[PATCH v2] resource: limit request_free_mem_region based on arch_get_mappable_range

[Dan Williams]

D Scott Phillips wrote:
[..]
> Hi Dan, sorry for my incredibly delayed response, I lost your message to
> a filter on my end :(
> 
> I'm happy to work toward your preferred approach here, though I'm not
> sure I know how to achieve it. I think I understand how cxl is keeping
> device_private_memory out, but I don't think I understand the resource
> system well enough to see how amdgpu can make a properly trimmed
> resource for request_free_mem_region. My novice attempt would be
> something like:
> 
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> index 8ee3d07ffbdfa..d84de6d66ac45 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> @@ -1038,7 +1039,14 @@ int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
>                 pgmap->range.end = adev->gmc.aper_base + adev->gmc.aper_size - 1;
>                 pgmap->type = MEMORY_DEVICE_COHERENT;
>         } else {
> -               res = devm_request_free_mem_region(adev->dev, &iomem_resource, size);
> +               struct range mappable;
> +               struct resource root;
> +
> +               mappable = arch_get_mappable_range();
> +               root.start = mappable.start;
> +               root.end = mappable.end;
> +               root.child = iomem_resource.child;
> +               res = devm_request_free_mem_region(adev->dev, &root, size);
>                 if (IS_ERR(res))
>                         return PTR_ERR(res);
>                 pgmap->range.start = res->start;
> 
> Apart from this being wrong with respect to resource_lock, is that sort
> of the idea? or am I missing the sensible way to hoist the vmemmap range
> into iomem_resource? or maybe I'm just totally off in the weeds.

You have the right idea, however, I think a better solution has appeared
in the meantime. See this recent fix from Thomas regarding collisions
between KASLR and request_free_mem_region():

http://lore.kernel.org/172418629773.2215.4158024254077335422.tip-bot2@tip-bot2

...in that case KASLR is limiting the maximum possible usable address
range that request_free_mem_region() can play. For this
arch_get_mappable_range() restriction can you adjust the new
@physmem_end variable for the same effect?