Re: Bad news: apparent bug in casMutVar going back to 7.2

Hey Ryan, i've made the leap to using 7.8 on my machine, so i'll first have to get some pull requests in on atomic-primops before I can test it locally :), expect those patches later today! looks like gcc's inline ASM logic is pretty correct, after testing it a bit locally, pardon my speculative jumping the gun. On Sat, Feb 1, 2014 at 9:10 AM, Ryan Newton wrote:

Hi Carter & others,

Carter, yes, this is CAS on pointers and in my next mail I'll try to come up with some hypotheses as to why we may have (remaining) problems there.

But first, I have been assured that on x86 there is no failure mode in which doing a comparison on the value read by CAS should not correctly diagnose success or failure (same as directly reading the Zero Flag) [1].

And yet, there's this discrepancy, where the modified casMutVar that I linked to does not have the failure. As for reproducing the failure, either of the two following tests will currently show problems:

- Two threads try to casIORef False->True, both succeed - 120 threads try to read, increment, CAS until they succeed. The total is often not 120 because multiple threads think the successfully incremented, say, 33->34.

Here's a specific recipe for the latter test on GHC 7.6.3 Mac or Linux:

*git clone git@github.com:rrnewton/haskell-lockfree-queue.git * *cd haskell-lockfree-queue/AtomicPrimops/*

*git checkout 1a1e7e55f6706f9e5754*

*cabal sandbox init*

*cabal install -f-withTH -fforeign ./ ./testing --enable-tests*

*./testing/dist/dist-sandbox-*/build/test-atomic-primops/test-atomic-primops -t n_threads*

You may have to run the last line several times to see the failure.

Best, -Ryan

[1] I guess the __sync_bool_compare_and_swap intrinsic which reads ZF is there just to avoid the extra comparison.

[2] P.S. I'd like to try this on GHC head, but the RHEL 6 machine I usually use to build it is currently not validating (below error, commit 65d05d7334). After I debug this gmp problem I'll confirm that the bug under discussion applies on the 7.8 branch.

./sync-all checkout ghc-7.8 sh validate ... /usr/bin/ld: libraries/integer-gmp/gmp/objs/aors.o: relocation R_X86_64_32 against `__gmpz_sub' can not be used when making a shared object; recompile with -fPIC libraries/integer-gmp/gmp/objs/aors.o: could not read symbols: Bad value

On Sat, Feb 1, 2014 at 2:55 AM, Carter Schonwald < carter.schonwald@gmail.com> wrote:

...

Ryan, is your benchmark using CAS on pointers, or immediate words? trying to get atomic primops to build on my 7.8 build on my mac

On Sat, Feb 1, 2014 at 2:44 AM, Carter Schonwald < carter.schonwald@gmail.com> wrote:

...

https://ghc.haskell.org/trac/ghc/ticket/8724#ticket is the ticket

when i'm more awake i'll experiment some more

On Sat, Feb 1, 2014 at 2:33 AM, Carter Schonwald < carter.schonwald@gmail.com> wrote:

...

i have a ticket for tracking this, though i'm thinking my initial attempt at a patch generates the same object code as it did before.

@ryan, what CPU variant are you testing this on? is this on a NUMA machine or something?

On Sat, Feb 1, 2014 at 1:58 AM, Carter Schonwald < carter.schonwald@gmail.com> wrote:

...

woops, i mean cmpxchgq

On Sat, Feb 1, 2014 at 1:36 AM, Carter Schonwald < carter.schonwald@gmail.com> wrote:

...

ok, i can confirm that on my 64bit mac, both clang and gcc use cmpxchgl rather than cmpxchg i'll whip up a strawman patch on head that can be cherrypicked / tested out by ryan et al

On Sat, Feb 1, 2014 at 1:12 AM, Carter Schonwald < carter.schonwald@gmail.com> wrote:

> Hey Ryan, > looking at this closely > Why isn't CAS using CMPXCHG8B on 64bit architectures? Could that be > the culprit? > > Could the issue be that we've not had a good stress test that would > create values that are equal on the 32bit range, but differ on the 64bit > range, and you're hitting that? > > Could you try seeing if doing that change fixes things up? > (I may be completely wrong, but just throwing this out as a naive > "obvious" guess) > > > On Sat, Feb 1, 2014 at 12:58 AM, Ryan Newton wrote: > >> Then again... I'm having trouble seeing how the spec on page 3-149 >> of the Intel manual would allow the behavior I'm seeing: >> >> >> http://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32... >> >> Nevertheless, this is exactly the behavior we're seeing with the >> current Haskell primops. Two threads simultaneously performing the same >> CAS(p,a,b) can both think that they succeeded. >> >> >> >> >> >> On Sat, Feb 1, 2014 at 12:33 AM, Ryan Newton wrote: >> >>> I commented on the commit here: >>> >>> >>> https://github.com/ghc/ghc/commit/521b792553bacbdb0eec138b150ab0626ea6f36b >>> >>> The problem is that our "cas" routine in SMP.h is similar to the C >>> compiler intrinsic __sync_val_compare_and_swap, in that it returns the old >>> value. But it seems we cannot use a comparison against that old value to >>> determine whether or not the CAS succeeded. (I believe the CAS may fail >>> due to contention, but the old value may happen to look like our old value.) >>> >>> Unfortunately, this didn't occur to me until it started causing >>> bugs [1] [2]. Fixing casMutVar# fixes these bugs. However, the way I'm >>> currently fixing CAS in the "atomic-primops" package is by using >>> __sync_bool_compare_and_swap: >>> >>> >>> https://github.com/rrnewton/haskell-lockfree/commit/f9716ddd94d5eff7420256de... >>> >>> What is the best fix for GHC itself? Would it be ok for GHC to >>> include a C compiler intrinsic like __sync_val_compare_and_swap? Otherwise >>> we need another big ifdbef'd function like "cas" in SMP.h that has the >>> architecture-specific inline asm across all architectures. I can write the >>> x86 one, but I'm not eager to try the others. >>> >>> Best, >>> -Ryan >>> >>> [1] https://github.com/iu-parfunc/lvars/issues/70 >>> [2] https://github.com/rrnewton/haskell-lockfree/issues/15 >>> >>> >> >> _______________________________________________ >> ghc-devs mailing list >> ghc-devs@haskell.org >> http://www.haskell.org/mailman/listinfo/ghc-devs >> >> >