manoj 99/02/17 00:13:16
Modified: . STATUS
pthreads/src/main http_main.c
Log:
Start to remove pthread calls from clean_child_exit to make it legal for
a signal handler. There is still a pthread_cancel loop, which seems to
work sometimes for killing hanging threads, but not all the time. The
real solution is to fix the hanging threads, which seem to be caused by
the (accidental?) usage of a cross-process alloc_mutex.
The last commit message was messed up. That patch didn't undo any
pthread calls during signal handling, but it is still necessary because
it prevents the thread from trying to join itself. I need someone to
show me how to change stupid commit messages.
Revision Changes Path
1.9 +3 -1 apache-apr/STATUS
Index: STATUS
===================================================================
RCS file: /home/cvs/apache-apr/STATUS,v
retrieving revision 1.8
retrieving revision 1.9
diff -u -u -r1.8 -r1.9
--- STATUS 1999/02/17 01:08:18 1.8
+++ STATUS 1999/02/17 08:13:13 1.9
@@ -1,5 +1,5 @@
Apache Portable Runtime STATUS:
-Last modified at [$Date: 1999/02/17 01:08:18 $]
+Last modified at [$Date: 1999/02/17 08:13:13 $]
Release:
@@ -40,6 +40,8 @@
Everything
Needs patch:
+ * A patch to make the alloc_mutex's scope per-process. It seems to be
+ causing trouble on Linux at least.
Open issues:
1.38 +17 -3 apache-apr/pthreads/src/main/http_main.c
Index: http_main.c
===================================================================
RCS file: /home/cvs/apache-apr/pthreads/src/main/http_main.c,v
retrieving revision 1.37
retrieving revision 1.38
diff -u -u -r1.37 -r1.38
--- http_main.c 1999/02/17 05:32:00 1.37
+++ http_main.c 1999/02/17 08:13:15 1.38
@@ -397,14 +397,29 @@
listen_rec *lr;
parent_score *ss = &ap_scoreboard_image->parent[child_num];
+ /* XXX - This should be taken care of by the parent process. When it
reaps
+ * the child, it will also update the scoreboard. We can't do this here,
+ * because clean_child_exit can be called from a signal handler, and
+ * ap_update_child_status calls pthread_self */
+#if 0
for (i = 0; i < ap_threads_per_child + ap_acceptors_per_child; i++)
ap_update_child_status(child_num, i, SERVER_DEAD, (request_rec *)
NULL);
+#endif
if (pchild) {
ap_child_exit_modules(pchild, server_conf);
ap_destroy_pool(pchild);
}
+ /* XXX - This is not allowed in a signal handler either, and it shouldn't
+ * be needed. But it's still in here, because without it, we get hanging
+ * threads on linux. The symptom isn't *too* bad on Linux: a core dump in
+ * the exit() library call.
+ *
+ * The root cause seems to be that alloc's pthread mutex is being used
+ * across processes, which is not kosher on some platforms (i.e. Linux),
+ * and a thread whacking a pool will hit an acquire_mutex and hang */
+
for (i = 0; i < ss->worker_threads + ss->acceptor_threads; i++) {
if (ap_scoreboard_image->servers[child_num][i].status !=
SERVER_DEAD) {
pthread_cancel(ap_scoreboard_image->servers[child_num][i].tid);
@@ -2811,9 +2826,8 @@
if (pid >= 0) {
child_slot = find_child_by_pid(0);
if (child_slot >= 0) {
- /* (void) ap_update_child_status(child_slot,
SERVER_DEAD,
- (request_rec *) NULL);
- LOOK INTO THIS */
+ for (i = 0; i < ap_threads_per_child +
ap_acceptors_per_child; i++)
+ ap_update_child_status(child_slot, i, SERVER_DEAD,
(request_rec *) NULL);
if (remaining_children_to_start
&& child_slot < ap_daemons_limit) {
