xref: /openbmc/qemu/tests/unit/test-bdrv-drain.c (revision a6caeee8)
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.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 bdrv_test_co_drain_begin(BlockDriverState *bs)
42 {
43     BDRVTestState *s = bs->opaque;
44     s->drain_count++;
45     if (s->sleep_in_drain_begin) {
46         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
47     }
48 }
49 
50 static void coroutine_fn bdrv_test_co_drain_end(BlockDriverState *bs)
51 {
52     BDRVTestState *s = bs->opaque;
53     s->drain_count--;
54 }
55 
56 static void bdrv_test_close(BlockDriverState *bs)
57 {
58     BDRVTestState *s = bs->opaque;
59     g_assert_cmpint(s->drain_count, >, 0);
60 }
61 
62 static void co_reenter_bh(void *opaque)
63 {
64     aio_co_wake(opaque);
65 }
66 
67 static int coroutine_fn bdrv_test_co_preadv(BlockDriverState *bs,
68                                             int64_t offset, int64_t bytes,
69                                             QEMUIOVector *qiov,
70                                             BdrvRequestFlags flags)
71 {
72     BDRVTestState *s = bs->opaque;
73 
74     /* We want this request to stay until the polling loop in drain waits for
75      * it to complete. We need to sleep a while as bdrv_drain_invoke() comes
76      * first and polls its result, too, but it shouldn't accidentally complete
77      * this request yet. */
78     qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
79 
80     if (s->bh_indirection_ctx) {
81         aio_bh_schedule_oneshot(s->bh_indirection_ctx, co_reenter_bh,
82                                 qemu_coroutine_self());
83         qemu_coroutine_yield();
84     }
85 
86     return 0;
87 }
88 
89 static int bdrv_test_change_backing_file(BlockDriverState *bs,
90                                          const char *backing_file,
91                                          const char *backing_fmt)
92 {
93     return 0;
94 }
95 
96 static BlockDriver bdrv_test = {
97     .format_name            = "test",
98     .instance_size          = sizeof(BDRVTestState),
99     .supports_backing       = true,
100 
101     .bdrv_close             = bdrv_test_close,
102     .bdrv_co_preadv         = bdrv_test_co_preadv,
103 
104     .bdrv_co_drain_begin    = bdrv_test_co_drain_begin,
105     .bdrv_co_drain_end      = bdrv_test_co_drain_end,
106 
107     .bdrv_child_perm        = bdrv_default_perms,
108 
109     .bdrv_change_backing_file = bdrv_test_change_backing_file,
110 };
111 
112 static void aio_ret_cb(void *opaque, int ret)
113 {
114     int *aio_ret = opaque;
115     *aio_ret = ret;
116 }
117 
118 typedef struct CallInCoroutineData {
119     void (*entry)(void);
120     bool done;
121 } CallInCoroutineData;
122 
123 static coroutine_fn void call_in_coroutine_entry(void *opaque)
124 {
125     CallInCoroutineData *data = opaque;
126 
127     data->entry();
128     data->done = true;
129 }
130 
131 static void call_in_coroutine(void (*entry)(void))
132 {
133     Coroutine *co;
134     CallInCoroutineData data = {
135         .entry  = entry,
136         .done   = false,
137     };
138 
139     co = qemu_coroutine_create(call_in_coroutine_entry, &data);
140     qemu_coroutine_enter(co);
141     while (!data.done) {
142         aio_poll(qemu_get_aio_context(), true);
143     }
144 }
145 
146 enum drain_type {
147     BDRV_DRAIN_ALL,
148     BDRV_DRAIN,
149     BDRV_SUBTREE_DRAIN,
150     DRAIN_TYPE_MAX,
151 };
152 
153 static void do_drain_begin(enum drain_type drain_type, BlockDriverState *bs)
154 {
155     switch (drain_type) {
156     case BDRV_DRAIN_ALL:        bdrv_drain_all_begin(); break;
157     case BDRV_DRAIN:            bdrv_drained_begin(bs); break;
158     case BDRV_SUBTREE_DRAIN:    bdrv_subtree_drained_begin(bs); break;
159     default:                    g_assert_not_reached();
160     }
161 }
162 
163 static void do_drain_end(enum drain_type drain_type, BlockDriverState *bs)
164 {
165     switch (drain_type) {
166     case BDRV_DRAIN_ALL:        bdrv_drain_all_end(); break;
167     case BDRV_DRAIN:            bdrv_drained_end(bs); break;
168     case BDRV_SUBTREE_DRAIN:    bdrv_subtree_drained_end(bs); break;
169     default:                    g_assert_not_reached();
170     }
171 }
172 
173 static void do_drain_begin_unlocked(enum drain_type drain_type, BlockDriverState *bs)
174 {
175     if (drain_type != BDRV_DRAIN_ALL) {
176         aio_context_acquire(bdrv_get_aio_context(bs));
177     }
178     do_drain_begin(drain_type, bs);
179     if (drain_type != BDRV_DRAIN_ALL) {
180         aio_context_release(bdrv_get_aio_context(bs));
181     }
182 }
183 
184 static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs)
185 {
186     if (drain_type != BDRV_DRAIN_ALL) {
187         aio_context_acquire(bdrv_get_aio_context(bs));
188     }
189     do_drain_end(drain_type, bs);
190     if (drain_type != BDRV_DRAIN_ALL) {
191         aio_context_release(bdrv_get_aio_context(bs));
192     }
193 }
194 
195 static void test_drv_cb_common(enum drain_type drain_type, bool recursive)
196 {
197     BlockBackend *blk;
198     BlockDriverState *bs, *backing;
199     BDRVTestState *s, *backing_s;
200     BlockAIOCB *acb;
201     int aio_ret;
202 
203     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
204 
205     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
206     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
207                               &error_abort);
208     s = bs->opaque;
209     blk_insert_bs(blk, bs, &error_abort);
210 
211     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
212     backing_s = backing->opaque;
213     bdrv_set_backing_hd(bs, backing, &error_abort);
214 
215     /* Simple bdrv_drain_all_begin/end pair, check that CBs are called */
216     g_assert_cmpint(s->drain_count, ==, 0);
217     g_assert_cmpint(backing_s->drain_count, ==, 0);
218 
219     do_drain_begin(drain_type, bs);
220 
221     g_assert_cmpint(s->drain_count, ==, 1);
222     g_assert_cmpint(backing_s->drain_count, ==, !!recursive);
223 
224     do_drain_end(drain_type, bs);
225 
226     g_assert_cmpint(s->drain_count, ==, 0);
227     g_assert_cmpint(backing_s->drain_count, ==, 0);
228 
229     /* Now do the same while a request is pending */
230     aio_ret = -EINPROGRESS;
231     acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
232     g_assert(acb != NULL);
233     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
234 
235     g_assert_cmpint(s->drain_count, ==, 0);
236     g_assert_cmpint(backing_s->drain_count, ==, 0);
237 
238     do_drain_begin(drain_type, bs);
239 
240     g_assert_cmpint(aio_ret, ==, 0);
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     bdrv_unref(backing);
250     bdrv_unref(bs);
251     blk_unref(blk);
252 }
253 
254 static void test_drv_cb_drain_all(void)
255 {
256     test_drv_cb_common(BDRV_DRAIN_ALL, true);
257 }
258 
259 static void test_drv_cb_drain(void)
260 {
261     test_drv_cb_common(BDRV_DRAIN, false);
262 }
263 
264 static void test_drv_cb_drain_subtree(void)
265 {
266     test_drv_cb_common(BDRV_SUBTREE_DRAIN, true);
267 }
268 
269 static void test_drv_cb_co_drain_all(void)
270 {
271     call_in_coroutine(test_drv_cb_drain_all);
272 }
273 
274 static void test_drv_cb_co_drain(void)
275 {
276     call_in_coroutine(test_drv_cb_drain);
277 }
278 
279 static void test_drv_cb_co_drain_subtree(void)
280 {
281     call_in_coroutine(test_drv_cb_drain_subtree);
282 }
283 
284 static void test_quiesce_common(enum drain_type drain_type, bool recursive)
285 {
286     BlockBackend *blk;
287     BlockDriverState *bs, *backing;
288 
289     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
290     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
291                               &error_abort);
292     blk_insert_bs(blk, bs, &error_abort);
293 
294     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
295     bdrv_set_backing_hd(bs, backing, &error_abort);
296 
297     g_assert_cmpint(bs->quiesce_counter, ==, 0);
298     g_assert_cmpint(backing->quiesce_counter, ==, 0);
299 
300     do_drain_begin(drain_type, bs);
301 
302     g_assert_cmpint(bs->quiesce_counter, ==, 1);
303     g_assert_cmpint(backing->quiesce_counter, ==, !!recursive);
304 
305     do_drain_end(drain_type, bs);
306 
307     g_assert_cmpint(bs->quiesce_counter, ==, 0);
308     g_assert_cmpint(backing->quiesce_counter, ==, 0);
309 
310     bdrv_unref(backing);
311     bdrv_unref(bs);
312     blk_unref(blk);
313 }
314 
315 static void test_quiesce_drain_all(void)
316 {
317     test_quiesce_common(BDRV_DRAIN_ALL, true);
318 }
319 
320 static void test_quiesce_drain(void)
321 {
322     test_quiesce_common(BDRV_DRAIN, false);
323 }
324 
325 static void test_quiesce_drain_subtree(void)
326 {
327     test_quiesce_common(BDRV_SUBTREE_DRAIN, true);
328 }
329 
330 static void test_quiesce_co_drain_all(void)
331 {
332     call_in_coroutine(test_quiesce_drain_all);
333 }
334 
335 static void test_quiesce_co_drain(void)
336 {
337     call_in_coroutine(test_quiesce_drain);
338 }
339 
340 static void test_quiesce_co_drain_subtree(void)
341 {
342     call_in_coroutine(test_quiesce_drain_subtree);
343 }
344 
345 static void test_nested(void)
346 {
347     BlockBackend *blk;
348     BlockDriverState *bs, *backing;
349     BDRVTestState *s, *backing_s;
350     enum drain_type outer, inner;
351 
352     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
353     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
354                               &error_abort);
355     s = bs->opaque;
356     blk_insert_bs(blk, bs, &error_abort);
357 
358     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
359     backing_s = backing->opaque;
360     bdrv_set_backing_hd(bs, backing, &error_abort);
361 
362     for (outer = 0; outer < DRAIN_TYPE_MAX; outer++) {
363         for (inner = 0; inner < DRAIN_TYPE_MAX; inner++) {
364             int backing_quiesce = (outer != BDRV_DRAIN) +
365                                   (inner != BDRV_DRAIN);
366 
367             g_assert_cmpint(bs->quiesce_counter, ==, 0);
368             g_assert_cmpint(backing->quiesce_counter, ==, 0);
369             g_assert_cmpint(s->drain_count, ==, 0);
370             g_assert_cmpint(backing_s->drain_count, ==, 0);
371 
372             do_drain_begin(outer, bs);
373             do_drain_begin(inner, bs);
374 
375             g_assert_cmpint(bs->quiesce_counter, ==, 2);
376             g_assert_cmpint(backing->quiesce_counter, ==, backing_quiesce);
377             g_assert_cmpint(s->drain_count, ==, 2);
378             g_assert_cmpint(backing_s->drain_count, ==, backing_quiesce);
379 
380             do_drain_end(inner, bs);
381             do_drain_end(outer, bs);
382 
383             g_assert_cmpint(bs->quiesce_counter, ==, 0);
384             g_assert_cmpint(backing->quiesce_counter, ==, 0);
385             g_assert_cmpint(s->drain_count, ==, 0);
386             g_assert_cmpint(backing_s->drain_count, ==, 0);
387         }
388     }
389 
390     bdrv_unref(backing);
391     bdrv_unref(bs);
392     blk_unref(blk);
393 }
394 
395 static void test_multiparent(void)
396 {
397     BlockBackend *blk_a, *blk_b;
398     BlockDriverState *bs_a, *bs_b, *backing;
399     BDRVTestState *a_s, *b_s, *backing_s;
400 
401     blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
402     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
403                                 &error_abort);
404     a_s = bs_a->opaque;
405     blk_insert_bs(blk_a, bs_a, &error_abort);
406 
407     blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
408     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
409                                 &error_abort);
410     b_s = bs_b->opaque;
411     blk_insert_bs(blk_b, bs_b, &error_abort);
412 
413     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
414     backing_s = backing->opaque;
415     bdrv_set_backing_hd(bs_a, backing, &error_abort);
416     bdrv_set_backing_hd(bs_b, backing, &error_abort);
417 
418     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
419     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
420     g_assert_cmpint(backing->quiesce_counter, ==, 0);
421     g_assert_cmpint(a_s->drain_count, ==, 0);
422     g_assert_cmpint(b_s->drain_count, ==, 0);
423     g_assert_cmpint(backing_s->drain_count, ==, 0);
424 
425     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
426 
427     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
428     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
429     g_assert_cmpint(backing->quiesce_counter, ==, 1);
430     g_assert_cmpint(a_s->drain_count, ==, 1);
431     g_assert_cmpint(b_s->drain_count, ==, 1);
432     g_assert_cmpint(backing_s->drain_count, ==, 1);
433 
434     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
435 
436     g_assert_cmpint(bs_a->quiesce_counter, ==, 2);
437     g_assert_cmpint(bs_b->quiesce_counter, ==, 2);
438     g_assert_cmpint(backing->quiesce_counter, ==, 2);
439     g_assert_cmpint(a_s->drain_count, ==, 2);
440     g_assert_cmpint(b_s->drain_count, ==, 2);
441     g_assert_cmpint(backing_s->drain_count, ==, 2);
442 
443     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
444 
445     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
446     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
447     g_assert_cmpint(backing->quiesce_counter, ==, 1);
448     g_assert_cmpint(a_s->drain_count, ==, 1);
449     g_assert_cmpint(b_s->drain_count, ==, 1);
450     g_assert_cmpint(backing_s->drain_count, ==, 1);
451 
452     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
453 
454     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
455     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
456     g_assert_cmpint(backing->quiesce_counter, ==, 0);
457     g_assert_cmpint(a_s->drain_count, ==, 0);
458     g_assert_cmpint(b_s->drain_count, ==, 0);
459     g_assert_cmpint(backing_s->drain_count, ==, 0);
460 
461     bdrv_unref(backing);
462     bdrv_unref(bs_a);
463     bdrv_unref(bs_b);
464     blk_unref(blk_a);
465     blk_unref(blk_b);
466 }
467 
468 static void test_graph_change_drain_subtree(void)
469 {
470     BlockBackend *blk_a, *blk_b;
471     BlockDriverState *bs_a, *bs_b, *backing;
472     BDRVTestState *a_s, *b_s, *backing_s;
473 
474     blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
475     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
476                                 &error_abort);
477     a_s = bs_a->opaque;
478     blk_insert_bs(blk_a, bs_a, &error_abort);
479 
480     blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
481     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
482                                 &error_abort);
483     b_s = bs_b->opaque;
484     blk_insert_bs(blk_b, bs_b, &error_abort);
485 
486     backing = bdrv_new_open_driver(&bdrv_test, "backing", 0, &error_abort);
487     backing_s = backing->opaque;
488     bdrv_set_backing_hd(bs_a, backing, &error_abort);
489 
490     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
491     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
492     g_assert_cmpint(backing->quiesce_counter, ==, 0);
493     g_assert_cmpint(a_s->drain_count, ==, 0);
494     g_assert_cmpint(b_s->drain_count, ==, 0);
495     g_assert_cmpint(backing_s->drain_count, ==, 0);
496 
497     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
498     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
499     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_a);
500     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
501     do_drain_begin(BDRV_SUBTREE_DRAIN, bs_b);
502 
503     bdrv_set_backing_hd(bs_b, backing, &error_abort);
504     g_assert_cmpint(bs_a->quiesce_counter, ==, 5);
505     g_assert_cmpint(bs_b->quiesce_counter, ==, 5);
506     g_assert_cmpint(backing->quiesce_counter, ==, 5);
507     g_assert_cmpint(a_s->drain_count, ==, 5);
508     g_assert_cmpint(b_s->drain_count, ==, 5);
509     g_assert_cmpint(backing_s->drain_count, ==, 5);
510 
511     bdrv_set_backing_hd(bs_b, NULL, &error_abort);
512     g_assert_cmpint(bs_a->quiesce_counter, ==, 3);
513     g_assert_cmpint(bs_b->quiesce_counter, ==, 2);
514     g_assert_cmpint(backing->quiesce_counter, ==, 3);
515     g_assert_cmpint(a_s->drain_count, ==, 3);
516     g_assert_cmpint(b_s->drain_count, ==, 2);
517     g_assert_cmpint(backing_s->drain_count, ==, 3);
518 
519     bdrv_set_backing_hd(bs_b, backing, &error_abort);
520     g_assert_cmpint(bs_a->quiesce_counter, ==, 5);
521     g_assert_cmpint(bs_b->quiesce_counter, ==, 5);
522     g_assert_cmpint(backing->quiesce_counter, ==, 5);
523     g_assert_cmpint(a_s->drain_count, ==, 5);
524     g_assert_cmpint(b_s->drain_count, ==, 5);
525     g_assert_cmpint(backing_s->drain_count, ==, 5);
526 
527     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
528     do_drain_end(BDRV_SUBTREE_DRAIN, bs_b);
529     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
530     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
531     do_drain_end(BDRV_SUBTREE_DRAIN, bs_a);
532 
533     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
534     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
535     g_assert_cmpint(backing->quiesce_counter, ==, 0);
536     g_assert_cmpint(a_s->drain_count, ==, 0);
537     g_assert_cmpint(b_s->drain_count, ==, 0);
538     g_assert_cmpint(backing_s->drain_count, ==, 0);
539 
540     bdrv_unref(backing);
541     bdrv_unref(bs_a);
542     bdrv_unref(bs_b);
543     blk_unref(blk_a);
544     blk_unref(blk_b);
545 }
546 
547 static void test_graph_change_drain_all(void)
548 {
549     BlockBackend *blk_a, *blk_b;
550     BlockDriverState *bs_a, *bs_b;
551     BDRVTestState *a_s, *b_s;
552 
553     /* Create node A with a BlockBackend */
554     blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
555     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
556                                 &error_abort);
557     a_s = bs_a->opaque;
558     blk_insert_bs(blk_a, bs_a, &error_abort);
559 
560     g_assert_cmpint(bs_a->quiesce_counter, ==, 0);
561     g_assert_cmpint(a_s->drain_count, ==, 0);
562 
563     /* Call bdrv_drain_all_begin() */
564     bdrv_drain_all_begin();
565 
566     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
567     g_assert_cmpint(a_s->drain_count, ==, 1);
568 
569     /* Create node B with a BlockBackend */
570     blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
571     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
572                                 &error_abort);
573     b_s = bs_b->opaque;
574     blk_insert_bs(blk_b, bs_b, &error_abort);
575 
576     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
577     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
578     g_assert_cmpint(a_s->drain_count, ==, 1);
579     g_assert_cmpint(b_s->drain_count, ==, 1);
580 
581     /* Unref and finally delete node A */
582     blk_unref(blk_a);
583 
584     g_assert_cmpint(bs_a->quiesce_counter, ==, 1);
585     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
586     g_assert_cmpint(a_s->drain_count, ==, 1);
587     g_assert_cmpint(b_s->drain_count, ==, 1);
588 
589     bdrv_unref(bs_a);
590 
591     g_assert_cmpint(bs_b->quiesce_counter, ==, 1);
592     g_assert_cmpint(b_s->drain_count, ==, 1);
593 
594     /* End the drained section */
595     bdrv_drain_all_end();
596 
597     g_assert_cmpint(bs_b->quiesce_counter, ==, 0);
598     g_assert_cmpint(b_s->drain_count, ==, 0);
599     g_assert_cmpint(qemu_get_aio_context()->external_disable_cnt, ==, 0);
600 
601     bdrv_unref(bs_b);
602     blk_unref(blk_b);
603 }
604 
605 struct test_iothread_data {
606     BlockDriverState *bs;
607     enum drain_type drain_type;
608     int *aio_ret;
609 };
610 
611 static void test_iothread_drain_entry(void *opaque)
612 {
613     struct test_iothread_data *data = opaque;
614 
615     aio_context_acquire(bdrv_get_aio_context(data->bs));
616     do_drain_begin(data->drain_type, data->bs);
617     g_assert_cmpint(*data->aio_ret, ==, 0);
618     do_drain_end(data->drain_type, data->bs);
619     aio_context_release(bdrv_get_aio_context(data->bs));
620 
621     qemu_event_set(&done_event);
622 }
623 
624 static void test_iothread_aio_cb(void *opaque, int ret)
625 {
626     int *aio_ret = opaque;
627     *aio_ret = ret;
628     qemu_event_set(&done_event);
629 }
630 
631 static void test_iothread_main_thread_bh(void *opaque)
632 {
633     struct test_iothread_data *data = opaque;
634 
635     /* Test that the AioContext is not yet locked in a random BH that is
636      * executed during drain, otherwise this would deadlock. */
637     aio_context_acquire(bdrv_get_aio_context(data->bs));
638     bdrv_flush(data->bs);
639     aio_context_release(bdrv_get_aio_context(data->bs));
640 }
641 
642 /*
643  * Starts an AIO request on a BDS that runs in the AioContext of iothread 1.
644  * The request involves a BH on iothread 2 before it can complete.
645  *
646  * @drain_thread = 0 means that do_drain_begin/end are called from the main
647  * thread, @drain_thread = 1 means that they are called from iothread 1. Drain
648  * for this BDS cannot be called from iothread 2 because only the main thread
649  * may do cross-AioContext polling.
650  */
651 static void test_iothread_common(enum drain_type drain_type, int drain_thread)
652 {
653     BlockBackend *blk;
654     BlockDriverState *bs;
655     BDRVTestState *s;
656     BlockAIOCB *acb;
657     int aio_ret;
658     struct test_iothread_data data;
659 
660     IOThread *a = iothread_new();
661     IOThread *b = iothread_new();
662     AioContext *ctx_a = iothread_get_aio_context(a);
663     AioContext *ctx_b = iothread_get_aio_context(b);
664 
665     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
666 
667     /* bdrv_drain_all() may only be called from the main loop thread */
668     if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) {
669         goto out;
670     }
671 
672     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
673     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
674                               &error_abort);
675     s = bs->opaque;
676     blk_insert_bs(blk, bs, &error_abort);
677     blk_set_disable_request_queuing(blk, true);
678 
679     blk_set_aio_context(blk, ctx_a, &error_abort);
680     aio_context_acquire(ctx_a);
681 
682     s->bh_indirection_ctx = ctx_b;
683 
684     aio_ret = -EINPROGRESS;
685     qemu_event_reset(&done_event);
686 
687     if (drain_thread == 0) {
688         acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret);
689     } else {
690         acb = blk_aio_preadv(blk, 0, &qiov, 0, aio_ret_cb, &aio_ret);
691     }
692     g_assert(acb != NULL);
693     g_assert_cmpint(aio_ret, ==, -EINPROGRESS);
694 
695     aio_context_release(ctx_a);
696 
697     data = (struct test_iothread_data) {
698         .bs         = bs,
699         .drain_type = drain_type,
700         .aio_ret    = &aio_ret,
701     };
702 
703     switch (drain_thread) {
704     case 0:
705         if (drain_type != BDRV_DRAIN_ALL) {
706             aio_context_acquire(ctx_a);
707         }
708 
709         aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data);
710 
711         /* The request is running on the IOThread a. Draining its block device
712          * will make sure that it has completed as far as the BDS is concerned,
713          * but the drain in this thread can continue immediately after
714          * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
715          * later. */
716         do_drain_begin(drain_type, bs);
717         g_assert_cmpint(bs->in_flight, ==, 0);
718 
719         if (drain_type != BDRV_DRAIN_ALL) {
720             aio_context_release(ctx_a);
721         }
722         qemu_event_wait(&done_event);
723         if (drain_type != BDRV_DRAIN_ALL) {
724             aio_context_acquire(ctx_a);
725         }
726 
727         g_assert_cmpint(aio_ret, ==, 0);
728         do_drain_end(drain_type, bs);
729 
730         if (drain_type != BDRV_DRAIN_ALL) {
731             aio_context_release(ctx_a);
732         }
733         break;
734     case 1:
735         aio_bh_schedule_oneshot(ctx_a, test_iothread_drain_entry, &data);
736         qemu_event_wait(&done_event);
737         break;
738     default:
739         g_assert_not_reached();
740     }
741 
742     aio_context_acquire(ctx_a);
743     blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort);
744     aio_context_release(ctx_a);
745 
746     bdrv_unref(bs);
747     blk_unref(blk);
748 
749 out:
750     iothread_join(a);
751     iothread_join(b);
752 }
753 
754 static void test_iothread_drain_all(void)
755 {
756     test_iothread_common(BDRV_DRAIN_ALL, 0);
757     test_iothread_common(BDRV_DRAIN_ALL, 1);
758 }
759 
760 static void test_iothread_drain(void)
761 {
762     test_iothread_common(BDRV_DRAIN, 0);
763     test_iothread_common(BDRV_DRAIN, 1);
764 }
765 
766 static void test_iothread_drain_subtree(void)
767 {
768     test_iothread_common(BDRV_SUBTREE_DRAIN, 0);
769     test_iothread_common(BDRV_SUBTREE_DRAIN, 1);
770 }
771 
772 
773 typedef struct TestBlockJob {
774     BlockJob common;
775     BlockDriverState *bs;
776     int run_ret;
777     int prepare_ret;
778     bool running;
779     bool should_complete;
780 } TestBlockJob;
781 
782 static int test_job_prepare(Job *job)
783 {
784     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
785 
786     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
787     bdrv_flush(s->bs);
788     return s->prepare_ret;
789 }
790 
791 static void test_job_commit(Job *job)
792 {
793     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
794 
795     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
796     bdrv_flush(s->bs);
797 }
798 
799 static void test_job_abort(Job *job)
800 {
801     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
802 
803     /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
804     bdrv_flush(s->bs);
805 }
806 
807 static int coroutine_fn test_job_run(Job *job, Error **errp)
808 {
809     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
810 
811     /* We are running the actual job code past the pause point in
812      * job_co_entry(). */
813     s->running = true;
814 
815     job_transition_to_ready(&s->common.job);
816     while (!s->should_complete) {
817         /* Avoid job_sleep_ns() because it marks the job as !busy. We want to
818          * emulate some actual activity (probably some I/O) here so that drain
819          * has to wait for this activity to stop. */
820         qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000);
821 
822         job_pause_point(&s->common.job);
823     }
824 
825     return s->run_ret;
826 }
827 
828 static void test_job_complete(Job *job, Error **errp)
829 {
830     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
831     s->should_complete = true;
832 }
833 
834 BlockJobDriver test_job_driver = {
835     .job_driver = {
836         .instance_size  = sizeof(TestBlockJob),
837         .free           = block_job_free,
838         .user_resume    = block_job_user_resume,
839         .run            = test_job_run,
840         .complete       = test_job_complete,
841         .prepare        = test_job_prepare,
842         .commit         = test_job_commit,
843         .abort          = test_job_abort,
844     },
845 };
846 
847 enum test_job_result {
848     TEST_JOB_SUCCESS,
849     TEST_JOB_FAIL_RUN,
850     TEST_JOB_FAIL_PREPARE,
851 };
852 
853 enum test_job_drain_node {
854     TEST_JOB_DRAIN_SRC,
855     TEST_JOB_DRAIN_SRC_CHILD,
856     TEST_JOB_DRAIN_SRC_PARENT,
857 };
858 
859 static void test_blockjob_common_drain_node(enum drain_type drain_type,
860                                             bool use_iothread,
861                                             enum test_job_result result,
862                                             enum test_job_drain_node drain_node)
863 {
864     BlockBackend *blk_src, *blk_target;
865     BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs;
866     BlockJob *job;
867     TestBlockJob *tjob;
868     IOThread *iothread = NULL;
869     AioContext *ctx;
870     int ret;
871 
872     src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR,
873                                &error_abort);
874     src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing",
875                                        BDRV_O_RDWR, &error_abort);
876     src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay",
877                                        BDRV_O_RDWR, &error_abort);
878 
879     bdrv_set_backing_hd(src_overlay, src, &error_abort);
880     bdrv_unref(src);
881     bdrv_set_backing_hd(src, src_backing, &error_abort);
882     bdrv_unref(src_backing);
883 
884     blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
885     blk_insert_bs(blk_src, src_overlay, &error_abort);
886 
887     switch (drain_node) {
888     case TEST_JOB_DRAIN_SRC:
889         drain_bs = src;
890         break;
891     case TEST_JOB_DRAIN_SRC_CHILD:
892         drain_bs = src_backing;
893         break;
894     case TEST_JOB_DRAIN_SRC_PARENT:
895         drain_bs = src_overlay;
896         break;
897     default:
898         g_assert_not_reached();
899     }
900 
901     if (use_iothread) {
902         iothread = iothread_new();
903         ctx = iothread_get_aio_context(iothread);
904         blk_set_aio_context(blk_src, ctx, &error_abort);
905     } else {
906         ctx = qemu_get_aio_context();
907     }
908 
909     target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR,
910                                   &error_abort);
911     blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
912     blk_insert_bs(blk_target, target, &error_abort);
913     blk_set_allow_aio_context_change(blk_target, true);
914 
915     aio_context_acquire(ctx);
916     tjob = block_job_create("job0", &test_job_driver, NULL, src,
917                             0, BLK_PERM_ALL,
918                             0, 0, NULL, NULL, &error_abort);
919     tjob->bs = src;
920     job = &tjob->common;
921     block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort);
922 
923     switch (result) {
924     case TEST_JOB_SUCCESS:
925         break;
926     case TEST_JOB_FAIL_RUN:
927         tjob->run_ret = -EIO;
928         break;
929     case TEST_JOB_FAIL_PREPARE:
930         tjob->prepare_ret = -EIO;
931         break;
932     }
933 
934     job_start(&job->job);
935     aio_context_release(ctx);
936 
937     if (use_iothread) {
938         /* job_co_entry() is run in the I/O thread, wait for the actual job
939          * code to start (we don't want to catch the job in the pause point in
940          * job_co_entry(). */
941         while (!tjob->running) {
942             aio_poll(qemu_get_aio_context(), false);
943         }
944     }
945 
946     g_assert_cmpint(job->job.pause_count, ==, 0);
947     g_assert_false(job->job.paused);
948     g_assert_true(tjob->running);
949     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
950 
951     do_drain_begin_unlocked(drain_type, drain_bs);
952 
953     if (drain_type == BDRV_DRAIN_ALL) {
954         /* bdrv_drain_all() drains both src and target */
955         g_assert_cmpint(job->job.pause_count, ==, 2);
956     } else {
957         g_assert_cmpint(job->job.pause_count, ==, 1);
958     }
959     g_assert_true(job->job.paused);
960     g_assert_false(job->job.busy); /* The job is paused */
961 
962     do_drain_end_unlocked(drain_type, drain_bs);
963 
964     if (use_iothread) {
965         /* paused is reset in the I/O thread, wait for it */
966         while (job->job.paused) {
967             aio_poll(qemu_get_aio_context(), false);
968         }
969     }
970 
971     g_assert_cmpint(job->job.pause_count, ==, 0);
972     g_assert_false(job->job.paused);
973     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
974 
975     do_drain_begin_unlocked(drain_type, target);
976 
977     if (drain_type == BDRV_DRAIN_ALL) {
978         /* bdrv_drain_all() drains both src and target */
979         g_assert_cmpint(job->job.pause_count, ==, 2);
980     } else {
981         g_assert_cmpint(job->job.pause_count, ==, 1);
982     }
983     g_assert_true(job->job.paused);
984     g_assert_false(job->job.busy); /* The job is paused */
985 
986     do_drain_end_unlocked(drain_type, target);
987 
988     if (use_iothread) {
989         /* paused is reset in the I/O thread, wait for it */
990         while (job->job.paused) {
991             aio_poll(qemu_get_aio_context(), false);
992         }
993     }
994 
995     g_assert_cmpint(job->job.pause_count, ==, 0);
996     g_assert_false(job->job.paused);
997     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
998 
999     aio_context_acquire(ctx);
1000     ret = job_complete_sync(&job->job, &error_abort);
1001     g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO));
1002 
1003     if (use_iothread) {
1004         blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort);
1005         assert(blk_get_aio_context(blk_target) == qemu_get_aio_context());
1006     }
1007     aio_context_release(ctx);
1008 
1009     blk_unref(blk_src);
1010     blk_unref(blk_target);
1011     bdrv_unref(src_overlay);
1012     bdrv_unref(target);
1013 
1014     if (iothread) {
1015         iothread_join(iothread);
1016     }
1017 }
1018 
1019 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread,
1020                                  enum test_job_result result)
1021 {
1022     test_blockjob_common_drain_node(drain_type, use_iothread, result,
1023                                     TEST_JOB_DRAIN_SRC);
1024     test_blockjob_common_drain_node(drain_type, use_iothread, result,
1025                                     TEST_JOB_DRAIN_SRC_CHILD);
1026     if (drain_type == BDRV_SUBTREE_DRAIN) {
1027         test_blockjob_common_drain_node(drain_type, use_iothread, result,
1028                                         TEST_JOB_DRAIN_SRC_PARENT);
1029     }
1030 }
1031 
1032 static void test_blockjob_drain_all(void)
1033 {
1034     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS);
1035 }
1036 
1037 static void test_blockjob_drain(void)
1038 {
1039     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS);
1040 }
1041 
1042 static void test_blockjob_drain_subtree(void)
1043 {
1044     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_SUCCESS);
1045 }
1046 
1047 static void test_blockjob_error_drain_all(void)
1048 {
1049     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN);
1050     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE);
1051 }
1052 
1053 static void test_blockjob_error_drain(void)
1054 {
1055     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN);
1056     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE);
1057 }
1058 
1059 static void test_blockjob_error_drain_subtree(void)
1060 {
1061     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_RUN);
1062     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_PREPARE);
1063 }
1064 
1065 static void test_blockjob_iothread_drain_all(void)
1066 {
1067     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS);
1068 }
1069 
1070 static void test_blockjob_iothread_drain(void)
1071 {
1072     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS);
1073 }
1074 
1075 static void test_blockjob_iothread_drain_subtree(void)
1076 {
1077     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_SUCCESS);
1078 }
1079 
1080 static void test_blockjob_iothread_error_drain_all(void)
1081 {
1082     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN);
1083     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE);
1084 }
1085 
1086 static void test_blockjob_iothread_error_drain(void)
1087 {
1088     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN);
1089     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE);
1090 }
1091 
1092 static void test_blockjob_iothread_error_drain_subtree(void)
1093 {
1094     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_RUN);
1095     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_PREPARE);
1096 }
1097 
1098 
1099 typedef struct BDRVTestTopState {
1100     BdrvChild *wait_child;
1101 } BDRVTestTopState;
1102 
1103 static void bdrv_test_top_close(BlockDriverState *bs)
1104 {
1105     BdrvChild *c, *next_c;
1106     QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1107         bdrv_unref_child(bs, c);
1108     }
1109 }
1110 
1111 static int coroutine_fn bdrv_test_top_co_preadv(BlockDriverState *bs,
1112                                                 int64_t offset, int64_t bytes,
1113                                                 QEMUIOVector *qiov,
1114                                                 BdrvRequestFlags flags)
1115 {
1116     BDRVTestTopState *tts = bs->opaque;
1117     return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags);
1118 }
1119 
1120 static BlockDriver bdrv_test_top_driver = {
1121     .format_name            = "test_top_driver",
1122     .instance_size          = sizeof(BDRVTestTopState),
1123 
1124     .bdrv_close             = bdrv_test_top_close,
1125     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
1126 
1127     .bdrv_child_perm        = bdrv_default_perms,
1128 };
1129 
1130 typedef struct TestCoDeleteByDrainData {
1131     BlockBackend *blk;
1132     bool detach_instead_of_delete;
1133     bool done;
1134 } TestCoDeleteByDrainData;
1135 
1136 static void coroutine_fn test_co_delete_by_drain(void *opaque)
1137 {
1138     TestCoDeleteByDrainData *dbdd = opaque;
1139     BlockBackend *blk = dbdd->blk;
1140     BlockDriverState *bs = blk_bs(blk);
1141     BDRVTestTopState *tts = bs->opaque;
1142     void *buffer = g_malloc(65536);
1143     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536);
1144 
1145     /* Pretend some internal write operation from parent to child.
1146      * Important: We have to read from the child, not from the parent!
1147      * Draining works by first propagating it all up the tree to the
1148      * root and then waiting for drainage from root to the leaves
1149      * (protocol nodes).  If we have a request waiting on the root,
1150      * everything will be drained before we go back down the tree, but
1151      * we do not want that.  We want to be in the middle of draining
1152      * when this following requests returns. */
1153     bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0);
1154 
1155     g_assert_cmpint(bs->refcnt, ==, 1);
1156 
1157     if (!dbdd->detach_instead_of_delete) {
1158         blk_unref(blk);
1159     } else {
1160         BdrvChild *c, *next_c;
1161         QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1162             bdrv_unref_child(bs, c);
1163         }
1164     }
1165 
1166     dbdd->done = true;
1167     g_free(buffer);
1168 }
1169 
1170 /**
1171  * Test what happens when some BDS has some children, you drain one of
1172  * them and this results in the BDS being deleted.
1173  *
1174  * If @detach_instead_of_delete is set, the BDS is not going to be
1175  * deleted but will only detach all of its children.
1176  */
1177 static void do_test_delete_by_drain(bool detach_instead_of_delete,
1178                                     enum drain_type drain_type)
1179 {
1180     BlockBackend *blk;
1181     BlockDriverState *bs, *child_bs, *null_bs;
1182     BDRVTestTopState *tts;
1183     TestCoDeleteByDrainData dbdd;
1184     Coroutine *co;
1185 
1186     bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR,
1187                               &error_abort);
1188     bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1189     tts = bs->opaque;
1190 
1191     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1192                         &error_abort);
1193     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds,
1194                       BDRV_CHILD_DATA, &error_abort);
1195 
1196     /* This child will be the one to pass to requests through to, and
1197      * it will stall until a drain occurs */
1198     child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR,
1199                                     &error_abort);
1200     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1201     /* Takes our reference to child_bs */
1202     tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child",
1203                                         &child_of_bds,
1204                                         BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY,
1205                                         &error_abort);
1206 
1207     /* This child is just there to be deleted
1208      * (for detach_instead_of_delete == true) */
1209     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1210                         &error_abort);
1211     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA,
1212                       &error_abort);
1213 
1214     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1215     blk_insert_bs(blk, bs, &error_abort);
1216 
1217     /* Referenced by blk now */
1218     bdrv_unref(bs);
1219 
1220     g_assert_cmpint(bs->refcnt, ==, 1);
1221     g_assert_cmpint(child_bs->refcnt, ==, 1);
1222     g_assert_cmpint(null_bs->refcnt, ==, 1);
1223 
1224 
1225     dbdd = (TestCoDeleteByDrainData){
1226         .blk = blk,
1227         .detach_instead_of_delete = detach_instead_of_delete,
1228         .done = false,
1229     };
1230     co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd);
1231     qemu_coroutine_enter(co);
1232 
1233     /* Drain the child while the read operation is still pending.
1234      * This should result in the operation finishing and
1235      * test_co_delete_by_drain() resuming.  Thus, @bs will be deleted
1236      * and the coroutine will exit while this drain operation is still
1237      * in progress. */
1238     switch (drain_type) {
1239     case BDRV_DRAIN:
1240         bdrv_ref(child_bs);
1241         bdrv_drain(child_bs);
1242         bdrv_unref(child_bs);
1243         break;
1244     case BDRV_SUBTREE_DRAIN:
1245         /* Would have to ref/unref bs here for !detach_instead_of_delete, but
1246          * then the whole test becomes pointless because the graph changes
1247          * don't occur during the drain any more. */
1248         assert(detach_instead_of_delete);
1249         bdrv_subtree_drained_begin(bs);
1250         bdrv_subtree_drained_end(bs);
1251         break;
1252     case BDRV_DRAIN_ALL:
1253         bdrv_drain_all_begin();
1254         bdrv_drain_all_end();
1255         break;
1256     default:
1257         g_assert_not_reached();
1258     }
1259 
1260     while (!dbdd.done) {
1261         aio_poll(qemu_get_aio_context(), true);
1262     }
1263 
1264     if (detach_instead_of_delete) {
1265         /* Here, the reference has not passed over to the coroutine,
1266          * so we have to delete the BB ourselves */
1267         blk_unref(blk);
1268     }
1269 }
1270 
1271 static void test_delete_by_drain(void)
1272 {
1273     do_test_delete_by_drain(false, BDRV_DRAIN);
1274 }
1275 
1276 static void test_detach_by_drain_all(void)
1277 {
1278     do_test_delete_by_drain(true, BDRV_DRAIN_ALL);
1279 }
1280 
1281 static void test_detach_by_drain(void)
1282 {
1283     do_test_delete_by_drain(true, BDRV_DRAIN);
1284 }
1285 
1286 static void test_detach_by_drain_subtree(void)
1287 {
1288     do_test_delete_by_drain(true, BDRV_SUBTREE_DRAIN);
1289 }
1290 
1291 
1292 struct detach_by_parent_data {
1293     BlockDriverState *parent_b;
1294     BdrvChild *child_b;
1295     BlockDriverState *c;
1296     BdrvChild *child_c;
1297     bool by_parent_cb;
1298 };
1299 static struct detach_by_parent_data detach_by_parent_data;
1300 
1301 static void detach_indirect_bh(void *opaque)
1302 {
1303     struct detach_by_parent_data *data = opaque;
1304 
1305     bdrv_unref_child(data->parent_b, data->child_b);
1306 
1307     bdrv_ref(data->c);
1308     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
1309                                       &child_of_bds, BDRV_CHILD_DATA,
1310                                       &error_abort);
1311 }
1312 
1313 static void detach_by_parent_aio_cb(void *opaque, int ret)
1314 {
1315     struct detach_by_parent_data *data = &detach_by_parent_data;
1316 
1317     g_assert_cmpint(ret, ==, 0);
1318     if (data->by_parent_cb) {
1319         detach_indirect_bh(data);
1320     }
1321 }
1322 
1323 static void detach_by_driver_cb_drained_begin(BdrvChild *child)
1324 {
1325     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1326                             detach_indirect_bh, &detach_by_parent_data);
1327     child_of_bds.drained_begin(child);
1328 }
1329 
1330 static BdrvChildClass detach_by_driver_cb_class;
1331 
1332 /*
1333  * Initial graph:
1334  *
1335  * PA     PB
1336  *    \ /   \
1337  *     A     B     C
1338  *
1339  * by_parent_cb == true:  Test that parent callbacks don't poll
1340  *
1341  *     PA has a pending write request whose callback changes the child nodes of
1342  *     PB: It removes B and adds C instead. The subtree of PB is drained, which
1343  *     will indirectly drain the write request, too.
1344  *
1345  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1346  *
1347  *     PA's BdrvChildClass has a .drained_begin callback that schedules a BH
1348  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
1349  *     state is messed up, but if it is only polled in the single
1350  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1351  */
1352 static void test_detach_indirect(bool by_parent_cb)
1353 {
1354     BlockBackend *blk;
1355     BlockDriverState *parent_a, *parent_b, *a, *b, *c;
1356     BdrvChild *child_a, *child_b;
1357     BlockAIOCB *acb;
1358 
1359     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
1360 
1361     if (!by_parent_cb) {
1362         detach_by_driver_cb_class = child_of_bds;
1363         detach_by_driver_cb_class.drained_begin =
1364             detach_by_driver_cb_drained_begin;
1365     }
1366 
1367     /* Create all involved nodes */
1368     parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR,
1369                                     &error_abort);
1370     parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0,
1371                                     &error_abort);
1372 
1373     a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort);
1374     b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort);
1375     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
1376 
1377     /* blk is a BB for parent-a */
1378     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1379     blk_insert_bs(blk, parent_a, &error_abort);
1380     bdrv_unref(parent_a);
1381 
1382     /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1383      * callback must not return immediately. */
1384     if (!by_parent_cb) {
1385         BDRVTestState *s = parent_a->opaque;
1386         s->sleep_in_drain_begin = true;
1387     }
1388 
1389     /* Set child relationships */
1390     bdrv_ref(b);
1391     bdrv_ref(a);
1392     child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds,
1393                                 BDRV_CHILD_DATA, &error_abort);
1394     child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds,
1395                                 BDRV_CHILD_COW, &error_abort);
1396 
1397     bdrv_ref(a);
1398     bdrv_attach_child(parent_a, a, "PA-A",
1399                       by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class,
1400                       BDRV_CHILD_DATA, &error_abort);
1401 
1402     g_assert_cmpint(parent_a->refcnt, ==, 1);
1403     g_assert_cmpint(parent_b->refcnt, ==, 1);
1404     g_assert_cmpint(a->refcnt, ==, 3);
1405     g_assert_cmpint(b->refcnt, ==, 2);
1406     g_assert_cmpint(c->refcnt, ==, 1);
1407 
1408     g_assert(QLIST_FIRST(&parent_b->children) == child_a);
1409     g_assert(QLIST_NEXT(child_a, next) == child_b);
1410     g_assert(QLIST_NEXT(child_b, next) == NULL);
1411 
1412     /* Start the evil write request */
1413     detach_by_parent_data = (struct detach_by_parent_data) {
1414         .parent_b = parent_b,
1415         .child_b = child_b,
1416         .c = c,
1417         .by_parent_cb = by_parent_cb,
1418     };
1419     acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL);
1420     g_assert(acb != NULL);
1421 
1422     /* Drain and check the expected result */
1423     bdrv_subtree_drained_begin(parent_b);
1424 
1425     g_assert(detach_by_parent_data.child_c != NULL);
1426 
1427     g_assert_cmpint(parent_a->refcnt, ==, 1);
1428     g_assert_cmpint(parent_b->refcnt, ==, 1);
1429     g_assert_cmpint(a->refcnt, ==, 3);
1430     g_assert_cmpint(b->refcnt, ==, 1);
1431     g_assert_cmpint(c->refcnt, ==, 2);
1432 
1433     g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c);
1434     g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a);
1435     g_assert(QLIST_NEXT(child_a, next) == NULL);
1436 
1437     g_assert_cmpint(parent_a->quiesce_counter, ==, 1);
1438     g_assert_cmpint(parent_b->quiesce_counter, ==, 1);
1439     g_assert_cmpint(a->quiesce_counter, ==, 1);
1440     g_assert_cmpint(b->quiesce_counter, ==, 0);
1441     g_assert_cmpint(c->quiesce_counter, ==, 1);
1442 
1443     bdrv_subtree_drained_end(parent_b);
1444 
1445     bdrv_unref(parent_b);
1446     blk_unref(blk);
1447 
1448     g_assert_cmpint(a->refcnt, ==, 1);
1449     g_assert_cmpint(b->refcnt, ==, 1);
1450     g_assert_cmpint(c->refcnt, ==, 1);
1451     bdrv_unref(a);
1452     bdrv_unref(b);
1453     bdrv_unref(c);
1454 }
1455 
1456 static void test_detach_by_parent_cb(void)
1457 {
1458     test_detach_indirect(true);
1459 }
1460 
1461 static void test_detach_by_driver_cb(void)
1462 {
1463     test_detach_indirect(false);
1464 }
1465 
1466 static void test_append_to_drained(void)
1467 {
1468     BlockBackend *blk;
1469     BlockDriverState *base, *overlay;
1470     BDRVTestState *base_s, *overlay_s;
1471 
1472     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1473     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
1474     base_s = base->opaque;
1475     blk_insert_bs(blk, base, &error_abort);
1476 
1477     overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR,
1478                                    &error_abort);
1479     overlay_s = overlay->opaque;
1480 
1481     do_drain_begin(BDRV_DRAIN, base);
1482     g_assert_cmpint(base->quiesce_counter, ==, 1);
1483     g_assert_cmpint(base_s->drain_count, ==, 1);
1484     g_assert_cmpint(base->in_flight, ==, 0);
1485 
1486     bdrv_append(overlay, base, &error_abort);
1487     g_assert_cmpint(base->in_flight, ==, 0);
1488     g_assert_cmpint(overlay->in_flight, ==, 0);
1489 
1490     g_assert_cmpint(base->quiesce_counter, ==, 1);
1491     g_assert_cmpint(base_s->drain_count, ==, 1);
1492     g_assert_cmpint(overlay->quiesce_counter, ==, 1);
1493     g_assert_cmpint(overlay_s->drain_count, ==, 1);
1494 
1495     do_drain_end(BDRV_DRAIN, base);
1496 
1497     g_assert_cmpint(base->quiesce_counter, ==, 0);
1498     g_assert_cmpint(base_s->drain_count, ==, 0);
1499     g_assert_cmpint(overlay->quiesce_counter, ==, 0);
1500     g_assert_cmpint(overlay_s->drain_count, ==, 0);
1501 
1502     bdrv_unref(overlay);
1503     bdrv_unref(base);
1504     blk_unref(blk);
1505 }
1506 
1507 static void test_set_aio_context(void)
1508 {
1509     BlockDriverState *bs;
1510     IOThread *a = iothread_new();
1511     IOThread *b = iothread_new();
1512     AioContext *ctx_a = iothread_get_aio_context(a);
1513     AioContext *ctx_b = iothread_get_aio_context(b);
1514 
1515     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
1516                               &error_abort);
1517 
1518     bdrv_drained_begin(bs);
1519     bdrv_try_set_aio_context(bs, ctx_a, &error_abort);
1520 
1521     aio_context_acquire(ctx_a);
1522     bdrv_drained_end(bs);
1523 
1524     bdrv_drained_begin(bs);
1525     bdrv_try_set_aio_context(bs, ctx_b, &error_abort);
1526     aio_context_release(ctx_a);
1527     aio_context_acquire(ctx_b);
1528     bdrv_try_set_aio_context(bs, qemu_get_aio_context(), &error_abort);
1529     aio_context_release(ctx_b);
1530     bdrv_drained_end(bs);
1531 
1532     bdrv_unref(bs);
1533     iothread_join(a);
1534     iothread_join(b);
1535 }
1536 
1537 
1538 typedef struct TestDropBackingBlockJob {
1539     BlockJob common;
1540     bool should_complete;
1541     bool *did_complete;
1542     BlockDriverState *detach_also;
1543     BlockDriverState *bs;
1544 } TestDropBackingBlockJob;
1545 
1546 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp)
1547 {
1548     TestDropBackingBlockJob *s =
1549         container_of(job, TestDropBackingBlockJob, common.job);
1550 
1551     while (!s->should_complete) {
1552         job_sleep_ns(job, 0);
1553     }
1554 
1555     return 0;
1556 }
1557 
1558 static void test_drop_backing_job_commit(Job *job)
1559 {
1560     TestDropBackingBlockJob *s =
1561         container_of(job, TestDropBackingBlockJob, common.job);
1562 
1563     bdrv_set_backing_hd(s->bs, NULL, &error_abort);
1564     bdrv_set_backing_hd(s->detach_also, NULL, &error_abort);
1565 
1566     *s->did_complete = true;
1567 }
1568 
1569 static const BlockJobDriver test_drop_backing_job_driver = {
1570     .job_driver = {
1571         .instance_size  = sizeof(TestDropBackingBlockJob),
1572         .free           = block_job_free,
1573         .user_resume    = block_job_user_resume,
1574         .run            = test_drop_backing_job_run,
1575         .commit         = test_drop_backing_job_commit,
1576     }
1577 };
1578 
1579 /**
1580  * Creates a child node with three parent nodes on it, and then runs a
1581  * block job on the final one, parent-node-2.
1582  *
1583  * The job is then asked to complete before a section where the child
1584  * is drained.
1585  *
1586  * Ending this section will undrain the child's parents, first
1587  * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
1588  * list is in reverse order of how they were added.  Ending the drain
1589  * on parent-node-2 will resume the job, thus completing it and
1590  * scheduling job_exit().
1591  *
1592  * Ending the drain on parent-node-1 will poll the AioContext, which
1593  * lets job_exit() and thus test_drop_backing_job_commit() run.  That
1594  * function first removes the child as parent-node-2's backing file.
1595  *
1596  * In old (and buggy) implementations, there are two problems with
1597  * that:
1598  * (A) bdrv_drain_invoke() polls for every node that leaves the
1599  *     drained section.  This means that job_exit() is scheduled
1600  *     before the child has left the drained section.  Its
1601  *     quiesce_counter is therefore still 1 when it is removed from
1602  *     parent-node-2.
1603  *
1604  * (B) bdrv_replace_child_noperm() calls drained_end() on the old
1605  *     child's parents as many times as the child is quiesced.  This
1606  *     means it will call drained_end() on parent-node-2 once.
1607  *     Because parent-node-2 is no longer quiesced at this point, this
1608  *     will fail.
1609  *
1610  * bdrv_replace_child_noperm() therefore must call drained_end() on
1611  * the parent only if it really is still drained because the child is
1612  * drained.
1613  *
1614  * If removing child from parent-node-2 was successful (as it should
1615  * be), test_drop_backing_job_commit() will then also remove the child
1616  * from parent-node-0.
1617  *
1618  * With an old version of our drain infrastructure ((A) above), that
1619  * resulted in the following flow:
1620  *
1621  * 1. child attempts to leave its drained section.  The call recurses
1622  *    to its parents.
1623  *
1624  * 2. parent-node-2 leaves the drained section.  Polling in
1625  *    bdrv_drain_invoke() will schedule job_exit().
1626  *
1627  * 3. parent-node-1 leaves the drained section.  Polling in
1628  *    bdrv_drain_invoke() will run job_exit(), thus disconnecting
1629  *    parent-node-0 from the child node.
1630  *
1631  * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
1632  *    iterate over the parents.  Thus, it now accesses the BdrvChild
1633  *    object that used to connect parent-node-0 and the child node.
1634  *    However, that object no longer exists, so it accesses a dangling
1635  *    pointer.
1636  *
1637  * The solution is to only poll once when running a bdrv_drained_end()
1638  * operation, specifically at the end when all drained_end()
1639  * operations for all involved nodes have been scheduled.
1640  * Note that this also solves (A) above, thus hiding (B).
1641  */
1642 static void test_blockjob_commit_by_drained_end(void)
1643 {
1644     BlockDriverState *bs_child, *bs_parents[3];
1645     TestDropBackingBlockJob *job;
1646     bool job_has_completed = false;
1647     int i;
1648 
1649     bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR,
1650                                     &error_abort);
1651 
1652     for (i = 0; i < 3; i++) {
1653         char name[32];
1654         snprintf(name, sizeof(name), "parent-node-%i", i);
1655         bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR,
1656                                              &error_abort);
1657         bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort);
1658     }
1659 
1660     job = block_job_create("job", &test_drop_backing_job_driver, NULL,
1661                            bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL,
1662                            &error_abort);
1663     job->bs = bs_parents[2];
1664 
1665     job->detach_also = bs_parents[0];
1666     job->did_complete = &job_has_completed;
1667 
1668     job_start(&job->common.job);
1669 
1670     job->should_complete = true;
1671     bdrv_drained_begin(bs_child);
1672     g_assert(!job_has_completed);
1673     bdrv_drained_end(bs_child);
1674     g_assert(job_has_completed);
1675 
1676     bdrv_unref(bs_parents[0]);
1677     bdrv_unref(bs_parents[1]);
1678     bdrv_unref(bs_parents[2]);
1679     bdrv_unref(bs_child);
1680 }
1681 
1682 
1683 typedef struct TestSimpleBlockJob {
1684     BlockJob common;
1685     bool should_complete;
1686     bool *did_complete;
1687 } TestSimpleBlockJob;
1688 
1689 static int coroutine_fn test_simple_job_run(Job *job, Error **errp)
1690 {
1691     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1692 
1693     while (!s->should_complete) {
1694         job_sleep_ns(job, 0);
1695     }
1696 
1697     return 0;
1698 }
1699 
1700 static void test_simple_job_clean(Job *job)
1701 {
1702     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1703     *s->did_complete = true;
1704 }
1705 
1706 static const BlockJobDriver test_simple_job_driver = {
1707     .job_driver = {
1708         .instance_size  = sizeof(TestSimpleBlockJob),
1709         .free           = block_job_free,
1710         .user_resume    = block_job_user_resume,
1711         .run            = test_simple_job_run,
1712         .clean          = test_simple_job_clean,
1713     },
1714 };
1715 
1716 static int drop_intermediate_poll_update_filename(BdrvChild *child,
1717                                                   BlockDriverState *new_base,
1718                                                   const char *filename,
1719                                                   Error **errp)
1720 {
1721     /*
1722      * We are free to poll here, which may change the block graph, if
1723      * it is not drained.
1724      */
1725 
1726     /* If the job is not drained: Complete it, schedule job_exit() */
1727     aio_poll(qemu_get_current_aio_context(), false);
1728     /* If the job is not drained: Run job_exit(), finish the job */
1729     aio_poll(qemu_get_current_aio_context(), false);
1730 
1731     return 0;
1732 }
1733 
1734 /**
1735  * Test a poll in the midst of bdrv_drop_intermediate().
1736  *
1737  * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(),
1738  * which can yield or poll.  This may lead to graph changes, unless
1739  * the whole subtree in question is drained.
1740  *
1741  * We test this on the following graph:
1742  *
1743  *                    Job
1744  *
1745  *                     |
1746  *                  job-node
1747  *                     |
1748  *                     v
1749  *
1750  *                  job-node
1751  *
1752  *                     |
1753  *                  backing
1754  *                     |
1755  *                     v
1756  *
1757  * node-2 --chain--> node-1 --chain--> node-0
1758  *
1759  * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0).
1760  *
1761  * This first updates node-2's backing filename by invoking
1762  * drop_intermediate_poll_update_filename(), which polls twice.  This
1763  * causes the job to finish, which in turns causes the job-node to be
1764  * deleted.
1765  *
1766  * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it
1767  * already has a pointer to the BdrvChild edge between job-node and
1768  * node-1.  When it tries to handle that edge, we probably get a
1769  * segmentation fault because the object no longer exists.
1770  *
1771  *
1772  * The solution is for bdrv_drop_intermediate() to drain top's
1773  * subtree.  This prevents graph changes from happening just because
1774  * BdrvChildClass.update_filename() yields or polls.  Thus, the block
1775  * job is paused during that drained section and must finish before or
1776  * after.
1777  *
1778  * (In addition, bdrv_replace_child() must keep the job paused.)
1779  */
1780 static void test_drop_intermediate_poll(void)
1781 {
1782     static BdrvChildClass chain_child_class;
1783     BlockDriverState *chain[3];
1784     TestSimpleBlockJob *job;
1785     BlockDriverState *job_node;
1786     bool job_has_completed = false;
1787     int i;
1788     int ret;
1789 
1790     chain_child_class = child_of_bds;
1791     chain_child_class.update_filename = drop_intermediate_poll_update_filename;
1792 
1793     for (i = 0; i < 3; i++) {
1794         char name[32];
1795         snprintf(name, 32, "node-%i", i);
1796 
1797         chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort);
1798     }
1799 
1800     job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR,
1801                                     &error_abort);
1802     bdrv_set_backing_hd(job_node, chain[1], &error_abort);
1803 
1804     /*
1805      * Establish the chain last, so the chain links are the first
1806      * elements in the BDS.parents lists
1807      */
1808     for (i = 0; i < 3; i++) {
1809         if (i) {
1810             /* Takes the reference to chain[i - 1] */
1811             chain[i]->backing = bdrv_attach_child(chain[i], chain[i - 1],
1812                                                   "chain", &chain_child_class,
1813                                                   BDRV_CHILD_COW, &error_abort);
1814         }
1815     }
1816 
1817     job = block_job_create("job", &test_simple_job_driver, NULL, job_node,
1818                            0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort);
1819 
1820     /* The job has a reference now */
1821     bdrv_unref(job_node);
1822 
1823     job->did_complete = &job_has_completed;
1824 
1825     job_start(&job->common.job);
1826     job->should_complete = true;
1827 
1828     g_assert(!job_has_completed);
1829     ret = bdrv_drop_intermediate(chain[1], chain[0], NULL);
1830     g_assert(ret == 0);
1831     g_assert(job_has_completed);
1832 
1833     bdrv_unref(chain[2]);
1834 }
1835 
1836 
1837 typedef struct BDRVReplaceTestState {
1838     bool was_drained;
1839     bool was_undrained;
1840     bool has_read;
1841 
1842     int drain_count;
1843 
1844     bool yield_before_read;
1845     Coroutine *io_co;
1846     Coroutine *drain_co;
1847 } BDRVReplaceTestState;
1848 
1849 static void bdrv_replace_test_close(BlockDriverState *bs)
1850 {
1851 }
1852 
1853 /**
1854  * If @bs has a backing file:
1855  *   Yield if .yield_before_read is true (and wait for drain_begin to
1856  *   wake us up).
1857  *   Forward the read to bs->backing.  Set .has_read to true.
1858  *   If drain_begin has woken us, wake it in turn.
1859  *
1860  * Otherwise:
1861  *   Set .has_read to true and return success.
1862  */
1863 static int coroutine_fn bdrv_replace_test_co_preadv(BlockDriverState *bs,
1864                                                     int64_t offset,
1865                                                     int64_t bytes,
1866                                                     QEMUIOVector *qiov,
1867                                                     BdrvRequestFlags flags)
1868 {
1869     BDRVReplaceTestState *s = bs->opaque;
1870 
1871     if (bs->backing) {
1872         int ret;
1873 
1874         g_assert(!s->drain_count);
1875 
1876         s->io_co = qemu_coroutine_self();
1877         if (s->yield_before_read) {
1878             s->yield_before_read = false;
1879             qemu_coroutine_yield();
1880         }
1881         s->io_co = NULL;
1882 
1883         ret = bdrv_co_preadv(bs->backing, offset, bytes, qiov, 0);
1884         s->has_read = true;
1885 
1886         /* Wake up drain_co if it runs */
1887         if (s->drain_co) {
1888             aio_co_wake(s->drain_co);
1889         }
1890 
1891         return ret;
1892     }
1893 
1894     s->has_read = true;
1895     return 0;
1896 }
1897 
1898 /**
1899  * If .drain_count is 0, wake up .io_co if there is one; and set
1900  * .was_drained.
1901  * Increment .drain_count.
1902  */
1903 static void coroutine_fn bdrv_replace_test_co_drain_begin(BlockDriverState *bs)
1904 {
1905     BDRVReplaceTestState *s = bs->opaque;
1906 
1907     if (!s->drain_count) {
1908         /* Keep waking io_co up until it is done */
1909         s->drain_co = qemu_coroutine_self();
1910         while (s->io_co) {
1911             aio_co_wake(s->io_co);
1912             s->io_co = NULL;
1913             qemu_coroutine_yield();
1914         }
1915         s->drain_co = NULL;
1916 
1917         s->was_drained = true;
1918     }
1919     s->drain_count++;
1920 }
1921 
1922 /**
1923  * Reduce .drain_count, set .was_undrained once it reaches 0.
1924  * If .drain_count reaches 0 and the node has a backing file, issue a
1925  * read request.
1926  */
1927 static void coroutine_fn bdrv_replace_test_co_drain_end(BlockDriverState *bs)
1928 {
1929     BDRVReplaceTestState *s = bs->opaque;
1930 
1931     g_assert(s->drain_count > 0);
1932     if (!--s->drain_count) {
1933         int ret;
1934 
1935         s->was_undrained = true;
1936 
1937         if (bs->backing) {
1938             char data;
1939             QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
1940 
1941             /* Queue a read request post-drain */
1942             ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
1943             g_assert(ret >= 0);
1944         }
1945     }
1946 }
1947 
1948 static BlockDriver bdrv_replace_test = {
1949     .format_name            = "replace_test",
1950     .instance_size          = sizeof(BDRVReplaceTestState),
1951 
1952     .bdrv_close             = bdrv_replace_test_close,
1953     .bdrv_co_preadv         = bdrv_replace_test_co_preadv,
1954 
1955     .bdrv_co_drain_begin    = bdrv_replace_test_co_drain_begin,
1956     .bdrv_co_drain_end      = bdrv_replace_test_co_drain_end,
1957 
1958     .bdrv_child_perm        = bdrv_default_perms,
1959 };
1960 
1961 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque)
1962 {
1963     int ret;
1964     char data;
1965 
1966     ret = blk_co_pread(opaque, 0, 1, &data, 0);
1967     g_assert(ret >= 0);
1968 }
1969 
1970 /**
1971  * We test two things:
1972  * (1) bdrv_replace_child_noperm() must not undrain the parent if both
1973  *     children are drained.
1974  * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
1975  *     drained child.  If the old child is drained, it must flush I/O
1976  *     requests after the new one has been attached.  If the new child
1977  *     is drained, it must flush I/O requests before the old one is
1978  *     detached.
1979  *
1980  * To do so, we create one parent node and two child nodes; then
1981  * attach one of the children (old_child_bs) to the parent, then
1982  * drain both old_child_bs and new_child_bs according to
1983  * old_drain_count and new_drain_count, respectively, and finally
1984  * we invoke bdrv_replace_node() to replace old_child_bs by
1985  * new_child_bs.
1986  *
1987  * The test block driver we use here (bdrv_replace_test) has a read
1988  * function that:
1989  * - For the parent node, can optionally yield, and then forwards the
1990  *   read to bdrv_preadv(),
1991  * - For the child node, just returns immediately.
1992  *
1993  * If the read yields, the drain_begin function will wake it up.
1994  *
1995  * The drain_end function issues a read on the parent once it is fully
1996  * undrained (which simulates requests starting to come in again).
1997  */
1998 static void do_test_replace_child_mid_drain(int old_drain_count,
1999                                             int new_drain_count)
2000 {
2001     BlockBackend *parent_blk;
2002     BlockDriverState *parent_bs;
2003     BlockDriverState *old_child_bs, *new_child_bs;
2004     BDRVReplaceTestState *parent_s;
2005     BDRVReplaceTestState *old_child_s, *new_child_s;
2006     Coroutine *io_co;
2007     int i;
2008 
2009     parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0,
2010                                      &error_abort);
2011     parent_s = parent_bs->opaque;
2012 
2013     parent_blk = blk_new(qemu_get_aio_context(),
2014                          BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
2015     blk_insert_bs(parent_blk, parent_bs, &error_abort);
2016 
2017     old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0,
2018                                         &error_abort);
2019     new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0,
2020                                         &error_abort);
2021     old_child_s = old_child_bs->opaque;
2022     new_child_s = new_child_bs->opaque;
2023 
2024     /* So that we can read something */
2025     parent_bs->total_sectors = 1;
2026     old_child_bs->total_sectors = 1;
2027     new_child_bs->total_sectors = 1;
2028 
2029     bdrv_ref(old_child_bs);
2030     parent_bs->backing = bdrv_attach_child(parent_bs, old_child_bs, "child",
2031                                            &child_of_bds, BDRV_CHILD_COW,
2032                                            &error_abort);
2033 
2034     for (i = 0; i < old_drain_count; i++) {
2035         bdrv_drained_begin(old_child_bs);
2036     }
2037     for (i = 0; i < new_drain_count; i++) {
2038         bdrv_drained_begin(new_child_bs);
2039     }
2040 
2041     if (!old_drain_count) {
2042         /*
2043          * Start a read operation that will yield, so it will not
2044          * complete before the node is drained.
2045          */
2046         parent_s->yield_before_read = true;
2047         io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co,
2048                                       parent_blk);
2049         qemu_coroutine_enter(io_co);
2050     }
2051 
2052     /* If we have started a read operation, it should have yielded */
2053     g_assert(!parent_s->has_read);
2054 
2055     /* Reset drained status so we can see what bdrv_replace_node() does */
2056     parent_s->was_drained = false;
2057     parent_s->was_undrained = false;
2058 
2059     g_assert(parent_bs->quiesce_counter == old_drain_count);
2060     bdrv_replace_node(old_child_bs, new_child_bs, &error_abort);
2061     g_assert(parent_bs->quiesce_counter == new_drain_count);
2062 
2063     if (!old_drain_count && !new_drain_count) {
2064         /*
2065          * From undrained to undrained drains and undrains the parent,
2066          * because bdrv_replace_node() contains a drained section for
2067          * @old_child_bs.
2068          */
2069         g_assert(parent_s->was_drained && parent_s->was_undrained);
2070     } else if (!old_drain_count && new_drain_count) {
2071         /*
2072          * From undrained to drained should drain the parent and keep
2073          * it that way.
2074          */
2075         g_assert(parent_s->was_drained && !parent_s->was_undrained);
2076     } else if (old_drain_count && !new_drain_count) {
2077         /*
2078          * From drained to undrained should undrain the parent and
2079          * keep it that way.
2080          */
2081         g_assert(!parent_s->was_drained && parent_s->was_undrained);
2082     } else /* if (old_drain_count && new_drain_count) */ {
2083         /*
2084          * From drained to drained must not undrain the parent at any
2085          * point
2086          */
2087         g_assert(!parent_s->was_drained && !parent_s->was_undrained);
2088     }
2089 
2090     if (!old_drain_count || !new_drain_count) {
2091         /*
2092          * If !old_drain_count, we have started a read request before
2093          * bdrv_replace_node().  If !new_drain_count, the parent must
2094          * have been undrained at some point, and
2095          * bdrv_replace_test_co_drain_end() starts a read request
2096          * then.
2097          */
2098         g_assert(parent_s->has_read);
2099     } else {
2100         /*
2101          * If the parent was never undrained, there is no way to start
2102          * a read request.
2103          */
2104         g_assert(!parent_s->has_read);
2105     }
2106 
2107     /* A drained child must have not received any request */
2108     g_assert(!(old_drain_count && old_child_s->has_read));
2109     g_assert(!(new_drain_count && new_child_s->has_read));
2110 
2111     for (i = 0; i < new_drain_count; i++) {
2112         bdrv_drained_end(new_child_bs);
2113     }
2114     for (i = 0; i < old_drain_count; i++) {
2115         bdrv_drained_end(old_child_bs);
2116     }
2117 
2118     /*
2119      * By now, bdrv_replace_test_co_drain_end() must have been called
2120      * at some point while the new child was attached to the parent.
2121      */
2122     g_assert(parent_s->has_read);
2123     g_assert(new_child_s->has_read);
2124 
2125     blk_unref(parent_blk);
2126     bdrv_unref(parent_bs);
2127     bdrv_unref(old_child_bs);
2128     bdrv_unref(new_child_bs);
2129 }
2130 
2131 static void test_replace_child_mid_drain(void)
2132 {
2133     int old_drain_count, new_drain_count;
2134 
2135     for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) {
2136         for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) {
2137             do_test_replace_child_mid_drain(old_drain_count, new_drain_count);
2138         }
2139     }
2140 }
2141 
2142 int main(int argc, char **argv)
2143 {
2144     int ret;
2145 
2146     bdrv_init();
2147     qemu_init_main_loop(&error_abort);
2148 
2149     g_test_init(&argc, &argv, NULL);
2150     qemu_event_init(&done_event, false);
2151 
2152     g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all);
2153     g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain);
2154     g_test_add_func("/bdrv-drain/driver-cb/drain_subtree",
2155                     test_drv_cb_drain_subtree);
2156 
2157     g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
2158                     test_drv_cb_co_drain_all);
2159     g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain);
2160     g_test_add_func("/bdrv-drain/driver-cb/co/drain_subtree",
2161                     test_drv_cb_co_drain_subtree);
2162 
2163 
2164     g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all);
2165     g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain);
2166     g_test_add_func("/bdrv-drain/quiesce/drain_subtree",
2167                     test_quiesce_drain_subtree);
2168 
2169     g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
2170                     test_quiesce_co_drain_all);
2171     g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain);
2172     g_test_add_func("/bdrv-drain/quiesce/co/drain_subtree",
2173                     test_quiesce_co_drain_subtree);
2174 
2175     g_test_add_func("/bdrv-drain/nested", test_nested);
2176     g_test_add_func("/bdrv-drain/multiparent", test_multiparent);
2177 
2178     g_test_add_func("/bdrv-drain/graph-change/drain_subtree",
2179                     test_graph_change_drain_subtree);
2180     g_test_add_func("/bdrv-drain/graph-change/drain_all",
2181                     test_graph_change_drain_all);
2182 
2183     g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all);
2184     g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain);
2185     g_test_add_func("/bdrv-drain/iothread/drain_subtree",
2186                     test_iothread_drain_subtree);
2187 
2188     g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all);
2189     g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain);
2190     g_test_add_func("/bdrv-drain/blockjob/drain_subtree",
2191                     test_blockjob_drain_subtree);
2192 
2193     g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
2194                     test_blockjob_error_drain_all);
2195     g_test_add_func("/bdrv-drain/blockjob/error/drain",
2196                     test_blockjob_error_drain);
2197     g_test_add_func("/bdrv-drain/blockjob/error/drain_subtree",
2198                     test_blockjob_error_drain_subtree);
2199 
2200     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
2201                     test_blockjob_iothread_drain_all);
2202     g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
2203                     test_blockjob_iothread_drain);
2204     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_subtree",
2205                     test_blockjob_iothread_drain_subtree);
2206 
2207     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
2208                     test_blockjob_iothread_error_drain_all);
2209     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
2210                     test_blockjob_iothread_error_drain);
2211     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_subtree",
2212                     test_blockjob_iothread_error_drain_subtree);
2213 
2214     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
2215     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
2216     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
2217     g_test_add_func("/bdrv-drain/detach/drain_subtree", test_detach_by_drain_subtree);
2218     g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb);
2219     g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb);
2220 
2221     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
2222 
2223     g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context);
2224 
2225     g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
2226                     test_blockjob_commit_by_drained_end);
2227 
2228     g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
2229                     test_drop_intermediate_poll);
2230 
2231     g_test_add_func("/bdrv-drain/replace_child/mid-drain",
2232                     test_replace_child_mid_drain);
2233 
2234     ret = g_test_run();
2235     qemu_event_destroy(&done_event);
2236     return ret;
2237 }
2238