[PATCH] fix ohci1394 rmmod

[PATCH] fix ohci1394 rmmod

[Paul Mackerras]

With the current ieee1394 code, if I try to rmmod the ohci1394 module,
it hangs forever.  The reason is that it is waiting for the knodemgr_n
kernel thread to exit.  The thread exits when it gets a signal, and
the mainline code has tried to send it a signal.  However, the thread
has called daemonize(), which sets the signal mask to all 1s, so no
signals ever get through.

The patch below fixes this by unblocking SIGTERM after the daemonize
call.  I looked around and found the same problem with the khpsbpkt
thread, so I applied the same fix there.  With this patch I can once
again successfully rmmod the ohci1394 module.

Ben, any comments on this patch?  I think it should go in.

Regards,
Paul.

diff -urN linux-2.5/drivers/ieee1394/ieee1394_core.c g5-ppc64/drivers/ieee1394/ieee1394_core.c
--- linux-2.5/drivers/ieee1394/ieee1394_core.c	2004-04-19 08:07:02.000000000 +1000
+++ g5-ppc64/drivers/ieee1394/ieee1394_core.c	2004-05-03 22:53:03.000000000 +1000
@@ -1017,9 +1017,15 @@
 	struct hpsb_packet *packet;
 	void (*complete_routine)(void*);
 	void *complete_data;
+	sigset_t sset;
 
 	daemonize("khpsbpkt");
 
+	/* Make us susceptible to SIGTERM */
+	sigemptyset(&sset);
+	sigaddset(&sset, SIGTERM);
+	sigprocmask(SIG_UNBLOCK, &sset, NULL);
+
 	while (!down_interruptible(&khpsbpkt_sig)) {
 		while ((skb = skb_dequeue(&hpsbpkt_queue)) != NULL) {
 			packet = (struct hpsb_packet *)skb->data;
diff -urN linux-2.5/drivers/ieee1394/nodemgr.c g5-ppc64/drivers/ieee1394/nodemgr.c
--- linux-2.5/drivers/ieee1394/nodemgr.c	2004-04-19 08:07:02.000000000 +1000
+++ g5-ppc64/drivers/ieee1394/nodemgr.c	2004-05-03 22:51:32.000000000 +1000
@@ -1464,10 +1464,16 @@
 	struct host_info *hi = (struct host_info *)__hi;
 	struct hpsb_host *host = hi->host;
 	int reset_cycles = 0;
+	sigset_t sset;
 
 	/* No userlevel access needed */
 	daemonize(hi->daemon_name);
 
+	/* Make us susceptible to SIGTERM */
+	sigemptyset(&sset);
+	sigaddset(&sset, SIGTERM);
+	sigprocmask(SIG_UNBLOCK, &sset, NULL);
+
 	/* Setup our device-model entries */
 	nodemgr_create_host_dev_files(host);
 

Re: [PATCH] fix ohci1394 rmmod

[Ben Collins]

> Ben, any comments on this patch?  I think it should go in.

Yeah, this is the wrong fix. I have a proper one in the repo that I'll
sync to Linus this week.

Point is, we need to ignore SIGTERM because init will send it during
shutdown, and that will kill these threads making it impossible to
unmount/sync sbp2 disks during the shutdown/reset process. My patch sets
a variable and then signals the semaphore in the thread, which checks
the variable and shuts down at that time.

-- 
Debian     - http://www.debian.org/
Linux 1394 - http://www.linux1394.org/
Subversion - http://subversion.tigris.org/
WatchGuard - http://www.watchguard.com/