Re: [patch 1/2] fs: fix d_validate (again)

[Nick Piggin <npiggin@kernel.dk>]

On Wed, Dec 15, 2010 at 12:53:15PM +0000, Al Viro wrote:
> On Tue, Nov 23, 2010 at 09:22:41PM +1100, Nick Piggin wrote:
> > Again, I'd like to please fix d_validate. It can be trivally fixed
> > without delving into filesystem code, and it does not prevent the proper
> > careful removal of d_validate and untrusted dentry caches from
> > filesystems.
> > 
> > The right way to merge is fix the bugs in core code now, and remove it
> > completely if and when filesystems stop using it. But it's no longer a
> > liability or problem to maintain, so it can stay as long as it likes.
> 
> TBH, I would prefer to kill the damn thing completely.  Never liked it,
> and I think I see how to get rid of it.

Yes that's the logical next step.

 
> Look: what happens is that ncpfs reads the entire directory from server
> and since it gets all metadata for free anyway, it creates all dentries
> as it goes (assiging i_ino, etc.) and sticks references to them into
> page cache for that directory.  readdir() goes through those pages and
> as long as all dentries it sees are still alive (and pages still present),
> it uses them.  If something's gone or the cache is old enough, we just
> evict all those pages and reread.  If some dentries are around during
> reread (or initial read, for that matter), we stick references to those
> back into the repopulated page cache.
> 
> The order of entries matches that in the metadata we'd got from server.
> 
> OK.  So let's do that:
> 	* stick reference back into the page in dentry->d_fsdata (the current
> uses are becoming obsolete with new scheme)
> 	* have d_iput() remove the reference to dentry from page under
> spinlock
> 	* have page eviction remove the references from corresponding dentries
> under the same spinlock
> 	* instead of d_validate() crap, grab the spinlock, lock dentry,
> check if it's still alive and do dget_locked() on it if it is (and unlock,
> of course).
> 
> Voila - d_validate() is no more.  IIRC, smbfs was essentially the same
> as ncpfs in that respect.  Comments?

TBH, it sounds plausible, but I have no way to test ncpfs, so I just
tried to stay away from such a big change after one look at trying to
rip the cache out.

It would be great if someone can do it, but can we not couple old
filesystem changes with this patch?  Ie. just merge the bugfixes no
upstream and in -stable (nobody will care about performance, I'd bet on
it). Then d_validate can go away naturally after the filesystem changes
are merged and it has no callers left.