Re: [PATCH RFC ticketlock] Auto-queued ticketlock

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Jun 12, 2013 at 01:26:25PM -0700, Linus Torvalds wrote:

> For similar reasons, I think you need to still maintain the d_lock in
> d_prune_aliases etc. That's a slow-path, so the fact that we add an
> atomic sequence there doesn't much matter.
> 
> However, one optimization missing from your patch is obvious in the
> profile. "dget_parent()" also needs to be optimized - you still have
> that as 99% of the spin-lock case. I think we could do something like
> 
>    rcu_read_lock();
>    parent = ACCESS_ONCE(dentry->d_parent);
>    if (atomic_inc_nonzero(&parent->d_count))
>       return parent;
>    .. get d_lock and do it the slow way ...
>    rcu_read_unlock();
> 
> to locklessly get the parent pointer. We know "parent" isn't going
> away (dentries are rcu-free'd and we hold the rcu read lock), and I
> think that we can optimistically take *any* parent dentry that
> happened to be valid at one point. As long as the refcount didn't go
> down to zero. Al?

What will you do with __d_rcu_to_refcount()?  Any such scheme has to
hold d_lock from zero->non-zero d_count changes, or atomic_dec_and_lock
in dput() won't help at all.  As it is, both comlete_walk() and unlazy_walk()
are grabbing ->d_lock on the dentry we'd reached, so they can call that
sucker.  And that'll give you ->d_lock contention when a bunch of threads
are hitting the same file; I don't see how atomics would avoid that
one...