block: Fix oops scsi_disk_get()
authorJan Kara <jack@suse.cz>
Thu, 23 Mar 2017 00:37:02 +0000 (01:37 +0100)
committerJens Axboe <axboe@fb.com>
Thu, 23 Mar 2017 02:11:37 +0000 (20:11 -0600)
commitd01b2dcb441b14c3b860d3d16ecb9ed6ed4ba7fb
treeb667a449649255a01e52a4dbb8219e7a53af7b14
parentc70c176ff8c3ff0ac6ef9a831cd591ea9a66bd1a
block: Fix oops scsi_disk_get()

When device open races with device shutdown, we can get the following
oops in scsi_disk_get():

[11863.044351] general protection fault: 0000 [#1] SMP
[11863.045561] Modules linked in: scsi_debug xfs libcrc32c netconsole btrfs raid6_pq zlib_deflate lzo_compress xor [last unloaded: loop]
[11863.047853] CPU: 3 PID: 13042 Comm: hald-probe-stor Tainted: G W      4.10.0-rc2-xen+ #35
[11863.048030] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
[11863.048030] task: ffff88007f438200 task.stack: ffffc90000fd0000
[11863.048030] RIP: 0010:scsi_disk_get+0x43/0x70
[11863.048030] RSP: 0018:ffffc90000fd3a08 EFLAGS: 00010202
[11863.048030] RAX: 6b6b6b6b6b6b6b6b RBX: ffff88007f56d000 RCX: 0000000000000000
[11863.048030] RDX: 0000000000000001 RSI: 0000000000000004 RDI: ffffffff81a8d880
[11863.048030] RBP: ffffc90000fd3a18 R08: 0000000000000000 R09: 0000000000000001
[11863.059217] R10: 0000000000000000 R11: 0000000000000000 R12: 00000000fffffffa
[11863.059217] R13: ffff880078872800 R14: ffff880070915540 R15: 000000000000001d
[11863.059217] FS:  00007f2611f71800(0000) GS:ffff88007f0c0000(0000) knlGS:0000000000000000
[11863.059217] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[11863.059217] CR2: 000000000060e048 CR3: 00000000778d4000 CR4: 00000000000006e0
[11863.059217] Call Trace:
[11863.059217]  ? disk_get_part+0x22/0x1f0
[11863.059217]  sd_open+0x39/0x130
[11863.059217]  __blkdev_get+0x69/0x430
[11863.059217]  ? bd_acquire+0x7f/0xc0
[11863.059217]  ? bd_acquire+0x96/0xc0
[11863.059217]  ? blkdev_get+0x350/0x350
[11863.059217]  blkdev_get+0x126/0x350
[11863.059217]  ? _raw_spin_unlock+0x2b/0x40
[11863.059217]  ? bd_acquire+0x7f/0xc0
[11863.059217]  ? blkdev_get+0x350/0x350
[11863.059217]  blkdev_open+0x65/0x80
...

As you can see RAX value is already poisoned showing that gendisk we got
is already freed. The problem is that get_gendisk() looks up device
number in ext_devt_idr and then does get_disk() which does kobject_get()
on the disks kobject. However the disk gets removed from ext_devt_idr
only in disk_release() (through blk_free_devt()) at which moment it has
already 0 refcount and is already on its way to be freed. Indeed we've
got a warning from kobject_get() about 0 refcount shortly before the
oops.

We fix the problem by using kobject_get_unless_zero() in get_disk() so
that get_disk() cannot get reference on a disk that is already being
freed.

Tested-by: Lekshmi Pillai <lekshmicpillai@in.ibm.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/genhd.c