Re: Q. hlist_bl_add_head_rcu() in d_alloc_parallel()

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

On Thu, Jun 23, 2016 at 10:57:09PM -0700, Linus Torvalds wrote:

> The non-RCU case basically never shows up in profiles (it used to,
> with symlinks, but you fixed that case), and if it ever does I suspect
> that the fix will be to make sure we don't fall out of rcu.
> 
> So don't worry too much about __d_lookup() being a hot case, I don't
> think it is.
> 
> (Of course, if you have a load that shows me wrong, let's look at it
> by all means. Maybe the loads I have used have been bad)

BTW, speaking of that area - is there any reason why dentry_cmp()
isn't simply
	return dentry_string_cmp(lockless_dereference(dentry->d_name.name),
				 ct, tcount);
other than "it predates lockless_dereference()"?

	The only difference is s/ACCESS_ONCE/READ_ONCE/, AFAICS -
lockless_dereference() uses the latter these days.  And that shouldn't
be a problem; as the matter of fact, *all* remaining ACCESS_ONCE in
fs/dcache.c look like they should become READ_ONCE.  Objections?

Al, digging through the barrier-related issues with dentries and peeling the
paint off the walls in process...