[PATCH] NUMA slab locking fixes: fix cpu down and up locking
authorRavikiran G Thirumalai <kiran@scalex86.org>
Sun, 5 Feb 2006 07:27:59 +0000 (23:27 -0800)
committerLinus Torvalds <torvalds@g5.osdl.org>
Sun, 5 Feb 2006 19:06:53 +0000 (11:06 -0800)
This fixes locking and bugs in cpu_down and cpu_up paths of the NUMA slab
allocator.  Sonny Rao <sonny@burdell.org> reported problems sometime back on
POWER5 boxes, when the last cpu on the nodes were being offlined.  We could
not reproduce the same on x86_64 because the cpumask (node_to_cpumask) was not
being updated on cpu down.  Since that issue is now fixed, we can reproduce
Sonny's problems on x86_64 NUMA, and here is the fix.

The problem earlier was on CPU_DOWN, if it was the last cpu on the node to go
down, the array_caches (shared, alien) and the kmem_list3 of the node were
being freed (kfree) with the kmem_list3 lock held.  If the l3 or the
array_caches were to come from the same cache being cleared, we hit on
badness.

This patch cleans up the locking in cpu_up and cpu_down path.  We cannot
really free l3 on cpu down because, there is no node offlining yet and even
though a cpu is not yet up, node local memory can be allocated for it.  So l3s
are usually allocated at keme_cache_create and destroyed at
kmem_cache_destroy.  Hence, we don't need cachep->spinlock protection to get
to the cachep->nodelist[nodeid] either.

Patch survived onlining and offlining on a 4 core 2 node Tyan box with a 4
dbench process running all the time.

Signed-off-by: Alok N Kataria <alokk@calsoftinc.com>
Signed-off-by: Ravikiran Thirumalai <kiran@scalex86.org>
Cc: Christoph Lameter <christoph@lameter.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/slab.c

index d3f68543f9f4e3fcb403192357d8d18056b75b03..9cc049a942c6b6dcfcdc7dacd5211c66f799b120 100644 (file)
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -884,14 +884,14 @@ static void __drain_alien_cache(struct kmem_cache *cachep,
        }
 }
 
-static void drain_alien_cache(struct kmem_cache *cachep, struct kmem_list3 *l3)
+static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien)
 {
        int i = 0;
        struct array_cache *ac;
        unsigned long flags;
 
        for_each_online_node(i) {
-               ac = l3->alien[i];
+               ac = alien[i];
                if (ac) {
                        spin_lock_irqsave(&ac->lock, flags);
                        __drain_alien_cache(cachep, ac, i);
@@ -901,8 +901,11 @@ static void drain_alien_cache(struct kmem_cache *cachep, struct kmem_list3 *l3)
 }
 #else
 #define alloc_alien_cache(node, limit) do { } while (0)
-#define free_alien_cache(ac_ptr) do { } while (0)
-#define drain_alien_cache(cachep, l3) do { } while (0)
+#define drain_alien_cache(cachep, alien) do { } while (0)
+
+static inline void free_alien_cache(struct array_cache **ac_ptr)
+{
+}
 #endif
 
 static int __devinit cpuup_callback(struct notifier_block *nfb,
@@ -936,6 +939,11 @@ static int __devinit cpuup_callback(struct notifier_block *nfb,
                                l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
                                    ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
 
+                               /*
+                                * The l3s don't come and go as CPUs come and
+                                * go.  cache_chain_mutex is sufficient
+                                * protection here.
+                                */
                                cachep->nodelists[node] = l3;
                        }
 
@@ -950,26 +958,47 @@ static int __devinit cpuup_callback(struct notifier_block *nfb,
                   & array cache's */
                list_for_each_entry(cachep, &cache_chain, next) {
                        struct array_cache *nc;
+                       struct array_cache *shared;
+                       struct array_cache **alien;
 
                        nc = alloc_arraycache(node, cachep->limit,
-                                             cachep->batchcount);
+                                               cachep->batchcount);
                        if (!nc)
                                goto bad;
+                       shared = alloc_arraycache(node,
+                                       cachep->shared * cachep->batchcount,
+                                       0xbaadf00d);
+                       if (!shared)
+                               goto bad;
+#ifdef CONFIG_NUMA
+                       alien = alloc_alien_cache(node, cachep->limit);
+                       if (!alien)
+                               goto bad;
+#endif
                        cachep->array[cpu] = nc;
 
                        l3 = cachep->nodelists[node];
                        BUG_ON(!l3);
-                       if (!l3->shared) {
-                               if (!(nc = alloc_arraycache(node,
-                                                           cachep->shared *
-                                                           cachep->batchcount,
-                                                           0xbaadf00d)))
-                                       goto bad;
 
-                               /* we are serialised from CPU_DEAD or
-                                  CPU_UP_CANCELLED by the cpucontrol lock */
-                               l3->shared = nc;
+                       spin_lock_irq(&l3->list_lock);
+                       if (!l3->shared) {
+                               /*
+                                * We are serialised from CPU_DEAD or
+                                * CPU_UP_CANCELLED by the cpucontrol lock
+                                */
+                               l3->shared = shared;
+                               shared = NULL;
                        }
+#ifdef CONFIG_NUMA
+                       if (!l3->alien) {
+                               l3->alien = alien;
+                               alien = NULL;
+                       }
+#endif
+                       spin_unlock_irq(&l3->list_lock);
+
+                       kfree(shared);
+                       free_alien_cache(alien);
                }
                mutex_unlock(&cache_chain_mutex);
                break;
@@ -978,23 +1007,32 @@ static int __devinit cpuup_callback(struct notifier_block *nfb,
                break;
 #ifdef CONFIG_HOTPLUG_CPU
        case CPU_DEAD:
+               /*
+                * Even if all the cpus of a node are down, we don't free the
+                * kmem_list3 of any cache. This to avoid a race between
+                * cpu_down, and a kmalloc allocation from another cpu for
+                * memory from the node of the cpu going down.  The list3
+                * structure is usually allocated from kmem_cache_create() and
+                * gets destroyed at kmem_cache_destroy().
+                */
                /* fall thru */
        case CPU_UP_CANCELED:
                mutex_lock(&cache_chain_mutex);
 
                list_for_each_entry(cachep, &cache_chain, next) {
                        struct array_cache *nc;
+                       struct array_cache *shared;
+                       struct array_cache **alien;
                        cpumask_t mask;
 
                        mask = node_to_cpumask(node);
-                       spin_lock(&cachep->spinlock);
                        /* cpu is dead; no one can alloc from it. */
                        nc = cachep->array[cpu];
                        cachep->array[cpu] = NULL;
                        l3 = cachep->nodelists[node];
 
                        if (!l3)
-                               goto unlock_cache;
+                               goto free_array_cache;
 
                        spin_lock_irq(&l3->list_lock);
 
@@ -1005,33 +1043,43 @@ static int __devinit cpuup_callback(struct notifier_block *nfb,
 
                        if (!cpus_empty(mask)) {
                                spin_unlock_irq(&l3->list_lock);
-                               goto unlock_cache;
+                               goto free_array_cache;
                        }
 
-                       if (l3->shared) {
+                       shared = l3->shared;
+                       if (shared) {
                                free_block(cachep, l3->shared->entry,
                                           l3->shared->avail, node);
-                               kfree(l3->shared);
                                l3->shared = NULL;
                        }
-                       if (l3->alien) {
-                               drain_alien_cache(cachep, l3);
-                               free_alien_cache(l3->alien);
-                               l3->alien = NULL;
-                       }
 
-                       /* free slabs belonging to this node */
-                       if (__node_shrink(cachep, node)) {
-                               cachep->nodelists[node] = NULL;
-                               spin_unlock_irq(&l3->list_lock);
-                               kfree(l3);
-                       } else {
-                               spin_unlock_irq(&l3->list_lock);
+                       alien = l3->alien;
+                       l3->alien = NULL;
+
+                       spin_unlock_irq(&l3->list_lock);
+
+                       kfree(shared);
+                       if (alien) {
+                               drain_alien_cache(cachep, alien);
+                               free_alien_cache(alien);
                        }
-                     unlock_cache:
-                       spin_unlock(&cachep->spinlock);
+free_array_cache:
                        kfree(nc);
                }
+               /*
+                * In the previous loop, all the objects were freed to
+                * the respective cache's slabs,  now we can go ahead and
+                * shrink each nodelist to its limit.
+                */
+               list_for_each_entry(cachep, &cache_chain, next) {
+                       l3 = cachep->nodelists[node];
+                       if (!l3)
+                               continue;
+                       spin_lock_irq(&l3->list_lock);
+                       /* free slabs belonging to this node */
+                       __node_shrink(cachep, node);
+                       spin_unlock_irq(&l3->list_lock);
+               }
                mutex_unlock(&cache_chain_mutex);
                break;
 #endif
@@ -2011,7 +2059,6 @@ static void drain_cpu_caches(struct kmem_cache *cachep)
 
        smp_call_function_all_cpus(do_drain, cachep);
        check_irq_on();
-       spin_lock(&cachep->spinlock);
        for_each_online_node(node) {
                l3 = cachep->nodelists[node];
                if (l3) {
@@ -2019,10 +2066,9 @@ static void drain_cpu_caches(struct kmem_cache *cachep)
                        drain_array_locked(cachep, l3->shared, 1, node);
                        spin_unlock_irq(&l3->list_lock);
                        if (l3->alien)
-                               drain_alien_cache(cachep, l3);
+                               drain_alien_cache(cachep, l3->alien);
                }
        }
-       spin_unlock(&cachep->spinlock);
 }
 
 static int __node_shrink(struct kmem_cache *cachep, int node)
@@ -3440,7 +3486,7 @@ static void cache_reap(void *unused)
 
                l3 = searchp->nodelists[numa_node_id()];
                if (l3->alien)
-                       drain_alien_cache(searchp, l3);
+                       drain_alien_cache(searchp, l3->alien);
                spin_lock_irq(&l3->list_lock);
 
                drain_array_locked(searchp, cpu_cache_get(searchp), 0,
@@ -3598,7 +3644,8 @@ static int s_show(struct seq_file *m, void *p)
                        num_slabs++;
                }
                free_objects += l3->free_objects;
-               shared_avail += l3->shared->avail;
+               if (l3->shared)
+                       shared_avail += l3->shared->avail;
 
                spin_unlock_irq(&l3->list_lock);
        }