Re: [PATCH v4 03/10] drm/amdgpu: add new IOCTL for usermode queue

[Shashank Sharma <shashank.sharma@amd.com>]

On 25/04/2023 14:14, Christian König wrote:
> Am 24.04.23 um 19:38 schrieb Shashank Sharma:
>> This patch adds:
>> - A new IOCTL function to create and destroy
>> - A new structure to keep all the user queue data in one place.
>> - A function to generate unique index for the queue.
>>
>> V1: Worked on review comments from RFC patch series:
>>    - Alex: Keep a list of queues, instead of single queue per process.
>>    - Christian: Use the queue manager instead of global ptrs,
>>             Don't keep the queue structure in amdgpu_ctx
>>
>> V2: Worked on review comments:
>>   - Christian:
>>     - Formatting of text
>>     - There is no need for queuing of userqueues, with idr in place
>>   - Alex:
>>     - Remove use_doorbell, its unnecessary
>>     - Reuse amdgpu_mqd_props for saving mqd fields
>>
>>   - Code formatting and re-arrangement
>>
>> V3:
>>   - Integration with doorbell manager
>>
>> V4:
>>   - Accommodate MQD union related changes in UAPI (Alex)
>>   - Do not set the queue size twice (Bas)
>>
>> Cc: Alex Deucher <alexander.deucher@amd.com>
>> Cc: Christian Koenig <christian.koenig@amd.com>
>> Signed-off-by: Shashank Sharma <shashank.sharma@amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c       |   1 +
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c | 126 ++++++++++++++++++
>>   .../gpu/drm/amd/include/amdgpu_userqueue.h    |   2 +
>>   3 files changed, 129 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>> index 2d6bcfd727c8..229976a2d0e7 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>> @@ -2749,6 +2749,7 @@ const struct drm_ioctl_desc amdgpu_ioctls_kms[] 
>> = {
>>       DRM_IOCTL_DEF_DRV(AMDGPU_GEM_VA, amdgpu_gem_va_ioctl, 
>> DRM_AUTH|DRM_RENDER_ALLOW),
>>       DRM_IOCTL_DEF_DRV(AMDGPU_GEM_OP, amdgpu_gem_op_ioctl, 
>> DRM_AUTH|DRM_RENDER_ALLOW),
>>       DRM_IOCTL_DEF_DRV(AMDGPU_GEM_USERPTR, amdgpu_gem_userptr_ioctl, 
>> DRM_AUTH|DRM_RENDER_ALLOW),
>> +    DRM_IOCTL_DEF_DRV(AMDGPU_USERQ, amdgpu_userq_ioctl, 
>> DRM_AUTH|DRM_RENDER_ALLOW),
>>   };
>>     static const struct drm_driver amdgpu_kms_driver = {
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c 
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
>> index effc0c7c02cf..333f31efbe7b 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
>> @@ -23,6 +23,132 @@
>>    */
>>     #include "amdgpu.h"
>> +#include "amdgpu_vm.h"
>> +#include "amdgpu_userqueue.h"
>> +
>> +static inline int
>> +amdgpu_userqueue_index(struct amdgpu_userq_mgr *uq_mgr, struct 
>> amdgpu_usermode_queue *queue)
>
> Maybe call that handle instead of index. And I don't think you 
> actually need those wrapper functions, see below.
>
>> +{
>> +    return idr_alloc(&uq_mgr->userq_idr, queue, 1, AMDGPU_MAX_USERQ, 
>> GFP_KERNEL);
>> +}
>> +
>> +static inline void
>> +amdgpu_userqueue_free_index(struct amdgpu_userq_mgr *uq_mgr, int 
>> queue_id)
>> +{
>> +    idr_remove(&uq_mgr->userq_idr, queue_id);
>> +}
>> +
>> +static struct amdgpu_usermode_queue *
>> +amdgpu_userqueue_find(struct amdgpu_userq_mgr *uq_mgr, int qid)
>> +{
>> +    return idr_find(&uq_mgr->userq_idr, qid);
>> +}
>> +
>> +static int amdgpu_userqueue_create_gfx(struct drm_file *filp, union 
>> drm_amdgpu_userq *args)
>> +{
>> +    struct amdgpu_usermode_queue *queue;
>> +    struct amdgpu_fpriv *fpriv = filp->driver_priv;
>> +    struct amdgpu_userq_mgr *uq_mgr = &fpriv->userq_mgr;
>> +    struct drm_amdgpu_userq_mqd_gfx *mqd_in = &args->in.mqd.gfx;
>> +    int r;
>> +
>> +    /* Do we support usermode queues on this GFX IP ? */
>> +    if (!uq_mgr->userq_funcs[AMDGPU_HW_IP_GFX]) {
>> +        DRM_ERROR("Usermode queue is not supported for this GFX IP\n");
>> +        return -EINVAL;
>> +    }
>> +
>> +    queue = kzalloc(sizeof(struct amdgpu_usermode_queue), GFP_KERNEL);
>> +    if (!queue) {
>> +        DRM_ERROR("Failed to allocate memory for queue\n");
>> +        return -ENOMEM;
>> +    }
>> +
>> +    mutex_lock(&uq_mgr->userq_mutex);
>> +    queue->userq_prop.wptr_gpu_addr = mqd_in->wptr_va;
>> +    queue->userq_prop.rptr_gpu_addr = mqd_in->rptr_va;
>> +    queue->userq_prop.queue_size = mqd_in->queue_size;
>> +    queue->userq_prop.hqd_base_gpu_addr = mqd_in->queue_va;
>> +
>> +    queue->doorbell_handle = mqd_in->doorbell_handle;
>> +    queue->queue_type = AMDGPU_HW_IP_GFX;
>> +    queue->flags = mqd_in->flags;
>> +    queue->vm = &fpriv->vm;
>> +    queue->queue_id = amdgpu_userqueue_index(uq_mgr, queue);
>
> It is usually good practice to *not* store those handles inside the 
> object they refer to, background is that this always creates a circle 
> reference (handle->object, object->handle).
>
> Additional to that I would move registering the userspace handle out 
> of this function.
I guess idr should be good enough to get me the queue ptr, and I don't 
have to save the queue id in queue.
>
>> +    if (queue->queue_id < 0) {
>> +        DRM_ERROR("Failed to allocate a queue id\n");
>> +        r = queue->queue_id;
>> +        goto free_queue;
>> +    }
>> +
>> +    args->out.queue_id = queue->queue_id;
>> +    args->out.flags = 0;
>> +    mutex_unlock(&uq_mgr->userq_mutex);
>> +    return 0;
>> +
>> +free_queue:
>> +    mutex_unlock(&uq_mgr->userq_mutex);
>> +    kfree(queue);
>> +    return r;
>> +}
>> +
>> +static int amdgpu_userqueue_create(struct drm_file *filp, union 
>> drm_amdgpu_userq *args)
>> +{
>> +    u32 ip_type = args->in.ip_type;
>> +
>> +    switch (ip_type) {
>> +    case AMDGPU_HW_IP_GFX:
>> +        return amdgpu_userqueue_create_gfx(filp, args);
>> +    default:
>> +        DRM_ERROR("Usermode queue is not supported for this IP 
>> (%u)\n", ip_type);
>> +    }
>
> Instead insert the idr_alloc() or more generally handle alloc here.

Noted,

- Shashank

>
> Christian.
>
>> +
>> +    return -EINVAL;
>> +}
>> +
>> +static void amdgpu_userqueue_destroy(struct drm_file *filp, int 
>> queue_id)
>> +{
>> +    struct amdgpu_fpriv *fpriv = filp->driver_priv;
>> +    struct amdgpu_userq_mgr *uq_mgr = &fpriv->userq_mgr;
>> +    struct amdgpu_usermode_queue *queue;
>> +
>> +    queue = amdgpu_userqueue_find(uq_mgr, queue_id);
>> +    if (!queue) {
>> +        DRM_DEBUG_DRIVER("Invalid queue id to destroy\n");
>> +        return;
>> +    }
>> +
>> +    mutex_lock(&uq_mgr->userq_mutex);
>> +    amdgpu_userqueue_free_index(uq_mgr, queue->queue_id);
>> +    mutex_unlock(&uq_mgr->userq_mutex);
>> +    kfree(queue);
>> +}
>> +
>> +int amdgpu_userq_ioctl(struct drm_device *dev, void *data,
>> +               struct drm_file *filp)
>> +{
>> +    union drm_amdgpu_userq *args = data;
>> +    int r = 0;
>> +
>> +    switch (args->in.op) {
>> +    case AMDGPU_USERQ_OP_CREATE:
>> +        r = amdgpu_userqueue_create(filp, args);
>> +        if (r)
>> +            DRM_ERROR("Failed to create usermode queue\n");
>> +        break;
>> +
>> +    case AMDGPU_USERQ_OP_FREE:
>> +        amdgpu_userqueue_destroy(filp, args->in.queue_id);
>> +        break;
>> +
>> +    default:
>> +        DRM_ERROR("Invalid user queue op specified: %d\n", 
>> args->in.op);
>> +        return -EINVAL;
>> +    }
>> +
>> +    return r;
>> +}
>> +
>>     int amdgpu_userq_mgr_init(struct amdgpu_userq_mgr *userq_mgr, 
>> struct amdgpu_device *adev)
>>   {
>> diff --git a/drivers/gpu/drm/amd/include/amdgpu_userqueue.h 
>> b/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
>> index 6fe5d8d73f37..8d8f6b3bcda5 100644
>> --- a/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
>> +++ b/drivers/gpu/drm/amd/include/amdgpu_userqueue.h
>> @@ -43,6 +43,8 @@ struct amdgpu_userq_funcs {
>>       void (*mqd_destroy)(struct amdgpu_userq_mgr *uq_mgr, struct 
>> amdgpu_usermode_queue *uq);
>>   };
>>   +int amdgpu_userq_ioctl(struct drm_device *dev, void *data, struct 
>> drm_file *filp);
>> +
>>   int amdgpu_userq_mgr_init(struct amdgpu_userq_mgr *userq_mgr, 
>> struct amdgpu_device *adev);
>>     void amdgpu_userq_mgr_fini(struct amdgpu_userq_mgr *userq_mgr);
>