]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/amdgpu: perform p-state switch after the whole hive initialized
authorEvan Quan <evan.quan@amd.com>
Tue, 5 Nov 2019 07:15:33 +0000 (15:15 +0800)
committerAlex Deucher <alexander.deucher@amd.com>
Wed, 6 Nov 2019 21:27:48 +0000 (16:27 -0500)
P-state switch should be performed after all devices from the hive
get initialized.

Signed-off-by: Evan Quan <evan.quan@amd.com>
Reviewed-by: Feifei Xu <Feifei.Xu@amd.com>
Reviewed-by: Jonathan Kim <Jonathan.Kim@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

index 7b33058c4e03355bef8c9e963345338e51e016b0..58f6b3b92831b4065b6a3c6568556e9cea50aa30 100644 (file)
@@ -2057,6 +2057,7 @@ out:
  */
 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
 {
+       struct amdgpu_gpu_instance *gpu_instance;
        int i = 0, r;
 
        for (i = 0; i < adev->num_ip_blocks; i++) {
@@ -2082,6 +2083,40 @@ static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
        if (r)
                DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
 
+
+       if (adev->gmc.xgmi.num_physical_nodes > 1) {
+               mutex_lock(&mgpu_info.mutex);
+
+               /*
+                * Reset device p-state to low as this was booted with high.
+                *
+                * This should be performed only after all devices from the same
+                * hive get initialized.
+                *
+                * However, it's unknown how many device in the hive in advance.
+                * As this is counted one by one during devices initializations.
+                *
+                * So, we wait for all XGMI interlinked devices initialized.
+                * This may bring some delays as those devices may come from
+                * different hives. But that should be OK.
+                */
+               if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
+                       for (i = 0; i < mgpu_info.num_gpu; i++) {
+                               gpu_instance = &(mgpu_info.gpu_ins[i]);
+                               if (gpu_instance->adev->flags & AMD_IS_APU)
+                                       continue;
+
+                               r = amdgpu_xgmi_set_pstate(gpu_instance->adev, 0);
+                               if (r) {
+                                       DRM_ERROR("pstate setting failed (%d).\n", r);
+                                       break;
+                               }
+                       }
+               }
+
+               mutex_unlock(&mgpu_info.mutex);
+       }
+
        return 0;
 }
 
@@ -2194,18 +2229,6 @@ static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
        r = amdgpu_ib_ring_tests(adev);
        if (r)
                DRM_ERROR("ib ring test failed (%d).\n", r);
-
-       /*
-        * set to low pstate by default
-        * This should be performed after all devices from
-        * XGMI finish their initializations. Thus it's moved
-        * to here.
-        * The time delay is 2S. TODO: confirm whether that
-        * is enough for all possible XGMI setups.
-        */
-       r = amdgpu_xgmi_set_pstate(adev, 0);
-       if (r)
-               DRM_ERROR("pstate setting failed (%d).\n", r);
 }
 
 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)