Commit c51bb53d authored by David Yat Sin's avatar David Yat Sin Committed by Alex Deucher
Browse files

drm/amdkfd: Add metadata ring buffer for compute 



Add support for separate ring-buffer for metadata packets when using
compute queues. Userspace application allocate the metadata ring-buffer
and the queue ring-buffer with a single allocation. The metadata
ring-buffer starts after the queue ring-buffer.

Signed-off-by: default avatarDavid Yat Sin <David.YatSin@amd.com>
Reviewed-by: default avatarPhilip Yang <Philip.Yang@amd.com>
Signed-off-by: default avatarAlex Deucher <alexander.deucher@amd.com>
parent d0c989a0

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c

+8 −0
Original line number Diff line number Diff line
@@ -221,6 +221,11 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties, 
		pr_debug("Size lower. clamped to KFD_MIN_QUEUE_RING_SIZE");

	}



	if ((args->metadata_ring_size != 0) && !is_power_of_2(args->metadata_ring_size)) {

		pr_err("Metadata ring size must be a power of 2 or 0\n");

		return -EINVAL;

	}



	if (!access_ok((const void __user *) args->read_pointer_address,

			sizeof(uint32_t))) {

		pr_err("Can't access read pointer\n");
@@ -255,6 +260,9 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties, 
	q_properties->priority = args->queue_priority;

	q_properties->queue_address = args->ring_base_address;

	q_properties->queue_size = args->ring_size;

	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)

		q_properties->metadata_queue_size = args->metadata_ring_size;



	q_properties->read_ptr = (void __user *)args->read_pointer_address;

	q_properties->write_ptr = (void __user *)args->write_pointer_address;

	q_properties->eop_ring_buffer_address = args->eop_buffer_address;

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c

+21 −0
Original line number Diff line number Diff line
@@ -266,6 +266,27 @@ static void update_mqd(struct mqd_manager *mm, void *mqd, 
	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);

	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);



	if (q->metadata_queue_size) {

		/* On GC 12.1 is 64 DWs which is 4 times size of AQL packet */

		if (q->metadata_queue_size == q->queue_size * 4) {

			/*

			 * User application allocates main queue ring and metadata queue ring

			 * with a single allocation. metadata queue ring starts after main

			 * queue ring.

			 */

			m->cp_hqd_kd_base =

				lower_32_bits((q->queue_address + q->queue_size) >> 8);

			m->cp_hqd_kd_base_hi =

				upper_32_bits((q->queue_address + q->queue_size) >> 8);



			m->cp_hqd_kd_cntl |= CP_HQD_KD_CNTL__KD_FETCHER_ENABLE_MASK;

			/* KD_SIZE = 2 for metadata packet = 64 DWs */

			m->cp_hqd_kd_cntl |= 2 << CP_HQD_KD_CNTL__KD_SIZE__SHIFT;

		} else {

			pr_warn("Invalid metadata ring size, metadata queue will be ignored\n");

		}

	}



	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);

	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);

	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);

drivers/gpu/drm/amd/amdkfd/kfd_priv.h

+2 −1
Original line number Diff line number Diff line
@@ -507,6 +507,7 @@ struct queue_properties { 
	unsigned int queue_id;

	uint64_t queue_address;

	uint64_t queue_size;

	uint64_t metadata_queue_size;

	uint32_t priority;

	uint32_t queue_percent;

	void __user *read_ptr;

drivers/gpu/drm/amd/amdkfd/kfd_queue.c

+5 −2
Original line number Diff line number Diff line
@@ -247,9 +247,12 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope 
	    properties->format == KFD_QUEUE_FORMAT_AQL &&

	    topo_dev->node_props.gfx_target_version >= 70000 &&

	    topo_dev->node_props.gfx_target_version < 90000)

		expected_queue_size = properties->queue_size / 2;

		/* metadata_queue_size not supported on GFX7/GFX8 */

		expected_queue_size =

			properties->queue_size / 2;

	else

		expected_queue_size = properties->queue_size;

		expected_queue_size =

			properties->queue_size + properties->metadata_queue_size;



	vm = drm_priv_to_vm(pdd->drm_priv);

	err = amdgpu_bo_reserve(vm->root.bo, false);

include/uapi/linux/kfd_ioctl.h

+3 −2
Original line number Diff line number Diff line
@@ -47,9 +47,10 @@ 
 * - 1.19 - Add a new ioctl to craete secondary kfd processes

 * - 1.20 - Trap handler support for expert scheduling mode available

 * - 1.21 - Debugger support to subscribe to LDS out-of-address exceptions

 * - 1.22 - Add queue creation with metadata ring base address

 */

#define KFD_IOCTL_MAJOR_VERSION 1

#define KFD_IOCTL_MINOR_VERSION 21

#define KFD_IOCTL_MINOR_VERSION 22



struct kfd_ioctl_get_version_args {

	__u32 major_version;	/* from KFD */
@@ -87,7 +88,7 @@ struct kfd_ioctl_create_queue_args { 
	__u32 ctx_save_restore_size;	/* to KFD */

	__u32 ctl_stack_size;		/* to KFD */

	__u32 sdma_engine_id;		/* to KFD */

	__u32 pad;

	__u32 metadata_ring_size;	/* to KFD */

};



struct kfd_ioctl_destroy_queue_args {