[Gc] Re: External thread suspension (again)

Keith Seitz keiths at redhat.com
Mon Jun 5 13:43:01 PDT 2006


Hans Boehm wrote:

> Some comments/questions:
> 
> 1) I don't see where the SUSPENDED flag gets set in response to
> GC_suspend_thread.  It gets set in GC_suspend_self(), but I don't
> think you get there unless it has already been set?

That gets set in GC_suspend_self. When GC_stop_world is called, 
GC_suspend_all will not signal the externally suspended thread (because 
GC_start_blocking will set thread_blocked).

> 2) You should probably call GC_brief_async_signal_safe_sleep().
> Last time I looked nanosleep wasn't technically async-signal-safe,
> but I think select() always has been.

I've changed that. Thanks.

> 3) Have you thought about the various possible signal races, and which
> signals are masked at what point?  I haven't yet had a chance to.  It
> may be fine, but if there's an obvious argument that it is, a
> comment would be great.

I've thought about it a bit, and the only problem that I see is in the 
signal handler code. *If* a non-blocked signal were to make it to an 
externally suspended thread (and we ended up in our signal handler), it 
would GC_suspend_self itself again. In which case, we would need to 
differentiate between a thread that is supposed to suspend itself and a 
thread which is already suspended.

But since we only install a signal handler for SUSPEND and THR_RESTART, 
so it shouldn't be an issue. A suspended thread should not get 
THR_RESTART, since blocked_thread is set by GC_start_blocking.

Or at least, that's all I could divine.

> 4) IIRC, GC_start_blocking() has a fundamental issue in that there isn't
> really a good way to guarantee that all callee-saves registers are visible
> onthe stack at the right point.  The result might be very occasional
> failures.  So long as this is used only for debugging, and since this
> is fixed with an interface change in 7.0, which can be easily accomodated
> by your patch, I'd lean towards ignoring this for now.  The more
> ambitious, and more correct, solution would be to backport the gc7
> do_blocking() code.

Okay, I will keep that in mind.

I've attached the latest revision of this patch.

Keith
-------------- next part --------------
A non-text attachment was scrubbed...
Name: external-suspend.patch
Type: text/x-patch
Size: 3338 bytes
Desc: not available
Url : http://napali.hpl.hp.com/pipermail/gc/attachments/20060605/9ddd6aba/external-suspend.bin


More information about the Gc mailing list