slab not freeing with current -git
Hi,
I just booted current -git on a POWER5 machine and quickly ran out of memory. The machine has 256MB of memory, and looking at the slab stats we are using almost all of it in unused slabs:
OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME 367860 7202 1% 0.12K 12262 30 49048K size-128 1362032 332 0% 0.03K 12161 112 48644K size-32 65714 26 0% 0.70K 5974 11 47792K blkdev_queue 47872 169 0% 0.50K 5984 8 23936K size-512 23012 181 0% 1.00K 5753 4 23012K size-1024
299 299 100% 16.00K 299 1 4784K size-16384
Backing out the "NUMA aware block device control structure allocation" patch fixes the problem. I wonder if kmalloc_node is broken.
FYI the kernel has NUMA enabled but the machine has only 1 NUMA node.
Anton
--
Author: Christoph Lameter <[email protected]>
Date: Thu, 23 Jun 2005 07:08:19 +0000 (-0700)
Source: http://www.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=1946089a109251655c5438d92c539bd2930e71ea
[PATCH] NUMA aware block device control structure allocation
Patch to allocate the control structures for for ide devices on the node of the device itself (for NUMA systems). The patch depends on the Slab API change patch by Manfred and me (in mm) and the pcidev_to_node patch that I posted today.
Does some realignment too.
drivers/block/as-iosched.c +5 -3
@@ -1871,20 +1871,22 @@ static int as_init_queue(request_queue_tif (!arq_pool)return -ENOMEM ; - ad =kmalloc (sizeof (*ad),GFP_KERNEL ); + ad =kmalloc_node (sizeof (*ad),GFP_KERNEL , q->node);if (!ad)return -ENOMEM ;memset (ad,0 ,sizeof (*ad)); ad->q = q;/* Identify what queue the data belongs to */ - ad->hash =kmalloc (sizeof (struct list_head )*AS_HASH_ENTRIES ,GFP_KERNEL ); + ad->hash =kmalloc_node (sizeof (struct list_head )*AS_HASH_ENTRIES , +GFP_KERNEL , q->node);if (!ad->hash) {kfree (ad);return -ENOMEM ; } - ad->arq_pool =mempool_create (BLKDEV_MIN_RQ , mempool_alloc_slab, mempool_free_slab, arq_pool); + ad->arq_pool =mempool_create_node (BLKDEV_MIN_RQ , mempool_alloc_slab, + mempool_free_slab, arq_pool, q->node);if (!ad->arq_pool) {kfree (ad->hash);kfree (ad);
drivers/block/deadline-iosched.c +5 -3
@@ -711,18 +711,20 @@ static int deadline_init_queue(request_qif (!drq_pool)return -ENOMEM ; - dd =kmalloc (sizeof (*dd),GFP_KERNEL ); + dd =kmalloc_node (sizeof (*dd),GFP_KERNEL , q->node);if (!dd)return -ENOMEM ;memset (dd,0 ,sizeof (*dd)); - dd->hash =kmalloc (sizeof (struct list_head )*DL_HASH_ENTRIES ,GFP_KERNEL ); + dd->hash =kmalloc_node (sizeof (struct list_head )*DL_HASH_ENTRIES , +GFP_KERNEL , q->node);if (!dd->hash) {kfree (dd);return -ENOMEM ; } - dd->drq_pool =mempool_create (BLKDEV_MIN_RQ , mempool_alloc_slab, mempool_free_slab, drq_pool); + dd->drq_pool =mempool_create_node (BLKDEV_MIN_RQ , mempool_alloc_slab, + mempool_free_slab, drq_pool, q->node);if (!dd->drq_pool) {kfree (dd->hash);kfree (dd);
drivers/block/genhd.c +10 -3
@@ -582,10 +582,16 @@ struct seq_operations diskstats_op = { .show = diskstats_show }; -struct gendisk *alloc_disk (int minors) { -struct gendisk *disk =kmalloc (sizeof (struct gendisk ),GFP_KERNEL ); +return alloc_disk_node (minors,-1 ); +} + +struct gendisk *alloc_disk_node (int minors,int node_id) +{ +struct gendisk *disk; + + disk =kmalloc_node (sizeof (struct gendisk ),GFP_KERNEL , node_id);if (disk) {memset (disk,0 ,sizeof (struct gendisk ));if (!init_disk_stats (disk)) {
@@ -594,7 +600,7 @@ struct gendisk *alloc_disk(int minors) }if (minors >1 ) {int size = (minors -1 ) *sizeof (struct hd_struct *); - disk->part =kmalloc (size,GFP_KERNEL ); + disk->part =kmalloc_node (size,GFP_KERNEL , node_id);if (!disk->part) {kfree (disk);return NULL ;
@@ -610,6 +616,7 @@ struct gendisk *alloc_disk(int minors) }EXPORT_SYMBOL (alloc_disk); +EXPORT_SYMBOL (alloc_disk_node);struct kobject *get_disk (struct gendisk *disk) {
drivers/block/ll_rw_blk.c +23 -7
@@ -28,6 +28,7 @@#include <linux/slab.h> #include <linux/swap.h> #include <linux/writeback.h> +#include <linux/blkdev.h> /* * for max sense size
@@ -1645,7 +1646,8 @@ static int blk_init_free_list(request_quinit_waitqueue_head (&rl->wait[WRITE ]);init_waitqueue_head (&rl->drain); - rl->rq_pool =mempool_create (BLKDEV_MIN_RQ , mempool_alloc_slab, mempool_free_slab, request_cachep); + rl->rq_pool =mempool_create_node (BLKDEV_MIN_RQ , mempool_alloc_slab, + mempool_free_slab, request_cachep, q->node);if (!rl->rq_pool)return -ENOMEM ;
@@ -1657,8 +1659,15 @@ static int __make_request(request_queue_request_queue_t *blk_alloc_queue (int gfp_mask) { -request_queue_t *q =kmem_cache_alloc (requestq_cachep, gfp_mask); +return blk_alloc_queue_node (gfp_mask,-1 ); +} +EXPORT_SYMBOL (blk_alloc_queue); +request_queue_t *blk_alloc_queue_node (int gfp_mask,int node_id) +{ +request_queue_t *q; + + q =kmem_cache_alloc_node (requestq_cachep, gfp_mask, node_id);if (!q)return NULL ;
@@ -1671,8 +1680,7 @@ request_queue_t *blk_alloc_queue(int gfpreturn q; } - -EXPORT_SYMBOL (blk_alloc_queue); +EXPORT_SYMBOL (blk_alloc_queue_node);/** * blk_init_queue - prepare a request queue for use with a block device
@@ -1705,13 +1713,22 @@ EXPORT_SYMBOL(blk_alloc_queue);* blk_init_queue() must be paired with a blk_cleanup_queue() call * when the block device is deactivated (such as at module unload). **/ +request_queue_t *blk_init_queue (request_fn_proc *rfn,spinlock_t *lock) { -request_queue_t *q =blk_alloc_queue (GFP_KERNEL ); +return blk_init_queue_node (rfn, lock,-1 ); +} +EXPORT_SYMBOL (blk_init_queue); + +request_queue_t * +blk_init_queue_node (request_fn_proc *rfn,spinlock_t *lock,int node_id) +{ +request_queue_t *q =blk_alloc_queue_node (GFP_KERNEL , node_id);if (!q)return NULL ; + q->node = node_id;if (blk_init_free_list (q)) goto out_init;
@@ -1754,8 +1771,7 @@ out_init:kmem_cache_free (requestq_cachep, q);return NULL ; } - -EXPORT_SYMBOL (blk_init_queue); +EXPORT_SYMBOL (blk_init_queue_node);int blk_get_queue (request_queue_t *q) {
drivers/ide/ide-disk.c +2 -1
@@ -1215,7 +1215,8 @@ static int ide_disk_probe(struct deviceif (!idkp) goto failed; - g =alloc_disk (1 <<PARTN_BITS ); + g =alloc_disk_node (1 <<PARTN_BITS , +pcibus_to_node (drive->hwif->pci_dev->bus));if (!g) goto out_free_idkp;
drivers/ide/ide-probe.c +5 -3
@@ -977,8 +977,9 @@ static int ide_init_queue(ide_drive_t *d* limits and LBA48 we could raise it but as yet * do not. */ - - q =blk_init_queue (do_ide_request, &ide_lock); + + q =blk_init_queue_node (do_ide_request, &ide_lock, +pcibus_to_node (drive->hwif->pci_dev->bus));if (!q)return 1 ;
@@ -1095,7 +1096,8 @@ static int init_irq (ide_hwif_t *hwif) hwgroup->hwif->next = hwif;spin_unlock_irq (&ide_lock); }else { - hwgroup =kmalloc (sizeof (ide_hwgroup_t),GFP_KERNEL ); + hwgroup =kmalloc_node (sizeof (ide_hwgroup_t),GFP_KERNEL , +pcibus_to_node (hwif->drives[0 ].hwif->pci_dev->bus));if (!hwgroup) goto out_up;
include/linux/blkdev.h +5 -1
@@ -396,6 +396,7 @@ struct request_queue*/ unsigned int sg_timeout;unsigned int sg_reserved_size; +int node;struct list_head drain_list;
@@ -615,6 +616,8 @@ static inline void blkdev_dequeue_reques/* * Access functions for manipulating queue properties */ +extern request_queue_t *blk_init_queue_node (request_fn_proc *rfn, +spinlock_t *lock,int node_id);extern request_queue_t *blk_init_queue (request_fn_proc *,spinlock_t *);extern void blk_cleanup_queue (request_queue_t *);extern void blk_queue_make_request (request_queue_t *,make_request_fn *);
@@ -646,7 +649,8 @@ extern void blk_wait_queue_drained(requeextern void blk_finish_queue_drain (request_queue_t *);int blk_get_queue (request_queue_t *); -request_queue_t *blk_alloc_queue (int ); +request_queue_t *blk_alloc_queue (int gfp_mask); +request_queue_t *blk_alloc_queue_node (int ,int );#define blk_put_queue (q) blk_cleanup_queue((q))/*
include/linux/genhd.h +1 -0
@@ -403,6 +403,7 @@ extern int rescan_partitions(struct gendextern void add_partition (struct gendisk *,int ,sector_t ,sector_t );extern void delete_partition (struct gendisk *,int ); +extern struct gendisk *alloc_disk_node (int minors,int node_id);extern struct gendisk *alloc_disk (int minors);extern struct kobject *get_disk (struct gendisk *disk);extern void put_disk (struct gendisk *disk);
include/linux/ide.h +1 -1
@@ -917,7 +917,7 @@ typedef struct hwif_s {unsigned dma;void (*led_act)(void *data,int rw); -} ide_hwif_t; +}____cacheline_maxaligned_in_smp ide_hwif_t;/* * internal ide interrupt handler type
include/linux/mempool.h +8 -3
@@ -20,9 +20,14 @@ typedef struct mempool_s {mempool_free_t *free;wait_queue_head_t wait; } mempool_t; -extern mempool_t *mempool_create (int min_nr,mempool_alloc_t *alloc_fn, -mempool_free_t *free_fn,void *pool_data); -extern int mempool_resize (mempool_t *pool,int new_min_nr,unsigned int __nocast gfp_mask); + +extern mempool_t *mempool_create (int min_nr,mempool_alloc_t *alloc_fn, +mempool_free_t *free_fn,void *pool_data); +extern mempool_t *mempool_create_node (int min_nr,mempool_alloc_t *alloc_fn, +mempool_free_t *free_fn,void *pool_data,int nid); + +extern int mempool_resize (mempool_t *pool,int new_min_nr, +unsigned int __nocast gfp_mask);extern void mempool_destroy (mempool_t *pool);extern void *mempool_alloc (mempool_t *pool,unsigned int __nocast gfp_mask);extern void mempool_free (void *element,mempool_t *pool);
mm/mempool.c +12 -5
@@ -51,16 +51,23 @@ static void free_pool(mempool_t *pool)* functions might sleep - as long as the mempool_alloc function is not called * from IRQ contexts. */ -mempool_t *mempool_create (int min_nr,mempool_alloc_t *alloc_fn, +mempool_t *mempool_create (int min_nr,mempool_alloc_t *alloc_fn,mempool_free_t *free_fn,void *pool_data) { -mempool_t *pool; +return mempool_create_node (min_nr,alloc_fn,free_fn, pool_data,-1 ); +} +EXPORT_SYMBOL (mempool_create); - pool =kmalloc (sizeof (*pool),GFP_KERNEL ); +mempool_t *mempool_create_node (int min_nr,mempool_alloc_t *alloc_fn, +mempool_free_t *free_fn,void *pool_data,int node_id) +{ +mempool_t *pool; + pool =kmalloc_node (sizeof (*pool),GFP_KERNEL , node_id);if (!pool)return NULL ;memset (pool,0 ,sizeof (*pool)); - pool->elements =kmalloc (min_nr *sizeof (void *),GFP_KERNEL ); + pool->elements =kmalloc_node (min_nr *sizeof (void *), +GFP_KERNEL , node_id);if (!pool->elements) {kfree (pool);return NULL ;
@@ -87,7 +94,7 @@ mempool_t * mempool_create(int min_nr, m }return pool; } -EXPORT_SYMBOL (mempool_create); +EXPORT_SYMBOL (mempool_create_node);/** * mempool_resize - resize an existing memory pool
Quoted 1 line · Anton Blanchard
FYI the kernel has NUMA enabled but the machine has only 1 NUMA node.
Ummm. But it has multiple memory nodes? We ran into trouble with the funky memory arrangement before.
Quoted 3 lines
Ummm. But it has multiple memory nodes? We ran into trouble with the funky memory arrangement before.
Its a shared processor partition (meaning the partition can get scheduled anywhere), so we throw all the cpus and memory into node 0.
This machine only has 1 cpu (2 threads) and they are both in node0:
# ls -l /sys/devices/system/node/node0/
lrwxrwxrwx 1 root root 0 Jul 5 17:56 cpu0 -> ../../../../devices/system/cpu/cpu0 lrwxrwxrwx 1 root root 0 Jul 5 17:56 cpu1 -> ../../../../devices/system/cpu/cpu1
Anton
Quoted 9 lines · Anton Blanchard
Its a shared processor partition (meaning the partition can get scheduled anywhere), so we throw all the cpus and memory into node 0.
This machine only has 1 cpu (2 threads) and they are both in node0:
# ls -l /sys/devices/system/node/node0/lrwxrwxrwx 1 root root 0 Jul 5 17:56 cpu0 -> ../../../../devices/system/cpu/cpu0 lrwxrwxrwx 1 root root 0 Jul 5 17:56 cpu1 -> ../../../../devices/system/cpu/cpu1
What does pcibus_to_node return for the pcibus device that you are trying to allocate for?
There must be something special here. We have run this patch for a long time on a variety of platforms (not on powerpc I am afraid). Maybe a broken pcibus_to_node function for ppc?
Quoted 3 lines
What does
pcibus_to_nodereturn for the pcibus device that you are trying to allocate for?
Its using the default. Looking at include/asm-generic/topology.h:
#ifndef pcibus_to_node
#define pcibus_to_node(node) (-1)
#endif
I wonder what kmalloc_node does when you pass it -1.
Anton
include/asm-generic/topology.h +1 -1
@@ -42,7 +42,7 @@#define node_to_first_cpu (node) (0)#endif #ifndef pcibus_to_node -#define pcibus_to_node (node) (-1) +#define pcibus_to_node (node) (0)#endif #ifndef pcibus_to_cpumask
Quoted 7 lines · Anton Blanchard
Its using the default. Looking at
include/asm-generic/topology.h:#ifndef
pcibus_to_node
#definepcibus_to_node(node)(-1)
#endifI wonder what
kmalloc_nodedoes when you pass it -1.
That is the issue. It should fall back to kmalloc. One piece of the patchset dropped out between Andrew and Linus.
mm/slab.c +3 -0
@@ -2443,6 +2443,9 @@ {kmem_cache_t *cachep; +if (node ==-1 ) +return kmalloc (size, flags); + cachep =kmem_find_general_cachep (size, flags);if (unlikely (cachep ==NULL ))return NULL ;
Quoted 3 lines
That is the issue. It should fall back to kmalloc. One piece of the patchset dropped out between Andrew and Linus.
That helped, but I still had lots of free memory in the request_queue slab. Giving kmem_cache_alloc_node the same treatment fixed it, although I wonder if we just shouldnt be passing in -1.
Anton
mm/slab.c +6 -0
@@ -2372,6 +2372,9 @@struct slab *slabp;kmem_bufctl_t next; +if (nodeid ==-1 ) +return kmem_cache_alloc (cachep, flags); +for (loop =0 ;;loop++) {struct list_head *q;
@@ -2443,6 +2446,9 @@ {kmem_cache_t *cachep; +if (node ==-1 ) +return kmalloc (size, flags); + cachep =kmem_find_general_cachep (size, flags);if (unlikely (cachep ==NULL ))return NULL ;
Quoted 7 lines · Anton Blanchard
Quoted 3 lines
That is the issue. It should fall back to kmalloc. One piece of the patchset dropped out between Andrew and Linus.
That helped, but I still had lots of free memory in the
request_queueslab. Givingkmem_cache_alloc_nodethe same treatment fixed it, although I wonder if we just shouldnt be passing in -1.
No. No. -1 is the sign for undetermined that is also used in other pieces of slab.c. Meaning do not do node specific allocs. The fix should only modify kmem_cache_alloc_node. But we wont need that if the new slab allocator would get in. Andrew?
Also if powerpc wants to use node specific allocations for block devices then the powerpc arch needs to supply a pcibus_to_node function that does return the correct node for the pcibus.
mm/slab.c +3 -0
@@ -2372,6 +2372,9 @@struct slab *slabp;kmem_bufctl_t next; +if (nodeid ==-1 ) +return kmem_cache_alloc_node (cachep, flags); +for (loop =0 ;;loop++) {struct list_head *q;