[PATCH] drm/amdkfd: don't add DOORBELL and MMIO BOs to validate list

[Felix Kuehling]

On 2022-05-25 06:37, Christian König wrote:
> Am 25.05.22 um 11:25 schrieb Lang Yu:
>> On 05/25/ , Christian König wrote:
>>> Am 25.05.22 um 10:43 schrieb Lang Yu:
>>>> DOORBELL and MMIO BOs never move once created.
>>>> No need to validate them after that.
>>> Yeah, but you still need to make sure their page tables are up to date.
>>>
>>> So this here might break horrible.
>> These BOs(and attachments) are validated when allocated and mapped.
>> Their page tables should be determined at this time.
>>
>> The kfd_bo_list is used to restore BOs after evictions.
>>
>> Do you mean their page tabes could be changed? Thanks.
>
> Yes, page tables can be destroyed under memory pressure as well.

Is that actually happening today, or is that some future optimization 
you have in mind? I know page tables can get evicted, but I didn't think 
they were destroyed unless the memory at that address is unmapped (which 
never happens for pinned BOs).


>
> Not sure how the KFD handles that, but in theory we should have every 
> BO used by a process on the validation list. Even the ones pinned.

Then we already have some other broken cases for the small number of 
kmapped BOs that are pinned and currently removed from the validation 
list (see amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel).

Regards,
   Felix


>
> Regards,
> Christian.
>
>>
>>
>>> Christian.
>>>
>>>> Signed-off-by: Lang Yu <Lang.Yu at amd.com>
>>>> ---
>>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 14 
>>>> +++++++++-----
>>>>    1 file changed, 9 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c 
>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
>>>> index 34ba9e776521..45de9cadd771 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
>>>> @@ -808,6 +808,10 @@ static void add_kgd_mem_to_kfd_bo_list(struct 
>>>> kgd_mem *mem,
>>>>        struct ttm_validate_buffer *entry = &mem->validate_list;
>>>>        struct amdgpu_bo *bo = mem->bo;
>>>> +    if (mem->alloc_flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
>>>> +                KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))
>>>> +        return;
>>>> +
>>>>        INIT_LIST_HEAD(&entry->head);
>>>>        entry->num_shared = 1;
>>>>        entry->bo = &bo->tbo;
>>>> @@ -824,6 +828,10 @@ static void 
>>>> remove_kgd_mem_from_kfd_bo_list(struct kgd_mem *mem,
>>>>    {
>>>>        struct ttm_validate_buffer *bo_list_entry;
>>>> +    if (mem->alloc_flags & (KFD_IOC_ALLOC_MEM_FLAGS_DOORBELL |
>>>> +                KFD_IOC_ALLOC_MEM_FLAGS_MMIO_REMAP))
>>>> +        return;
>>>> +
>>>>        bo_list_entry = &mem->validate_list;
>>>>        mutex_lock(&process_info->lock);
>>>>        list_del(&bo_list_entry->head);
>>>> @@ -1649,7 +1657,6 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
>>>>        unsigned long bo_size = mem->bo->tbo.base.size;
>>>>        struct kfd_mem_attachment *entry, *tmp;
>>>>        struct bo_vm_reservation_context ctx;
>>>> -    struct ttm_validate_buffer *bo_list_entry;
>>>>        unsigned int mapped_to_gpu_memory;
>>>>        int ret;
>>>>        bool is_imported = false;
>>>> @@ -1677,10 +1684,7 @@ int amdgpu_amdkfd_gpuvm_free_memory_of_gpu(
>>>>        }
>>>>        /* Make sure restore workers don't access the BO any more */
>>>> -    bo_list_entry = &mem->validate_list;
>>>> -    mutex_lock(&process_info->lock);
>>>> -    list_del(&bo_list_entry->head);
>>>> -    mutex_unlock(&process_info->lock);
>>>> +    remove_kgd_mem_from_kfd_bo_list(mem, process_info);
>>>>        /* No more MMU notifiers */
>>>>        amdgpu_mn_unregister(mem->bo);
>