Re: Union mount and lockdep design issues
From: Michal Suchanek
Date: Tue Jul 12 2011 - 14:49:38 EST
On 12 July 2011 13:45, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
> Michal Suchanek <hramrach@xxxxxxxxxx> writes:
>
>> On 12 July 2011 10:30, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>>> Michal Suchanek <hramrach@xxxxxxxxxx> writes:
>>>
>>>> The locking order is likely determined by the structure of the union
>>>> and not some system-wide order of filesystems so assuming the readonly
>>>> layers are locked as well you will probably get a deadlock with
>>>> technically correct mount:
>>>>
>>>> mount -t overlayfs overlayfs -olowerdir=/lower2,upperdir=/upper /tmpoverlay
>>>> mount -t overlayfs overlayfs -olowerdir=/lower1,upperdir=/tmpoverlay /overlay
>>>>
>>>> mount -t overlayfs overlayfs -olowerdir=/lower1,upperdir=/upper2 /tmpoverlay2
>>>> mount -t overlayfs overlayfs -olowerdir=/lower2,upperdir=/tmpoverlay2 /overlay2
>>>>
>>>> because now lower1 and lower2 are differently ordered in the two
>>>> overlays.
>>>
>>> Overlayfs never locks both upper and lower at the same time, which means
>>> there's no AB-BA locking dependency. ÂThe lock orderings are:
>>>
>>> -> /overlay
>>> Â-> /lower1
>>> Â-> /tmpoverlay
>>> Â Â-> /lower2
>>> Â Â-> /upper
>>> -> /overlay2
>>> Â-> /lower2
>>> Â-> /tmpoverlay2
>>> Â Â-> /lower1
>>> Â Â-> /upper2
>>>
>>> As you can see there's no nesting of lower2 and lower1 into each other.
>>>
>>> When you combine two filesystems, a completely new ordering is created
>>> each time, there's no possibility to make an AB-BA nesting. ÂAt least I
>>> cannot see one.
>>
>> Except you can get in situation where overlay locks lower1 and
>> tmpoverlay waits for lower2
>
> Note: tmpoverlay lock does *not* nest into lower1 lock, they are both on
> the same nesting level. ÂThere's no dependency between the two.
>
>> Âwhich is held by overlay2 waiting for
>> lower1 in tmpoverlay2.
>
> No deadlock there.
>
That's nice.
You can still do
mount --bind /lower1 /lower2/lower1
mount --bind /lower2 /lower1/lower2
Which is technically not against usage guidelines, unlike mount --bind
/upper /lower1/upper
If crossing mount boundaries is forbidden try with symlinks or hardlinks.
Thanks
Michal
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/