diff --git a/block.c b/block.c
index c8586f41ba270bd5ce267a521719c198a4c79eae..6c128007fdc2ec008812930bc24eaff9019c478a 100644
--- a/block.c
+++ b/block.c
@@ -333,6 +333,10 @@ BlockDriverState *bdrv_new(void)
 
     qemu_co_queue_init(&bs->flush_queue);
 
+    for (i = 0; i < bdrv_drain_all_count; i++) {
+        bdrv_drained_begin(bs);
+    }
+
     QTAILQ_INSERT_TAIL(&all_bdrv_states, bs, bs_list);
 
     return bs;
@@ -1164,7 +1168,7 @@ static int bdrv_open_driver(BlockDriverState *bs, BlockDriver *drv,
                             int open_flags, Error **errp)
 {
     Error *local_err = NULL;
-    int ret;
+    int i, ret;
 
     bdrv_assign_node_name(bs, node_name, &local_err);
     if (local_err) {
@@ -1212,6 +1216,12 @@ static int bdrv_open_driver(BlockDriverState *bs, BlockDriver *drv,
     assert(bdrv_min_mem_align(bs) != 0);
     assert(is_power_of_2(bs->bl.request_alignment));
 
+    for (i = 0; i < bs->quiesce_counter; i++) {
+        if (drv->bdrv_co_drain_begin) {
+            drv->bdrv_co_drain_begin(bs);
+        }
+    }
+
     return 0;
 open_failed:
     bs->drv = NULL;
@@ -2033,7 +2043,12 @@ static void bdrv_replace_child_noperm(BdrvChild *child,
             child->role->detach(child);
         }
         if (old_bs->quiesce_counter && child->role->drained_end) {
-            for (i = 0; i < old_bs->quiesce_counter; i++) {
+            int num = old_bs->quiesce_counter;
+            if (child->role->parent_is_bds) {
+                num -= bdrv_drain_all_count;
+            }
+            assert(num >= 0);
+            for (i = 0; i < num; i++) {
                 child->role->drained_end(child);
             }
         }
@@ -2045,7 +2060,12 @@ static void bdrv_replace_child_noperm(BdrvChild *child,
     if (new_bs) {
         QLIST_INSERT_HEAD(&new_bs->parents, child, next_parent);
         if (new_bs->quiesce_counter && child->role->drained_begin) {
-            for (i = 0; i < new_bs->quiesce_counter; i++) {
+            int num = new_bs->quiesce_counter;
+            if (child->role->parent_is_bds) {
+                num -= bdrv_drain_all_count;
+            }
+            assert(num >= 0);
+            for (i = 0; i < num; i++) {
                 child->role->drained_begin(child);
             }
         }
@@ -4049,6 +4069,14 @@ BlockDriverState *bdrv_next_node(BlockDriverState *bs)
     return QTAILQ_NEXT(bs, node_list);
 }
 
+BlockDriverState *bdrv_next_all_states(BlockDriverState *bs)
+{
+    if (!bs) {
+        return QTAILQ_FIRST(&all_bdrv_states);
+    }
+    return QTAILQ_NEXT(bs, bs_list);
+}
+
 const char *bdrv_get_node_name(const BlockDriverState *bs)
 {
     return bs->node_name;
diff --git a/block/io.c b/block/io.c
index 1834a14aa6a868db156be4d3dd9cf91fa871f73d..ef4fedd36416d2569331f4905aaad39063fcaf8b 100644
--- a/block/io.c
+++ b/block/io.c
@@ -38,6 +38,8 @@
 /* Maximum bounce buffer for copy-on-read and write zeroes, in bytes */
 #define MAX_BOUNCE_BUFFER (32768 << BDRV_SECTOR_BITS)
 
+static AioWait drain_all_aio_wait;
+
 static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
     int64_t offset, int bytes, BdrvRequestFlags flags);
 
@@ -472,6 +474,29 @@ static void bdrv_drain_assert_idle(BlockDriverState *bs)
     }
 }
 
+unsigned int bdrv_drain_all_count = 0;
+
+static bool bdrv_drain_all_poll(void)
+{
+    BlockDriverState *bs = NULL;
+    bool result = false;
+
+    /* Execute pending BHs first (may modify the graph) and check everything
+     * else only after the BHs have executed. */
+    while (aio_poll(qemu_get_aio_context(), false));
+
+    /* bdrv_drain_poll() can't make changes to the graph and we are holding the
+     * main AioContext lock, so iterating bdrv_next_all_states() is safe. */
+    while ((bs = bdrv_next_all_states(bs))) {
+        AioContext *aio_context = bdrv_get_aio_context(bs);
+        aio_context_acquire(aio_context);
+        result |= bdrv_drain_poll(bs, false, NULL, true);
+        aio_context_release(aio_context);
+    }
+
+    return result;
+}
+
 /*
  * Wait for pending requests to complete across all BlockDriverStates
  *
@@ -486,45 +511,51 @@ static void bdrv_drain_assert_idle(BlockDriverState *bs)
  */
 void bdrv_drain_all_begin(void)
 {
-    BlockDriverState *bs;
-    BdrvNextIterator it;
+    BlockDriverState *bs = NULL;
 
     if (qemu_in_coroutine()) {
-        bdrv_co_yield_to_drain(NULL, true, false, NULL, false, true);
+        bdrv_co_yield_to_drain(NULL, true, false, NULL, true, true);
         return;
     }
 
-    /* BDRV_POLL_WHILE() for a node can only be called from its own I/O thread
-     * or the main loop AioContext. We potentially use BDRV_POLL_WHILE() on
-     * nodes in several different AioContexts, so make sure we're in the main
-     * context. */
+    /* AIO_WAIT_WHILE() with a NULL context can only be called from the main
+     * loop AioContext, so make sure we're in the main context. */
     assert(qemu_get_current_aio_context() == qemu_get_aio_context());
+    assert(bdrv_drain_all_count < INT_MAX);
+    bdrv_drain_all_count++;
 
-    for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
+    /* Quiesce all nodes, without polling in-flight requests yet. The graph
+     * cannot change during this loop. */
+    while ((bs = bdrv_next_all_states(bs))) {
         AioContext *aio_context = bdrv_get_aio_context(bs);
 
         aio_context_acquire(aio_context);
-        bdrv_do_drained_begin(bs, true, NULL, false, true);
+        bdrv_do_drained_begin(bs, false, NULL, true, false);
         aio_context_release(aio_context);
     }
 
-    for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
+    /* Now poll the in-flight requests */
+    AIO_WAIT_WHILE(&drain_all_aio_wait, NULL, bdrv_drain_all_poll());
+
+    while ((bs = bdrv_next_all_states(bs))) {
         bdrv_drain_assert_idle(bs);
     }
 }
 
 void bdrv_drain_all_end(void)
 {
-    BlockDriverState *bs;
-    BdrvNextIterator it;
+    BlockDriverState *bs = NULL;
 
-    for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
+    while ((bs = bdrv_next_all_states(bs))) {
         AioContext *aio_context = bdrv_get_aio_context(bs);
 
         aio_context_acquire(aio_context);
-        bdrv_do_drained_end(bs, true, NULL, false);
+        bdrv_do_drained_end(bs, false, NULL, true);
         aio_context_release(aio_context);
     }
+
+    assert(bdrv_drain_all_count > 0);
+    bdrv_drain_all_count--;
 }
 
 void bdrv_drain_all(void)
@@ -647,6 +678,7 @@ void bdrv_inc_in_flight(BlockDriverState *bs)
 void bdrv_wakeup(BlockDriverState *bs)
 {
     aio_wait_kick(bdrv_get_aio_wait(bs));
+    aio_wait_kick(&drain_all_aio_wait);
 }
 
 void bdrv_dec_in_flight(BlockDriverState *bs)
diff --git a/include/block/block.h b/include/block/block.h
index 836746e4e1841126904bd108818d426655d58fa4..b1d6fdb97a92f9e3ad832e9b123a4a7361bec33d 100644
--- a/include/block/block.h
+++ b/include/block/block.h
@@ -421,6 +421,7 @@ BlockDriverState *bdrv_lookup_bs(const char *device,
                                  Error **errp);
 bool bdrv_chain_contains(BlockDriverState *top, BlockDriverState *base);
 BlockDriverState *bdrv_next_node(BlockDriverState *bs);
+BlockDriverState *bdrv_next_all_states(BlockDriverState *bs);
 
 typedef struct BdrvNextIterator {
     enum {
diff --git a/include/block/block_int.h b/include/block/block_int.h
index 1abfc26d76efb1f0f1b1c26823a3c13c3b458a97..7cd7eed83b09b3d8a6e5af382e0e89954c1e0099 100644
--- a/include/block/block_int.h
+++ b/include/block/block_int.h
@@ -854,6 +854,7 @@ int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
     int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
     BdrvRequestFlags flags);
 
+extern unsigned int bdrv_drain_all_count;
 void bdrv_apply_subtree_drain(BdrvChild *child, BlockDriverState *new_parent);
 void bdrv_unapply_subtree_drain(BdrvChild *child, BlockDriverState *old_parent);