Re: [PATCH 15/16] drm/amdgpu: use doorbell mgr for MES kernel doorbells

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

On 30/03/2023 22:58, Alex Deucher wrote:
> On Wed, Mar 29, 2023 at 11:48 AM Shashank Sharma
> <shashank.sharma@amd.com> wrote:
>> This patch:
>> - adds a doorbell object in MES structure, to manage the MES
>>    doorbell requirements in kernel.
>> - Removes the doorbell management code, and its variables from
>>    the doorbell_init function, it will be done in doorbell manager
>>    now.
>> - creates doorbell pages for MES kernel level needs (doorbells
>>    for MES self tests)
>> - current MES code was allocating MES doorbells in MES process context,
>>    but those were rung using kernel doorbell calls. This patch allocates
>>    MES kernel doorbells instead for this in add_hw_queue.
>>
>> Cc: Alex Deucher <alexander.deucher@amd.com>
>> Cc: Christian Koenig <christian.koenig@amd.com>
>> Signed-off-by: Shashank Sharma <shashank.sharma@amd.com>
>> Signed-off-by: Arvind Yadav <arvind.yadav@amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 105 ++++++++++++------------
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h |   5 +-
>>   2 files changed, 56 insertions(+), 54 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
>> index 0c546245793b..423cd642647c 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c
>> @@ -65,91 +65,89 @@ unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar(
>>                  doorbell_id * 2);
>>   }
>>
>> -static int amdgpu_mes_queue_doorbell_get(struct amdgpu_device *adev,
>> +static int amdgpu_mes_kernel_doorbell_get(struct amdgpu_device *adev,
>>                                           struct amdgpu_mes_process *process,
>>                                           int ip_type, uint64_t *doorbell_index)
>>   {
>>          unsigned int offset, found;
>> +       struct amdgpu_doorbell_obj *doorbells = &adev->mes.kernel_doorbells;
>>
>> -       if (ip_type == AMDGPU_RING_TYPE_SDMA) {
>> +       if (ip_type == AMDGPU_RING_TYPE_SDMA)
>>                  offset = adev->doorbell_index.sdma_engine[0];
>> -               found = find_next_zero_bit(process->doorbell_bitmap,
>> -                                          AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS,
>> -                                          offset);
>> -       } else {
>> -               found = find_first_zero_bit(process->doorbell_bitmap,
>> -                                           AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS);
>> -       }
>> +       else
>> +               offset = 0;
>>
>> +       found = find_next_zero_bit(doorbells->doorbell_bitmap,
>> +                                  AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS,
>> +                                  offset);
>>          if (found >= AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS) {
>>                  DRM_WARN("No doorbell available\n");
>>                  return -ENOSPC;
>>          }
>>
>> -       set_bit(found, process->doorbell_bitmap);
>> +       set_bit(found, doorbells->doorbell_bitmap);
>>
>> -       *doorbell_index = amdgpu_mes_get_doorbell_dw_offset_in_bar(adev,
>> -                               process->doorbell_index, found);
>> +       *doorbell_index = amdgpu_doorbell_index_on_bar(adev, doorbells->bo, found);
>>
>>          return 0;
>>   }
>>
>> -static void amdgpu_mes_queue_doorbell_free(struct amdgpu_device *adev,
>> +static void amdgpu_mes_kernel_doorbell_free(struct amdgpu_device *adev,
>>                                             struct amdgpu_mes_process *process,
>>                                             uint32_t doorbell_index)
>>   {
>>          unsigned int old, doorbell_id;
>> +       struct amdgpu_doorbell_obj *doorbells = &adev->mes.kernel_doorbells;
>>
>> -       doorbell_id = doorbell_index -
>> -               (process->doorbell_index *
>> -                amdgpu_mes_doorbell_process_slice(adev)) / sizeof(u32);
>> +       /* Find the relative index of the doorbell in this object */
>> +       doorbell_id = doorbell_index - doorbells->start;
>>          doorbell_id /= 2;
>>
>> -       old = test_and_clear_bit(doorbell_id, process->doorbell_bitmap);
>> +       old = test_and_clear_bit(doorbell_id, doorbells->doorbell_bitmap);
>>          WARN_ON(!old);
>>   }
>>
>>   static int amdgpu_mes_doorbell_init(struct amdgpu_device *adev)
>>   {
>> -       size_t doorbell_start_offset;
>> -       size_t doorbell_aperture_size;
>> -       size_t doorbell_process_limit;
>> -       size_t aggregated_doorbell_start;
>> -       int i;
>> -
>> -       aggregated_doorbell_start = (adev->doorbell_index.max_assignment + 1) * sizeof(u32);
>> -       aggregated_doorbell_start =
>> -               roundup(aggregated_doorbell_start, PAGE_SIZE);
>> -
>> -       doorbell_start_offset = aggregated_doorbell_start + PAGE_SIZE;
>> -       doorbell_start_offset =
>> -               roundup(doorbell_start_offset,
>> -                       amdgpu_mes_doorbell_process_slice(adev));
>> -
>> -       doorbell_aperture_size = adev->doorbell.size;
>> -       doorbell_aperture_size =
>> -                       rounddown(doorbell_aperture_size,
>> -                                 amdgpu_mes_doorbell_process_slice(adev));
>> -
>> -       if (doorbell_aperture_size > doorbell_start_offset)
>> -               doorbell_process_limit =
>> -                       (doorbell_aperture_size - doorbell_start_offset) /
>> -                       amdgpu_mes_doorbell_process_slice(adev);
>> -       else
>> -               return -ENOSPC;
>> +       int i, r;
>> +       u32 agg_db_start_index, nbits;
>> +       struct amdgpu_doorbell_obj *mes_doorbells = &adev->mes.kernel_doorbells;
>>
>> -       adev->mes.doorbell_id_offset = doorbell_start_offset / sizeof(u32);
>> -       adev->mes.max_doorbell_slices = doorbell_process_limit;
>> +               /* Allocated one page doorbells for MES kernel usages */
>> +       mes_doorbells->size = PAGE_SIZE;
>>
>> -       /* allocate Qword range for aggregated doorbell */
>> -       for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++)
>> -               adev->mes.aggregated_doorbells[i] =
>> -                       aggregated_doorbell_start / sizeof(u32) + i * 2;
>> +       nbits = DIV_ROUND_UP(mes_doorbells->size, sizeof(u32));
>> +       mes_doorbells->doorbell_bitmap = bitmap_zalloc(nbits, GFP_KERNEL);
>> +       if (!mes_doorbells->doorbell_bitmap) {
>> +               DRM_ERROR("Failed to allocate MES doorbell bitmap\n");
>> +               return -ENOMEM;
>> +       }
>> +
>> +       r = amdgpu_doorbell_alloc_page(adev, mes_doorbells);
> Rather than allocating a separate page here, just allocate two pages
> in the earlier patch where you allocate the KGD doorbell and then just
> use the second page here.

got it,

- Shashank

> Alex
>
>> +       if (r) {
>> +               DRM_ERROR("Failed to create MES doorbell object\n, err=%d", r);
>> +               bitmap_free(mes_doorbells->doorbell_bitmap);
>> +               return r;
>> +       }
>> +
>> +       /* Get the absolute doorbell index for aggregated doobells */
>> +       agg_db_start_index = mes_doorbells->start;
>> +       for (i = 0; i < AMDGPU_MES_PRIORITY_NUM_LEVELS; i++) {
>> +               adev->mes.aggregated_doorbells[i] = agg_db_start_index + i;
>> +               set_bit(agg_db_start_index + i, mes_doorbells->doorbell_bitmap);
>> +       }
>>
>> -       DRM_INFO("max_doorbell_slices=%zu\n", doorbell_process_limit);
>>          return 0;
>>   }
>>
>> +static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
>> +{
>> +       struct amdgpu_doorbell_obj *mes_doorbells = &adev->mes.kernel_doorbells;
>> +
>> +       bitmap_free(mes_doorbells->doorbell_bitmap);
>> +       amdgpu_doorbell_free_page(adev, mes_doorbells);
>> +}
>> +
>>   int amdgpu_mes_init(struct amdgpu_device *adev)
>>   {
>>          int i, r;
>> @@ -248,6 +246,7 @@ void amdgpu_mes_fini(struct amdgpu_device *adev)
>>          amdgpu_device_wb_free(adev, adev->mes.sch_ctx_offs);
>>          amdgpu_device_wb_free(adev, adev->mes.query_status_fence_offs);
>>          amdgpu_device_wb_free(adev, adev->mes.read_val_offs);
>> +       amdgpu_mes_doorbell_free(adev);
>>
>>          idr_destroy(&adev->mes.pasid_idr);
>>          idr_destroy(&adev->mes.gang_id_idr);
>> @@ -677,7 +676,7 @@ int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
>>          *queue_id = queue->queue_id = r;
>>
>>          /* allocate a doorbell index for the queue */
>> -       r = amdgpu_mes_queue_doorbell_get(adev, gang->process,
>> +       r = amdgpu_mes_kernel_doorbell_get(adev, gang->process,
>>                                            qprops->queue_type,
>>                                            &qprops->doorbell_off);
>>          if (r)
>> @@ -735,7 +734,7 @@ int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
>>          return 0;
>>
>>   clean_up_doorbell:
>> -       amdgpu_mes_queue_doorbell_free(adev, gang->process,
>> +       amdgpu_mes_kernel_doorbell_free(adev, gang->process,
>>                                         qprops->doorbell_off);
>>   clean_up_queue_id:
>>          spin_lock_irqsave(&adev->mes.queue_id_lock, flags);
>> @@ -790,7 +789,7 @@ int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id)
>>                            queue_id);
>>
>>          list_del(&queue->list);
>> -       amdgpu_mes_queue_doorbell_free(adev, gang->process,
>> +       amdgpu_mes_kernel_doorbell_free(adev, gang->process,
>>                                         queue->doorbell_off);
>>          amdgpu_mes_unlock(&adev->mes);
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h
>> index 97c05d08a551..e7e9dfe44c99 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h
>> @@ -27,6 +27,7 @@
>>   #include "amdgpu_irq.h"
>>   #include "kgd_kfd_interface.h"
>>   #include "amdgpu_gfx.h"
>> +#include "amdgpu_doorbell.h"
>>   #include <linux/sched/mm.h>
>>
>>   #define AMDGPU_MES_MAX_COMPUTE_PIPES        8
>> @@ -76,7 +77,6 @@ struct amdgpu_mes {
>>          uint32_t                        kiq_version;
>>
>>          uint32_t                        total_max_queue;
>> -       uint32_t                        doorbell_id_offset;
>>          uint32_t                        max_doorbell_slices;
>>
>>          uint64_t                        default_process_quantum;
>> @@ -128,6 +128,9 @@ struct amdgpu_mes {
>>          int                             (*kiq_hw_init)(struct amdgpu_device *adev);
>>          int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
>>
>> +       /* MES Kernel doorbells */
>> +       struct amdgpu_doorbell_obj      kernel_doorbells;
>> +
>>          /* ip specific functions */
>>          const struct amdgpu_mes_funcs   *funcs;
>>   };
>> --
>> 2.40.0
>>