Since fw version 25.161, GuC lets us know when an engine had to be reset
due to a hang in another dependent engine, by setting BIT(engine_class) in
the queue_engine_error field. GuC will ignore any other wq item until this
flag is cleared.
To restart the workqueue processing for that engine, we must insert a
special wq item called resume-parsing and wait until the queue_engine_error
field is updated.
Co-Developed-by: Michel Thierry <michel.thierry@xxxxxxxxx>
Co-Developed-by: Michal Wajdeczko <michal.wajdeczko@xxxxxxxxx>
Signed-off-by: Michel Thierry <michel.thierry@xxxxxxxxx>
Signed-off-by: Michal Wajdeczko <michal.wajdeczko@xxxxxxxxx>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@xxxxxxxxx>
Cc: Vinay Belgaumkar <vinay.belgaumkar@xxxxxxxxx>
Cc: Michal Winiarski <michal.winiarski@xxxxxxxxx>
Cc: Tomasz Lis <tomasz.lis@xxxxxxxxx>
---
drivers/gpu/drm/i915/intel_guc_submission.c | 92 +++++++++++++++++++--
1 file changed, 86 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/i915/intel_guc_submission.c b/drivers/gpu/drm/i915/intel_guc_submission.c
index 1a27352b73a9..b0b10cac9b9b 100644
--- a/drivers/gpu/drm/i915/intel_guc_submission.c
+++ b/drivers/gpu/drm/i915/intel_guc_submission.c
@@ -427,6 +427,76 @@ static void guc_proxy_stage_desc_fini(struct intel_guc_client *client)
memset(desc, 0, sizeof(*desc));
}
+static u32 get_wq_offset(struct guc_process_desc *desc)
+{
+ const size_t wqi_size = sizeof(struct guc_wq_item);
+ u32 wq_off;
+
+ /*
+ * Free space is guaranteed, either by normal port submission or
+ * because we waited for the wq_resume to be processed.
+ */
+ wq_off = READ_ONCE(desc->tail);
+ GEM_BUG_ON(CIRC_SPACE(wq_off, READ_ONCE(desc->head),
+ GUC_WQ_SIZE) < wqi_size);
+ GEM_BUG_ON(wq_off & (wqi_size - 1));
+
+ return wq_off;
+}
+
+static void write_wqi(struct guc_process_desc *desc, u32 wq_off)
+{
+ const size_t wqi_size = sizeof(struct guc_wq_item);
+
+ WRITE_ONCE(desc->tail, (wq_off + wqi_size) & (GUC_WQ_SIZE - 1));
+}
+
+static void guc_append_wq_resume_parsing_item(struct intel_guc_client *client,
+ struct intel_engine_cs *engine)
+{
+ struct guc_process_desc *desc = __get_process_desc(client);
+ const u32 target_engine = engine->guc_id;
+ const u32 wqi_resume = WQ_TYPE_RESUME |
+ (target_engine << WQ_TARGET_SHIFT) |
+ (0 << WQ_LEN_SHIFT);
+ const u32 wqi_noop = WQ_TYPE_NOOP |
+ (target_engine << WQ_TARGET_SHIFT) |
+ (0 << WQ_LEN_SHIFT);
+ struct guc_wq_item *wqi;
+ u32 wq_off;
+
+ lockdep_assert_held(&client->wq_lock);
+
+ wq_off = get_wq_offset(desc);
+ wqi = client->vaddr + wq_off + GUC_DB_SIZE;
+
+ /*
+ * Submit 4 wq_items (1 RESUME_WQ_PARSING followed by 3 NOOPs) in
+ * order to keep it the same size as a 'normal' wq_item.
+ */
+ wqi->header = wqi_resume;
+ wqi->context_desc = wqi_noop;
+ wqi->submit_element_info = wqi_noop;
+ wqi->fence_id = wqi_noop;
+
+ write_wqi(desc, wq_off);
+}
+
+#define GUC_WAIT_FOR_ENGINE_ERROR_CLEANED_MS 10
+
+static void guc_wait_wq_resumed(struct intel_guc_client *client,
+ struct intel_engine_cs *engine)
+{
+ struct guc_process_desc *desc = __get_process_desc(client);
+ u8 guc_class = engine->class;
+
+ lockdep_assert_held(&client->wq_lock);
+
+ /* must wait for the flag to be cleared */
+ WARN_ON(wait_for_atomic(!(desc->queue_engine_error & BIT(guc_class)),
+ GUC_WAIT_FOR_ENGINE_ERROR_CLEANED_MS));
+}
+
/* Construct a Work Item and append it to the GuC's Work Queue */
static void guc_wq_item_append(struct intel_guc_client *client,
struct intel_context *ce,
@@ -454,11 +524,7 @@ static void guc_wq_item_append(struct intel_guc_client *client,
/* We expect the WQ to be active if we're appending items to it */
GEM_BUG_ON(desc->wq_status != WQ_STATUS_ACTIVE);
- /* Free space is guaranteed. */
- wq_off = READ_ONCE(desc->tail);
- GEM_BUG_ON(CIRC_SPACE(wq_off, READ_ONCE(desc->head),
- GUC_WQ_SIZE) < wqi_size);
- GEM_BUG_ON(wq_off & (wqi_size - 1));
+ wq_off = get_wq_offset(desc);
/* WQ starts from the page after doorbell / process_desc */
wqi = client->vaddr + wq_off + GUC_DB_SIZE;
@@ -481,7 +547,7 @@ static void guc_wq_item_append(struct intel_guc_client *client,
}
/* Make the update visible to GuC */
- WRITE_ONCE(desc->tail, (wq_off + wqi_size) & (GUC_WQ_SIZE - 1));
+ write_wqi(desc, wq_off);
}
static void guc_ring_doorbell(struct intel_guc_client *client)
@@ -505,6 +571,14 @@ static void guc_ring_doorbell(struct intel_guc_client *client)
GEM_BUG_ON(db->db_status != GUC_DOORBELL_ENABLED);
}
+static bool guc_needs_wq_resume_parsing_item(struct intel_guc_client *client,
+ u32 target_engine_class)
+{
+ struct guc_process_desc *desc = __get_process_desc(client);
+
+ return desc->queue_engine_error & BIT(target_engine_class);
+}
+
static void guc_add_request(struct intel_guc *guc, struct i915_request *rq)
{
struct intel_guc_client *client = guc->execbuf_client;
@@ -512,9 +586,15 @@ static void guc_add_request(struct intel_guc *guc, struct i915_request *rq)
struct intel_context *ce = rq->hw_context;
u32 ctx_desc = lower_32_bits(rq->hw_context->lrc_desc);
u32 ring_tail = intel_ring_set_tail(rq->ring, rq->tail) / sizeof(u64);
+ u8 guc_class = engine->class;
spin_lock(&client->wq_lock);
+ if (guc_needs_wq_resume_parsing_item(client, guc_class)) {
+ guc_append_wq_resume_parsing_item(client, engine);
+ guc_wait_wq_resumed(client, engine);
+ }
+
guc_wq_item_append(client, ce, engine->guc_id, ctx_desc,
ring_tail, rq->fence.seqno);
guc_ring_doorbell(client);
--
2.19.2
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
