Re: [Libunwind-devel] Re: [patch 2/2] Allow caller to block signals.

[Paul Pluzhnikov]

On Fri, Sep 25, 2009 at 12:00 PM, Mark Rabkin <address@hidden> wrote:

> I read this:
> http://lists.nongnu.org/archive/html/libunwind-devel/2009-09/msg00000.html
>
> But I was a little unclear on the main motivation.
>
> Was the motivation that the existing signal blocking mechanism didn't
> prevent re-entry?  If it didn't prevent it, did it cause crashes or aborted
> stack traces or what was the symptom?

Existing mechanism did prevent re-entry, but at a cost of lots of
sigprocmask()s. We have one particular (emulated) environment, where these
are quite costly.

> If you're now planning to block additional signals in your own client code
> ("at the top level"),

We don't actually need to block them, we just need to make sure that the
same thread doesn't reenter libunwind from a signal handler....

> are you removing the calls inside libunwind to ensure
> it doesn't unblock them (as it does SETMASK),

The remaining sigprocmask calls are following this pattern:

  sigprocmask (SIG_SETMASK, &unwi_full_mask, &saved_mask);
  ret = dl_iterate_phdr (check_callback, as);
  sigprocmask (SIG_SETMASK, &saved_mask, NULL);

so (AFAICT) they wouldn't unblock anything that was blocked before.

But I see now that I should have eliminated these remaining sigprocmasks
as well. Patch attached.

> or are you just removing them
> for performance since they're unnecessary as your initial message implies?

Yes.

> Finally, I just wanted to get a clarification on the effects of linking
> libunwind -- it was my impression that linking it may also affect the stack
> unwinding done when, for instance, a C++ exception is thrown.

Not by default. In default configuration on x86_64, I get:

nm -D src/.libs/libunwind.so | grep ' T ' | grep -v ' _U'
0000000000001380 T backtrace

So only calls to (glibc) backtrace would be affected.

Cheers,
-- 
Paul Pluzhnikov

libunwind-sigprocmask-20090925-2.txt
Description: Text document