Re: [PATCH v5 04/10] drm/amdgpu: create GFX-gen11 usermode queue

[Felix Kuehling <felix.kuehling@xxxxxxx>]

Am 2023-07-12 um 11:55 schrieb Shashank Sharma:
> On 11/07/2023 21:51, Felix Kuehling wrote:
> > On 2023-07-06 09:39, Christian König wrote:
> > > Am 06.07.23 um 15:37 schrieb Shashank Sharma:
> > > > On 06/07/2023 15:22, Christian König wrote:
> > > > > Am 06.07.23 um 14:35 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.
> > > > > >
> > > > > > Cc: Alex Deucher <alexander.deucher@xxxxxxx>
> > > > > > Cc: Christian Koenig <christian.koenig@xxxxxxx>
> > > > > > Signed-off-by: Shashank Sharma <shashank.sharma@xxxxxxx>
> > > > > > Signed-off-by: Arvind Yadav <arvind.yadav@xxxxxxx>
> > > > > > ---
> > > > > >   drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c | 16 ++++
> > > > > >   drivers/gpu/drm/amd/amdgpu/gfx_v11_0.c        | 73 
> > > > > > +++++++++++++++++++
> > > > > >   .../gpu/drm/amd/include/amdgpu_userqueue.h    |  7 ++
> > > > > >   3 files changed, 96 insertions(+)
> > > > > >
> > > > > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c 
> > > > > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
> > > > > > index e37b5da5a0d0..bb774144c372 100644
> > > > > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
> > > > > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userqueue.c
> > > > > > @@ -134,12 +134,28 @@ 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;
> > > > > > +}
> > > > > > +
> > > > > >   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 c4940b6ea1c4..e76e1b86b434 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"
> > > > > > @@ -6486,3 +6487,75 @@ 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 */
> > > > > > +    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;
> > > > > > +    }
> > > > >
> > > > > Using amdgpu_bo_create_kernel() for the MQD is most likely not a 
> > > > > good idea in the long term, but should work for now.
> > > > I was a bit curious about this, the scope of this MQD object is 
> > > > kernel internal and used for queue mapping only, userspace doesn't 
> > > > know much about it. Do you still think we should not create a 
> > > > kernel object for it ?
> > >
> > >
> > > Well we should use a kernel BO. But amdgpu_bo_create_kernel() not 
> > > only creates a kernel BO but also pins it! And that is problematic 
> > > because it allows userspace to do a deny of service attach on the 
> > > kernel module.
> > >
> > > What we need is an eviction fence, e.g. what KFD is already using. 
> > > Then the BO is created similar to how VM page tables are created, 
> > > maybe even using the same reservation object.
> >
> > KFD doesn't currently use eviction fences on MQDs. Those are pinned. 
> > I guess you could treat the MQDs more like we treat page tables. They 
> > are allocated in kernel mode but protected with fences rather than 
> > pinning.
> >
> > I'm not sure if MES needs to be able to access MQDs while queues are 
> > not mapped. If that's the case, pinning can't be avoided.
> I was planning to do something like this keep userqueue objects from 
> eviction:
>
> - Add a new mutex in queue struct (say userq_eviction_fence)
>
> - lock this while mapping the queue, and unlock it while unmapping of 
> the queue
>
> - add checks in amdgpu_vm_evictable() to add a 
> mutex_trylock(userq_eviction_fence)
>
> Does it sound like what we want to do here or should I need something 
> else ?

A mutex is not a fence. The eviction fences we have in KFD are 
dma_fences. They interact with TTM's memory eviction logic, which allows 
us to stop user mode queues before TTM moves memory. A mutex cannot do that.

Regards,
  Felix


> > Regards,
> >   Felix
> >
> >
> > > But for a test this here is probably ok.
> > >
> > > Christian.
> > >
> > > > - Shashank
> > > >
> > > > > Probably best to add a comment here that this needs to be improved.
> > > > >
> > > > > Apart from that looks good to me,
> > > > > Christian.
> > > > >
> > > > > > +    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 {