summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorChristian König <christian.koenig@amd.com>2022-10-26 13:26:37 +0300
committerAlex Deucher <alexander.deucher@amd.com>2022-11-15 23:25:22 +0300
commit06a2d7cc3f0476be4682ef90eb09a28fa3daed37 (patch)
tree17c7f45421a398fbe7137fa627127321cf457ea8
parentd27252b5706e51188aed7647126e44dcf9e940c1 (diff)
downloadlinux-06a2d7cc3f0476be4682ef90eb09a28fa3daed37.tar.xz
drm/amdgpu: revert "implement tdr advanced mode"
This reverts commit e6c6338f393b74ac0b303d567bb918b44ae7ad75. This feature basically re-submits one job after another to figure out which one was the one causing a hang. This is obviously incompatible with gang-submit which requires that multiple jobs run at the same time. It's also absolutely not helpful to crash the hardware multiple times if a clean recovery is desired. For testing and debugging environments we should rather disable recovery alltogether to be able to inspect the state with a hw debugger. Additional to that the sw implementation is clearly buggy and causes reference count issues for the hardware fence. Signed-off-by: Christian König <christian.koenig@amd.com> Reviewed-by: Alex Deucher <alexander.deucher@amd.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_device.c103
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c2
-rw-r--r--drivers/gpu/drm/scheduler/sched_main.c58
-rw-r--r--include/drm/gpu_scheduler.h3
4 files changed, 10 insertions, 156 deletions
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index 947ed5d97942..2a8166fb69f2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -5088,94 +5088,6 @@ static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
return 0;
}
-static void amdgpu_device_recheck_guilty_jobs(
- struct amdgpu_device *adev, struct list_head *device_list_handle,
- struct amdgpu_reset_context *reset_context)
-{
- int i, r = 0;
-
- for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
- struct amdgpu_ring *ring = adev->rings[i];
- int ret = 0;
- struct drm_sched_job *s_job;
-
- if (!ring || !ring->sched.thread)
- continue;
-
- s_job = list_first_entry_or_null(&ring->sched.pending_list,
- struct drm_sched_job, list);
- if (s_job == NULL)
- continue;
-
- /* clear job's guilty and depend the folowing step to decide the real one */
- drm_sched_reset_karma(s_job);
- drm_sched_resubmit_jobs_ext(&ring->sched, 1);
-
- if (!s_job->s_fence->parent) {
- DRM_WARN("Failed to get a HW fence for job!");
- continue;
- }
-
- ret = dma_fence_wait_timeout(s_job->s_fence->parent, false, ring->sched.timeout);
- if (ret == 0) { /* timeout */
- DRM_ERROR("Found the real bad job! ring:%s, job_id:%llx\n",
- ring->sched.name, s_job->id);
-
-
- amdgpu_fence_driver_isr_toggle(adev, true);
-
- /* Clear this failed job from fence array */
- amdgpu_fence_driver_clear_job_fences(ring);
-
- amdgpu_fence_driver_isr_toggle(adev, false);
-
- /* Since the job won't signal and we go for
- * another resubmit drop this parent pointer
- */
- dma_fence_put(s_job->s_fence->parent);
- s_job->s_fence->parent = NULL;
-
- /* set guilty */
- drm_sched_increase_karma(s_job);
- amdgpu_reset_prepare_hwcontext(adev, reset_context);
-retry:
- /* do hw reset */
- if (amdgpu_sriov_vf(adev)) {
- amdgpu_virt_fini_data_exchange(adev);
- r = amdgpu_device_reset_sriov(adev, false);
- if (r)
- adev->asic_reset_res = r;
- } else {
- clear_bit(AMDGPU_SKIP_HW_RESET,
- &reset_context->flags);
- r = amdgpu_do_asic_reset(device_list_handle,
- reset_context);
- if (r && r == -EAGAIN)
- goto retry;
- }
-
- /*
- * add reset counter so that the following
- * resubmitted job could flush vmid
- */
- atomic_inc(&adev->gpu_reset_counter);
- continue;
- }
-
- /* got the hw fence, signal finished fence */
- atomic_dec(ring->sched.score);
- dma_fence_get(&s_job->s_fence->finished);
- dma_fence_signal(&s_job->s_fence->finished);
- dma_fence_put(&s_job->s_fence->finished);
-
- /* remove node from list and free the job */
- spin_lock(&ring->sched.job_list_lock);
- list_del_init(&s_job->list);
- spin_unlock(&ring->sched.job_list_lock);
- ring->sched.ops->free_job(s_job);
- }
-}
-
static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev)
{
struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
@@ -5196,7 +5108,6 @@ static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev)
}
-
/**
* amdgpu_device_gpu_recover - reset the asic and recover scheduler
*
@@ -5219,7 +5130,6 @@ int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
int i, r = 0;
bool need_emergency_restart = false;
bool audio_suspended = false;
- int tmp_vram_lost_counter;
bool gpu_reset_for_dev_remove = false;
gpu_reset_for_dev_remove =
@@ -5365,7 +5275,6 @@ retry: /* Rest of adevs pre asic reset from XGMI hive. */
amdgpu_device_stop_pending_resets(tmp_adev);
}
- tmp_vram_lost_counter = atomic_read(&((adev)->vram_lost_counter));
/* Actual ASIC resets if needed.*/
/* Host driver will handle XGMI hive reset for SRIOV */
if (amdgpu_sriov_vf(adev)) {
@@ -5390,18 +5299,6 @@ skip_hw_reset:
/* Post ASIC reset for all devs .*/
list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
- /*
- * Sometimes a later bad compute job can block a good gfx job as gfx
- * and compute ring share internal GC HW mutually. We add an additional
- * guilty jobs recheck step to find the real guilty job, it synchronously
- * submits and pends for the first job being signaled. If it gets timeout,
- * we identify it as a real guilty job.
- */
- if (amdgpu_gpu_recovery == 2 &&
- !(tmp_vram_lost_counter < atomic_read(&adev->vram_lost_counter)))
- amdgpu_device_recheck_guilty_jobs(
- tmp_adev, device_list_handle, reset_context);
-
for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
struct amdgpu_ring *ring = tmp_adev->rings[i];
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index 8281d37220cf..72b77b3ac4fb 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -533,7 +533,7 @@ module_param_named(compute_multipipe, amdgpu_compute_multipipe, int, 0444);
* DOC: gpu_recovery (int)
* Set to enable GPU recovery mechanism (1 = enable, 0 = disable). The default is -1 (auto, disabled except SRIOV).
*/
-MODULE_PARM_DESC(gpu_recovery, "Enable GPU recovery mechanism, (2 = advanced tdr mode, 1 = enable, 0 = disable, -1 = auto)");
+MODULE_PARM_DESC(gpu_recovery, "Enable GPU recovery mechanism, (1 = enable, 0 = disable, -1 = auto)");
module_param_named(gpu_recovery, amdgpu_gpu_recovery, int, 0444);
/**
diff --git a/drivers/gpu/drm/scheduler/sched_main.c b/drivers/gpu/drm/scheduler/sched_main.c
index d0ff9e11cb69..0150ac7fbf40 100644
--- a/drivers/gpu/drm/scheduler/sched_main.c
+++ b/drivers/gpu/drm/scheduler/sched_main.c
@@ -443,27 +443,6 @@ static void drm_sched_job_timedout(struct work_struct *work)
}
}
- /**
- * drm_sched_increase_karma - Update sched_entity guilty flag
- *
- * @bad: The job guilty of time out
- *
- * Increment on every hang caused by the 'bad' job. If this exceeds the hang
- * limit of the scheduler then the respective sched entity is marked guilty and
- * jobs from it will not be scheduled further
- */
-void drm_sched_increase_karma(struct drm_sched_job *bad)
-{
- drm_sched_increase_karma_ext(bad, 1);
-}
-EXPORT_SYMBOL(drm_sched_increase_karma);
-
-void drm_sched_reset_karma(struct drm_sched_job *bad)
-{
- drm_sched_increase_karma_ext(bad, 0);
-}
-EXPORT_SYMBOL(drm_sched_reset_karma);
-
/**
* drm_sched_stop - stop the scheduler
*
@@ -605,31 +584,14 @@ EXPORT_SYMBOL(drm_sched_start);
*/
void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched)
{
- drm_sched_resubmit_jobs_ext(sched, INT_MAX);
-}
-EXPORT_SYMBOL(drm_sched_resubmit_jobs);
-
-/**
- * drm_sched_resubmit_jobs_ext - helper to relunch certain number of jobs from mirror ring list
- *
- * @sched: scheduler instance
- * @max: job numbers to relaunch
- *
- */
-void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)
-{
struct drm_sched_job *s_job, *tmp;
uint64_t guilty_context;
bool found_guilty = false;
struct dma_fence *fence;
- int i = 0;
list_for_each_entry_safe(s_job, tmp, &sched->pending_list, list) {
struct drm_sched_fence *s_fence = s_job->s_fence;
- if (i >= max)
- break;
-
if (!found_guilty && atomic_read(&s_job->karma) > sched->hang_limit) {
found_guilty = true;
guilty_context = s_job->s_fence->scheduled.context;
@@ -639,7 +601,6 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)
dma_fence_set_error(&s_fence->finished, -ECANCELED);
fence = sched->ops->run_job(s_job);
- i++;
if (IS_ERR_OR_NULL(fence)) {
if (IS_ERR(fence))
@@ -655,7 +616,7 @@ void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max)
}
}
}
-EXPORT_SYMBOL(drm_sched_resubmit_jobs_ext);
+EXPORT_SYMBOL(drm_sched_resubmit_jobs);
/**
* drm_sched_job_init - init a scheduler job
@@ -1172,13 +1133,15 @@ void drm_sched_fini(struct drm_gpu_scheduler *sched)
EXPORT_SYMBOL(drm_sched_fini);
/**
- * drm_sched_increase_karma_ext - Update sched_entity guilty flag
+ * drm_sched_increase_karma - Update sched_entity guilty flag
*
* @bad: The job guilty of time out
- * @type: type for increase/reset karma
*
+ * Increment on every hang caused by the 'bad' job. If this exceeds the hang
+ * limit of the scheduler then the respective sched entity is marked guilty and
+ * jobs from it will not be scheduled further
*/
-void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type)
+void drm_sched_increase_karma(struct drm_sched_job *bad)
{
int i;
struct drm_sched_entity *tmp;
@@ -1190,10 +1153,7 @@ void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type)
* corrupt but keep in mind that kernel jobs always considered good.
*/
if (bad->s_priority != DRM_SCHED_PRIORITY_KERNEL) {
- if (type == 0)
- atomic_set(&bad->karma, 0);
- else if (type == 1)
- atomic_inc(&bad->karma);
+ atomic_inc(&bad->karma);
for (i = DRM_SCHED_PRIORITY_MIN; i < DRM_SCHED_PRIORITY_KERNEL;
i++) {
@@ -1204,7 +1164,7 @@ void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type)
if (bad->s_fence->scheduled.context ==
entity->fence_context) {
if (entity->guilty)
- atomic_set(entity->guilty, type);
+ atomic_set(entity->guilty, 1);
break;
}
}
@@ -1214,4 +1174,4 @@ void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type)
}
}
}
-EXPORT_SYMBOL(drm_sched_increase_karma_ext);
+EXPORT_SYMBOL(drm_sched_increase_karma);
diff --git a/include/drm/gpu_scheduler.h b/include/drm/gpu_scheduler.h
index ca11716d084a..0168ff469ae0 100644
--- a/include/drm/gpu_scheduler.h
+++ b/include/drm/gpu_scheduler.h
@@ -528,10 +528,7 @@ void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
-void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max);
void drm_sched_increase_karma(struct drm_sched_job *bad);
-void drm_sched_reset_karma(struct drm_sched_job *bad);
-void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
bool drm_sched_dependency_optimized(struct dma_fence* fence,
struct drm_sched_entity *entity);
void drm_sched_fault(struct drm_gpu_scheduler *sched);