On Mon, Jun 09, 2014 at 12:30:34PM +0900, J. R. Okajima wrote:
> Linus Torvalds:
> > So I ended up doing an rc8 because I was a bit worried about some
> > last-minute dcache fixes, but it turns out that nobody seemed to even
> > notice those. We did have other issues during the week, though, so it
>       :::
> I am afraid there is a problem in dcache. Please read
> http://marc.info/?l=linux-fsdevel&m=140214911608925&w=2

There is a problem, all right, but your fix doesn't really fix it - just
narrows the race window ;-/  I would really like to detach the bugger
as soon as __dentry_kill() removes it from the list of children, but
unfortunately NFS wants it to be still valid in ->d_iput() (BTW,
nfs_can_unlink() is doing something very odd -
        parent = dget_parent(dentry);
        if (parent == NULL)
                goto out_free;
is pointless, since dget_parent() never returns NULL; what was that
check trying to accomplish?)

Your scenario isn't feasible as described, but something similar can
happen with *two* shrinkers racing - dirB might've ended up on one list,
with fileC looked up a bit later, ending up on another list right after
that.  So the problem is real; unfortunately, DCACHE_DENTRY_KILLED might
have appeared right after we'd dropped ->d_lock.

So I suspect that the right fix is a bit trickier - in addition to check
on the fast path (i.e. when trylock gets us the lock on parent), we need
        * get rcu_read_lock() before dropping ->d_lock.
        * check if dentry is already doomed right after taking rcu_read_lock();
if not, any value we might see in ->d_parent afterwards will point to object
not freed until we drop rcu_read_lock.

IOW, something like the delta below.  Comments?

PS: apologies for being MIA; caught some crap, spent the last week being
very unhappy ;-/  I'll send a pull request tomorrow morning.

diff --git a/fs/dcache.c b/fs/dcache.c
index be2bea8..65ec10f 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -532,10 +532,16 @@ static inline struct dentry *lock_parent(struct dentry 
        struct dentry *parent = dentry->d_parent;
        if (IS_ROOT(dentry))
                return NULL;
+       if (unlikely((int)dentry->d_lockref.count < 0))
+               return NULL;
        if (likely(spin_trylock(&parent->d_lock)))
                return parent;
-       spin_unlock(&dentry->d_lock);
+       if (unlikely((int)dentry->d_lockref.count < 0)) {
+               rcu_read_unlock();
+               return NULL;
+       }
+       spin_unlock(&dentry->d_lock);
        parent = ACCESS_ONCE(dentry->d_parent);
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Reply via email to