Re: [Lse-tech] lockmeter results comparing 2.4.17, 2.5.3, and 2.5.5

[Andrew Morton <akpm@zip.com.au>]

"Martin J. Bligh" wrote:
> 
> > inode_lock hold times are a problem for other reasons.  Leaving this
> > unfixed makes the preepmtible kernel rather pointless....  An ideal
> > fix would be to release inodes based on VM pressure against their backing
> > page.  But I don't think anyone's started looking at inode_lock yet.
> >
> > The big one is lru_list_lock, of course.  I'll be releasing code in
> > the next couple of days which should take that off the map.  Testing
> > would be appreciated.
> 
> Seeing as people seem to be interested ... there are some big holders
> of BKL around too - do_exit shows up badly (50ms in the data Hanna
> posted, and I've seen that a lot before).

That'll be where exit() takes down the tasks's address spaces.  
zap_page_range().  That's a nasty one.

> I've seen sync_old_buffers
> hold the BKL for 64ms on an 8way Specweb99 run (22Gb of RAM?)
> (though this was on an older 2.4 kernel, and might be fixed by now).

That will still be there - presumably it's where we walk the
per-superblock dirty inode list.  hmm.

For lru_list_lock we can do an end-around by not using
buffers at all.

The other big one is truncate_inode_pages().  With ratcache
it's not a contention problem, but it is a latency problem.
I expect that we can drastically reduce the lock hold time
there by simply snipping the wholly-truncated pages out of
the tree, and thus privatising them so they can be disposed
of outside any locking.

-