xref: /openbmc/qemu/tests/unit/test-bdrv-drain.c (revision a1fadbcf)
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_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_change_backing_file = bdrv_test_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     if (drain_type != BDRV_DRAIN_ALL) {
183         aio_context_acquire(bdrv_get_aio_context(bs));
184     }
185     do_drain_begin(drain_type, bs);
186     if (drain_type != BDRV_DRAIN_ALL) {
187         aio_context_release(bdrv_get_aio_context(bs));
188     }
189 }
190 
191 static BlockBackend * no_coroutine_fn test_setup(void)
192 {
193     BlockBackend *blk;
194     BlockDriverState *bs, *backing;
195 
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,
198                               &error_abort);
199     blk_insert_bs(blk, bs, &error_abort);
200 
201     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
202     bdrv_set_backing_hd(bs, backing, &error_abort);
203 
204     bdrv_unref(backing);
205     bdrv_unref(bs);
206 
207     return blk;
208 }
209 
210 static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs)
211 {
212     if (drain_type != BDRV_DRAIN_ALL) {
213         aio_context_acquire(bdrv_get_aio_context(bs));
214     }
215     do_drain_end(drain_type, bs);
216     if (drain_type != BDRV_DRAIN_ALL) {
217         aio_context_release(bdrv_get_aio_context(bs));
218     }
219 }
220 
221 static void test_drv_cb_common(BlockBackend *blk, enum drain_type drain_type,
222                                bool recursive)
223 {
224     BlockDriverState *bs = blk_bs(blk);
225     BlockDriverState *backing = bs->backing->bs;
226     BDRVTestState *s, *backing_s;
227     BlockAIOCB *acb;
228     int aio_ret;
229 
230     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
231 
232     s = bs->opaque;
233     backing_s = backing->opaque;
234 
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);
238 
239     do_drain_begin(drain_type, bs);
240 
241     g_assert_cmpint(s->drain_count, ==, 1);
242     g_assert_cmpint(backing_s->drain_count, ==, !!recursive);
243 
244     do_drain_end(drain_type, bs);
245 
246     g_assert_cmpint(s->drain_count, ==, 0);
247     g_assert_cmpint(backing_s->drain_count, ==, 0);
248 
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);
254 
255     g_assert_cmpint(s->drain_count, ==, 0);
256     g_assert_cmpint(backing_s->drain_count, ==, 0);
257 
258     do_drain_begin(drain_type, bs);
259 
260     g_assert_cmpint(aio_ret, ==, 0);
261     g_assert_cmpint(s->drain_count, ==, 1);
262     g_assert_cmpint(backing_s->drain_count, ==, !!recursive);
263 
264     do_drain_end(drain_type, bs);
265 
266     g_assert_cmpint(s->drain_count, ==, 0);
267     g_assert_cmpint(backing_s->drain_count, ==, 0);
268 }
269 
270 static void test_drv_cb_drain_all(void)
271 {
272     BlockBackend *blk = test_setup();
273     test_drv_cb_common(blk, BDRV_DRAIN_ALL, true);
274     blk_unref(blk);
275 }
276 
277 static void test_drv_cb_drain(void)
278 {
279     BlockBackend *blk = test_setup();
280     test_drv_cb_common(blk, BDRV_DRAIN, false);
281     blk_unref(blk);
282 }
283 
284 static void coroutine_fn test_drv_cb_co_drain_all_entry(void)
285 {
286     BlockBackend *blk = blk_all_next(NULL);
287     test_drv_cb_common(blk, BDRV_DRAIN_ALL, true);
288 }
289 
290 static void test_drv_cb_co_drain_all(void)
291 {
292     BlockBackend *blk = test_setup();
293     call_in_coroutine(test_drv_cb_co_drain_all_entry);
294     blk_unref(blk);
295 }
296 
297 static void coroutine_fn test_drv_cb_co_drain_entry(void)
298 {
299     BlockBackend *blk = blk_all_next(NULL);
300     test_drv_cb_common(blk, BDRV_DRAIN, false);
301 }
302 
303 static void test_drv_cb_co_drain(void)
304 {
305     BlockBackend *blk = test_setup();
306     call_in_coroutine(test_drv_cb_co_drain_entry);
307     blk_unref(blk);
308 }
309 
310 static void 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     /* 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     bdrv_dec_in_flight(data->bs); /* incremented by test_iothread_common() */
516     aio_context_release(bdrv_get_aio_context(data->bs));
517 }
518 
519 /*
520  * Starts an AIO request on a BDS that runs in the AioContext of iothread 1.
521  * The request involves a BH on iothread 2 before it can complete.
522  *
523  * @drain_thread = 0 means that do_drain_begin/end are called from the main
524  * thread, @drain_thread = 1 means that they are called from iothread 1. Drain
525  * for this BDS cannot be called from iothread 2 because only the main thread
526  * may do cross-AioContext polling.
527  */
528 static void test_iothread_common(enum drain_type drain_type, int drain_thread)
529 {
530     BlockBackend *blk;
531     BlockDriverState *bs;
532     BDRVTestState *s;
533     BlockAIOCB *acb;
534     Coroutine *co;
535     int aio_ret;
536     struct test_iothread_data data;
537 
538     IOThread *a = iothread_new();
539     IOThread *b = iothread_new();
540     AioContext *ctx_a = iothread_get_aio_context(a);
541     AioContext *ctx_b = iothread_get_aio_context(b);
542 
543     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
544 
545     /* bdrv_drain_all() may only be called from the main loop thread */
546     if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) {
547         goto out;
548     }
549 
550     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
551     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
552                               &error_abort);
553     s = bs->opaque;
554     blk_insert_bs(blk, bs, &error_abort);
555     blk_set_disable_request_queuing(blk, true);
556 
557     blk_set_aio_context(blk, ctx_a, &error_abort);
558     aio_context_acquire(ctx_a);
559 
560     s->bh_indirection_ctx = ctx_b;
561 
562     aio_ret = -EINPROGRESS;
563     qemu_event_reset(&done_event);
564 
565     if (drain_thread == 0) {
566         acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret);
567     } else {
568         acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
569     }
570     g_assert(acb != NULL);
571     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
572 
573     aio_context_release(ctx_a);
574 
575     data = (struct test_iothread_data) {
576         .bs         = bs,
577         .drain_type = drain_type,
578         .aio_ret    = &aio_ret,
579     };
580 
581     switch (drain_thread) {
582     case 0:
583         if (drain_type != BDRV_DRAIN_ALL) {
584             aio_context_acquire(ctx_a);
585         }
586 
587         /*
588          * Increment in_flight so that do_drain_begin() waits for
589          * test_iothread_main_thread_bh(). This prevents the race between
590          * test_iothread_main_thread_bh() in IOThread a and do_drain_begin() in
591          * this thread. test_iothread_main_thread_bh() decrements in_flight.
592          */
593         bdrv_inc_in_flight(bs);
594         aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data);
595 
596         /* The request is running on the IOThread a. Draining its block device
597          * will make sure that it has completed as far as the BDS is concerned,
598          * but the drain in this thread can continue immediately after
599          * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
600          * later. */
601         do_drain_begin(drain_type, bs);
602         g_assert_cmpint(bs->in_flight, ==, 0);
603 
604         if (drain_type != BDRV_DRAIN_ALL) {
605             aio_context_release(ctx_a);
606         }
607         qemu_event_wait(&done_event);
608         if (drain_type != BDRV_DRAIN_ALL) {
609             aio_context_acquire(ctx_a);
610         }
611 
612         g_assert_cmpint(aio_ret, ==, 0);
613         do_drain_end(drain_type, bs);
614 
615         if (drain_type != BDRV_DRAIN_ALL) {
616             aio_context_release(ctx_a);
617         }
618         break;
619     case 1:
620         co = qemu_coroutine_create(test_iothread_drain_co_entry, &data);
621         aio_co_enter(ctx_a, co);
622         AIO_WAIT_WHILE_UNLOCKED(NULL, !data.co_done);
623         break;
624     default:
625         g_assert_not_reached();
626     }
627 
628     aio_context_acquire(ctx_a);
629     blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort);
630     aio_context_release(ctx_a);
631 
632     bdrv_unref(bs);
633     blk_unref(blk);
634 
635 out:
636     iothread_join(a);
637     iothread_join(b);
638 }
639 
640 static void test_iothread_drain_all(void)
641 {
642     test_iothread_common(BDRV_DRAIN_ALL, 0);
643     test_iothread_common(BDRV_DRAIN_ALL, 1);
644 }
645 
646 static void test_iothread_drain(void)
647 {
648     test_iothread_common(BDRV_DRAIN, 0);
649     test_iothread_common(BDRV_DRAIN, 1);
650 }
651 
652 
653 typedef struct TestBlockJob {
654     BlockJob common;
655     BlockDriverState *bs;
656     int run_ret;
657     int prepare_ret;
658     bool running;
659     bool should_complete;
660 } TestBlockJob;
661 
662 static int test_job_prepare(Job *job)
663 {
664     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
665 
666     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
667     bdrv_flush(s->bs);
668     return s->prepare_ret;
669 }
670 
671 static void test_job_commit(Job *job)
672 {
673     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
674 
675     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
676     bdrv_flush(s->bs);
677 }
678 
679 static void test_job_abort(Job *job)
680 {
681     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
682 
683     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
684     bdrv_flush(s->bs);
685 }
686 
687 static int coroutine_fn test_job_run(Job *job, Error **errp)
688 {
689     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
690 
691     /* We are running the actual job code past the pause point in
692      * job_co_entry(). */
693     s->running = true;
694 
695     job_transition_to_ready(&s->common.job);
696     while (!s->should_complete) {
697         /* Avoid job_sleep_ns() because it marks the job as !busy. We want to
698          * emulate some actual activity (probably some I/O) here so that drain
699          * has to wait for this activity to stop. */
700         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000);
701 
702         job_pause_point(&s->common.job);
703     }
704 
705     return s->run_ret;
706 }
707 
708 static void test_job_complete(Job *job, Error **errp)
709 {
710     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
711     s->should_complete = true;
712 }
713 
714 BlockJobDriver test_job_driver = {
715     .job_driver = {
716         .instance_size  = sizeof(TestBlockJob),
717         .free           = block_job_free,
718         .user_resume    = block_job_user_resume,
719         .run            = test_job_run,
720         .complete       = test_job_complete,
721         .prepare        = test_job_prepare,
722         .commit         = test_job_commit,
723         .abort          = test_job_abort,
724     },
725 };
726 
727 enum test_job_result {
728     TEST_JOB_SUCCESS,
729     TEST_JOB_FAIL_RUN,
730     TEST_JOB_FAIL_PREPARE,
731 };
732 
733 enum test_job_drain_node {
734     TEST_JOB_DRAIN_SRC,
735     TEST_JOB_DRAIN_SRC_CHILD,
736 };
737 
738 static void test_blockjob_common_drain_node(enum drain_type drain_type,
739                                             bool use_iothread,
740                                             enum test_job_result result,
741                                             enum test_job_drain_node drain_node)
742 {
743     BlockBackend *blk_src, *blk_target;
744     BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs;
745     BlockJob *job;
746     TestBlockJob *tjob;
747     IOThread *iothread = NULL;
748     AioContext *ctx;
749     int ret;
750 
751     src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR,
752                                &error_abort);
753     src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing",
754                                        BDRV_O_RDWR, &error_abort);
755     src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay",
756                                        BDRV_O_RDWR, &error_abort);
757 
758     bdrv_set_backing_hd(src_overlay, src, &error_abort);
759     bdrv_unref(src);
760     bdrv_set_backing_hd(src, src_backing, &error_abort);
761     bdrv_unref(src_backing);
762 
763     blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
764     blk_insert_bs(blk_src, src_overlay, &error_abort);
765 
766     switch (drain_node) {
767     case TEST_JOB_DRAIN_SRC:
768         drain_bs = src;
769         break;
770     case TEST_JOB_DRAIN_SRC_CHILD:
771         drain_bs = src_backing;
772         break;
773     default:
774         g_assert_not_reached();
775     }
776 
777     if (use_iothread) {
778         iothread = iothread_new();
779         ctx = iothread_get_aio_context(iothread);
780         blk_set_aio_context(blk_src, ctx, &error_abort);
781     } else {
782         ctx = qemu_get_aio_context();
783     }
784 
785     target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR,
786                                   &error_abort);
787     blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
788     blk_insert_bs(blk_target, target, &error_abort);
789     blk_set_allow_aio_context_change(blk_target, true);
790 
791     aio_context_acquire(ctx);
792     tjob = block_job_create("job0", &test_job_driver, NULL, src,
793                             0, BLK_PERM_ALL,
794                             0, 0, NULL, NULL, &error_abort);
795     tjob->bs = src;
796     job = &tjob->common;
797     block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort);
798 
799     switch (result) {
800     case TEST_JOB_SUCCESS:
801         break;
802     case TEST_JOB_FAIL_RUN:
803         tjob->run_ret = -EIO;
804         break;
805     case TEST_JOB_FAIL_PREPARE:
806         tjob->prepare_ret = -EIO;
807         break;
808     }
809     aio_context_release(ctx);
810 
811     job_start(&job->job);
812 
813     if (use_iothread) {
814         /* job_co_entry() is run in the I/O thread, wait for the actual job
815          * code to start (we don't want to catch the job in the pause point in
816          * job_co_entry(). */
817         while (!tjob->running) {
818             aio_poll(qemu_get_aio_context(), false);
819         }
820     }
821 
822     WITH_JOB_LOCK_GUARD() {
823         g_assert_cmpint(job->job.pause_count, ==, 0);
824         g_assert_false(job->job.paused);
825         g_assert_true(tjob->running);
826         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
827     }
828 
829     do_drain_begin_unlocked(drain_type, drain_bs);
830 
831     WITH_JOB_LOCK_GUARD() {
832         if (drain_type == BDRV_DRAIN_ALL) {
833             /* bdrv_drain_all() drains both src and target */
834             g_assert_cmpint(job->job.pause_count, ==, 2);
835         } else {
836             g_assert_cmpint(job->job.pause_count, ==, 1);
837         }
838         g_assert_true(job->job.paused);
839         g_assert_false(job->job.busy); /* The job is paused */
840     }
841 
842     do_drain_end_unlocked(drain_type, drain_bs);
843 
844     if (use_iothread) {
845         /*
846          * Here we are waiting for the paused status to change,
847          * so don't bother protecting the read every time.
848          *
849          * paused is reset in the I/O thread, wait for it
850          */
851         while (job->job.paused) {
852             aio_poll(qemu_get_aio_context(), false);
853         }
854     }
855 
856     WITH_JOB_LOCK_GUARD() {
857         g_assert_cmpint(job->job.pause_count, ==, 0);
858         g_assert_false(job->job.paused);
859         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
860     }
861 
862     do_drain_begin_unlocked(drain_type, target);
863 
864     WITH_JOB_LOCK_GUARD() {
865         if (drain_type == BDRV_DRAIN_ALL) {
866             /* bdrv_drain_all() drains both src and target */
867             g_assert_cmpint(job->job.pause_count, ==, 2);
868         } else {
869             g_assert_cmpint(job->job.pause_count, ==, 1);
870         }
871         g_assert_true(job->job.paused);
872         g_assert_false(job->job.busy); /* The job is paused */
873     }
874 
875     do_drain_end_unlocked(drain_type, target);
876 
877     if (use_iothread) {
878         /*
879          * Here we are waiting for the paused status to change,
880          * so don't bother protecting the read every time.
881          *
882          * paused is reset in the I/O thread, wait for it
883          */
884         while (job->job.paused) {
885             aio_poll(qemu_get_aio_context(), false);
886         }
887     }
888 
889     WITH_JOB_LOCK_GUARD() {
890         g_assert_cmpint(job->job.pause_count, ==, 0);
891         g_assert_false(job->job.paused);
892         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
893     }
894 
895     WITH_JOB_LOCK_GUARD() {
896         ret = job_complete_sync_locked(&job->job, &error_abort);
897     }
898     g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO));
899 
900     aio_context_acquire(ctx);
901     if (use_iothread) {
902         blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort);
903         assert(blk_get_aio_context(blk_target) == qemu_get_aio_context());
904     }
905     aio_context_release(ctx);
906 
907     blk_unref(blk_src);
908     blk_unref(blk_target);
909     bdrv_unref(src_overlay);
910     bdrv_unref(target);
911 
912     if (iothread) {
913         iothread_join(iothread);
914     }
915 }
916 
917 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread,
918                                  enum test_job_result result)
919 {
920     test_blockjob_common_drain_node(drain_type, use_iothread, result,
921                                     TEST_JOB_DRAIN_SRC);
922     test_blockjob_common_drain_node(drain_type, use_iothread, result,
923                                     TEST_JOB_DRAIN_SRC_CHILD);
924 }
925 
926 static void test_blockjob_drain_all(void)
927 {
928     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS);
929 }
930 
931 static void test_blockjob_drain(void)
932 {
933     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS);
934 }
935 
936 static void test_blockjob_error_drain_all(void)
937 {
938     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN);
939     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE);
940 }
941 
942 static void test_blockjob_error_drain(void)
943 {
944     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN);
945     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE);
946 }
947 
948 static void test_blockjob_iothread_drain_all(void)
949 {
950     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS);
951 }
952 
953 static void test_blockjob_iothread_drain(void)
954 {
955     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS);
956 }
957 
958 static void test_blockjob_iothread_error_drain_all(void)
959 {
960     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN);
961     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE);
962 }
963 
964 static void test_blockjob_iothread_error_drain(void)
965 {
966     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN);
967     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE);
968 }
969 
970 
971 typedef struct BDRVTestTopState {
972     BdrvChild *wait_child;
973 } BDRVTestTopState;
974 
975 static void bdrv_test_top_close(BlockDriverState *bs)
976 {
977     BdrvChild *c, *next_c;
978 
979     bdrv_graph_wrlock(NULL);
980     QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
981         bdrv_unref_child(bs, c);
982     }
983     bdrv_graph_wrunlock();
984 }
985 
986 static int coroutine_fn GRAPH_RDLOCK
987 bdrv_test_top_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
988                         QEMUIOVector *qiov, BdrvRequestFlags flags)
989 {
990     BDRVTestTopState *tts = bs->opaque;
991     return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags);
992 }
993 
994 static BlockDriver bdrv_test_top_driver = {
995     .format_name            = "test_top_driver",
996     .instance_size          = sizeof(BDRVTestTopState),
997 
998     .bdrv_close             = bdrv_test_top_close,
999     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
1000 
1001     .bdrv_child_perm        = bdrv_default_perms,
1002 };
1003 
1004 typedef struct TestCoDeleteByDrainData {
1005     BlockBackend *blk;
1006     bool detach_instead_of_delete;
1007     bool done;
1008 } TestCoDeleteByDrainData;
1009 
1010 static void coroutine_fn test_co_delete_by_drain(void *opaque)
1011 {
1012     TestCoDeleteByDrainData *dbdd = opaque;
1013     BlockBackend *blk = dbdd->blk;
1014     BlockDriverState *bs = blk_bs(blk);
1015     BDRVTestTopState *tts = bs->opaque;
1016     void *buffer = g_malloc(65536);
1017     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536);
1018 
1019     /* Pretend some internal write operation from parent to child.
1020      * Important: We have to read from the child, not from the parent!
1021      * Draining works by first propagating it all up the tree to the
1022      * root and then waiting for drainage from root to the leaves
1023      * (protocol nodes).  If we have a request waiting on the root,
1024      * everything will be drained before we go back down the tree, but
1025      * we do not want that.  We want to be in the middle of draining
1026      * when this following requests returns. */
1027     bdrv_graph_co_rdlock();
1028     bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0);
1029     bdrv_graph_co_rdunlock();
1030 
1031     g_assert_cmpint(bs->refcnt, ==, 1);
1032 
1033     if (!dbdd->detach_instead_of_delete) {
1034         blk_co_unref(blk);
1035     } else {
1036         BdrvChild *c, *next_c;
1037         QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1038             bdrv_co_unref_child(bs, c);
1039         }
1040     }
1041 
1042     dbdd->done = true;
1043     g_free(buffer);
1044 }
1045 
1046 /**
1047  * Test what happens when some BDS has some children, you drain one of
1048  * them and this results in the BDS being deleted.
1049  *
1050  * If @detach_instead_of_delete is set, the BDS is not going to be
1051  * deleted but will only detach all of its children.
1052  */
1053 static void do_test_delete_by_drain(bool detach_instead_of_delete,
1054                                     enum drain_type drain_type)
1055 {
1056     BlockBackend *blk;
1057     BlockDriverState *bs, *child_bs, *null_bs;
1058     BDRVTestTopState *tts;
1059     TestCoDeleteByDrainData dbdd;
1060     Coroutine *co;
1061 
1062     bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR,
1063                               &error_abort);
1064     bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1065     tts = bs->opaque;
1066 
1067     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1068                         &error_abort);
1069     bdrv_graph_wrlock(NULL);
1070     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds,
1071                       BDRV_CHILD_DATA, &error_abort);
1072     bdrv_graph_wrunlock();
1073 
1074     /* This child will be the one to pass to requests through to, and
1075      * it will stall until a drain occurs */
1076     child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR,
1077                                     &error_abort);
1078     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1079     /* Takes our reference to child_bs */
1080     bdrv_graph_wrlock(NULL);
1081     tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child",
1082                                         &child_of_bds,
1083                                         BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY,
1084                                         &error_abort);
1085     bdrv_graph_wrunlock();
1086 
1087     /* This child is just there to be deleted
1088      * (for detach_instead_of_delete == true) */
1089     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1090                         &error_abort);
1091     bdrv_graph_wrlock(NULL);
1092     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA,
1093                       &error_abort);
1094     bdrv_graph_wrunlock();
1095 
1096     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1097     blk_insert_bs(blk, bs, &error_abort);
1098 
1099     /* Referenced by blk now */
1100     bdrv_unref(bs);
1101 
1102     g_assert_cmpint(bs->refcnt, ==, 1);
1103     g_assert_cmpint(child_bs->refcnt, ==, 1);
1104     g_assert_cmpint(null_bs->refcnt, ==, 1);
1105 
1106 
1107     dbdd = (TestCoDeleteByDrainData){
1108         .blk = blk,
1109         .detach_instead_of_delete = detach_instead_of_delete,
1110         .done = false,
1111     };
1112     co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd);
1113     qemu_coroutine_enter(co);
1114 
1115     /* Drain the child while the read operation is still pending.
1116      * This should result in the operation finishing and
1117      * test_co_delete_by_drain() resuming.  Thus, @bs will be deleted
1118      * and the coroutine will exit while this drain operation is still
1119      * in progress. */
1120     switch (drain_type) {
1121     case BDRV_DRAIN:
1122         bdrv_ref(child_bs);
1123         bdrv_drain(child_bs);
1124         bdrv_unref(child_bs);
1125         break;
1126     case BDRV_DRAIN_ALL:
1127         bdrv_drain_all_begin();
1128         bdrv_drain_all_end();
1129         break;
1130     default:
1131         g_assert_not_reached();
1132     }
1133 
1134     while (!dbdd.done) {
1135         aio_poll(qemu_get_aio_context(), true);
1136     }
1137 
1138     if (detach_instead_of_delete) {
1139         /* Here, the reference has not passed over to the coroutine,
1140          * so we have to delete the BB ourselves */
1141         blk_unref(blk);
1142     }
1143 }
1144 
1145 static void test_delete_by_drain(void)
1146 {
1147     do_test_delete_by_drain(false, BDRV_DRAIN);
1148 }
1149 
1150 static void test_detach_by_drain_all(void)
1151 {
1152     do_test_delete_by_drain(true, BDRV_DRAIN_ALL);
1153 }
1154 
1155 static void test_detach_by_drain(void)
1156 {
1157     do_test_delete_by_drain(true, BDRV_DRAIN);
1158 }
1159 
1160 
1161 struct detach_by_parent_data {
1162     BlockDriverState *parent_b;
1163     BdrvChild *child_b;
1164     BlockDriverState *c;
1165     BdrvChild *child_c;
1166     bool by_parent_cb;
1167     bool detach_on_drain;
1168 };
1169 static struct detach_by_parent_data detach_by_parent_data;
1170 
1171 static void detach_indirect_bh(void *opaque)
1172 {
1173     struct detach_by_parent_data *data = opaque;
1174 
1175     bdrv_dec_in_flight(data->child_b->bs);
1176 
1177     bdrv_graph_wrlock(NULL);
1178     bdrv_unref_child(data->parent_b, data->child_b);
1179 
1180     bdrv_ref(data->c);
1181     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
1182                                       &child_of_bds, BDRV_CHILD_DATA,
1183                                       &error_abort);
1184     bdrv_graph_wrunlock();
1185 }
1186 
1187 static void detach_by_parent_aio_cb(void *opaque, int ret)
1188 {
1189     struct detach_by_parent_data *data = &detach_by_parent_data;
1190 
1191     g_assert_cmpint(ret, ==, 0);
1192     if (data->by_parent_cb) {
1193         bdrv_inc_in_flight(data->child_b->bs);
1194         detach_indirect_bh(data);
1195     }
1196 }
1197 
1198 static void detach_by_driver_cb_drained_begin(BdrvChild *child)
1199 {
1200     struct detach_by_parent_data *data = &detach_by_parent_data;
1201 
1202     if (!data->detach_on_drain) {
1203         return;
1204     }
1205     data->detach_on_drain = false;
1206 
1207     bdrv_inc_in_flight(data->child_b->bs);
1208     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1209                             detach_indirect_bh, &detach_by_parent_data);
1210     child_of_bds.drained_begin(child);
1211 }
1212 
1213 static BdrvChildClass detach_by_driver_cb_class;
1214 
1215 /*
1216  * Initial graph:
1217  *
1218  * PA     PB
1219  *    \ /   \
1220  *     A     B     C
1221  *
1222  * by_parent_cb == true:  Test that parent callbacks don't poll
1223  *
1224  *     PA has a pending write request whose callback changes the child nodes of
1225  *     PB: It removes B and adds C instead. The subtree of PB is drained, which
1226  *     will indirectly drain the write request, too.
1227  *
1228  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1229  *
1230  *     PA's BdrvChildClass has a .drained_begin callback that schedules a BH
1231  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
1232  *     state is messed up, but if it is only polled in the single
1233  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1234  */
1235 static void test_detach_indirect(bool by_parent_cb)
1236 {
1237     BlockBackend *blk;
1238     BlockDriverState *parent_a, *parent_b, *a, *b, *c;
1239     BdrvChild *child_a, *child_b;
1240     BlockAIOCB *acb;
1241 
1242     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
1243 
1244     if (!by_parent_cb) {
1245         detach_by_driver_cb_class = child_of_bds;
1246         detach_by_driver_cb_class.drained_begin =
1247             detach_by_driver_cb_drained_begin;
1248         detach_by_driver_cb_class.drained_end = NULL;
1249         detach_by_driver_cb_class.drained_poll = NULL;
1250     }
1251 
1252     detach_by_parent_data = (struct detach_by_parent_data) {
1253         .detach_on_drain = false,
1254     };
1255 
1256     /* Create all involved nodes */
1257     parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR,
1258                                     &error_abort);
1259     parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0,
1260                                     &error_abort);
1261 
1262     a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort);
1263     b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort);
1264     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
1265 
1266     /* blk is a BB for parent-a */
1267     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1268     blk_insert_bs(blk, parent_a, &error_abort);
1269     bdrv_unref(parent_a);
1270 
1271     /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1272      * callback must not return immediately. */
1273     if (!by_parent_cb) {
1274         BDRVTestState *s = parent_a->opaque;
1275         s->sleep_in_drain_begin = true;
1276     }
1277 
1278     /* Set child relationships */
1279     bdrv_ref(b);
1280     bdrv_ref(a);
1281     bdrv_graph_wrlock(NULL);
1282     child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds,
1283                                 BDRV_CHILD_DATA, &error_abort);
1284     child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds,
1285                                 BDRV_CHILD_COW, &error_abort);
1286 
1287     bdrv_ref(a);
1288     bdrv_attach_child(parent_a, a, "PA-A",
1289                       by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class,
1290                       BDRV_CHILD_DATA, &error_abort);
1291     bdrv_graph_wrunlock();
1292 
1293     g_assert_cmpint(parent_a->refcnt, ==, 1);
1294     g_assert_cmpint(parent_b->refcnt, ==, 1);
1295     g_assert_cmpint(a->refcnt, ==, 3);
1296     g_assert_cmpint(b->refcnt, ==, 2);
1297     g_assert_cmpint(c->refcnt, ==, 1);
1298 
1299     g_assert(QLIST_FIRST(&parent_b->children) == child_a);
1300     g_assert(QLIST_NEXT(child_a, next) == child_b);
1301     g_assert(QLIST_NEXT(child_b, next) == NULL);
1302 
1303     /* Start the evil write request */
1304     detach_by_parent_data = (struct detach_by_parent_data) {
1305         .parent_b = parent_b,
1306         .child_b = child_b,
1307         .c = c,
1308         .by_parent_cb = by_parent_cb,
1309         .detach_on_drain = true,
1310     };
1311     acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL);
1312     g_assert(acb != NULL);
1313 
1314     /* Drain and check the expected result */
1315     bdrv_drained_begin(parent_b);
1316     bdrv_drained_begin(a);
1317     bdrv_drained_begin(b);
1318     bdrv_drained_begin(c);
1319 
1320     g_assert(detach_by_parent_data.child_c != NULL);
1321 
1322     g_assert_cmpint(parent_a->refcnt, ==, 1);
1323     g_assert_cmpint(parent_b->refcnt, ==, 1);
1324     g_assert_cmpint(a->refcnt, ==, 3);
1325     g_assert_cmpint(b->refcnt, ==, 1);
1326     g_assert_cmpint(c->refcnt, ==, 2);
1327 
1328     g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c);
1329     g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a);
1330     g_assert(QLIST_NEXT(child_a, next) == NULL);
1331 
1332     g_assert_cmpint(parent_a->quiesce_counter, ==, 1);
1333     g_assert_cmpint(parent_b->quiesce_counter, ==, 3);
1334     g_assert_cmpint(a->quiesce_counter, ==, 1);
1335     g_assert_cmpint(b->quiesce_counter, ==, 1);
1336     g_assert_cmpint(c->quiesce_counter, ==, 1);
1337 
1338     bdrv_drained_end(parent_b);
1339     bdrv_drained_end(a);
1340     bdrv_drained_end(b);
1341     bdrv_drained_end(c);
1342 
1343     bdrv_unref(parent_b);
1344     blk_unref(blk);
1345 
1346     g_assert_cmpint(a->refcnt, ==, 1);
1347     g_assert_cmpint(b->refcnt, ==, 1);
1348     g_assert_cmpint(c->refcnt, ==, 1);
1349     bdrv_unref(a);
1350     bdrv_unref(b);
1351     bdrv_unref(c);
1352 }
1353 
1354 static void test_detach_by_parent_cb(void)
1355 {
1356     test_detach_indirect(true);
1357 }
1358 
1359 static void test_detach_by_driver_cb(void)
1360 {
1361     test_detach_indirect(false);
1362 }
1363 
1364 static void test_append_to_drained(void)
1365 {
1366     BlockBackend *blk;
1367     BlockDriverState *base, *overlay;
1368     BDRVTestState *base_s, *overlay_s;
1369 
1370     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1371     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
1372     base_s = base->opaque;
1373     blk_insert_bs(blk, base, &error_abort);
1374 
1375     overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR,
1376                                    &error_abort);
1377     overlay_s = overlay->opaque;
1378 
1379     do_drain_begin(BDRV_DRAIN, base);
1380     g_assert_cmpint(base->quiesce_counter, ==, 1);
1381     g_assert_cmpint(base_s->drain_count, ==, 1);
1382     g_assert_cmpint(base->in_flight, ==, 0);
1383 
1384     aio_context_acquire(qemu_get_aio_context());
1385     bdrv_append(overlay, base, &error_abort);
1386     aio_context_release(qemu_get_aio_context());
1387 
1388     g_assert_cmpint(base->in_flight, ==, 0);
1389     g_assert_cmpint(overlay->in_flight, ==, 0);
1390 
1391     g_assert_cmpint(base->quiesce_counter, ==, 1);
1392     g_assert_cmpint(base_s->drain_count, ==, 1);
1393     g_assert_cmpint(overlay->quiesce_counter, ==, 1);
1394     g_assert_cmpint(overlay_s->drain_count, ==, 1);
1395 
1396     do_drain_end(BDRV_DRAIN, base);
1397 
1398     g_assert_cmpint(base->quiesce_counter, ==, 0);
1399     g_assert_cmpint(base_s->drain_count, ==, 0);
1400     g_assert_cmpint(overlay->quiesce_counter, ==, 0);
1401     g_assert_cmpint(overlay_s->drain_count, ==, 0);
1402 
1403     bdrv_unref(overlay);
1404     bdrv_unref(base);
1405     blk_unref(blk);
1406 }
1407 
1408 static void test_set_aio_context(void)
1409 {
1410     BlockDriverState *bs;
1411     IOThread *a = iothread_new();
1412     IOThread *b = iothread_new();
1413     AioContext *ctx_a = iothread_get_aio_context(a);
1414     AioContext *ctx_b = iothread_get_aio_context(b);
1415 
1416     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
1417                               &error_abort);
1418 
1419     bdrv_drained_begin(bs);
1420     bdrv_try_change_aio_context(bs, ctx_a, NULL, &error_abort);
1421 
1422     aio_context_acquire(ctx_a);
1423     bdrv_drained_end(bs);
1424 
1425     bdrv_drained_begin(bs);
1426     bdrv_try_change_aio_context(bs, ctx_b, NULL, &error_abort);
1427     aio_context_release(ctx_a);
1428     aio_context_acquire(ctx_b);
1429     bdrv_try_change_aio_context(bs, qemu_get_aio_context(), NULL, &error_abort);
1430     aio_context_release(ctx_b);
1431     bdrv_drained_end(bs);
1432 
1433     bdrv_unref(bs);
1434     iothread_join(a);
1435     iothread_join(b);
1436 }
1437 
1438 
1439 typedef struct TestDropBackingBlockJob {
1440     BlockJob common;
1441     bool should_complete;
1442     bool *did_complete;
1443     BlockDriverState *detach_also;
1444     BlockDriverState *bs;
1445 } TestDropBackingBlockJob;
1446 
1447 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp)
1448 {
1449     TestDropBackingBlockJob *s =
1450         container_of(job, TestDropBackingBlockJob, common.job);
1451 
1452     while (!s->should_complete) {
1453         job_sleep_ns(job, 0);
1454     }
1455 
1456     return 0;
1457 }
1458 
1459 static void test_drop_backing_job_commit(Job *job)
1460 {
1461     TestDropBackingBlockJob *s =
1462         container_of(job, TestDropBackingBlockJob, common.job);
1463 
1464     bdrv_set_backing_hd(s->bs, NULL, &error_abort);
1465     bdrv_set_backing_hd(s->detach_also, NULL, &error_abort);
1466 
1467     *s->did_complete = true;
1468 }
1469 
1470 static const BlockJobDriver test_drop_backing_job_driver = {
1471     .job_driver = {
1472         .instance_size  = sizeof(TestDropBackingBlockJob),
1473         .free           = block_job_free,
1474         .user_resume    = block_job_user_resume,
1475         .run            = test_drop_backing_job_run,
1476         .commit         = test_drop_backing_job_commit,
1477     }
1478 };
1479 
1480 /**
1481  * Creates a child node with three parent nodes on it, and then runs a
1482  * block job on the final one, parent-node-2.
1483  *
1484  * The job is then asked to complete before a section where the child
1485  * is drained.
1486  *
1487  * Ending this section will undrain the child's parents, first
1488  * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
1489  * list is in reverse order of how they were added.  Ending the drain
1490  * on parent-node-2 will resume the job, thus completing it and
1491  * scheduling job_exit().
1492  *
1493  * Ending the drain on parent-node-1 will poll the AioContext, which
1494  * lets job_exit() and thus test_drop_backing_job_commit() run.  That
1495  * function first removes the child as parent-node-2's backing file.
1496  *
1497  * In old (and buggy) implementations, there are two problems with
1498  * that:
1499  * (A) bdrv_drain_invoke() polls for every node that leaves the
1500  *     drained section.  This means that job_exit() is scheduled
1501  *     before the child has left the drained section.  Its
1502  *     quiesce_counter is therefore still 1 when it is removed from
1503  *     parent-node-2.
1504  *
1505  * (B) bdrv_replace_child_noperm() calls drained_end() on the old
1506  *     child's parents as many times as the child is quiesced.  This
1507  *     means it will call drained_end() on parent-node-2 once.
1508  *     Because parent-node-2 is no longer quiesced at this point, this
1509  *     will fail.
1510  *
1511  * bdrv_replace_child_noperm() therefore must call drained_end() on
1512  * the parent only if it really is still drained because the child is
1513  * drained.
1514  *
1515  * If removing child from parent-node-2 was successful (as it should
1516  * be), test_drop_backing_job_commit() will then also remove the child
1517  * from parent-node-0.
1518  *
1519  * With an old version of our drain infrastructure ((A) above), that
1520  * resulted in the following flow:
1521  *
1522  * 1. child attempts to leave its drained section.  The call recurses
1523  *    to its parents.
1524  *
1525  * 2. parent-node-2 leaves the drained section.  Polling in
1526  *    bdrv_drain_invoke() will schedule job_exit().
1527  *
1528  * 3. parent-node-1 leaves the drained section.  Polling in
1529  *    bdrv_drain_invoke() will run job_exit(), thus disconnecting
1530  *    parent-node-0 from the child node.
1531  *
1532  * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
1533  *    iterate over the parents.  Thus, it now accesses the BdrvChild
1534  *    object that used to connect parent-node-0 and the child node.
1535  *    However, that object no longer exists, so it accesses a dangling
1536  *    pointer.
1537  *
1538  * The solution is to only poll once when running a bdrv_drained_end()
1539  * operation, specifically at the end when all drained_end()
1540  * operations for all involved nodes have been scheduled.
1541  * Note that this also solves (A) above, thus hiding (B).
1542  */
1543 static void test_blockjob_commit_by_drained_end(void)
1544 {
1545     BlockDriverState *bs_child, *bs_parents[3];
1546     TestDropBackingBlockJob *job;
1547     bool job_has_completed = false;
1548     int i;
1549 
1550     bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR,
1551                                     &error_abort);
1552 
1553     for (i = 0; i < 3; i++) {
1554         char name[32];
1555         snprintf(name, sizeof(name), "parent-node-%i", i);
1556         bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR,
1557                                              &error_abort);
1558         bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort);
1559     }
1560 
1561     job = block_job_create("job", &test_drop_backing_job_driver, NULL,
1562                            bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL,
1563                            &error_abort);
1564     job->bs = bs_parents[2];
1565 
1566     job->detach_also = bs_parents[0];
1567     job->did_complete = &job_has_completed;
1568 
1569     job_start(&job->common.job);
1570 
1571     job->should_complete = true;
1572     bdrv_drained_begin(bs_child);
1573     g_assert(!job_has_completed);
1574     bdrv_drained_end(bs_child);
1575     aio_poll(qemu_get_aio_context(), false);
1576     g_assert(job_has_completed);
1577 
1578     bdrv_unref(bs_parents[0]);
1579     bdrv_unref(bs_parents[1]);
1580     bdrv_unref(bs_parents[2]);
1581     bdrv_unref(bs_child);
1582 }
1583 
1584 
1585 typedef struct TestSimpleBlockJob {
1586     BlockJob common;
1587     bool should_complete;
1588     bool *did_complete;
1589 } TestSimpleBlockJob;
1590 
1591 static int coroutine_fn test_simple_job_run(Job *job, Error **errp)
1592 {
1593     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1594 
1595     while (!s->should_complete) {
1596         job_sleep_ns(job, 0);
1597     }
1598 
1599     return 0;
1600 }
1601 
1602 static void test_simple_job_clean(Job *job)
1603 {
1604     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1605     *s->did_complete = true;
1606 }
1607 
1608 static const BlockJobDriver test_simple_job_driver = {
1609     .job_driver = {
1610         .instance_size  = sizeof(TestSimpleBlockJob),
1611         .free           = block_job_free,
1612         .user_resume    = block_job_user_resume,
1613         .run            = test_simple_job_run,
1614         .clean          = test_simple_job_clean,
1615     },
1616 };
1617 
1618 static int drop_intermediate_poll_update_filename(BdrvChild *child,
1619                                                   BlockDriverState *new_base,
1620                                                   const char *filename,
1621                                                   Error **errp)
1622 {
1623     /*
1624      * We are free to poll here, which may change the block graph, if
1625      * it is not drained.
1626      */
1627 
1628     /* If the job is not drained: Complete it, schedule job_exit() */
1629     aio_poll(qemu_get_current_aio_context(), false);
1630     /* If the job is not drained: Run job_exit(), finish the job */
1631     aio_poll(qemu_get_current_aio_context(), false);
1632 
1633     return 0;
1634 }
1635 
1636 /**
1637  * Test a poll in the midst of bdrv_drop_intermediate().
1638  *
1639  * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(),
1640  * which can yield or poll.  This may lead to graph changes, unless
1641  * the whole subtree in question is drained.
1642  *
1643  * We test this on the following graph:
1644  *
1645  *                    Job
1646  *
1647  *                     |
1648  *                  job-node
1649  *                     |
1650  *                     v
1651  *
1652  *                  job-node
1653  *
1654  *                     |
1655  *                  backing
1656  *                     |
1657  *                     v
1658  *
1659  * node-2 --chain--> node-1 --chain--> node-0
1660  *
1661  * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0).
1662  *
1663  * This first updates node-2's backing filename by invoking
1664  * drop_intermediate_poll_update_filename(), which polls twice.  This
1665  * causes the job to finish, which in turns causes the job-node to be
1666  * deleted.
1667  *
1668  * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it
1669  * already has a pointer to the BdrvChild edge between job-node and
1670  * node-1.  When it tries to handle that edge, we probably get a
1671  * segmentation fault because the object no longer exists.
1672  *
1673  *
1674  * The solution is for bdrv_drop_intermediate() to drain top's
1675  * subtree.  This prevents graph changes from happening just because
1676  * BdrvChildClass.update_filename() yields or polls.  Thus, the block
1677  * job is paused during that drained section and must finish before or
1678  * after.
1679  *
1680  * (In addition, bdrv_replace_child() must keep the job paused.)
1681  */
1682 static void test_drop_intermediate_poll(void)
1683 {
1684     static BdrvChildClass chain_child_class;
1685     BlockDriverState *chain[3];
1686     TestSimpleBlockJob *job;
1687     BlockDriverState *job_node;
1688     bool job_has_completed = false;
1689     int i;
1690     int ret;
1691 
1692     chain_child_class = child_of_bds;
1693     chain_child_class.update_filename = drop_intermediate_poll_update_filename;
1694 
1695     for (i = 0; i < 3; i++) {
1696         char name[32];
1697         snprintf(name, 32, "node-%i", i);
1698 
1699         chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort);
1700     }
1701 
1702     job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR,
1703                                     &error_abort);
1704     bdrv_set_backing_hd(job_node, chain[1], &error_abort);
1705 
1706     /*
1707      * Establish the chain last, so the chain links are the first
1708      * elements in the BDS.parents lists
1709      */
1710     bdrv_graph_wrlock(NULL);
1711     for (i = 0; i < 3; i++) {
1712         if (i) {
1713             /* Takes the reference to chain[i - 1] */
1714             bdrv_attach_child(chain[i], chain[i - 1], "chain",
1715                               &chain_child_class, BDRV_CHILD_COW, &error_abort);
1716         }
1717     }
1718     bdrv_graph_wrunlock();
1719 
1720     job = block_job_create("job", &test_simple_job_driver, NULL, job_node,
1721                            0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort);
1722 
1723     /* The job has a reference now */
1724     bdrv_unref(job_node);
1725 
1726     job->did_complete = &job_has_completed;
1727 
1728     job_start(&job->common.job);
1729     job->should_complete = true;
1730 
1731     g_assert(!job_has_completed);
1732     ret = bdrv_drop_intermediate(chain[1], chain[0], NULL);
1733     aio_poll(qemu_get_aio_context(), false);
1734     g_assert(ret == 0);
1735     g_assert(job_has_completed);
1736 
1737     bdrv_unref(chain[2]);
1738 }
1739 
1740 
1741 typedef struct BDRVReplaceTestState {
1742     bool setup_completed;
1743     bool was_drained;
1744     bool was_undrained;
1745     bool has_read;
1746 
1747     int drain_count;
1748 
1749     bool yield_before_read;
1750     Coroutine *io_co;
1751     Coroutine *drain_co;
1752 } BDRVReplaceTestState;
1753 
1754 static void bdrv_replace_test_close(BlockDriverState *bs)
1755 {
1756 }
1757 
1758 /**
1759  * If @bs has a backing file:
1760  *   Yield if .yield_before_read is true (and wait for drain_begin to
1761  *   wake us up).
1762  *   Forward the read to bs->backing.  Set .has_read to true.
1763  *   If drain_begin has woken us, wake it in turn.
1764  *
1765  * Otherwise:
1766  *   Set .has_read to true and return success.
1767  */
1768 static int coroutine_fn GRAPH_RDLOCK
1769 bdrv_replace_test_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1770                             QEMUIOVector *qiov, BdrvRequestFlags flags)
1771 {
1772     BDRVReplaceTestState *s = bs->opaque;
1773 
1774     if (bs->backing) {
1775         int ret;
1776 
1777         g_assert(!s->drain_count);
1778 
1779         s->io_co = qemu_coroutine_self();
1780         if (s->yield_before_read) {
1781             s->yield_before_read = false;
1782             qemu_coroutine_yield();
1783         }
1784         s->io_co = NULL;
1785 
1786         ret = bdrv_co_preadv(bs->backing, offset, bytes, qiov, 0);
1787         s->has_read = true;
1788 
1789         /* Wake up drain_co if it runs */
1790         if (s->drain_co) {
1791             aio_co_wake(s->drain_co);
1792         }
1793 
1794         return ret;
1795     }
1796 
1797     s->has_read = true;
1798     return 0;
1799 }
1800 
1801 static void coroutine_fn bdrv_replace_test_drain_co(void *opaque)
1802 {
1803     BlockDriverState *bs = opaque;
1804     BDRVReplaceTestState *s = bs->opaque;
1805 
1806     /* Keep waking io_co up until it is done */
1807     while (s->io_co) {
1808         aio_co_wake(s->io_co);
1809         s->io_co = NULL;
1810         qemu_coroutine_yield();
1811     }
1812     s->drain_co = NULL;
1813     bdrv_dec_in_flight(bs);
1814 }
1815 
1816 /**
1817  * If .drain_count is 0, wake up .io_co if there is one; and set
1818  * .was_drained.
1819  * Increment .drain_count.
1820  */
1821 static void bdrv_replace_test_drain_begin(BlockDriverState *bs)
1822 {
1823     BDRVReplaceTestState *s = bs->opaque;
1824 
1825     if (!s->setup_completed) {
1826         return;
1827     }
1828 
1829     if (!s->drain_count) {
1830         s->drain_co = qemu_coroutine_create(bdrv_replace_test_drain_co, bs);
1831         bdrv_inc_in_flight(bs);
1832         aio_co_enter(bdrv_get_aio_context(bs), s->drain_co);
1833         s->was_drained = true;
1834     }
1835     s->drain_count++;
1836 }
1837 
1838 static void coroutine_fn bdrv_replace_test_read_entry(void *opaque)
1839 {
1840     BlockDriverState *bs = opaque;
1841     char data;
1842     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
1843     int ret;
1844 
1845     /* Queue a read request post-drain */
1846     bdrv_graph_co_rdlock();
1847     ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
1848     bdrv_graph_co_rdunlock();
1849 
1850     g_assert(ret >= 0);
1851     bdrv_dec_in_flight(bs);
1852 }
1853 
1854 /**
1855  * Reduce .drain_count, set .was_undrained once it reaches 0.
1856  * If .drain_count reaches 0 and the node has a backing file, issue a
1857  * read request.
1858  */
1859 static void bdrv_replace_test_drain_end(BlockDriverState *bs)
1860 {
1861     BDRVReplaceTestState *s = bs->opaque;
1862 
1863     if (!s->setup_completed) {
1864         return;
1865     }
1866 
1867     g_assert(s->drain_count > 0);
1868     if (!--s->drain_count) {
1869         s->was_undrained = true;
1870 
1871         if (bs->backing) {
1872             Coroutine *co = qemu_coroutine_create(bdrv_replace_test_read_entry,
1873                                                   bs);
1874             bdrv_inc_in_flight(bs);
1875             aio_co_enter(bdrv_get_aio_context(bs), co);
1876         }
1877     }
1878 }
1879 
1880 static BlockDriver bdrv_replace_test = {
1881     .format_name            = "replace_test",
1882     .instance_size          = sizeof(BDRVReplaceTestState),
1883     .supports_backing       = true,
1884 
1885     .bdrv_close             = bdrv_replace_test_close,
1886     .bdrv_co_preadv         = bdrv_replace_test_co_preadv,
1887 
1888     .bdrv_drain_begin       = bdrv_replace_test_drain_begin,
1889     .bdrv_drain_end         = bdrv_replace_test_drain_end,
1890 
1891     .bdrv_child_perm        = bdrv_default_perms,
1892 };
1893 
1894 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque)
1895 {
1896     int ret;
1897     char data;
1898 
1899     ret = blk_co_pread(opaque, 0, 1, &data, 0);
1900     g_assert(ret >= 0);
1901 }
1902 
1903 /**
1904  * We test two things:
1905  * (1) bdrv_replace_child_noperm() must not undrain the parent if both
1906  *     children are drained.
1907  * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
1908  *     drained child.  If the old child is drained, it must flush I/O
1909  *     requests after the new one has been attached.  If the new child
1910  *     is drained, it must flush I/O requests before the old one is
1911  *     detached.
1912  *
1913  * To do so, we create one parent node and two child nodes; then
1914  * attach one of the children (old_child_bs) to the parent, then
1915  * drain both old_child_bs and new_child_bs according to
1916  * old_drain_count and new_drain_count, respectively, and finally
1917  * we invoke bdrv_replace_node() to replace old_child_bs by
1918  * new_child_bs.
1919  *
1920  * The test block driver we use here (bdrv_replace_test) has a read
1921  * function that:
1922  * - For the parent node, can optionally yield, and then forwards the
1923  *   read to bdrv_preadv(),
1924  * - For the child node, just returns immediately.
1925  *
1926  * If the read yields, the drain_begin function will wake it up.
1927  *
1928  * The drain_end function issues a read on the parent once it is fully
1929  * undrained (which simulates requests starting to come in again).
1930  */
1931 static void do_test_replace_child_mid_drain(int old_drain_count,
1932                                             int new_drain_count)
1933 {
1934     BlockBackend *parent_blk;
1935     BlockDriverState *parent_bs;
1936     BlockDriverState *old_child_bs, *new_child_bs;
1937     BDRVReplaceTestState *parent_s;
1938     BDRVReplaceTestState *old_child_s, *new_child_s;
1939     Coroutine *io_co;
1940     int i;
1941 
1942     parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0,
1943                                      &error_abort);
1944     parent_s = parent_bs->opaque;
1945 
1946     parent_blk = blk_new(qemu_get_aio_context(),
1947                          BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
1948     blk_insert_bs(parent_blk, parent_bs, &error_abort);
1949 
1950     old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0,
1951                                         &error_abort);
1952     new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0,
1953                                         &error_abort);
1954     old_child_s = old_child_bs->opaque;
1955     new_child_s = new_child_bs->opaque;
1956 
1957     /* So that we can read something */
1958     parent_bs->total_sectors = 1;
1959     old_child_bs->total_sectors = 1;
1960     new_child_bs->total_sectors = 1;
1961 
1962     bdrv_ref(old_child_bs);
1963     bdrv_graph_wrlock(NULL);
1964     bdrv_attach_child(parent_bs, old_child_bs, "child", &child_of_bds,
1965                       BDRV_CHILD_COW, &error_abort);
1966     bdrv_graph_wrunlock();
1967     parent_s->setup_completed = true;
1968 
1969     for (i = 0; i < old_drain_count; i++) {
1970         bdrv_drained_begin(old_child_bs);
1971     }
1972     for (i = 0; i < new_drain_count; i++) {
1973         bdrv_drained_begin(new_child_bs);
1974     }
1975 
1976     if (!old_drain_count) {
1977         /*
1978          * Start a read operation that will yield, so it will not
1979          * complete before the node is drained.
1980          */
1981         parent_s->yield_before_read = true;
1982         io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co,
1983                                       parent_blk);
1984         qemu_coroutine_enter(io_co);
1985     }
1986 
1987     /* If we have started a read operation, it should have yielded */
1988     g_assert(!parent_s->has_read);
1989 
1990     /* Reset drained status so we can see what bdrv_replace_node() does */
1991     parent_s->was_drained = false;
1992     parent_s->was_undrained = false;
1993 
1994     g_assert(parent_bs->quiesce_counter == old_drain_count);
1995     bdrv_replace_node(old_child_bs, new_child_bs, &error_abort);
1996     g_assert(parent_bs->quiesce_counter == new_drain_count);
1997 
1998     if (!old_drain_count && !new_drain_count) {
1999         /*
2000          * From undrained to undrained drains and undrains the parent,
2001          * because bdrv_replace_node() contains a drained section for
2002          * @old_child_bs.
2003          */
2004         g_assert(parent_s->was_drained && parent_s->was_undrained);
2005     } else if (!old_drain_count && new_drain_count) {
2006         /*
2007          * From undrained to drained should drain the parent and keep
2008          * it that way.
2009          */
2010         g_assert(parent_s->was_drained && !parent_s->was_undrained);
2011     } else if (old_drain_count && !new_drain_count) {
2012         /*
2013          * From drained to undrained should undrain the parent and
2014          * keep it that way.
2015          */
2016         g_assert(!parent_s->was_drained && parent_s->was_undrained);
2017     } else /* if (old_drain_count && new_drain_count) */ {
2018         /*
2019          * From drained to drained must not undrain the parent at any
2020          * point
2021          */
2022         g_assert(!parent_s->was_drained && !parent_s->was_undrained);
2023     }
2024 
2025     if (!old_drain_count || !new_drain_count) {
2026         /*
2027          * If !old_drain_count, we have started a read request before
2028          * bdrv_replace_node().  If !new_drain_count, the parent must
2029          * have been undrained at some point, and
2030          * bdrv_replace_test_co_drain_end() starts a read request
2031          * then.
2032          */
2033         g_assert(parent_s->has_read);
2034     } else {
2035         /*
2036          * If the parent was never undrained, there is no way to start
2037          * a read request.
2038          */
2039         g_assert(!parent_s->has_read);
2040     }
2041 
2042     /* A drained child must have not received any request */
2043     g_assert(!(old_drain_count && old_child_s->has_read));
2044     g_assert(!(new_drain_count && new_child_s->has_read));
2045 
2046     for (i = 0; i < new_drain_count; i++) {
2047         bdrv_drained_end(new_child_bs);
2048     }
2049     for (i = 0; i < old_drain_count; i++) {
2050         bdrv_drained_end(old_child_bs);
2051     }
2052 
2053     /*
2054      * By now, bdrv_replace_test_co_drain_end() must have been called
2055      * at some point while the new child was attached to the parent.
2056      */
2057     g_assert(parent_s->has_read);
2058     g_assert(new_child_s->has_read);
2059 
2060     blk_unref(parent_blk);
2061     bdrv_unref(parent_bs);
2062     bdrv_unref(old_child_bs);
2063     bdrv_unref(new_child_bs);
2064 }
2065 
2066 static void test_replace_child_mid_drain(void)
2067 {
2068     int old_drain_count, new_drain_count;
2069 
2070     for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) {
2071         for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) {
2072             do_test_replace_child_mid_drain(old_drain_count, new_drain_count);
2073         }
2074     }
2075 }
2076 
2077 int main(int argc, char **argv)
2078 {
2079     int ret;
2080 
2081     bdrv_init();
2082     qemu_init_main_loop(&error_abort);
2083 
2084     g_test_init(&argc, &argv, NULL);
2085     qemu_event_init(&done_event, false);
2086 
2087     g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all);
2088     g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain);
2089 
2090     g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
2091                     test_drv_cb_co_drain_all);
2092     g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain);
2093 
2094     g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all);
2095     g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain);
2096 
2097     g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
2098                     test_quiesce_co_drain_all);
2099     g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain);
2100 
2101     g_test_add_func("/bdrv-drain/nested", test_nested);
2102 
2103     g_test_add_func("/bdrv-drain/graph-change/drain_all",
2104                     test_graph_change_drain_all);
2105 
2106     g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all);
2107     g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain);
2108 
2109     g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all);
2110     g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain);
2111 
2112     g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
2113                     test_blockjob_error_drain_all);
2114     g_test_add_func("/bdrv-drain/blockjob/error/drain",
2115                     test_blockjob_error_drain);
2116 
2117     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
2118                     test_blockjob_iothread_drain_all);
2119     g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
2120                     test_blockjob_iothread_drain);
2121 
2122     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
2123                     test_blockjob_iothread_error_drain_all);
2124     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
2125                     test_blockjob_iothread_error_drain);
2126 
2127     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
2128     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
2129     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
2130     g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb);
2131     g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb);
2132 
2133     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
2134 
2135     g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context);
2136 
2137     g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
2138                     test_blockjob_commit_by_drained_end);
2139 
2140     g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
2141                     test_drop_intermediate_poll);
2142 
2143     g_test_add_func("/bdrv-drain/replace_child/mid-drain",
2144                     test_replace_child_mid_drain);
2145 
2146     ret = g_test_run();
2147     qemu_event_destroy(&done_event);
2148     return ret;
2149 }
2150