[PATCH 1/2] drm/amdgpu: increase the reserved VM size to 2MB

[Chauhan, Madhav]

[AMD Public Use]

-----Original Message-----
From: Christian König <ckoenig.leichtzumerken at gmail.com> 
Sent: Wednesday, October 14, 2020 2:06 PM
To: Chauhan, Madhav <Madhav.Chauhan at amd.com>; amd-gfx at lists.freedesktop.org
Cc: Pan, Xinhui <Xinhui.Pan at amd.com>
Subject: Re: [PATCH 1/2] drm/amdgpu: increase the reserved VM size to 2MB

Am 14.10.20 um 10:26 schrieb Chauhan, Madhav:
> [AMD Public Use]
>
> -----Original Message-----
> From: Christian König <ckoenig.leichtzumerken at gmail.com>
> Sent: Tuesday, October 13, 2020 10:38 PM
> To: amd-gfx at lists.freedesktop.org
> Cc: Chauhan, Madhav <Madhav.Chauhan at amd.com>; Pan, Xinhui 
> <Xinhui.Pan at amd.com>
> Subject: [PATCH 1/2] drm/amdgpu: increase the reserved VM size to 2MB
>
> Ideally this should be a multiple of the VM block size.
> 2MB should at least fit for Vega/Navi.
>
> Signed-off-by: Christian König <christian.koenig at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index 7c46937c1c0e..81ccd0a0c3db 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -112,8 +112,8 @@ struct amdgpu_bo_list_entry;
>   #define AMDGPU_MMHUB_0				1
>   #define AMDGPU_MMHUB_1				2
>   
> -/* hardcode that limit for now */
> -#define AMDGPU_VA_RESERVED_SIZE			(1ULL << 20)
> +/* Reserve 2MB at top/bottom of address space for kernel use */
> +#define AMDGPU_VA_RESERVED_SIZE			(2ULL << 20)
>
> Looks fine to me: Reviewed-by: Madhav Chauhan <madhav.chauhan at amd.com> 
> Clarification on comment:
> We check va_address < AMDGPU_VA_RESERVED_SIZE for invalid reservations, shouldn’t this be "bottom"instead of "top/bottom" of address space??

In amdgpu_info_ioctl() we report AMDGPU_VA_RESERVED_SIZE as start of the usable address space and vm_size - AMDGPU_VA_RESERVED_SIZE as end.

Could be that we don't check if the address in the reserved hole at the end of the address space. That would be a bug and should probably be fixed.

Thanks. Inside amdgpu_gem_va_ioctl, Shouldn’t we also validate addresses bigger than AMDGPU_GMC_HOLE_END??
Currently we allow them and just remove last 16 bits. What will happen if User passed 0x ffff ffff ffff ffff, it will be treated as 0x ffff ffff ffff
While address space end is 0x ffff ffff 0000??

Regards,
Madhav

Christian.

>
> Regards,
> Madhav
>   
>   /* max vmids dedicated for process */
>   #define AMDGPU_VM_MAX_RESERVED_VMID	1