On 2022-03-05 13:40, Christian König
wrote:
Am 04.03.22 um 18:10 schrieb Andrey Grodzovsky:
On 2022-03-03 03:23, Christian König wrote:
Allows submitting jobs as gang which needs to run on multiple
engines at the same time.
Basic idea is that we have a global gang submit fence representing when the
gang leader is finally pushed to run on the hardware last.
Jobs submitted as gang are never re-submitted in case of a GPU reset since this
won't work and will just deadlock the hardware immediately again.
Signed-off-by: Christian König <christian.koenig@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu.h | 3 ++
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 34 ++++++++++++++++++++++
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 28 ++++++++++++++++--
drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 3 ++
4 files changed, 66 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
index 7f447ed7a67f..a664d43d7502 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
@@ -852,6 +852,7 @@ struct amdgpu_device {
u64 fence_context;
unsigned num_rings;
struct amdgpu_ring *rings[AMDGPU_MAX_RINGS];
+ struct dma_fence __rcu *gang_submit;
bool ib_pool_ready;
struct amdgpu_sa_manager ib_pools[AMDGPU_IB_POOL_MAX];
struct amdgpu_sched gpu_sched[AMDGPU_HW_IP_NUM][AMDGPU_RING_PRIO_MAX];
@@ -1233,6 +1234,8 @@ void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev,
struct amdgpu_ring *ring);
void amdgpu_device_halt(struct amdgpu_device *adev);
+struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
+ struct dma_fence *gang);
/* atpx handler */
#if defined(CONFIG_VGA_SWITCHEROO)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index d78141e2c509..a116b8c08827 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -3512,6 +3512,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
adev->gmc.gart_size = 512 * 1024 * 1024;
adev->accel_working = false;
adev->num_rings = 0;
+ RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub());
adev->mman.buffer_funcs = NULL;
adev->mman.buffer_funcs_ring = NULL;
adev->vm_manager.vm_pte_funcs = NULL;
@@ -3989,6 +3990,7 @@ void amdgpu_device_fini_sw(struct amdgpu_device *adev)
release_firmware(adev->firmware.gpu_info_fw);
adev->firmware.gpu_info_fw = NULL;
adev->accel_working = false;
+ dma_fence_put(rcu_dereference_protected(adev->gang_submit, true));
amdgpu_reset_fini(adev);
@@ -5744,3 +5746,35 @@ void amdgpu_device_halt(struct amdgpu_device *adev)
pci_disable_device(pdev);
pci_wait_for_pending_transaction(pdev);
}
+
+/**
+ * amdgpu_device_switch_gang - switch to a new gang
+ * @adev: amdgpu_device pointer
+ * @gang: the gang to switch to
+ *
+ * Try to switch to a new gang or return a reference to the current gang if that
+ * isn't possible.
+ * Returns: Either NULL if we switched correctly or a reference to the existing
+ * gang.
+ */
+struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
+ struct dma_fence *gang)
+{
+ struct dma_fence *old = NULL;
+
+ do {
+ dma_fence_put(old);
+ old = dma_fence_get_rcu_safe(&adev->gang_submit);
+
+ if (old == gang)
+ break;
+
+ if (!dma_fence_is_signaled(old))
+ return old;
+
+ } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,
+ old, gang) != old);
+
+ dma_fence_put(old);
+ return NULL;
+}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index e07ceae36a5c..059e11c7898c 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -169,11 +169,29 @@ static void amdgpu_job_free_cb(struct drm_sched_job *s_job)
kfree(job);
}
+void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
+ struct amdgpu_job *leader)
+{
+ struct dma_fence *fence = &leader->base.s_fence->scheduled;
+
+ WARN_ON(job->gang_submit);
+
+ /*
+ * Don't add a reference when we are the gang leader to avoid circle
+ * dependency.
+ */
+ if (job != leader)
+ dma_fence_get(fence);
+ job->gang_submit = fence;
+}
+
void amdgpu_job_free(struct amdgpu_job *job)
{
amdgpu_job_free_resources(job);
amdgpu_sync_free(&job->sync);
amdgpu_sync_free(&job->sched_sync);
+ if (job->gang_submit != &job->base.s_fence->scheduled)
+ dma_fence_put(job->gang_submit);
/* only put the hw fence if has embedded fence */
if (job->hw_fence.ops != NULL)
@@ -247,12 +265,16 @@ static struct dma_fence *amdgpu_job_dependency(struct drm_sched_job *sched_job,
fence = amdgpu_sync_get_fence(&job->sync);
}
+ if (!fence && !job->gang_submit)
+ fence = amdgpu_device_switch_gang(ring->adev, job->gang_submit);
+
Why job->gang_submit should be NULL in the check above ? Don't you want to switch to an actual new gang fence here ?
Jobs that don't have gang_submit fence set are not gang jobs anyway and we don't care for this dependency
for them right ?
Well exactly that's the point. That a job is not part of a gang submit is signaled by setting the pointer to NULL.
No dispute on this
If we don't check for NULL here we would just crash.
But you go into the 'if clause' if job->gang_submit is equal
to NULL, i would expect to see here
if (!fence && job->gang_submit) because you want
to switch to an actual new gang (not to NULL)
Andrey
Christian.
Andrey
return fence;
}
static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
{
struct amdgpu_ring *ring = to_amdgpu_ring(sched_job->sched);
+ struct amdgpu_device *adev = ring->adev;
struct dma_fence *fence = NULL, *finished;
struct amdgpu_job *job;
int r = 0;
@@ -264,8 +286,10 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
trace_amdgpu_sched_run_job(job);
- if (job->vram_lost_counter != atomic_read(&ring->adev->vram_lost_counter))
- dma_fence_set_error(finished, -ECANCELED);/* skip IB as well if VRAM lost */
+ /* Skip job if VRAM is lost and never resubmit gangs */
+ if (job->vram_lost_counter != atomic_read(&adev->vram_lost_counter) ||
+ (job->job_run_counter && job->gang_submit))
+ dma_fence_set_error(finished, -ECANCELED);
if (finished->error < 0) {
DRM_INFO("Skip scheduling IBs!\n");
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
index 0bab8fe0d419..615328130615 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.h
@@ -51,6 +51,7 @@ struct amdgpu_job {
struct amdgpu_sync sched_sync;
struct dma_fence hw_fence;
struct dma_fence *external_hw_fence;
+ struct dma_fence *gang_submit;
uint32_t preamble_status;
uint32_t preemption_status;
bool vm_needs_flush;
@@ -80,6 +81,8 @@ int amdgpu_job_alloc_with_ib(struct amdgpu_device *adev, unsigned size,
void amdgpu_job_set_resources(struct amdgpu_job *job, struct amdgpu_bo *gds,
struct amdgpu_bo *gws, struct amdgpu_bo *oa);
void amdgpu_job_free_resources(struct amdgpu_job *job);
+void amdgpu_job_set_gang_leader(struct amdgpu_job *job,
+ struct amdgpu_job *leader);
void amdgpu_job_free(struct amdgpu_job *job);
int amdgpu_job_submit(struct amdgpu_job *job, struct drm_sched_entity *entity,
void *owner, struct dma_fence **f);
