Re: [RFC][PATCH] HWPOISON: remove the unsafe __set_page_locked()

[Nick Piggin <npiggin@suse.de>]

On Mon, Sep 28, 2009 at 04:44:01PM +0800, Wu Fengguang wrote:
> On Mon, Sep 28, 2009 at 03:20:25AM +0800, Nick Piggin wrote:
> > On Sun, Sep 27, 2009 at 06:47:39PM +0800, Wu Fengguang wrote:
> > > > 
> > > > And standard deviation is 0.04%, much larger than the difference 0.008% ..
> > > 
> > > Sorry that's not correct. I improved the accounting by treating
> > > function0+function1 from two CPUs as an integral entity:
> > > 
> > >                  total time      add_to_page_cache_lru   percent  stddev
> > >          before  3880166848.722  9683329.610             0.250%   0.014%
> > >          after   3828516894.376  9778088.870             0.256%   0.012%
> > >          delta                                           0.006%
> > 
> > I don't understand why you're doing this NFS workload to measure?
> 
> Because it is the first convenient workload hit my mind, and avoids
> real disk IO :)

Using tmpfs or sparse files is probably a lot easier.

 
> > I see significant nfs, networking protocol and device overheads in
> > your profiles, also you're hitting some locks or something which
> > is causing massive context switching. So I don't think this is a
> > good test.
> 
> Yes there are overheads. However it is a real and common workload.

Right, but so are lots of other workloads that don't hit
add_to_page_cache heavily :)
 

> > But anyway as Hugh points out, you need to compare with a
> > *completely* fixed kernel, which includes auditing all users of page
> > flags non-atomically (slab, notably, but possibly also other
> > places).
> 
> That's good point. We can do more benchmarks when more fixes are
> available. However I suspect their design goal will be "fix them
> without introducing noticeable overheads" :)

s/noticeable//

The problem with all the non-noticeable overheads that we're
continually adding to the kernel is that we're adding them to
the kernel. Non-noticeable part only makes it worse because
you can't bisect them :)
 

> > One other thing to keep in mind that I will mention is that I am
> > going to push in a patch to the page allocator to allow callers
> > to avoid the refcounting (atomic_dec_and_test) in page lifetime,
> > which is especially important for SLUB and takes more cycles off
> > the page allocator...
> >
> > I don't know exactly what you're going to do after that to get a
> > stable reference to slab pages. I guess you can read the page
> > flags and speculatively take some slab locks and recheck etc...
> 
> For reliably we could skip page lock on zero refcounted pages.
> 
> We may lose the PG_hwpoison bit on races with __SetPageSlub*, however
> it should be an acceptable imperfection.

I think if you're wiling to accept these problems, then it is
completely reasonable to also accept similar races with kernel
fastpaths to avoid extra overheads there.