]> git.proxmox.com Git - mirror_qemu.git/blobdiff - block.c
Merge remote-tracking branch 'remotes/mst/tags/for_upstream' into staging
[mirror_qemu.git] / block.c
diff --git a/block.c b/block.c
index c6d4564bcf30058e5f67392f50ca4dd18c86897a..35e78e2172d52f1a75fd1c6fe0e24179e9b594db 100644 (file)
--- a/block.c
+++ b/block.c
@@ -152,53 +152,53 @@ int path_is_absolute(const char *path)
 #endif
 }
 
-/* if filename is absolute, just copy it to dest. Otherwise, build a
+/* if filename is absolute, just return its duplicate. Otherwise, build a
    path to it by considering it is relative to base_path. URL are
    supported. */
-void path_combine(char *dest, int dest_size,
-                  const char *base_path,
-                  const char *filename)
+char *path_combine(const char *base_path, const char *filename)
 {
+    const char *protocol_stripped = NULL;
     const char *p, *p1;
+    char *result;
     int len;
 
-    if (dest_size <= 0)
-        return;
     if (path_is_absolute(filename)) {
-        pstrcpy(dest, dest_size, filename);
-    } else {
-        const char *protocol_stripped = NULL;
+        return g_strdup(filename);
+    }
 
-        if (path_has_protocol(base_path)) {
-            protocol_stripped = strchr(base_path, ':');
-            if (protocol_stripped) {
-                protocol_stripped++;
-            }
+    if (path_has_protocol(base_path)) {
+        protocol_stripped = strchr(base_path, ':');
+        if (protocol_stripped) {
+            protocol_stripped++;
         }
-        p = protocol_stripped ?: base_path;
+    }
+    p = protocol_stripped ?: base_path;
 
-        p1 = strrchr(base_path, '/');
+    p1 = strrchr(base_path, '/');
 #ifdef _WIN32
-        {
-            const char *p2;
-            p2 = strrchr(base_path, '\\');
-            if (!p1 || p2 > p1)
-                p1 = p2;
+    {
+        const char *p2;
+        p2 = strrchr(base_path, '\\');
+        if (!p1 || p2 > p1) {
+            p1 = p2;
         }
+    }
 #endif
-        if (p1)
-            p1++;
-        else
-            p1 = base_path;
-        if (p1 > p)
-            p = p1;
-        len = p - base_path;
-        if (len > dest_size - 1)
-            len = dest_size - 1;
-        memcpy(dest, base_path, len);
-        dest[len] = '\0';
-        pstrcat(dest, dest_size, filename);
+    if (p1) {
+        p1++;
+    } else {
+        p1 = base_path;
+    }
+    if (p1 > p) {
+        p = p1;
     }
+    len = p - base_path;
+
+    result = g_malloc(len + strlen(filename) + 1);
+    memcpy(result, base_path, len);
+    strcpy(result + len, filename);
+
+    return result;
 }
 
 /*
@@ -303,30 +303,61 @@ fail:
     return -EACCES;
 }
 
-void bdrv_get_full_backing_filename_from_filename(const char *backed,
-                                                  const char *backing,
-                                                  char *dest, size_t sz,
-                                                  Error **errp)
+/*
+ * If @backing is empty, this function returns NULL without setting
+ * @errp.  In all other cases, NULL will only be returned with @errp
+ * set.
+ *
+ * Therefore, a return value of NULL without @errp set means that
+ * there is no backing file; if @errp is set, there is one but its
+ * absolute filename cannot be generated.
+ */
+char *bdrv_get_full_backing_filename_from_filename(const char *backed,
+                                                   const char *backing,
+                                                   Error **errp)
 {
-    if (backing[0] == '\0' || path_has_protocol(backing) ||
-        path_is_absolute(backing))
-    {
-        pstrcpy(dest, sz, backing);
+    if (backing[0] == '\0') {
+        return NULL;
+    } else if (path_has_protocol(backing) || path_is_absolute(backing)) {
+        return g_strdup(backing);
     } else if (backed[0] == '\0' || strstart(backed, "json:", NULL)) {
         error_setg(errp, "Cannot use relative backing file names for '%s'",
                    backed);
+        return NULL;
     } else {
-        path_combine(dest, sz, backed, backing);
+        return path_combine(backed, backing);
     }
 }
 
-void bdrv_get_full_backing_filename(BlockDriverState *bs, char *dest, size_t sz,
-                                    Error **errp)
+/*
+ * If @filename is empty or NULL, this function returns NULL without
+ * setting @errp.  In all other cases, NULL will only be returned with
+ * @errp set.
+ */
+static char *bdrv_make_absolute_filename(BlockDriverState *relative_to,
+                                         const char *filename, Error **errp)
 {
-    char *backed = bs->exact_filename[0] ? bs->exact_filename : bs->filename;
+    char *dir, *full_name;
+
+    if (!filename || filename[0] == '\0') {
+        return NULL;
+    } else if (path_has_protocol(filename) || path_is_absolute(filename)) {
+        return g_strdup(filename);
+    }
+
+    dir = bdrv_dirname(relative_to, errp);
+    if (!dir) {
+        return NULL;
+    }
+
+    full_name = g_strconcat(dir, filename, NULL);
+    g_free(dir);
+    return full_name;
+}
 
-    bdrv_get_full_backing_filename_from_filename(backed, bs->backing_file,
-                                                 dest, sz, errp);
+char *bdrv_get_full_backing_filename(BlockDriverState *bs, Error **errp)
+{
+    return bdrv_make_absolute_filename(bs, bs->backing_file, errp);
 }
 
 void bdrv_register(BlockDriver *bdrv)
@@ -1004,6 +1035,8 @@ static void bdrv_backing_attach(BdrvChild *c)
                "node is used as backing hd of '%s'",
                bdrv_get_device_or_node_name(parent));
 
+    bdrv_refresh_filename(backing_hd);
+
     parent->open_flags &= ~BDRV_O_NO_BACKING;
     pstrcpy(parent->backing_file, sizeof(parent->backing_file),
             backing_hd->filename);
@@ -1413,6 +1446,7 @@ static int bdrv_open_common(BlockDriverState *bs, BlockBackend *file,
     }
 
     if (file != NULL) {
+        bdrv_refresh_filename(blk_bs(file));
         filename = blk_bs(file)->filename;
     } else {
         /*
@@ -1438,13 +1472,19 @@ static int bdrv_open_common(BlockDriverState *bs, BlockBackend *file,
     bs->read_only = !(bs->open_flags & BDRV_O_RDWR);
 
     if (use_bdrv_whitelist && !bdrv_is_whitelisted(drv, bs->read_only)) {
-        error_setg(errp,
-                   !bs->read_only && bdrv_is_whitelisted(drv, true)
-                        ? "Driver '%s' can only be used for read-only devices"
-                        : "Driver '%s' is not whitelisted",
-                   drv->format_name);
-        ret = -ENOTSUP;
-        goto fail_opts;
+        if (!bs->read_only && bdrv_is_whitelisted(drv, true)) {
+            ret = bdrv_apply_auto_read_only(bs, NULL, NULL);
+        } else {
+            ret = -ENOTSUP;
+        }
+        if (ret < 0) {
+            error_setg(errp,
+                       !bs->read_only && bdrv_is_whitelisted(drv, true)
+                       ? "Driver '%s' can only be used for read-only devices"
+                       : "Driver '%s' is not whitelisted",
+                       drv->format_name);
+            goto fail_opts;
+        }
     }
 
     /* bdrv_new() and bdrv_close() make it so */
@@ -1948,13 +1988,32 @@ static int bdrv_child_check_perm(BdrvChild *c, BlockReopenQueue *q,
     ret = bdrv_check_update_perm(c->bs, q, perm, shared, ignore_children, errp);
     g_slist_free(ignore_children);
 
-    return ret;
+    if (ret < 0) {
+        return ret;
+    }
+
+    if (!c->has_backup_perm) {
+        c->has_backup_perm = true;
+        c->backup_perm = c->perm;
+        c->backup_shared_perm = c->shared_perm;
+    }
+    /*
+     * Note: it's OK if c->has_backup_perm was already set, as we can find the
+     * same child twice during check_perm procedure
+     */
+
+    c->perm = perm;
+    c->shared_perm = shared;
+
+    return 0;
 }
 
 static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared)
 {
     uint64_t cumulative_perms, cumulative_shared_perms;
 
+    c->has_backup_perm = false;
+
     c->perm = perm;
     c->shared_perm = shared;
 
@@ -1965,6 +2024,12 @@ static void bdrv_child_set_perm(BdrvChild *c, uint64_t perm, uint64_t shared)
 
 static void bdrv_child_abort_perm_update(BdrvChild *c)
 {
+    if (c->has_backup_perm) {
+        c->perm = c->backup_perm;
+        c->shared_perm = c->backup_shared_perm;
+        c->has_backup_perm = false;
+    }
+
     bdrv_abort_perm_update(c->bs);
 }
 
@@ -2303,8 +2368,6 @@ void bdrv_set_backing_hd(BlockDriverState *bs, BlockDriverState *backing_hd,
         bdrv_unref(backing_hd);
     }
 
-    bdrv_refresh_filename(bs);
-
 out:
     bdrv_refresh_limits(bs, NULL);
 }
@@ -2322,10 +2385,11 @@ out:
 int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options,
                            const char *bdref_key, Error **errp)
 {
-    char *backing_filename = g_malloc0(PATH_MAX);
+    char *backing_filename = NULL;
     char *bdref_key_dot;
     const char *reference = NULL;
     int ret = 0;
+    bool implicit_backing = false;
     BlockDriverState *backing_hd;
     QDict *options;
     QDict *tmp_parent_options = NULL;
@@ -2356,13 +2420,22 @@ int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options,
      */
     reference = qdict_get_try_str(parent_options, bdref_key);
     if (reference || qdict_haskey(options, "file.filename")) {
-        backing_filename[0] = '\0';
+        /* keep backing_filename NULL */
     } else if (bs->backing_file[0] == '\0' && qdict_size(options) == 0) {
         qobject_unref(options);
         goto free_exit;
     } else {
-        bdrv_get_full_backing_filename(bs, backing_filename, PATH_MAX,
-                                       &local_err);
+        if (qdict_size(options) == 0) {
+            /* If the user specifies options that do not modify the
+             * backing file's behavior, we might still consider it the
+             * implicit backing file.  But it's easier this way, and
+             * just specifying some of the backing BDS's options is
+             * only possible with -drive anyway (otherwise the QAPI
+             * schema forces the user to specify everything). */
+            implicit_backing = !strcmp(bs->auto_backing_file, bs->backing_file);
+        }
+
+        backing_filename = bdrv_get_full_backing_filename(bs, &local_err);
         if (local_err) {
             ret = -EINVAL;
             error_propagate(errp, local_err);
@@ -2383,9 +2456,8 @@ int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options,
         qdict_put_str(options, "driver", bs->backing_format);
     }
 
-    backing_hd = bdrv_open_inherit(*backing_filename ? backing_filename : NULL,
-                                   reference, options, 0, bs, &child_backing,
-                                   errp);
+    backing_hd = bdrv_open_inherit(backing_filename, reference, options, 0, bs,
+                                   &child_backing, errp);
     if (!backing_hd) {
         bs->open_flags |= BDRV_O_NO_BACKING;
         error_prepend(errp, "Could not open backing file: ");
@@ -2394,6 +2466,12 @@ int bdrv_open_backing_file(BlockDriverState *bs, QDict *parent_options,
     }
     bdrv_set_aio_context(backing_hd, bdrv_get_aio_context(bs));
 
+    if (implicit_backing) {
+        bdrv_refresh_filename(backing_hd);
+        pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file),
+                backing_hd->filename);
+    }
+
     /* Hook up the backing file link; drop our reference, bs owns the
      * backing_hd reference now */
     bdrv_set_backing_hd(bs, backing_hd, &local_err);
@@ -2833,8 +2911,6 @@ static BlockDriverState *bdrv_open_inherit(const char *filename,
         g_free(child_key_dot);
     }
 
-    bdrv_refresh_filename(bs);
-
     /* Check if any unknown options were used */
     if (qdict_size(options) != 0) {
         const QDictEntry *entry = qdict_first(options);
@@ -3197,6 +3273,7 @@ int bdrv_reopen_prepare(BDRVReopenState *reopen_state, BlockReopenQueue *queue,
                         Error **errp)
 {
     int ret = -1;
+    int old_flags;
     Error *local_err = NULL;
     BlockDriver *drv;
     QemuOpts *opts;
@@ -3223,7 +3300,12 @@ int bdrv_reopen_prepare(BDRVReopenState *reopen_state, BlockReopenQueue *queue,
         goto error;
     }
 
+    /* This was already called in bdrv_reopen_queue_child() so the flags
+     * are up-to-date. This time we simply want to remove the options from
+     * QemuOpts in order to indicate that they have been processed. */
+    old_flags = reopen_state->flags;
     update_flags_from_options(&reopen_state->flags, opts);
+    assert(old_flags == reopen_state->flags);
 
     discard = qemu_opt_get_del(opts, BDRV_OPT_DISCARD);
     if (discard != NULL) {
@@ -3273,6 +3355,7 @@ int bdrv_reopen_prepare(BDRVReopenState *reopen_state, BlockReopenQueue *queue,
             if (local_err != NULL) {
                 error_propagate(errp, local_err);
             } else {
+                bdrv_refresh_filename(reopen_state->bs);
                 error_setg(errp, "failed while preparing to reopen image '%s'",
                            reopen_state->bs->filename);
             }
@@ -3530,7 +3613,9 @@ void bdrv_close_all(void)
 
 static bool should_update_child(BdrvChild *c, BlockDriverState *to)
 {
-    BdrvChild *to_c;
+    GQueue *queue;
+    GHashTable *found;
+    bool ret;
 
     if (c->role->stay_at_node) {
         return false;
@@ -3566,14 +3651,43 @@ static bool should_update_child(BdrvChild *c, BlockDriverState *to)
      * if A is a child of B, that means we cannot replace A by B there
      * because that would create a loop.  Silently detaching A from B
      * is also not really an option.  So overall just leaving A in
-     * place there is the most sensible choice. */
-    QLIST_FOREACH(to_c, &to->children, next) {
-        if (to_c == c) {
-            return false;
+     * place there is the most sensible choice.
+     *
+     * We would also create a loop in any cases where @c is only
+     * indirectly referenced by @to. Prevent this by returning false
+     * if @c is found (by breadth-first search) anywhere in the whole
+     * subtree of @to.
+     */
+
+    ret = true;
+    found = g_hash_table_new(NULL, NULL);
+    g_hash_table_add(found, to);
+    queue = g_queue_new();
+    g_queue_push_tail(queue, to);
+
+    while (!g_queue_is_empty(queue)) {
+        BlockDriverState *v = g_queue_pop_head(queue);
+        BdrvChild *c2;
+
+        QLIST_FOREACH(c2, &v->children, next) {
+            if (c2 == c) {
+                ret = false;
+                break;
+            }
+
+            if (g_hash_table_contains(found, c2->bs)) {
+                continue;
+            }
+
+            g_queue_push_tail(queue, c2->bs);
+            g_hash_table_add(found, c2->bs);
         }
     }
 
-    return true;
+    g_queue_free(queue);
+    g_hash_table_destroy(found);
+
+    return ret;
 }
 
 void bdrv_replace_node(BlockDriverState *from, BlockDriverState *to,
@@ -3719,6 +3833,7 @@ static void bdrv_check_co_entry(void *opaque)
 {
     CheckCo *cco = opaque;
     cco->ret = bdrv_co_check(cco->bs, cco->res, cco->fix);
+    aio_wait_kick();
 }
 
 int bdrv_check(BlockDriverState *bs,
@@ -3737,7 +3852,7 @@ int bdrv_check(BlockDriverState *bs,
         bdrv_check_co_entry(&cco);
     } else {
         co = qemu_coroutine_create(bdrv_check_co_entry, &cco);
-        qemu_coroutine_enter(co);
+        bdrv_coroutine_enter(bs, co);
         BDRV_POLL_WHILE(bs, cco.ret == -EINPROGRESS);
     }
 
@@ -3776,6 +3891,8 @@ int bdrv_change_backing_file(BlockDriverState *bs,
     if (ret == 0) {
         pstrcpy(bs->backing_file, sizeof(bs->backing_file), backing_file ?: "");
         pstrcpy(bs->backing_format, sizeof(bs->backing_format), backing_fmt ?: "");
+        pstrcpy(bs->auto_backing_file, sizeof(bs->auto_backing_file),
+                backing_file ?: "");
     }
     return ret;
 }
@@ -3868,7 +3985,10 @@ int bdrv_drop_intermediate(BlockDriverState *top, BlockDriverState *base,
     /* success - we can delete the intermediate states, and link top->base */
     /* TODO Check graph modification op blockers (BLK_PERM_GRAPH_MOD) once
      * we've figured out how they should work. */
-    backing_file_str = backing_file_str ? backing_file_str : base->filename;
+    if (!backing_file_str) {
+        bdrv_refresh_filename(base);
+        backing_file_str = base->filename;
+    }
 
     QLIST_FOREACH_SAFE(c, &top->parents, next_parent, next) {
         /* Check whether we are allowed to switch c from top to base */
@@ -4113,6 +4233,154 @@ BlockDeviceInfoList *bdrv_named_nodes_list(Error **errp)
     return list;
 }
 
+#define QAPI_LIST_ADD(list, element) do { \
+    typeof(list) _tmp = g_new(typeof(*(list)), 1); \
+    _tmp->value = (element); \
+    _tmp->next = (list); \
+    (list) = _tmp; \
+} while (0)
+
+typedef struct XDbgBlockGraphConstructor {
+    XDbgBlockGraph *graph;
+    GHashTable *graph_nodes;
+} XDbgBlockGraphConstructor;
+
+static XDbgBlockGraphConstructor *xdbg_graph_new(void)
+{
+    XDbgBlockGraphConstructor *gr = g_new(XDbgBlockGraphConstructor, 1);
+
+    gr->graph = g_new0(XDbgBlockGraph, 1);
+    gr->graph_nodes = g_hash_table_new(NULL, NULL);
+
+    return gr;
+}
+
+static XDbgBlockGraph *xdbg_graph_finalize(XDbgBlockGraphConstructor *gr)
+{
+    XDbgBlockGraph *graph = gr->graph;
+
+    g_hash_table_destroy(gr->graph_nodes);
+    g_free(gr);
+
+    return graph;
+}
+
+static uintptr_t xdbg_graph_node_num(XDbgBlockGraphConstructor *gr, void *node)
+{
+    uintptr_t ret = (uintptr_t)g_hash_table_lookup(gr->graph_nodes, node);
+
+    if (ret != 0) {
+        return ret;
+    }
+
+    /*
+     * Start counting from 1, not 0, because 0 interferes with not-found (NULL)
+     * answer of g_hash_table_lookup.
+     */
+    ret = g_hash_table_size(gr->graph_nodes) + 1;
+    g_hash_table_insert(gr->graph_nodes, node, (void *)ret);
+
+    return ret;
+}
+
+static void xdbg_graph_add_node(XDbgBlockGraphConstructor *gr, void *node,
+                                XDbgBlockGraphNodeType type, const char *name)
+{
+    XDbgBlockGraphNode *n;
+
+    n = g_new0(XDbgBlockGraphNode, 1);
+
+    n->id = xdbg_graph_node_num(gr, node);
+    n->type = type;
+    n->name = g_strdup(name);
+
+    QAPI_LIST_ADD(gr->graph->nodes, n);
+}
+
+static void xdbg_graph_add_edge(XDbgBlockGraphConstructor *gr, void *parent,
+                                const BdrvChild *child)
+{
+    typedef struct {
+        unsigned int flag;
+        BlockPermission num;
+    } PermissionMap;
+
+    static const PermissionMap permissions[] = {
+        { BLK_PERM_CONSISTENT_READ, BLOCK_PERMISSION_CONSISTENT_READ },
+        { BLK_PERM_WRITE,           BLOCK_PERMISSION_WRITE },
+        { BLK_PERM_WRITE_UNCHANGED, BLOCK_PERMISSION_WRITE_UNCHANGED },
+        { BLK_PERM_RESIZE,          BLOCK_PERMISSION_RESIZE },
+        { BLK_PERM_GRAPH_MOD,       BLOCK_PERMISSION_GRAPH_MOD },
+        { 0, 0 }
+    };
+    const PermissionMap *p;
+    XDbgBlockGraphEdge *edge;
+
+    QEMU_BUILD_BUG_ON(1UL << (ARRAY_SIZE(permissions) - 1) != BLK_PERM_ALL + 1);
+
+    edge = g_new0(XDbgBlockGraphEdge, 1);
+
+    edge->parent = xdbg_graph_node_num(gr, parent);
+    edge->child = xdbg_graph_node_num(gr, child->bs);
+    edge->name = g_strdup(child->name);
+
+    for (p = permissions; p->flag; p++) {
+        if (p->flag & child->perm) {
+            QAPI_LIST_ADD(edge->perm, p->num);
+        }
+        if (p->flag & child->shared_perm) {
+            QAPI_LIST_ADD(edge->shared_perm, p->num);
+        }
+    }
+
+    QAPI_LIST_ADD(gr->graph->edges, edge);
+}
+
+
+XDbgBlockGraph *bdrv_get_xdbg_block_graph(Error **errp)
+{
+    BlockBackend *blk;
+    BlockJob *job;
+    BlockDriverState *bs;
+    BdrvChild *child;
+    XDbgBlockGraphConstructor *gr = xdbg_graph_new();
+
+    for (blk = blk_all_next(NULL); blk; blk = blk_all_next(blk)) {
+        char *allocated_name = NULL;
+        const char *name = blk_name(blk);
+
+        if (!*name) {
+            name = allocated_name = blk_get_attached_dev_id(blk);
+        }
+        xdbg_graph_add_node(gr, blk, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_BACKEND,
+                           name);
+        g_free(allocated_name);
+        if (blk_root(blk)) {
+            xdbg_graph_add_edge(gr, blk, blk_root(blk));
+        }
+    }
+
+    for (job = block_job_next(NULL); job; job = block_job_next(job)) {
+        GSList *el;
+
+        xdbg_graph_add_node(gr, job, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_JOB,
+                           job->job.id);
+        for (el = job->nodes; el; el = el->next) {
+            xdbg_graph_add_edge(gr, job, (BdrvChild *)el->data);
+        }
+    }
+
+    QTAILQ_FOREACH(bs, &graph_bdrv_states, node_list) {
+        xdbg_graph_add_node(gr, bs, X_DBG_BLOCK_GRAPH_NODE_TYPE_BLOCK_DRIVER,
+                           bs->node_name);
+        QLIST_FOREACH(child, &bs->children, next) {
+            xdbg_graph_add_edge(gr, bs, child);
+        }
+    }
+
+    return xdbg_graph_finalize(gr);
+}
+
 BlockDriverState *bdrv_lookup_bs(const char *device,
                                  const char *node_name,
                                  Error **errp)
@@ -4268,16 +4536,6 @@ bool bdrv_can_write_zeroes_with_unmap(BlockDriverState *bs)
     return bs->supported_zero_flags & BDRV_REQ_MAY_UNMAP;
 }
 
-const char *bdrv_get_encrypted_filename(BlockDriverState *bs)
-{
-    if (bs->backing && bs->backing->bs->encrypted)
-        return bs->backing_file;
-    else if (bs->encrypted)
-        return bs->filename;
-    else
-        return NULL;
-}
-
 void bdrv_get_backing_filename(BlockDriverState *bs,
                                char *filename, int filename_size)
 {
@@ -4301,11 +4559,12 @@ int bdrv_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
     return drv->bdrv_get_info(bs, bdi);
 }
 
-ImageInfoSpecific *bdrv_get_specific_info(BlockDriverState *bs)
+ImageInfoSpecific *bdrv_get_specific_info(BlockDriverState *bs,
+                                          Error **errp)
 {
     BlockDriver *drv = bs->drv;
     if (drv && drv->bdrv_get_specific_info) {
-        return drv->bdrv_get_specific_info(bs);
+        return drv->bdrv_get_specific_info(bs, errp);
     }
     return NULL;
 }
@@ -4385,7 +4644,6 @@ BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs,
     int is_protocol = 0;
     BlockDriverState *curr_bs = NULL;
     BlockDriverState *retval = NULL;
-    Error *local_error = NULL;
 
     if (!bs || !bs->drv || !backing_file) {
         return NULL;
@@ -4393,7 +4651,6 @@ BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs,
 
     filename_full     = g_malloc(PATH_MAX);
     backing_file_full = g_malloc(PATH_MAX);
-    filename_tmp      = g_malloc(PATH_MAX);
 
     is_protocol = path_has_protocol(backing_file);
 
@@ -4402,41 +4659,43 @@ BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs,
         /* If either of the filename paths is actually a protocol, then
          * compare unmodified paths; otherwise make paths relative */
         if (is_protocol || path_has_protocol(curr_bs->backing_file)) {
+            char *backing_file_full_ret;
+
             if (strcmp(backing_file, curr_bs->backing_file) == 0) {
                 retval = curr_bs->backing->bs;
                 break;
             }
             /* Also check against the full backing filename for the image */
-            bdrv_get_full_backing_filename(curr_bs, backing_file_full, PATH_MAX,
-                                           &local_error);
-            if (local_error == NULL) {
-                if (strcmp(backing_file, backing_file_full) == 0) {
+            backing_file_full_ret = bdrv_get_full_backing_filename(curr_bs,
+                                                                   NULL);
+            if (backing_file_full_ret) {
+                bool equal = strcmp(backing_file, backing_file_full_ret) == 0;
+                g_free(backing_file_full_ret);
+                if (equal) {
                     retval = curr_bs->backing->bs;
                     break;
                 }
-            } else {
-                error_free(local_error);
-                local_error = NULL;
             }
         } else {
             /* If not an absolute filename path, make it relative to the current
              * image's filename path */
-            path_combine(filename_tmp, PATH_MAX, curr_bs->filename,
-                         backing_file);
-
-            /* We are going to compare absolute pathnames */
-            if (!realpath(filename_tmp, filename_full)) {
+            filename_tmp = bdrv_make_absolute_filename(curr_bs, backing_file,
+                                                       NULL);
+            /* We are going to compare canonicalized absolute pathnames */
+            if (!filename_tmp || !realpath(filename_tmp, filename_full)) {
+                g_free(filename_tmp);
                 continue;
             }
+            g_free(filename_tmp);
 
             /* We need to make sure the backing filename we are comparing against
              * is relative to the current image filename (or absolute) */
-            path_combine(filename_tmp, PATH_MAX, curr_bs->filename,
-                         curr_bs->backing_file);
-
-            if (!realpath(filename_tmp, backing_file_full)) {
+            filename_tmp = bdrv_get_full_backing_filename(curr_bs, NULL);
+            if (!filename_tmp || !realpath(filename_tmp, backing_file_full)) {
+                g_free(filename_tmp);
                 continue;
             }
+            g_free(filename_tmp);
 
             if (strcmp(backing_file_full, filename_full) == 0) {
                 retval = curr_bs->backing->bs;
@@ -4447,7 +4706,6 @@ BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs,
 
     g_free(filename_full);
     g_free(backing_file_full);
-    g_free(filename_tmp);
     return retval;
 }
 
@@ -4536,6 +4794,7 @@ static void coroutine_fn bdrv_co_invalidate_cache(BlockDriverState *bs,
         if (parent->role->activate) {
             parent->role->activate(parent, &local_err);
             if (local_err) {
+                bs->open_flags |= BDRV_O_INACTIVE;
                 error_propagate(errp, local_err);
                 return;
             }
@@ -4554,6 +4813,7 @@ static void coroutine_fn bdrv_invalidate_cache_co_entry(void *opaque)
     InvalidateCacheCo *ico = opaque;
     bdrv_co_invalidate_cache(ico->bs, ico->errp);
     ico->done = true;
+    aio_wait_kick();
 }
 
 void bdrv_invalidate_cache(BlockDriverState *bs, Error **errp)
@@ -4570,7 +4830,7 @@ void bdrv_invalidate_cache(BlockDriverState *bs, Error **errp)
         bdrv_invalidate_cache_co_entry(&ico);
     } else {
         co = qemu_coroutine_create(bdrv_invalidate_cache_co_entry, &ico);
-        qemu_coroutine_enter(co);
+        bdrv_coroutine_enter(bs, co);
         BDRV_POLL_WHILE(bs, !ico.done);
     }
 }
@@ -4932,17 +5192,17 @@ void bdrv_img_create(const char *filename, const char *fmt,
     size = qemu_opt_get_size(opts, BLOCK_OPT_SIZE, img_size);
     if (backing_file && !(flags & BDRV_O_NO_BACKING)) {
         BlockDriverState *bs;
-        char *full_backing = g_new0(char, PATH_MAX);
+        char *full_backing;
         int back_flags;
         QDict *backing_options = NULL;
 
-        bdrv_get_full_backing_filename_from_filename(filename, backing_file,
-                                                     full_backing, PATH_MAX,
-                                                     &local_err);
+        full_backing =
+            bdrv_get_full_backing_filename_from_filename(filename, backing_file,
+                                                         &local_err);
         if (local_err) {
-            g_free(full_backing);
             goto out;
         }
+        assert(full_backing);
 
         /* backing files always opened read-only */
         back_flags = flags;
@@ -5063,6 +5323,9 @@ void bdrv_detach_aio_context(BlockDriverState *bs)
         bdrv_detach_aio_context(child->bs);
     }
 
+    if (bs->quiesce_counter) {
+        aio_enable_external(bs->aio_context);
+    }
     bs->aio_context = NULL;
 }
 
@@ -5076,6 +5339,10 @@ void bdrv_attach_aio_context(BlockDriverState *bs,
         return;
     }
 
+    if (bs->quiesce_counter) {
+        aio_disable_external(new_context);
+    }
+
     bs->aio_context = new_context;
 
     QLIST_FOREACH(child, &bs->children, next) {
@@ -5097,18 +5364,16 @@ 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)
 {
-    AioContext *ctx = bdrv_get_aio_context(bs);
-
-    aio_disable_external(ctx);
-    bdrv_parent_drained_begin(bs, NULL, false);
-    bdrv_drain(bs); /* ensure there are no in-flight requests */
-
-    while (aio_poll(ctx, false)) {
-        /* wait for all bottom halves to execute */
+    if (bdrv_get_aio_context(bs) == new_context) {
+        return;
     }
 
+    bdrv_drained_begin(bs);
     bdrv_detach_aio_context(bs);
 
     /* This function executes in the old AioContext so acquire the new one in
@@ -5116,8 +5381,7 @@ void bdrv_set_aio_context(BlockDriverState *bs, AioContext *new_context)
      */
     aio_context_acquire(new_context);
     bdrv_attach_aio_context(bs, new_context);
-    bdrv_parent_drained_end(bs, NULL, false);
-    aio_enable_external(ctx);
+    bdrv_drained_end(bs);
     aio_context_release(new_context);
 }
 
@@ -5271,33 +5535,113 @@ out:
     return to_replace_bs;
 }
 
-static bool append_open_options(QDict *d, BlockDriverState *bs)
+/**
+ * Iterates through the list of runtime option keys that are said to
+ * be "strong" for a BDS.  An option is called "strong" if it changes
+ * a BDS's data.  For example, the null block driver's "size" and
+ * "read-zeroes" options are strong, but its "latency-ns" option is
+ * not.
+ *
+ * If a key returned by this function ends with a dot, all options
+ * starting with that prefix are strong.
+ */
+static const char *const *strong_options(BlockDriverState *bs,
+                                         const char *const *curopt)
+{
+    static const char *const global_options[] = {
+        "driver", "filename", NULL
+    };
+
+    if (!curopt) {
+        return &global_options[0];
+    }
+
+    curopt++;
+    if (curopt == &global_options[ARRAY_SIZE(global_options) - 1] && bs->drv) {
+        curopt = bs->drv->strong_runtime_opts;
+    }
+
+    return (curopt && *curopt) ? curopt : NULL;
+}
+
+/**
+ * Copies all strong runtime options from bs->options to the given
+ * QDict.  The set of strong option keys is determined by invoking
+ * strong_options().
+ *
+ * Returns true iff any strong option was present in bs->options (and
+ * thus copied to the target QDict) with the exception of "filename"
+ * and "driver".  The caller is expected to use this value to decide
+ * whether the existence of strong options prevents the generation of
+ * a plain filename.
+ */
+static bool append_strong_runtime_options(QDict *d, BlockDriverState *bs)
 {
-    const QDictEntry *entry;
-    QemuOptDesc *desc;
     bool found_any = false;
+    const char *const *option_name = NULL;
 
-    for (entry = qdict_first(bs->options); entry;
-         entry = qdict_next(bs->options, entry))
-    {
-        /* Exclude all non-driver-specific options */
-        for (desc = bdrv_runtime_opts.desc; desc->name; desc++) {
-            if (!strcmp(qdict_entry_key(entry), desc->name)) {
-                break;
+    if (!bs->drv) {
+        return false;
+    }
+
+    while ((option_name = strong_options(bs, option_name))) {
+        bool option_given = false;
+
+        assert(strlen(*option_name) > 0);
+        if ((*option_name)[strlen(*option_name) - 1] != '.') {
+            QObject *entry = qdict_get(bs->options, *option_name);
+            if (!entry) {
+                continue;
+            }
+
+            qdict_put_obj(d, *option_name, qobject_ref(entry));
+            option_given = true;
+        } else {
+            const QDictEntry *entry;
+            for (entry = qdict_first(bs->options); entry;
+                 entry = qdict_next(bs->options, entry))
+            {
+                if (strstart(qdict_entry_key(entry), *option_name, NULL)) {
+                    qdict_put_obj(d, qdict_entry_key(entry),
+                                  qobject_ref(qdict_entry_value(entry)));
+                    option_given = true;
+                }
             }
         }
-        if (desc->name) {
-            continue;
+
+        /* While "driver" and "filename" need to be included in a JSON filename,
+         * their existence does not prohibit generation of a plain filename. */
+        if (!found_any && option_given &&
+            strcmp(*option_name, "driver") && strcmp(*option_name, "filename"))
+        {
+            found_any = true;
         }
+    }
 
-        qdict_put_obj(d, qdict_entry_key(entry),
-                      qobject_ref(qdict_entry_value(entry)));
-        found_any = true;
+    if (!qdict_haskey(d, "driver")) {
+        /* Drivers created with bdrv_new_open_driver() may not have a
+         * @driver option.  Add it here. */
+        qdict_put_str(d, "driver", bs->drv->format_name);
     }
 
     return found_any;
 }
 
+/* Note: This function may return false positives; it may return true
+ * even if opening the backing file specified by bs's image header
+ * would result in exactly bs->backing. */
+static bool bdrv_backing_overridden(BlockDriverState *bs)
+{
+    if (bs->backing) {
+        return strcmp(bs->auto_backing_file,
+                      bs->backing->bs->filename);
+    } else {
+        /* No backing BDS, so if the image header reports any backing
+         * file, it must have been suppressed */
+        return bs->auto_backing_file[0] != '\0';
+    }
+}
+
 /* Updates the following BDS fields:
  *  - exact_filename: A filename which may be used for opening a block device
  *                    which (mostly) equals the given BDS (even without any
@@ -5313,92 +5657,108 @@ static bool append_open_options(QDict *d, BlockDriverState *bs)
 void bdrv_refresh_filename(BlockDriverState *bs)
 {
     BlockDriver *drv = bs->drv;
+    BdrvChild *child;
     QDict *opts;
+    bool backing_overridden;
+    bool generate_json_filename; /* Whether our default implementation should
+                                    fill exact_filename (false) or not (true) */
 
     if (!drv) {
         return;
     }
 
-    /* This BDS's file name will most probably depend on its file's name, so
-     * refresh that first */
-    if (bs->file) {
-        bdrv_refresh_filename(bs->file->bs);
+    /* This BDS's file name may depend on any of its children's file names, so
+     * refresh those first */
+    QLIST_FOREACH(child, &bs->children, next) {
+        bdrv_refresh_filename(child->bs);
+    }
+
+    if (bs->implicit) {
+        /* For implicit nodes, just copy everything from the single child */
+        child = QLIST_FIRST(&bs->children);
+        assert(QLIST_NEXT(child, next) == NULL);
+
+        pstrcpy(bs->exact_filename, sizeof(bs->exact_filename),
+                child->bs->exact_filename);
+        pstrcpy(bs->filename, sizeof(bs->filename), child->bs->filename);
+
+        bs->full_open_options = qobject_ref(child->bs->full_open_options);
+
+        return;
+    }
+
+    backing_overridden = bdrv_backing_overridden(bs);
+
+    if (bs->open_flags & BDRV_O_NO_IO) {
+        /* Without I/O, the backing file does not change anything.
+         * Therefore, in such a case (primarily qemu-img), we can
+         * pretend the backing file has not been overridden even if
+         * it technically has been. */
+        backing_overridden = false;
+    }
+
+    /* Gather the options QDict */
+    opts = qdict_new();
+    generate_json_filename = append_strong_runtime_options(opts, bs);
+    generate_json_filename |= backing_overridden;
+
+    if (drv->bdrv_gather_child_options) {
+        /* Some block drivers may not want to present all of their children's
+         * options, or name them differently from BdrvChild.name */
+        drv->bdrv_gather_child_options(bs, opts, backing_overridden);
+    } else {
+        QLIST_FOREACH(child, &bs->children, next) {
+            if (child->role == &child_backing && !backing_overridden) {
+                /* We can skip the backing BDS if it has not been overridden */
+                continue;
+            }
+
+            qdict_put(opts, child->name,
+                      qobject_ref(child->bs->full_open_options));
+        }
+
+        if (backing_overridden && !bs->backing) {
+            /* Force no backing file */
+            qdict_put_null(opts, "backing");
+        }
     }
 
+    qobject_unref(bs->full_open_options);
+    bs->full_open_options = opts;
+
     if (drv->bdrv_refresh_filename) {
         /* Obsolete information is of no use here, so drop the old file name
          * information before refreshing it */
         bs->exact_filename[0] = '\0';
-        if (bs->full_open_options) {
-            qobject_unref(bs->full_open_options);
-            bs->full_open_options = NULL;
-        }
 
-        opts = qdict_new();
-        append_open_options(opts, bs);
-        drv->bdrv_refresh_filename(bs, opts);
-        qobject_unref(opts);
+        drv->bdrv_refresh_filename(bs);
     } else if (bs->file) {
         /* Try to reconstruct valid information from the underlying file */
-        bool has_open_options;
 
         bs->exact_filename[0] = '\0';
-        if (bs->full_open_options) {
-            qobject_unref(bs->full_open_options);
-            bs->full_open_options = NULL;
-        }
 
-        opts = qdict_new();
-        has_open_options = append_open_options(opts, bs);
-
-        /* If no specific options have been given for this BDS, the filename of
-         * the underlying file should suffice for this one as well */
-        if (bs->file->bs->exact_filename[0] && !has_open_options) {
+        /*
+         * We can use the underlying file's filename if:
+         * - it has a filename,
+         * - the file is a protocol BDS, and
+         * - opening that file (as this BDS's format) will automatically create
+         *   the BDS tree we have right now, that is:
+         *   - the user did not significantly change this BDS's behavior with
+         *     some explicit (strong) options
+         *   - no non-file child of this BDS has been overridden by the user
+         *   Both of these conditions are represented by generate_json_filename.
+         */
+        if (bs->file->bs->exact_filename[0] &&
+            bs->file->bs->drv->bdrv_file_open &&
+            !generate_json_filename)
+        {
             strcpy(bs->exact_filename, bs->file->bs->exact_filename);
         }
-        /* Reconstructing the full options QDict is simple for most format block
-         * drivers, as long as the full options are known for the underlying
-         * file BDS. The full options QDict of that file BDS should somehow
-         * contain a representation of the filename, therefore the following
-         * suffices without querying the (exact_)filename of this BDS. */
-        if (bs->file->bs->full_open_options) {
-            qdict_put_str(opts, "driver", drv->format_name);
-            qdict_put(opts, "file",
-                      qobject_ref(bs->file->bs->full_open_options));
-
-            bs->full_open_options = opts;
-        } else {
-            qobject_unref(opts);
-        }
-    } else if (!bs->full_open_options && qdict_size(bs->options)) {
-        /* There is no underlying file BDS (at least referenced by BDS.file),
-         * so the full options QDict should be equal to the options given
-         * specifically for this block device when it was opened (plus the
-         * driver specification).
-         * Because those options don't change, there is no need to update
-         * full_open_options when it's already set. */
-
-        opts = qdict_new();
-        append_open_options(opts, bs);
-        qdict_put_str(opts, "driver", drv->format_name);
-
-        if (bs->exact_filename[0]) {
-            /* This may not work for all block protocol drivers (some may
-             * require this filename to be parsed), but we have to find some
-             * default solution here, so just include it. If some block driver
-             * does not support pure options without any filename at all or
-             * needs some special format of the options QDict, it needs to
-             * implement the driver-specific bdrv_refresh_filename() function.
-             */
-            qdict_put_str(opts, "filename", bs->exact_filename);
-        }
-
-        bs->full_open_options = opts;
     }
 
     if (bs->exact_filename[0]) {
         pstrcpy(bs->filename, sizeof(bs->filename), bs->exact_filename);
-    } else if (bs->full_open_options) {
+    } else {
         QString *json = qobject_to_json(QOBJECT(bs->full_open_options));
         snprintf(bs->filename, sizeof(bs->filename), "json:%s",
                  qstring_get_str(json));
@@ -5406,6 +5766,33 @@ void bdrv_refresh_filename(BlockDriverState *bs)
     }
 }
 
+char *bdrv_dirname(BlockDriverState *bs, Error **errp)
+{
+    BlockDriver *drv = bs->drv;
+
+    if (!drv) {
+        error_setg(errp, "Node '%s' is ejected", bs->node_name);
+        return NULL;
+    }
+
+    if (drv->bdrv_dirname) {
+        return drv->bdrv_dirname(bs, errp);
+    }
+
+    if (bs->file) {
+        return bdrv_dirname(bs->file->bs, errp);
+    }
+
+    bdrv_refresh_filename(bs);
+    if (bs->exact_filename[0] != '\0') {
+        return path_combine(bs->exact_filename, "");
+    }
+
+    error_setg(errp, "Cannot generate a base directory for %s nodes",
+               drv->format_name);
+    return NULL;
+}
+
 /*
  * Hot add/remove a BDS's child. So the user can take a child offline when
  * it is broken and take a new child online