[PATCH v2] drm/amdgpu: Fix the EPERM error when get user pages

[Felix Kuehling]

On 2023-05-19 08:02, Ma Jun wrote:
> Check and pass the readonly flags when set amdgpu_ttm_tt flags
>
> for readonly ptr pages. Otherwise, there is EPERM error returned
>
> during the KFDExceptionTest.PermissionFaultUserPointer test on
>
> ploaris10.

I remember you asked me about this maybe 2 weeks ago. I'm still not sure 
how your test is supposed to work correctly. Someone needs to decide 
whether to set KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE when registering the 
userptr. How does this happen? I think you need some code that checks 
that the VMA is read-only and then clears the WRITABLE flag from 
mem->alloc_flags. But I don't see that anywhere. I don't think we ever 
added proper support for read-only userptrs in KFD. Your patch is a 
partial fix at best, and I don't see how this helps.

Two more nit-picks inline.


>
> Signed-off-by: Ma Jun <Jun.Ma2 at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 9 ++++++++-
>   1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index d426333e865a..85d1087439c0 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -1036,10 +1036,17 @@ static int init_user_pages(struct kgd_mem *mem, uint64_t user_addr,
>   	struct ttm_operation_ctx ctx = { true, false };
>   	struct hmm_range *range;
>   	int ret = 0;
> +	uint32_t tt_flags = 0;
>   
>   	mutex_lock(&process_info->lock);
>   
> -	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, 0);
> +	if (!(mem->alloc_flags & KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE))
> +		tt_flags |= AMDGPU_GEM_USERPTR_READONLY;
> +	else
> +		tt_flags = 0;

This else-branch is unnecessary because you initialized tt_flags to 0 above.


> +
> +

Double blank lines. One is enough.

Regards,
   Felix


> +	ret = amdgpu_ttm_tt_set_userptr(&bo->tbo, user_addr, tt_flags);
>   	if (ret) {
>   		pr_err("%s: Failed to set userptr: %d\n", __func__, ret);
>   		goto out;