drm/amdgpu: implement ring set_priority for gfx_v8 compute v8wip-high-priority-v9

Programming CP_HQD_QUEUE_PRIORITY enables a queue to take priority over
other queues on the same pipe. Multiple queues on a pipe are timesliced
so this gives us full precedence over other queues.

Programming CP_HQD_PIPE_PRIORITY changes the SPI_ARB_PRIORITY of the
wave as follows:
        0x2: CS_H
        0x1: CS_M
        0x0: CS_L

The SPI block will then dispatch work according to the policy set by
SPI_ARB_PRIORITY. In the current policy CS_H is higher priority than
gfx.

In order to prevent getting stuck in loops of CUs bouncing between GFX
and high priority compute and introducing further latency, we reserve
CUs 2+ for high priority compute on-demand.

v2: fix srbm_select to ring->queue and use ring->funcs->type
v3: use AMD_SCHED_PRIORITY_* instead of AMDGPU_CTX_PRIORITY_*
v4: switch int to enum amd_sched_priority
v5: corresponding changes for srbm_lock
v6: change CU reservation to PIPE_PERCENT allocation
v7: use kiq instead of MMIO
v8: back to MMIO, and make the implementation sleep safe.

Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>

3 files changed, 119 insertions, 0 deletions

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 8ae867137d84..25c51ac35ce2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -1097,6 +1097,10 @@ struct amdgpu_gfx {
 	bool                            in_suspend;
 	/* NGG */
 	struct amdgpu_ngg		ngg;
+
+	/* pipe reservation */
+	struct mutex			pipe_reserve_mutex;
+	DECLARE_BITMAP			(pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
 };
 
 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index c93516a6901f..7fb70d514f75 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -1826,6 +1826,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 	adev->vm_manager.vm_pte_num_rings = 0;
 	adev->gart.gart_funcs = NULL;
 	adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
+	bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
 
 	adev->smc_rreg = &amdgpu_invalid_rreg;
 	adev->smc_wreg = &amdgpu_invalid_wreg;
@@ -1854,6 +1855,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 	mutex_init(&adev->pm.mutex);
 	mutex_init(&adev->gfx.gpu_clock_mutex);
 	mutex_init(&adev->srbm_mutex);
+	mutex_init(&adev->gfx.pipe_reserve_mutex);
 	mutex_init(&adev->grbm_idx_mutex);
 	mutex_init(&adev->mn_lock);
 	hash_init(adev->mn_hash);
diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
index 4ba2aebc6384..d75c8b48bd0c 100644
--- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
@@ -6649,6 +6649,118 @@ static void gfx_v8_0_ring_set_wptr_compute(struct amdgpu_ring *ring)
 	WDOORBELL32(ring->doorbell_index, lower_32_bits(ring->wptr));
 }
 
+static void gfx_v8_0_ring_set_pipe_percent(struct amdgpu_ring *ring,
+					   bool acquire)
+{
+	struct amdgpu_device *adev = ring->adev;
+	int pipe_num, tmp, reg;
+	int pipe_percent = acquire ? SPI_WCL_PIPE_PERCENT_GFX__VALUE_MASK : 0x1;
+
+	pipe_num = ring->me * adev->gfx.mec.num_pipe_per_mec + ring->pipe;
+
+	/* first me only has 2 entries, GFX and HP3D */
+	if (ring->me > 0)
+		pipe_num -= 2;
+
+	/* There is a bug in the GFX pipes that results in a HS
+	 * deadlock if the pipe is restricted to a percentage
+	 * lower than 17 */
+	if (ring->funcs->type != AMDGPU_RING_TYPE_COMPUTE && pipe_percent < 17)
+		pipe_percent = 17;
+
+	reg = mmSPI_WCL_PIPE_PERCENT_GFX + pipe_num;
+	tmp = RREG32(reg);
+	tmp = REG_SET_FIELD(tmp, SPI_WCL_PIPE_PERCENT_GFX, VALUE, pipe_percent);
+	WREG32(reg, tmp);
+}
+
+static void gfx_v8_0_pipe_reserve_resources(struct amdgpu_device *adev,
+					    struct amdgpu_ring *ring,
+					    bool acquire)
+{
+	int i, pipe;
+	bool reserve;
+	struct amdgpu_ring *iring;
+
+	mutex_lock(&adev->gfx.pipe_reserve_mutex);
+	pipe = amdgpu_queue_to_bit(adev, ring->me, ring->pipe, 0);
+	if (acquire)
+		set_bit(pipe, adev->gfx.pipe_reserve_bitmap);
+	else
+		clear_bit(pipe, adev->gfx.pipe_reserve_bitmap);
+
+	if (!bitmap_weight(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES)) {
+		/* Clear all reservations - everyone reacquires all resources */
+		for (i = 0; i < adev->gfx.num_gfx_rings; ++i)
+			gfx_v8_0_ring_set_pipe_percent(&adev->gfx.gfx_ring[i],
+						       true);
+
+		for (i = 0; i < adev->gfx.num_compute_rings; ++i)
+			gfx_v8_0_ring_set_pipe_percent(&adev->gfx.compute_ring[i],
+						       true);
+	} else {
+		/* Lower all pipes without a current reservation */
+		for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
+			iring = &adev->gfx.gfx_ring[i];
+			pipe = amdgpu_queue_to_bit(adev,
+						   iring->me,
+						   iring->pipe,
+						   0);
+			reserve = test_bit(pipe, adev->gfx.pipe_reserve_bitmap);
+			gfx_v8_0_ring_set_pipe_percent(iring, reserve);
+		}
+
+		for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
+			iring = &adev->gfx.compute_ring[i];
+			pipe = amdgpu_queue_to_bit(adev,
+						   iring->me,
+						   iring->pipe,
+						   0);
+			reserve = test_bit(pipe, adev->gfx.pipe_reserve_bitmap);
+			gfx_v8_0_ring_set_pipe_percent(iring, reserve);
+		}
+	}
+
+	mutex_unlock(&adev->gfx.pipe_reserve_mutex);
+}
+
+static void gfx_v8_0_hqd_set_priority(struct amdgpu_device *adev,
+				      struct amdgpu_ring *ring,
+				      enum amd_sched_priority priority)
+{
+	mutex_lock(&adev->srbm_mutex);
+	vi_srbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
+
+	switch (priority) {
+	case AMD_SCHED_PRIORITY_NORMAL:
+		WREG32(mmCP_HQD_PIPE_PRIORITY, 0x0);
+		WREG32(mmCP_HQD_QUEUE_PRIORITY, 0x0);
+		break;
+	case AMD_SCHED_PRIORITY_HIGH:
+		WREG32(mmCP_HQD_PIPE_PRIORITY, 0x2);
+		WREG32(mmCP_HQD_QUEUE_PRIORITY, 0xf);
+		break;
+	default:
+		WARN(1, "Attempt to set invalid SPI priority:%d for ring:%d\n",
+				priority, ring->idx);
+		break;
+	}
+
+	vi_srbm_select(adev, 0, 0, 0, 0);
+	mutex_unlock(&adev->srbm_mutex);
+}
+static void gfx_v8_0_ring_set_priority_compute(struct amdgpu_ring *ring,
+					       enum amd_sched_priority priority)
+{
+	struct amdgpu_device *adev = ring->adev;
+
+	if (ring->funcs->type != AMDGPU_RING_TYPE_COMPUTE)
+		return;
+
+	gfx_v8_0_hqd_set_priority(adev, ring, priority);
+	gfx_v8_0_pipe_reserve_resources(adev, ring, priority >= AMD_SCHED_PRIORITY_HIGH);
+}
+
 static void gfx_v8_0_ring_emit_fence_compute(struct amdgpu_ring *ring,
 					     u64 addr, u64 seq,
 					     unsigned flags)
@@ -7075,6 +7187,7 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_compute = {
 	.test_ib = gfx_v8_0_ring_test_ib,
 	.insert_nop = amdgpu_ring_insert_nop,
 	.pad_ib = amdgpu_ring_generic_pad_ib,
+	.set_priority = gfx_v8_0_ring_set_priority_compute,
 };
 
 static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_kiq = {