[PATCH] numa: Slab handle memoryless nodes Against: 2.6.33-mmotm-100302-1838 Example usage of generic "numa_mem_id()": The mainline slab code, since ~ 2.6.19, does not handle memoryless nodes well. Specifically, the "fast path"--____cache_alloc()--will never succeed as slab doesn't cache offnode objects on the per cpu queues, and for memoryless nodes, all memory will be "off node" relative to numa_node_id(). This adds significant overhead to all kmem cache allocations, incurring a significant regression relative to earlier kernels [from before slab.c was reorganized]. This patch uses the generic topology function "numa_mem_id()" to return the "effective local memory node" for the calling context. This is the first node in the local node's generic fallback zonelist-- i.e., the same node that "local" mempolicy-based allocations would use. This lets slab cache these "local" allocations and avoid fallback/refill on every allocation. N.B.: Slab will need to handle node and memory hotplug events that could change the value returned by numa_mem_id() for any given node. E.g., flush all per cpu slab queues before rebuilding the zonelists. Andi Kleen and David Rientjes are currently working on patch series to improve slab support for memory hotplug. When that effort settles down, and if there is general agreement on this approach, I'll prepare another patch to address possible change in "local memory node", if still necessary. Performance impact on "hackbench 400 process 200" 2.6.33+mmotm-100302-1838 no-patch this-patch [series] ia64 no memoryless nodes [avg of 10]: 11.853 11.739 (secs) ia64 cpus all on memless nodes [10]: 264.909 27.938 ~10x speedup The slowdown of the patched kernel from ~12 sec to ~28 seconds when configured with memoryless nodes is the result of all cpus allocating from a single node's mm pagepool. The cache lines of the single node are distributed/interleaved over the memory of the real physical nodes, but the zone locks of the single node with memory still each live in a single cache line that is accessed from all processors. x86_64 [8x6 AMD] [avg of 10]: 3.322 3.148 Signed-off-by: Lee Schermerhorn <lee.schermerhorn@xxxxxx> mm/slab.c | 27 ++++++++++++++------------- 1 file changed, 14 insertions(+), 13 deletions(-) Index: linux-2.6.33-mmotm-100302-1838/mm/slab.c =================================================================== --- linux-2.6.33-mmotm-100302-1838.orig/mm/slab.c +++ linux-2.6.33-mmotm-100302-1838/mm/slab.c @@ -1073,7 +1073,7 @@ static inline int cache_free_alien(struc struct array_cache *alien = NULL; int node; - node = numa_node_id(); + node = numa_mem_id(); /* * Make sure we are not freeing a object from another node to the array @@ -1418,7 +1418,7 @@ void __init kmem_cache_init(void) * 6) Resize the head arrays of the kmalloc caches to their final sizes. */ - node = numa_node_id(); + node = numa_mem_id(); /* 1) create the cache_cache */ INIT_LIST_HEAD(&cache_chain); @@ -2052,7 +2052,7 @@ static int __init_refok setup_cpu_cache( } } } - cachep->nodelists[numa_node_id()]->next_reap = + cachep->nodelists[numa_mem_id()]->next_reap = jiffies + REAPTIMEOUT_LIST3 + ((unsigned long)cachep) % REAPTIMEOUT_LIST3; @@ -2383,7 +2383,7 @@ static void check_spinlock_acquired(stru { #ifdef CONFIG_SMP check_irq_off(); - assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); + assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); #endif } @@ -2410,7 +2410,7 @@ static void do_drain(void *arg) { struct kmem_cache *cachep = arg; struct array_cache *ac; - int node = numa_node_id(); + int node = numa_mem_id(); check_irq_off(); ac = cpu_cache_get(cachep); @@ -2943,7 +2943,7 @@ static void *cache_alloc_refill(struct k retry: check_irq_off(); - node = numa_node_id(); + node = numa_mem_id(); ac = cpu_cache_get(cachep); batchcount = ac->batchcount; if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { @@ -3147,7 +3147,7 @@ static void *alternate_node_alloc(struct if (in_interrupt() || (flags & __GFP_THISNODE)) return NULL; - nid_alloc = nid_here = numa_node_id(); + nid_alloc = nid_here = numa_mem_id(); if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) nid_alloc = cpuset_mem_spread_node(); else if (current->mempolicy) @@ -3316,6 +3316,7 @@ __cache_alloc_node(struct kmem_cache *ca { unsigned long save_flags; void *ptr; + int slab_node = numa_mem_id(); flags &= gfp_allowed_mask; @@ -3328,7 +3329,7 @@ __cache_alloc_node(struct kmem_cache *ca local_irq_save(save_flags); if (nodeid == -1) - nodeid = numa_node_id(); + nodeid = slab_node; if (unlikely(!cachep->nodelists[nodeid])) { /* Node not bootstrapped yet */ @@ -3336,7 +3337,7 @@ __cache_alloc_node(struct kmem_cache *ca goto out; } - if (nodeid == numa_node_id()) { + if (nodeid == slab_node) { /* * Use the locally cached objects if possible. * However ____cache_alloc does not allow fallback @@ -3380,8 +3381,8 @@ __do_cache_alloc(struct kmem_cache *cach * We may just have run out of memory on the local node. * ____cache_alloc_node() knows how to locate memory on other nodes */ - if (!objp) - objp = ____cache_alloc_node(cache, flags, numa_node_id()); + if (!objp) + objp = ____cache_alloc_node(cache, flags, numa_mem_id()); out: return objp; @@ -3478,7 +3479,7 @@ static void cache_flusharray(struct kmem { int batchcount; struct kmem_list3 *l3; - int node = numa_node_id(); + int node = numa_mem_id(); batchcount = ac->batchcount; #if DEBUG @@ -4053,7 +4054,7 @@ static void cache_reap(struct work_struc { struct kmem_cache *searchp; struct kmem_list3 *l3; - int node = numa_node_id(); + int node = numa_mem_id(); struct delayed_work *work = to_delayed_work(w); if (!mutex_trylock(&cache_chain_mutex)) -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html