[patch] Lockup on wait on exited child with exited child

[Jan Kratochvil]

Hi Roland,

you were right, updated.


Thanks,
Jan


On Sun, 04 Nov 2007 00:55:14 +0100, Roland McGrath wrote:
> I think that drops it early (while still waiting for a report from the
> parent) in case of a normal parent that just happens to be almost finished
> exiting.  Unless I'm missing something, this extra droptcb should only be
> required in the case of a zombie group leader that has already reported
> and been short-circuited at the top of droptcb.  So I think it want the
> same test used at the top of detach.
> 
> 
> Thanks,
> Roland
-------------- next part --------------
2007-11-06  Jan Kratochvil  <jan.kratochvil at redhat.com>

	* strace.c [LINUX] (droptcb): DROPTCB even PARENT if PARENT is
	TCB_EXITING zombie group leader.
	* test/childthread.c: New file.
	* test/.cvsignore, test/Makefile: Add it.
	Code advisory: Roland McGrath
	Fixes RH#354261.

--- strace.c	3 Nov 2007 23:34:11 -0000	1.84
+++ strace.c	6 Nov 2007 20:59:48 -0000
@@ -1297,6 +1297,17 @@ struct tcb *tcp;
 		if (!(tcp->flags & TCB_CLONE_DETACHED))
 #endif
 			tcp->parent->nzombies++;
+#ifdef LINUX
+		/* Update `tcp->parent->parent->nchildren' and the other fields
+		   like NCLONE_DETACHED, only for zombie group leader that has
+		   already reported and been short-circuited at the top of this
+		   function.  The same condition as at the top of DETACH.  */
+		if ((tcp->flags & TCB_CLONE_THREAD) &&
+		    tcp->parent->nclone_threads == 0 &&
+		    (tcp->parent->flags & TCB_EXITING)) {
+			droptcb(tcp->parent);
+		}
+#endif
 		tcp->parent = NULL;
 	}
 
--- test/.cvsignore	5 Jul 2007 18:43:18 -0000	1.3
+++ test/.cvsignore	6 Nov 2007 20:59:49 -0000
@@ -4,3 +4,4 @@ skodic
 vfork
 clone
 leaderkill
+childthread
--- test/Makefile	5 Jul 2007 18:43:18 -0000	1.5
+++ test/Makefile	6 Nov 2007 20:59:49 -0000
@@ -2,10 +2,12 @@
 # $Id: Makefile,v 1.5 2007/07/05 18:43:18 roland Exp $
 #
 
-all: vfork fork sig skodic clone leaderkill
+all: vfork fork sig skodic clone leaderkill childthread
 
 leaderkill: LDFLAGS += -pthread
 
+childthread: LDFLAGS += -pthread
+
 clean distclean:
 	rm -f clone vfork fork sig leaderkill *.o core
 
--- /dev/null	1 Jan 1970 00:00:00 -0000
+++ test/childthread.c	6 Nov 2007 20:59:49 -0000
@@ -0,0 +1,59 @@
+/* Test exit of a child of a TCB_EXITING child where the toplevel process starts
+ * waiting on it.  The middle one gets detached and strace must update the
+ * toplevel process'es number of attached children to 0.
+ * 
+ * gcc -o test/childthread test/childthread.c -Wall -ggdb2 -pthread;./strace -f ./test/childthread
+ * It must print: write(1, "OK\n", ...
+ */
+
+#include <pthread.h>
+#include <assert.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <sys/wait.h>
+
+static void *start0 (void *arg)
+{
+  pause ();
+  /* NOTREACHED */
+  assert (0);
+}
+
+int main (void)
+{
+  pthread_t thread0;
+  int i;
+  pid_t child, got_pid;
+  int status;
+
+  child = fork ();
+  switch (child)
+    {
+    case -1:
+      assert (0);
+    case 0:
+      i = pthread_create (&thread0, NULL, start0, NULL);
+      assert (i == 0);
+      /* The thread must be initialized, it becomes thread-child of this
+	 process-child (child of a child of the toplevel process).  */
+      sleep (1);
+      /* Here the child TCB cannot be deallocated as there still exist
+       * children (the thread child in START0).  */
+      exit (42);
+      /* NOTREACHED */
+      assert (0);
+    default:
+      /* We must not be waiting in WAITPID when the child double-exits.  */
+      sleep (2);
+      /* PID must be -1.  */
+      got_pid = waitpid (-1, &status, 0);
+      assert (got_pid == child);
+      assert (WIFEXITED (status));
+      assert (WEXITSTATUS (status) == 42);
+      puts ("OK");
+      exit (0);
+    }
+  /* NOTREACHED */
+  assert (0);
+}