[PATCH 0/7 V2] xfs: use generic percpu counters for icsb

[Dave Chinner]

On Thu, Feb 05, 2015 at 09:08:40AM -0500, Brian Foster wrote:
> On Thu, Feb 05, 2015 at 07:54:02AM +1100, Dave Chinner wrote:
> > Hi folks,
> > 
> > This is the second version of the generic per-cpu counter rework
> > patch series. The first version can be found here:
> > 
> > http://oss.sgi.com/archives/xfs/2015-02/msg00000.html
> > 
> > New in V2:
> > 
> > - drop the moving of the struct xfs_sb to xfs_super.h
> > - fixed all the little things that Christoph and Brian noted.
> > - keep the per-cpu counters in the struct xfs_mount and keep the
> >   functions to sync them with the struct xfs_sb values when read
> >   from disk or written to disk.
> > - integrated Christoph Hellwig's additional cleanup patch. This was
> >   done by:
> > 	- intergating xfs_mod_XXX factoring into the relevant percpu
> > 	  counter conversion patch
> > 	- separating out xfs_mod_frextents into it's won patch
> > 	- separating out the replacement of
> > 	  xfs_mod_incore_sb_batched
> > 	- doing all the now unused API removal in a separate patch
> > 
> > The series passes xfstests without regressions, and no scalability
> > issues have been seen in my performance tests on a 16p VM. SGI - you
> > still need to test this, though. :)
> > 
> > Thoughts, comments?
> > 
> 
> All in all this looks pretty good to me save a couple notes pointed out
> in the patches. In a quick test, this handles the imaxct overshoot
> problem Eric noted much better than the current implementation. That
> said, it's still not precise.

Right.

> My understanding is that's fine, but I wonder if we want to tack on a
> variant of Eric's original patch as well so when we still do overshoot
> imaxpct (albeit by much less than before: I reproduce an overshoot of
> <100 inodes vs several thousand) we at least report an accurate inode
> count. Thoughts?

Yes, Eric's patch is still necessary.

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com