Teach the testsuite that GDBserver reliably reports program exits.
authorPedro Alves <palves@redhat.com>
Wed, 2 Oct 2013 11:44:20 +0000 (11:44 +0000)
committerPedro Alves <palves@redhat.com>
Wed, 2 Oct 2013 11:44:20 +0000 (11:44 +0000)
commitb477a5e649150a2ed9c5090b01989e746d65cef5
treee2f1366c7a649dbd8c47d2df3057b15f8a66c5ea
parent1a3d890bcc0e6f9fe0116b3274023ae0af847593
Teach the testsuite that GDBserver reliably reports program exits.

Running catch-syscall.exp against a gdbserver that actually supports
it, we get:

 FAIL: gdb.base/catch-syscall.exp: continue until exit (the program exited)
 FAIL: gdb.base/catch-syscall.exp: continue until exit (the program exited)
 FAIL: gdb.base/catch-syscall.exp: continue until exit (the program exited)
 FAIL: gdb.base/catch-syscall.exp: continue until exit at catch syscall with unused syscall (mlock) (the program exited)
 FAIL: gdb.base/catch-syscall.exp: continue until exit (the program exited)

The fail pattern is:

 Catchpoint 2 (call to syscall exit_group), 0x000000323d4baa29 in _exit () from /lib64/libc.so.6
 (gdb) PASS: gdb.base/catch-syscall.exp: program has called exit_group
 delete breakpoints
 Delete all breakpoints? (y or n) y
 (gdb) info breakpoints
 No breakpoints or watchpoints.
 (gdb) break exit
 Breakpoint 3 at 0x323d438bf0
 (gdb) continue
 Continuing.
 [Inferior 1 (process 21081) exited normally]

That "break exit" + "continue" comes from:

> # gdb_continue_to_end:
> # The case where the target uses stubs has to be handled specially. If a
> #       stub is used, we set a breakpoint at exit because we cannot rely on
> #       exit() behavior of a remote target.
> #

The native-gdbserver.exp board, used to test against gdbserver in
"target remote" mode, triggers that case ($use_gdb_stub is true).  So
gdb_continue_to_end doesn't work for catch-syscall.exp as here we
catch the exit_group and continue from that, expecting to see a real
program exit.  I was about to post a patch that changes
catch-syscall.exp to call a new function that just always does what
gdb_continue_to_end does in the !$use_gdb_stub case.  But, since
GDBserver doesn't really need this, in the end I thought it better to
teach the testsuite that there are stubs that know how to report
program exits, by adding a new "exit_is_reliable" board variable that
then gdb_continue_to_end checks.

Tested on x86_64 Fedora 17, native and gdbserver.

gdb/testsuite/
2013-10-02  Pedro Alves  <palves@redhat.com>

* README (Board Settings): Document "exit_is_reliable".
* lib/gdb.exp (gdb_continue_to_end): Check whether the board says
running to exit reliably reports program exits.
* boards/native-gdbserver.exp: Set exit_is_reliable in the board
info.
* boards/native-stdio-gdbserver.exp: Likewise.
gdb/testsuite/ChangeLog
gdb/testsuite/README
gdb/testsuite/boards/native-gdbserver.exp
gdb/testsuite/boards/native-stdio-gdbserver.exp
gdb/testsuite/lib/gdb.exp