Re: ATA identify retries exceeded - kernel panic

[Matthew Dillon <dillon@xxxxxxxxxxxxxxxxxxxx>]

    Ok, how about this.  If the simple form of objcache_create is used
    we install simple_objsize and allow M_ZERO.  Otherwise we panic if
    M_ZERO is specified.

    I let the second layer case fall through... if the allocator happens
    to be malloc, then M_ZERO will work.  If it doesn't, then the problem
    will be caught later on when the first layer has cached objects and
    the caller tries to get a new object.  Not perfect but it should catch
    any problems.

						-Matt

Index: kern_objcache.c
===================================================================
RCS file: /cvs/src/sys/kern/kern_objcache.c,v
retrieving revision 1.12
diff -u -r1.12 kern_objcache.c
--- kern_objcache.c	20 Nov 2006 22:19:54 -0000	1.12
+++ kern_objcache.c	29 Nov 2006 23:33:00 -0000
@@ -129,6 +129,7 @@
 	objcache_alloc_fn	*alloc;
 	objcache_free_fn	*free;
 	void			*allocator_args;
+	size_t			simple_objsize;
 
 	SLIST_ENTRY(objcache)	oc_next;
 
@@ -228,6 +229,12 @@
 			     null_ctor, null_dtor, NULL,
 			     objcache_malloc_alloc, objcache_malloc_free,
 			     margs);
+
+	/*
+	 * This indicates that we are a simple objcache and allows
+	 * objcache_get() calls with M_ZERO.
+	 */
+	oc->simple_objsize = objsize;
 	return (oc);
 }
 
@@ -264,7 +271,14 @@
 	loadedmag = cpucache->loaded_magazine;
 	if (MAGAZINE_NOTEMPTY(loadedmag)) {
 		obj = loadedmag->objects[--loadedmag->rounds];
+done:
 		crit_exit();
+		if (ocflags & M_ZERO) {
+			if (oc->simple_objsize)
+				bzero(obj, oc->simple_objsize);
+			else
+				panic("objcache_get(): M_ZERO illegal here");
+		}
 		return (obj);
 	}
 
@@ -275,8 +289,7 @@
 		swap(cpucache->loaded_magazine, cpucache->previous_magazine);
 		loadedmag = cpucache->loaded_magazine;
 		obj = loadedmag->objects[--loadedmag->rounds];
-		crit_exit();
-		return (obj);
+		goto done;
 	}
 
 	/*
@@ -284,6 +297,8 @@
 	 * move one of the empty ones to the depot.
 	 *
 	 * Obtain the depot spinlock.
+	 *
+	 * NOTE: Beyond this point, M_ZERO is handled via oc->alloc()
 	 */
 	depot = &oc->depot[myclusterid];
 	spin_lock_wr(&depot->spin);