This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [musl] Re: [PATCH] Fix invalid sigprocmask call


On Fri, Mar 24, 2017 at 09:05:15PM +0800, Yousong Zhou wrote:
> On 24 March 2017 at 20:55, Pedro Alves <palves@redhat.com> wrote:
> > On 03/24/2017 12:23 PM, Yousong Zhou wrote:
> >> On 24 March 2017 at 18:47, Pedro Alves <palves@redhat.com> wrote:
> >>> On 03/24/2017 03:01 AM, Yousong Zhou wrote:
> >>>> The POSIX document says
> >>>>
> >>>>     The pthread_sigmask() and sigprocmask() functions shall fail if:
> >>>>
> >>>>     [EINVAL]
> >>>>     The value of the how argument is not equal to one of the defined values.
> >>>>
> >>>> and this is how musl-libc is currently doing.  Fix the call to be safe
> >>>> and correct
> >>>>
> >>>>  [1] http://pubs.opengroup.org/onlinepubs/9699919799/functions/pthread_sigmask.html
> >>>>
> >>>
> >>> I don't agree.  It's a musl bug.  Please fix it / file a musl bug.
> >>
> >> I already did that before sending to gdb-patches
> >>
> >>   http://www.openwall.com/lists/musl/2017/03/24/1
> >>
> >> I am aware of the fact that the current code works with glibc and mac
> >> osx 10.11.6.  The Linux kernel code at the moment also accepts the
> >> call with how==0
> >
> > Cool.
> >
> >>
> >> But this is more about interpretation of POSIX document itself.  And
> >> it says, clearly without pre-condition words or ambiguity in the
> >> ERRORS section of that page, to return EINVAL if how is not equal to
> >> one of the defined values.
> >
> > The standard wasn't built on a vacuum.  It starts by ratifying common
> > implementation behavior.  If no historical implementation behaves like what
> > you're suggesting, what's the point of enforcing that, when it's clearly
> > NOT the intent?  You're just causing porting pain for no good reason.
> > Please file a bug against the standard to have the error section clarified instead.
> 
> Lol, now I will consider the idea of bumping the door of POSIX committee ;)

If you file a report and it's deemed a bug in the standard and
changed, I'm happy to change this on the musl side. Just keep me
posted on what happens. I don't have any preference on what the
behavior "should" be here (IMO imposing a behavior when the caller has
violated constraints like passed a wrong value for how is pointless
anyway) but I do want to conform to the standard.

> >> I also tried to find some posix-compliant testsuite and to search the
> >> github code for samples of pthread_sigmask call.  The first I came
> >> across was the following code snippet at link
> >> https://github.com/juj/posixtestsuite/blob/master/conformance/interfaces/pthread_sigmask/8-1.c#L57
> >>
> >>         pthread_sigmask(SIG_BLOCK, NULL, &oactl);
> >
> > The fact that that call includes SIG_BLOCK doesn't say whether
> > passing 0 should be rejected.
> >
> > So I cloned that repo, and did a quick grep.  And lo:
> >
> > https://github.com/juj/posixtestsuite/blob/26372421f53aeeeeeb4b23561c417886f1930ef6/conformance/interfaces/fork/12-1.c#L187
> >
> >                 /* Examine the current blocked signal set. USR1 & USR2 shall be present */
> >                 ret = sigprocmask( 0, NULL, &mask );
> >
> >                 if ( ret != 0 )
> >                 {
> >                         UNRESOLVED( errno, "Sigprocmask failed in child" );
> >                 }
> >
> > Thanks,
> > Pedro Alves
> >
> 
> Okay, then another fact that the posixtestsuite project also expects
> to accept how==0
> 
> I am cc-ing musl-libc list now.

If you're talking about "Open POSIX Test Suite", which the above link
seems to point to a fork of, the majority of its tests are invalid,
invoking undefined behavior then asserting that the error/effect they
wrongly expect happens. Without a thorough audit of its test validity
it's less than worthless.

Rich


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]