]> git.proxmox.com Git - mirror_qemu.git/blobdiff - block.c
Merge remote-tracking branch 'remotes/bonzini/tags/for-upstream' into staging
[mirror_qemu.git] / block.c
diff --git a/block.c b/block.c
index 6999aad4460b1ae1651217374f9575d12d22eb32..c139540f2bac727c17c0c6d150dc1d5c2f813618 100644 (file)
--- a/block.c
+++ b/block.c
@@ -936,6 +936,20 @@ static int bdrv_child_cb_inactivate(BdrvChild *child)
     return 0;
 }
 
+static bool bdrv_child_cb_can_set_aio_ctx(BdrvChild *child, AioContext *ctx,
+                                          GSList **ignore, Error **errp)
+{
+    BlockDriverState *bs = child->opaque;
+    return bdrv_can_set_aio_context(bs, ctx, ignore, errp);
+}
+
+static void bdrv_child_cb_set_aio_ctx(BdrvChild *child, AioContext *ctx,
+                                      GSList **ignore)
+{
+    BlockDriverState *bs = child->opaque;
+    return bdrv_set_aio_context_ignore(bs, ctx, ignore);
+}
+
 /*
  * Returns the options and flags that a temporary snapshot should get, based on
  * the originally requested flags (the originally requested image will have
@@ -1003,6 +1017,8 @@ const BdrvChildRole child_file = {
     .attach          = bdrv_child_cb_attach,
     .detach          = bdrv_child_cb_detach,
     .inactivate      = bdrv_child_cb_inactivate,
+    .can_set_aio_ctx = bdrv_child_cb_can_set_aio_ctx,
+    .set_aio_ctx     = bdrv_child_cb_set_aio_ctx,
 };
 
 /*
@@ -1029,6 +1045,8 @@ const BdrvChildRole child_format = {
     .attach          = bdrv_child_cb_attach,
     .detach          = bdrv_child_cb_detach,
     .inactivate      = bdrv_child_cb_inactivate,
+    .can_set_aio_ctx = bdrv_child_cb_can_set_aio_ctx,
+    .set_aio_ctx     = bdrv_child_cb_set_aio_ctx,
 };
 
 static void bdrv_backing_attach(BdrvChild *c)
@@ -1152,6 +1170,8 @@ const BdrvChildRole child_backing = {
     .drained_end     = bdrv_child_cb_drained_end,
     .inactivate      = bdrv_child_cb_inactivate,
     .update_filename = bdrv_backing_update_filename,
+    .can_set_aio_ctx = bdrv_child_cb_can_set_aio_ctx,
+    .set_aio_ctx     = bdrv_child_cb_set_aio_ctx,
 };
 
 static int bdrv_open_flags(BlockDriverState *bs, int flags)
@@ -1686,9 +1706,12 @@ static int bdrv_fill_options(QDict **options, const char *filename,
 
 static int bdrv_child_check_perm(BdrvChild *c, BlockReopenQueue *q,
                                  uint64_t perm, uint64_t shared,
-                                 GSList *ignore_children, Error **errp);
+                                 GSList *ignore_children,
+                                 bool *tighten_restrictions, Error **errp);
 static void bdrv_child_abort_perm_update(BdrvChild *c);
 static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared);
+static void bdrv_get_cumulative_perm(BlockDriverState *bs, uint64_t *perm,
+                                     uint64_t *shared_perm);
 
 typedef struct BlockReopenQueueEntry {
      bool prepared;
@@ -1759,23 +1782,61 @@ static void bdrv_child_perm(BlockDriverState *bs, BlockDriverState *child_bs,
  * permissions of all its parents. This involves checking whether all necessary
  * permission changes to child nodes can be performed.
  *
+ * Will set *tighten_restrictions to true if and only if new permissions have to
+ * be taken or currently shared permissions are to be unshared.  Otherwise,
+ * errors are not fatal as long as the caller accepts that the restrictions
+ * remain tighter than they need to be.  The caller still has to abort the
+ * transaction.
+ * @tighten_restrictions cannot be used together with @q: When reopening, we may
+ * encounter fatal errors even though no restrictions are to be tightened.  For
+ * example, changing a node from RW to RO will fail if the WRITE permission is
+ * to be kept.
+ *
  * A call to this function must always be followed by a call to bdrv_set_perm()
  * or bdrv_abort_perm_update().
  */
 static int bdrv_check_perm(BlockDriverState *bs, BlockReopenQueue *q,
                            uint64_t cumulative_perms,
                            uint64_t cumulative_shared_perms,
-                           GSList *ignore_children, Error **errp)
+                           GSList *ignore_children,
+                           bool *tighten_restrictions, Error **errp)
 {
     BlockDriver *drv = bs->drv;
     BdrvChild *c;
     int ret;
 
+    assert(!q || !tighten_restrictions);
+
+    if (tighten_restrictions) {
+        uint64_t current_perms, current_shared;
+        uint64_t added_perms, removed_shared_perms;
+
+        bdrv_get_cumulative_perm(bs, &current_perms, &current_shared);
+
+        added_perms = cumulative_perms & ~current_perms;
+        removed_shared_perms = current_shared & ~cumulative_shared_perms;
+
+        *tighten_restrictions = added_perms || removed_shared_perms;
+    }
+
     /* Write permissions never work with read-only images */
     if ((cumulative_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) &&
         !bdrv_is_writable_after_reopen(bs, q))
     {
-        error_setg(errp, "Block node is read-only");
+        if (!bdrv_is_writable_after_reopen(bs, NULL)) {
+            error_setg(errp, "Block node is read-only");
+        } else {
+            uint64_t current_perms, current_shared;
+            bdrv_get_cumulative_perm(bs, &current_perms, &current_shared);
+            if (current_perms & (BLK_PERM_WRITE | BLK_PERM_WRITE_UNCHANGED)) {
+                error_setg(errp, "Cannot make block node read-only, there is "
+                           "a writer on it");
+            } else {
+                error_setg(errp, "Cannot make block node read-only and create "
+                           "a writer on it");
+            }
+        }
+
         return -EPERM;
     }
 
@@ -1798,11 +1859,18 @@ static int bdrv_check_perm(BlockDriverState *bs, BlockReopenQueue *q,
     /* Check all children */
     QLIST_FOREACH(c, &bs->children, next) {
         uint64_t cur_perm, cur_shared;
+        bool child_tighten_restr;
+
         bdrv_child_perm(bs, c->bs, c, c->role, q,
                         cumulative_perms, cumulative_shared_perms,
                         &cur_perm, &cur_shared);
-        ret = bdrv_child_check_perm(c, q, cur_perm, cur_shared,
-                                    ignore_children, errp);
+        ret = bdrv_child_check_perm(c, q, cur_perm, cur_shared, ignore_children,
+                                    tighten_restrictions ? &child_tighten_restr
+                                                         : NULL,
+                                    errp);
+        if (tighten_restrictions) {
+            *tighten_restrictions |= child_tighten_restr;
+        }
         if (ret < 0) {
             return ret;
         }
@@ -1926,17 +1994,23 @@ char *bdrv_perm_names(uint64_t perm)
  * set, the BdrvChild objects in this list are ignored in the calculations;
  * this allows checking permission updates for an existing reference.
  *
+ * See bdrv_check_perm() for the semantics of @tighten_restrictions.
+ *
  * Needs to be followed by a call to either bdrv_set_perm() or
  * bdrv_abort_perm_update(). */
 static int bdrv_check_update_perm(BlockDriverState *bs, BlockReopenQueue *q,
                                   uint64_t new_used_perm,
                                   uint64_t new_shared_perm,
-                                  GSList *ignore_children, Error **errp)
+                                  GSList *ignore_children,
+                                  bool *tighten_restrictions,
+                                  Error **errp)
 {
     BdrvChild *c;
     uint64_t cumulative_perms = new_used_perm;
     uint64_t cumulative_shared_perms = new_shared_perm;
 
+    assert(!q || !tighten_restrictions);
+
     /* There is no reason why anyone couldn't tolerate write_unchanged */
     assert(new_shared_perm & BLK_PERM_WRITE_UNCHANGED);
 
@@ -1948,6 +2022,11 @@ static int bdrv_check_update_perm(BlockDriverState *bs, BlockReopenQueue *q,
         if ((new_used_perm & c->shared_perm) != new_used_perm) {
             char *user = bdrv_child_user_desc(c);
             char *perm_names = bdrv_perm_names(new_used_perm & ~c->shared_perm);
+
+            if (tighten_restrictions) {
+                *tighten_restrictions = true;
+            }
+
             error_setg(errp, "Conflicts with use by %s as '%s', which does not "
                              "allow '%s' on %s",
                        user, c->name, perm_names, bdrv_get_node_name(c->bs));
@@ -1959,6 +2038,11 @@ static int bdrv_check_update_perm(BlockDriverState *bs, BlockReopenQueue *q,
         if ((c->perm & new_shared_perm) != c->perm) {
             char *user = bdrv_child_user_desc(c);
             char *perm_names = bdrv_perm_names(c->perm & ~new_shared_perm);
+
+            if (tighten_restrictions) {
+                *tighten_restrictions = true;
+            }
+
             error_setg(errp, "Conflicts with use by %s as '%s', which uses "
                              "'%s' on %s",
                        user, c->name, perm_names, bdrv_get_node_name(c->bs));
@@ -1972,19 +2056,21 @@ static int bdrv_check_update_perm(BlockDriverState *bs, BlockReopenQueue *q,
     }
 
     return bdrv_check_perm(bs, q, cumulative_perms, cumulative_shared_perms,
-                           ignore_children, errp);
+                           ignore_children, tighten_restrictions, errp);
 }
 
 /* Needs to be followed by a call to either bdrv_child_set_perm() or
  * bdrv_child_abort_perm_update(). */
 static int bdrv_child_check_perm(BdrvChild *c, BlockReopenQueue *q,
                                  uint64_t perm, uint64_t shared,
-                                 GSList *ignore_children, Error **errp)
+                                 GSList *ignore_children,
+                                 bool *tighten_restrictions, Error **errp)
 {
     int ret;
 
     ignore_children = g_slist_prepend(g_slist_copy(ignore_children), c);
-    ret = bdrv_check_update_perm(c->bs, q, perm, shared, ignore_children, errp);
+    ret = bdrv_check_update_perm(c->bs, q, perm, shared, ignore_children,
+                                 tighten_restrictions, errp);
     g_slist_free(ignore_children);
 
     if (ret < 0) {
@@ -2035,11 +2121,26 @@ static void bdrv_child_abort_perm_update(BdrvChild *c)
 int bdrv_child_try_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared,
                             Error **errp)
 {
+    Error *local_err = NULL;
     int ret;
+    bool tighten_restrictions;
 
-    ret = bdrv_child_check_perm(c, NULL, perm, shared, NULL, errp);
+    ret = bdrv_child_check_perm(c, NULL, perm, shared, NULL,
+                                &tighten_restrictions, &local_err);
     if (ret < 0) {
         bdrv_child_abort_perm_update(c);
+        if (tighten_restrictions) {
+            error_propagate(errp, local_err);
+        } else {
+            /*
+             * Our caller may intend to only loosen restrictions and
+             * does not expect this function to fail.  Errors are not
+             * fatal in such a case, so we can just hide them from our
+             * caller.
+             */
+            error_free(local_err);
+            ret = 0;
+        }
         return ret;
     }
 
@@ -2048,6 +2149,18 @@ int bdrv_child_try_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared,
     return 0;
 }
 
+int bdrv_child_refresh_perms(BlockDriverState *bs, BdrvChild *c, Error **errp)
+{
+    uint64_t parent_perms, parent_shared;
+    uint64_t perms, shared;
+
+    bdrv_get_cumulative_perm(bs, &parent_perms, &parent_shared);
+    bdrv_child_perm(bs, c->bs, c, c->role, NULL, parent_perms, parent_shared,
+                    &perms, &shared);
+
+    return bdrv_child_try_set_perm(c, perms, shared, errp);
+}
+
 void bdrv_filter_default_perms(BlockDriverState *bs, BdrvChild *c,
                                const BdrvChildRole *role,
                                BlockReopenQueue *reopen_queue,
@@ -2193,33 +2306,69 @@ static void bdrv_replace_child(BdrvChild *child, BlockDriverState *new_bs)
 
     bdrv_replace_child_noperm(child, new_bs);
 
+    /*
+     * Start with the new node's permissions.  If @new_bs is a (direct
+     * or indirect) child of @old_bs, we must complete the permission
+     * update on @new_bs before we loosen the restrictions on @old_bs.
+     * Otherwise, bdrv_check_perm() on @old_bs would re-initiate
+     * updating the permissions of @new_bs, and thus not purely loosen
+     * restrictions.
+     */
+    if (new_bs) {
+        bdrv_get_cumulative_perm(new_bs, &perm, &shared_perm);
+        bdrv_set_perm(new_bs, perm, shared_perm);
+    }
+
     if (old_bs) {
         /* Update permissions for old node. This is guaranteed to succeed
          * because we're just taking a parent away, so we're loosening
          * restrictions. */
+        bool tighten_restrictions;
+        int ret;
+
         bdrv_get_cumulative_perm(old_bs, &perm, &shared_perm);
-        bdrv_check_perm(old_bs, NULL, perm, shared_perm, NULL, &error_abort);
-        bdrv_set_perm(old_bs, perm, shared_perm);
-    }
+        ret = bdrv_check_perm(old_bs, NULL, perm, shared_perm, NULL,
+                              &tighten_restrictions, NULL);
+        assert(tighten_restrictions == false);
+        if (ret < 0) {
+            /* We only tried to loosen restrictions, so errors are not fatal */
+            bdrv_abort_perm_update(old_bs);
+        } else {
+            bdrv_set_perm(old_bs, perm, shared_perm);
+        }
 
-    if (new_bs) {
-        bdrv_get_cumulative_perm(new_bs, &perm, &shared_perm);
-        bdrv_set_perm(new_bs, perm, shared_perm);
+        /* When the parent requiring a non-default AioContext is removed, the
+         * node moves back to the main AioContext */
+        bdrv_try_set_aio_context(old_bs, qemu_get_aio_context(), NULL);
     }
 }
 
+/*
+ * This function steals the reference to child_bs from the caller.
+ * That reference is later dropped by bdrv_root_unref_child().
+ *
+ * On failure NULL is returned, errp is set and the reference to
+ * child_bs is also dropped.
+ *
+ * The caller must hold the AioContext lock @child_bs, but not that of @ctx
+ * (unless @child_bs is already in @ctx).
+ */
 BdrvChild *bdrv_root_attach_child(BlockDriverState *child_bs,
                                   const char *child_name,
                                   const BdrvChildRole *child_role,
+                                  AioContext *ctx,
                                   uint64_t perm, uint64_t shared_perm,
                                   void *opaque, Error **errp)
 {
     BdrvChild *child;
+    Error *local_err = NULL;
     int ret;
 
-    ret = bdrv_check_update_perm(child_bs, NULL, perm, shared_perm, NULL, errp);
+    ret = bdrv_check_update_perm(child_bs, NULL, perm, shared_perm, NULL, NULL,
+                                 errp);
     if (ret < 0) {
         bdrv_abort_perm_update(child_bs);
+        bdrv_unref(child_bs);
         return NULL;
     }
 
@@ -2233,12 +2382,48 @@ BdrvChild *bdrv_root_attach_child(BlockDriverState *child_bs,
         .opaque         = opaque,
     };
 
+    /* If the AioContexts don't match, first try to move the subtree of
+     * child_bs into the AioContext of the new parent. If this doesn't work,
+     * try moving the parent into the AioContext of child_bs instead. */
+    if (bdrv_get_aio_context(child_bs) != ctx) {
+        ret = bdrv_try_set_aio_context(child_bs, ctx, &local_err);
+        if (ret < 0 && child_role->can_set_aio_ctx) {
+            GSList *ignore = g_slist_prepend(NULL, child);;
+            ctx = bdrv_get_aio_context(child_bs);
+            if (child_role->can_set_aio_ctx(child, ctx, &ignore, NULL)) {
+                error_free(local_err);
+                ret = 0;
+                g_slist_free(ignore);
+                ignore = g_slist_prepend(NULL, child);;
+                child_role->set_aio_ctx(child, ctx, &ignore);
+            }
+            g_slist_free(ignore);
+        }
+        if (ret < 0) {
+            error_propagate(errp, local_err);
+            g_free(child);
+            bdrv_abort_perm_update(child_bs);
+            return NULL;
+        }
+    }
+
     /* This performs the matching bdrv_set_perm() for the above check. */
     bdrv_replace_child(child, child_bs);
 
     return child;
 }
 
+/*
+ * This function transfers the reference to child_bs from the caller
+ * to parent_bs. That reference is later dropped by parent_bs on
+ * bdrv_close() or if someone calls bdrv_unref_child().
+ *
+ * On failure NULL is returned, errp is set and the reference to
+ * child_bs is also dropped.
+ *
+ * If @parent_bs and @child_bs are in different AioContexts, the caller must
+ * hold the AioContext lock for @child_bs, but not for @parent_bs.
+ */
 BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
                              BlockDriverState *child_bs,
                              const char *child_name,
@@ -2251,11 +2436,11 @@ BdrvChild *bdrv_attach_child(BlockDriverState *parent_bs,
     bdrv_get_cumulative_perm(parent_bs, &perm, &shared_perm);
 
     assert(parent_bs->drv);
-    assert(bdrv_get_aio_context(parent_bs) == bdrv_get_aio_context(child_bs));
     bdrv_child_perm(parent_bs, child_bs, NULL, child_role, NULL,
                     perm, shared_perm, &perm, &shared_perm);
 
     child = bdrv_root_attach_child(child_bs, child_name, child_role,
+                                   bdrv_get_aio_context(parent_bs),
                                    perm, shared_perm, parent_bs, errp);
     if (child == NULL) {
         return NULL;
@@ -2366,12 +2551,9 @@ void bdrv_set_backing_hd(BlockDriverState *bs, BlockDriverState *backing_hd,
     /* If backing_hd was already part of bs's backing chain, and
      * inherits_from pointed recursively to bs then let's update it to
      * point directly to bs (else it will become NULL). */
-    if (update_inherits_from) {
+    if (bs->backing && update_inherits_from) {
         backing_hd->inherits_from = bs;
     }
-    if (!bs->backing) {
-        bdrv_unref(backing_hd);
-    }
 
 out:
     bdrv_refresh_limits(bs, NULL);
@@ -2469,7 +2651,6 @@ int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options,
         ret = -EINVAL;
         goto free_exit;
     }
-    bdrv_set_aio_context(backing_hd, bdrv_get_aio_context(bs));
 
     if (implicit_backing) {
         bdrv_refresh_filename(backing_hd);
@@ -2559,7 +2740,6 @@ BdrvChild *bdrv_open_child(const char *filename,
                            const BdrvChildRole *child_role,
                            bool allow_none, Error **errp)
 {
-    BdrvChild *c;
     BlockDriverState *bs;
 
     bs = bdrv_open_child_bs(filename, options, bdref_key, parent, child_role,
@@ -2568,13 +2748,7 @@ BdrvChild *bdrv_open_child(const char *filename,
         return NULL;
     }
 
-    c = bdrv_attach_child(parent, bs, bdref_key, child_role, errp);
-    if (!c) {
-        bdrv_unref(bs);
-        return NULL;
-    }
-
-    return c;
+    return bdrv_attach_child(parent, bs, bdref_key, child_role, errp);
 }
 
 /* TODO Future callers may need to specify parent/child_role in order for
@@ -2843,7 +3017,7 @@ static BlockDriverState *bdrv_open_inherit(const char *filename,
             /* Not requesting BLK_PERM_CONSISTENT_READ because we're only
              * looking at the header to guess the image format. This works even
              * in cases where a guest would not see a consistent state. */
-            file = blk_new(0, BLK_PERM_ALL);
+            file = blk_new(bdrv_get_aio_context(file_bs), 0, BLK_PERM_ALL);
             blk_insert_bs(file, file_bs, &local_err);
             bdrv_unref(file_bs);
             if (local_err) {
@@ -3272,7 +3446,7 @@ int bdrv_reopen_multiple(BlockReopenQueue *bs_queue, Error **errp)
     QSIMPLEQ_FOREACH(bs_entry, bs_queue, entry) {
         BDRVReopenState *state = &bs_entry->state;
         ret = bdrv_check_perm(state->bs, bs_queue, state->perm,
-                              state->shared_perm, NULL, errp);
+                              state->shared_perm, NULL, NULL, errp);
         if (ret < 0) {
             goto cleanup_perm;
         }
@@ -3284,7 +3458,7 @@ int bdrv_reopen_multiple(BlockReopenQueue *bs_queue, Error **errp)
                             state->perm, state->shared_perm,
                             &nperm, &nshared);
             ret = bdrv_check_update_perm(state->new_backing_bs, NULL,
-                                         nperm, nshared, NULL, errp);
+                                         nperm, nshared, NULL, NULL, errp);
             if (ret < 0) {
                 goto cleanup_perm;
             }
@@ -3828,7 +4002,6 @@ static void bdrv_close(BlockDriverState *bs)
     BdrvAioNotifier *ban, *ban_next;
     BdrvChild *child, *next;
 
-    assert(!bs->job);
     assert(!bs->refcnt);
 
     bdrv_drained_begin(bs); /* complete I/O */
@@ -3842,22 +4015,12 @@ static void bdrv_close(BlockDriverState *bs)
         bs->drv = NULL;
     }
 
-    bdrv_set_backing_hd(bs, NULL, &error_abort);
-
-    if (bs->file != NULL) {
-        bdrv_unref_child(bs, bs->file);
-        bs->file = NULL;
-    }
-
     QLIST_FOREACH_SAFE(child, &bs->children, next, next) {
-        /* TODO Remove bdrv_unref() from drivers' close function and use
-         * bdrv_unref_child() here */
-        if (child->bs->inherits_from == bs) {
-            child->bs->inherits_from = NULL;
-        }
-        bdrv_detach_child(child);
+        bdrv_unref_child(bs, child);
     }
 
+    bs->backing = NULL;
+    bs->file = NULL;
     g_free(bs->opaque);
     bs->opaque = NULL;
     atomic_set(&bs->copy_on_read, 0);
@@ -3986,13 +4149,13 @@ void bdrv_replace_node(BlockDriverState *from, BlockDriverState *to,
     uint64_t perm = 0, shared = BLK_PERM_ALL;
     int ret;
 
-    assert(!atomic_read(&from->in_flight));
-    assert(!atomic_read(&to->in_flight));
-
     /* Make sure that @from doesn't go away until we have successfully attached
      * all of its parents to @to. */
     bdrv_ref(from);
 
+    assert(qemu_get_current_aio_context() == qemu_get_aio_context());
+    bdrv_drained_begin(from);
+
     /* Put all parents into @list and calculate their cumulative permissions */
     QLIST_FOREACH_SAFE(c, &from->parents, next_parent, next) {
         assert(c->bs == from);
@@ -4011,7 +4174,7 @@ void bdrv_replace_node(BlockDriverState *from, BlockDriverState *to,
 
     /* Check whether the required permissions can be granted on @to, ignoring
      * all BdrvChild in @list so that they can't block themselves. */
-    ret = bdrv_check_update_perm(to, NULL, perm, shared, list, errp);
+    ret = bdrv_check_update_perm(to, NULL, perm, shared, list, NULL, errp);
     if (ret < 0) {
         bdrv_abort_perm_update(to);
         goto out;
@@ -4033,6 +4196,7 @@ void bdrv_replace_node(BlockDriverState *from, BlockDriverState *to,
 
 out:
     g_slist_free(list);
+    bdrv_drained_end(from);
     bdrv_unref(from);
 }
 
@@ -4078,7 +4242,6 @@ out:
 
 static void bdrv_delete(BlockDriverState *bs)
 {
-    assert(!bs->job);
     assert(bdrv_op_blocker_is_empty(bs));
     assert(!bs->refcnt);
 
@@ -4358,7 +4521,7 @@ int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base,
         /* Check whether we are allowed to switch c from top to base */
         GSList *ignore_children = g_slist_prepend(NULL, c);
         ret = bdrv_check_update_perm(base, NULL, c->perm, c->shared_perm,
-                                     ignore_children, &local_err);
+                                     ignore_children, NULL, &local_err);
         g_slist_free(ignore_children);
         if (ret < 0) {
             error_report_err(local_err);
@@ -5133,7 +5296,7 @@ static void coroutine_fn bdrv_co_invalidate_cache(BlockDriverState *bs,
      */
     bs->open_flags &= ~BDRV_O_INACTIVE;
     bdrv_get_cumulative_perm(bs, &perm, &shared_perm);
-    ret = bdrv_check_perm(bs, NULL, perm, shared_perm, NULL, &local_err);
+    ret = bdrv_check_perm(bs, NULL, perm, shared_perm, NULL, NULL, &local_err);
     if (ret < 0) {
         bs->open_flags |= BDRV_O_INACTIVE;
         error_propagate(errp, local_err);
@@ -5247,6 +5410,7 @@ static bool bdrv_has_bds_parent(BlockDriverState *bs, bool only_active)
 static int bdrv_inactivate_recurse(BlockDriverState *bs)
 {
     BdrvChild *child, *parent;
+    bool tighten_restrictions;
     uint64_t perm, shared_perm;
     int ret;
 
@@ -5283,8 +5447,15 @@ static int bdrv_inactivate_recurse(BlockDriverState *bs)
 
     /* Update permissions, they may differ for inactive nodes */
     bdrv_get_cumulative_perm(bs, &perm, &shared_perm);
-    bdrv_check_perm(bs, NULL, perm, shared_perm, NULL, &error_abort);
-    bdrv_set_perm(bs, perm, shared_perm);
+    ret = bdrv_check_perm(bs, NULL, perm, shared_perm, NULL,
+                          &tighten_restrictions, NULL);
+    assert(tighten_restrictions == false);
+    if (ret < 0) {
+        /* We only tried to loosen restrictions, so errors are not fatal */
+        bdrv_abort_perm_update(bs);
+    } else {
+        bdrv_set_perm(bs, perm, shared_perm);
+    }
 
 
     /* Recursively inactivate children */
@@ -5666,10 +5837,9 @@ static void bdrv_do_remove_aio_context_notifier(BdrvAioNotifier *ban)
     g_free(ban);
 }
 
-void bdrv_detach_aio_context(BlockDriverState *bs)
+static void bdrv_detach_aio_context(BlockDriverState *bs)
 {
     BdrvAioNotifier *baf, *baf_tmp;
-    BdrvChild *child;
 
     assert(!bs->walking_aio_notifiers);
     bs->walking_aio_notifiers = true;
@@ -5688,9 +5858,6 @@ void bdrv_detach_aio_context(BlockDriverState *bs)
     if (bs->drv && bs->drv->bdrv_detach_aio_context) {
         bs->drv->bdrv_detach_aio_context(bs);
     }
-    QLIST_FOREACH(child, &bs->children, next) {
-        bdrv_detach_aio_context(child->bs);
-    }
 
     if (bs->quiesce_counter) {
         aio_enable_external(bs->aio_context);
@@ -5698,11 +5865,10 @@ void bdrv_detach_aio_context(BlockDriverState *bs)
     bs->aio_context = NULL;
 }
 
-void bdrv_attach_aio_context(BlockDriverState *bs,
-                             AioContext *new_context)
+static void bdrv_attach_aio_context(BlockDriverState *bs,
+                                    AioContext *new_context)
 {
     BdrvAioNotifier *ban, *ban_tmp;
-    BdrvChild *child;
 
     if (bs->quiesce_counter) {
         aio_disable_external(new_context);
@@ -5710,9 +5876,6 @@ void bdrv_attach_aio_context(BlockDriverState *bs,
 
     bs->aio_context = new_context;
 
-    QLIST_FOREACH(child, &bs->children, next) {
-        bdrv_attach_aio_context(child->bs, new_context);
-    }
     if (bs->drv && bs->drv->bdrv_attach_aio_context) {
         bs->drv->bdrv_attach_aio_context(bs, new_context);
     }
@@ -5729,16 +5892,44 @@ void bdrv_attach_aio_context(BlockDriverState *bs,
     bs->walking_aio_notifiers = false;
 }
 
-/* The caller must own the AioContext lock for the old AioContext of bs, but it
- * must not own the AioContext lock for new_context (unless new_context is
- * the same as the current context of bs). */
-void bdrv_set_aio_context(BlockDriverState *bs, AioContext *new_context)
+/*
+ * Changes the AioContext used for fd handlers, timers, and BHs by this
+ * BlockDriverState and all its children and parents.
+ *
+ * The caller must own the AioContext lock for the old AioContext of bs, but it
+ * must not own the AioContext lock for new_context (unless new_context is the
+ * same as the current context of bs).
+ *
+ * @ignore will accumulate all visited BdrvChild object. The caller is
+ * responsible for freeing the list afterwards.
+ */
+void bdrv_set_aio_context_ignore(BlockDriverState *bs,
+                                 AioContext *new_context, GSList **ignore)
 {
+    BdrvChild *child;
+
     if (bdrv_get_aio_context(bs) == new_context) {
         return;
     }
 
     bdrv_drained_begin(bs);
+
+    QLIST_FOREACH(child, &bs->children, next) {
+        if (g_slist_find(*ignore, child)) {
+            continue;
+        }
+        *ignore = g_slist_prepend(*ignore, child);
+        bdrv_set_aio_context_ignore(child->bs, new_context, ignore);
+    }
+    QLIST_FOREACH(child, &bs->parents, next_parent) {
+        if (g_slist_find(*ignore, child)) {
+            continue;
+        }
+        assert(child->role->set_aio_ctx);
+        *ignore = g_slist_prepend(*ignore, child);
+        child->role->set_aio_ctx(child, new_context, ignore);
+    }
+
     bdrv_detach_aio_context(bs);
 
     /* This function executes in the old AioContext so acquire the new one in
@@ -5750,6 +5941,91 @@ void bdrv_set_aio_context(BlockDriverState *bs, AioContext *new_context)
     aio_context_release(new_context);
 }
 
+static bool bdrv_parent_can_set_aio_context(BdrvChild *c, AioContext *ctx,
+                                            GSList **ignore, Error **errp)
+{
+    if (g_slist_find(*ignore, c)) {
+        return true;
+    }
+    *ignore = g_slist_prepend(*ignore, c);
+
+    /* A BdrvChildRole that doesn't handle AioContext changes cannot
+     * tolerate any AioContext changes */
+    if (!c->role->can_set_aio_ctx) {
+        char *user = bdrv_child_user_desc(c);
+        error_setg(errp, "Changing iothreads is not supported by %s", user);
+        g_free(user);
+        return false;
+    }
+    if (!c->role->can_set_aio_ctx(c, ctx, ignore, errp)) {
+        assert(!errp || *errp);
+        return false;
+    }
+    return true;
+}
+
+bool bdrv_child_can_set_aio_context(BdrvChild *c, AioContext *ctx,
+                                    GSList **ignore, Error **errp)
+{
+    if (g_slist_find(*ignore, c)) {
+        return true;
+    }
+    *ignore = g_slist_prepend(*ignore, c);
+    return bdrv_can_set_aio_context(c->bs, ctx, ignore, errp);
+}
+
+/* @ignore will accumulate all visited BdrvChild object. The caller is
+ * responsible for freeing the list afterwards. */
+bool bdrv_can_set_aio_context(BlockDriverState *bs, AioContext *ctx,
+                              GSList **ignore, Error **errp)
+{
+    BdrvChild *c;
+
+    if (bdrv_get_aio_context(bs) == ctx) {
+        return true;
+    }
+
+    QLIST_FOREACH(c, &bs->parents, next_parent) {
+        if (!bdrv_parent_can_set_aio_context(c, ctx, ignore, errp)) {
+            return false;
+        }
+    }
+    QLIST_FOREACH(c, &bs->children, next) {
+        if (!bdrv_child_can_set_aio_context(c, ctx, ignore, errp)) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+int bdrv_child_try_set_aio_context(BlockDriverState *bs, AioContext *ctx,
+                                   BdrvChild *ignore_child, Error **errp)
+{
+    GSList *ignore;
+    bool ret;
+
+    ignore = ignore_child ? g_slist_prepend(NULL, ignore_child) : NULL;
+    ret = bdrv_can_set_aio_context(bs, ctx, &ignore, errp);
+    g_slist_free(ignore);
+
+    if (!ret) {
+        return -EPERM;
+    }
+
+    ignore = ignore_child ? g_slist_prepend(NULL, ignore_child) : NULL;
+    bdrv_set_aio_context_ignore(bs, ctx, &ignore);
+    g_slist_free(ignore);
+
+    return 0;
+}
+
+int bdrv_try_set_aio_context(BlockDriverState *bs, AioContext *ctx,
+                             Error **errp)
+{
+    return bdrv_child_try_set_aio_context(bs, ctx, NULL, errp);
+}
+
 void bdrv_add_aio_context_notifier(BlockDriverState *bs,
         void (*attached_aio_context)(AioContext *new_context, void *opaque),
         void (*detach_aio_context)(void *opaque), void *opaque)