[PATCH] drm/amdgpu: add VM update fences back to the root PD v2

Tom St Denis tom.stdenis at amd.com
Wed Feb 19 15:47:44 UTC 2020


The tip of origin/amd-staging-drm-next for me is:

commit 7fd3b632e17e55c5ffd008f9f025754e7daa1b66
Refs: {origin/amd-staging-drm-next}, v5.4-rc7-2847-g7fd3b632e17e
Author:     Monk Liu <Monk.Liu at amd.com>
AuthorDate: Thu Feb 6 23:55:58 2020 +0800
Commit:     Monk Liu <Monk.Liu at amd.com>
CommitDate: Wed Feb 19 13:33:02 2020 +0800

     drm/amdgpu: fix colliding of preemption

     what:
     some os preemption path is messed up with world switch preemption

     fix:
     cleanup those logics so os preemption not mixed with world switch

     this patch is a general fix for issues comes from SRIOV MCBP, but
     there is still UMD side issues not resovlved yet, so this patch
     cannot fix all world switch bug.

     Signed-off-by: Monk Liu <Monk.Liu at amd.com>
     Acked-by: Hawking Zhang <Hawking.Zhang at amd.com>

Which I had fetched just an hour ago.

On 2020-02-19 10:41 a.m., Christian König wrote:
> Well what branch are you trying to merge that into?
>
> The conflict resolution should be simple, just keep the 
> vm->update_funcs->prepare(...) line as it is in your branch.
>
> When you get those errors then something went wrong in your rebase.
>
> Christian.
>
> Am 19.02.20 um 16:14 schrieb Tom St Denis:
>> Doesn't build even with conflict resolved:
>>
>> [root at raven linux]# make
>>   CALL    scripts/checksyscalls.sh
>>   CALL    scripts/atomic/check-atomics.sh
>>   DESCEND  objtool
>>   CHK     include/generated/compile.h
>>   CC [M]  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.o
>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c: In function 
>> ‘amdgpu_vm_bo_update_mapping’:
>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:1612:41: error: ‘owner’ 
>> undeclared (first use in this function)
>>  1612 |  r = vm->update_funcs->prepare(&params, owner, exclusive);
>>       |                                         ^~~~~
>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:1612:41: note: each undeclared 
>> identifier is reported only once for each function it appears in
>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:1612:48: error: ‘exclusive’ 
>> undeclared (first use in this function)
>>  1612 |  r = vm->update_funcs->prepare(&params, owner, exclusive);
>>       |                                                ^~~~~~~~~
>> make[4]: *** [scripts/Makefile.build:266: 
>> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.o] Error 1
>> make[3]: *** [scripts/Makefile.build:509: drivers/gpu/drm/amd/amdgpu] 
>> Error 2
>> make[2]: *** [scripts/Makefile.build:509: drivers/gpu/drm] Error 2
>> make[1]: *** [scripts/Makefile.build:509: drivers/gpu] Error 2
>> make: *** [Makefile:1649: drivers] Error 2
>>
>> Should I just move to drm-misc-next?
>>
>> tom
>>
>> On 2020-02-19 10:02 a.m., Christian König wrote:
>>> Add update fences to the root PD while mapping BOs.
>>>
>>> Otherwise PDs freed during the mapping won't wait for
>>> updates to finish and can cause corruptions.
>>>
>>> v2: rebased on drm-misc-next
>>>
>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>> Fixes: 90b69cdc5f159 drm/amdgpu: stop adding VM updates fences to 
>>> the resv obj
>>> ---
>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 14 ++++++++++++--
>>>   1 file changed, 12 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> index d16231d6a790..ef73fa94f357 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>> @@ -588,8 +588,8 @@ void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm,
>>>   {
>>>       entry->priority = 0;
>>>       entry->tv.bo = &vm->root.base.bo->tbo;
>>> -    /* One for TTM and one for the CS job */
>>> -    entry->tv.num_shared = 2;
>>> +    /* Two for VM updates, one for TTM and one for the CS job */
>>> +    entry->tv.num_shared = 4;
>>>       entry->user_pages = NULL;
>>>       list_add(&entry->tv.head, validated);
>>>   }
>>> @@ -1591,6 +1591,16 @@ static int amdgpu_vm_bo_update_mapping(struct 
>>> amdgpu_device *adev,
>>>           goto error_unlock;
>>>       }
>>>   +    if (flags & AMDGPU_PTE_VALID) {
>>> +        struct amdgpu_bo *root = vm->root.base.bo;
>>> +
>>> +        if (!dma_fence_is_signaled(vm->last_direct))
>>> +            amdgpu_bo_fence(root, vm->last_direct, true);
>>> +
>>> +        if (!dma_fence_is_signaled(vm->last_delayed))
>>> +            amdgpu_bo_fence(root, vm->last_delayed, true);
>>> +    }
>>> +
>>>       r = vm->update_funcs->prepare(&params, owner, exclusive);
>>>       if (r)
>>>           goto error_unlock;
>


More information about the amd-gfx mailing list