*sigpause hanging on 8.x+

[Garrett Cooper]

On Sun, Jul 11, 2010 at 2:08 PM, Kostik Belousov <kostikbel at gmail.com> wrote:
> On Sun, Jul 11, 2010 at 12:39:39PM -0700, Garrett Cooper wrote:
>> So, long story short... I've basically ported the open posix testsuite
>> to FreeBSD, and one of the tests tests out sigpause. Unfortunately the
>> sucker hangs on my dev box at home.
>>
>> I've written a short testcase that demonstrates this. It prints out:
>>
>> $ ~/test_sigpause
>> 0
>>
>> And proceeds to be unresponsive to signals (except SIGSTOP / SIGKILL,
>> as expected).
>>
>> When I monkey around with libc's compat4.3 stuff a bit, this is what comes up:
>>
>> $ env LD_LIBRARY_PATH=$PWD:/usr/src/lib/libc/../libthr ~/test_sigpause
>> 0
>> before sigemptyset
>> before _sigsuspend
>>
>> So it's getting stuck after calling _sigsuspend.
>>
>> I tried the same thing on a i386 8-STABLE VM and it hangs as well.
>>
>> I tried applying similar printfs in libthr but it's not hitting that
>> code at all (it's now responding to SIGTERM though, which is
>> interesting, but not too interesting to me).
>>
>> I also wrote similar code that exercised the functionality in
>> sigsuspend, by calling sigprocmask beforehand, and it works.
>>
>> Thoughts?
>>
>> -Garrett
>>
>> Dev machine:
>> FreeBSD bayonetta.local 9.0-CURRENT FreeBSD 9.0-CURRENT #1
>> r206173:209901M: Sun Jul 11 04:18:42 PDT 2010
>> root@:/usr/obj/usr/src/sys/BAYONETTA  amd64
>> VM:
>> FreeBSD starr-bastion.localdomain 8.0-STABLE FreeBSD 8.0-STABLE #0
>> r207913: Tue May 11 06:21:57 UTC 2010
>> root at starr-bastion.localdomain:/usr/obj/usr/src/sys/GENERIC  i386
>>
>> Index: compat-43/sigcompat.c
>> ===================================================================
>> --- compat-43/sigcompat.c     (revision 206173)
>> +++ compat-43/sigcompat.c     (working copy)
>> @@ -36,6 +36,7 @@
>>  #include "namespace.h"
>>  #include <sys/param.h>
>>  #include <signal.h>
>> +#include <stdio.h>
>>  #include <string.h>
>>  #include "un-namespace.h"
>>  #include "libc_private.h"
>> @@ -102,7 +103,9 @@
>>  {
>>       sigset_t set;
>>
>> +     printf("before sigemptyset\n");
>>       sigemptyset(&set);
>> +     printf("before _sigsuspend\n");
>>       set.__bits[0] = mask;
>>       return (_sigsuspend(&set));
>>  }
>> @@ -111,10 +114,16 @@
>>  xsi_sigpause(int sig)
>>  {
>>       sigset_t set;
>> +     int rc;
>>
>> +     printf("before sigemptyset\n");
>>       sigemptyset(&set);
>> +     printf("before sigaddset\n");
>>       sigaddset(&set, sig);
>> -     return (_sigsuspend(&set));
>> +     printf("before _sigsuspend\n");
>> +     rc = (_sigsuspend(&set));
>> +     printf("after _sigsuspend\n");
>> +     return rc;
>>  }
>>
>>  int
>>
>> $ cat ~/test_sigpause.c
>> #include <signal.h>
>> #include <stdio.h>
>>
>> int
>> main (void)
>> {
>>         printf("0\n");
>>         fflush(stdout);
>>         (void) sigpause(1);
>>         return 0;
>> }
>> $ cat ~/test_sigsuspend.c
>> #include <err.h>
>> #include <signal.h>
>>
>> int
>> main (void)
>> {
>>         sigset_t oset;
>>         sigset_t nset;
>>         if (sigprocmask(1, &nset, &oset) == -1)
>>                 err(1, "sigprocmask(-1, &nset, &oset)");
>>         if (sigprocmask(-1, &nset, &oset) == -1)
>>                 err(1, "sigprocmask(-1, &nset, &oset)");
>>         return (sigsuspend(&nset));
>> }
>
> It seems I got a sigmask for sigpause inside the xsi_sigpause() backward.
> On the other hand, I do not understand what is your issue with sigpause().

The negative testcase from the open posix testsuite was setup so that
setting sigpause(-1) would return -1 with EINVAL, according to the
sig* manpages (-1 is an invalid signal of course). That isn't being
triggered with either function today.

0 seems a bit wonky too (it's an invalid signal number).

My bet is that values greater than SIGRTMAX aren't interpreted properly either.

> diff --git a/lib/libc/compat-43/sigcompat.c b/lib/libc/compat-43/sigcompat.c
> index c3ba30a..bab9d5c 100644
> --- a/lib/libc/compat-43/sigcompat.c
> +++ b/lib/libc/compat-43/sigcompat.c
> @@ -111,9 +111,12 @@ int
>  xsi_sigpause(int sig)
>  {
>        sigset_t set;
> +       int error;
>
> -       sigemptyset(&set);
> -       sigaddset(&set, sig);
> +       error = _sigprocmask(SIG_BLOCK, NULL, &set);
> +       if (error != 0)
> +               return (error);
> +       sigdelset(&set, sig);
>        return (_sigsuspend(&set));
>  }

Doesn't this violate the restore clause noted in the manpage?

     The xsi_sigpause() function removes sig from the signal mask of the call-
     ing process and suspend the calling process until a signal is received.
     The xsi_sigpause() function restores the signal mask of the process to
     its original state before returning.

So if I had a sigset defined above with sig, then redefined it, I
would be whacking the previous handler by passing in NULL to
_sigprocmask, correct? If so, sigpause has issues too in its
implementation.

There's also some interesting SIGDELSET action going on in libthr's
copy of _sigsuspend's with SIGCANCEL (apparently that's the unofficial
alias for SIGRTMIN as defined by libthr), but that's a sidenote for
the actual issue seen here.

Thanks!
-Garrett