On 2020/11/11 17:11, Christoph Hellwig wrote:
>> + spin_lock_init(&dev->zone_res_lock);
>> + if (dev->memory_backed)
>> dev->zone_locks = bitmap_zalloc(dev->nr_zones, GFP_KERNEL);
>> + else
>> + dev->zone_locks = kmalloc_array(dev->nr_zones,
>> + sizeof(spinlock_t), GFP_KERNEL);
>> + if (!dev->zone_locks) {
>
> Using the same pointer for different types is pretty horrible. At very
> least we should use a union.
>
> But I think we'd end up with less contention if we add a new
>
> struct nullb_zone
>
> that contains the fields of blk_zone that we actuall care about plus:
>
> union {
> spinlock_t spinlock;
> mutex_t mutex;
> }
>
> and then use the proper lock primitive also for the memory backed case.
OK. Cooking V3 with that.
--
Damien Le Moal
Western Digital Research
