On Sun, Jul 17, 2011 at 7:08 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Sun, Jul 17, 2011 at 06:13:34PM -0700, Hugh Dickins wrote:
>
>> And in the genuine negative dentry case: at present (with or without
>> my or Linus's patches) there is no final nd->seq check, is there?
>
> For stat() - no, there isn't. We really bail out with -ENOENT, no matter
> what. Racy, which is what you are hitting.
The fundamental bug is in the dentry pruning. Not in lookup. If the
lookup sees a negative dentry, it will currently return ENOENT. Not
unlazy the thing like you claim. Your patch makes it unlazy the RCU
walk for a negative dentry and then retry that it currently doesn't -
for both the racy case (nobody cares, since that isn't performance
critical) _and_ for the normal real negative case.
And it's that "normal real negative case" that I really think that
your "goto unlazy" is just wrong. It's right for the sequence number
change, but not for the "!inode" case.
Now, I do agree that maybe that case simply should check the dentry
sequence count. I wish all cases did. Hugh patch did that. But the
reason I dislike Hugh's patch is that when I say "I wish they all
did", I mean that I dislike the special casing. And Hugh's patch just
adds *more* special casing for that NULL entry - I'd wish we just
always did it regardless of whether it was NULL or not.
So I claim that your patch changes *way* more than mine does. I think
I'd prefer Hugh's over that one. I still think that right now, my
one-liner is actually the one that changes things the least (and I
don't mean in size of patch, but in behavior and logic).
Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
