Re: [Gc] Assertions when glibc uses lock elision (#51)

On Thu, 11 Sep 2014 21:11:38 +0400, Ivan Maidanski <[email protected]> wrote:
> 
> Hi Jan,
> I have merged fix-tsx-bug commit to master (as well as release-7_4 branch).
> Note that this fix (workaround) enabled only for linux/x64 at this moment. To enable it for x86, use -DFIX_GLIBC_TSX_BUG . Please test it whether it is OK for x86. If it solves the issue I'll change the code to define the macro implicitly for linux/x86.
> OTOH, this workaround has performance drawback (in locking) for linux/intel platforms which do not have TSX. Could we exclude such targets by checking TSX presence at runtime?
> Regards,
> Ivan
> -----
> Monday, 08 Sept 2014, 12:22 +04:00 from Jan Alexander Steffens <[email protected]>:
> guile: ../nptl/pthread_mutex_lock.c:80: __pthread_mutex_cond_lock: Assertion mutex->__data.__owner == 0' failed.
> gc's testsuite also randomly asserts. Strangely, it seems to happen soon after start or not at all.
> The fix_tsx_bug branch in its current state (commit  757af8a ) works on x86_64 but not i686, which continues to assert.
> Meanwhile, the glibc maintainers assert that this is not their bug.

Hi Ivan,

At least up to now it is not necessary to use PTHREAD_MUTEX_ERRORCHECK:
http://repo.or.cz/w/glibc.git/blob/HEAD:/nptl/pthread_mutexattr_settype.c

Peter
_______________________________________________
bdwgc mailing list
[email protected]
https://lists.opendylan.org/mailman/listinfo/bdwgc