slub: only IPI CPUs that have per cpu obj to flush
[linux-2.6.git] / mm / slub.c
index 4056d29..dcbb192 100644 (file)
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -366,9 +366,10 @@ static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page
                const char *n)
 {
        VM_BUG_ON(!irqs_disabled());
-#ifdef CONFIG_CMPXCHG_DOUBLE
+#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
+    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
        if (s->flags & __CMPXCHG_DOUBLE) {
-               if (cmpxchg_double(&page->freelist,
+               if (cmpxchg_double(&page->freelist, &page->counters,
                        freelist_old, counters_old,
                        freelist_new, counters_new))
                return 1;
@@ -400,9 +401,10 @@ static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
                void *freelist_new, unsigned long counters_new,
                const char *n)
 {
-#ifdef CONFIG_CMPXCHG_DOUBLE
+#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
+    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
        if (s->flags & __CMPXCHG_DOUBLE) {
-               if (cmpxchg_double(&page->freelist,
+               if (cmpxchg_double(&page->freelist, &page->counters,
                        freelist_old, counters_old,
                        freelist_new, counters_new))
                return 1;
@@ -570,7 +572,7 @@ static void slab_bug(struct kmem_cache *s, char *fmt, ...)
        va_end(args);
        printk(KERN_ERR "========================================"
                        "=====================================\n");
-       printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
+       printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
        printk(KERN_ERR "----------------------------------------"
                        "-------------------------------------\n\n");
 }
@@ -1579,6 +1581,7 @@ static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
        struct zone *zone;
        enum zone_type high_zoneidx = gfp_zone(flags);
        void *object;
+       unsigned int cpuset_mems_cookie;
 
        /*
         * The defrag ratio allows a configuration of the tradeoffs between
@@ -1602,23 +1605,32 @@ static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
                        get_cycles() % 1024 > s->remote_node_defrag_ratio)
                return NULL;
 
-       get_mems_allowed();
-       zonelist = node_zonelist(slab_node(current->mempolicy), flags);
-       for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
-               struct kmem_cache_node *n;
-
-               n = get_node(s, zone_to_nid(zone));
-
-               if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
-                               n->nr_partial > s->min_partial) {
-                       object = get_partial_node(s, n, c);
-                       if (object) {
-                               put_mems_allowed();
-                               return object;
+       do {
+               cpuset_mems_cookie = get_mems_allowed();
+               zonelist = node_zonelist(slab_node(current->mempolicy), flags);
+               for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
+                       struct kmem_cache_node *n;
+
+                       n = get_node(s, zone_to_nid(zone));
+
+                       if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
+                                       n->nr_partial > s->min_partial) {
+                               object = get_partial_node(s, n, c);
+                               if (object) {
+                                       /*
+                                        * Return the object even if
+                                        * put_mems_allowed indicated that
+                                        * the cpuset mems_allowed was
+                                        * updated in parallel. It's a
+                                        * harmless race between the alloc
+                                        * and the cpuset update.
+                                        */
+                                       put_mems_allowed(cpuset_mems_cookie);
+                                       return object;
+                               }
                        }
                }
-       }
-       put_mems_allowed();
+       } while (!put_mems_allowed(cpuset_mems_cookie));
 #endif
        return NULL;
 }
@@ -1901,11 +1913,14 @@ static void unfreeze_partials(struct kmem_cache *s)
                        }
 
                        if (l != m) {
-                               if (l == M_PARTIAL)
+                               if (l == M_PARTIAL) {
                                        remove_partial(n, page);
-                               else
+                                       stat(s, FREE_REMOVE_PARTIAL);
+                               } else {
                                        add_partial(n, page,
                                                DEACTIVATE_TO_TAIL);
+                                       stat(s, FREE_ADD_PARTIAL);
+                               }
 
                                l = m;
                        }
@@ -1978,7 +1993,7 @@ int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
                page->pobjects = pobjects;
                page->next = oldpage;
 
-       } while (irqsafe_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
+       } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
        stat(s, CPU_PARTIAL_FREE);
        return pobjects;
 }
@@ -2013,9 +2028,17 @@ static void flush_cpu_slab(void *d)
        __flush_cpu_slab(s, smp_processor_id());
 }
 
+static bool has_cpu_slab(int cpu, void *info)
+{
+       struct kmem_cache *s = info;
+       struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
+
+       return !!(c->page);
+}
+
 static void flush_all(struct kmem_cache *s)
 {
-       on_each_cpu(flush_cpu_slab, s, 1);
+       on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
 }
 
 /*
@@ -2124,6 +2147,37 @@ static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
 }
 
 /*
+ * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
+ * or deactivate the page.
+ *
+ * The page is still frozen if the return value is not NULL.
+ *
+ * If this function returns NULL then the page has been unfrozen.
+ */
+static inline void *get_freelist(struct kmem_cache *s, struct page *page)
+{
+       struct page new;
+       unsigned long counters;
+       void *freelist;
+
+       do {
+               freelist = page->freelist;
+               counters = page->counters;
+               new.counters = counters;
+               VM_BUG_ON(!new.frozen);
+
+               new.inuse = page->objects;
+               new.frozen = freelist != NULL;
+
+       } while (!cmpxchg_double_slab(s, page,
+               freelist, counters,
+               NULL, new.counters,
+               "get_freelist"));
+
+       return freelist;
+}
+
+/*
  * Slow path. The lockless freelist is empty or we need to perform
  * debugging duties.
  *
@@ -2144,8 +2198,6 @@ static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
 {
        void **object;
        unsigned long flags;
-       struct page new;
-       unsigned long counters;
 
        local_irq_save(flags);
 #ifdef CONFIG_PREEMPT
@@ -2166,31 +2218,14 @@ redo:
                goto new_slab;
        }
 
-       stat(s, ALLOC_SLOWPATH);
-
-       do {
-               object = c->page->freelist;
-               counters = c->page->counters;
-               new.counters = counters;
-               VM_BUG_ON(!new.frozen);
-
-               /*
-                * If there is no object left then we use this loop to
-                * deactivate the slab which is simple since no objects
-                * are left in the slab and therefore we do not need to
-                * put the page back onto the partial list.
-                *
-                * If there are objects left then we retrieve them
-                * and use them to refill the per cpu queue.
-                */
+       /* must check again c->freelist in case of cpu migration or IRQ */
+       object = c->freelist;
+       if (object)
+               goto load_freelist;
 
-               new.inuse = c->page->objects;
-               new.frozen = object != NULL;
+       stat(s, ALLOC_SLOWPATH);
 
-       } while (!__cmpxchg_double_slab(s, c->page,
-                       object, counters,
-                       NULL, new.counters,
-                       "__slab_alloc"));
+       object = get_freelist(s, c->page);
 
        if (!object) {
                c->page = NULL;
@@ -2304,7 +2339,7 @@ redo:
                 * Since this is without lock semantics the protection is only against
                 * code executing on this cpu *not* from access by other cpus.
                 */
-               if (unlikely(!irqsafe_cpu_cmpxchg_double(
+               if (unlikely(!this_cpu_cmpxchg_double(
                                s->cpu_slab->freelist, s->cpu_slab->tid,
                                object, tid,
                                get_freepointer_safe(s, object), next_tid(tid)))) {
@@ -2534,7 +2569,7 @@ redo:
        if (likely(page == c->page)) {
                set_freepointer(s, object, c->freelist);
 
-               if (unlikely(!irqsafe_cpu_cmpxchg_double(
+               if (unlikely(!this_cpu_cmpxchg_double(
                                s->cpu_slab->freelist, s->cpu_slab->tid,
                                c->freelist, tid,
                                object, next_tid(tid)))) {
@@ -2999,7 +3034,8 @@ static int kmem_cache_open(struct kmem_cache *s,
                }
        }
 
-#ifdef CONFIG_CMPXCHG_DOUBLE
+#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
+    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
        if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
                /* Enable fast mode */
                s->flags |= __CMPXCHG_DOUBLE;
@@ -3656,6 +3692,9 @@ void __init kmem_cache_init(void)
        struct kmem_cache *temp_kmem_cache_node;
        unsigned long kmalloc_size;
 
+       if (debug_guardpage_minorder())
+               slub_max_order = 0;
+
        kmem_size = offsetof(struct kmem_cache, node) +
                                nr_node_ids * sizeof(struct kmem_cache_node *);
 
@@ -4636,6 +4675,8 @@ static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
        err = strict_strtoul(buf, 10, &objects);
        if (err)
                return err;
+       if (objects && kmem_cache_debug(s))
+               return -EINVAL;
 
        s->cpu_partial = objects;
        flush_all(s);