Re: [PATCH] kmalloc_percpu

[Jens Axboe <axboe@suse.de>]

On Mon, May 05 2003, Andrew Morton wrote:
> "David S. Miller" <davem@redhat.com> wrote:
> >
> >    From: Andrew Morton <akpm@digeo.com>
> >    Date: Mon, 5 May 2003 23:42:48 -0700
> >    
> >    Can't think of anything very clever there, except to go and un-percpuify the
> >    disk stats.  I think that's best, really - disk requests only come in at 100
> >    to 200 per second - atomic_t's or int-plus-per-disk-spinlock will be fine.
> >    
> > Use some spinlock we already have to be holding during the
> > counter bumps.
> 
> Last time we looked at that, q->lock was already held in almost all the right
> places so yes, that'd work.

As far as I can see, queue lock _is_ held in all the right spot. At
least where it matters, adding new samples.

> > Frankly, these things don't need to be %100 accurate.  Using
> > a new spinlock or an atomic_t for this seems rediculious.
> 
> The disk_stats structure has an "in flight" member.  If we don't have proper
> locking around that, disks will appear to have -3 requests in flight for all
> time, which would look a tad odd.

So check for < 0 in flight? I totally agree with davem here.

-- 
Jens Axboe