1 files changed, 201 insertions, 36 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
index 1d155463d044..72af5e5a894a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
@@ -515,7 +515,7 @@ int amdgpu_gfx_disable_kcq(struct amdgpu_device *adev, int xcc_id)
 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
 		return -EINVAL;
 
-	if (!kiq_ring->sched.ready || adev->job_hang || amdgpu_in_reset(adev))
+	if (!kiq_ring->sched.ready || amdgpu_in_reset(adev))
 		return 0;
 
 	spin_lock(&kiq->ring_lock);
@@ -567,7 +567,7 @@ int amdgpu_gfx_disable_kgq(struct amdgpu_device *adev, int xcc_id)
 	if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
 		return -EINVAL;
 
-	if (!adev->gfx.kiq[0].ring.sched.ready || adev->job_hang)
+	if (!adev->gfx.kiq[0].ring.sched.ready || amdgpu_in_reset(adev))
 		return 0;
 
 	if (amdgpu_gfx_is_master_xcc(adev, xcc_id)) {
@@ -771,18 +771,8 @@ int amdgpu_gfx_enable_kgq(struct amdgpu_device *adev, int xcc_id)
 	return r;
 }
 
-/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
- *
- * @adev: amdgpu_device pointer
- * @bool enable true: enable gfx off feature, false: disable gfx off feature
- *
- * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
- * 2. other client can send request to disable gfx off feature, the request should be honored.
- * 3. other client can cancel their request of disable gfx off feature
- * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
- */
-
-void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
+static void amdgpu_gfx_do_off_ctrl(struct amdgpu_device *adev, bool enable,
+				   bool no_delay)
 {
 	unsigned long delay = GFX_OFF_DELAY_ENABLE;
 
@@ -804,9 +794,9 @@ void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
 		if (adev->gfx.gfx_off_req_count == 0 &&
 		    !adev->gfx.gfx_off_state) {
 			/* If going to s2idle, no need to wait */
-			if (adev->in_s0ix) {
+			if (no_delay) {
 				if (!amdgpu_dpm_set_powergating_by_smu(adev,
-						AMD_IP_BLOCK_TYPE_GFX, true))
+						AMD_IP_BLOCK_TYPE_GFX, true, 0))
 					adev->gfx.gfx_off_state = true;
 			} else {
 				schedule_delayed_work(&adev->gfx.gfx_off_delay_work,
@@ -818,7 +808,7 @@ void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
 			cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
 
 			if (adev->gfx.gfx_off_state &&
-			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
+			    !amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false, 0)) {
 				adev->gfx.gfx_off_state = false;
 
 				if (adev->gfx.funcs->init_spm_golden) {
@@ -836,6 +826,43 @@ unlock:
 	mutex_unlock(&adev->gfx.gfx_off_mutex);
 }
 
+/* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
+ *
+ * @adev: amdgpu_device pointer
+ * @bool enable true: enable gfx off feature, false: disable gfx off feature
+ *
+ * 1. gfx off feature will be enabled by gfx ip after gfx cg pg enabled.
+ * 2. other client can send request to disable gfx off feature, the request should be honored.
+ * 3. other client can cancel their request of disable gfx off feature
+ * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
+ *
+ * gfx off allow will be delayed by GFX_OFF_DELAY_ENABLE ms.
+ */
+void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
+{
+	/* If going to s2idle, no need to wait */
+	bool no_delay = adev->in_s0ix ? true : false;
+
+	amdgpu_gfx_do_off_ctrl(adev, enable, no_delay);
+}
+
+/* amdgpu_gfx_off_ctrl_immediate - Handle gfx off feature enable/disable
+ *
+ * @adev: amdgpu_device pointer
+ * @bool enable true: enable gfx off feature, false: disable gfx off feature
+ *
+ * 1. gfx off feature will be enabled by gfx ip after gfx cg pg enabled.
+ * 2. other client can send request to disable gfx off feature, the request should be honored.
+ * 3. other client can cancel their request of disable gfx off feature
+ * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
+ *
+ * gfx off allow will be issued immediately.
+ */
+void amdgpu_gfx_off_ctrl_immediate(struct amdgpu_device *adev, bool enable)
+{
+	amdgpu_gfx_do_off_ctrl(adev, enable, true);
+}
+
 int amdgpu_set_gfx_off_residency(struct amdgpu_device *adev, bool value)
 {
 	int r = 0;
@@ -1484,6 +1511,24 @@ static int amdgpu_gfx_run_cleaner_shader(struct amdgpu_device *adev, int xcp_id)
 	return 0;
 }
 
+/**
+ * amdgpu_gfx_set_run_cleaner_shader - Execute the AMDGPU GFX Cleaner Shader
+ * @dev: The device structure
+ * @attr: The device attribute structure
+ * @buf: The buffer containing the input data
+ * @count: The size of the input data
+ *
+ * Provides the sysfs interface to manually run a cleaner shader, which is
+ * used to clear the GPU state between different tasks. Writing a value to the
+ * 'run_cleaner_shader' sysfs file triggers the cleaner shader execution.
+ * The value written corresponds to the partition index on multi-partition
+ * devices. On single-partition devices, the value should be '0'.
+ *
+ * The cleaner shader clears the Local Data Store (LDS) and General Purpose
+ * Registers (GPRs) to ensure data isolation between GPU workloads.
+ *
+ * Return: The number of bytes written to the sysfs file.
+ */
 static ssize_t amdgpu_gfx_set_run_cleaner_shader(struct device *dev,
 						 struct device_attribute *attr,
 						 const char *buf,
@@ -1532,6 +1577,19 @@ static ssize_t amdgpu_gfx_set_run_cleaner_shader(struct device *dev,
 	return count;
 }
 
+/**
+ * amdgpu_gfx_get_enforce_isolation - Query AMDGPU GFX Enforce Isolation Settings
+ * @dev: The device structure
+ * @attr: The device attribute structure
+ * @buf: The buffer to store the output data
+ *
+ * Provides the sysfs read interface to get the current settings of the 'enforce_isolation'
+ * feature for each GPU partition. Reading from the 'enforce_isolation'
+ * sysfs file returns the isolation settings for all partitions, where '0'
+ * indicates disabled and '1' indicates enabled.
+ *
+ * Return: The number of bytes read from the sysfs file.
+ */
 static ssize_t amdgpu_gfx_get_enforce_isolation(struct device *dev,
 						struct device_attribute *attr,
 						char *buf)
@@ -1555,6 +1613,20 @@ static ssize_t amdgpu_gfx_get_enforce_isolation(struct device *dev,
 	return size;
 }
 
+/**
+ * amdgpu_gfx_set_enforce_isolation - Control AMDGPU GFX Enforce Isolation
+ * @dev: The device structure
+ * @attr: The device attribute structure
+ * @buf: The buffer containing the input data
+ * @count: The size of the input data
+ *
+ * This function allows control over the 'enforce_isolation' feature, which
+ * serializes access to the graphics engine. Writing '1' or '0' to the
+ * 'enforce_isolation' sysfs file enables or disables process isolation for
+ * each partition. The input should specify the setting for all partitions.
+ *
+ * Return: The number of bytes written to the sysfs file.
+ */
 static ssize_t amdgpu_gfx_set_enforce_isolation(struct device *dev,
 						struct device_attribute *attr,
 						const char *buf, size_t count)
@@ -1593,22 +1665,12 @@ static ssize_t amdgpu_gfx_set_enforce_isolation(struct device *dev,
 	}
 
 	mutex_lock(&adev->enforce_isolation_mutex);
-
-	for (i = 0; i < num_partitions; i++) {
-		if (adev->enforce_isolation[i] && !partition_values[i]) {
-			/* Going from enabled to disabled */
-			amdgpu_vmid_free_reserved(adev, AMDGPU_GFXHUB(i));
-			amdgpu_mes_set_enforce_isolation(adev, i, false);
-		} else if (!adev->enforce_isolation[i] && partition_values[i]) {
-			/* Going from disabled to enabled */
-			amdgpu_vmid_alloc_reserved(adev, AMDGPU_GFXHUB(i));
-			amdgpu_mes_set_enforce_isolation(adev, i, true);
-		}
+	for (i = 0; i < num_partitions; i++)
 		adev->enforce_isolation[i] = partition_values[i];
-	}
-
 	mutex_unlock(&adev->enforce_isolation_mutex);
 
+	amdgpu_mes_update_enforce_isolation(adev);
+
 	return count;
 }
 
@@ -1933,13 +1995,24 @@ void amdgpu_gfx_enforce_isolation_handler(struct work_struct *work)
 		if (adev->kfd.init_complete) {
 			WARN_ON_ONCE(!adev->gfx.kfd_sch_inactive[idx]);
 			WARN_ON_ONCE(adev->gfx.kfd_sch_req_count[idx]);
-				amdgpu_amdkfd_start_sched(adev, idx);
-				adev->gfx.kfd_sch_inactive[idx] = false;
+			amdgpu_amdkfd_start_sched(adev, idx);
+			adev->gfx.kfd_sch_inactive[idx] = false;
 		}
 	}
 	mutex_unlock(&adev->enforce_isolation_mutex);
 }
 
+/**
+ * amdgpu_gfx_enforce_isolation_wait_for_kfd - Manage KFD wait period for process isolation
+ * @adev: amdgpu_device pointer
+ * @idx: Index of the GPU partition
+ *
+ * When kernel submissions come in, the jobs are given a time slice and once
+ * that time slice is up, if there are KFD user queues active, kernel
+ * submissions are blocked until KFD has had its time slice. Once the KFD time
+ * slice is up, KFD user queues are preempted and kernel submissions are
+ * unblocked and allowed to run again.
+ */
 static void
 amdgpu_gfx_enforce_isolation_wait_for_kfd(struct amdgpu_device *adev,
 					  u32 idx)
@@ -1985,6 +2058,15 @@ amdgpu_gfx_enforce_isolation_wait_for_kfd(struct amdgpu_device *adev,
 		msleep(GFX_SLICE_PERIOD_MS);
 }
 
+/**
+ * amdgpu_gfx_enforce_isolation_ring_begin_use - Begin use of a ring with enforced isolation
+ * @ring: Pointer to the amdgpu_ring structure
+ *
+ * Ring begin_use helper implementation for gfx which serializes access to the
+ * gfx IP between kernel submission IOCTLs and KFD user queues when isolation
+ * enforcement is enabled. The kernel submission IOCTLs and KFD user queues
+ * each get a time slice when both are active.
+ */
 void amdgpu_gfx_enforce_isolation_ring_begin_use(struct amdgpu_ring *ring)
 {
 	struct amdgpu_device *adev = ring->adev;
@@ -2016,6 +2098,15 @@ void amdgpu_gfx_enforce_isolation_ring_begin_use(struct amdgpu_ring *ring)
 		amdgpu_gfx_kfd_sch_ctrl(adev, idx, false);
 }
 
+/**
+ * amdgpu_gfx_enforce_isolation_ring_end_use - End use of a ring with enforced isolation
+ * @ring: Pointer to the amdgpu_ring structure
+ *
+ * Ring end_use helper implementation for gfx which serializes access to the
+ * gfx IP between kernel submission IOCTLs and KFD user queues when isolation
+ * enforcement is enabled. The kernel submission IOCTLs and KFD user queues
+ * each get a time slice when both are active.
+ */
 void amdgpu_gfx_enforce_isolation_ring_end_use(struct amdgpu_ring *ring)
 {
 	struct amdgpu_device *adev = ring->adev;
@@ -2044,6 +2135,80 @@ void amdgpu_gfx_enforce_isolation_ring_end_use(struct amdgpu_ring *ring)
 		amdgpu_gfx_kfd_sch_ctrl(adev, idx, true);
 }
 
+void amdgpu_gfx_profile_idle_work_handler(struct work_struct *work)
+{
+	struct amdgpu_device *adev =
+		container_of(work, struct amdgpu_device, gfx.idle_work.work);
+	enum PP_SMC_POWER_PROFILE profile;
+	u32 i, fences = 0;
+	int r;
+
+	if (adev->gfx.num_gfx_rings)
+		profile = PP_SMC_POWER_PROFILE_FULLSCREEN3D;
+	else
+		profile = PP_SMC_POWER_PROFILE_COMPUTE;
+
+	for (i = 0; i < AMDGPU_MAX_GFX_RINGS; ++i)
+		fences += amdgpu_fence_count_emitted(&adev->gfx.gfx_ring[i]);
+	for (i = 0; i < (AMDGPU_MAX_COMPUTE_RINGS * AMDGPU_MAX_GC_INSTANCES); ++i)
+		fences += amdgpu_fence_count_emitted(&adev->gfx.compute_ring[i]);
+	if (!fences && !atomic_read(&adev->gfx.total_submission_cnt)) {
+		mutex_lock(&adev->gfx.workload_profile_mutex);
+		if (adev->gfx.workload_profile_active) {
+			r = amdgpu_dpm_switch_power_profile(adev, profile, false);
+			if (r)
+				dev_warn(adev->dev, "(%d) failed to disable %s power profile mode\n", r,
+					 profile == PP_SMC_POWER_PROFILE_FULLSCREEN3D ?
+					 "fullscreen 3D" : "compute");
+			adev->gfx.workload_profile_active = false;
+		}
+		mutex_unlock(&adev->gfx.workload_profile_mutex);
+	} else {
+		schedule_delayed_work(&adev->gfx.idle_work, GFX_PROFILE_IDLE_TIMEOUT);
+	}
+}
+
+void amdgpu_gfx_profile_ring_begin_use(struct amdgpu_ring *ring)
+{
+	struct amdgpu_device *adev = ring->adev;
+	enum PP_SMC_POWER_PROFILE profile;
+	int r;
+
+	if (adev->gfx.num_gfx_rings)
+		profile = PP_SMC_POWER_PROFILE_FULLSCREEN3D;
+	else
+		profile = PP_SMC_POWER_PROFILE_COMPUTE;
+
+	atomic_inc(&adev->gfx.total_submission_cnt);
+
+	cancel_delayed_work_sync(&adev->gfx.idle_work);
+
+	/* We can safely return early here because we've cancelled the
+	 * the delayed work so there is no one else to set it to false
+	 * and we don't care if someone else sets it to true.
+	 */
+	if (adev->gfx.workload_profile_active)
+		return;
+
+	mutex_lock(&adev->gfx.workload_profile_mutex);
+	if (!adev->gfx.workload_profile_active) {
+		r = amdgpu_dpm_switch_power_profile(adev, profile, true);
+		if (r)
+			dev_warn(adev->dev, "(%d) failed to disable %s power profile mode\n", r,
+				 profile == PP_SMC_POWER_PROFILE_FULLSCREEN3D ?
+				 "fullscreen 3D" : "compute");
+		adev->gfx.workload_profile_active = true;
+	}
+	mutex_unlock(&adev->gfx.workload_profile_mutex);
+}
+
+void amdgpu_gfx_profile_ring_end_use(struct amdgpu_ring *ring)
+{
+	atomic_dec(&ring->adev->gfx.total_submission_cnt);
+
+	schedule_delayed_work(&ring->adev->gfx.idle_work, GFX_PROFILE_IDLE_TIMEOUT);
+}
+
 /*
  * debugfs for to enable/disable gfx job submission to specific core.
  */
@@ -2058,7 +2223,7 @@ static int amdgpu_debugfs_gfx_sched_mask_set(void *data, u64 val)
 	if (!adev)
 		return -ENODEV;
 
-	mask = (1 << adev->gfx.num_gfx_rings) - 1;
+	mask = (1ULL << adev->gfx.num_gfx_rings) - 1;
 	if ((val & mask) == 0)
 		return -EINVAL;
 
@@ -2086,7 +2251,7 @@ static int amdgpu_debugfs_gfx_sched_mask_get(void *data, u64 *val)
 	for (i = 0; i < adev->gfx.num_gfx_rings; ++i) {
 		ring = &adev->gfx.gfx_ring[i];
 		if (ring->sched.ready)
-			mask |= 1 << i;
+			mask |= 1ULL << i;
 	}
 
 	*val = mask;
@@ -2128,7 +2293,7 @@ static int amdgpu_debugfs_compute_sched_mask_set(void *data, u64 val)
 	if (!adev)
 		return -ENODEV;
 
-	mask = (1 << adev->gfx.num_compute_rings) - 1;
+	mask = (1ULL << adev->gfx.num_compute_rings) - 1;
 	if ((val & mask) == 0)
 		return -EINVAL;
 
@@ -2157,7 +2322,7 @@ static int amdgpu_debugfs_compute_sched_mask_get(void *data, u64 *val)
 	for (i = 0; i < adev->gfx.num_compute_rings; ++i) {
 		ring = &adev->gfx.compute_ring[i];
 		if (ring->sched.ready)
-			mask |= 1 << i;
+			mask |= 1ULL << i;
 	}
 
 	*val = mask;