md: Flush workqueue md_rdev_misc_wq in md_alloc()
authorDavid Sloan <david.sloan@eideticom.com>
Thu, 11 Aug 2022 17:14:13 +0000 (11:14 -0600)
committerSong Liu <song@kernel.org>
Wed, 24 Aug 2022 17:26:35 +0000 (10:26 -0700)
A race condition still exists when removing and re-creating md devices
in test cases. However, it is only seen on some setups.

The race condition was tracked down to a reference still being held
to the kobject by the rdev in the md_rdev_misc_wq which will be released
in rdev_delayed_delete().

md_alloc() waits for previous deletions by waiting on the md_misc_wq,
but the md_rdev_misc_wq may still be holding a reference to a recently
removed device.

To fix this, also flush the md_rdev_misc_wq in md_alloc().

Signed-off-by: David Sloan <david.sloan@eideticom.com>
[logang@deltatee.com: rewrote commit message]
Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
Signed-off-by: Song Liu <song@kernel.org>
drivers/md/md.c

index afaf36b2f6ab8b32d53edc24e12ba17fa7a16884..71d221601bf8032182d65deadd47c8a4da208839 100644 (file)
@@ -5620,6 +5620,7 @@ struct mddev *md_alloc(dev_t dev, char *name)
         * removed (mddev_delayed_delete).
         */
        flush_workqueue(md_misc_wq);
+       flush_workqueue(md_rdev_misc_wq);
 
        mutex_lock(&disks_mutex);
        mddev = mddev_alloc(dev);