xref: /openbmc/qemu/tests/unit/test-bdrv-drain.c (revision 701bff24)
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     aio_context_release(ctx);
934 
935     job_start(&job->job);
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     WITH_JOB_LOCK_GUARD() {
947         g_assert_cmpint(job->job.pause_count, ==, 0);
948         g_assert_false(job->job.paused);
949         g_assert_true(tjob->running);
950         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
951     }
952 
953     do_drain_begin_unlocked(drain_type, drain_bs);
954 
955     WITH_JOB_LOCK_GUARD() {
956         if (drain_type == BDRV_DRAIN_ALL) {
957             /* bdrv_drain_all() drains both src and target */
958             g_assert_cmpint(job->job.pause_count, ==, 2);
959         } else {
960             g_assert_cmpint(job->job.pause_count, ==, 1);
961         }
962         g_assert_true(job->job.paused);
963         g_assert_false(job->job.busy); /* The job is paused */
964     }
965 
966     do_drain_end_unlocked(drain_type, drain_bs);
967 
968     if (use_iothread) {
969         /*
970          * Here we are waiting for the paused status to change,
971          * so don't bother protecting the read every time.
972          *
973          * paused is reset in the I/O thread, wait for it
974          */
975         while (job->job.paused) {
976             aio_poll(qemu_get_aio_context(), false);
977         }
978     }
979 
980     WITH_JOB_LOCK_GUARD() {
981         g_assert_cmpint(job->job.pause_count, ==, 0);
982         g_assert_false(job->job.paused);
983         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
984     }
985 
986     do_drain_begin_unlocked(drain_type, target);
987 
988     WITH_JOB_LOCK_GUARD() {
989         if (drain_type == BDRV_DRAIN_ALL) {
990             /* bdrv_drain_all() drains both src and target */
991             g_assert_cmpint(job->job.pause_count, ==, 2);
992         } else {
993             g_assert_cmpint(job->job.pause_count, ==, 1);
994         }
995         g_assert_true(job->job.paused);
996         g_assert_false(job->job.busy); /* The job is paused */
997     }
998 
999     do_drain_end_unlocked(drain_type, target);
1000 
1001     if (use_iothread) {
1002         /*
1003          * Here we are waiting for the paused status to change,
1004          * so don't bother protecting the read every time.
1005          *
1006          * paused is reset in the I/O thread, wait for it
1007          */
1008         while (job->job.paused) {
1009             aio_poll(qemu_get_aio_context(), false);
1010         }
1011     }
1012 
1013     WITH_JOB_LOCK_GUARD() {
1014         g_assert_cmpint(job->job.pause_count, ==, 0);
1015         g_assert_false(job->job.paused);
1016         g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
1017     }
1018 
1019     WITH_JOB_LOCK_GUARD() {
1020         ret = job_complete_sync_locked(&job->job, &error_abort);
1021     }
1022     g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO));
1023 
1024     aio_context_acquire(ctx);
1025     if (use_iothread) {
1026         blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort);
1027         assert(blk_get_aio_context(blk_target) == qemu_get_aio_context());
1028     }
1029     aio_context_release(ctx);
1030 
1031     blk_unref(blk_src);
1032     blk_unref(blk_target);
1033     bdrv_unref(src_overlay);
1034     bdrv_unref(target);
1035 
1036     if (iothread) {
1037         iothread_join(iothread);
1038     }
1039 }
1040 
1041 static void test_blockjob_common(enum drain_type drain_type, bool use_iothread,
1042                                  enum test_job_result result)
1043 {
1044     test_blockjob_common_drain_node(drain_type, use_iothread, result,
1045                                     TEST_JOB_DRAIN_SRC);
1046     test_blockjob_common_drain_node(drain_type, use_iothread, result,
1047                                     TEST_JOB_DRAIN_SRC_CHILD);
1048     if (drain_type == BDRV_SUBTREE_DRAIN) {
1049         test_blockjob_common_drain_node(drain_type, use_iothread, result,
1050                                         TEST_JOB_DRAIN_SRC_PARENT);
1051     }
1052 }
1053 
1054 static void test_blockjob_drain_all(void)
1055 {
1056     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS);
1057 }
1058 
1059 static void test_blockjob_drain(void)
1060 {
1061     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS);
1062 }
1063 
1064 static void test_blockjob_drain_subtree(void)
1065 {
1066     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_SUCCESS);
1067 }
1068 
1069 static void test_blockjob_error_drain_all(void)
1070 {
1071     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN);
1072     test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE);
1073 }
1074 
1075 static void test_blockjob_error_drain(void)
1076 {
1077     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN);
1078     test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE);
1079 }
1080 
1081 static void test_blockjob_error_drain_subtree(void)
1082 {
1083     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_RUN);
1084     test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_PREPARE);
1085 }
1086 
1087 static void test_blockjob_iothread_drain_all(void)
1088 {
1089     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS);
1090 }
1091 
1092 static void test_blockjob_iothread_drain(void)
1093 {
1094     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS);
1095 }
1096 
1097 static void test_blockjob_iothread_drain_subtree(void)
1098 {
1099     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_SUCCESS);
1100 }
1101 
1102 static void test_blockjob_iothread_error_drain_all(void)
1103 {
1104     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN);
1105     test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE);
1106 }
1107 
1108 static void test_blockjob_iothread_error_drain(void)
1109 {
1110     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN);
1111     test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE);
1112 }
1113 
1114 static void test_blockjob_iothread_error_drain_subtree(void)
1115 {
1116     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_RUN);
1117     test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_PREPARE);
1118 }
1119 
1120 
1121 typedef struct BDRVTestTopState {
1122     BdrvChild *wait_child;
1123 } BDRVTestTopState;
1124 
1125 static void bdrv_test_top_close(BlockDriverState *bs)
1126 {
1127     BdrvChild *c, *next_c;
1128     QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1129         bdrv_unref_child(bs, c);
1130     }
1131 }
1132 
1133 static int coroutine_fn bdrv_test_top_co_preadv(BlockDriverState *bs,
1134                                                 int64_t offset, int64_t bytes,
1135                                                 QEMUIOVector *qiov,
1136                                                 BdrvRequestFlags flags)
1137 {
1138     BDRVTestTopState *tts = bs->opaque;
1139     return bdrv_co_preadv(tts->wait_child, offset, bytes, qiov, flags);
1140 }
1141 
1142 static BlockDriver bdrv_test_top_driver = {
1143     .format_name            = "test_top_driver",
1144     .instance_size          = sizeof(BDRVTestTopState),
1145 
1146     .bdrv_close             = bdrv_test_top_close,
1147     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
1148 
1149     .bdrv_child_perm        = bdrv_default_perms,
1150 };
1151 
1152 typedef struct TestCoDeleteByDrainData {
1153     BlockBackend *blk;
1154     bool detach_instead_of_delete;
1155     bool done;
1156 } TestCoDeleteByDrainData;
1157 
1158 static void coroutine_fn test_co_delete_by_drain(void *opaque)
1159 {
1160     TestCoDeleteByDrainData *dbdd = opaque;
1161     BlockBackend *blk = dbdd->blk;
1162     BlockDriverState *bs = blk_bs(blk);
1163     BDRVTestTopState *tts = bs->opaque;
1164     void *buffer = g_malloc(65536);
1165     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536);
1166 
1167     /* Pretend some internal write operation from parent to child.
1168      * Important: We have to read from the child, not from the parent!
1169      * Draining works by first propagating it all up the tree to the
1170      * root and then waiting for drainage from root to the leaves
1171      * (protocol nodes).  If we have a request waiting on the root,
1172      * everything will be drained before we go back down the tree, but
1173      * we do not want that.  We want to be in the middle of draining
1174      * when this following requests returns. */
1175     bdrv_co_preadv(tts->wait_child, 0, 65536, &qiov, 0);
1176 
1177     g_assert_cmpint(bs->refcnt, ==, 1);
1178 
1179     if (!dbdd->detach_instead_of_delete) {
1180         blk_unref(blk);
1181     } else {
1182         BdrvChild *c, *next_c;
1183         QLIST_FOREACH_SAFE(c, &bs->children, next, next_c) {
1184             bdrv_unref_child(bs, c);
1185         }
1186     }
1187 
1188     dbdd->done = true;
1189     g_free(buffer);
1190 }
1191 
1192 /**
1193  * Test what happens when some BDS has some children, you drain one of
1194  * them and this results in the BDS being deleted.
1195  *
1196  * If @detach_instead_of_delete is set, the BDS is not going to be
1197  * deleted but will only detach all of its children.
1198  */
1199 static void do_test_delete_by_drain(bool detach_instead_of_delete,
1200                                     enum drain_type drain_type)
1201 {
1202     BlockBackend *blk;
1203     BlockDriverState *bs, *child_bs, *null_bs;
1204     BDRVTestTopState *tts;
1205     TestCoDeleteByDrainData dbdd;
1206     Coroutine *co;
1207 
1208     bs = bdrv_new_open_driver(&bdrv_test_top_driver, "top", BDRV_O_RDWR,
1209                               &error_abort);
1210     bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1211     tts = bs->opaque;
1212 
1213     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1214                         &error_abort);
1215     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds,
1216                       BDRV_CHILD_DATA, &error_abort);
1217 
1218     /* This child will be the one to pass to requests through to, and
1219      * it will stall until a drain occurs */
1220     child_bs = bdrv_new_open_driver(&bdrv_test, "child", BDRV_O_RDWR,
1221                                     &error_abort);
1222     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
1223     /* Takes our reference to child_bs */
1224     tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child",
1225                                         &child_of_bds,
1226                                         BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY,
1227                                         &error_abort);
1228 
1229     /* This child is just there to be deleted
1230      * (for detach_instead_of_delete == true) */
1231     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
1232                         &error_abort);
1233     bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA,
1234                       &error_abort);
1235 
1236     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1237     blk_insert_bs(blk, bs, &error_abort);
1238 
1239     /* Referenced by blk now */
1240     bdrv_unref(bs);
1241 
1242     g_assert_cmpint(bs->refcnt, ==, 1);
1243     g_assert_cmpint(child_bs->refcnt, ==, 1);
1244     g_assert_cmpint(null_bs->refcnt, ==, 1);
1245 
1246 
1247     dbdd = (TestCoDeleteByDrainData){
1248         .blk = blk,
1249         .detach_instead_of_delete = detach_instead_of_delete,
1250         .done = false,
1251     };
1252     co = qemu_coroutine_create(test_co_delete_by_drain, &dbdd);
1253     qemu_coroutine_enter(co);
1254 
1255     /* Drain the child while the read operation is still pending.
1256      * This should result in the operation finishing and
1257      * test_co_delete_by_drain() resuming.  Thus, @bs will be deleted
1258      * and the coroutine will exit while this drain operation is still
1259      * in progress. */
1260     switch (drain_type) {
1261     case BDRV_DRAIN:
1262         bdrv_ref(child_bs);
1263         bdrv_drain(child_bs);
1264         bdrv_unref(child_bs);
1265         break;
1266     case BDRV_SUBTREE_DRAIN:
1267         /* Would have to ref/unref bs here for !detach_instead_of_delete, but
1268          * then the whole test becomes pointless because the graph changes
1269          * don't occur during the drain any more. */
1270         assert(detach_instead_of_delete);
1271         bdrv_subtree_drained_begin(bs);
1272         bdrv_subtree_drained_end(bs);
1273         break;
1274     case BDRV_DRAIN_ALL:
1275         bdrv_drain_all_begin();
1276         bdrv_drain_all_end();
1277         break;
1278     default:
1279         g_assert_not_reached();
1280     }
1281 
1282     while (!dbdd.done) {
1283         aio_poll(qemu_get_aio_context(), true);
1284     }
1285 
1286     if (detach_instead_of_delete) {
1287         /* Here, the reference has not passed over to the coroutine,
1288          * so we have to delete the BB ourselves */
1289         blk_unref(blk);
1290     }
1291 }
1292 
1293 static void test_delete_by_drain(void)
1294 {
1295     do_test_delete_by_drain(false, BDRV_DRAIN);
1296 }
1297 
1298 static void test_detach_by_drain_all(void)
1299 {
1300     do_test_delete_by_drain(true, BDRV_DRAIN_ALL);
1301 }
1302 
1303 static void test_detach_by_drain(void)
1304 {
1305     do_test_delete_by_drain(true, BDRV_DRAIN);
1306 }
1307 
1308 static void test_detach_by_drain_subtree(void)
1309 {
1310     do_test_delete_by_drain(true, BDRV_SUBTREE_DRAIN);
1311 }
1312 
1313 
1314 struct detach_by_parent_data {
1315     BlockDriverState *parent_b;
1316     BdrvChild *child_b;
1317     BlockDriverState *c;
1318     BdrvChild *child_c;
1319     bool by_parent_cb;
1320 };
1321 static struct detach_by_parent_data detach_by_parent_data;
1322 
1323 static void detach_indirect_bh(void *opaque)
1324 {
1325     struct detach_by_parent_data *data = opaque;
1326 
1327     bdrv_unref_child(data->parent_b, data->child_b);
1328 
1329     bdrv_ref(data->c);
1330     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
1331                                       &child_of_bds, BDRV_CHILD_DATA,
1332                                       &error_abort);
1333 }
1334 
1335 static void detach_by_parent_aio_cb(void *opaque, int ret)
1336 {
1337     struct detach_by_parent_data *data = &detach_by_parent_data;
1338 
1339     g_assert_cmpint(ret, ==, 0);
1340     if (data->by_parent_cb) {
1341         detach_indirect_bh(data);
1342     }
1343 }
1344 
1345 static void detach_by_driver_cb_drained_begin(BdrvChild *child)
1346 {
1347     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
1348                             detach_indirect_bh, &detach_by_parent_data);
1349     child_of_bds.drained_begin(child);
1350 }
1351 
1352 static BdrvChildClass detach_by_driver_cb_class;
1353 
1354 /*
1355  * Initial graph:
1356  *
1357  * PA     PB
1358  *    \ /   \
1359  *     A     B     C
1360  *
1361  * by_parent_cb == true:  Test that parent callbacks don't poll
1362  *
1363  *     PA has a pending write request whose callback changes the child nodes of
1364  *     PB: It removes B and adds C instead. The subtree of PB is drained, which
1365  *     will indirectly drain the write request, too.
1366  *
1367  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
1368  *
1369  *     PA's BdrvChildClass has a .drained_begin callback that schedules a BH
1370  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
1371  *     state is messed up, but if it is only polled in the single
1372  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
1373  */
1374 static void test_detach_indirect(bool by_parent_cb)
1375 {
1376     BlockBackend *blk;
1377     BlockDriverState *parent_a, *parent_b, *a, *b, *c;
1378     BdrvChild *child_a, *child_b;
1379     BlockAIOCB *acb;
1380 
1381     QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
1382 
1383     if (!by_parent_cb) {
1384         detach_by_driver_cb_class = child_of_bds;
1385         detach_by_driver_cb_class.drained_begin =
1386             detach_by_driver_cb_drained_begin;
1387     }
1388 
1389     /* Create all involved nodes */
1390     parent_a = bdrv_new_open_driver(&bdrv_test, "parent-a", BDRV_O_RDWR,
1391                                     &error_abort);
1392     parent_b = bdrv_new_open_driver(&bdrv_test, "parent-b", 0,
1393                                     &error_abort);
1394 
1395     a = bdrv_new_open_driver(&bdrv_test, "a", BDRV_O_RDWR, &error_abort);
1396     b = bdrv_new_open_driver(&bdrv_test, "b", BDRV_O_RDWR, &error_abort);
1397     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
1398 
1399     /* blk is a BB for parent-a */
1400     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1401     blk_insert_bs(blk, parent_a, &error_abort);
1402     bdrv_unref(parent_a);
1403 
1404     /* If we want to get bdrv_drain_invoke() to call aio_poll(), the driver
1405      * callback must not return immediately. */
1406     if (!by_parent_cb) {
1407         BDRVTestState *s = parent_a->opaque;
1408         s->sleep_in_drain_begin = true;
1409     }
1410 
1411     /* Set child relationships */
1412     bdrv_ref(b);
1413     bdrv_ref(a);
1414     child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds,
1415                                 BDRV_CHILD_DATA, &error_abort);
1416     child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds,
1417                                 BDRV_CHILD_COW, &error_abort);
1418 
1419     bdrv_ref(a);
1420     bdrv_attach_child(parent_a, a, "PA-A",
1421                       by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class,
1422                       BDRV_CHILD_DATA, &error_abort);
1423 
1424     g_assert_cmpint(parent_a->refcnt, ==, 1);
1425     g_assert_cmpint(parent_b->refcnt, ==, 1);
1426     g_assert_cmpint(a->refcnt, ==, 3);
1427     g_assert_cmpint(b->refcnt, ==, 2);
1428     g_assert_cmpint(c->refcnt, ==, 1);
1429 
1430     g_assert(QLIST_FIRST(&parent_b->children) == child_a);
1431     g_assert(QLIST_NEXT(child_a, next) == child_b);
1432     g_assert(QLIST_NEXT(child_b, next) == NULL);
1433 
1434     /* Start the evil write request */
1435     detach_by_parent_data = (struct detach_by_parent_data) {
1436         .parent_b = parent_b,
1437         .child_b = child_b,
1438         .c = c,
1439         .by_parent_cb = by_parent_cb,
1440     };
1441     acb = blk_aio_preadv(blk, 0, &qiov, 0, detach_by_parent_aio_cb, NULL);
1442     g_assert(acb != NULL);
1443 
1444     /* Drain and check the expected result */
1445     bdrv_subtree_drained_begin(parent_b);
1446 
1447     g_assert(detach_by_parent_data.child_c != NULL);
1448 
1449     g_assert_cmpint(parent_a->refcnt, ==, 1);
1450     g_assert_cmpint(parent_b->refcnt, ==, 1);
1451     g_assert_cmpint(a->refcnt, ==, 3);
1452     g_assert_cmpint(b->refcnt, ==, 1);
1453     g_assert_cmpint(c->refcnt, ==, 2);
1454 
1455     g_assert(QLIST_FIRST(&parent_b->children) == detach_by_parent_data.child_c);
1456     g_assert(QLIST_NEXT(detach_by_parent_data.child_c, next) == child_a);
1457     g_assert(QLIST_NEXT(child_a, next) == NULL);
1458 
1459     g_assert_cmpint(parent_a->quiesce_counter, ==, 1);
1460     g_assert_cmpint(parent_b->quiesce_counter, ==, 1);
1461     g_assert_cmpint(a->quiesce_counter, ==, 1);
1462     g_assert_cmpint(b->quiesce_counter, ==, 0);
1463     g_assert_cmpint(c->quiesce_counter, ==, 1);
1464 
1465     bdrv_subtree_drained_end(parent_b);
1466 
1467     bdrv_unref(parent_b);
1468     blk_unref(blk);
1469 
1470     g_assert_cmpint(a->refcnt, ==, 1);
1471     g_assert_cmpint(b->refcnt, ==, 1);
1472     g_assert_cmpint(c->refcnt, ==, 1);
1473     bdrv_unref(a);
1474     bdrv_unref(b);
1475     bdrv_unref(c);
1476 }
1477 
1478 static void test_detach_by_parent_cb(void)
1479 {
1480     test_detach_indirect(true);
1481 }
1482 
1483 static void test_detach_by_driver_cb(void)
1484 {
1485     test_detach_indirect(false);
1486 }
1487 
1488 static void test_append_to_drained(void)
1489 {
1490     BlockBackend *blk;
1491     BlockDriverState *base, *overlay;
1492     BDRVTestState *base_s, *overlay_s;
1493 
1494     blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
1495     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
1496     base_s = base->opaque;
1497     blk_insert_bs(blk, base, &error_abort);
1498 
1499     overlay = bdrv_new_open_driver(&bdrv_test, "overlay", BDRV_O_RDWR,
1500                                    &error_abort);
1501     overlay_s = overlay->opaque;
1502 
1503     do_drain_begin(BDRV_DRAIN, base);
1504     g_assert_cmpint(base->quiesce_counter, ==, 1);
1505     g_assert_cmpint(base_s->drain_count, ==, 1);
1506     g_assert_cmpint(base->in_flight, ==, 0);
1507 
1508     bdrv_append(overlay, base, &error_abort);
1509     g_assert_cmpint(base->in_flight, ==, 0);
1510     g_assert_cmpint(overlay->in_flight, ==, 0);
1511 
1512     g_assert_cmpint(base->quiesce_counter, ==, 1);
1513     g_assert_cmpint(base_s->drain_count, ==, 1);
1514     g_assert_cmpint(overlay->quiesce_counter, ==, 1);
1515     g_assert_cmpint(overlay_s->drain_count, ==, 1);
1516 
1517     do_drain_end(BDRV_DRAIN, base);
1518 
1519     g_assert_cmpint(base->quiesce_counter, ==, 0);
1520     g_assert_cmpint(base_s->drain_count, ==, 0);
1521     g_assert_cmpint(overlay->quiesce_counter, ==, 0);
1522     g_assert_cmpint(overlay_s->drain_count, ==, 0);
1523 
1524     bdrv_unref(overlay);
1525     bdrv_unref(base);
1526     blk_unref(blk);
1527 }
1528 
1529 static void test_set_aio_context(void)
1530 {
1531     BlockDriverState *bs;
1532     IOThread *a = iothread_new();
1533     IOThread *b = iothread_new();
1534     AioContext *ctx_a = iothread_get_aio_context(a);
1535     AioContext *ctx_b = iothread_get_aio_context(b);
1536 
1537     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
1538                               &error_abort);
1539 
1540     bdrv_drained_begin(bs);
1541     bdrv_try_set_aio_context(bs, ctx_a, &error_abort);
1542 
1543     aio_context_acquire(ctx_a);
1544     bdrv_drained_end(bs);
1545 
1546     bdrv_drained_begin(bs);
1547     bdrv_try_set_aio_context(bs, ctx_b, &error_abort);
1548     aio_context_release(ctx_a);
1549     aio_context_acquire(ctx_b);
1550     bdrv_try_set_aio_context(bs, qemu_get_aio_context(), &error_abort);
1551     aio_context_release(ctx_b);
1552     bdrv_drained_end(bs);
1553 
1554     bdrv_unref(bs);
1555     iothread_join(a);
1556     iothread_join(b);
1557 }
1558 
1559 
1560 typedef struct TestDropBackingBlockJob {
1561     BlockJob common;
1562     bool should_complete;
1563     bool *did_complete;
1564     BlockDriverState *detach_also;
1565     BlockDriverState *bs;
1566 } TestDropBackingBlockJob;
1567 
1568 static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp)
1569 {
1570     TestDropBackingBlockJob *s =
1571         container_of(job, TestDropBackingBlockJob, common.job);
1572 
1573     while (!s->should_complete) {
1574         job_sleep_ns(job, 0);
1575     }
1576 
1577     return 0;
1578 }
1579 
1580 static void test_drop_backing_job_commit(Job *job)
1581 {
1582     TestDropBackingBlockJob *s =
1583         container_of(job, TestDropBackingBlockJob, common.job);
1584 
1585     bdrv_set_backing_hd(s->bs, NULL, &error_abort);
1586     bdrv_set_backing_hd(s->detach_also, NULL, &error_abort);
1587 
1588     *s->did_complete = true;
1589 }
1590 
1591 static const BlockJobDriver test_drop_backing_job_driver = {
1592     .job_driver = {
1593         .instance_size  = sizeof(TestDropBackingBlockJob),
1594         .free           = block_job_free,
1595         .user_resume    = block_job_user_resume,
1596         .run            = test_drop_backing_job_run,
1597         .commit         = test_drop_backing_job_commit,
1598     }
1599 };
1600 
1601 /**
1602  * Creates a child node with three parent nodes on it, and then runs a
1603  * block job on the final one, parent-node-2.
1604  *
1605  * The job is then asked to complete before a section where the child
1606  * is drained.
1607  *
1608  * Ending this section will undrain the child's parents, first
1609  * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
1610  * list is in reverse order of how they were added.  Ending the drain
1611  * on parent-node-2 will resume the job, thus completing it and
1612  * scheduling job_exit().
1613  *
1614  * Ending the drain on parent-node-1 will poll the AioContext, which
1615  * lets job_exit() and thus test_drop_backing_job_commit() run.  That
1616  * function first removes the child as parent-node-2's backing file.
1617  *
1618  * In old (and buggy) implementations, there are two problems with
1619  * that:
1620  * (A) bdrv_drain_invoke() polls for every node that leaves the
1621  *     drained section.  This means that job_exit() is scheduled
1622  *     before the child has left the drained section.  Its
1623  *     quiesce_counter is therefore still 1 when it is removed from
1624  *     parent-node-2.
1625  *
1626  * (B) bdrv_replace_child_noperm() calls drained_end() on the old
1627  *     child's parents as many times as the child is quiesced.  This
1628  *     means it will call drained_end() on parent-node-2 once.
1629  *     Because parent-node-2 is no longer quiesced at this point, this
1630  *     will fail.
1631  *
1632  * bdrv_replace_child_noperm() therefore must call drained_end() on
1633  * the parent only if it really is still drained because the child is
1634  * drained.
1635  *
1636  * If removing child from parent-node-2 was successful (as it should
1637  * be), test_drop_backing_job_commit() will then also remove the child
1638  * from parent-node-0.
1639  *
1640  * With an old version of our drain infrastructure ((A) above), that
1641  * resulted in the following flow:
1642  *
1643  * 1. child attempts to leave its drained section.  The call recurses
1644  *    to its parents.
1645  *
1646  * 2. parent-node-2 leaves the drained section.  Polling in
1647  *    bdrv_drain_invoke() will schedule job_exit().
1648  *
1649  * 3. parent-node-1 leaves the drained section.  Polling in
1650  *    bdrv_drain_invoke() will run job_exit(), thus disconnecting
1651  *    parent-node-0 from the child node.
1652  *
1653  * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
1654  *    iterate over the parents.  Thus, it now accesses the BdrvChild
1655  *    object that used to connect parent-node-0 and the child node.
1656  *    However, that object no longer exists, so it accesses a dangling
1657  *    pointer.
1658  *
1659  * The solution is to only poll once when running a bdrv_drained_end()
1660  * operation, specifically at the end when all drained_end()
1661  * operations for all involved nodes have been scheduled.
1662  * Note that this also solves (A) above, thus hiding (B).
1663  */
1664 static void test_blockjob_commit_by_drained_end(void)
1665 {
1666     BlockDriverState *bs_child, *bs_parents[3];
1667     TestDropBackingBlockJob *job;
1668     bool job_has_completed = false;
1669     int i;
1670 
1671     bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR,
1672                                     &error_abort);
1673 
1674     for (i = 0; i < 3; i++) {
1675         char name[32];
1676         snprintf(name, sizeof(name), "parent-node-%i", i);
1677         bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR,
1678                                              &error_abort);
1679         bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort);
1680     }
1681 
1682     job = block_job_create("job", &test_drop_backing_job_driver, NULL,
1683                            bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL,
1684                            &error_abort);
1685     job->bs = bs_parents[2];
1686 
1687     job->detach_also = bs_parents[0];
1688     job->did_complete = &job_has_completed;
1689 
1690     job_start(&job->common.job);
1691 
1692     job->should_complete = true;
1693     bdrv_drained_begin(bs_child);
1694     g_assert(!job_has_completed);
1695     bdrv_drained_end(bs_child);
1696     g_assert(job_has_completed);
1697 
1698     bdrv_unref(bs_parents[0]);
1699     bdrv_unref(bs_parents[1]);
1700     bdrv_unref(bs_parents[2]);
1701     bdrv_unref(bs_child);
1702 }
1703 
1704 
1705 typedef struct TestSimpleBlockJob {
1706     BlockJob common;
1707     bool should_complete;
1708     bool *did_complete;
1709 } TestSimpleBlockJob;
1710 
1711 static int coroutine_fn test_simple_job_run(Job *job, Error **errp)
1712 {
1713     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1714 
1715     while (!s->should_complete) {
1716         job_sleep_ns(job, 0);
1717     }
1718 
1719     return 0;
1720 }
1721 
1722 static void test_simple_job_clean(Job *job)
1723 {
1724     TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
1725     *s->did_complete = true;
1726 }
1727 
1728 static const BlockJobDriver test_simple_job_driver = {
1729     .job_driver = {
1730         .instance_size  = sizeof(TestSimpleBlockJob),
1731         .free           = block_job_free,
1732         .user_resume    = block_job_user_resume,
1733         .run            = test_simple_job_run,
1734         .clean          = test_simple_job_clean,
1735     },
1736 };
1737 
1738 static int drop_intermediate_poll_update_filename(BdrvChild *child,
1739                                                   BlockDriverState *new_base,
1740                                                   const char *filename,
1741                                                   Error **errp)
1742 {
1743     /*
1744      * We are free to poll here, which may change the block graph, if
1745      * it is not drained.
1746      */
1747 
1748     /* If the job is not drained: Complete it, schedule job_exit() */
1749     aio_poll(qemu_get_current_aio_context(), false);
1750     /* If the job is not drained: Run job_exit(), finish the job */
1751     aio_poll(qemu_get_current_aio_context(), false);
1752 
1753     return 0;
1754 }
1755 
1756 /**
1757  * Test a poll in the midst of bdrv_drop_intermediate().
1758  *
1759  * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(),
1760  * which can yield or poll.  This may lead to graph changes, unless
1761  * the whole subtree in question is drained.
1762  *
1763  * We test this on the following graph:
1764  *
1765  *                    Job
1766  *
1767  *                     |
1768  *                  job-node
1769  *                     |
1770  *                     v
1771  *
1772  *                  job-node
1773  *
1774  *                     |
1775  *                  backing
1776  *                     |
1777  *                     v
1778  *
1779  * node-2 --chain--> node-1 --chain--> node-0
1780  *
1781  * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0).
1782  *
1783  * This first updates node-2's backing filename by invoking
1784  * drop_intermediate_poll_update_filename(), which polls twice.  This
1785  * causes the job to finish, which in turns causes the job-node to be
1786  * deleted.
1787  *
1788  * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it
1789  * already has a pointer to the BdrvChild edge between job-node and
1790  * node-1.  When it tries to handle that edge, we probably get a
1791  * segmentation fault because the object no longer exists.
1792  *
1793  *
1794  * The solution is for bdrv_drop_intermediate() to drain top's
1795  * subtree.  This prevents graph changes from happening just because
1796  * BdrvChildClass.update_filename() yields or polls.  Thus, the block
1797  * job is paused during that drained section and must finish before or
1798  * after.
1799  *
1800  * (In addition, bdrv_replace_child() must keep the job paused.)
1801  */
1802 static void test_drop_intermediate_poll(void)
1803 {
1804     static BdrvChildClass chain_child_class;
1805     BlockDriverState *chain[3];
1806     TestSimpleBlockJob *job;
1807     BlockDriverState *job_node;
1808     bool job_has_completed = false;
1809     int i;
1810     int ret;
1811 
1812     chain_child_class = child_of_bds;
1813     chain_child_class.update_filename = drop_intermediate_poll_update_filename;
1814 
1815     for (i = 0; i < 3; i++) {
1816         char name[32];
1817         snprintf(name, 32, "node-%i", i);
1818 
1819         chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort);
1820     }
1821 
1822     job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR,
1823                                     &error_abort);
1824     bdrv_set_backing_hd(job_node, chain[1], &error_abort);
1825 
1826     /*
1827      * Establish the chain last, so the chain links are the first
1828      * elements in the BDS.parents lists
1829      */
1830     for (i = 0; i < 3; i++) {
1831         if (i) {
1832             /* Takes the reference to chain[i - 1] */
1833             chain[i]->backing = bdrv_attach_child(chain[i], chain[i - 1],
1834                                                   "chain", &chain_child_class,
1835                                                   BDRV_CHILD_COW, &error_abort);
1836         }
1837     }
1838 
1839     job = block_job_create("job", &test_simple_job_driver, NULL, job_node,
1840                            0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort);
1841 
1842     /* The job has a reference now */
1843     bdrv_unref(job_node);
1844 
1845     job->did_complete = &job_has_completed;
1846 
1847     job_start(&job->common.job);
1848     job->should_complete = true;
1849 
1850     g_assert(!job_has_completed);
1851     ret = bdrv_drop_intermediate(chain[1], chain[0], NULL);
1852     g_assert(ret == 0);
1853     g_assert(job_has_completed);
1854 
1855     bdrv_unref(chain[2]);
1856 }
1857 
1858 
1859 typedef struct BDRVReplaceTestState {
1860     bool was_drained;
1861     bool was_undrained;
1862     bool has_read;
1863 
1864     int drain_count;
1865 
1866     bool yield_before_read;
1867     Coroutine *io_co;
1868     Coroutine *drain_co;
1869 } BDRVReplaceTestState;
1870 
1871 static void bdrv_replace_test_close(BlockDriverState *bs)
1872 {
1873 }
1874 
1875 /**
1876  * If @bs has a backing file:
1877  *   Yield if .yield_before_read is true (and wait for drain_begin to
1878  *   wake us up).
1879  *   Forward the read to bs->backing.  Set .has_read to true.
1880  *   If drain_begin has woken us, wake it in turn.
1881  *
1882  * Otherwise:
1883  *   Set .has_read to true and return success.
1884  */
1885 static int coroutine_fn bdrv_replace_test_co_preadv(BlockDriverState *bs,
1886                                                     int64_t offset,
1887                                                     int64_t bytes,
1888                                                     QEMUIOVector *qiov,
1889                                                     BdrvRequestFlags flags)
1890 {
1891     BDRVReplaceTestState *s = bs->opaque;
1892 
1893     if (bs->backing) {
1894         int ret;
1895 
1896         g_assert(!s->drain_count);
1897 
1898         s->io_co = qemu_coroutine_self();
1899         if (s->yield_before_read) {
1900             s->yield_before_read = false;
1901             qemu_coroutine_yield();
1902         }
1903         s->io_co = NULL;
1904 
1905         ret = bdrv_co_preadv(bs->backing, offset, bytes, qiov, 0);
1906         s->has_read = true;
1907 
1908         /* Wake up drain_co if it runs */
1909         if (s->drain_co) {
1910             aio_co_wake(s->drain_co);
1911         }
1912 
1913         return ret;
1914     }
1915 
1916     s->has_read = true;
1917     return 0;
1918 }
1919 
1920 /**
1921  * If .drain_count is 0, wake up .io_co if there is one; and set
1922  * .was_drained.
1923  * Increment .drain_count.
1924  */
1925 static void coroutine_fn bdrv_replace_test_co_drain_begin(BlockDriverState *bs)
1926 {
1927     BDRVReplaceTestState *s = bs->opaque;
1928 
1929     if (!s->drain_count) {
1930         /* Keep waking io_co up until it is done */
1931         s->drain_co = qemu_coroutine_self();
1932         while (s->io_co) {
1933             aio_co_wake(s->io_co);
1934             s->io_co = NULL;
1935             qemu_coroutine_yield();
1936         }
1937         s->drain_co = NULL;
1938 
1939         s->was_drained = true;
1940     }
1941     s->drain_count++;
1942 }
1943 
1944 /**
1945  * Reduce .drain_count, set .was_undrained once it reaches 0.
1946  * If .drain_count reaches 0 and the node has a backing file, issue a
1947  * read request.
1948  */
1949 static void coroutine_fn bdrv_replace_test_co_drain_end(BlockDriverState *bs)
1950 {
1951     BDRVReplaceTestState *s = bs->opaque;
1952 
1953     g_assert(s->drain_count > 0);
1954     if (!--s->drain_count) {
1955         int ret;
1956 
1957         s->was_undrained = true;
1958 
1959         if (bs->backing) {
1960             char data;
1961             QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
1962 
1963             /* Queue a read request post-drain */
1964             ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
1965             g_assert(ret >= 0);
1966         }
1967     }
1968 }
1969 
1970 static BlockDriver bdrv_replace_test = {
1971     .format_name            = "replace_test",
1972     .instance_size          = sizeof(BDRVReplaceTestState),
1973 
1974     .bdrv_close             = bdrv_replace_test_close,
1975     .bdrv_co_preadv         = bdrv_replace_test_co_preadv,
1976 
1977     .bdrv_co_drain_begin    = bdrv_replace_test_co_drain_begin,
1978     .bdrv_co_drain_end      = bdrv_replace_test_co_drain_end,
1979 
1980     .bdrv_child_perm        = bdrv_default_perms,
1981 };
1982 
1983 static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque)
1984 {
1985     int ret;
1986     char data;
1987 
1988     ret = blk_co_pread(opaque, 0, 1, &data, 0);
1989     g_assert(ret >= 0);
1990 }
1991 
1992 /**
1993  * We test two things:
1994  * (1) bdrv_replace_child_noperm() must not undrain the parent if both
1995  *     children are drained.
1996  * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
1997  *     drained child.  If the old child is drained, it must flush I/O
1998  *     requests after the new one has been attached.  If the new child
1999  *     is drained, it must flush I/O requests before the old one is
2000  *     detached.
2001  *
2002  * To do so, we create one parent node and two child nodes; then
2003  * attach one of the children (old_child_bs) to the parent, then
2004  * drain both old_child_bs and new_child_bs according to
2005  * old_drain_count and new_drain_count, respectively, and finally
2006  * we invoke bdrv_replace_node() to replace old_child_bs by
2007  * new_child_bs.
2008  *
2009  * The test block driver we use here (bdrv_replace_test) has a read
2010  * function that:
2011  * - For the parent node, can optionally yield, and then forwards the
2012  *   read to bdrv_preadv(),
2013  * - For the child node, just returns immediately.
2014  *
2015  * If the read yields, the drain_begin function will wake it up.
2016  *
2017  * The drain_end function issues a read on the parent once it is fully
2018  * undrained (which simulates requests starting to come in again).
2019  */
2020 static void do_test_replace_child_mid_drain(int old_drain_count,
2021                                             int new_drain_count)
2022 {
2023     BlockBackend *parent_blk;
2024     BlockDriverState *parent_bs;
2025     BlockDriverState *old_child_bs, *new_child_bs;
2026     BDRVReplaceTestState *parent_s;
2027     BDRVReplaceTestState *old_child_s, *new_child_s;
2028     Coroutine *io_co;
2029     int i;
2030 
2031     parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0,
2032                                      &error_abort);
2033     parent_s = parent_bs->opaque;
2034 
2035     parent_blk = blk_new(qemu_get_aio_context(),
2036                          BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
2037     blk_insert_bs(parent_blk, parent_bs, &error_abort);
2038 
2039     old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0,
2040                                         &error_abort);
2041     new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0,
2042                                         &error_abort);
2043     old_child_s = old_child_bs->opaque;
2044     new_child_s = new_child_bs->opaque;
2045 
2046     /* So that we can read something */
2047     parent_bs->total_sectors = 1;
2048     old_child_bs->total_sectors = 1;
2049     new_child_bs->total_sectors = 1;
2050 
2051     bdrv_ref(old_child_bs);
2052     parent_bs->backing = bdrv_attach_child(parent_bs, old_child_bs, "child",
2053                                            &child_of_bds, BDRV_CHILD_COW,
2054                                            &error_abort);
2055 
2056     for (i = 0; i < old_drain_count; i++) {
2057         bdrv_drained_begin(old_child_bs);
2058     }
2059     for (i = 0; i < new_drain_count; i++) {
2060         bdrv_drained_begin(new_child_bs);
2061     }
2062 
2063     if (!old_drain_count) {
2064         /*
2065          * Start a read operation that will yield, so it will not
2066          * complete before the node is drained.
2067          */
2068         parent_s->yield_before_read = true;
2069         io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co,
2070                                       parent_blk);
2071         qemu_coroutine_enter(io_co);
2072     }
2073 
2074     /* If we have started a read operation, it should have yielded */
2075     g_assert(!parent_s->has_read);
2076 
2077     /* Reset drained status so we can see what bdrv_replace_node() does */
2078     parent_s->was_drained = false;
2079     parent_s->was_undrained = false;
2080 
2081     g_assert(parent_bs->quiesce_counter == old_drain_count);
2082     bdrv_replace_node(old_child_bs, new_child_bs, &error_abort);
2083     g_assert(parent_bs->quiesce_counter == new_drain_count);
2084 
2085     if (!old_drain_count && !new_drain_count) {
2086         /*
2087          * From undrained to undrained drains and undrains the parent,
2088          * because bdrv_replace_node() contains a drained section for
2089          * @old_child_bs.
2090          */
2091         g_assert(parent_s->was_drained && parent_s->was_undrained);
2092     } else if (!old_drain_count && new_drain_count) {
2093         /*
2094          * From undrained to drained should drain the parent and keep
2095          * it that way.
2096          */
2097         g_assert(parent_s->was_drained && !parent_s->was_undrained);
2098     } else if (old_drain_count && !new_drain_count) {
2099         /*
2100          * From drained to undrained should undrain the parent and
2101          * keep it that way.
2102          */
2103         g_assert(!parent_s->was_drained && parent_s->was_undrained);
2104     } else /* if (old_drain_count && new_drain_count) */ {
2105         /*
2106          * From drained to drained must not undrain the parent at any
2107          * point
2108          */
2109         g_assert(!parent_s->was_drained && !parent_s->was_undrained);
2110     }
2111 
2112     if (!old_drain_count || !new_drain_count) {
2113         /*
2114          * If !old_drain_count, we have started a read request before
2115          * bdrv_replace_node().  If !new_drain_count, the parent must
2116          * have been undrained at some point, and
2117          * bdrv_replace_test_co_drain_end() starts a read request
2118          * then.
2119          */
2120         g_assert(parent_s->has_read);
2121     } else {
2122         /*
2123          * If the parent was never undrained, there is no way to start
2124          * a read request.
2125          */
2126         g_assert(!parent_s->has_read);
2127     }
2128 
2129     /* A drained child must have not received any request */
2130     g_assert(!(old_drain_count && old_child_s->has_read));
2131     g_assert(!(new_drain_count && new_child_s->has_read));
2132 
2133     for (i = 0; i < new_drain_count; i++) {
2134         bdrv_drained_end(new_child_bs);
2135     }
2136     for (i = 0; i < old_drain_count; i++) {
2137         bdrv_drained_end(old_child_bs);
2138     }
2139 
2140     /*
2141      * By now, bdrv_replace_test_co_drain_end() must have been called
2142      * at some point while the new child was attached to the parent.
2143      */
2144     g_assert(parent_s->has_read);
2145     g_assert(new_child_s->has_read);
2146 
2147     blk_unref(parent_blk);
2148     bdrv_unref(parent_bs);
2149     bdrv_unref(old_child_bs);
2150     bdrv_unref(new_child_bs);
2151 }
2152 
2153 static void test_replace_child_mid_drain(void)
2154 {
2155     int old_drain_count, new_drain_count;
2156 
2157     for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) {
2158         for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) {
2159             do_test_replace_child_mid_drain(old_drain_count, new_drain_count);
2160         }
2161     }
2162 }
2163 
2164 int main(int argc, char **argv)
2165 {
2166     int ret;
2167 
2168     bdrv_init();
2169     qemu_init_main_loop(&error_abort);
2170 
2171     g_test_init(&argc, &argv, NULL);
2172     qemu_event_init(&done_event, false);
2173 
2174     g_test_add_func("/bdrv-drain/driver-cb/drain_all", test_drv_cb_drain_all);
2175     g_test_add_func("/bdrv-drain/driver-cb/drain", test_drv_cb_drain);
2176     g_test_add_func("/bdrv-drain/driver-cb/drain_subtree",
2177                     test_drv_cb_drain_subtree);
2178 
2179     g_test_add_func("/bdrv-drain/driver-cb/co/drain_all",
2180                     test_drv_cb_co_drain_all);
2181     g_test_add_func("/bdrv-drain/driver-cb/co/drain", test_drv_cb_co_drain);
2182     g_test_add_func("/bdrv-drain/driver-cb/co/drain_subtree",
2183                     test_drv_cb_co_drain_subtree);
2184 
2185 
2186     g_test_add_func("/bdrv-drain/quiesce/drain_all", test_quiesce_drain_all);
2187     g_test_add_func("/bdrv-drain/quiesce/drain", test_quiesce_drain);
2188     g_test_add_func("/bdrv-drain/quiesce/drain_subtree",
2189                     test_quiesce_drain_subtree);
2190 
2191     g_test_add_func("/bdrv-drain/quiesce/co/drain_all",
2192                     test_quiesce_co_drain_all);
2193     g_test_add_func("/bdrv-drain/quiesce/co/drain", test_quiesce_co_drain);
2194     g_test_add_func("/bdrv-drain/quiesce/co/drain_subtree",
2195                     test_quiesce_co_drain_subtree);
2196 
2197     g_test_add_func("/bdrv-drain/nested", test_nested);
2198     g_test_add_func("/bdrv-drain/multiparent", test_multiparent);
2199 
2200     g_test_add_func("/bdrv-drain/graph-change/drain_subtree",
2201                     test_graph_change_drain_subtree);
2202     g_test_add_func("/bdrv-drain/graph-change/drain_all",
2203                     test_graph_change_drain_all);
2204 
2205     g_test_add_func("/bdrv-drain/iothread/drain_all", test_iothread_drain_all);
2206     g_test_add_func("/bdrv-drain/iothread/drain", test_iothread_drain);
2207     g_test_add_func("/bdrv-drain/iothread/drain_subtree",
2208                     test_iothread_drain_subtree);
2209 
2210     g_test_add_func("/bdrv-drain/blockjob/drain_all", test_blockjob_drain_all);
2211     g_test_add_func("/bdrv-drain/blockjob/drain", test_blockjob_drain);
2212     g_test_add_func("/bdrv-drain/blockjob/drain_subtree",
2213                     test_blockjob_drain_subtree);
2214 
2215     g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
2216                     test_blockjob_error_drain_all);
2217     g_test_add_func("/bdrv-drain/blockjob/error/drain",
2218                     test_blockjob_error_drain);
2219     g_test_add_func("/bdrv-drain/blockjob/error/drain_subtree",
2220                     test_blockjob_error_drain_subtree);
2221 
2222     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
2223                     test_blockjob_iothread_drain_all);
2224     g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
2225                     test_blockjob_iothread_drain);
2226     g_test_add_func("/bdrv-drain/blockjob/iothread/drain_subtree",
2227                     test_blockjob_iothread_drain_subtree);
2228 
2229     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
2230                     test_blockjob_iothread_error_drain_all);
2231     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
2232                     test_blockjob_iothread_error_drain);
2233     g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_subtree",
2234                     test_blockjob_iothread_error_drain_subtree);
2235 
2236     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
2237     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
2238     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
2239     g_test_add_func("/bdrv-drain/detach/drain_subtree", test_detach_by_drain_subtree);
2240     g_test_add_func("/bdrv-drain/detach/parent_cb", test_detach_by_parent_cb);
2241     g_test_add_func("/bdrv-drain/detach/driver_cb", test_detach_by_driver_cb);
2242 
2243     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
2244 
2245     g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context);
2246 
2247     g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
2248                     test_blockjob_commit_by_drained_end);
2249 
2250     g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
2251                     test_drop_intermediate_poll);
2252 
2253     g_test_add_func("/bdrv-drain/replace_child/mid-drain",
2254                     test_replace_child_mid_drain);
2255 
2256     ret = g_test_run();
2257     qemu_event_destroy(&done_event);
2258     return ret;
2259 }
2260