]> git.proxmox.com Git - mirror_qemu.git/blobdiff - tests/test-bdrv-drain.c
Merge remote-tracking branch 'remotes/mst/tags/for_upstream' into staging
[mirror_qemu.git] / tests / test-bdrv-drain.c
index 89ac15e88a0285e7b6c69ab01080105318a89207..11072718400aecf6f7ead81f470949eb654bc164 100644 (file)
@@ -27,6 +27,7 @@
 #include "block/blockjob_int.h"
 #include "sysemu/block-backend.h"
 #include "qapi/error.h"
+#include "qemu/main-loop.h"
 #include "iothread.h"
 
 static QemuEvent done_event;
@@ -84,20 +85,11 @@ static int coroutine_fn bdrv_test_co_preadv(BlockDriverState *bs,
     return 0;
 }
 
-static void bdrv_test_child_perm(BlockDriverState *bs, BdrvChild *c,
-                                 const BdrvChildRole *role,
-                                 BlockReopenQueue *reopen_queue,
-                                 uint64_t perm, uint64_t shared,
-                                 uint64_t *nperm, uint64_t *nshared)
+static int bdrv_test_change_backing_file(BlockDriverState *bs,
+                                         const char *backing_file,
+                                         const char *backing_fmt)
 {
-    /* bdrv_format_default_perms() accepts only these two, so disguise
-     * detach_by_driver_cb_role as one of them. */
-    if (role != &child_file && role != &child_backing) {
-        role = &child_file;
-    }
-
-    bdrv_format_default_perms(bs, c, role, reopen_queue, perm, shared,
-                              nperm, nshared);
+    return 0;
 }
 
 static BlockDriver bdrv_test = {
@@ -110,7 +102,9 @@ static BlockDriver bdrv_test = {
     .bdrv_co_drain_begin    = bdrv_test_co_drain_begin,
     .bdrv_co_drain_end      = bdrv_test_co_drain_end,
 
-    .bdrv_child_perm        = bdrv_test_child_perm,
+    .bdrv_child_perm        = bdrv_default_perms,
+
+    .bdrv_change_backing_file = bdrv_test_change_backing_file,
 };
 
 static void aio_ret_cb(void *opaque, int ret)
@@ -174,6 +168,28 @@ static void do_drain_end(enum drain_type drain_type, BlockDriverState *bs)
     }
 }
 
+static void do_drain_begin_unlocked(enum drain_type drain_type, BlockDriverState *bs)
+{
+    if (drain_type != BDRV_DRAIN_ALL) {
+        aio_context_acquire(bdrv_get_aio_context(bs));
+    }
+    do_drain_begin(drain_type, bs);
+    if (drain_type != BDRV_DRAIN_ALL) {
+        aio_context_release(bdrv_get_aio_context(bs));
+    }
+}
+
+static void do_drain_end_unlocked(enum drain_type drain_type, BlockDriverState *bs)
+{
+    if (drain_type != BDRV_DRAIN_ALL) {
+        aio_context_acquire(bdrv_get_aio_context(bs));
+    }
+    do_drain_end(drain_type, bs);
+    if (drain_type != BDRV_DRAIN_ALL) {
+        aio_context_release(bdrv_get_aio_context(bs));
+    }
+}
+
 static void test_drv_cb_common(enum drain_type drain_type, bool recursive)
 {
     BlockBackend *blk;
@@ -182,14 +198,9 @@ static void test_drv_cb_common(enum drain_type drain_type, bool recursive)
     BlockAIOCB *acb;
     int aio_ret;
 
-    QEMUIOVector qiov;
-    struct iovec iov = {
-        .iov_base = NULL,
-        .iov_len = 0,
-    };
-    qemu_iovec_init_external(&qiov, &iov, 1);
+    QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
 
-    blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
                               &error_abort);
     s = bs->opaque;
@@ -273,7 +284,7 @@ static void test_quiesce_common(enum drain_type drain_type, bool recursive)
     BlockBackend *blk;
     BlockDriverState *bs, *backing;
 
-    blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
                               &error_abort);
     blk_insert_bs(blk, bs, &error_abort);
@@ -336,7 +347,7 @@ static void test_nested(void)
     BDRVTestState *s, *backing_s;
     enum drain_type outer, inner;
 
-    blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
                               &error_abort);
     s = bs->opaque;
@@ -385,13 +396,13 @@ static void test_multiparent(void)
     BlockDriverState *bs_a, *bs_b, *backing;
     BDRVTestState *a_s, *b_s, *backing_s;
 
-    blk_a = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
                                 &error_abort);
     a_s = bs_a->opaque;
     blk_insert_bs(blk_a, bs_a, &error_abort);
 
-    blk_b = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
                                 &error_abort);
     b_s = bs_b->opaque;
@@ -458,13 +469,13 @@ static void test_graph_change_drain_subtree(void)
     BlockDriverState *bs_a, *bs_b, *backing;
     BDRVTestState *a_s, *b_s, *backing_s;
 
-    blk_a = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
                                 &error_abort);
     a_s = bs_a->opaque;
     blk_insert_bs(blk_a, bs_a, &error_abort);
 
-    blk_b = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
                                 &error_abort);
     b_s = bs_b->opaque;
@@ -538,7 +549,7 @@ static void test_graph_change_drain_all(void)
     BDRVTestState *a_s, *b_s;
 
     /* Create node A with a BlockBackend */
-    blk_a = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_a = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs_a = bdrv_new_open_driver(&bdrv_test, "test-node-a", BDRV_O_RDWR,
                                 &error_abort);
     a_s = bs_a->opaque;
@@ -554,7 +565,7 @@ static void test_graph_change_drain_all(void)
     g_assert_cmpint(a_s->drain_count, ==, 1);
 
     /* Create node B with a BlockBackend */
-    blk_b = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_b = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs_b = bdrv_new_open_driver(&bdrv_test, "test-node-b", BDRV_O_RDWR,
                                 &error_abort);
     b_s = bs_b->opaque;
@@ -614,6 +625,17 @@ static void test_iothread_aio_cb(void *opaque, int ret)
     qemu_event_set(&done_event);
 }
 
+static void test_iothread_main_thread_bh(void *opaque)
+{
+    struct test_iothread_data *data = opaque;
+
+    /* Test that the AioContext is not yet locked in a random BH that is
+     * executed during drain, otherwise this would deadlock. */
+    aio_context_acquire(bdrv_get_aio_context(data->bs));
+    bdrv_flush(data->bs);
+    aio_context_release(bdrv_get_aio_context(data->bs));
+}
+
 /*
  * Starts an AIO request on a BDS that runs in the AioContext of iothread 1.
  * The request involves a BH on iothread 2 before it can complete.
@@ -637,30 +659,28 @@ static void test_iothread_common(enum drain_type drain_type, int drain_thread)
     AioContext *ctx_a = iothread_get_aio_context(a);
     AioContext *ctx_b = iothread_get_aio_context(b);
 
-    QEMUIOVector qiov;
-    struct iovec iov = {
-        .iov_base = NULL,
-        .iov_len = 0,
-    };
-    qemu_iovec_init_external(&qiov, &iov, 1);
+    QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
 
     /* bdrv_drain_all() may only be called from the main loop thread */
     if (drain_type == BDRV_DRAIN_ALL && drain_thread != 0) {
         goto out;
     }
 
-    blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
                               &error_abort);
     s = bs->opaque;
     blk_insert_bs(blk, bs, &error_abort);
+    blk_set_disable_request_queuing(blk, true);
 
-    blk_set_aio_context(blk, ctx_a);
+    blk_set_aio_context(blk, ctx_a, &error_abort);
     aio_context_acquire(ctx_a);
 
     s->bh_indirection_ctx = ctx_b;
 
     aio_ret = -EINPROGRESS;
+    qemu_event_reset(&done_event);
+
     if (drain_thread == 0) {
         acb = blk_aio_preadv(blk, 0, &qiov, 0, test_iothread_aio_cb, &aio_ret);
     } else {
@@ -683,12 +703,13 @@ static void test_iothread_common(enum drain_type drain_type, int drain_thread)
             aio_context_acquire(ctx_a);
         }
 
+        aio_bh_schedule_oneshot(ctx_a, test_iothread_main_thread_bh, &data);
+
         /* The request is running on the IOThread a. Draining its block device
          * will make sure that it has completed as far as the BDS is concerned,
          * but the drain in this thread can continue immediately after
          * bdrv_dec_in_flight() and aio_ret might be assigned only slightly
          * later. */
-        qemu_event_reset(&done_event);
         do_drain_begin(drain_type, bs);
         g_assert_cmpint(bs->in_flight, ==, 0);
 
@@ -708,7 +729,6 @@ static void test_iothread_common(enum drain_type drain_type, int drain_thread)
         }
         break;
     case 1:
-        qemu_event_reset(&done_event);
         aio_bh_schedule_oneshot(ctx_a, test_iothread_drain_entry, &data);
         qemu_event_wait(&done_event);
         break;
@@ -717,7 +737,7 @@ static void test_iothread_common(enum drain_type drain_type, int drain_thread)
     }
 
     aio_context_acquire(ctx_a);
-    blk_set_aio_context(blk, qemu_get_aio_context());
+    blk_set_aio_context(blk, qemu_get_aio_context(), &error_abort);
     aio_context_release(ctx_a);
 
     bdrv_unref(bs);
@@ -749,23 +769,56 @@ static void test_iothread_drain_subtree(void)
 
 typedef struct TestBlockJob {
     BlockJob common;
+    int run_ret;
+    int prepare_ret;
+    bool running;
     bool should_complete;
 } TestBlockJob;
 
+static int test_job_prepare(Job *job)
+{
+    TestBlockJob *s = container_of(job, TestBlockJob, common.job);
+
+    /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
+    blk_flush(s->common.blk);
+    return s->prepare_ret;
+}
+
+static void test_job_commit(Job *job)
+{
+    TestBlockJob *s = container_of(job, TestBlockJob, common.job);
+
+    /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
+    blk_flush(s->common.blk);
+}
+
+static void test_job_abort(Job *job)
+{
+    TestBlockJob *s = container_of(job, TestBlockJob, common.job);
+
+    /* Provoke an AIO_WAIT_WHILE() call to verify there is no deadlock */
+    blk_flush(s->common.blk);
+}
+
 static int coroutine_fn test_job_run(Job *job, Error **errp)
 {
     TestBlockJob *s = container_of(job, TestBlockJob, common.job);
 
+    /* We are running the actual job code past the pause point in
+     * job_co_entry(). */
+    s->running = true;
+
     job_transition_to_ready(&s->common.job);
     while (!s->should_complete) {
-        /* Avoid block_job_sleep_ns() because it marks the job as !busy. We
-         * want to emulate some actual activity (probably some I/O) here so
-         * that drain has to wait for this acitivity to stop. */
-        qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 100000);
+        /* Avoid job_sleep_ns() because it marks the job as !busy. We want to
+         * emulate some actual activity (probably some I/O) here so that drain
+         * has to wait for this activity to stop. */
+        qemu_co_sleep_ns(QEMU_CLOCK_REALTIME, 1000000);
+
         job_pause_point(&s->common.job);
     }
 
-    return 0;
+    return s->run_ret;
 }
 
 static void test_job_complete(Job *job, Error **errp)
@@ -779,39 +832,118 @@ BlockJobDriver test_job_driver = {
         .instance_size  = sizeof(TestBlockJob),
         .free           = block_job_free,
         .user_resume    = block_job_user_resume,
-        .drain          = block_job_drain,
         .run            = test_job_run,
         .complete       = test_job_complete,
+        .prepare        = test_job_prepare,
+        .commit         = test_job_commit,
+        .abort          = test_job_abort,
     },
 };
 
-static void test_blockjob_common(enum drain_type drain_type)
+enum test_job_result {
+    TEST_JOB_SUCCESS,
+    TEST_JOB_FAIL_RUN,
+    TEST_JOB_FAIL_PREPARE,
+};
+
+enum test_job_drain_node {
+    TEST_JOB_DRAIN_SRC,
+    TEST_JOB_DRAIN_SRC_CHILD,
+    TEST_JOB_DRAIN_SRC_PARENT,
+};
+
+static void test_blockjob_common_drain_node(enum drain_type drain_type,
+                                            bool use_iothread,
+                                            enum test_job_result result,
+                                            enum test_job_drain_node drain_node)
 {
     BlockBackend *blk_src, *blk_target;
-    BlockDriverState *src, *target;
+    BlockDriverState *src, *src_backing, *src_overlay, *target, *drain_bs;
     BlockJob *job;
+    TestBlockJob *tjob;
+    IOThread *iothread = NULL;
+    AioContext *ctx;
     int ret;
 
     src = bdrv_new_open_driver(&bdrv_test, "source", BDRV_O_RDWR,
                                &error_abort);
-    blk_src = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
-    blk_insert_bs(blk_src, src, &error_abort);
+    src_backing = bdrv_new_open_driver(&bdrv_test, "source-backing",
+                                       BDRV_O_RDWR, &error_abort);
+    src_overlay = bdrv_new_open_driver(&bdrv_test, "source-overlay",
+                                       BDRV_O_RDWR, &error_abort);
+
+    bdrv_set_backing_hd(src_overlay, src, &error_abort);
+    bdrv_unref(src);
+    bdrv_set_backing_hd(src, src_backing, &error_abort);
+    bdrv_unref(src_backing);
+
+    blk_src = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_insert_bs(blk_src, src_overlay, &error_abort);
+
+    switch (drain_node) {
+    case TEST_JOB_DRAIN_SRC:
+        drain_bs = src;
+        break;
+    case TEST_JOB_DRAIN_SRC_CHILD:
+        drain_bs = src_backing;
+        break;
+    case TEST_JOB_DRAIN_SRC_PARENT:
+        drain_bs = src_overlay;
+        break;
+    default:
+        g_assert_not_reached();
+    }
+
+    if (use_iothread) {
+        iothread = iothread_new();
+        ctx = iothread_get_aio_context(iothread);
+        blk_set_aio_context(blk_src, ctx, &error_abort);
+    } else {
+        ctx = qemu_get_aio_context();
+    }
 
     target = bdrv_new_open_driver(&bdrv_test, "target", BDRV_O_RDWR,
                                   &error_abort);
-    blk_target = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk_target = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     blk_insert_bs(blk_target, target, &error_abort);
+    blk_set_allow_aio_context_change(blk_target, true);
 
-    job = block_job_create("job0", &test_job_driver, NULL, src, 0, BLK_PERM_ALL,
-                           0, 0, NULL, NULL, &error_abort);
+    aio_context_acquire(ctx);
+    tjob = block_job_create("job0", &test_job_driver, NULL, src,
+                            0, BLK_PERM_ALL,
+                            0, 0, NULL, NULL, &error_abort);
+    job = &tjob->common;
     block_job_add_bdrv(job, "target", target, 0, BLK_PERM_ALL, &error_abort);
+
+    switch (result) {
+    case TEST_JOB_SUCCESS:
+        break;
+    case TEST_JOB_FAIL_RUN:
+        tjob->run_ret = -EIO;
+        break;
+    case TEST_JOB_FAIL_PREPARE:
+        tjob->prepare_ret = -EIO;
+        break;
+    }
+
     job_start(&job->job);
+    aio_context_release(ctx);
+
+    if (use_iothread) {
+        /* job_co_entry() is run in the I/O thread, wait for the actual job
+         * code to start (we don't want to catch the job in the pause point in
+         * job_co_entry(). */
+        while (!tjob->running) {
+            aio_poll(qemu_get_aio_context(), false);
+        }
+    }
 
     g_assert_cmpint(job->job.pause_count, ==, 0);
     g_assert_false(job->job.paused);
-    g_assert_true(job->job.busy); /* We're in job_sleep_ns() */
+    g_assert_true(tjob->running);
+    g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
 
-    do_drain_begin(drain_type, src);
+    do_drain_begin_unlocked(drain_type, drain_bs);
 
     if (drain_type == BDRV_DRAIN_ALL) {
         /* bdrv_drain_all() drains both src and target */
@@ -822,13 +954,20 @@ static void test_blockjob_common(enum drain_type drain_type)
     g_assert_true(job->job.paused);
     g_assert_false(job->job.busy); /* The job is paused */
 
-    do_drain_end(drain_type, src);
+    do_drain_end_unlocked(drain_type, drain_bs);
+
+    if (use_iothread) {
+        /* paused is reset in the I/O thread, wait for it */
+        while (job->job.paused) {
+            aio_poll(qemu_get_aio_context(), false);
+        }
+    }
 
     g_assert_cmpint(job->job.pause_count, ==, 0);
     g_assert_false(job->job.paused);
     g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
 
-    do_drain_begin(drain_type, target);
+    do_drain_begin_unlocked(drain_type, target);
 
     if (drain_type == BDRV_DRAIN_ALL) {
         /* bdrv_drain_all() drains both src and target */
@@ -839,34 +978,116 @@ static void test_blockjob_common(enum drain_type drain_type)
     g_assert_true(job->job.paused);
     g_assert_false(job->job.busy); /* The job is paused */
 
-    do_drain_end(drain_type, target);
+    do_drain_end_unlocked(drain_type, target);
+
+    if (use_iothread) {
+        /* paused is reset in the I/O thread, wait for it */
+        while (job->job.paused) {
+            aio_poll(qemu_get_aio_context(), false);
+        }
+    }
 
     g_assert_cmpint(job->job.pause_count, ==, 0);
     g_assert_false(job->job.paused);
-    g_assert_true(job->job.busy); /* We're in job_sleep_ns() */
+    g_assert_true(job->job.busy); /* We're in qemu_co_sleep_ns() */
 
+    aio_context_acquire(ctx);
     ret = job_complete_sync(&job->job, &error_abort);
-    g_assert_cmpint(ret, ==, 0);
+    g_assert_cmpint(ret, ==, (result == TEST_JOB_SUCCESS ? 0 : -EIO));
+
+    if (use_iothread) {
+        blk_set_aio_context(blk_src, qemu_get_aio_context(), &error_abort);
+        assert(blk_get_aio_context(blk_target) == qemu_get_aio_context());
+    }
+    aio_context_release(ctx);
 
     blk_unref(blk_src);
     blk_unref(blk_target);
-    bdrv_unref(src);
+    bdrv_unref(src_overlay);
     bdrv_unref(target);
+
+    if (iothread) {
+        iothread_join(iothread);
+    }
+}
+
+static void test_blockjob_common(enum drain_type drain_type, bool use_iothread,
+                                 enum test_job_result result)
+{
+    test_blockjob_common_drain_node(drain_type, use_iothread, result,
+                                    TEST_JOB_DRAIN_SRC);
+    test_blockjob_common_drain_node(drain_type, use_iothread, result,
+                                    TEST_JOB_DRAIN_SRC_CHILD);
+    if (drain_type == BDRV_SUBTREE_DRAIN) {
+        test_blockjob_common_drain_node(drain_type, use_iothread, result,
+                                        TEST_JOB_DRAIN_SRC_PARENT);
+    }
 }
 
 static void test_blockjob_drain_all(void)
 {
-    test_blockjob_common(BDRV_DRAIN_ALL);
+    test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_SUCCESS);
 }
 
 static void test_blockjob_drain(void)
 {
-    test_blockjob_common(BDRV_DRAIN);
+    test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_SUCCESS);
 }
 
 static void test_blockjob_drain_subtree(void)
 {
-    test_blockjob_common(BDRV_SUBTREE_DRAIN);
+    test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_SUCCESS);
+}
+
+static void test_blockjob_error_drain_all(void)
+{
+    test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_RUN);
+    test_blockjob_common(BDRV_DRAIN_ALL, false, TEST_JOB_FAIL_PREPARE);
+}
+
+static void test_blockjob_error_drain(void)
+{
+    test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_RUN);
+    test_blockjob_common(BDRV_DRAIN, false, TEST_JOB_FAIL_PREPARE);
+}
+
+static void test_blockjob_error_drain_subtree(void)
+{
+    test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_RUN);
+    test_blockjob_common(BDRV_SUBTREE_DRAIN, false, TEST_JOB_FAIL_PREPARE);
+}
+
+static void test_blockjob_iothread_drain_all(void)
+{
+    test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_SUCCESS);
+}
+
+static void test_blockjob_iothread_drain(void)
+{
+    test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_SUCCESS);
+}
+
+static void test_blockjob_iothread_drain_subtree(void)
+{
+    test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_SUCCESS);
+}
+
+static void test_blockjob_iothread_error_drain_all(void)
+{
+    test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_RUN);
+    test_blockjob_common(BDRV_DRAIN_ALL, true, TEST_JOB_FAIL_PREPARE);
+}
+
+static void test_blockjob_iothread_error_drain(void)
+{
+    test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_RUN);
+    test_blockjob_common(BDRV_DRAIN, true, TEST_JOB_FAIL_PREPARE);
+}
+
+static void test_blockjob_iothread_error_drain_subtree(void)
+{
+    test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_RUN);
+    test_blockjob_common(BDRV_SUBTREE_DRAIN, true, TEST_JOB_FAIL_PREPARE);
 }
 
 
@@ -897,7 +1118,7 @@ static BlockDriver bdrv_test_top_driver = {
     .bdrv_close             = bdrv_test_top_close,
     .bdrv_co_preadv         = bdrv_test_top_co_preadv,
 
-    .bdrv_child_perm        = bdrv_format_default_perms,
+    .bdrv_child_perm        = bdrv_default_perms,
 };
 
 typedef struct TestCoDeleteByDrainData {
@@ -913,13 +1134,7 @@ static void coroutine_fn test_co_delete_by_drain(void *opaque)
     BlockDriverState *bs = blk_bs(blk);
     BDRVTestTopState *tts = bs->opaque;
     void *buffer = g_malloc(65536);
-    QEMUIOVector qiov;
-    struct iovec iov = {
-        .iov_base = buffer,
-        .iov_len  = 65536,
-    };
-
-    qemu_iovec_init_external(&qiov, &iov, 1);
+    QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buffer, 65536);
 
     /* Pretend some internal write operation from parent to child.
      * Important: We have to read from the child, not from the parent!
@@ -969,7 +1184,8 @@ static void do_test_delete_by_drain(bool detach_instead_of_delete,
 
     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
                         &error_abort);
-    bdrv_attach_child(bs, null_bs, "null-child", &child_file, &error_abort);
+    bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds,
+                      BDRV_CHILD_DATA, &error_abort);
 
     /* This child will be the one to pass to requests through to, and
      * it will stall until a drain occurs */
@@ -977,16 +1193,19 @@ static void do_test_delete_by_drain(bool detach_instead_of_delete,
                                     &error_abort);
     child_bs->total_sectors = 65536 >> BDRV_SECTOR_BITS;
     /* Takes our reference to child_bs */
-    tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child", &child_file,
+    tts->wait_child = bdrv_attach_child(bs, child_bs, "wait-child",
+                                        &child_of_bds,
+                                        BDRV_CHILD_DATA | BDRV_CHILD_PRIMARY,
                                         &error_abort);
 
     /* This child is just there to be deleted
      * (for detach_instead_of_delete == true) */
     null_bs = bdrv_open("null-co://", NULL, NULL, BDRV_O_RDWR | BDRV_O_PROTOCOL,
                         &error_abort);
-    bdrv_attach_child(bs, null_bs, "null-child", &child_file, &error_abort);
+    bdrv_attach_child(bs, null_bs, "null-child", &child_of_bds, BDRV_CHILD_DATA,
+                      &error_abort);
 
-    blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     blk_insert_bs(blk, bs, &error_abort);
 
     /* Referenced by blk now */
@@ -1081,7 +1300,8 @@ static void detach_indirect_bh(void *opaque)
 
     bdrv_ref(data->c);
     data->child_c = bdrv_attach_child(data->parent_b, data->c, "PB-C",
-                                      &child_file, &error_abort);
+                                      &child_of_bds, BDRV_CHILD_DATA,
+                                      &error_abort);
 }
 
 static void detach_by_parent_aio_cb(void *opaque, int ret)
@@ -1098,10 +1318,10 @@ static void detach_by_driver_cb_drained_begin(BdrvChild *child)
 {
     aio_bh_schedule_oneshot(qemu_get_current_aio_context(),
                             detach_indirect_bh, &detach_by_parent_data);
-    child_file.drained_begin(child);
+    child_of_bds.drained_begin(child);
 }
 
-static BdrvChildRole detach_by_driver_cb_role;
+static BdrvChildClass detach_by_driver_cb_class;
 
 /*
  * Initial graph:
@@ -1118,7 +1338,7 @@ static BdrvChildRole detach_by_driver_cb_role;
  *
  * by_parent_cb == false: Test that bdrv_drain_invoke() doesn't poll
  *
- *     PA's BdrvChildRole has a .drained_begin callback that schedules a BH
+ *     PA's BdrvChildClass has a .drained_begin callback that schedules a BH
  *     that does the same graph change. If bdrv_drain_invoke() calls it, the
  *     state is messed up, but if it is only polled in the single
  *     BDRV_POLL_WHILE() at the end of the drain, this should work fine.
@@ -1130,16 +1350,11 @@ static void test_detach_indirect(bool by_parent_cb)
     BdrvChild *child_a, *child_b;
     BlockAIOCB *acb;
 
-    QEMUIOVector qiov;
-    struct iovec iov = {
-        .iov_base = NULL,
-        .iov_len = 0,
-    };
-    qemu_iovec_init_external(&qiov, &iov, 1);
+    QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, NULL, 0);
 
     if (!by_parent_cb) {
-        detach_by_driver_cb_role = child_file;
-        detach_by_driver_cb_role.drained_begin =
+        detach_by_driver_cb_class = child_of_bds;
+        detach_by_driver_cb_class.drained_begin =
             detach_by_driver_cb_drained_begin;
     }
 
@@ -1154,7 +1369,7 @@ static void test_detach_indirect(bool by_parent_cb)
     c = bdrv_new_open_driver(&bdrv_test, "c", BDRV_O_RDWR, &error_abort);
 
     /* blk is a BB for parent-a */
-    blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     blk_insert_bs(blk, parent_a, &error_abort);
     bdrv_unref(parent_a);
 
@@ -1168,13 +1383,15 @@ static void test_detach_indirect(bool by_parent_cb)
     /* Set child relationships */
     bdrv_ref(b);
     bdrv_ref(a);
-    child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_file, &error_abort);
-    child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_backing, &error_abort);
+    child_b = bdrv_attach_child(parent_b, b, "PB-B", &child_of_bds,
+                                BDRV_CHILD_DATA, &error_abort);
+    child_a = bdrv_attach_child(parent_b, a, "PB-A", &child_of_bds,
+                                BDRV_CHILD_COW, &error_abort);
 
     bdrv_ref(a);
     bdrv_attach_child(parent_a, a, "PA-A",
-                      by_parent_cb ? &child_file : &detach_by_driver_cb_role,
-                      &error_abort);
+                      by_parent_cb ? &child_of_bds : &detach_by_driver_cb_class,
+                      BDRV_CHILD_DATA, &error_abort);
 
     g_assert_cmpint(parent_a->refcnt, ==, 1);
     g_assert_cmpint(parent_b->refcnt, ==, 1);
@@ -1222,12 +1439,6 @@ static void test_detach_indirect(bool by_parent_cb)
     bdrv_unref(parent_b);
     blk_unref(blk);
 
-    /* XXX Once bdrv_close() unref's children instead of just detaching them,
-     * this won't be necessary any more. */
-    bdrv_unref(a);
-    bdrv_unref(a);
-    bdrv_unref(c);
-
     g_assert_cmpint(a->refcnt, ==, 1);
     g_assert_cmpint(b->refcnt, ==, 1);
     g_assert_cmpint(c->refcnt, ==, 1);
@@ -1252,7 +1463,7 @@ static void test_append_to_drained(void)
     BlockDriverState *base, *overlay;
     BDRVTestState *base_s, *overlay_s;
 
-    blk = blk_new(BLK_PERM_ALL, BLK_PERM_ALL);
+    blk = blk_new(qemu_get_aio_context(), BLK_PERM_ALL, BLK_PERM_ALL);
     base = bdrv_new_open_driver(&bdrv_test, "base", BDRV_O_RDWR, &error_abort);
     base_s = base->opaque;
     blk_insert_bs(blk, base, &error_abort);
@@ -1287,6 +1498,639 @@ static void test_append_to_drained(void)
     blk_unref(blk);
 }
 
+static void test_set_aio_context(void)
+{
+    BlockDriverState *bs;
+    IOThread *a = iothread_new();
+    IOThread *b = iothread_new();
+    AioContext *ctx_a = iothread_get_aio_context(a);
+    AioContext *ctx_b = iothread_get_aio_context(b);
+
+    bs = bdrv_new_open_driver(&bdrv_test, "test-node", BDRV_O_RDWR,
+                              &error_abort);
+
+    bdrv_drained_begin(bs);
+    bdrv_try_set_aio_context(bs, ctx_a, &error_abort);
+
+    aio_context_acquire(ctx_a);
+    bdrv_drained_end(bs);
+
+    bdrv_drained_begin(bs);
+    bdrv_try_set_aio_context(bs, ctx_b, &error_abort);
+    aio_context_release(ctx_a);
+    aio_context_acquire(ctx_b);
+    bdrv_try_set_aio_context(bs, qemu_get_aio_context(), &error_abort);
+    aio_context_release(ctx_b);
+    bdrv_drained_end(bs);
+
+    bdrv_unref(bs);
+    iothread_join(a);
+    iothread_join(b);
+}
+
+
+typedef struct TestDropBackingBlockJob {
+    BlockJob common;
+    bool should_complete;
+    bool *did_complete;
+    BlockDriverState *detach_also;
+} TestDropBackingBlockJob;
+
+static int coroutine_fn test_drop_backing_job_run(Job *job, Error **errp)
+{
+    TestDropBackingBlockJob *s =
+        container_of(job, TestDropBackingBlockJob, common.job);
+
+    while (!s->should_complete) {
+        job_sleep_ns(job, 0);
+    }
+
+    return 0;
+}
+
+static void test_drop_backing_job_commit(Job *job)
+{
+    TestDropBackingBlockJob *s =
+        container_of(job, TestDropBackingBlockJob, common.job);
+
+    bdrv_set_backing_hd(blk_bs(s->common.blk), NULL, &error_abort);
+    bdrv_set_backing_hd(s->detach_also, NULL, &error_abort);
+
+    *s->did_complete = true;
+}
+
+static const BlockJobDriver test_drop_backing_job_driver = {
+    .job_driver = {
+        .instance_size  = sizeof(TestDropBackingBlockJob),
+        .free           = block_job_free,
+        .user_resume    = block_job_user_resume,
+        .run            = test_drop_backing_job_run,
+        .commit         = test_drop_backing_job_commit,
+    }
+};
+
+/**
+ * Creates a child node with three parent nodes on it, and then runs a
+ * block job on the final one, parent-node-2.
+ *
+ * The job is then asked to complete before a section where the child
+ * is drained.
+ *
+ * Ending this section will undrain the child's parents, first
+ * parent-node-2, then parent-node-1, then parent-node-0 -- the parent
+ * list is in reverse order of how they were added.  Ending the drain
+ * on parent-node-2 will resume the job, thus completing it and
+ * scheduling job_exit().
+ *
+ * Ending the drain on parent-node-1 will poll the AioContext, which
+ * lets job_exit() and thus test_drop_backing_job_commit() run.  That
+ * function first removes the child as parent-node-2's backing file.
+ *
+ * In old (and buggy) implementations, there are two problems with
+ * that:
+ * (A) bdrv_drain_invoke() polls for every node that leaves the
+ *     drained section.  This means that job_exit() is scheduled
+ *     before the child has left the drained section.  Its
+ *     quiesce_counter is therefore still 1 when it is removed from
+ *     parent-node-2.
+ *
+ * (B) bdrv_replace_child_noperm() calls drained_end() on the old
+ *     child's parents as many times as the child is quiesced.  This
+ *     means it will call drained_end() on parent-node-2 once.
+ *     Because parent-node-2 is no longer quiesced at this point, this
+ *     will fail.
+ *
+ * bdrv_replace_child_noperm() therefore must call drained_end() on
+ * the parent only if it really is still drained because the child is
+ * drained.
+ *
+ * If removing child from parent-node-2 was successful (as it should
+ * be), test_drop_backing_job_commit() will then also remove the child
+ * from parent-node-0.
+ *
+ * With an old version of our drain infrastructure ((A) above), that
+ * resulted in the following flow:
+ *
+ * 1. child attempts to leave its drained section.  The call recurses
+ *    to its parents.
+ *
+ * 2. parent-node-2 leaves the drained section.  Polling in
+ *    bdrv_drain_invoke() will schedule job_exit().
+ *
+ * 3. parent-node-1 leaves the drained section.  Polling in
+ *    bdrv_drain_invoke() will run job_exit(), thus disconnecting
+ *    parent-node-0 from the child node.
+ *
+ * 4. bdrv_parent_drained_end() uses a QLIST_FOREACH_SAFE() loop to
+ *    iterate over the parents.  Thus, it now accesses the BdrvChild
+ *    object that used to connect parent-node-0 and the child node.
+ *    However, that object no longer exists, so it accesses a dangling
+ *    pointer.
+ *
+ * The solution is to only poll once when running a bdrv_drained_end()
+ * operation, specifically at the end when all drained_end()
+ * operations for all involved nodes have been scheduled.
+ * Note that this also solves (A) above, thus hiding (B).
+ */
+static void test_blockjob_commit_by_drained_end(void)
+{
+    BlockDriverState *bs_child, *bs_parents[3];
+    TestDropBackingBlockJob *job;
+    bool job_has_completed = false;
+    int i;
+
+    bs_child = bdrv_new_open_driver(&bdrv_test, "child-node", BDRV_O_RDWR,
+                                    &error_abort);
+
+    for (i = 0; i < 3; i++) {
+        char name[32];
+        snprintf(name, sizeof(name), "parent-node-%i", i);
+        bs_parents[i] = bdrv_new_open_driver(&bdrv_test, name, BDRV_O_RDWR,
+                                             &error_abort);
+        bdrv_set_backing_hd(bs_parents[i], bs_child, &error_abort);
+    }
+
+    job = block_job_create("job", &test_drop_backing_job_driver, NULL,
+                           bs_parents[2], 0, BLK_PERM_ALL, 0, 0, NULL, NULL,
+                           &error_abort);
+
+    job->detach_also = bs_parents[0];
+    job->did_complete = &job_has_completed;
+
+    job_start(&job->common.job);
+
+    job->should_complete = true;
+    bdrv_drained_begin(bs_child);
+    g_assert(!job_has_completed);
+    bdrv_drained_end(bs_child);
+    g_assert(job_has_completed);
+
+    bdrv_unref(bs_parents[0]);
+    bdrv_unref(bs_parents[1]);
+    bdrv_unref(bs_parents[2]);
+    bdrv_unref(bs_child);
+}
+
+
+typedef struct TestSimpleBlockJob {
+    BlockJob common;
+    bool should_complete;
+    bool *did_complete;
+} TestSimpleBlockJob;
+
+static int coroutine_fn test_simple_job_run(Job *job, Error **errp)
+{
+    TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
+
+    while (!s->should_complete) {
+        job_sleep_ns(job, 0);
+    }
+
+    return 0;
+}
+
+static void test_simple_job_clean(Job *job)
+{
+    TestSimpleBlockJob *s = container_of(job, TestSimpleBlockJob, common.job);
+    *s->did_complete = true;
+}
+
+static const BlockJobDriver test_simple_job_driver = {
+    .job_driver = {
+        .instance_size  = sizeof(TestSimpleBlockJob),
+        .free           = block_job_free,
+        .user_resume    = block_job_user_resume,
+        .run            = test_simple_job_run,
+        .clean          = test_simple_job_clean,
+    },
+};
+
+static int drop_intermediate_poll_update_filename(BdrvChild *child,
+                                                  BlockDriverState *new_base,
+                                                  const char *filename,
+                                                  Error **errp)
+{
+    /*
+     * We are free to poll here, which may change the block graph, if
+     * it is not drained.
+     */
+
+    /* If the job is not drained: Complete it, schedule job_exit() */
+    aio_poll(qemu_get_current_aio_context(), false);
+    /* If the job is not drained: Run job_exit(), finish the job */
+    aio_poll(qemu_get_current_aio_context(), false);
+
+    return 0;
+}
+
+/**
+ * Test a poll in the midst of bdrv_drop_intermediate().
+ *
+ * bdrv_drop_intermediate() calls BdrvChildClass.update_filename(),
+ * which can yield or poll.  This may lead to graph changes, unless
+ * the whole subtree in question is drained.
+ *
+ * We test this on the following graph:
+ *
+ *                    Job
+ *
+ *                     |
+ *                  job-node
+ *                     |
+ *                     v
+ *
+ *                  job-node
+ *
+ *                     |
+ *                  backing
+ *                     |
+ *                     v
+ *
+ * node-2 --chain--> node-1 --chain--> node-0
+ *
+ * We drop node-1 with bdrv_drop_intermediate(top=node-1, base=node-0).
+ *
+ * This first updates node-2's backing filename by invoking
+ * drop_intermediate_poll_update_filename(), which polls twice.  This
+ * causes the job to finish, which in turns causes the job-node to be
+ * deleted.
+ *
+ * bdrv_drop_intermediate() uses a QLIST_FOREACH_SAFE() loop, so it
+ * already has a pointer to the BdrvChild edge between job-node and
+ * node-1.  When it tries to handle that edge, we probably get a
+ * segmentation fault because the object no longer exists.
+ *
+ *
+ * The solution is for bdrv_drop_intermediate() to drain top's
+ * subtree.  This prevents graph changes from happening just because
+ * BdrvChildClass.update_filename() yields or polls.  Thus, the block
+ * job is paused during that drained section and must finish before or
+ * after.
+ *
+ * (In addition, bdrv_replace_child() must keep the job paused.)
+ */
+static void test_drop_intermediate_poll(void)
+{
+    static BdrvChildClass chain_child_class;
+    BlockDriverState *chain[3];
+    TestSimpleBlockJob *job;
+    BlockDriverState *job_node;
+    bool job_has_completed = false;
+    int i;
+    int ret;
+
+    chain_child_class = child_of_bds;
+    chain_child_class.update_filename = drop_intermediate_poll_update_filename;
+
+    for (i = 0; i < 3; i++) {
+        char name[32];
+        snprintf(name, 32, "node-%i", i);
+
+        chain[i] = bdrv_new_open_driver(&bdrv_test, name, 0, &error_abort);
+    }
+
+    job_node = bdrv_new_open_driver(&bdrv_test, "job-node", BDRV_O_RDWR,
+                                    &error_abort);
+    bdrv_set_backing_hd(job_node, chain[1], &error_abort);
+
+    /*
+     * Establish the chain last, so the chain links are the first
+     * elements in the BDS.parents lists
+     */
+    for (i = 0; i < 3; i++) {
+        if (i) {
+            /* Takes the reference to chain[i - 1] */
+            chain[i]->backing = bdrv_attach_child(chain[i], chain[i - 1],
+                                                  "chain", &chain_child_class,
+                                                  BDRV_CHILD_COW, &error_abort);
+        }
+    }
+
+    job = block_job_create("job", &test_simple_job_driver, NULL, job_node,
+                           0, BLK_PERM_ALL, 0, 0, NULL, NULL, &error_abort);
+
+    /* The job has a reference now */
+    bdrv_unref(job_node);
+
+    job->did_complete = &job_has_completed;
+
+    job_start(&job->common.job);
+    job->should_complete = true;
+
+    g_assert(!job_has_completed);
+    ret = bdrv_drop_intermediate(chain[1], chain[0], NULL);
+    g_assert(ret == 0);
+    g_assert(job_has_completed);
+
+    bdrv_unref(chain[2]);
+}
+
+
+typedef struct BDRVReplaceTestState {
+    bool was_drained;
+    bool was_undrained;
+    bool has_read;
+
+    int drain_count;
+
+    bool yield_before_read;
+    Coroutine *io_co;
+    Coroutine *drain_co;
+} BDRVReplaceTestState;
+
+static void bdrv_replace_test_close(BlockDriverState *bs)
+{
+}
+
+/**
+ * If @bs has a backing file:
+ *   Yield if .yield_before_read is true (and wait for drain_begin to
+ *   wake us up).
+ *   Forward the read to bs->backing.  Set .has_read to true.
+ *   If drain_begin has woken us, wake it in turn.
+ *
+ * Otherwise:
+ *   Set .has_read to true and return success.
+ */
+static int coroutine_fn bdrv_replace_test_co_preadv(BlockDriverState *bs,
+                                                    uint64_t offset,
+                                                    uint64_t bytes,
+                                                    QEMUIOVector *qiov,
+                                                    int flags)
+{
+    BDRVReplaceTestState *s = bs->opaque;
+
+    if (bs->backing) {
+        int ret;
+
+        g_assert(!s->drain_count);
+
+        s->io_co = qemu_coroutine_self();
+        if (s->yield_before_read) {
+            s->yield_before_read = false;
+            qemu_coroutine_yield();
+        }
+        s->io_co = NULL;
+
+        ret = bdrv_preadv(bs->backing, offset, qiov);
+        s->has_read = true;
+
+        /* Wake up drain_co if it runs */
+        if (s->drain_co) {
+            aio_co_wake(s->drain_co);
+        }
+
+        return ret;
+    }
+
+    s->has_read = true;
+    return 0;
+}
+
+/**
+ * If .drain_count is 0, wake up .io_co if there is one; and set
+ * .was_drained.
+ * Increment .drain_count.
+ */
+static void coroutine_fn bdrv_replace_test_co_drain_begin(BlockDriverState *bs)
+{
+    BDRVReplaceTestState *s = bs->opaque;
+
+    if (!s->drain_count) {
+        /* Keep waking io_co up until it is done */
+        s->drain_co = qemu_coroutine_self();
+        while (s->io_co) {
+            aio_co_wake(s->io_co);
+            s->io_co = NULL;
+            qemu_coroutine_yield();
+        }
+        s->drain_co = NULL;
+
+        s->was_drained = true;
+    }
+    s->drain_count++;
+}
+
+/**
+ * Reduce .drain_count, set .was_undrained once it reaches 0.
+ * If .drain_count reaches 0 and the node has a backing file, issue a
+ * read request.
+ */
+static void coroutine_fn bdrv_replace_test_co_drain_end(BlockDriverState *bs)
+{
+    BDRVReplaceTestState *s = bs->opaque;
+
+    g_assert(s->drain_count > 0);
+    if (!--s->drain_count) {
+        int ret;
+
+        s->was_undrained = true;
+
+        if (bs->backing) {
+            char data;
+            QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, &data, 1);
+
+            /* Queue a read request post-drain */
+            ret = bdrv_replace_test_co_preadv(bs, 0, 1, &qiov, 0);
+            g_assert(ret >= 0);
+        }
+    }
+}
+
+static BlockDriver bdrv_replace_test = {
+    .format_name            = "replace_test",
+    .instance_size          = sizeof(BDRVReplaceTestState),
+
+    .bdrv_close             = bdrv_replace_test_close,
+    .bdrv_co_preadv         = bdrv_replace_test_co_preadv,
+
+    .bdrv_co_drain_begin    = bdrv_replace_test_co_drain_begin,
+    .bdrv_co_drain_end      = bdrv_replace_test_co_drain_end,
+
+    .bdrv_child_perm        = bdrv_default_perms,
+};
+
+static void coroutine_fn test_replace_child_mid_drain_read_co(void *opaque)
+{
+    int ret;
+    char data;
+
+    ret = blk_co_pread(opaque, 0, 1, &data, 0);
+    g_assert(ret >= 0);
+}
+
+/**
+ * We test two things:
+ * (1) bdrv_replace_child_noperm() must not undrain the parent if both
+ *     children are drained.
+ * (2) bdrv_replace_child_noperm() must never flush I/O requests to a
+ *     drained child.  If the old child is drained, it must flush I/O
+ *     requests after the new one has been attached.  If the new child
+ *     is drained, it must flush I/O requests before the old one is
+ *     detached.
+ *
+ * To do so, we create one parent node and two child nodes; then
+ * attach one of the children (old_child_bs) to the parent, then
+ * drain both old_child_bs and new_child_bs according to
+ * old_drain_count and new_drain_count, respectively, and finally
+ * we invoke bdrv_replace_node() to replace old_child_bs by
+ * new_child_bs.
+ *
+ * The test block driver we use here (bdrv_replace_test) has a read
+ * function that:
+ * - For the parent node, can optionally yield, and then forwards the
+ *   read to bdrv_preadv(),
+ * - For the child node, just returns immediately.
+ *
+ * If the read yields, the drain_begin function will wake it up.
+ *
+ * The drain_end function issues a read on the parent once it is fully
+ * undrained (which simulates requests starting to come in again).
+ */
+static void do_test_replace_child_mid_drain(int old_drain_count,
+                                            int new_drain_count)
+{
+    BlockBackend *parent_blk;
+    BlockDriverState *parent_bs;
+    BlockDriverState *old_child_bs, *new_child_bs;
+    BDRVReplaceTestState *parent_s;
+    BDRVReplaceTestState *old_child_s, *new_child_s;
+    Coroutine *io_co;
+    int i;
+
+    parent_bs = bdrv_new_open_driver(&bdrv_replace_test, "parent", 0,
+                                     &error_abort);
+    parent_s = parent_bs->opaque;
+
+    parent_blk = blk_new(qemu_get_aio_context(),
+                         BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
+    blk_insert_bs(parent_blk, parent_bs, &error_abort);
+
+    old_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "old-child", 0,
+                                        &error_abort);
+    new_child_bs = bdrv_new_open_driver(&bdrv_replace_test, "new-child", 0,
+                                        &error_abort);
+    old_child_s = old_child_bs->opaque;
+    new_child_s = new_child_bs->opaque;
+
+    /* So that we can read something */
+    parent_bs->total_sectors = 1;
+    old_child_bs->total_sectors = 1;
+    new_child_bs->total_sectors = 1;
+
+    bdrv_ref(old_child_bs);
+    parent_bs->backing = bdrv_attach_child(parent_bs, old_child_bs, "child",
+                                           &child_of_bds, BDRV_CHILD_COW,
+                                           &error_abort);
+
+    for (i = 0; i < old_drain_count; i++) {
+        bdrv_drained_begin(old_child_bs);
+    }
+    for (i = 0; i < new_drain_count; i++) {
+        bdrv_drained_begin(new_child_bs);
+    }
+
+    if (!old_drain_count) {
+        /*
+         * Start a read operation that will yield, so it will not
+         * complete before the node is drained.
+         */
+        parent_s->yield_before_read = true;
+        io_co = qemu_coroutine_create(test_replace_child_mid_drain_read_co,
+                                      parent_blk);
+        qemu_coroutine_enter(io_co);
+    }
+
+    /* If we have started a read operation, it should have yielded */
+    g_assert(!parent_s->has_read);
+
+    /* Reset drained status so we can see what bdrv_replace_node() does */
+    parent_s->was_drained = false;
+    parent_s->was_undrained = false;
+
+    g_assert(parent_bs->quiesce_counter == old_drain_count);
+    bdrv_replace_node(old_child_bs, new_child_bs, &error_abort);
+    g_assert(parent_bs->quiesce_counter == new_drain_count);
+
+    if (!old_drain_count && !new_drain_count) {
+        /*
+         * From undrained to undrained drains and undrains the parent,
+         * because bdrv_replace_node() contains a drained section for
+         * @old_child_bs.
+         */
+        g_assert(parent_s->was_drained && parent_s->was_undrained);
+    } else if (!old_drain_count && new_drain_count) {
+        /*
+         * From undrained to drained should drain the parent and keep
+         * it that way.
+         */
+        g_assert(parent_s->was_drained && !parent_s->was_undrained);
+    } else if (old_drain_count && !new_drain_count) {
+        /*
+         * From drained to undrained should undrain the parent and
+         * keep it that way.
+         */
+        g_assert(!parent_s->was_drained && parent_s->was_undrained);
+    } else /* if (old_drain_count && new_drain_count) */ {
+        /*
+         * From drained to drained must not undrain the parent at any
+         * point
+         */
+        g_assert(!parent_s->was_drained && !parent_s->was_undrained);
+    }
+
+    if (!old_drain_count || !new_drain_count) {
+        /*
+         * If !old_drain_count, we have started a read request before
+         * bdrv_replace_node().  If !new_drain_count, the parent must
+         * have been undrained at some point, and
+         * bdrv_replace_test_co_drain_end() starts a read request
+         * then.
+         */
+        g_assert(parent_s->has_read);
+    } else {
+        /*
+         * If the parent was never undrained, there is no way to start
+         * a read request.
+         */
+        g_assert(!parent_s->has_read);
+    }
+
+    /* A drained child must have not received any request */
+    g_assert(!(old_drain_count && old_child_s->has_read));
+    g_assert(!(new_drain_count && new_child_s->has_read));
+
+    for (i = 0; i < new_drain_count; i++) {
+        bdrv_drained_end(new_child_bs);
+    }
+    for (i = 0; i < old_drain_count; i++) {
+        bdrv_drained_end(old_child_bs);
+    }
+
+    /*
+     * By now, bdrv_replace_test_co_drain_end() must have been called
+     * at some point while the new child was attached to the parent.
+     */
+    g_assert(parent_s->has_read);
+    g_assert(new_child_s->has_read);
+
+    blk_unref(parent_blk);
+    bdrv_unref(parent_bs);
+    bdrv_unref(old_child_bs);
+    bdrv_unref(new_child_bs);
+}
+
+static void test_replace_child_mid_drain(void)
+{
+    int old_drain_count, new_drain_count;
+
+    for (old_drain_count = 0; old_drain_count < 2; old_drain_count++) {
+        for (new_drain_count = 0; new_drain_count < 2; new_drain_count++) {
+            do_test_replace_child_mid_drain(old_drain_count, new_drain_count);
+        }
+    }
+}
+
 int main(int argc, char **argv)
 {
     int ret;
@@ -1338,6 +2182,27 @@ int main(int argc, char **argv)
     g_test_add_func("/bdrv-drain/blockjob/drain_subtree",
                     test_blockjob_drain_subtree);
 
+    g_test_add_func("/bdrv-drain/blockjob/error/drain_all",
+                    test_blockjob_error_drain_all);
+    g_test_add_func("/bdrv-drain/blockjob/error/drain",
+                    test_blockjob_error_drain);
+    g_test_add_func("/bdrv-drain/blockjob/error/drain_subtree",
+                    test_blockjob_error_drain_subtree);
+
+    g_test_add_func("/bdrv-drain/blockjob/iothread/drain_all",
+                    test_blockjob_iothread_drain_all);
+    g_test_add_func("/bdrv-drain/blockjob/iothread/drain",
+                    test_blockjob_iothread_drain);
+    g_test_add_func("/bdrv-drain/blockjob/iothread/drain_subtree",
+                    test_blockjob_iothread_drain_subtree);
+
+    g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_all",
+                    test_blockjob_iothread_error_drain_all);
+    g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain",
+                    test_blockjob_iothread_error_drain);
+    g_test_add_func("/bdrv-drain/blockjob/iothread/error/drain_subtree",
+                    test_blockjob_iothread_error_drain_subtree);
+
     g_test_add_func("/bdrv-drain/deletion/drain", test_delete_by_drain);
     g_test_add_func("/bdrv-drain/detach/drain_all", test_detach_by_drain_all);
     g_test_add_func("/bdrv-drain/detach/drain", test_detach_by_drain);
@@ -1347,6 +2212,17 @@ int main(int argc, char **argv)
 
     g_test_add_func("/bdrv-drain/attach/drain", test_append_to_drained);
 
+    g_test_add_func("/bdrv-drain/set_aio_context", test_set_aio_context);
+
+    g_test_add_func("/bdrv-drain/blockjob/commit_by_drained_end",
+                    test_blockjob_commit_by_drained_end);
+
+    g_test_add_func("/bdrv-drain/bdrv_drop_intermediate/poll",
+                    test_drop_intermediate_poll);
+
+    g_test_add_func("/bdrv-drain/replace_child/mid-drain",
+                    test_replace_child_mid_drain);
+
     ret = g_test_run();
     qemu_event_destroy(&done_event);
     return ret;