Commit ca1130de authored by Dave Airlie's avatar Dave Airlie
Browse files

Merge tag 'drm-amdkfd-next-2015-01-21' of git://people.freedesktop.org/~gabbayo/linux into drm-next 

- Infrastructure work in amdkfd to prepare for VI support. This work mainly
  includes separating modules into ASIC-specific functionality, adding
  new properties that are relevant for VI, making sure that shared code is
  reused, etc.

- Improve mechanism of submitting packets to HIQ (the kernel queue that amdkfd
  uses to issue commands to the GPU). The driver used to verify that each CS
  was read by the GPU. However, this proved to be both unnecessary and erroneous.
  Therefore, we cancelled this verification.

- Moved initialization of compute VMIDs into radeon driver

- Various minor fixes

* tag 'drm-amdkfd-next-2015-01-21' of git://people.freedesktop.org/~gabbayo/linux: (22 commits)
  drm/amdkfd: Fix description of sched_policy module parameter
  drm/amdkfd: Remove sync_with_hw() from amdkfd
  drm/amdkfd: Remove unused function busy_wait()
  drm/amdkfd: Replace cpu_relax() with schedule() in DQM
  drm/amdkfd: Fix for-loop when allocating HQD (non-HWS)
  drm/amdkfd: Add initial VI support for KQ
  drm/amdkfd: Encapsulate KQ functions in ops structure
  drm/amdkfd: Add initial VI support for DQM
  drm/amdkfd: Encapsulate DQM functions in ops structure
  drm/amdkfd: Don't BUG on freeing GART sub-allocation
  drm/amdkfd: Fix logic of destroy_queue_nocpsch()
  MAINTAINERS: Update amdkfd files
  drm/amdkfd: Change MQD manager to be H/W specific
  drm/amdkfd: Add asic property to kfd_device_info
  drm/amdkfd: Make KFD_MQD_TYPE enum types H/W agnostic
  drm/amdkfd: Add new VI-specific queue properties
  drm/radeon: Use new cik_structs.h file
  drm/amdkfd: Don't include header files from radeon
  drm/amd: Put cik structures in a common place
  drm/radeon: Don't use relative paths in #include
  ...
parents fc839753 cb2ac441

MAINTAINERS

+2 −0
Original line number Diff line number Diff line
@@ -624,6 +624,8 @@ L: dri-devel@lists.freedesktop.org 
T:      git git://people.freedesktop.org/~gabbayo/linux.git

S:      Supported

F:      drivers/gpu/drm/amd/amdkfd/

F:	drivers/gpu/drm/amd/include/cik_structs.h

F:	drivers/gpu/drm/amd/include/kgd_kfd_interface.h

F:      drivers/gpu/drm/radeon/radeon_kfd.c

F:      drivers/gpu/drm/radeon/radeon_kfd.h

F:      include/uapi/linux/kfd_ioctl.h

drivers/gpu/drm/amd/amdkfd/Makefile

+4 −1
Original line number Diff line number Diff line
@@ -7,8 +7,11 @@ ccflags-y := -Iinclude/drm -Idrivers/gpu/drm/amd/include/ 
amdkfd-y	:= kfd_module.o kfd_device.o kfd_chardev.o kfd_topology.o \

		kfd_pasid.o kfd_doorbell.o kfd_flat_memory.o \

		kfd_process.o kfd_queue.o kfd_mqd_manager.o \

		kfd_kernel_queue.o kfd_packet_manager.o \

		kfd_mqd_manager_cik.o kfd_mqd_manager_vi.o \

		kfd_kernel_queue.o kfd_kernel_queue_cik.o \

		kfd_kernel_queue_vi.o kfd_packet_manager.o \

		kfd_process_queue_manager.o kfd_device_queue_manager.o \

		kfd_device_queue_manager_cik.o kfd_device_queue_manager_vi.o \

		kfd_interrupt.o



obj-$(CONFIG_HSA_AMD)	+= amdkfd.o

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

+13 −0
Original line number Diff line number Diff line
@@ -168,6 +168,8 @@ 
#define	IB_ATC_EN					(1U << 23)

#define	DEFAULT_MIN_IB_AVAIL_SIZE			(3U << 20)



#define	AQL_ENABLE					1



#define CP_HQD_DEQUEUE_REQUEST				0xC974

#define	DEQUEUE_REQUEST_DRAIN				1

#define DEQUEUE_REQUEST_RESET				2
@@ -188,6 +190,17 @@ 
#define	MQD_VMID_MASK					(0xf << 0)

#define	MQD_CONTROL_PRIV_STATE_EN			(1U << 8)



#define	SDMA_RB_VMID(x)					(x << 24)

#define	SDMA_RB_ENABLE					(1 << 0)

#define	SDMA_RB_SIZE(x)					((x) << 1) /* log2 */

#define	SDMA_RPTR_WRITEBACK_ENABLE			(1 << 12)

#define	SDMA_RPTR_WRITEBACK_TIMER(x)			((x) << 16) /* log2 */

#define	SDMA_OFFSET(x)					(x << 0)

#define	SDMA_DB_ENABLE					(1 << 28)

#define	SDMA_ATC					(1 << 0)

#define	SDMA_VA_PTR32					(1 << 4)

#define	SDMA_VA_SHARED_BASE(x)				(x << 8)



#define GRBM_GFX_INDEX					0x30800

#define	INSTANCE_INDEX(x)				((x) << 0)

#define	SH_INDEX(x)					((x) << 8)

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

+31 −2
Original line number Diff line number Diff line
@@ -145,6 +145,8 @@ static long kfd_ioctl_get_version(struct file *filep, struct kfd_process *p, 
static int set_queue_properties_from_user(struct queue_properties *q_properties,

				struct kfd_ioctl_create_queue_args *args)

{

	void *tmp;



	if (args->queue_percentage > KFD_MAX_QUEUE_PERCENTAGE) {

		pr_err("kfd: queue percentage must be between 0 to KFD_MAX_QUEUE_PERCENTAGE\n");

		return -EINVAL;
@@ -182,6 +184,20 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties, 
		return -EFAULT;

	}



	tmp = (void *)(uintptr_t)args->eop_buffer_address;

	if (tmp != NULL &&

		!access_ok(VERIFY_WRITE, tmp, sizeof(uint32_t))) {

		pr_debug("kfd: can't access eop buffer");

		return -EFAULT;

	}



	tmp = (void *)(uintptr_t)args->ctx_save_restore_address;

	if (tmp != NULL &&

		!access_ok(VERIFY_WRITE, tmp, sizeof(uint32_t))) {

		pr_debug("kfd: can't access ctx save restore buffer");

		return -EFAULT;

	}



	q_properties->is_interop = false;

	q_properties->queue_percent = args->queue_percentage;

	q_properties->priority = args->queue_priority;
@@ -189,6 +205,11 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties, 
	q_properties->queue_size = args->ring_size;

	q_properties->read_ptr = (uint32_t *) args->read_pointer_address;

	q_properties->write_ptr = (uint32_t *) args->write_pointer_address;

	q_properties->eop_ring_buffer_address = args->eop_buffer_address;

	q_properties->eop_ring_buffer_size = args->eop_buffer_size;

	q_properties->ctx_save_restore_area_address =

			args->ctx_save_restore_address;

	q_properties->ctx_save_restore_area_size = args->ctx_save_restore_size;

	if (args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE ||

		args->queue_type == KFD_IOC_QUEUE_TYPE_COMPUTE_AQL)

		q_properties->type = KFD_QUEUE_TYPE_COMPUTE;
@@ -220,6 +241,11 @@ static int set_queue_properties_from_user(struct queue_properties *q_properties, 


	pr_debug("Queue Format (%d)\n", q_properties->format);



	pr_debug("Queue EOP (0x%llX)\n", q_properties->eop_ring_buffer_address);



	pr_debug("Queue CTX save arex (0x%llX)\n",

			q_properties->ctx_save_restore_area_address);



	return 0;

}
@@ -244,9 +270,12 @@ static long kfd_ioctl_create_queue(struct file *filep, struct kfd_process *p, 
	if (err)

		return err;



	pr_debug("kfd: looking for gpu id 0x%x\n", args.gpu_id);

	dev = kfd_device_by_id(args.gpu_id);

	if (dev == NULL)

	if (dev == NULL) {

		pr_debug("kfd: gpu id 0x%x was not found\n", args.gpu_id);

		return -EINVAL;

	}



	mutex_lock(&p->mutex);
@@ -410,7 +439,7 @@ static long kfd_ioctl_set_memory_policy(struct file *filep, 
		(args.alternate_policy == KFD_IOC_CACHE_POLICY_COHERENT)

		   ? cache_policy_coherent : cache_policy_noncoherent;



	if (!dev->dqm->set_cache_memory_policy(dev->dqm,

	if (!dev->dqm->ops.set_cache_memory_policy(dev->dqm,

				&pdd->qpd,

				default_policy,

				alternate_policy,

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

+16 −5
Original line number Diff line number Diff line
@@ -31,6 +31,14 @@ 
#define MQD_SIZE_ALIGNED 768



static const struct kfd_device_info kaveri_device_info = {

	.asic_family = CHIP_KAVERI,

	.max_pasid_bits = 16,

	.ih_ring_entry_size = 4 * sizeof(uint32_t),

	.mqd_size_aligned = MQD_SIZE_ALIGNED

};



static const struct kfd_device_info carrizo_device_info = {

	.asic_family = CHIP_CARRIZO,

	.max_pasid_bits = 16,

	.ih_ring_entry_size = 4 * sizeof(uint32_t),

	.num_of_watch_points = 4,
@@ -65,7 +73,7 @@ static const struct kfd_deviceid supported_devices[] = { 
	{ 0x1318, &kaveri_device_info },	/* Kaveri */

	{ 0x131B, &kaveri_device_info },	/* Kaveri */

	{ 0x131C, &kaveri_device_info },	/* Kaveri */

	{ 0x131D, &kaveri_device_info },	/* Kaveri */

	{ 0x131D, &kaveri_device_info }		/* Kaveri */

};



static int kfd_gtt_sa_init(struct kfd_dev *kfd, unsigned int buf_size,
@@ -245,7 +253,7 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd, 
		goto device_queue_manager_error;

	}



	if (kfd->dqm->start(kfd->dqm) != 0) {

	if (kfd->dqm->ops.start(kfd->dqm) != 0) {

		dev_err(kfd_device,

			"Error starting queuen manager for device (%x:%x)\n",

			kfd->pdev->vendor, kfd->pdev->device);
@@ -299,7 +307,7 @@ void kgd2kfd_suspend(struct kfd_dev *kfd) 
	BUG_ON(kfd == NULL);



	if (kfd->init_complete) {

		kfd->dqm->stop(kfd->dqm);

		kfd->dqm->ops.stop(kfd->dqm);

		amd_iommu_set_invalidate_ctx_cb(kfd->pdev, NULL);

		amd_iommu_free_device(kfd->pdev);

	}
@@ -320,7 +328,7 @@ int kgd2kfd_resume(struct kfd_dev *kfd) 
			return -ENXIO;

		amd_iommu_set_invalidate_ctx_cb(kfd->pdev,

						iommu_pasid_shutdown_callback);

		kfd->dqm->start(kfd->dqm);

		kfd->dqm->ops.start(kfd->dqm);

	}



	return 0;
@@ -503,7 +511,10 @@ int kfd_gtt_sa_free(struct kfd_dev *kfd, struct kfd_mem_obj *mem_obj) 
	unsigned int bit;



	BUG_ON(!kfd);

	BUG_ON(!mem_obj);



	/* Act like kfree when trying to free a NULL object */

	if (!mem_obj)

		return 0;



	pr_debug("kfd: free mem_obj = %p, range_start = %d, range_end = %d\n",

			mem_obj, mem_obj->range_start, mem_obj->range_end);

CRA Git | Maintained and supported by SUSTech CRA and CCSE