[lkml]   [2008]   [Nov]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH] fix pktcdvd breakage from commit e105b8bfc769b0545b6f0f395179d1e43cbee822
On Sun, Nov 30, 2008 at 02:52:21PM +0000, Al Viro wrote:
> On Sun, Nov 30, 2008 at 03:28:15PM +0100, Kay Sievers wrote:
> > > So we need to preserve the layout, with the easiest way probably being "add
> > > one more ktype and use kobject_init_and_add() instead of that device_create()".
> > > Sigh...
> >
> > What do you mean? We just need to replace the bogus "pd->pkt_dev" with
> > MKDEV(0, 0) and we are fine.
> Userland-visible change - right now cat /sys/class/pktcdvd/pktcdvd3/dev will
> give you dev_t of the block device in question.

Looking at the locking scheme there, it appears that we have an unpleasant
race of the same kind as discussed in md-with-eviction thread. If removal
hits between finding (and grabbing) gendisk and actual call of ->open(),
it will succeed just fine and we might get an open for _different_ object,
while holding a reference to disk that had already gone through del_gendisk().
Call ioctl() on that sucker and you've got
struct pktcdvd_device *pd = bdev->bd_disk->private_data;
which will point to already freed object.

It really appears that we need to
* add ->use() and ->unuse() callbacks
* have exact_lock() call ->use() in addition to get_disk() - either
on each hit, or on the "it's currently not in use" ones.
* have ->unuse() called on failure exits in __blkdev_get() and in
blkdev_put(), with rules matching those for calls of ->use().

That would allow to close that kind of races, both here and in md. I'd
probably prefer the second variant for calling rules - we would be able
to bracket the "it's associated with underlying objects" intervals by
those calls... Hell knows; I really need to get some sleep and look through
the ->open() and ->release() instances for block devices. Later...

 \ /
  Last update: 2008-11-30 16:57    [W:0.230 / U:0.124 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site