Re: [RFC][PATCH] vfs: In mntput run deactivate_super on a shallow stack.

[ebiederm@xmission.com (Eric W. Biederman)]

Al Viro <viro@ZenIV.linux.org.uk> writes:

> On Wed, Apr 09, 2014 at 03:58:25PM -0700, Eric W. Biederman wrote:
>> 
>> mntput as part of pathput is called from all over the vfs sometimes as
>> in the case of symlink chasing from some rather deep call chains.
>> During filesystem unmount with the right set of races those innocuous
>> little mntput calls that take very little stack space can become calls
>> become mosters calling deactivate_super that can take up 3k or more of
>> stack space as syncrhonous filesystem I/O is performed, through
>> multiple levels of the I/O stack.
>> 
>> Avoid deactivate_super being called from a deep stack by converting
>> mntput to use task_work_add when the mnt_count goes to 0.  The
>> filesystem is still unmounted synchronously preserving the semantics
>> that system calls like umount require.
>
> Careful.  For one thing, you've just introduced a massive leak in knfsd
> and any other kernel thread that might do mntput().  task_work_add()
> makes no sense there - there is no userland to return to.  For another,
> in things like cleanup of failing modprobe we might end up delaying fs
> shutdown too much.  So it's not that simple, unfortunately.

Unfortunately.

> I agree that fs shutdown is better dealt with on mostly empty stack, of
> course - moreover, done right that has a potential to make mntput()
> safe in atomic contexts (there's also acct_auto_close_mnt() to deal
> with; that might take some work to get right, but I think it's not
> fatal).

I am slowly digging into this.

With this patch I was was able to do an A/B comparison of what the stack
cost on my unmounting my minimal ext4 filesystem from d_invalidate
called with in a context with maximum symlink recursion depth, without
and without a changed mntput.

I used sysfs instead of nfs to mount my minimal ext4 filesystem on as I
was a lazy bum and didn't have a nfs server setup handy.

With just my detach_mounts branched merged into 3.15-rc0
I saw 4880 stack bytes left before calling detach_mounts from d_invalidate
I saw 3904 stack bytes left after calling detach_mounts from d_invalidate

Which means in practice unmounting my mininal ext4 filesystem image only
used 976 additional bytes of stack.

With the same kernel plus my change to mntput
I saw 4880 stack bytes left before calling detach_mounts from d_invalidate
I saw 4880 stack bytes left after calling detach_mounts from d_invalidate

Which at least confirms that a change to mntput is enough to make deep
stacks safe.

With 3904 bytes of headroom from ext4 I may have to measure some of the
nastier cases just to be certain there actually is a problem here.

Eric