Hi Andrew,

> * Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE> [2019-09-05 14:04:06 +0200]:
>
>> Quite a number of tests FAIL on Solaris due to a mismatch between
>> expected and received messages: the testsuite expects something like
>> 
>> 	Program received signal SIGABRT, Aborted.
>> 
>> while on Solaris it gets
>> 
>> 	Thread 2 received signal SIGABRT, Aborted.
>> 
>> For a simple testcase, info threads shows
>> 
>> (gdb) info threads 
>>   Id   Target Id         Frame 
>>   1    LWP    1          main () at /vol/src/gnu/gdb/doc/bugs/ua.c:5
>> * 2    Thread 1 (LWP 1)  main () at /vol/src/gnu/gdb/doc/bugs/ua.c:5
>> 
>> I suspect this is due to support for the old pre-Solaris 9 MxN thread
>> model where user level threads were mapped to a different set of lwps.
>> 
>> For the moment, I'm dealing with this by allowing both forms of the
>> message in the testsuite.  The patch is almost completely mechanical,
>> with the exception of gdb.base/sigbpt.exp where the introduction of a
>> new group in the RE required adjustments in the $expect_out indices.
>
> I'm a little nervous about just allowing either "Thread" or "Program"
> for all tests for all targets.  Maybe others will disagree and think
> I'm worrying about nothing, but I wonder if we could be more
> conservative by adding a support function into lib/gdb.exp that takes
> the name of a signal and returns the string we expect from GDB, which
> we can then change based on Solaris/non-Solaris.
>
> I haven't looked through the patch in enough detail to know if there's
> any reason why this wouldn't work, so please push back if you think
> the idea is unworkable.

sorry for letting the ball drop on this one.  Only recently did I
stumble across it again when looking into a related issue and now I
finally understand why Solaris is different here.

[Thread starting at https://sourceware.org/ml/gdb-patches/2019-09/msg00050.html]

* Consider the following testcase:

$ cat selfkill.c 
#include <sys/types.h>
#include <signal.h>
#include <unistd.h>
#include <pthread.h>

void *
selfkill (void *arg)
{
  kill (getpid (), SIGINT);
  return NULL;
}

int
main (void)
{
#ifdef _REENTRANT
  pthread_t tid;
  pthread_create (&tid, NULL, selfkill, NULL);
  pthread_join (tid, NULL);
#else
  selfkill (NULL);
#endif
  return 0;
}

* Now compile on Solaris 9, both without and with -pthread:

$ gcc -o selfkill selfkill.c
$ gcc -pthread -o selfkill-mt selfkill.c

* Run the identical binaries and versions of gdb (7.11 here) on both
  Solaris 9 and Solaris 10:

$ gdb -q --batch -ex run selfkill{,-mt}

** Solaris 9, selfkill:

Program received signal SIGINT, Interrupt.
0xb5d54186 in _libc_kill () from /usr/lib/libc.so.1

** Solaris 9, selfkill-mt:

[Thread debugging using libthread_db enabled]
[New Thread 1 (LWP 1)]
[New LWP    2        ]
[New Thread 2 (LWP 2)]

Thread 2 received signal SIGINT, Interrupt.
[Switching to Thread 1 (LWP 1)]
0xb5c9fad5 in _lwp_wait () from /usr/lib/libc.so.1

** Solaris 10, selfkill:

[Thread debugging using libthread_db enabled]
[New Thread 1 (LWP 1)]

Thread 2 received signal SIGINT, Interrupt.
[Switching to Thread 1 (LWP 1)]
0xfef0c165 in kill () from /lib/libc.so.1

** Solaris 10, selfkill-mt:

[Thread debugging using libthread_db enabled]
[New Thread 1 (LWP 1)]
[New LWP    2        ]
[New Thread 2 (LWP 2)]

Thread 2 received signal SIGINT, Interrupt.
[Switching to Thread 1 (LWP 1)]
0xfeedca05 in __lwp_wait () from /lib/libc.so.1

** Trying the same on Linux/x86_64, one sees the same behaviour as on
   Solaris 9: non-threaded and threaded programs behave differently.

* As you can see, on Solaris 10 even the not explicitly threaded version
  of the test is shown as threaded, explaining the difference in the
  "... received signal" messages.

  This is a consequence of the Thread Model Unification Project in
  Solaris 10, which removed the difference between non-threaded and
  threaded processes.  This has nothing to do with the removal of the
  pre-Solaris 9 MxN multilevel thread model as I'd originally
  suspected.

That said, I've reworked the patch to use a support function in gdb.exp
(signal_recv_msg) to hide the difference between the Solaris and
non-Solaris cases.  One could think about introducing an additional arg
(threaded) so both the explicitly threaded and other cases could be
handled similarly.

So far lightly tested on amd64-pc-solaris2.11 and x86_64-pc-linux-gnu.

	Rainer

-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University