Re: [PATCH 1/2] SLUB: Out-of-memory diagnostics

From: David Rientjes
Date: Thu Jun 11 2009 - 15:39:51 EST


On Thu, 11 Jun 2009, Pekka J Enberg wrote:

> diff --git a/mm/slub.c b/mm/slub.c
> index 65ffda5..2bbacfc 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -1484,6 +1484,56 @@ static inline int node_match(struct kmem_cache_cpu *c, int node)
> return 1;
> }
>
> +static int count_free(struct page *page)
> +{
> + return page->objects - page->inuse;
> +}
> +
> +static unsigned long count_partial(struct kmem_cache_node *n,
> + int (*get_count)(struct page *))
> +{
> + unsigned long flags;
> + unsigned long x = 0;
> + struct page *page;
> +
> + spin_lock_irqsave(&n->list_lock, flags);
> + list_for_each_entry(page, &n->partial, lru)
> + x += get_count(page);
> + spin_unlock_irqrestore(&n->list_lock, flags);
> + return x;
> +}
> +
> +static noinline void
> +slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
> +{
> + int node;
> +
> + printk(KERN_WARNING
> + "SLUB: Unable to allocate memory on node %d (gfp=%x)\n",
> + nid, gfpflags);
> + printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
> + "default order: %d, min order: %d\n", s->name, s->objsize,
> + s->size, oo_order(s->oo), oo_order(s->min));
> +
> + for_each_online_node(node) {
> + struct kmem_cache_node *n = get_node(s, node);
> + unsigned long nr_slabs;
> + unsigned long nr_objs;
> + unsigned long nr_free;
> +
> + if (!n)
> + continue;
> +
> + nr_slabs = atomic_long_read(&n->nr_slabs);
> + nr_objs = atomic_long_read(&n->total_objects);

This won't compile unless CONFIG_SLUB_DEBUG is enabled.

Perhaps slab_out_of_memory()'s partial list scan should only be declared
for CONFIG_SLUB_DEBUG? Otherwise, nr_slabs and nr_objs will always be 0
since we don't increment them and nr_free will appear out of sync. I
suspect debuggers will be asking if CONFIG_SLUB_DEBUG is enabled often in
such scenarios.

> + nr_free = count_partial(n, count_free);
> +
> + printk(KERN_WARNING
> + " node %d: slabs: %ld, objs: %ld, free: %ld\n",
> + node, nr_slabs, nr_objs, nr_free);
> + }
> +}
> +
> /*
> * Slow path. The lockless freelist is empty or we need to perform
> * debugging duties.
> @@ -1565,6 +1615,7 @@ new_slab:
> c->page = new;
> goto load_freelist;
> }
> + slab_out_of_memory(s, gfpflags, node);

I assume this is being added to __slab_alloc() and not new_slab() because
you need to reenable irqs before using count_partial()?

Do you really need to move count_partial() and count_free() out from under
CONFIG_SLUB_DEBUG? Couldn't you just add this to new_slab():

out:
if (!page)
slab_out_of_memory(s, flags, node);
return page;

and then add your own partial list scanner in slab_out_of_memory():

nr_free = 0;
spin_lock(&n->list_lock);
list_for_each_entry(page, &n->partial, lru)
nr_free += page->objects - page->inuse;
spin_unlock(&n->list_lock);
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/