Re: [PATCH] Busy inodes after unmount, be more verbose in generic_shutdown_super

[Balbir Singh <balbir@in.ibm.com>]

On Tue, Mar 07, 2006 at 09:16:22AM +0300, Kirill Korotaev wrote:
> >>The code changes look big, have you looked at
> >>http://marc.theaimsgroup.com/?l=linux-kernel&m=113817279225962&w=2
> >
> >
> >No I haven't.  I like it.
> > - Holding the semaphore shouldn't be a problem.
> > - calling down_read_trylock ought to be fast
> > - I *think* the unwanted calls to prune_dcache are always under
> >   PF_MEMALLOC - they certainly seem to be.
> No, it looks as it is not :(
> Have you noticed my comment about "count" argument to prune_dcache()?
> For example, prune_dcache() is called from shrink_dcache_parent() which 
> is called in many places and not all of them have PF_MEMALLOC or 
> s_umount semaphore for write. But prune_dcache() doesn't care for super 
> blocks etc. It simply shrinks N dentries which are found _first_.
> 
> So the condition:
> +		if ((current->flags & PF_MEMALLOC) &&
> +			!(ret = down_read_trylock(&s->s_umount))) {
> is not always true when the race occurs, as PF_MEMALLOC is not always set.

I understand your comment about shrink_dcache_parent() being called
from several places. prune_one_dentry() would eventually dput the parent,
but unmount would go ahead and unmount the filesystem before the
dput of the parent could happen.

Given that background, I thought our main concern was with respect to
unmount. The race was between shrink_dcache_parent() (called from unmount)
and shrink_dcache_memory() (called from the allocator), hence the fix
for the race condition.

I just noticied that 2.6.16-rc* now seems to have drop_slab() where
PF_MEMALLOC is not set. So, we can still race with my fix if there
if /proc/sys/vm/drop_caches is written to and unmount is done in parallel.

A simple hack would be to set PF_MEMALLOC in drop_slab(), but I do not
think it is a good idea.

> 
> >And it is a nice small change.
> >Have you had any other feedback on this?
> here it is :)
> 

Thanks for your detailed feedback

> Thanks,
> Kirill
> 

Regards,
Balbir