Re: [Libunwind-devel] Eliminating the sigprocmask system call

[David Mosberger-Tang]

On 6/4/08, Arun Sharma <address@hidden> wrote:
> When we were analyzing the performance of stack unwinding (especially
>  compared to a much simpler unwinder based on frame pointers), one
>  thing that stood out is the cost of the sigprocmask(2) system call
>  which is invoked from the getcontext() in libc. It appears to me that
>  this is necessary only if the caller wants to unw_resume() at some
>  point.
>
>  However, for callers who just want to examine the stack, this could
>  potentially be eliminated. Currently there is no way for the caller to
>  express the intent at unw_getcontext() time. Looking at some of the
>  other archs which have a hand coded getcontext.S (ia64, arm etc) - I
>  don't see any calls to sigprocmask. Does this mean that
>  Ltest-resume-sig is failing on those archs?

I don't know about ARM, but I have never seen Ltest-resume-sig fail on
ia64 and I don't think it should.

>  If we must restore signals state on unw_resume(), I was thinking
>  something along the lines of:
>
>  /* For users who want to unw_resume */
>  unw_getcontext(&uc);
>
>  /* For users who want to just examine the stack */
>  unw_get_mcontext(&uc);
>
>  Comments?

I don't particularly like making the API more complicated (and
error-prone to use) for such reasons.  Maybe not using the libc
getcontext() would be better?

  --david