[PATCH] lockdep: annotate SLAB code
authorIngo Molnar <mingo@elte.hu>
Mon, 3 Jul 2006 07:25:28 +0000 (00:25 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Mon, 3 Jul 2006 22:27:09 +0000 (15:27 -0700)
Teach special (recursive) locking code to the lock validator.  Has no effect
on non-lockdep kernels.

Fix initialize-locks-via-memcpy assumptions.

Effects on non-lockdep kernels: the subclass nesting parameter is passed into
cache_free_alien() and __cache_free(), and turns one internal
kmem_cache_free() call into an open-coded __cache_free() call.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Cc: Pekka Enberg <penberg@cs.helsinki.fi>
Cc: Christoph Lameter <clameter@engr.sgi.com>
Cc: Manfred Spraul <manfred@colorfullife.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/slab.c

index 3936af3..85c2e03 100644 (file)
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -1021,7 +1021,8 @@ static void drain_alien_cache(struct kmem_cache *cachep,
        }
 }
 
-static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
+static inline int cache_free_alien(struct kmem_cache *cachep, void *objp,
+                                  int nesting)
 {
        struct slab *slabp = virt_to_slab(objp);
        int nodeid = slabp->nodeid;
@@ -1039,7 +1040,7 @@ static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
        STATS_INC_NODEFREES(cachep);
        if (l3->alien && l3->alien[nodeid]) {
                alien = l3->alien[nodeid];
-               spin_lock(&alien->lock);
+               spin_lock_nested(&alien->lock, nesting);
                if (unlikely(alien->avail == alien->limit)) {
                        STATS_INC_ACOVERFLOW(cachep);
                        __drain_alien_cache(cachep, alien, nodeid);
@@ -1068,7 +1069,8 @@ static inline void free_alien_cache(struct array_cache **ac_ptr)
 {
 }
 
-static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
+static inline int cache_free_alien(struct kmem_cache *cachep, void *objp,
+                                  int nesting)
 {
        return 0;
 }
@@ -1272,6 +1274,11 @@ static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
 
        local_irq_disable();
        memcpy(ptr, list, sizeof(struct kmem_list3));
+       /*
+        * Do not assume that spinlocks can be initialized via memcpy:
+        */
+       spin_lock_init(&ptr->list_lock);
+
        MAKE_ALL_LISTS(cachep, ptr, nodeid);
        cachep->nodelists[nodeid] = ptr;
        local_irq_enable();
@@ -1398,7 +1405,7 @@ void __init kmem_cache_init(void)
        }
        /* 4) Replace the bootstrap head arrays */
        {
-               void *ptr;
+               struct array_cache *ptr;
 
                ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
 
@@ -1406,6 +1413,11 @@ void __init kmem_cache_init(void)
                BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
                memcpy(ptr, cpu_cache_get(&cache_cache),
                       sizeof(struct arraycache_init));
+               /*
+                * Do not assume that spinlocks can be initialized via memcpy:
+                */
+               spin_lock_init(&ptr->lock);
+
                cache_cache.array[smp_processor_id()] = ptr;
                local_irq_enable();
 
@@ -1416,6 +1428,11 @@ void __init kmem_cache_init(void)
                       != &initarray_generic.cache);
                memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
                       sizeof(struct arraycache_init));
+               /*
+                * Do not assume that spinlocks can be initialized via memcpy:
+                */
+               spin_lock_init(&ptr->lock);
+
                malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
                    ptr;
                local_irq_enable();
@@ -1743,6 +1760,8 @@ static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
 }
 #endif
 
+static void __cache_free(struct kmem_cache *cachep, void *objp, int nesting);
+
 /**
  * slab_destroy - destroy and release all objects in a slab
  * @cachep: cache pointer being destroyed
@@ -1766,8 +1785,17 @@ static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
                call_rcu(&slab_rcu->head, kmem_rcu_free);
        } else {
                kmem_freepages(cachep, addr);
-               if (OFF_SLAB(cachep))
-                       kmem_cache_free(cachep->slabp_cache, slabp);
+               if (OFF_SLAB(cachep)) {
+                       unsigned long flags;
+
+                       /*
+                        * lockdep: we may nest inside an already held
+                        * ac->lock, so pass in a nesting flag:
+                        */
+                       local_irq_save(flags);
+                       __cache_free(cachep->slabp_cache, slabp, 1);
+                       local_irq_restore(flags);
+               }
        }
 }
 
@@ -3072,7 +3100,16 @@ static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
                if (slabp->inuse == 0) {
                        if (l3->free_objects > l3->free_limit) {
                                l3->free_objects -= cachep->num;
+                               /*
+                                * It is safe to drop the lock. The slab is
+                                * no longer linked to the cache. cachep
+                                * cannot disappear - we are using it and
+                                * all destruction of caches must be
+                                * serialized properly by the user.
+                                */
+                               spin_unlock(&l3->list_lock);
                                slab_destroy(cachep, slabp);
+                               spin_lock(&l3->list_lock);
                        } else {
                                list_add(&slabp->list, &l3->slabs_free);
                        }
@@ -3098,7 +3135,7 @@ static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
 #endif
        check_irq_off();
        l3 = cachep->nodelists[node];
-       spin_lock(&l3->list_lock);
+       spin_lock_nested(&l3->list_lock, SINGLE_DEPTH_NESTING);
        if (l3->shared) {
                struct array_cache *shared_array = l3->shared;
                int max = shared_array->limit - shared_array->avail;
@@ -3141,14 +3178,14 @@ free_done:
  * Release an obj back to its cache. If the obj has a constructed state, it must
  * be in this state _before_ it is released.  Called with disabled ints.
  */
-static inline void __cache_free(struct kmem_cache *cachep, void *objp)
+static void __cache_free(struct kmem_cache *cachep, void *objp, int nesting)
 {
        struct array_cache *ac = cpu_cache_get(cachep);
 
        check_irq_off();
        objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
 
-       if (cache_free_alien(cachep, objp))
+       if (cache_free_alien(cachep, objp, nesting))
                return;
 
        if (likely(ac->avail < ac->limit)) {
@@ -3387,7 +3424,7 @@ void kmem_cache_free(struct kmem_cache *cachep, void *objp)
        BUG_ON(virt_to_cache(objp) != cachep);
 
        local_irq_save(flags);
-       __cache_free(cachep, objp);
+       __cache_free(cachep, objp, 0);
        local_irq_restore(flags);
 }
 EXPORT_SYMBOL(kmem_cache_free);
@@ -3412,7 +3449,7 @@ void kfree(const void *objp)
        kfree_debugcheck(objp);
        c = virt_to_cache(objp);
        debug_check_no_locks_freed(objp, obj_size(c));
-       __cache_free(c, (void *)objp);
+       __cache_free(c, (void *)objp, 0);
        local_irq_restore(flags);
 }
 EXPORT_SYMBOL(kfree);