[PATCH RFA] lin-lwp.c: Block SIGCHLD events when attaching

Kevin Buettner kevinb@cygnus.com
Wed Nov 7 15:07:00 GMT 2001


When doing a lin_lwp_attach_lwp(), it is sometimes possible to receive
a SIGCHLD signal thus causing the waitpid() call to fail with EINTR. 
This in turn causes the second assert() in lin_lwp_attach_lwp() to
fail.

Reproducing this problem can be somewhat difficult.  I've only been
able to reproduce it on a dual processor Linux/x86 machine.  I did
manage to reproduce it using the linux-dp program as follows:

    1) Start linux-dp (again, running on an SMP machine).

    2) Determine the process id.  Let's call it PID.  (Make appropriate
       substitution below.)

    3) Invoke gdb:

	    gdb linux-dp

    4) Do:

	    set height 0
	    break print_philosopher

    5) Do:
	    while 1
	     attach PID
	     continue
	     detach
	    end
    6) Go get a cup of coffee, take a nap, etc.  When you come back, you
       may be fortunate enough to see the following:

	    ...
	    Breakpoint 1, print_philosopher (n=2, left=95 '_', right=95 '_')
		at ../../../src/gdb/testsuite/gdb.threads/linux-dp.c:105
	    105       shared_printf ("%*s%c %d %c\n", (n * 4) + 2, "", left, n, right);
	    ../../src/gdb/lin-lwp.c:416: gdb-internal-error: lin_lwp_attach: Assertion `pid == GET_PID (inferior_ptid) && WIFSTOPPED (status) && WSTOPSIG (status) == SIGSTOP' failed.
	    An internal GDB error was detected.  This may make further
	    debugging unreliable.  Continue this debugging session? (y or n) 

Jim Blandy deserves credit for arriving at the above procedure for
reproducing this bug.  (The actual test case that we used for tracking
down this problem is different and is able to demonstrate it in a very
short period of time.  Having *lots* of running threads greatly
increases the chances of being able to reproduce it quickly.)

The fix is below.  Okay to commit?

	    * lin-lwp.c (lin_lwp_attach_lwp): Make sure SIGCHLD is in set of
	    blocked signals.

Index: lin-lwp.c
===================================================================
RCS file: /cvs/src/src/gdb/lin-lwp.c,v
retrieving revision 1.30
diff -u -p -r1.30 lin-lwp.c
--- lin-lwp.c	2001/10/14 11:30:37	1.30
+++ lin-lwp.c	2001/11/19 19:08:10
@@ -352,6 +352,14 @@ lin_lwp_attach_lwp (ptid_t ptid, int ver
 
   gdb_assert (is_lwp (ptid));
 
+  /* Make sure SIGCHLD is blocked.  We don't want SIGCHLD events
+     to interrupt either the ptrace() or waitpid() calls below.  */
+  if (! sigismember (&blocked_mask, SIGCHLD))
+    {
+      sigaddset (&blocked_mask, SIGCHLD);
+      sigprocmask (SIG_BLOCK, &blocked_mask, NULL);
+    }
+
   if (verbose)
     printf_filtered ("[New %s]\n", target_pid_to_str (ptid));
 



More information about the Gdb-patches mailing list