Re: [PATCH 0/4] MR260: vkd3d: Prepare for buffering of command list commands.

[Evan Tang (＠etang-cw)]

Evan Tang (@etang-cw) commented about libs/vkd3d/command.c:
> - vkd3d_private_store_destroy(&signature->private_store);
> -
> - vkd3d_free((void *)signature->desc.pArgumentDescs);
> - vkd3d_free(signature);
> -
> - d3d12_device_release(device);
>
> +}
> -
>
> +static void d3d12_command_signature_add_pending(struct d3d12_command_signature *signature)
> +{
> - vkd3d_atomic_increment(&signature->pending_count);
>
> +}
> -
>
> +static void d3d12_command_signature_release_pending(struct d3d12_command_signature *signature)
> +{
> - if (!vkd3d_atomic_decrement(&signature->pending_count) && !signature->refcount)

Having two separate refcounts seems race-prone

I think you can double-free if one thread decrements the pending count while another decrements the refcount, then both see the other as zero