Re: Linux 2.4.33-rc1

[Willy Tarreau]

Hi Marcelo,

On Wed, Jun 21, 2006 at 10:50:15AM -0300, Marcelo Tosatti wrote:
> Hi Willy,
> 
> On Wed, Jun 21, 2006 at 12:23:57AM +0200, Willy Tarreau wrote:
> > On Mon, Jun 19, 2006 at 08:45:06PM -0300, Marcelo Tosatti wrote:
> > > > 
> > > > ---- from here ----
> > > > 
> > > > 
> > > > > +		inode = dentry->d_inode;
> > > > > +		if (inode)
> > > > > +			atomic_inc(&inode->i_count);
> > > > >  		error = vfs_unlink(nd.dentry->d_inode, dentry);
> > > > >  	exit2:
> > > > >  		dput(dentry);
> > > > >  	}
> > > > >  	up(&nd.dentry->d_inode->i_sem);
> > > > > +	if (inode)
> > > > > +		iput(inode);
> > > > 
> > > > ---- to here ----
> > > > 
> > > > I believe that nd.dentry->d_inode cannot vanish because it is protected by the
> > > > down(->i_sem) before and the up(->i_sem) after. Am I right or am I missing
> > > > something important ?
> > > 
> > > Indeed it can't, but dentry->d_inode will be set to NULL by
> > > nfs_unlink->nfs_safe_remove->d_delete. Thus the problem.
> > 
> > What puzzles me is how are we supposed to up(&nd.dentry->d_inode->i_sem) if
> > dentry->d_inode can become NULL ? 
> 
>         down(&nd.dentry->d_inode->i_sem);
>         dentry = lookup_hash(&nd.last, nd.dentry);
>         error = vfs_unlink(nd.dentry->d_inode, dentry);
> 					       ^^^^^^^
> 
> It does vfs_unlink(parent, child). The child is killed.
>
> > simply by keeping a copy of it ? I thought
> > that the down() protected the whole thing, but may be that's stupid anyway.
> 
> It does not protect from last reference going away, which is what
> happens when NFS d_delete's the dentry. dentry_iput() gets called, and
> maybe the last reference to "struct inode" might be gone too.

OK for this one, but what I mean is that vfs_unlink() will kill dentry
(its second argument, which your fix protects), while the down()/up()
we initially suspected, and tried to protect, play with nd.dentry->d_inode,
which is the directory in which we are removing dentry. And I can't see how
nd.dentry->d_inode can be killed during vfs_unlink(), reason why I thought
the fix above did not provide any benefit.

> > I've been running rc1 without this patch for a few hours and during kernel
> > compiles without a problem, so I'm not sure about what to think about the
> > other changes which were apparently harmless too :-/
> 
> Use NFS... :)

I'm using it at home right now, doing lots of git-checkout -f and make -j
dep ; make -j 4 bzImage modules on SMP without a glitch. Right now I'm
running rc2 without trouble and I noticed you did not merge the fix for
sys_unlink() that we discussed above. May be you simply have forgotten it
or may be we are in sync but were talking about different things.

> > Well, if I resume it right, we only need to merge your patch and mine and
> > it *should* be OK.
> > 
> > BTW, I've been reviewing the PaX patch and found *at least* one patch
> > that should be merged (fix for oops). I'll send it separately, and it's
> > queued in -upstream.
> 
> Merged, thanks.

Thanks,
Willy

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/