drm/radeon: adding synchronization for GRBM GFX

Implementing a lock for selecting and accessing shader engines and arrays.
This lock will make sure that radeon and amdkfd are not colliding when
accessing shader engines and arrays with GRBM_GFX_INDEX register.
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NOded Gabbay <oded.gabbay@amd.com>

drivers/gpu/drm/radeon/cik.c

@@ -1563,6 +1563,8 @@ static const u32 godavari_golden_registers[] =
 
 static void cik_init_golden_registers(struct radeon_device *rdev)
 {
+	/* Some of the registers might be dependent on GRBM_GFX_INDEX */
+	mutex_lock(&rdev->grbm_idx_mutex);
 	switch (rdev->family) {
 	case CHIP_BONAIRE:
 		radeon_program_register_sequence(rdev,
@@ -1637,6 +1639,7 @@ static void cik_init_golden_registers(struct radeon_device *rdev)
 	default:
 		break;
 	}
+	mutex_unlock(&rdev->grbm_idx_mutex);
 }
 
 /**
@@ -3428,6 +3431,7 @@ static void cik_setup_rb(struct radeon_device *rdev,
 	u32 disabled_rbs = 0;
 	u32 enabled_rbs = 0;
 
+	mutex_lock(&rdev->grbm_idx_mutex);
 	for (i = 0; i < se_num; i++) {
 		for (j = 0; j < sh_per_se; j++) {
 			cik_select_se_sh(rdev, i, j);
@@ -3439,6 +3443,7 @@ static void cik_setup_rb(struct radeon_device *rdev,
 		}
 	}
 	cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
+	mutex_unlock(&rdev->grbm_idx_mutex);
 
 	mask = 1;
 	for (i = 0; i < max_rb_num_per_se * se_num; i++) {
@@ -3449,6 +3454,7 @@ static void cik_setup_rb(struct radeon_device *rdev,
 
 	rdev->config.cik.backend_enable_mask = enabled_rbs;
 
+	mutex_lock(&rdev->grbm_idx_mutex);
 	for (i = 0; i < se_num; i++) {
 		cik_select_se_sh(rdev, i, 0xffffffff);
 		data = 0;
@@ -3476,6 +3482,7 @@ static void cik_setup_rb(struct radeon_device *rdev,
 		WREG32(PA_SC_RASTER_CONFIG, data);
 	}
 	cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
+	mutex_unlock(&rdev->grbm_idx_mutex);
 }
 
 /**
@@ -3693,6 +3700,12 @@ static void cik_gpu_init(struct radeon_device *rdev)
 	/* set HW defaults for 3D engine */
 	WREG32(CP_MEQ_THRESHOLDS, MEQ1_START(0x30) | MEQ2_START(0x60));
 
+	mutex_lock(&rdev->grbm_idx_mutex);
+	/*
+	 * making sure that the following register writes will be broadcasted
+	 * to all the shaders
+	 */
+	cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
 	WREG32(SX_DEBUG_1, 0x20);
 
 	WREG32(TA_CNTL_AUX, 0x00010000);
@@ -3748,6 +3761,7 @@ static void cik_gpu_init(struct radeon_device *rdev)
 
 	WREG32(PA_CL_ENHANCE, CLIP_VTX_REORDER_ENA | NUM_CLIP_SEQ(3));
 	WREG32(PA_SC_ENHANCE, ENABLE_PA_SC_OUT_OF_ORDER);
+	mutex_unlock(&rdev->grbm_idx_mutex);
 
 	udelay(50);
 }
@@ -6068,6 +6082,7 @@ static void cik_wait_for_rlc_serdes(struct radeon_device *rdev)
 	u32 i, j, k;
 	u32 mask;
 
+	mutex_lock(&rdev->grbm_idx_mutex);
 	for (i = 0; i < rdev->config.cik.max_shader_engines; i++) {
 		for (j = 0; j < rdev->config.cik.max_sh_per_se; j++) {
 			cik_select_se_sh(rdev, i, j);
@@ -6079,6 +6094,7 @@ static void cik_wait_for_rlc_serdes(struct radeon_device *rdev)
 		}
 	}
 	cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
+	mutex_unlock(&rdev->grbm_idx_mutex);
 
 	mask = SE_MASTER_BUSY_MASK | GC_MASTER_BUSY | TC0_MASTER_BUSY | TC1_MASTER_BUSY;
 	for (k = 0; k < rdev->usec_timeout; k++) {
@@ -6213,10 +6229,12 @@ static int cik_rlc_resume(struct radeon_device *rdev)
 	WREG32(RLC_LB_CNTR_INIT, 0);
 	WREG32(RLC_LB_CNTR_MAX, 0x00008000);
 
+	mutex_lock(&rdev->grbm_idx_mutex);
 	cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
 	WREG32(RLC_LB_INIT_CU_MASK, 0xffffffff);
 	WREG32(RLC_LB_PARAMS, 0x00600408);
 	WREG32(RLC_LB_CNTL, 0x80000004);
+	mutex_unlock(&rdev->grbm_idx_mutex);
 
 	WREG32(RLC_MC_CNTL, 0);
 	WREG32(RLC_UCODE_CNTL, 0);
@@ -6283,11 +6301,13 @@ static void cik_enable_cgcg(struct radeon_device *rdev, bool enable)
 
 		tmp = cik_halt_rlc(rdev);
 
+		mutex_lock(&rdev->grbm_idx_mutex);
 		cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
 		WREG32(RLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
 		WREG32(RLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
 		tmp2 = BPM_ADDR_MASK | CGCG_OVERRIDE_0 | CGLS_ENABLE;
 		WREG32(RLC_SERDES_WR_CTRL, tmp2);
+		mutex_unlock(&rdev->grbm_idx_mutex);
 
 		cik_update_rlc(rdev, tmp);
 
@@ -6329,11 +6349,13 @@ static void cik_enable_mgcg(struct radeon_device *rdev, bool enable)
 
 		tmp = cik_halt_rlc(rdev);
 
+		mutex_lock(&rdev->grbm_idx_mutex);
 		cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
 		WREG32(RLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
 		WREG32(RLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
 		data = BPM_ADDR_MASK | MGCG_OVERRIDE_0;
 		WREG32(RLC_SERDES_WR_CTRL, data);
+		mutex_unlock(&rdev->grbm_idx_mutex);
 
 		cik_update_rlc(rdev, tmp);
 
@@ -6377,11 +6399,13 @@ static void cik_enable_mgcg(struct radeon_device *rdev, bool enable)
 
 		tmp = cik_halt_rlc(rdev);
 
+		mutex_lock(&rdev->grbm_idx_mutex);
 		cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
 		WREG32(RLC_SERDES_WR_CU_MASTER_MASK, 0xffffffff);
 		WREG32(RLC_SERDES_WR_NONCU_MASTER_MASK, 0xffffffff);
 		data = BPM_ADDR_MASK | MGCG_OVERRIDE_1;
 		WREG32(RLC_SERDES_WR_CTRL, data);
+		mutex_unlock(&rdev->grbm_idx_mutex);
 
 		cik_update_rlc(rdev, tmp);
 	}
@@ -6810,10 +6834,12 @@ static u32 cik_get_cu_active_bitmap(struct radeon_device *rdev, u32 se, u32 sh)
 	u32 mask = 0, tmp, tmp1;
 	int i;
 
+	mutex_lock(&rdev->grbm_idx_mutex);
 	cik_select_se_sh(rdev, se, sh);
 	tmp = RREG32(CC_GC_SHADER_ARRAY_CONFIG);
 	tmp1 = RREG32(GC_USER_SHADER_ARRAY_CONFIG);
 	cik_select_se_sh(rdev, 0xffffffff, 0xffffffff);
+	mutex_unlock(&rdev->grbm_idx_mutex);
 
 	tmp &= 0xffff0000;
 

drivers/gpu/drm/radeon/radeon.h

@@ -2397,6 +2397,8 @@ struct radeon_device {
 	struct radeon_atcs		atcs;
 	/* srbm instance registers */
 	struct mutex			srbm_mutex;
+	/* GRBM index mutex. Protects concurrents access to GRBM index */
+	struct mutex			grbm_idx_mutex;
 	/* clock, powergating flags */
 	u32 cg_flags;
 	u32 pg_flags;

drivers/gpu/drm/radeon/radeon_device.c

@@ -1303,6 +1303,7 @@ int radeon_device_init(struct radeon_device *rdev,
 	mutex_init(&rdev->pm.mutex);
 	mutex_init(&rdev->gpu_clock_mutex);
 	mutex_init(&rdev->srbm_mutex);
+	mutex_init(&rdev->grbm_idx_mutex);
 	init_rwsem(&rdev->pm.mclk_lock);
 	init_rwsem(&rdev->exclusive_lock);
 	init_waitqueue_head(&rdev->irq.vblank_queue);