Re: remove zero_page (was Re: -mm merge plans for 2.6.24)

[Nick Piggin <nickpiggin@yahoo.com.au>]

On Thursday 04 October 2007 01:21, Linus Torvalds wrote:
> On Wed, 3 Oct 2007, Nick Piggin wrote:
> > I don't know if Linus actually disliked the patch itself, or disliked
> > my (maybe confusingly worded) rationale?
>
> Yes. I'd happily accept the patch, but I'd want it clarified and made
> obvious what the problem was - and it wasn't the zero page itself, it was
> a regression in the VM that made it less palatable.

OK, revised changelog at the end of this mail...


> I also thought that there were potentially better solutions, namely to
> simply avoid the VM regression, but I also acknowledge that they may not
> be worth it - I just want them to be on the table.
>
> In short: the real cost of the zero page was the reference counting on the
> page that we do these days. For example, I really do believe that the
> problem could fairly easily be fixed by simply not considering zero_page
> to be a "vm_normal_page()". We already *do* have support for pages not
> getting ref-counted (since we need it for other things), and I think that
> zero_page very naturally falls into exactly that situation.
>
> So in many ways, I would think that turning zero-page into a nonrefcounted
> page (the same way we really do have to do for other things anyway) would
> be the much more *direct* solution, and in many ways the obvious one.

That was my first approach. It isn't completely trivial, but
vm_normal_page() does play a part (but we end up needing a
vm_normal_page() variant -- IIRC vm_normal_or_zero_page()). But taken
as a whole, non-refcounted zero_page is obviously a lot more work than
no zero page at all :)


> HOWEVER - if people think that it's easier to remove zero_page, and want
> to do it for other reasons, *AND* can hopefully even back up the claim
> that it never matters with numbers (ie that the extra pagefaults just make
> the whole zero-page thing pointless), then I'd certainly accept the patch.

I have done some tests which indicate a couple of very basic common tools
don't do much zero-page activity (ie. kbuild). And also combined with some
logical arguments to say that a "sane" app wouldn't be using zero_page much.
(basically -- if the app cares about memory or cache footprint and is using
many pages of zeroes, then it should have a more compressed representation
of zeroes anyway).

However there is a window for some "insane" code to regress without the
zero_page. I'm not arguing that we don't care about those, however I have
no way to guarantee they don't exist. I hope we wouldn't get a potentially
useless complexity like this stuck in the VM forever just because we don't
_know_ whether it's useful to anybody...

How about something like this?

---
From: Nick Piggin <npiggin@suse.de>

The commit b5810039a54e5babf428e9a1e89fc1940fabff11 contains the note

  A last caveat: the ZERO_PAGE is now refcounted and managed with rmap
  (and thus mapcounted and count towards shared rss).  These writes to
  the struct page could cause excessive cacheline bouncing on big
  systems.  There are a number of ways this could be addressed if it is
  an issue.

And indeed this cacheline bouncing has shown up on large SGI systems.
There was a situation where an Altix system was essentially livelocked
tearing down ZERO_PAGE pagetables when an HPC app aborted during startup.
This situation can be avoided in userspace, but it does highlight the
potential scalability problem with refcounting ZERO_PAGE, and corner
cases where it can really hurt (we don't want the system to livelock!).

There are several broad ways to fix this problem:
1. add back some special casing to avoid refcounting ZERO_PAGE
2. per-node or per-cpu ZERO_PAGES
3. remove the ZERO_PAGE completely

I will argue for 3. The others should also fix the problem, but they
result in more complex code than does 3, with little or no real benefit
that I can see. Why?

Inserting a ZERO_PAGE for anonymous read faults appears to be a false
optimisation: if an application is performance critical, it would not be
doing many read faults of new memory, or at least it could be expected to
write to that memory soon afterwards. If cache or memory use is critical,
it should not be working with a significant number of ZERO_PAGEs anyway
(a more compact representation of zeroes should be used).

As a sanity check -- mesuring on my desktop system, there are never many
mappings to the ZERO_PAGE (eg. 2 or 3), thus memory usage here should not
increase much without it.

When running a make -j4 kernel compile on my dual core system, there are
about 1,000 mappings to the ZERO_PAGE created per second, but about 1,000
ZERO_PAGE COW faults per second (less than 1 ZERO_PAGE mapping per second
is torn down without being COWed). So removing ZERO_PAGE will save 1,000
page faults per second, and 2,000 bounces of the ZERO_PAGE struct page
cacheline per second when running kbuild, while saving less than 1 page
clearing operation per second (even 1 page clear is far cheaper than a
thousand cacheline bounces between CPUs).

Of course, neither the logical argument nor these checks give anything
like a guarantee of no regressions. However, I think this is a reasonable
opportunity to remove the ZERO_PAGE from the pagefault path.

The /dev/zero ZERO_PAGE usage and TLB tricks also get nuked.  I don't see
much use to them except complexity and useless benchmarks.  All other users
of ZERO_PAGE are converted just to use ZERO_PAGE(0) for simplicity.  We can
look at replacing them all and ripping out ZERO_PAGE completely if/when
this patch gets in.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/