[PATCH v6 4/9] drm/amdgpu: create GFX-gen11 usermode queue

Christian König christian.koenig at amd.com
Mon Oct 2 09:42:55 UTC 2023


Am 28.09.23 um 15:54 schrieb Shashank Sharma:
>
> On 28/09/2023 15:52, Alex Deucher wrote:
>> On Thu, Sep 28, 2023 at 9:40 AM Shashank Sharma 
>> <shashank.sharma at amd.com> wrote:
>>>
>>> On 28/09/2023 15:27, Alex Deucher wrote:
>>>> On Thu, Sep 28, 2023 at 9:22 AM Shashank Sharma 
>>>> <shashank.sharma at amd.com> wrote:
>>>>> On 14/09/2023 10:24, Shashank Sharma wrote:
>>>>>> On 14/09/2023 09:45, Christian König wrote:
>>>>>>> Am 08.09.23 um 18:04 schrieb Shashank Sharma:
>>>>>>>> A Memory queue descriptor (MQD) of a userqueue defines it in
>>>>>>>> the hw's context. As MQD format can vary between different
>>>>>>>> graphics IPs, we need gfx GEN specific handlers to create MQDs.
>>>>>>>>
>>>>>>>> This patch:
>>>>>>>> - Introduces MQD handler functions for the usermode queues.
>>>>>>>> - Adds new functions to create and destroy userqueue MQD for
>>>>>>>>      GFX-GEN-11 IP
>>>>>>>>
>>>>>>>> V1: Worked on review comments from Alex:
>>>>>>>>        - Make MQD functions GEN and IP specific
>>>>>>>>
>>>>>>>> V2: Worked on review comments from Alex:
>>>>>>>>        - Reuse the existing adev->mqd[ip] for MQD creation
>>>>>>>>        - Formatting and arrangement of code
>>>>>>>>
>>>>>>>> V3:
>>>>>>>>        - Integration with doorbell manager
>>>>>>>>
>>>>>>>> V4: Review comments addressed:
>>>>>>>>        - Do not create a new file for userq, reuse gfx_v11_0.c 
>>>>>>>> (Alex)
>>>>>>>>        - Align name of structure members (Luben)
>>>>>>>>        - Don't break up the Cc tag list and the Sob tag list in 
>>>>>>>> commit
>>>>>>>>          message (Luben)
>>>>>>>> V5:
>>>>>>>>       - No need to reserve the bo for MQD (Christian).
>>>>>>>>       - Some more changes to support IP specific MQD creation.
>>>>>>>>
>>>>>>>> V6:
>>>>>>>>       - Add a comment reminding us to replace the
>>>>>>>> amdgpu_bo_create_kernel()
>>>>>>>>         calls while creating MQD object to amdgpu_bo_create() once
>>>>>>>> eviction
>>>>>>>>         fences are ready (Christian).
>>>>>>>>
>>>>>>>> Cc: Alex Deucher <alexander.deucher at amd.com>
>>>>>>>> Cc: Christian Koenig <christian.koenig at amd.com>
>>>>>>>> Signed-off-by: Shashank Sharma <shashank.sharma at amd.com>
>>>>>>>> Signed-off-by: Arvind Yadav <arvind.yadav at amd.com>
>>>>>>>> ---
>>>>>>>>     drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c | 16 ++++
>>>>>>>>     drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c        | 77
>>>>>>>> +++++++++++++++++++
>>>>>>>>     .../gpu/drm/amd/include/amdgpu_userqueue.h    | 7 ++
>>>>>>>>     3 files changed, 100 insertions(+)
>>>>>>>>
>>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
>>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
>>>>>>>> index 44769423ba30..03fc8e89eafb 100644
>>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
>>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
>>>>>>>> @@ -140,12 +140,28 @@ int amdgpu_userq_ioctl(struct drm_device 
>>>>>>>> *dev,
>>>>>>>> void *data,
>>>>>>>>         return r;
>>>>>>>>     }
>>>>>>>>     +extern const struct amdgpu_userq_funcs userq_gfx_v11_funcs;
>>>>>>>> +
>>>>>>>> +static void
>>>>>>>> +amdgpu_userqueue_setup_gfx(struct amdgpu_userq_mgr *uq_mgr)
>>>>>>>> +{
>>>>>>>> +    int maj;
>>>>>>>> +    struct amdgpu_device *adev = uq_mgr->adev;
>>>>>>>> +    uint32_t version = adev->ip_versions[GC_HWIP][0];
>>>>>>>> +
>>>>>>>> +    /* We support usermode queue only for GFX V11 as of now */
>>>>>>>> +    maj = IP_VERSION_MAJ(version);
>>>>>>>> +    if (maj == 11)
>>>>>>>> +        uq_mgr->userq_funcs[AMDGPU_HW_IP_GFX] = 
>>>>>>>> &userq_gfx_v11_funcs;
>>>>>>>> +}
>>>>>>> That belongs into gfx_v11.c and not here.
>>>>>> Agree,
>>>>> On a second thought, we can't move it to gfx_v11.c, as this is the 
>>>>> place
>>>>> where we are setting up the gfx_userqueue functions in 
>>>>> fpriv->uq_mgr()
>>>>> for the first time, and we do not have another option but to check 
>>>>> the
>>>>> IP and setup the functions here. The only other option to do this 
>>>>> will
>>>>> be to move  uq_mgr->userq_funcs to adev->gfx.userq_funcs and setup 
>>>>> them
>>>>> with the IP init (as Alex once suggested). Please let me know your
>>>>> thoughts on this.
>>>> That seems cleaner to me.  They should be global anyway and could be
>>>> set as part of the individual IP init sequences.  Then the presence of
>>>> a pointer could be used to determine whether or not a particular IP
>>>> type supports user queues.
>>>>
>>>> Alex
>>>>
>>> So if I understand this correctly, this is how we are looking to 
>>> arrange
>>> the userqueue IP functions:
>>>
>>> - Presence of adev->gfx.funcs.userqueue_funcs() will decide if this IP
>>> supports userqueue or not.
>>>
>>> - sw_init function of the IP will setup these fptrs like:
>>>
>>>     in gfx_v11_0_sw_init :
>>>
>>>       if (GFX_MAJ == 11)
>>>
>>>           adev->gfx.funcs.userqueue_funcs = gfx_v11_0_userqueue_funcs
>> I was thinking something more like:
>>
>> adev->userq_funcs[AMD_IP_BLOCK_TYPE_GFX] = gfx_v11_0_userqueue_funcs;
>>
>> That way there would be one place for all of the all of the fptrs and
>> you could use the IP type to query the right one.
>>
>> And then in the IOCTLs, you could just check if the pointer is 
>> valid.  E.g.,
>>
>> if (!adev->userq_funcs[ip_block_type])
>>     return -EINVAL;
>>
>> etc.
>>
>> You could store any metadata relevant to each userq in the per fd user
>> queue manager and then just pass the state to the global userq
>> functions for each IP.
>
> Makes sense, we can do that, hope this works for Christian as well, 
> @Christian ?

Perfectly fine with me.

Regards,
Christian.

>
> - Shashank
>
>>
>> Alex
>>
>>>
>>> In amdgpu_userqueue_ioctl:
>>>
>>>       |
>>>
>>> CASE: create:
>>>
>>>       amdgpu_userqueue_create()
>>>
>>>       if (adev->gfx.funcs.userqueue_funcs) {
>>>
>>>           adev->gfx.funcs.userqueue_funcs.create_mqd();
>>>
>>>       }
>>>
>>>
>>> CASE: destroy:
>>>
>>>       amdgpu_userqueue_destroy()
>>>
>>>       if (adev->gfx.funcs.userqueue_funcs) {
>>>
>>>           adev->gfx.funcs.userqueue_funcs.destroy_mqd();
>>>
>>>       }
>>>
>>> and so on ...
>>>
>>> Am I getting this right ?
>>>
>>> - Shashank
>>>
>>>>> - Shashank
>>>>>
>>>>>>>> +
>>>>>>>>     int amdgpu_userq_mgr_init(struct amdgpu_userq_mgr *userq_mgr,
>>>>>>>> struct amdgpu_device *adev)
>>>>>>>>     {
>>>>>>>>         mutex_init(&userq_mgr->userq_mutex);
>>>>>>>>         idr_init_base(&userq_mgr->userq_idr, 1);
>>>>>>>>         userq_mgr->adev = adev;
>>>>>>>>     +    amdgpu_userqueue_setup_gfx(userq_mgr);
>>>>>>>>         return 0;
>>>>>>>>     }
>>>>>>>>     diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
>>>>>>>> b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
>>>>>>>> index 0451533ddde4..6760abda08df 100644
>>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
>>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c
>>>>>>>> @@ -30,6 +30,7 @@
>>>>>>>>     #include "amdgpu_psp.h"
>>>>>>>>     #include "amdgpu_smu.h"
>>>>>>>>     #include "amdgpu_atomfirmware.h"
>>>>>>>> +#include "amdgpu_userqueue.h"
>>>>>>>>     #include "imu_v11_0.h"
>>>>>>>>     #include "soc21.h"
>>>>>>>>     #include "nvd.h"
>>>>>>>> @@ -6422,3 +6423,79 @@ const struct amdgpu_ip_block_version
>>>>>>>> gfx_v11_0_ip_block =
>>>>>>>>         .rev = 0,
>>>>>>>>         .funcs = &gfx_v11_0_ip_funcs,
>>>>>>>>     };
>>>>>>>> +
>>>>>>>> +static int gfx_v11_0_userq_mqd_create(struct amdgpu_userq_mgr 
>>>>>>>> *uq_mgr,
>>>>>>>> +                      struct drm_amdgpu_userq_in *args_in,
>>>>>>>> +                      struct amdgpu_usermode_queue *queue)
>>>>>>>> +{
>>>>>>>> +    struct amdgpu_device *adev = uq_mgr->adev;
>>>>>>>> +    struct amdgpu_mqd *mqd_gfx_generic =
>>>>>>>> &adev->mqds[AMDGPU_HW_IP_GFX];
>>>>>>>> +    struct drm_amdgpu_userq_mqd_gfx_v11_0 mqd_user;
>>>>>>>> +    struct amdgpu_mqd_prop userq_props;
>>>>>>>> +    int r;
>>>>>>>> +
>>>>>>>> +    /* Incoming MQD parameters from userspace to be saved here */
>>>>>>>> +    memset(&mqd_user, 0, sizeof(mqd_user));
>>>>>>>> +
>>>>>>>> +    /* Structure to initialize MQD for userqueue using generic 
>>>>>>>> MQD
>>>>>>>> init function */
>>>>>>>> +    memset(&userq_props, 0, sizeof(userq_props));
>>>>>>>> +
>>>>>>>> +    if (args_in->mqd_size != sizeof(struct
>>>>>>>> drm_amdgpu_userq_mqd_gfx_v11_0)) {
>>>>>>>> +        DRM_ERROR("MQD size mismatch\n");
>>>>>>>> +        return -EINVAL;
>>>>>>>> +    }
>>>>>>>> +
>>>>>>>> +    if (copy_from_user(&mqd_user, u64_to_user_ptr(args_in->mqd),
>>>>>>>> args_in->mqd_size)) {
>>>>>>>> +        DRM_ERROR("Failed to get user MQD\n");
>>>>>>>> +        return -EFAULT;
>>>>>>>> +    }
>>>>>>>> +
>>>>>>>> +    /*
>>>>>>>> +     * Create BO for actual Userqueue MQD now
>>>>>>>> +     * Todo: replace the calls to bo_create_kernel() with
>>>>>>>> bo_create() and use
>>>>>>>> +     * implicit pinning for the MQD buffers.
>>>>>>> Well not implicit pinning, but rather fencing of the BO.
>>>>>>>
>>>>>> Noted.
>>>>>>
>>>>>> - Shashank
>>>>>>
>>>>>>
>>>>>>> Regards,
>>>>>>> Christian.
>>>>>>>
>>>>>>>> +     */
>>>>>>>> +    r = amdgpu_bo_create_kernel(adev, mqd_gfx_generic->mqd_size,
>>>>>>>> PAGE_SIZE,
>>>>>>>> +                    AMDGPU_GEM_DOMAIN_GTT,
>>>>>>>> +                    &queue->mqd.obj,
>>>>>>>> +                    &queue->mqd.gpu_addr,
>>>>>>>> +                    &queue->mqd.cpu_ptr);
>>>>>>>> +    if (r) {
>>>>>>>> +        DRM_ERROR("Failed to allocate BO for userqueue (%d)", r);
>>>>>>>> +        return -ENOMEM;
>>>>>>>> +    }
>>>>>>>> +    memset(queue->mqd.cpu_ptr, 0, mqd_gfx_generic->mqd_size);
>>>>>>>> +
>>>>>>>> +    /* Initialize the MQD BO with user given values */
>>>>>>>> +    userq_props.wptr_gpu_addr = mqd_user.wptr_va;
>>>>>>>> +    userq_props.rptr_gpu_addr = mqd_user.rptr_va;
>>>>>>>> +    userq_props.queue_size = mqd_user.queue_size;
>>>>>>>> +    userq_props.hqd_base_gpu_addr = mqd_user.queue_va;
>>>>>>>> +    userq_props.mqd_gpu_addr = queue->mqd.gpu_addr;
>>>>>>>> +    userq_props.use_doorbell = true;
>>>>>>>> +
>>>>>>>> +    r = mqd_gfx_generic->init_mqd(adev, (void 
>>>>>>>> *)queue->mqd.cpu_ptr,
>>>>>>>> &userq_props);
>>>>>>>> +    if (r) {
>>>>>>>> +        DRM_ERROR("Failed to initialize MQD for userqueue\n");
>>>>>>>> +        goto free_mqd;
>>>>>>>> +    }
>>>>>>>> +
>>>>>>>> +    return 0;
>>>>>>>> +
>>>>>>>> +free_mqd:
>>>>>>>> +    amdgpu_bo_free_kernel(&queue->mqd.obj, &queue->mqd.gpu_addr,
>>>>>>>> &queue->mqd.cpu_ptr);
>>>>>>>> +    return r;
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +static void
>>>>>>>> +gfx_v11_0_userq_mqd_destroy(struct amdgpu_userq_mgr *uq_mgr, 
>>>>>>>> struct
>>>>>>>> amdgpu_usermode_queue *queue)
>>>>>>>> +{
>>>>>>>> +    struct amdgpu_userq_obj *mqd = &queue->mqd;
>>>>>>>> +
>>>>>>>> +    amdgpu_bo_free_kernel(&mqd->obj, &mqd->gpu_addr, 
>>>>>>>> &mqd->cpu_ptr);
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> +const struct amdgpu_userq_funcs userq_gfx_v11_funcs = {
>>>>>>>> +    .mqd_create = gfx_v11_0_userq_mqd_create,
>>>>>>>> +    .mqd_destroy = gfx_v11_0_userq_mqd_destroy,
>>>>>>>> +};
>>>>>>>> diff --git a/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
>>>>>>>> b/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
>>>>>>>> index 55ed6512a565..240f92796f00 100644
>>>>>>>> --- a/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
>>>>>>>> +++ b/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
>>>>>>>> @@ -29,6 +29,12 @@
>>>>>>>>       struct amdgpu_mqd_prop;
>>>>>>>>     +struct amdgpu_userq_obj {
>>>>>>>> +    void         *cpu_ptr;
>>>>>>>> +    uint64_t     gpu_addr;
>>>>>>>> +    struct amdgpu_bo *obj;
>>>>>>>> +};
>>>>>>>> +
>>>>>>>>     struct amdgpu_usermode_queue {
>>>>>>>>         int            queue_type;
>>>>>>>>         uint64_t        doorbell_handle;
>>>>>>>> @@ -37,6 +43,7 @@ struct amdgpu_usermode_queue {
>>>>>>>>         struct amdgpu_mqd_prop    *userq_prop;
>>>>>>>>         struct amdgpu_userq_mgr *userq_mgr;
>>>>>>>>         struct amdgpu_vm    *vm;
>>>>>>>> +    struct amdgpu_userq_obj mqd;
>>>>>>>>     };
>>>>>>>>       struct amdgpu_userq_funcs {



More information about the amd-gfx mailing list