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

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

On Sun, Sep 01, 2013 at 03:16:24PM -0700, Linus Torvalds wrote:

> Does not seem to matter. Still 66% mntput_no_expire, 31% path_init.
> And that lg_local_lock() takes 5-6% of CPU, pretty much all of which
> is that single xadd instruction that implements the spinlock.
> 
> This is on /tmp, which is tmpfs. But I don't see how any of that could
> matter. "mntput()" does an unconditional call to mntput_no_expire(),
> and mntput_no_expire() does that br_read_lock() unconditionally too.
> 
> Note that I'm talking about that "cheap" *read* lock being expensive.
> It's the local one, not the global one. So it's not what Waiman saw
> with the global lock. This is a local per-cpu thing.
> 
> That read-lock is supposed to be very cheap - it's just a per-cpu
> spinlock. But it ends up being very expensive for some reason. I'm not
> quite sure why - I don't see any lg_global_lock() calls at all, so...
> 
> I wonder if there is some false sharing going on. But I don't see that
> either, this is the percpu offset map afaik:
> 
>   000000000000f560 d files_lglock_lock
>   000000000000f564 d nr_dentry
>   000000000000f568 d last_ino
>   000000000000f56c d nr_unused
>   000000000000f570 d nr_inodes
>   000000000000f574 d vfsmount_lock_lock
>   000000000000f580 d bh_accounting
> 
> and I don't see anything there that would get cross-cpu accesses, so
> there shouldn't be any cacheline bouncing. That's the whole point of
> percpu variables, after all.

Hell knows...  Are you sure you don't see br_write_lock() at all?  I don't
see anything else that would cause cross-cpu traffic with that layout...