Re: [PATCH] vkd3d: Don't allow concurrent writes to descriptors.

[Derek Lesho]

On 9/11/19 11:11 PM, Conor McCarthy wrote:
> On Thu, Sep 12, 2019 at 5:18 AM Derek Lesho <[e-mail] 
> > wrote:
> +    for (unsigned int i = 0; i < desc->NumDescriptors; i++)
> +    {
> +        struct d3d12_desc *cur_desc = (struct d3d12_desc *)
> (object->descriptors + (i * descriptor_size));
> +        pthread_spinlock_t *lock =
> vkd3d_malloc(sizeof(pthread_spinlock_t));
> +
> +        pthread_spin_init(lock, PTHREAD_PROCESS_PRIVATE);
> +        cur_desc->lock = lock;
> +    }
>
>
> On Linux, pthread_spinlock_t is just a single int (not sure about the 
> Mac equivalent but probably similar). It's more efficient to declare 
> cur_desc->lock as a pthread_spinlock_t rather than a pointer to one in 
> allocated memory.

Yes, however the entire d3d12_desc structure is memset away, including 
the lock.  Another solution to this problem would involve using 
"descriptor headers" that go in between the _desc structures in 
heap->descriptors.