[PATCH v2 4/5] Subject: drm/amdgpu: Redo XGMI reset synchronization.
Andrey Grodzovsky
andrey.grodzovsky at amd.com
Fri Dec 13 16:54:08 UTC 2019
Use task barrier in XGMI hive to synchronize ASIC resets
across devices in XGMI hive.
v2: Retrun right away with a warning if no xgmi hive, update doc.
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky at amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 37 +++++++++++++++++++++++++-----
1 file changed, 31 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 1d19edfa..2ae944c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -67,6 +67,7 @@
#include "amdgpu_tmz.h"
#include <linux/suspend.h>
+#include <drm/task_barrier.h>
MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
@@ -2663,14 +2664,38 @@ static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
{
struct amdgpu_device *adev =
container_of(__work, struct amdgpu_device, xgmi_reset_work);
+ struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0);
- if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO)
- adev->asic_reset_res = (adev->in_baco == false) ?
- amdgpu_device_baco_enter(adev->ddev) :
- qamdgpu_device_baco_exit(adev->ddev);
- else
- adev->asic_reset_res = amdgpu_asic_reset(adev);
+ /* It's a bug to not have a hive within this function */
+ if (WARN_ON(!hive))
+ return;
+
+ /*
+ * Use task barrier to synchronize all xgmi reset works across the
+ * hive. task_barrier_enter and task_barrier_exit will block
+ * until all the threads running the xgmi reset works reach
+ * those points. task_barrier_full will do both blocks.
+ */
+ if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
+
+ task_barrier_enter(&hive->tb);
+ adev->asic_reset_res = amdgpu_device_baco_enter(adev->ddev);
+
+ if (adev->asic_reset_res)
+ goto fail;
+
+ task_barrier_exit(&hive->tb);
+ adev->asic_reset_res = amdgpu_device_baco_exit(adev->ddev);
+
+ if (adev->asic_reset_res)
+ goto fail;
+ } else {
+
+ task_barrier_full(&hive->tb);
+ adev->asic_reset_res = amdgpu_asic_reset(adev);
+ }
+fail:
if (adev->asic_reset_res)
DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
adev->asic_reset_res, adev->ddev->unique);
--
2.7.4
More information about the amd-gfx
mailing list