xref: /openbmc/qemu/tests/unit/test-bdrv-drain.c (revision 370ed600)
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     g_assert_cmpint(qemu_get_aio_context()->external_disable_cnt, ==, 0);
477 
478     bdrv_unref(bs_b);
479     blk_unref(blk_b);
480 }
481 
482 struct test_iothread_data {
483     BlockDriverState *bs;
484     enum drain_type drain_type;
485     int *aio_ret;
486 };
487 
488 static void test_iothread_drain_entry(void *opaque)
489 {
490     struct test_iothread_data *data = opaque;
491 
492     aio_context_acquire(bdrv_get_aio_context(data->bs));
493     do_drain_begin(data->drain_type, data->bs);
494     g_assert_cmpint(*data->aio_ret, ==, 0);
495     do_drain_end(data->drain_type, data->bs);
496     aio_context_release(bdrv_get_aio_context(data->bs));
497 
498     qemu_event_set(&done_event);
499 }
500 
501 static void test_iothread_aio_cb(void *opaque, int ret)
502 {
503     int *aio_ret = opaque;
504     *aio_ret = ret;
505     qemu_event_set(&done_event);
506 }
507 
508 static void test_iothread_main_thread_bh(void *opaque)
509 {
510     struct test_iothread_data *data = opaque;
511 
512     /* Test that the AioContext is not yet locked in a random BH that is
513      * executed during drain, otherwise this would deadlock. */
514     aio_context_acquire(bdrv_get_aio_context(data->bs));
515     bdrv_flush(data->bs);
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     int aio_ret;
535     struct test_iothread_data data;
536 
537     IOThread *a = iothread_new();
538     IOThread *b = iothread_new();
539     AioContext *ctx_a = iothread_get_aio_context(a);
540     AioContext *ctx_b = iothread_get_aio_context(b);
541 
542     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
543 
544     /* bdrv_drain_all() may only be called from the main loop thread */
545     if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) {
546         goto out;
547     }
548 
549     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
550     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
551                               &error_abort);
552     s = bs->opaque;
553     blk_insert_bs(blk, bs, &error_abort);
554     blk_set_disable_request_queuing(blk, true);
555 
556     blk_set_aio_context(blk, ctx_a, &error_abort);
557     aio_context_acquire(ctx_a);
558 
559     s->bh_indirection_ctx = ctx_b;
560 
561     aio_ret = -EINPROGRESS;
562     qemu_event_reset(&done_event);
563 
564     if (drain_thread == 0) {
565         acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret);
566     } else {
567         acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
568     }
569     g_assert(acb != NULL);
570     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
571 
572     aio_context_release(ctx_a);
573 
574     data = (struct test_iothread_data) {
575         .bs         = bs,
576         .drain_type = drain_type,
577         .aio_ret    = &aio_ret,
578     };
579 
580     switch (drain_thread) {
581     case 0:
582         if (drain_type != BDRV_DRAIN_ALL) {
583             aio_context_acquire(ctx_a);
584         }
585 
586         aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data);
587 
588         /* The request is running on the IOThread a. Draining its block device
589          * will make sure that it has completed as far as the BDS is concerned,
590          * but the drain in this thread can continue immediately after
591          * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
592          * later. */
593         do_drain_begin(drain_type, bs);
594         g_assert_cmpint(bs->in_flight, ==, 0);
595 
596         if (drain_type != BDRV_DRAIN_ALL) {
597             aio_context_release(ctx_a);
598         }
599         qemu_event_wait(&done_event);
600         if (drain_type != BDRV_DRAIN_ALL) {
601             aio_context_acquire(ctx_a);
602         }
603 
604         g_assert_cmpint(aio_ret, ==, 0);
605         do_drain_end(drain_type, bs);
606 
607         if (drain_type != BDRV_DRAIN_ALL) {
608             aio_context_release(ctx_a);
609         }
610         break;
611     case 1:
612         aio_bh_schedule_oneshot(ctx_a, test_iothread_drain_entry, &data);
613         qemu_event_wait(&done_event);
614         break;
615     default:
616         g_assert_not_reached();
617     }
618 
619     aio_context_acquire(ctx_a);
620     blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort);
621     aio_context_release(ctx_a);
622 
623     bdrv_unref(bs);
624     blk_unref(blk);
625 
626 out:
627     iothread_join(a);
628     iothread_join(b);
629 }
630 
631 static void test_iothread_drain_all(void)
632 {
633     test_iothread_common(BDRV_DRAIN_ALL, 0);
634     test_iothread_common(BDRV_DRAIN_ALL, 1);
635 }
636 
637 static void test_iothread_drain(void)
638 {
639     test_iothread_common(BDRV_DRAIN, 0);
640     test_iothread_common(BDRV_DRAIN, 1);
641 }
642 
643 
644 typedef struct TestBlockJob {
645     BlockJob common;
646     BlockDriverState *bs;
647     int run_ret;
648     int prepare_ret;
649     bool running;
650     bool should_complete;
651 } TestBlockJob;
652 
653 static int test_job_prepare(Job *job)
654 {
655     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
656 
657     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
658     bdrv_flush(s->bs);
659     return s->prepare_ret;
660 }
661 
662 static void test_job_commit(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 }
669 
670 static void test_job_abort(Job *job)
671 {
672     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
673 
674     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
675     bdrv_flush(s->bs);
676 }
677 
678 static int coroutine_fn test_job_run(Job *job, Error **errp)
679 {
680     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
681 
682     /* We are running the actual job code past the pause point in
683      * job_co_entry(). */
684     s->running = true;
685 
686     job_transition_to_ready(&s->common.job);
687     while (!s->should_complete) {
688         /* Avoid job_sleep_ns() because it marks the job as !busy. We want to
689          * emulate some actual activity (probably some I/O) here so that drain
690          * has to wait for this activity to stop. */
691         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000);
692 
693         job_pause_point(&s->common.job);
694     }
695 
696     return s->run_ret;
697 }
698 
699 static void test_job_complete(Job *job, Error **errp)
700 {
701     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
702     s->should_complete = true;
703 }
704 
705 BlockJobDriver test_job_driver = {
706     .job_driver = {
707         .instance_size  = sizeof(TestBlockJob),
708         .free           = block_job_free,
709         .user_resume    = block_job_user_resume,
710         .run            = test_job_run,
711         .complete       = test_job_complete,
712         .prepare        = test_job_prepare,
713         .commit         = test_job_commit,
714         .abort          = test_job_abort,
715     },
716 };
717 
718 enum test_job_result {
719     TEST_JOB_SUCCESS,
720     TEST_JOB_FAIL_RUN,
721     TEST_JOB_FAIL_PREPARE,
722 };
723 
724 enum test_job_drain_node {
725     TEST_JOB_DRAIN_SRC,
726     TEST_JOB_DRAIN_SRC_CHILD,
727 };
728 
729 static void test_blockjob_common_drain_node(enum drain_type drain_type,
730                                             bool use_iothread,
731                                             enum test_job_result result,
732                                             enum test_job_drain_node drain_node)
733 {
734     BlockBackend *blk_src, *blk_target;
735     BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs;
736     BlockJob *job;
737     TestBlockJob *tjob;
738     IOThread *iothread = NULL;
739     AioContext *ctx;
740     int ret;
741 
742     src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR,
743                                &error_abort);
744     src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing",
745                                        BDRV_O_RDWR, &error_abort);
746     src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay",
747                                        BDRV_O_RDWR, &error_abort);
748 
749     bdrv_set_backing_hd(src_overlay, src, &error_abort);
750     bdrv_unref(src);
751     bdrv_set_backing_hd(src, src_backing, &error_abort);
752     bdrv_unref(src_backing);
753 
754     blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
755     blk_insert_bs(blk_src, src_overlay, &error_abort);
756 
757     switch (drain_node) {
758     case TEST_JOB_DRAIN_SRC:
759         drain_bs = src;
760         break;
761     case TEST_JOB_DRAIN_SRC_CHILD:
762         drain_bs = src_backing;
763         break;
764     default:
765         g_assert_not_reached();
766     }
767 
768     if (use_iothread) {
769         iothread = iothread_new();
770         ctx = iothread_get_aio_context(iothread);
771         blk_set_aio_context(blk_src, ctx, &error_abort);
772     } else {
773         ctx = qemu_get_aio_context();
774     }
775 
776     target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR,
777                                   &error_abort);
778     blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
779     blk_insert_bs(blk_target, target, &error_abort);
780     blk_set_allow_aio_context_change(blk_target, true);
781 
782     aio_context_acquire(ctx);
783     tjob = block_job_create("job0", &test_job_driver, NULL, src,
784                             0, BLK_PERM_ALL,
785                             0, 0, NULL, NULL, &error_abort);
786     tjob->bs = src;
787     job = &tjob->common;
788     block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort);
789 
790     switch (result) {
791     case TEST_JOB_SUCCESS:
792         break;
793     case TEST_JOB_FAIL_RUN:
794         tjob->run_ret = -EIO;
795         break;
796     case TEST_JOB_FAIL_PREPARE:
797         tjob->prepare_ret = -EIO;
798         break;
799     }
800     aio_context_release(ctx);
801 
802     job_start(&job->job);
803 
804     if (use_iothread) {
805         /* job_co_entry() is run in the I/O thread, wait for the actual job
806          * code to start (we don't want to catch the job in the pause point in
807          * job_co_entry(). */
808         while (!tjob->running) {
809             aio_poll(qemu_get_aio_context(), false);
810         }
811     }
812 
813     WITH_JOB_LOCK_GUARD() {
814         g_assert_cmpint(job->job.pause_count, ==, 0);
815         g_assert_false(job->job.paused);
816         g_assert_true(tjob->running);
817         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
818     }
819 
820     do_drain_begin_unlocked(drain_type, drain_bs);
821 
822     WITH_JOB_LOCK_GUARD() {
823         if (drain_type == BDRV_DRAIN_ALL) {
824             /* bdrv_drain_all() drains both src and target */
825             g_assert_cmpint(job->job.pause_count, ==, 2);
826         } else {
827             g_assert_cmpint(job->job.pause_count, ==, 1);
828         }
829         g_assert_true(job->job.paused);
830         g_assert_false(job->job.busy); /* The job is paused */
831     }
832 
833     do_drain_end_unlocked(drain_type, drain_bs);
834 
835     if (use_iothread) {
836         /*
837          * Here we are waiting for the paused status to change,
838          * so don't bother protecting the read every time.
839          *
840          * paused is reset in the I/O thread, wait for it
841          */
842         while (job->job.paused) {
843             aio_poll(qemu_get_aio_context(), false);
844         }
845     }
846 
847     WITH_JOB_LOCK_GUARD() {
848         g_assert_cmpint(job->job.pause_count, ==, 0);
849         g_assert_false(job->job.paused);
850         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
851     }
852 
853     do_drain_begin_unlocked(drain_type, target);
854 
855     WITH_JOB_LOCK_GUARD() {
856         if (drain_type == BDRV_DRAIN_ALL) {
857             /* bdrv_drain_all() drains both src and target */
858             g_assert_cmpint(job->job.pause_count, ==, 2);
859         } else {
860             g_assert_cmpint(job->job.pause_count, ==, 1);
861         }
862         g_assert_true(job->job.paused);
863         g_assert_false(job->job.busy); /* The job is paused */
864     }
865 
866     do_drain_end_unlocked(drain_type, target);
867 
868     if (use_iothread) {
869         /*
870          * Here we are waiting for the paused status to change,
871          * so don't bother protecting the read every time.
872          *
873          * paused is reset in the I/O thread, wait for it
874          */
875         while (job->job.paused) {
876             aio_poll(qemu_get_aio_context(), false);
877         }
878     }
879 
880     WITH_JOB_LOCK_GUARD() {
881         g_assert_cmpint(job->job.pause_count, ==, 0);
882         g_assert_false(job->job.paused);
883         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
884     }
885 
886     WITH_JOB_LOCK_GUARD() {
887         ret = job_complete_sync_locked(&job->job, &error_abort);
888     }
889     g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO));
890 
891     aio_context_acquire(ctx);
892     if (use_iothread) {
893         blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort);
894         assert(blk_get_aio_context(blk_target) == qemu_get_aio_context());
895     }
896     aio_context_release(ctx);
897 
898     blk_unref(blk_src);
899     blk_unref(blk_target);
900     bdrv_unref(src_overlay);
901     bdrv_unref(target);
902 
903     if (iothread) {
904         iothread_join(iothread);
905     }
906 }
907 
908 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread,
909                                  enum test_job_result result)
910 {
911     test_blockjob_common_drain_node(drain_type, use_iothread, result,
912                                     TEST_JOB_DRAIN_SRC);
913     test_blockjob_common_drain_node(drain_type, use_iothread, result,
914                                     TEST_JOB_DRAIN_SRC_CHILD);
915 }
916 
917 static void test_blockjob_drain_all(void)
918 {
919     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS);
920 }
921 
922 static void test_blockjob_drain(void)
923 {
924     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS);
925 }
926 
927 static void test_blockjob_error_drain_all(void)
928 {
929     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN);
930     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE);
931 }
932 
933 static void test_blockjob_error_drain(void)
934 {
935     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN);
936     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE);
937 }
938 
939 static void test_blockjob_iothread_drain_all(void)
940 {
941     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS);
942 }
943 
944 static void test_blockjob_iothread_drain(void)
945 {
946     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS);
947 }
948 
949 static void test_blockjob_iothread_error_drain_all(void)
950 {
951     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN);
952     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE);
953 }
954 
955 static void test_blockjob_iothread_error_drain(void)
956 {
957     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN);
958     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE);
959 }
960 
961 
962 typedef struct BDRVTestTopState {
963     BdrvChild *wait_child;
964 } BDRVTestTopState;
965 
966 static void bdrv_test_top_close(BlockDriverState *bs)
967 {
968     BdrvChild *c, *next_c;
969     QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
970         bdrv_unref_child(bs, c);
971     }
972 }
973 
974 static int coroutine_fn GRAPH_RDLOCK
975 bdrv_test_top_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
976                         QEMUIOVector *qiov, BdrvRequestFlags flags)
977 {
978     BDRVTestTopState *tts = bs->opaque;
979     return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags);
980 }
981 
982 static BlockDriver bdrv_test_top_driver = {
983     .format_name            = "test_top_driver",
984     .instance_size          = sizeof(BDRVTestTopState),
985 
986     .bdrv_close             = bdrv_test_top_close,
987     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
988 
989     .bdrv_child_perm        = bdrv_default_perms,
990 };
991 
992 typedef struct TestCoDeleteByDrainData {
993     BlockBackend *blk;
994     bool detach_instead_of_delete;
995     bool done;
996 } TestCoDeleteByDrainData;
997 
998 static void coroutine_fn test_co_delete_by_drain(void *opaque)
999 {
1000     TestCoDeleteByDrainData *dbdd = opaque;
1001     BlockBackend *blk = dbdd->blk;
1002     BlockDriverState *bs = blk_bs(blk);
1003     BDRVTestTopState *tts = bs->opaque;
1004     void *buffer = g_malloc(65536);
1005     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536);
1006 
1007     GRAPH_RDLOCK_GUARD();
1008 
1009     /* Pretend some internal write operation from parent to child.
1010      * Important: We have to read from the child, not from the parent!
1011      * Draining works by first propagating it all up the tree to the
1012      * root and then waiting for drainage from root to the leaves
1013      * (protocol nodes).  If we have a request waiting on the root,
1014      * everything will be drained before we go back down the tree, but
1015      * we do not want that.  We want to be in the middle of draining
1016      * when this following requests returns. */
1017     bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0);
1018 
1019     g_assert_cmpint(bs->refcnt, ==, 1);
1020 
1021     if (!dbdd->detach_instead_of_delete) {
1022         blk_unref(blk);
1023     } else {
1024         BdrvChild *c, *next_c;
1025         QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1026             bdrv_unref_child(bs, c);
1027         }
1028     }
1029 
1030     dbdd->done = true;
1031     g_free(buffer);
1032 }
1033 
1034 /**
1035  * Test what happens when some BDS has some children, you drain one of
1036  * them and this results in the BDS being deleted.
1037  *
1038  * If @detach_instead_of_delete is set, the BDS is not going to be
1039  * deleted but will only detach all of its children.
1040  */
1041 static void do_test_delete_by_drain(bool detach_instead_of_delete,
1042                                     enum drain_type drain_type)
1043 {
1044     BlockBackend *blk;
1045     BlockDriverState *bs, *child_bs, *null_bs;
1046     BDRVTestTopState *tts;
1047     TestCoDeleteByDrainData dbdd;
1048     Coroutine *co;
1049 
1050     bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR,
1051                               &error_abort);
1052     bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1053     tts = bs->opaque;
1054 
1055     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1056                         &error_abort);
1057     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds,
1058                       BDRV_CHILD_DATA, &error_abort);
1059 
1060     /* This child will be the one to pass to requests through to, and
1061      * it will stall until a drain occurs */
1062     child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR,
1063                                     &error_abort);
1064     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1065     /* Takes our reference to child_bs */
1066     tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child",
1067                                         &child_of_bds,
1068                                         BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY,
1069                                         &error_abort);
1070 
1071     /* This child is just there to be deleted
1072      * (for detach_instead_of_delete == true) */
1073     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1074                         &error_abort);
1075     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA,
1076                       &error_abort);
1077 
1078     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1079     blk_insert_bs(blk, bs, &error_abort);
1080 
1081     /* Referenced by blk now */
1082     bdrv_unref(bs);
1083 
1084     g_assert_cmpint(bs->refcnt, ==, 1);
1085     g_assert_cmpint(child_bs->refcnt, ==, 1);
1086     g_assert_cmpint(null_bs->refcnt, ==, 1);
1087 
1088 
1089     dbdd = (TestCoDeleteByDrainData){
1090         .blk = blk,
1091         .detach_instead_of_delete = detach_instead_of_delete,
1092         .done = false,
1093     };
1094     co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd);
1095     qemu_coroutine_enter(co);
1096 
1097     /* Drain the child while the read operation is still pending.
1098      * This should result in the operation finishing and
1099      * test_co_delete_by_drain() resuming.  Thus, @bs will be deleted
1100      * and the coroutine will exit while this drain operation is still
1101      * in progress. */
1102     switch (drain_type) {
1103     case BDRV_DRAIN:
1104         bdrv_ref(child_bs);
1105         bdrv_drain(child_bs);
1106         bdrv_unref(child_bs);
1107         break;
1108     case BDRV_DRAIN_ALL:
1109         bdrv_drain_all_begin();
1110         bdrv_drain_all_end();
1111         break;
1112     default:
1113         g_assert_not_reached();
1114     }
1115 
1116     while (!dbdd.done) {
1117         aio_poll(qemu_get_aio_context(), true);
1118     }
1119 
1120     if (detach_instead_of_delete) {
1121         /* Here, the reference has not passed over to the coroutine,
1122          * so we have to delete the BB ourselves */
1123         blk_unref(blk);
1124     }
1125 }
1126 
1127 static void test_delete_by_drain(void)
1128 {
1129     do_test_delete_by_drain(false, BDRV_DRAIN);
1130 }
1131 
1132 static void test_detach_by_drain_all(void)
1133 {
1134     do_test_delete_by_drain(true, BDRV_DRAIN_ALL);
1135 }
1136 
1137 static void test_detach_by_drain(void)
1138 {
1139     do_test_delete_by_drain(true, BDRV_DRAIN);
1140 }
1141 
1142 
1143 struct detach_by_parent_data {
1144     BlockDriverState *parent_b;
1145     BdrvChild *child_b;
1146     BlockDriverState *c;
1147     BdrvChild *child_c;
1148     bool by_parent_cb;
1149     bool detach_on_drain;
1150 };
1151 static struct detach_by_parent_data detach_by_parent_data;
1152 
1153 static void detach_indirect_bh(void *opaque)
1154 {
1155     struct detach_by_parent_data *data = opaque;
1156 
1157     bdrv_dec_in_flight(data->child_b->bs);
1158     bdrv_unref_child(data->parent_b, data->child_b);
1159 
1160     bdrv_ref(data->c);
1161     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
1162                                       &child_of_bds, BDRV_CHILD_DATA,
1163                                       &error_abort);
1164 }
1165 
1166 static void 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         detach_indirect_bh(data);
1174     }
1175 }
1176 
1177 static void detach_by_driver_cb_drained_begin(BdrvChild *child)
1178 {
1179     struct detach_by_parent_data *data = &detach_by_parent_data;
1180 
1181     if (!data->detach_on_drain) {
1182         return;
1183     }
1184     data->detach_on_drain = false;
1185 
1186     bdrv_inc_in_flight(data->child_b->bs);
1187     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1188                             detach_indirect_bh, &detach_by_parent_data);
1189     child_of_bds.drained_begin(child);
1190 }
1191 
1192 static BdrvChildClass detach_by_driver_cb_class;
1193 
1194 /*
1195  * Initial graph:
1196  *
1197  * PA     PB
1198  *    \ /   \
1199  *     A     B     C
1200  *
1201  * by_parent_cb == true:  Test that parent callbacks don't poll
1202  *
1203  *     PA has a pending write request whose callback changes the child nodes of
1204  *     PB: It removes B and adds C instead. The subtree of PB is drained, which
1205  *     will indirectly drain the write request, too.
1206  *
1207  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1208  *
1209  *     PA's BdrvChildClass has a .drained_begin callback that schedules a BH
1210  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
1211  *     state is messed up, but if it is only polled in the single
1212  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1213  */
1214 static void test_detach_indirect(bool by_parent_cb)
1215 {
1216     BlockBackend *blk;
1217     BlockDriverState *parent_a, *parent_b, *a, *b, *c;
1218     BdrvChild *child_a, *child_b;
1219     BlockAIOCB *acb;
1220 
1221     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
1222 
1223     if (!by_parent_cb) {
1224         detach_by_driver_cb_class = child_of_bds;
1225         detach_by_driver_cb_class.drained_begin =
1226             detach_by_driver_cb_drained_begin;
1227         detach_by_driver_cb_class.drained_end = NULL;
1228         detach_by_driver_cb_class.drained_poll = NULL;
1229     }
1230 
1231     detach_by_parent_data = (struct detach_by_parent_data) {
1232         .detach_on_drain = false,
1233     };
1234 
1235     /* Create all involved nodes */
1236     parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR,
1237                                     &error_abort);
1238     parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0,
1239                                     &error_abort);
1240 
1241     a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort);
1242     b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort);
1243     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
1244 
1245     /* blk is a BB for parent-a */
1246     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1247     blk_insert_bs(blk, parent_a, &error_abort);
1248     bdrv_unref(parent_a);
1249 
1250     /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1251      * callback must not return immediately. */
1252     if (!by_parent_cb) {
1253         BDRVTestState *s = parent_a->opaque;
1254         s->sleep_in_drain_begin = true;
1255     }
1256 
1257     /* Set child relationships */
1258     bdrv_ref(b);
1259     bdrv_ref(a);
1260     child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds,
1261                                 BDRV_CHILD_DATA, &error_abort);
1262     child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds,
1263                                 BDRV_CHILD_COW, &error_abort);
1264 
1265     bdrv_ref(a);
1266     bdrv_attach_child(parent_a, a, "PA-A",
1267                       by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class,
1268                       BDRV_CHILD_DATA, &error_abort);
1269 
1270     g_assert_cmpint(parent_a->refcnt, ==, 1);
1271     g_assert_cmpint(parent_b->refcnt, ==, 1);
1272     g_assert_cmpint(a->refcnt, ==, 3);
1273     g_assert_cmpint(b->refcnt, ==, 2);
1274     g_assert_cmpint(c->refcnt, ==, 1);
1275 
1276     g_assert(QLIST_FIRST(&parent_b->children) == child_a);
1277     g_assert(QLIST_NEXT(child_a, next) == child_b);
1278     g_assert(QLIST_NEXT(child_b, next) == NULL);
1279 
1280     /* Start the evil write request */
1281     detach_by_parent_data = (struct detach_by_parent_data) {
1282         .parent_b = parent_b,
1283         .child_b = child_b,
1284         .c = c,
1285         .by_parent_cb = by_parent_cb,
1286         .detach_on_drain = true,
1287     };
1288     acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL);
1289     g_assert(acb != NULL);
1290 
1291     /* Drain and check the expected result */
1292     bdrv_drained_begin(parent_b);
1293     bdrv_drained_begin(a);
1294     bdrv_drained_begin(b);
1295     bdrv_drained_begin(c);
1296 
1297     g_assert(detach_by_parent_data.child_c != NULL);
1298 
1299     g_assert_cmpint(parent_a->refcnt, ==, 1);
1300     g_assert_cmpint(parent_b->refcnt, ==, 1);
1301     g_assert_cmpint(a->refcnt, ==, 3);
1302     g_assert_cmpint(b->refcnt, ==, 1);
1303     g_assert_cmpint(c->refcnt, ==, 2);
1304 
1305     g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c);
1306     g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a);
1307     g_assert(QLIST_NEXT(child_a, next) == NULL);
1308 
1309     g_assert_cmpint(parent_a->quiesce_counter, ==, 1);
1310     g_assert_cmpint(parent_b->quiesce_counter, ==, 3);
1311     g_assert_cmpint(a->quiesce_counter, ==, 1);
1312     g_assert_cmpint(b->quiesce_counter, ==, 1);
1313     g_assert_cmpint(c->quiesce_counter, ==, 1);
1314 
1315     bdrv_drained_end(parent_b);
1316     bdrv_drained_end(a);
1317     bdrv_drained_end(b);
1318     bdrv_drained_end(c);
1319 
1320     bdrv_unref(parent_b);
1321     blk_unref(blk);
1322 
1323     g_assert_cmpint(a->refcnt, ==, 1);
1324     g_assert_cmpint(b->refcnt, ==, 1);
1325     g_assert_cmpint(c->refcnt, ==, 1);
1326     bdrv_unref(a);
1327     bdrv_unref(b);
1328     bdrv_unref(c);
1329 }
1330 
1331 static void test_detach_by_parent_cb(void)
1332 {
1333     test_detach_indirect(true);
1334 }
1335 
1336 static void test_detach_by_driver_cb(void)
1337 {
1338     test_detach_indirect(false);
1339 }
1340 
1341 static void test_append_to_drained(void)
1342 {
1343     BlockBackend *blk;
1344     BlockDriverState *base, *overlay;
1345     BDRVTestState *base_s, *overlay_s;
1346 
1347     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1348     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
1349     base_s = base->opaque;
1350     blk_insert_bs(blk, base, &error_abort);
1351 
1352     overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR,
1353                                    &error_abort);
1354     overlay_s = overlay->opaque;
1355 
1356     do_drain_begin(BDRV_DRAIN, base);
1357     g_assert_cmpint(base->quiesce_counter, ==, 1);
1358     g_assert_cmpint(base_s->drain_count, ==, 1);
1359     g_assert_cmpint(base->in_flight, ==, 0);
1360 
1361     bdrv_append(overlay, base, &error_abort);
1362     g_assert_cmpint(base->in_flight, ==, 0);
1363     g_assert_cmpint(overlay->in_flight, ==, 0);
1364 
1365     g_assert_cmpint(base->quiesce_counter, ==, 1);
1366     g_assert_cmpint(base_s->drain_count, ==, 1);
1367     g_assert_cmpint(overlay->quiesce_counter, ==, 1);
1368     g_assert_cmpint(overlay_s->drain_count, ==, 1);
1369 
1370     do_drain_end(BDRV_DRAIN, base);
1371 
1372     g_assert_cmpint(base->quiesce_counter, ==, 0);
1373     g_assert_cmpint(base_s->drain_count, ==, 0);
1374     g_assert_cmpint(overlay->quiesce_counter, ==, 0);
1375     g_assert_cmpint(overlay_s->drain_count, ==, 0);
1376 
1377     bdrv_unref(overlay);
1378     bdrv_unref(base);
1379     blk_unref(blk);
1380 }
1381 
1382 static void test_set_aio_context(void)
1383 {
1384     BlockDriverState *bs;
1385     IOThread *a = iothread_new();
1386     IOThread *b = iothread_new();
1387     AioContext *ctx_a = iothread_get_aio_context(a);
1388     AioContext *ctx_b = iothread_get_aio_context(b);
1389 
1390     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
1391                               &error_abort);
1392 
1393     bdrv_drained_begin(bs);
1394     bdrv_try_change_aio_context(bs, ctx_a, NULL, &error_abort);
1395 
1396     aio_context_acquire(ctx_a);
1397     bdrv_drained_end(bs);
1398 
1399     bdrv_drained_begin(bs);
1400     bdrv_try_change_aio_context(bs, ctx_b, NULL, &error_abort);
1401     aio_context_release(ctx_a);
1402     aio_context_acquire(ctx_b);
1403     bdrv_try_change_aio_context(bs, qemu_get_aio_context(), NULL, &error_abort);
1404     aio_context_release(ctx_b);
1405     bdrv_drained_end(bs);
1406 
1407     bdrv_unref(bs);
1408     iothread_join(a);
1409     iothread_join(b);
1410 }
1411 
1412 
1413 typedef struct TestDropBackingBlockJob {
1414     BlockJob common;
1415     bool should_complete;
1416     bool *did_complete;
1417     BlockDriverState *detach_also;
1418     BlockDriverState *bs;
1419 } TestDropBackingBlockJob;
1420 
1421 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp)
1422 {
1423     TestDropBackingBlockJob *s =
1424         container_of(job, TestDropBackingBlockJob, common.job);
1425 
1426     while (!s->should_complete) {
1427         job_sleep_ns(job, 0);
1428     }
1429 
1430     return 0;
1431 }
1432 
1433 static void test_drop_backing_job_commit(Job *job)
1434 {
1435     TestDropBackingBlockJob *s =
1436         container_of(job, TestDropBackingBlockJob, common.job);
1437 
1438     bdrv_set_backing_hd(s->bs, NULL, &error_abort);
1439     bdrv_set_backing_hd(s->detach_also, NULL, &error_abort);
1440 
1441     *s->did_complete = true;
1442 }
1443 
1444 static const BlockJobDriver test_drop_backing_job_driver = {
1445     .job_driver = {
1446         .instance_size  = sizeof(TestDropBackingBlockJob),
1447         .free           = block_job_free,
1448         .user_resume    = block_job_user_resume,
1449         .run            = test_drop_backing_job_run,
1450         .commit         = test_drop_backing_job_commit,
1451     }
1452 };
1453 
1454 /**
1455  * Creates a child node with three parent nodes on it, and then runs a
1456  * block job on the final one, parent-node-2.
1457  *
1458  * The job is then asked to complete before a section where the child
1459  * is drained.
1460  *
1461  * Ending this section will undrain the child's parents, first
1462  * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
1463  * list is in reverse order of how they were added.  Ending the drain
1464  * on parent-node-2 will resume the job, thus completing it and
1465  * scheduling job_exit().
1466  *
1467  * Ending the drain on parent-node-1 will poll the AioContext, which
1468  * lets job_exit() and thus test_drop_backing_job_commit() run.  That
1469  * function first removes the child as parent-node-2's backing file.
1470  *
1471  * In old (and buggy) implementations, there are two problems with
1472  * that:
1473  * (A) bdrv_drain_invoke() polls for every node that leaves the
1474  *     drained section.  This means that job_exit() is scheduled
1475  *     before the child has left the drained section.  Its
1476  *     quiesce_counter is therefore still 1 when it is removed from
1477  *     parent-node-2.
1478  *
1479  * (B) bdrv_replace_child_noperm() calls drained_end() on the old
1480  *     child's parents as many times as the child is quiesced.  This
1481  *     means it will call drained_end() on parent-node-2 once.
1482  *     Because parent-node-2 is no longer quiesced at this point, this
1483  *     will fail.
1484  *
1485  * bdrv_replace_child_noperm() therefore must call drained_end() on
1486  * the parent only if it really is still drained because the child is
1487  * drained.
1488  *
1489  * If removing child from parent-node-2 was successful (as it should
1490  * be), test_drop_backing_job_commit() will then also remove the child
1491  * from parent-node-0.
1492  *
1493  * With an old version of our drain infrastructure ((A) above), that
1494  * resulted in the following flow:
1495  *
1496  * 1. child attempts to leave its drained section.  The call recurses
1497  *    to its parents.
1498  *
1499  * 2. parent-node-2 leaves the drained section.  Polling in
1500  *    bdrv_drain_invoke() will schedule job_exit().
1501  *
1502  * 3. parent-node-1 leaves the drained section.  Polling in
1503  *    bdrv_drain_invoke() will run job_exit(), thus disconnecting
1504  *    parent-node-0 from the child node.
1505  *
1506  * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
1507  *    iterate over the parents.  Thus, it now accesses the BdrvChild
1508  *    object that used to connect parent-node-0 and the child node.
1509  *    However, that object no longer exists, so it accesses a dangling
1510  *    pointer.
1511  *
1512  * The solution is to only poll once when running a bdrv_drained_end()
1513  * operation, specifically at the end when all drained_end()
1514  * operations for all involved nodes have been scheduled.
1515  * Note that this also solves (A) above, thus hiding (B).
1516  */
1517 static void test_blockjob_commit_by_drained_end(void)
1518 {
1519     BlockDriverState *bs_child, *bs_parents[3];
1520     TestDropBackingBlockJob *job;
1521     bool job_has_completed = false;
1522     int i;
1523 
1524     bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR,
1525                                     &error_abort);
1526 
1527     for (i = 0; i < 3; i++) {
1528         char name[32];
1529         snprintf(name, sizeof(name), "parent-node-%i", i);
1530         bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR,
1531                                              &error_abort);
1532         bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort);
1533     }
1534 
1535     job = block_job_create("job", &test_drop_backing_job_driver, NULL,
1536                            bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL,
1537                            &error_abort);
1538     job->bs = bs_parents[2];
1539 
1540     job->detach_also = bs_parents[0];
1541     job->did_complete = &job_has_completed;
1542 
1543     job_start(&job->common.job);
1544 
1545     job->should_complete = true;
1546     bdrv_drained_begin(bs_child);
1547     g_assert(!job_has_completed);
1548     bdrv_drained_end(bs_child);
1549     aio_poll(qemu_get_aio_context(), false);
1550     g_assert(job_has_completed);
1551 
1552     bdrv_unref(bs_parents[0]);
1553     bdrv_unref(bs_parents[1]);
1554     bdrv_unref(bs_parents[2]);
1555     bdrv_unref(bs_child);
1556 }
1557 
1558 
1559 typedef struct TestSimpleBlockJob {
1560     BlockJob common;
1561     bool should_complete;
1562     bool *did_complete;
1563 } TestSimpleBlockJob;
1564 
1565 static int coroutine_fn test_simple_job_run(Job *job, Error **errp)
1566 {
1567     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1568 
1569     while (!s->should_complete) {
1570         job_sleep_ns(job, 0);
1571     }
1572 
1573     return 0;
1574 }
1575 
1576 static void test_simple_job_clean(Job *job)
1577 {
1578     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1579     *s->did_complete = true;
1580 }
1581 
1582 static const BlockJobDriver test_simple_job_driver = {
1583     .job_driver = {
1584         .instance_size  = sizeof(TestSimpleBlockJob),
1585         .free           = block_job_free,
1586         .user_resume    = block_job_user_resume,
1587         .run            = test_simple_job_run,
1588         .clean          = test_simple_job_clean,
1589     },
1590 };
1591 
1592 static int drop_intermediate_poll_update_filename(BdrvChild *child,
1593                                                   BlockDriverState *new_base,
1594                                                   const char *filename,
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     for (i = 0; i < 3; i++) {
1685         if (i) {
1686             /* Takes the reference to chain[i - 1] */
1687             bdrv_attach_child(chain[i], chain[i - 1], "chain",
1688                               &chain_child_class, BDRV_CHILD_COW, &error_abort);
1689         }
1690     }
1691 
1692     job = block_job_create("job", &test_simple_job_driver, NULL, job_node,
1693                            0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort);
1694 
1695     /* The job has a reference now */
1696     bdrv_unref(job_node);
1697 
1698     job->did_complete = &job_has_completed;
1699 
1700     job_start(&job->common.job);
1701     job->should_complete = true;
1702 
1703     g_assert(!job_has_completed);
1704     ret = bdrv_drop_intermediate(chain[1], chain[0], NULL);
1705     aio_poll(qemu_get_aio_context(), false);
1706     g_assert(ret == 0);
1707     g_assert(job_has_completed);
1708 
1709     bdrv_unref(chain[2]);
1710 }
1711 
1712 
1713 typedef struct BDRVReplaceTestState {
1714     bool setup_completed;
1715     bool was_drained;
1716     bool was_undrained;
1717     bool has_read;
1718 
1719     int drain_count;
1720 
1721     bool yield_before_read;
1722     Coroutine *io_co;
1723     Coroutine *drain_co;
1724 } BDRVReplaceTestState;
1725 
1726 static void bdrv_replace_test_close(BlockDriverState *bs)
1727 {
1728 }
1729 
1730 /**
1731  * If @bs has a backing file:
1732  *   Yield if .yield_before_read is true (and wait for drain_begin to
1733  *   wake us up).
1734  *   Forward the read to bs->backing.  Set .has_read to true.
1735  *   If drain_begin has woken us, wake it in turn.
1736  *
1737  * Otherwise:
1738  *   Set .has_read to true and return success.
1739  */
1740 static int coroutine_fn GRAPH_RDLOCK
1741 bdrv_replace_test_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
1742                             QEMUIOVector *qiov, BdrvRequestFlags flags)
1743 {
1744     BDRVReplaceTestState *s = bs->opaque;
1745 
1746     if (bs->backing) {
1747         int ret;
1748 
1749         g_assert(!s->drain_count);
1750 
1751         s->io_co = qemu_coroutine_self();
1752         if (s->yield_before_read) {
1753             s->yield_before_read = false;
1754             qemu_coroutine_yield();
1755         }
1756         s->io_co = NULL;
1757 
1758         ret = bdrv_co_preadv(bs->backing, offset, bytes, qiov, 0);
1759         s->has_read = true;
1760 
1761         /* Wake up drain_co if it runs */
1762         if (s->drain_co) {
1763             aio_co_wake(s->drain_co);
1764         }
1765 
1766         return ret;
1767     }
1768 
1769     s->has_read = true;
1770     return 0;
1771 }
1772 
1773 static void coroutine_fn bdrv_replace_test_drain_co(void *opaque)
1774 {
1775     BlockDriverState *bs = opaque;
1776     BDRVReplaceTestState *s = bs->opaque;
1777 
1778     /* Keep waking io_co up until it is done */
1779     while (s->io_co) {
1780         aio_co_wake(s->io_co);
1781         s->io_co = NULL;
1782         qemu_coroutine_yield();
1783     }
1784     s->drain_co = NULL;
1785     bdrv_dec_in_flight(bs);
1786 }
1787 
1788 /**
1789  * If .drain_count is 0, wake up .io_co if there is one; and set
1790  * .was_drained.
1791  * Increment .drain_count.
1792  */
1793 static void bdrv_replace_test_drain_begin(BlockDriverState *bs)
1794 {
1795     BDRVReplaceTestState *s = bs->opaque;
1796 
1797     if (!s->setup_completed) {
1798         return;
1799     }
1800 
1801     if (!s->drain_count) {
1802         s->drain_co = qemu_coroutine_create(bdrv_replace_test_drain_co, bs);
1803         bdrv_inc_in_flight(bs);
1804         aio_co_enter(bdrv_get_aio_context(bs), s->drain_co);
1805         s->was_drained = true;
1806     }
1807     s->drain_count++;
1808 }
1809 
1810 static void coroutine_fn bdrv_replace_test_read_entry(void *opaque)
1811 {
1812     BlockDriverState *bs = opaque;
1813     char data;
1814     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
1815     int ret;
1816 
1817     /* Queue a read request post-drain */
1818     bdrv_graph_co_rdlock();
1819     ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
1820     bdrv_graph_co_rdunlock();
1821 
1822     g_assert(ret >= 0);
1823     bdrv_dec_in_flight(bs);
1824 }
1825 
1826 /**
1827  * Reduce .drain_count, set .was_undrained once it reaches 0.
1828  * If .drain_count reaches 0 and the node has a backing file, issue a
1829  * read request.
1830  */
1831 static void bdrv_replace_test_drain_end(BlockDriverState *bs)
1832 {
1833     BDRVReplaceTestState *s = bs->opaque;
1834 
1835     if (!s->setup_completed) {
1836         return;
1837     }
1838 
1839     g_assert(s->drain_count > 0);
1840     if (!--s->drain_count) {
1841         s->was_undrained = true;
1842 
1843         if (bs->backing) {
1844             Coroutine *co = qemu_coroutine_create(bdrv_replace_test_read_entry,
1845                                                   bs);
1846             bdrv_inc_in_flight(bs);
1847             aio_co_enter(bdrv_get_aio_context(bs), co);
1848         }
1849     }
1850 }
1851 
1852 static BlockDriver bdrv_replace_test = {
1853     .format_name            = "replace_test",
1854     .instance_size          = sizeof(BDRVReplaceTestState),
1855     .supports_backing       = true,
1856 
1857     .bdrv_close             = bdrv_replace_test_close,
1858     .bdrv_co_preadv         = bdrv_replace_test_co_preadv,
1859 
1860     .bdrv_drain_begin       = bdrv_replace_test_drain_begin,
1861     .bdrv_drain_end         = bdrv_replace_test_drain_end,
1862 
1863     .bdrv_child_perm        = bdrv_default_perms,
1864 };
1865 
1866 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque)
1867 {
1868     int ret;
1869     char data;
1870 
1871     ret = blk_co_pread(opaque, 0, 1, &data, 0);
1872     g_assert(ret >= 0);
1873 }
1874 
1875 /**
1876  * We test two things:
1877  * (1) bdrv_replace_child_noperm() must not undrain the parent if both
1878  *     children are drained.
1879  * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
1880  *     drained child.  If the old child is drained, it must flush I/O
1881  *     requests after the new one has been attached.  If the new child
1882  *     is drained, it must flush I/O requests before the old one is
1883  *     detached.
1884  *
1885  * To do so, we create one parent node and two child nodes; then
1886  * attach one of the children (old_child_bs) to the parent, then
1887  * drain both old_child_bs and new_child_bs according to
1888  * old_drain_count and new_drain_count, respectively, and finally
1889  * we invoke bdrv_replace_node() to replace old_child_bs by
1890  * new_child_bs.
1891  *
1892  * The test block driver we use here (bdrv_replace_test) has a read
1893  * function that:
1894  * - For the parent node, can optionally yield, and then forwards the
1895  *   read to bdrv_preadv(),
1896  * - For the child node, just returns immediately.
1897  *
1898  * If the read yields, the drain_begin function will wake it up.
1899  *
1900  * The drain_end function issues a read on the parent once it is fully
1901  * undrained (which simulates requests starting to come in again).
1902  */
1903 static void do_test_replace_child_mid_drain(int old_drain_count,
1904                                             int new_drain_count)
1905 {
1906     BlockBackend *parent_blk;
1907     BlockDriverState *parent_bs;
1908     BlockDriverState *old_child_bs, *new_child_bs;
1909     BDRVReplaceTestState *parent_s;
1910     BDRVReplaceTestState *old_child_s, *new_child_s;
1911     Coroutine *io_co;
1912     int i;
1913 
1914     parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0,
1915                                      &error_abort);
1916     parent_s = parent_bs->opaque;
1917 
1918     parent_blk = blk_new(qemu_get_aio_context(),
1919                          BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
1920     blk_insert_bs(parent_blk, parent_bs, &error_abort);
1921 
1922     old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0,
1923                                         &error_abort);
1924     new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0,
1925                                         &error_abort);
1926     old_child_s = old_child_bs->opaque;
1927     new_child_s = new_child_bs->opaque;
1928 
1929     /* So that we can read something */
1930     parent_bs->total_sectors = 1;
1931     old_child_bs->total_sectors = 1;
1932     new_child_bs->total_sectors = 1;
1933 
1934     bdrv_ref(old_child_bs);
1935     bdrv_attach_child(parent_bs, old_child_bs, "child", &child_of_bds,
1936                       BDRV_CHILD_COW, &error_abort);
1937     parent_s->setup_completed = true;
1938 
1939     for (i = 0; i < old_drain_count; i++) {
1940         bdrv_drained_begin(old_child_bs);
1941     }
1942     for (i = 0; i < new_drain_count; i++) {
1943         bdrv_drained_begin(new_child_bs);
1944     }
1945 
1946     if (!old_drain_count) {
1947         /*
1948          * Start a read operation that will yield, so it will not
1949          * complete before the node is drained.
1950          */
1951         parent_s->yield_before_read = true;
1952         io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co,
1953                                       parent_blk);
1954         qemu_coroutine_enter(io_co);
1955     }
1956 
1957     /* If we have started a read operation, it should have yielded */
1958     g_assert(!parent_s->has_read);
1959 
1960     /* Reset drained status so we can see what bdrv_replace_node() does */
1961     parent_s->was_drained = false;
1962     parent_s->was_undrained = false;
1963 
1964     g_assert(parent_bs->quiesce_counter == old_drain_count);
1965     bdrv_replace_node(old_child_bs, new_child_bs, &error_abort);
1966     g_assert(parent_bs->quiesce_counter == new_drain_count);
1967 
1968     if (!old_drain_count && !new_drain_count) {
1969         /*
1970          * From undrained to undrained drains and undrains the parent,
1971          * because bdrv_replace_node() contains a drained section for
1972          * @old_child_bs.
1973          */
1974         g_assert(parent_s->was_drained && parent_s->was_undrained);
1975     } else if (!old_drain_count && new_drain_count) {
1976         /*
1977          * From undrained to drained should drain the parent and keep
1978          * it that way.
1979          */
1980         g_assert(parent_s->was_drained && !parent_s->was_undrained);
1981     } else if (old_drain_count && !new_drain_count) {
1982         /*
1983          * From drained to undrained should undrain the parent and
1984          * keep it that way.
1985          */
1986         g_assert(!parent_s->was_drained && parent_s->was_undrained);
1987     } else /* if (old_drain_count && new_drain_count) */ {
1988         /*
1989          * From drained to drained must not undrain the parent at any
1990          * point
1991          */
1992         g_assert(!parent_s->was_drained && !parent_s->was_undrained);
1993     }
1994 
1995     if (!old_drain_count || !new_drain_count) {
1996         /*
1997          * If !old_drain_count, we have started a read request before
1998          * bdrv_replace_node().  If !new_drain_count, the parent must
1999          * have been undrained at some point, and
2000          * bdrv_replace_test_co_drain_end() starts a read request
2001          * then.
2002          */
2003         g_assert(parent_s->has_read);
2004     } else {
2005         /*
2006          * If the parent was never undrained, there is no way to start
2007          * a read request.
2008          */
2009         g_assert(!parent_s->has_read);
2010     }
2011 
2012     /* A drained child must have not received any request */
2013     g_assert(!(old_drain_count && old_child_s->has_read));
2014     g_assert(!(new_drain_count && new_child_s->has_read));
2015 
2016     for (i = 0; i < new_drain_count; i++) {
2017         bdrv_drained_end(new_child_bs);
2018     }
2019     for (i = 0; i < old_drain_count; i++) {
2020         bdrv_drained_end(old_child_bs);
2021     }
2022 
2023     /*
2024      * By now, bdrv_replace_test_co_drain_end() must have been called
2025      * at some point while the new child was attached to the parent.
2026      */
2027     g_assert(parent_s->has_read);
2028     g_assert(new_child_s->has_read);
2029 
2030     blk_unref(parent_blk);
2031     bdrv_unref(parent_bs);
2032     bdrv_unref(old_child_bs);
2033     bdrv_unref(new_child_bs);
2034 }
2035 
2036 static void test_replace_child_mid_drain(void)
2037 {
2038     int old_drain_count, new_drain_count;
2039 
2040     for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) {
2041         for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) {
2042             do_test_replace_child_mid_drain(old_drain_count, new_drain_count);
2043         }
2044     }
2045 }
2046 
2047 int main(int argc, char **argv)
2048 {
2049     int ret;
2050 
2051     bdrv_init();
2052     qemu_init_main_loop(&error_abort);
2053 
2054     g_test_init(&argc, &argv, NULL);
2055     qemu_event_init(&done_event, false);
2056 
2057     g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all);
2058     g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain);
2059 
2060     g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
2061                     test_drv_cb_co_drain_all);
2062     g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain);
2063 
2064     g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all);
2065     g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain);
2066 
2067     g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
2068                     test_quiesce_co_drain_all);
2069     g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain);
2070 
2071     g_test_add_func("/bdrv-drain/nested", test_nested);
2072 
2073     g_test_add_func("/bdrv-drain/graph-change/drain_all",
2074                     test_graph_change_drain_all);
2075 
2076     g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all);
2077     g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain);
2078 
2079     g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all);
2080     g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain);
2081 
2082     g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
2083                     test_blockjob_error_drain_all);
2084     g_test_add_func("/bdrv-drain/blockjob/error/drain",
2085                     test_blockjob_error_drain);
2086 
2087     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
2088                     test_blockjob_iothread_drain_all);
2089     g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
2090                     test_blockjob_iothread_drain);
2091 
2092     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
2093                     test_blockjob_iothread_error_drain_all);
2094     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
2095                     test_blockjob_iothread_error_drain);
2096 
2097     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
2098     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
2099     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
2100     g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb);
2101     g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb);
2102 
2103     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
2104 
2105     g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context);
2106 
2107     g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
2108                     test_blockjob_commit_by_drained_end);
2109 
2110     g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
2111                     test_drop_intermediate_poll);
2112 
2113     g_test_add_func("/bdrv-drain/replace_child/mid-drain",
2114                     test_replace_child_mid_drain);
2115 
2116     ret = g_test_run();
2117     qemu_event_destroy(&done_event);
2118     return ret;
2119 }
2120