[PATCH libdrm 2/2] amdgpu: add amdgpu_bo_va_op_raw

[Christian König]

Am 08.02.2017 um 13:55 schrieb Nicolai Hähnle:
> On 08.02.2017 13:44, Nicolai Hähnle wrote:
>> On 08.02.2017 13:39, Christian König wrote:
>>> Am 08.02.2017 um 13:34 schrieb Nicolai Hähnle:
>>>> From: Nicolai Hähnle <nicolai.haehnle at amd.com>
>>>>
>>>> This variant allows the caller full control over flags and size, and
>>>> allows passing a NULL bo (for PRT support).
>>>>
>>>> Cc: Christian König <christian.koenig at amd.com>
>>>> Cc: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
>>>> Cc: Jerry Zhang <Jerry.Zhang at amd.com>
>>>> Signed-off-by: Nicolai Hähnle <nicolai.haehnle at amd.com>
>>>
>>> Reviewed-by: Christian König <christian.koenig at amd.com>.
>>>
>>> You have a perfect timing, just wanted to ask if anybody has already
>>> done the unit test or if I should tackle it?
>>
>> I actually started hacking on radeonsi support while traveling and
>> decided to push that towards a somewhat usable state, so a series for
>> Mesa is going out now :)
>>
>> A unit test in libdrm is still a good idea, but I haven't done that.
>
> FWIW, I don't plan to write such a unit test, but I did send out a 
> basic piglit test as well that works for me with the Mesa series -- 
> minus the fact I still get GPU VM faults when I change the test to not 
> commit pages.

Ok, I've hacked together a quick unit test to write a few dw to a PRT 
mapping using the SDMA.

Seems to cause a VM fault as well without a message that we turned PRT 
handling on.

Going to investigate, looks like another issue with the kernel patches 
to me.

Regards,
Christian.

>
> Nicolai
>
>
>>
>> Cheers,
>> Nicolai
>>
>>>
>>> Regards,
>>> Christian.
>>>
>>>> ---
>>>>   amdgpu/amdgpu.h    | 28 ++++++++++++++++++++++++++++
>>>>   amdgpu/amdgpu_bo.c | 25 ++++++++++++++++++++-----
>>>>   2 files changed, 48 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/amdgpu/amdgpu.h b/amdgpu/amdgpu.h
>>>> index 7b26a04..6b2ded8 100644
>>>> --- a/amdgpu/amdgpu.h
>>>> +++ b/amdgpu/amdgpu.h
>>>> @@ -1186,6 +1186,34 @@ int amdgpu_bo_va_op(amdgpu_bo_handle bo,
>>>>               uint32_t ops);
>>>>     /**
>>>> + *  VA mapping/unmapping for a buffer object or PRT region.
>>>> + *
>>>> + * This is not a simple drop-in extension for amdgpu_bo_va_op;
>>>> instead, all
>>>> + * parameters are treated "raw", i.e. size is not automatically
>>>> aligned, and
>>>> + * all flags must be specified explicitly.
>>>> + *
>>>> + * \param  dev        - \c [in] device handle
>>>> + * \param  bo        - \c [in] BO handle (may be NULL)
>>>> + * \param  offset    - \c [in] Start offset to map
>>>> + * \param  size        - \c [in] Size to map
>>>> + * \param  addr        - \c [in] Start virtual address.
>>>> + * \param  flags    - \c [in] Supported flags for mapping/unmapping
>>>> + * \param  ops        - \c [in] AMDGPU_VA_OP_MAP or 
>>>> AMDGPU_VA_OP_UNMAP
>>>> + *
>>>> + * \return   0 on success\n
>>>> + *          <0 - Negative POSIX Error code
>>>> + *
>>>> +*/
>>>> +
>>>> +int amdgpu_bo_va_op_raw(amdgpu_device_handle dev,
>>>> +            amdgpu_bo_handle bo,
>>>> +            uint64_t offset,
>>>> +            uint64_t size,
>>>> +            uint64_t addr,
>>>> +            uint64_t flags,
>>>> +            uint32_t ops);
>>>> +
>>>> +/**
>>>>    *  create semaphore
>>>>    *
>>>>    * \param   sem       - \c [out] semaphore handle
>>>> diff --git a/amdgpu/amdgpu_bo.c b/amdgpu/amdgpu_bo.c
>>>> index d30fd1e..f725bfd 100644
>>>> --- a/amdgpu/amdgpu_bo.c
>>>> +++ b/amdgpu/amdgpu_bo.c
>>>> @@ -683,6 +683,23 @@ int amdgpu_bo_va_op(amdgpu_bo_handle bo,
>>>>                uint32_t ops)
>>>>   {
>>>>       amdgpu_device_handle dev = bo->dev;
>>>> +
>>>> +    size = ALIGN(size, getpagesize());
>>>> +
>>>> +    return amdgpu_bo_va_op_raw(dev, bo, offset, size, addr,
>>>> +                   AMDGPU_VM_PAGE_READABLE |
>>>> +                   AMDGPU_VM_PAGE_WRITEABLE |
>>>> +                   AMDGPU_VM_PAGE_EXECUTABLE, ops);
>>>> +}
>>>> +
>>>> +int amdgpu_bo_va_op_raw(amdgpu_device_handle dev,
>>>> +            amdgpu_bo_handle bo,
>>>> +            uint64_t offset,
>>>> +            uint64_t size,
>>>> +            uint64_t addr,
>>>> +            uint64_t flags,
>>>> +            uint32_t ops)
>>>> +{
>>>>       struct drm_amdgpu_gem_va va;
>>>>       int r;
>>>>   @@ -690,14 +707,12 @@ int amdgpu_bo_va_op(amdgpu_bo_handle bo,
>>>>           return -EINVAL;
>>>>         memset(&va, 0, sizeof(va));
>>>> -    va.handle = bo->handle;
>>>> +    va.handle = bo ? bo->handle : 0;
>>>>       va.operation = ops;
>>>> -    va.flags = AMDGPU_VM_PAGE_READABLE |
>>>> -           AMDGPU_VM_PAGE_WRITEABLE |
>>>> -           AMDGPU_VM_PAGE_EXECUTABLE;
>>>> +    va.flags = flags;
>>>>       va.va_address = addr;
>>>>       va.offset_in_bo = offset;
>>>> -    va.map_size = ALIGN(size, getpagesize());
>>>> +    va.map_size = size;
>>>>         r = drmCommandWriteRead(dev->fd, DRM_AMDGPU_GEM_VA, &va,
>>>> sizeof(va));
>>>>
>>>
>>>
>>
>