Re: [PATCH v7 1/4] spinlock: A new lockref structure for lockless update of refcount

[Linus Torvalds <torvalds@linux-foundation.org>]

On Sun, Sep 8, 2013 at 5:03 PM, Al Viro <viro@zeniv.linux.org.uk> wrote:
>
> There's one exception - basically, we decide to put duplicates of
> reference(s) we hold into (a bunch of) structures being created.  If
> we decide that we'd failed and need to roll back, the structures
> need to be taken out of whatever lists, etc. they'd been already
> put on and references held in them - dropped.  That removal gets done
> under a spinlock.  Sure, we can string those structures on some kind
> of temp list, drop the spinlock and do dput() on everything in there,
> but it's much more convenient to just free them as we are evicting
> them, doing dput() as we go.  Which is safe, since we are still have
> the references used to create these buggers pinned down.

Hmm. Which codepath does this? Because I got curious and added back
the __might_sleep() unconditionally to dput() just to see (now that I
think that the dput() bugs are gone), and at least under normal load
it doesn't trigger. I even wrote a thing that just constantly creates
and renames the target file concurrently with looking it up, so that
I've stress-tested the RCU sequence number failure path (and verified
with a profile that yes, it does trigger the "oops, need to retry"
case). I didn't test anything odd at all (just my dentry stress tests
and a regular graphical desktop), though.

And I have too much memory to sanely stress any out-of-memory situations.

#firstworldkerneldeveloperproblems

            Linus