_drain_all);
2081     g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain);
2082 
2083     g_test_add_func("/bdrv-drain/nested", test_nested);
2084 
2085     g_test_add_func("/bdrv-drain/graph-change/drain_all",
2086                     test_graph_change_drain_all);
2087 
2088     g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all);
2089     g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain);
2090 
2091     g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all);
2092     g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain);
2093 
2094     g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
2095                     test_blockjob_error_drain_all);
2096     g_test_add_func("/bdrv-drain/blockjob/error/drain",
2097                     test_blockjob_error_drain);
2098 
2099     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
2100                     test_blockjob_iothread_drain_all);
2101     g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
2102                     test_blockjob_iothread_drain);
2103 
2104     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
2105                     test_blockjob_iothread_error_drain_all);
2106     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
2107                     test_blockjob_iothread_error_drain);
2108 
2109     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
2110     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
2111     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
2112     g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb);
2113     g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb);
2114 
2115     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
2116 
2117     g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context);
2118 
2119     g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
2120                     test_blockjob_commit_by_drained_end);
2121 
2122     g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
2123                     test_drop_intermediate_poll);
2124 
2125     g_test_add_func("/bdrv-drain/replace_child/mid-drain",
2126                     test_replace_child_mid_drain);
2127 
2128     ret = g_test_run();
2129     qemu_event_destroy(&done_event);
2130     return ret;
2131 }
2132