[PATCH] mtd: fix double free and WARN_ON in add_mtd_device() error paths

[Xue Lei]

When device_register() or mtd_nvmem_add() fails inside
add_mtd_device() for a partition, the error handling triggers
mtd_release() via put_device() or device_unregister(). mtd_release()
calls release_mtd_partition() which frees the mtd_info structure.
However, callers such as mtd_add_partition() and add_mtd_partitions()
also call free_partition() in their error paths, resulting in a double
free.

Additionally, release_mtd_partition() hits WARN_ON(!list_empty(
&mtd->part.node)) because the partition node is still linked in the
parent's partitions list when the release callback fires from the
add_mtd_device() error path.

Fix this by overriding dev->type and dev->release before put_device()
in the error paths, so that device_release() invokes a no-op function
instead of mtd_release(). For the mtd_nvmem_add() failure case,
device_unregister() is replaced with device_del() to separate the
device removal from the final kobject reference drop, allowing the
override to take effect before put_device() is called.

The callers' error paths (list_del + free_partition) remain the sole
owners of mtd_info lifetime on add_mtd_device() failure, which is the
expected contract.

The normal partition teardown path is not affected: del_mtd_device()
goes through kref_put() -> mtd_device_release() -> device_unregister()
with dev->type still set to &mtd_devtype, so mtd_release() ->
release_mtd_partition() continues to work correctly for the regular
removal case.

Reported-by: syzbot+e9c76b56dc05023b8117 at syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=e9c76b56dc05023b8117
Fixes: 19bfa9ebebb5 ("mtd: use refcount to prevent corruption")
Signed-off-by: Xue Lei <Xue.Lei at windriver.com>
---
 drivers/mtd/mtdcore.c | 23 +++++++++++++++++++----
 1 file changed, 19 insertions(+), 4 deletions(-)

diff --git a/drivers/mtd/mtdcore.c b/drivers/mtd/mtdcore.c
index 576537774628..16629382a787 100644
--- a/drivers/mtd/mtdcore.c
+++ b/drivers/mtd/mtdcore.c
@@ -105,6 +105,15 @@ static void mtd_release(struct device *dev)
 	device_destroy(&mtd_class, index + 1);
 }
 
+/*
+ * No-op device release used in add_mtd_device() error paths.
+ * Prevents mtd_release() from being called via device_release(),
+ * which would free the mtd_info that the caller still manages.
+ */
+static void mtd_dev_release_nop(struct device *dev)
+{
+}
+
 static void mtd_device_release(struct kref *kref)
 {
 	struct mtd_info *mtd = container_of(kref, struct mtd_info, refcnt);
@@ -799,10 +808,8 @@ int add_mtd_device(struct mtd_info *mtd)
 	mtd_check_of_node(mtd);
 	of_node_get(mtd_get_of_node(mtd));
 	error = device_register(&mtd->dev);
-	if (error) {
-		put_device(&mtd->dev);
+	if (error)
 		goto fail_added;
-	}
 
 	/* Add the nvmem provider */
 	error = mtd_nvmem_add(mtd);
@@ -840,8 +847,16 @@ int add_mtd_device(struct mtd_info *mtd)
 	return 0;
 
 fail_nvmem_add:
-	device_unregister(&mtd->dev);
+	device_del(&mtd->dev);
 fail_added:
+	/*
+	 * Clear type and set nop release to prevent mtd_release() ->
+	 * release_mtd_partition() -> free_partition() from freeing mtd.
+	 * The caller handles cleanup on failure.
+	 */
+	mtd->dev.type = NULL;
+	mtd->dev.release = mtd_dev_release_nop;
+	put_device(&mtd->dev);
 	of_node_put(mtd_get_of_node(mtd));
 fail_devname:
 	idr_remove(&mtd_idr, i);
-- 
2.43.0