Re: [patch 20/25] debugobjects: Prepare kmem_cache allocations for batching

[Thomas Gleixner]

On Thu, Oct 10 2024 at 16:40, Leizhen wrote:
> On 2024/10/8 0:50, Thomas Gleixner wrote:
>> +static bool kmem_alloc_batch(struct hlist_head *head, struct kmem_cache *cache, gfp_t gfp)
>> +{
>> +	struct hlist_node *last = NULL;
>> +	struct debug_obj *obj;
>> +
>> +	for (int cnt = 0; cnt < ODEBUG_BATCH_SIZE; cnt++) {
>> +		obj = kmem_cache_zalloc(cache, gfp);
>> +		if (!obj) {
>> +			free_object_list(head);
>> +			return false;
>> +		}
>> +		debug_objects_allocated++;
>> +
>> +		if (!last)
>> +			last = &obj->node;
>> +		obj->batch_last = last;
>> +
>> +		hlist_add_head(&obj->node, head);
>
> There seems to be a need to use hlist_splice_init() outside the loop like patch 2/25.
> Prepare the local list first, if all ODEBUG_BATCH_SIZE is successfully allocated,
> attach it to 'head' at a time, including account debug_objects_allocated. Then back
> to patch 8/25, keep debug_objects_freed still in lock protection and do not move it
> into free_object_list(). Seems like it should be.

In case that the allocation of 16 object fails, the accounting
concurrency problem vs. free_object_list is pretty irrelevant.

It's debug information and moving it under lock protection for a corner
case is just making the code more complex.

Thanks,

        tglx