[PATCH 08/31] drm/amdgpu: track ring state associated with a job
Alex Deucher
alexander.deucher at amd.com
Thu Jun 5 01:45:38 UTC 2025
We need to know the wptr and sequence number associated
with a job so that we can re-emit the unprocessed state
after a ring reset. Pre-allocate storage space for
the ring buffer contents and add a helper to save off
the unprocessed state so that it can be re-emitted
after the queue is reset.
Add a helper that ring reset callbacks can use to verify
that the ring has reset successfully and to reemit any
unprocessed ring contents from subsequent jobs.
Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 15 ++++
drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 13 +++-
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 11 ++-
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 92 +++++++++++++++++++++++
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 15 ++++
5 files changed, 143 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
index 569e0e5373927..25a664273bf0d 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
@@ -141,6 +141,9 @@ int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f, struct amd
}
fence = &am_fence->base;
am_fence->ring = ring;
+ am_fence->start_ring_wptr = 0;
+ am_fence->end_ring_wptr = 0;
+ am_fence->context = 0;
seq = ++ring->fence_drv.sync_seq;
if (job && job->job_run_counter) {
@@ -748,6 +751,18 @@ void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring)
amdgpu_fence_process(ring);
}
+/**
+ * amdgpu_fence_driver_seq_force_completion - force signal of specified sequence
+ *
+ * @ring: fence of the ring to signal
+ *
+ */
+void amdgpu_fence_driver_seq_force_completion(struct amdgpu_ring *ring, u32 seq)
+{
+ amdgpu_fence_write(ring, seq);
+ amdgpu_fence_process(ring);
+}
+
/*
* Common fence implementation
*/
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
index 802743efa3b39..636941697a740 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
@@ -126,7 +126,9 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
struct dma_fence **f)
{
struct amdgpu_device *adev = ring->adev;
+ u64 start_ring_wptr, end_ring_wptr;
struct amdgpu_ib *ib = &ibs[0];
+ struct amdgpu_fence *am_fence;
struct dma_fence *tmp = NULL;
bool need_ctx_switch;
struct amdgpu_vm *vm;
@@ -138,7 +140,6 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
int vmid = AMDGPU_JOB_GET_VMID(job);
bool need_pipe_sync = false;
unsigned int cond_exec;
-
unsigned int i;
int r = 0;
@@ -187,6 +188,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
dev_err(adev->dev, "scheduling IB failed (%d).\n", r);
return r;
}
+ start_ring_wptr = ring->wptr;
need_ctx_switch = ring->current_ctx != fence_ctx;
if (ring->funcs->emit_pipeline_sync && job &&
@@ -306,6 +308,15 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned int num_ibs,
amdgpu_ring_ib_end(ring);
amdgpu_ring_commit(ring);
+ /* This must be last for resets to work properly
+ * as we need to save the wptr associated with this
+ * fence.
+ */
+ end_ring_wptr = ring->wptr;
+ am_fence = container_of(*f, struct amdgpu_fence, base);
+ am_fence->start_ring_wptr = start_ring_wptr;
+ am_fence->end_ring_wptr = end_ring_wptr;
+
return 0;
}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
index 23b6a0fc0c691..73c26e2e01647 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_job.c
@@ -89,8 +89,9 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
{
struct amdgpu_ring *ring = to_amdgpu_ring(s_job->sched);
struct amdgpu_job *job = to_amdgpu_job(s_job);
- struct amdgpu_task_info *ti;
+ struct amdgpu_fence *am_fence = &job->hw_fence;
struct amdgpu_device *adev = ring->adev;
+ struct amdgpu_task_info *ti;
bool set_error = false;
int idx, r;
@@ -154,7 +155,8 @@ static enum drm_gpu_sched_stat amdgpu_job_timedout(struct drm_sched_job *s_job)
else
is_guilty = true;
- if (is_guilty)
+ amdgpu_ring_backup_unprocessed_jobs(ring, is_guilty, am_fence);
+ if (is_guilty) {
dma_fence_set_error(&s_job->s_fence->finished, -ETIME);
set_error = true;
}
@@ -409,6 +411,7 @@ 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_fence *am_fence;
struct amdgpu_job *job;
int r = 0;
@@ -433,6 +436,10 @@ static struct dma_fence *amdgpu_job_run(struct drm_sched_job *sched_job)
"Error scheduling IBs (%d) in ring(%s)", r,
ring->name);
}
+ if (fence && finished) {
+ am_fence = container_of(fence, struct amdgpu_fence, base);
+ am_fence->context = finished->context;
+ }
job->job_run_counter++;
amdgpu_job_free_resources(job);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
index 426834806fbf2..3a0e0883bd8e7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
@@ -333,6 +333,12 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
/* Initialize cached_rptr to 0 */
ring->cached_rptr = 0;
+ if (!ring->ring_backup) {
+ ring->ring_backup = kvzalloc(ring->ring_size, GFP_KERNEL);
+ if (!ring->ring_backup)
+ return -ENOMEM;
+ }
+
/* Allocate ring buffer */
if (ring->ring_obj == NULL) {
r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE,
@@ -342,6 +348,7 @@ int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
(void **)&ring->ring);
if (r) {
dev_err(adev->dev, "(%d) ring create failed\n", r);
+ kvfree(ring->ring_backup);
return r;
}
amdgpu_ring_clear_ring(ring);
@@ -385,6 +392,8 @@ void amdgpu_ring_fini(struct amdgpu_ring *ring)
amdgpu_bo_free_kernel(&ring->ring_obj,
&ring->gpu_addr,
(void **)&ring->ring);
+ kvfree(ring->ring_backup);
+ ring->ring_backup = NULL;
dma_fence_put(ring->vmid_wait);
ring->vmid_wait = NULL;
@@ -753,3 +762,86 @@ bool amdgpu_ring_sched_ready(struct amdgpu_ring *ring)
return true;
}
+
+static void amdgpu_ring_backup_unprocessed_job(struct amdgpu_ring *ring,
+ unsigned int idx,
+ u64 start_wptr, u32 end_wptr)
+{
+ unsigned int first_idx = start_wptr & ring->buf_mask;
+ unsigned int last_idx = end_wptr & ring->buf_mask;
+ unsigned int i, j, entries_to_copy;
+
+ if (last_idx < first_idx) {
+ entries_to_copy = ring->buf_mask + 1 - first_idx;
+ for (i = 0; i < entries_to_copy; i++)
+ ring->ring_backup[idx + i] = ring->ring[first_idx + i];
+ ring->ring_backup_entries_to_copy += entries_to_copy;
+ entries_to_copy = last_idx;
+ for (j = 0; j < entries_to_copy; j++)
+ ring->ring_backup[idx + i + j] = ring->ring[j];
+ ring->ring_backup_entries_to_copy += entries_to_copy;
+ } else {
+ entries_to_copy = last_idx - first_idx;
+ for (i = 0; i < entries_to_copy; i++)
+ ring->ring_backup[idx + i] = ring->ring[first_idx + i];
+ ring->ring_backup_entries_to_copy += entries_to_copy;
+ }
+}
+
+void amdgpu_ring_backup_unprocessed_jobs(struct amdgpu_ring *ring,
+ bool is_guilty,
+ struct amdgpu_fence *bad_fence)
+{
+ struct amdgpu_fence *fence;
+ struct dma_fence *old, **ptr;
+ int i;
+
+ ring->ring_backup_entries_to_copy = 0;
+ for (i = 0; i <= ring->fence_drv.num_fences_mask; i++) {
+ ptr = &ring->fence_drv.fences[i];
+ rcu_read_lock();
+ old = rcu_dereference(*ptr);
+
+ if (old && !dma_fence_is_signaled(old)) {
+ fence = container_of(old, struct amdgpu_fence, base);
+ /* save everything if the ring is not guilty, otherwise
+ * just save the content from other contexts.
+ */
+ if (!is_guilty || (fence->context != bad_fence->context))
+ amdgpu_ring_backup_unprocessed_job(ring,
+ ring->ring_backup_entries_to_copy,
+ fence->start_ring_wptr,
+ fence->end_ring_wptr);
+ }
+ rcu_read_unlock();
+ }
+
+ ring->ring_bad_seq = bad_fence->base.seqno;
+}
+
+int amdgpu_ring_reemit_unprocessed_jobs(struct amdgpu_ring *ring)
+{
+ unsigned int i;
+ int r;
+
+ /* verify that the ring is functional */
+ r = amdgpu_ring_test_ring(ring);
+ if (r)
+ return r;
+ /* re-emit the unprocessed ring contents */
+ if (ring->ring_backup_entries_to_copy) {
+ r = amdgpu_ring_alloc(ring, ring->ring_backup_entries_to_copy);
+ if (r)
+ return r;
+ /* signal the fence of the bad job */
+ amdgpu_fence_driver_seq_force_completion(ring, ring->ring_bad_seq);
+ for (i = 0; i < ring->ring_backup_entries_to_copy; i++)
+ amdgpu_ring_write(ring, ring->ring_backup[i]);
+ amdgpu_ring_commit(ring);
+ } else {
+ /* signal the fence of the bad job */
+ amdgpu_fence_driver_seq_force_completion(ring, ring->ring_bad_seq);
+ }
+
+ return 0;
+}
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
index e1f25218943a4..69b71401adb7a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
@@ -141,6 +141,12 @@ struct amdgpu_fence {
/* RB, DMA, etc. */
struct amdgpu_ring *ring;
ktime_t start_timestamp;
+
+ /* wptrs for the fence for resets */
+ u64 start_ring_wptr;
+ u64 end_ring_wptr;
+ /* fence context for resets */
+ u64 context;
};
extern const struct drm_sched_backend_ops amdgpu_sched_ops;
@@ -148,6 +154,8 @@ extern const struct drm_sched_backend_ops amdgpu_sched_ops;
void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring);
void amdgpu_fence_driver_set_error(struct amdgpu_ring *ring, int error);
void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring);
+void amdgpu_fence_driver_seq_force_completion(struct amdgpu_ring *ring,
+ u32 seq);
int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring);
int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
@@ -284,6 +292,9 @@ struct amdgpu_ring {
struct amdgpu_bo *ring_obj;
uint32_t *ring;
+ uint32_t *ring_backup;
+ unsigned int ring_backup_entries_to_copy;
+ uint64_t ring_bad_seq;
unsigned rptr_offs;
u64 rptr_gpu_addr;
volatile u32 *rptr_cpu_addr;
@@ -550,4 +561,8 @@ int amdgpu_ib_pool_init(struct amdgpu_device *adev);
void amdgpu_ib_pool_fini(struct amdgpu_device *adev);
int amdgpu_ib_ring_tests(struct amdgpu_device *adev);
bool amdgpu_ring_sched_ready(struct amdgpu_ring *ring);
+void amdgpu_ring_backup_unprocessed_jobs(struct amdgpu_ring *ring,
+ bool is_guilty,
+ struct amdgpu_fence *bad_fence);
+int amdgpu_ring_reemit_unprocessed_jobs(struct amdgpu_ring *ring);
#endif
--
2.49.0
More information about the amd-gfx
mailing list