[Cluster-devel] cluster/gfs-kernel/src/dlm lock_dlm.h plock.c

[teigland at sourceware.org]

CVSROOT:	/cvs/cluster
Module name:	cluster
Branch: 	RHEL4
Changes by:	teigland at sourceware.org	2007-12-03 16:40:41

Modified files:
	gfs-kernel/src/dlm: lock_dlm.h plock.c 

Log message:
	Posix locks don't work between threads, but it seems some programs do
	involve threads sharing plocks incidentally (with meaningless results).
	Given the plock state we keep in lock_dlm, multiple threads accessing the
	locks would not only be meaningless, but could corrupt the state, leaving
	the threads permanently hung.  This patch tries to keep threads from
	stomping on each other, hopefully preventing most hangs.
	
	bz 383391

Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/dlm/lock_dlm.h.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.18.2.7&r2=1.18.2.8
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/dlm/plock.c.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.12.2.4&r2=1.12.2.5

--- cluster/gfs-kernel/src/dlm/Attic/lock_dlm.h	2007/03/22 21:23:18	1.18.2.7
+++ cluster/gfs-kernel/src/dlm/Attic/lock_dlm.h	2007/12/03 16:40:41	1.18.2.8
@@ -152,6 +152,7 @@
 	uint64_t                end;
 	int                     count;
 	int                     ex;
+	int			busy;
 };
 
 #define LFL_NOBLOCK             0
--- cluster/gfs-kernel/src/dlm/Attic/plock.c	2007/03/22 21:23:18	1.12.2.4
+++ cluster/gfs-kernel/src/dlm/Attic/plock.c	2007/12/03 16:40:41	1.12.2.5
@@ -488,8 +488,10 @@
 
 	if (found) {
 		DLM_ASSERT(po->lp, );
+		po->busy = 1;
 		error = wait_async(po->lp);
 		DLM_ASSERT(!error, );
+		po->busy = 0;
 		goto restart;
 	}
 }
@@ -1154,8 +1156,15 @@
 	int found = FALSE;
 
 	list_for_each_entry(po, &r->locks, list) {
-		if (po->owner == owner)
-			continue;
+		if (po->owner == owner) {
+			if (!po->busy)
+				continue;
+			log_debug("po busy %llx %llu %u",
+				  (unsigned long long)name->ln_number,
+				  (unsigned long long)owner, po->pid);
+			found = TRUE;
+			break;
+		}
 		if (!ranges_overlap(po->start, po->end, *start, *end))
 			continue;