[PATCH v9 01/14] drm/amdgpu: UAPI for user queue management

Sharma, Shashank shashank.sharma at amd.com
Thu May 2 12:53:19 UTC 2024


On 02/05/2024 07:23, Sharma, Shashank wrote:
> Hey Alex,
>
> On 01/05/2024 22:39, Alex Deucher wrote:
>> On Fri, Apr 26, 2024 at 10:07 AM Shashank Sharma
>> <shashank.sharma at amd.com> wrote:
>>> From: Alex Deucher <alexander.deucher at amd.com>
>>>
>>> This patch intorduces new UAPI/IOCTL for usermode graphics
>>> queue. The userspace app will fill this structure and request
>>> the graphics driver to add a graphics work queue for it. The
>>> output of this UAPI is a queue id.
>>>
>>> This UAPI maps the queue into GPU, so the graphics app can start
>>> submitting work to the queue as soon as the call returns.
>>>
>>> V2: Addressed review comments from Alex and Christian
>>>      - Make the doorbell offset's comment clearer
>>>      - Change the output parameter name to queue_id
>>>
>>> V3: Integration with doorbell manager
>>>
>>> V4:
>>>      - Updated the UAPI doc (Pierre-Eric)
>>>      - Created a Union for engine specific MQDs (Alex)
>>>      - Added Christian's R-B
>>> V5:
>>>      - Add variables for GDS and CSA in MQD structure (Alex)
>>>      - Make MQD data a ptr-size pair instead of union (Alex)
>>>
>>> V9:
>>>     - renamed struct drm_amdgpu_userq_mqd_gfx_v11 to struct
>>>       drm_amdgpu_userq_mqd as its being used for SDMA and
>>>       compute queues as well
>>>
>>> Cc: Alex Deucher <alexander.deucher at amd.com>
>>> Cc: Christian Koenig <christian.koenig at amd.com>
>>> Reviewed-by: Christian König <christian.koenig at amd.com>
>>> Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
>>> Signed-off-by: Shashank Sharma <shashank.sharma at amd.com>
>>> ---
>>>   include/uapi/drm/amdgpu_drm.h | 110 
>>> ++++++++++++++++++++++++++++++++++
>>>   1 file changed, 110 insertions(+)
>>>
>>> diff --git a/include/uapi/drm/amdgpu_drm.h 
>>> b/include/uapi/drm/amdgpu_drm.h
>>> index 96e32dafd4f0..22f56a30f7cb 100644
>>> --- a/include/uapi/drm/amdgpu_drm.h
>>> +++ b/include/uapi/drm/amdgpu_drm.h
>>> @@ -54,6 +54,7 @@ extern "C" {
>>>   #define DRM_AMDGPU_VM                  0x13
>>>   #define DRM_AMDGPU_FENCE_TO_HANDLE     0x14
>>>   #define DRM_AMDGPU_SCHED               0x15
>>> +#define DRM_AMDGPU_USERQ               0x16
>>>
>>>   #define DRM_IOCTL_AMDGPU_GEM_CREATE DRM_IOWR(DRM_COMMAND_BASE + 
>>> DRM_AMDGPU_GEM_CREATE, union drm_amdgpu_gem_create)
>>>   #define DRM_IOCTL_AMDGPU_GEM_MMAP DRM_IOWR(DRM_COMMAND_BASE + 
>>> DRM_AMDGPU_GEM_MMAP, union drm_amdgpu_gem_mmap)
>>> @@ -71,6 +72,7 @@ extern "C" {
>>>   #define DRM_IOCTL_AMDGPU_VM DRM_IOWR(DRM_COMMAND_BASE + 
>>> DRM_AMDGPU_VM, union drm_amdgpu_vm)
>>>   #define DRM_IOCTL_AMDGPU_FENCE_TO_HANDLE DRM_IOWR(DRM_COMMAND_BASE 
>>> + DRM_AMDGPU_FENCE_TO_HANDLE, union drm_amdgpu_fence_to_handle)
>>>   #define DRM_IOCTL_AMDGPU_SCHED DRM_IOW(DRM_COMMAND_BASE + 
>>> DRM_AMDGPU_SCHED, union drm_amdgpu_sched)
>>> +#define DRM_IOCTL_AMDGPU_USERQ DRM_IOW(DRM_COMMAND_BASE + 
>>> DRM_AMDGPU_USERQ, union drm_amdgpu_userq)
>>>
>>>   /**
>>>    * DOC: memory domains
>>> @@ -317,6 +319,114 @@ union drm_amdgpu_ctx {
>>>          union drm_amdgpu_ctx_out out;
>>>   };
>>>
>>> +/* user queue IOCTL */
>>> +#define AMDGPU_USERQ_OP_CREATE 1
>>> +#define AMDGPU_USERQ_OP_FREE   2
>>> +
>>> +/* Flag to indicate secure buffer related workload, unused for now */
>>> +#define AMDGPU_USERQ_MQD_FLAGS_SECURE  (1 << 0)
>>> +/* Flag to indicate AQL workload, unused for now */
>>> +#define AMDGPU_USERQ_MQD_FLAGS_AQL     (1 << 1)
>>> +
>>> +/*
>>> + * MQD (memory queue descriptor) is a set of parameters which allow
>>> + * the GPU to uniquely define and identify a usermode queue. This
>>> + * structure defines the MQD for GFX-V11 IP ver 0.
>>> + */
>>> +struct drm_amdgpu_userq_mqd {
>> Maybe rename this to drm_amdgpu_gfx_userq_mqd since it's gfx specific.
>> Then we can add different MQDs for SDMA, compute, etc. as they have
>> different metadata.  E.g., the shadow and CSA are gfx only.
>
>
> Actually this was named drm_amdgpu_userq_mqd_gfx_v11_0 until the last 
> patchset, but then I realized that apart from the objects (gds/shadow 
> va) nothing is gfx specific, its actually required for every userqueue 
> IP which is MES based, so I thought it would be an overkill to create 
> multiple structures for almost the same data. If you feel strong about 
> this, I can change it again.
>
> - Shashank


Please ignore my last comment, I understand what you are mentioning, and 
I have reformatted the patches accordingly. Now, I am keeping everything 
reqd for MES in one basic struture (drm_amdgpu_userq_in) and creating  
drm_amdgpu_userq_mqd_gfx_v11 for GFX specific things (like CSA, Shadow 
and GDS areas). Now there will be one separate patch which will enabled 
GFX_IP on MES code, just like how we have separate patches for SDMA and 
Compute IP in this series.  I will send the V10 patches with this 
reformatting in some time.

- Shashank

>
>> Alex
>>
>>
>>> +       /**
>>> +        * @queue_va: Virtual address of the GPU memory which holds 
>>> the queue
>>> +        * object. The queue holds the workload packets.
>>> +        */
>>> +       __u64   queue_va;
>>> +       /**
>>> +        * @queue_size: Size of the queue in bytes, this needs to be 
>>> 256-byte
>>> +        * aligned.
>>> +        */
>>> +       __u64   queue_size;
>>> +       /**
>>> +        * @rptr_va : Virtual address of the GPU memory which holds 
>>> the ring RPTR.
>>> +        * This object must be at least 8 byte in size and aligned 
>>> to 8-byte offset.
>>> +        */
>>> +       __u64   rptr_va;
>>> +       /**
>>> +        * @wptr_va : Virtual address of the GPU memory which holds 
>>> the ring WPTR.
>>> +        * This object must be at least 8 byte in size and aligned 
>>> to 8-byte offset.
>>> +        *
>>> +        * Queue, RPTR and WPTR can come from the same object, as 
>>> long as the size
>>> +        * and alignment related requirements are met.
>>> +        */
>>> +       __u64   wptr_va;
>>> +       /**
>>> +        * @shadow_va: Virtual address of the GPU memory to hold the 
>>> shadow buffer.
>>> +        * This must be a from a separate GPU object, and must be at 
>>> least 4-page
>>> +        * sized.
>>> +        */
>>> +       __u64   shadow_va;
>>> +       /**
>>> +        * @gds_va: Virtual address of the GPU memory to hold the 
>>> GDS buffer.
>>> +        * This must be a from a separate GPU object, and must be at 
>>> least 1-page
>>> +        * sized.
>>> +        */
>>> +       __u64   gds_va;
>>> +       /**
>>> +        * @csa_va: Virtual address of the GPU memory to hold the 
>>> CSA buffer.
>>> +        * This must be a from a separate GPU object, and must be at 
>>> least 1-page
>>> +        * sized.
>>> +        */
>>> +       __u64   csa_va;
>>> +};
>>> +
>>> +struct drm_amdgpu_userq_in {
>>> +       /** AMDGPU_USERQ_OP_* */
>>> +       __u32   op;
>>> +       /** Queue handle for USERQ_OP_FREE */
>>> +       __u32   queue_id;
>>> +       /** the target GPU engine to execute workload 
>>> (AMDGPU_HW_IP_*) */
>>> +       __u32   ip_type;
>>> +       /**
>>> +        * @flags: flags to indicate special function for queue like 
>>> secure
>>> +        * buffer (TMZ). Unused for now.
>>> +        */
>>> +       __u32   flags;
>>> +       /**
>>> +        * @doorbell_handle: the handle of doorbell GEM object
>>> +        * associated to this client.
>>> +        */
>>> +       __u32   doorbell_handle;
>>> +       /**
>>> +        * @doorbell_offset: 32-bit offset of the doorbell in the 
>>> doorbell bo.
>>> +        * Kernel will generate absolute doorbell offset using 
>>> doorbell_handle
>>> +        * and doorbell_offset in the doorbell bo.
>>> +        */
>>> +       __u32   doorbell_offset;
>>> +       /**
>>> +        * @mqd: Queue descriptor for USERQ_OP_CREATE
>>> +        * MQD data can be of different size for different GPU 
>>> IP/engine and
>>> +        * their respective versions/revisions, so this points to a 
>>> __u64 *
>>> +        * which holds MQD of this usermode queue.
>>> +        */
>>> +       __u64 mqd;
>>> +       /**
>>> +        * @size: size of MQD data in bytes, it must match the MQD 
>>> structure
>>> +        * size of the respective engine/revision defined in UAPI 
>>> for ex, for
>>> +        * gfx_v11 workloads, size = 
>>> sizeof(drm_amdgpu_userq_mqd_gfx_v11).
>>> +        */
>>> +       __u64 mqd_size;
>>> +};
>>> +
>>> +struct drm_amdgpu_userq_out {
>>> +       /** Queue handle */
>>> +       __u32   queue_id;
>>> +       /** Flags */
>>> +       __u32   flags;
>>> +};
>>> +
>>> +union drm_amdgpu_userq {
>>> +       struct drm_amdgpu_userq_in in;
>>> +       struct drm_amdgpu_userq_out out;
>>> +};
>>> +
>>>   /* vm ioctl */
>>>   #define AMDGPU_VM_OP_RESERVE_VMID      1
>>>   #define AMDGPU_VM_OP_UNRESERVE_VMID    2
>>> -- 
>>> 2.43.2
>>>


More information about the amd-gfx mailing list