Re: [PATCH 2/7] proc: Implement support for automounts in task directories

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 06 Nov 2008 19:51:23 -0800 ebiederm@xmission.com (Eric W. Biederman) wrote:

> If we could do all of this with reference counting so that the
> mount would persist exactly until the last user of it has gone
> away without a periodic poll I would love it.  But the infrastructure
> doesn't support that today,

Well that sucks.  The free-on-last-put idiom occurs in so many places
and serves us so well.  I wonder what went wrong here?

I guess it has interactions with dentry and inode cache aging which
could get tricky.

> and where this is at least partially
> a bug fix I would rather not have the change depend on enhancing
> the VFS.
> 
> The algorithm is actually very aggressive and in practice you don't
> see any /proc/<pid>/net showing up as a mount point.

Do you think it has failure modes?  Most particularly: obscure usage
patterns which can cause memory exhaustion?

> > Obviously, that becomes clearer as one spends more time with the code,
> > but I wonder whether this has all been made as maintainble as it
> > possibly could be.
> 
> Good question.
> 
> In the sense of will we have to go through and futz with the code all
> of the time.  The abstraction seems good.   You put a mount on
> the proc_automounts list with do_add_mounts and it goes away eventually
> with all of the vfs rules maintained.
> 
> In the sense of can the code be read?    Perhaps it could be better.
> I expect it helps to have run the code and see /proc/net as a filesystem.
> that is magically mounted.

'twould be a useful contribution if you were to enshrine your
discoveries in /*these things*/.  You knew I was working up to that :)