Re: [PATCH 2/2] ovl: allow distributed fs as lower layer

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

Miklos Szeredi <miklos@szeredi.hu> writes:

> On Fri, Jun 05, 2015 at 01:07:15AM +0100, Al Viro wrote:
>
>> Umm...  Cosmetical point is that this
>> 
>> > +static bool ovl_remote(struct dentry *root)
>> > +{
>> > +	const struct dentry_operations *dop = root->d_op;
>> > +
>> > +	return dop && (dop->d_revalidate || dop->d_weak_revalidate);
>> > +}
>> 
>> is better done as
>> 	root->d_flags & (DCACHE_OP_REVALIDATE | DCACHE_OP_WEAK_REVALIDATE)
>
> Okay.
>
>> 
>> More interesting question is whether anything in the system relies on
>> existing behaviour that follows ->d_revalidate() returning 0.
>
> Hmm, d_invalidate() almost always follows ->d_revalidate().  Almost, becuase RCU
> lookup can get aborted at that point.  We can easily stick d_invalidate() in
> there for the non-RCU case.
>
> Regular lookup also almost always follows ->d_revalidate().  Except if
> allocation of new dentry fails.  So relying on this would be buggy (which is not
> to say nobody does it).
>
>>  Have you tried to mount e.g. procfs as underlying layer and torture it for a
>> while?
>
> I did try now.  Nothing bad happened during the test (parallel stat(1) of the
> whole overlayed proc tree).
>
> My laptop froze while trying to write this mail.  But it's 8 years old and when
> the fan starts to make noises and the weather is hot, it does this sometimes.  I
> don't think that has anything to do with overlayfs, but will do more
> testing...

A nasty corner case to be aware of (and I think this is part of what Al
was warning about).  /proc/sys/net is different depending upon which
current->nsproxy->net_ns.

Eric