On 2020/7/7 下午2:59, Christopher Lameter wrote:
> On Thu, 2 Jul 2020, Xunlei Pang wrote:
>
>> This patch introduces two counters to maintain the actual number
>> of partial objects dynamically instead of iterating the partial
>> page lists with list_lock held.
>>
>> New counters of kmem_cache_node are: pfree_objects, ptotal_objects.
>> The main operations are under list_lock in slow path, its performance
>> impact is minimal.
>
>
> If at all then these counters need to be under CONFIG_SLUB_DEBUG.
>
>> --- a/mm/slab.h
>> +++ b/mm/slab.h
>> @@ -616,6 +616,8 @@ struct kmem_cache_node {
>> #ifdef CONFIG_SLUB
>> unsigned long nr_partial;
>> struct list_head partial;
>> + atomic_long_t pfree_objects; /* partial free objects */
>> + atomic_long_t ptotal_objects; /* partial total objects */
>
> Please in the CONFIG_SLUB_DEBUG. Without CONFIG_SLUB_DEBUG we need to
> build with minimal memory footprint.
Thanks for the comments.
show_slab_objects() also calls it with CONFIG_SYSFS:
if (flags & SO_PARTIAL) {
struct kmem_cache_node *n;
for_each_kmem_cache_node(s, node, n) {
if (flags & SO_TOTAL)
x = count_partial(n, count_total);
else if (flags & SO_OBJECTS)
x = count_partial(n, count_inuse);
else
x = n->nr_partial;
total += x;
nodes[node] += x;
}
}
I'm not sure if it's due to some historical reason, it works without
CONFIG_SLUB_DEBUG.
>
>> #ifdef CONFIG_SLUB_DEBUG
>> atomic_long_t nr_slabs;
>> atomic_long_t total_objects;
>> diff --git a/mm/slub.c b/mm/slub.c
>
>
>
> Also this looks to be quite heavy on the cache and on execution time. Note
> that the list_lock could be taken frequently in the performance sensitive
> case of freeing an object that is not in the partial lists.
>
Yes, the concurrent __slab_free() has potential lock/atomic contention,
how about using percpu variable for partial free like below?
static inline void
__update_partial_free(struct kmem_cache_node *n, long delta)
{
atomic_long_add(delta, this_cpu_ptr(n->partial_free_objs));
}
-Xunlei
