Re: Something is very wrong with disk caching in 7.0

Martin Nilsson wrote:
> Julian Elischer wrote:
> 

try this patch

Index: sys/vm/vm_pageout.c
===================================================================
RCS file: /home/ncvs/src/sys/vm/vm_pageout.c,v
retrieving revision 1.268
diff -u -r1.268 vm_pageout.c
--- sys/vm/vm_pageout.c	7 Jan 2005 02:29:27 -0000	1.268
+++ sys/vm/vm_pageout.c	30 Jul 2005 03:12:37 -0000
_at__at_ -5,6 +5,8 _at__at_
  * All rights reserved.
  * Copyright (c) 1994 David Greenman
  * All rights reserved.
+ * Copyright (c) 2005 Yahoo! Technologies Norway AS
+ * All rights reserved.
  *
  * This code is derived from software contributed to Berkeley by
  * The Mach Operating System project at Carnegie-Mellon University.
_at__at_ -210,6 +212,16 _at__at_
 static void vm_pageout_page_stats(void);
 
 /*
+ * Experimental VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK option, which will cause
+ * pagedaemon to fall back to blocking locking of vm objects if nonblocking
+ * lock attempt failed.  Lock order violation is avoided by unlocking 
+ * the page queues before locking the object.  marker pages are used
+ * to detect changes and allow for continued page queue traversal even
+ * when changes had occurred.
+ */
+#define VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+
+/*
  * vm_pageout_clean:
  *
  * Clean the page and remove it from the laundry.
_at__at_ -750,8 +762,37 _at__at_
 		 * queue, most likely are being paged out.
 		 */
 		if (!VM_OBJECT_TRYLOCK(object)) {
+#ifdef VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+			/*
+			 * Cannot lock object while holding page queue lock.
+			 * Depend on both struct vm_object and normal
+			 * struct vm_page being type stable and sanity
+			 * check after reobtaining page queue lock.
+			 */
+			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl,
+					   m, &marker, pageq);
+			vm_page_unlock_queues();
+			VM_OBJECT_LOCK(object);
+			vm_page_lock_queues();
+			/* Page queue might have changed. */
+			next = TAILQ_NEXT(&marker, pageq);
+			if (m->queue != PQ_INACTIVE ||
+			    m->object != object ||
+			    m->hold_count != 0 ||
+			    &marker != TAILQ_NEXT(m, pageq)) {
+				/* Page changed. */
+				VM_OBJECT_UNLOCK(object);
+ 				TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
+					     &marker, pageq);
+				addl_page_shortage++;
+				continue;
+			}
+			TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl,
+				     &marker, pageq);
+#else
 			addl_page_shortage++;
 			continue;
+#endif
 		}
 		if (m->busy || (m->flags & PG_BUSY)) {
 			VM_OBJECT_UNLOCK(object);
_at__at_ -1024,10 +1065,44 _at__at_
 
 		next = TAILQ_NEXT(m, pageq);
 		object = m->object;
+#ifdef VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+		if ((m->flags & PG_MARKER) != 0) {
+			m = next;
+			continue;
+		}
+#endif
 		if (!VM_OBJECT_TRYLOCK(object)) {
+#ifdef VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+			/*
+			 * Cannot lock object while holding page queue lock.
+			 * Depend on both struct vm_object and normal
+			 * struct vm_page being type stable and sanity
+			 * check after reobtaining page queue lock.
+			 */
+			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE].pl,
+					   m, &marker, pageq);
+			vm_page_unlock_queues();
+			VM_OBJECT_LOCK(object);
+			vm_page_lock_queues();
+			/* Page queue might have changed. */
+			next = TAILQ_NEXT(&marker, pageq);
+			if (m->queue != PQ_ACTIVE ||
+			    m->object != object ||
+			    &marker != TAILQ_NEXT(m, pageq)) {
+				/* Page changed. */
+				VM_OBJECT_UNLOCK(object);
+ 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl,
+					     &marker, pageq);
+				m = next;
+				continue;
+			}
+			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl,
+				     &marker, pageq);
+#else
 			vm_pageq_requeue(m);
 			m = next;
 			continue;
+#endif
 		}
 
 		/*
_at__at_ -1264,6 +1339,9 _at__at_
 vm_pageout_page_stats()
 {
 	vm_object_t object;
+#ifdef VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+	struct vm_page marker;
+#endif
 	vm_page_t m,next;
 	int pcount,tpcount;		/* Number of pages to check */
 	static int fullintervalcount = 0;
_at__at_ -1287,6 +1365,16 _at__at_
 		fullintervalcount = 0;
 	}
 
+#ifdef VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+	/*
+	 * Initialize our marker
+	 */
+	bzero(&marker, sizeof(marker));
+	marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER;
+	marker.queue = PQ_INACTIVE;
+	marker.wire_count = 1;
+#endif
+
 	m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl);
 	while ((m != NULL) && (pcount-- > 0)) {
 		int actcount;
_at__at_ -1296,10 +1384,45 _at__at_
 
 		next = TAILQ_NEXT(m, pageq);
 		object = m->object;
+
+#ifdef VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+		if ((m->flags & PG_MARKER) != 0) {
+			m = next;
+			continue;
+		}
+#endif
 		if (!VM_OBJECT_TRYLOCK(object)) {
+#ifdef VM_PAGEOUT_FORCE_BLOCKING_OBJLOCK
+			/*
+			 * Cannot lock object while holding page queue lock.
+			 * Depend on both struct vm_object and normal
+			 * struct vm_page being type stable and sanity
+			 * check after reobtaining page queue lock.
+			 */
+			TAILQ_INSERT_AFTER(&vm_page_queues[PQ_ACTIVE].pl,
+					   m, &marker, pageq);
+			vm_page_unlock_queues();
+			VM_OBJECT_LOCK(object);
+			vm_page_lock_queues();
+			/* Page queue might have changed. */
+			next = TAILQ_NEXT(&marker, pageq);
+			if (m->queue != PQ_ACTIVE ||
+			    m->object != object ||
+			    &marker != TAILQ_NEXT(m, pageq)) {
+				/* Page changed. */
+				VM_OBJECT_UNLOCK(object);
+ 				TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl,
+					     &marker, pageq);
+				m = next;
+				continue;
+			}
+			TAILQ_REMOVE(&vm_page_queues[PQ_ACTIVE].pl,
+				     &marker, pageq);
+#else
 			vm_pageq_requeue(m);
 			m = next;
 			continue;
+#endif
 		}
 
 		/*