Re: [patchset] Lockfree fd lookup 0 of 5

[viro@parcelfarce.linux.theplanet.co.uk]

On Tue, Aug 03, 2004 at 02:53:17PM +0530, Ravikiran G Thirumalai wrote:
> > How about this for comparison?  That's just a dumb "convert to rwlock"
> > patch; we can be smarter in e.g. close_on_exec handling, but that's a
> > separate story.
> > 
> 
> I ran tiobench on this patch and here is the comparison:
> 
> 
> Kernel		Seqread		Randread	Seqwrite	Randwrite
> --------------------------------------------------------------------------
> 2.6.7		410.33		234.15		254.39		189.36
> rwlocks-viro	401.84		232.69		254.09		194.62
> refcount (kref)	455.72		281.75		272.87		230.10

Thanks.  IOW, we are really seeing cacheline bounces - not contention...

I'm still not sure that in the current form patch is a good idea.  The thing
is, existing checks for ->f_count value are bogus in practically all cases;
IMO we should sort that out before making any decision based on the need for
such checks.  Ditto for uses of fcheck() (open-coded or not) in arch/*.

I agree that some form of "postpone freeing and make fget() lockless" would
make sense, but I'd rather clean the area *before* doing that; afterwards it
will be harder and results of cleanup can affect the patches in non-trivial
way.