|
From: Jeremy F. <je...@go...> - 2005-03-18 07:37:39
|
Julian Seward wrote:
>On reflection, I'm a little puzzled. I thought one of the aims of
>your recent threading rework was to decouple V from the precise
>behaviour of the threading libraries. So long as they use sys_clone
>in a way we can cope with, we are happy to let the threading library
>do whatever it wants, right? But the implication of this comment
>is that in fact we do have to be concerned whether we're running
>NPTL, LinuxThreads or some hybrid. Ummmmm ....
>
In general that's true, but exit and exit_group always need to be
emulated, and I'm worried about non-standard behaviour in exit_group,
particularly. Though its hard to tell, your trace seems to suggest that
it expects exit_group to kill the whole process, even though they
threads are not in the same thread group (they weren't created with
CLONE_THREAD). Valgrind, naturally, emulates standard 2.6 behaviour of
exit_group.
It may be moot because Alex Ivershen's report pointed out a real bug in
the handling of sigsuspend, which may have some bearing on your bug
(though I can't convince myself of that based on the trace). I'm about
to check in the fix for that one, so give it a go.
J
|