On 9/17/25 11:42 AM, Alex Deucher wrote:
When in S0i3, the GFX state is retained, so all we need to do
is stop the runlist so GFX can enter gfxoff.
Signed-off-by: Alex Deucher <alexander.deuc...@amd.com>
Confirmed this doesn't run into the problem we hit with GFX10 on my
attempt. Also it fixes the problem properly on GFX11.
For all 3 patches:
Reviewed-by: Mario Limonciello (AMD) <supe...@kernel.org>
Tested-By: David Perry <david.pe...@amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 16 +++++++---
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h | 12 ++++++++
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 36 ++++++++++++++++++++++
3 files changed, 60 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
index fbe7616555c83..a2879d2b7c8ec 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
@@ -250,16 +250,24 @@ void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
void amdgpu_amdkfd_suspend(struct amdgpu_device *adev, bool suspend_proc)
{
- if (adev->kfd.dev)
- kgd2kfd_suspend(adev->kfd.dev, suspend_proc);
+ if (adev->kfd.dev) {
+ if (adev->in_s0ix)
+ kgd2kfd_stop_sched_all_nodes(adev->kfd.dev);
+ else
+ kgd2kfd_suspend(adev->kfd.dev, suspend_proc);
+ }
}
int amdgpu_amdkfd_resume(struct amdgpu_device *adev, bool resume_proc)
{
int r = 0;
- if (adev->kfd.dev)
- r = kgd2kfd_resume(adev->kfd.dev, resume_proc);
+ if (adev->kfd.dev) {
+ if (adev->in_s0ix)
+ r = kgd2kfd_start_sched_all_nodes(adev->kfd.dev);
+ else
+ r = kgd2kfd_resume(adev->kfd.dev, resume_proc);
+ }
return r;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 127927b16ee29..9e120c934cc17 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -428,7 +428,9 @@ void kgd2kfd_smi_event_throttle(struct kfd_dev *kfd,
uint64_t throttle_bitmask);
int kgd2kfd_check_and_lock_kfd(struct kfd_dev *kfd);
void kgd2kfd_unlock_kfd(struct kfd_dev *kfd);
int kgd2kfd_start_sched(struct kfd_dev *kfd, uint32_t node_id);
+int kgd2kfd_start_sched_all_nodes(struct kfd_dev *kfd);
int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t node_id);
+int kgd2kfd_stop_sched_all_nodes(struct kfd_dev *kfd);
bool kgd2kfd_compute_active(struct kfd_dev *kfd, uint32_t node_id);
bool kgd2kfd_vmfault_fast_path(struct amdgpu_device *adev, struct
amdgpu_iv_entry *entry,
bool retry_fault);
@@ -518,11 +520,21 @@ static inline int kgd2kfd_start_sched(struct kfd_dev
*kfd, uint32_t node_id)
return 0;
}
+static inline int kgd2kfd_start_sched_all_nodes(struct kfd_dev *kfd)
+{
+ return 0;
+}
+
static inline int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t node_id)
{
return 0;
}
+static inline int kgd2kfd_stop_sched_all_nodes(struct kfd_dev *kfd)
+{
+ return 0;
+}
+
static inline bool kgd2kfd_compute_active(struct kfd_dev *kfd, uint32_t
node_id)
{
return false;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 7e749f9b6d69d..349c351e242b5 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -1550,6 +1550,25 @@ int kgd2kfd_start_sched(struct kfd_dev *kfd, uint32_t
node_id)
return ret;
}
+int kgd2kfd_start_sched_all_nodes(struct kfd_dev *kfd)
+{
+ struct kfd_node *node;
+ int i, r;
+
+ if (!kfd->init_complete)
+ return 0;
+
+ for (i = 0; i < kfd->num_nodes; i++) {
+ node = kfd->nodes[i];
+ r = node->dqm->ops.unhalt(node->dqm);
+ if (r) {
+ dev_err(kfd_device, "Error in starting scheduler\n");
+ return r;
+ }
+ }
+ return 0;
+}
+
int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t node_id)
{
struct kfd_node *node;
@@ -1567,6 +1586,23 @@ int kgd2kfd_stop_sched(struct kfd_dev *kfd, uint32_t
node_id)
return node->dqm->ops.halt(node->dqm);
}
+int kgd2kfd_stop_sched_all_nodes(struct kfd_dev *kfd)
+{
+ struct kfd_node *node;
+ int i, r;
+
+ if (!kfd->init_complete)
+ return 0;
+
+ for (i = 0; i < kfd->num_nodes; i++) {
+ node = kfd->nodes[i];
+ r = node->dqm->ops.halt(node->dqm);
+ if (r)
+ return r;
+ }
+ return 0;
+}
+
bool kgd2kfd_compute_active(struct kfd_dev *kfd, uint32_t node_id)
{
struct kfd_node *node;
