Re: [PATCH 5/6] vfs: reorganize do_lookup

From: Miklos Szeredi
Date: Thu Apr 05 2012 - 10:44:20 EST


Al Viro <viro@xxxxxxxxxxxxxxxxxx> writes:

> On Tue, Apr 03, 2012 at 10:13:18AM +0200, Miklos Szeredi wrote:
>
>> The do_last() reorganization in the atomic-open series needs to be split
>> up, I realize. Do you have any other high level comments about that
>> series?
>
> Yes.
> a) Please, pull removal of open-from-d_revalidate() as far in front
> of the queue as possible, *along* *with* -EOPENSTALE stuff. Without the
> latter the former is simply broken - we might have hit -ESTALE before and
> LOOKUP_REVAL might have already been set, so just failing ->open() with
> ESTALE may end up not repeating it.
> TBH, had that thing been in front of the queue, I would've put
> it into the last pull request; that particular idiocy (NFS4 doing open
> from all methods involved, except for ->open()) had been a serious source
> of annoyance for a long time. It really needs killing...

Will post updated series in a minute.

>
> b) opendata is simply bogus. You need to pass caller-allocated
> struct file in any case, right? So why not use it to pass what you
> need to pass? We want to be able to tell "it's a symlink, here's the
> vfsmount/dentry, now sod off and handle it yourself"?

My concern is about the visibility of a half cooked file pointer to the
filesystem. The compiler won't notice anything wrong with the
following, neither will casual testing, but it will go boom at
inopportune times:

foo_atomic_open(..., struct file *filp, ...)
{
/* ... */

if (unlikely error)
goto cleanup;

filp = finish_open(filp, dentry, NULL);

/* ... */

if (some other error)
goto cleanup;

return filp;

cleanup:
fput(filp);
return ERR_PTR(error);
}


Thanks,
Miklos

--
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/