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

From: Wu Fengguang
Date: Fri Oct 02 2009 - 06:55:37 EST


On Thu, Oct 01, 2009 at 10:02:07AM +0800, Nick Piggin wrote:
> 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.

Good ideas. In fact I tried them in the very beginning.
The ratios are roughly at the same level (which is somehow unexpected):

total time add_to_page_cache_lru percent stddev
tmpfs 1579056274.576 2615476.234 0.1656354036338758%
sparse 1074931917.425 3001273 0.27920586888791538%

Workload is to copy 1G /dev/zero to /dev/shm/, or 1G sparse file
(ext2) to /dev/null.

echo 1 > /debug/tracing/function_profile_enabled
cp /dev/zero /dev/shm/
echo 0 > /debug/tracing/function_profile_enabled

dd if=/dev/zero of=/mnt/test bs=1k count=1 seek=1048575
echo 1 > /debug/tracing/function_profile_enabled
cp /mnt/test/sparse /dev/null
echo 0 > /debug/tracing/function_profile_enabled

> > > 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 :)

Yes it makes sense.

> > > 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.

Yes I do. Even better, for this perticular race, we managed to avoid
it completely without introducing overheads in fast path :)

Thanks,
Fengguang

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/