[v2,4/5] drm/i915/guc: Make adding GuC work items lockless

Commit Message

Michał Winiarski Sept. 12, 2017, 5:40 p.m. UTC

We can get rid of a spinlock by updating the tail directly using
cmpxchg. We can also put guc client on a diet by removing some constants
from the struct.
This causes a small change in one of GuC debugfs files.
We're no longer reporting constant values (which I don't think is a
problem), but we're also no longer reporting the tail (does anyone care?).

v2: Free space calculation inside the loop, compiler barriers. (Chris)

Cc: Chris Wilson <[e-mail]>
Cc: Daniele Ceraolo Spurio <[e-mail]>
Cc: Michal Wajdeczko <[e-mail]>
Cc: Oscar Mateo <[e-mail]>
Suggested-by: Chris Wilson <[e-mail]>
Signed-off-by: Michał Winiarski <[e-mail]>
Reviewed-by: Chris Wilson <[e-mail]>
---
 drivers/gpu/drm/i915/i915_debugfs.c        |  2 --
 drivers/gpu/drm/i915/i915_guc_submission.c | 41 ++++++++++++------------------
 drivers/gpu/drm/i915/intel_uc.h            |  5 ----
 3 files changed, 16 insertions(+), 32 deletions(-)

Comments

Chris Wilson Sept. 12, 2017, 9:22 p.m. UTC | #1

Quoting Michał Winiarski (2017-09-12 18:40:55)
> @@ -434,11 +430,16 @@ static void guc_wq_item_append(struct i915_guc_client *client,
>          */
>         BUILD_BUG_ON(wqi_size != 16);
>  
> -       /* postincrement WQ tail for next time */
> -       wq_off = client->wq_tail;
> +       /* Find our offset and postincrement WQ tail for next time, free space
> +        * is guaranteed.
> +        */
> +       do {
> +               wq_off = READ_ONCE(desc->tail);
> +               wq_next = (wq_off + wqi_size) & (GUC_WQ_SIZE - 1);
> +               GEM_BUG_ON(CIRC_SPACE(wq_off, READ_ONCE(desc->head),
> +                                     GUC_WQ_SIZE) < wqi_size);
> +       } while (cmpxchg(&desc->tail, wq_off, wq_next) != wq_off);
>         GEM_BUG_ON(wq_off & (wqi_size - 1));

I forgot the trick in cmpxchg to avoid the READ_ONCE in every loop:
	wq_next = READ_ONCE(desc->tail);
	do {
		wq_off = wq_next;
		wq_next = (wq_off + wqi_size) & (GUC_WQ_SIZE - 1);
		GEM_BUG_ON(...);
	} while ((wq_next = cmpxchg(&desc->tail, wq_off, wq_next) != wq_off);
-Chris

Patch

diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index 40a315fc0e31..0bd5cacd288b 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -2447,8 +2447,6 @@  static void i915_guc_client_info(struct seq_file *m,
 		client->priority, client->stage_id, client->proc_desc_offset);
 	seq_printf(m, "\tDoorbell id %d, offset: 0x%lx\n",
 		client->doorbell_id, client->doorbell_offset);
-	seq_printf(m, "\tWQ size %d, offset: 0x%x, tail %d\n",
-		client->wq_size, client->wq_offset, client->wq_tail);
 
 	for_each_engine(engine, dev_priv, id) {
 		u64 submissions = client->submissions[id];
diff --git a/drivers/gpu/drm/i915/i915_guc_submission.c b/drivers/gpu/drm/i915/i915_guc_submission.c
index 674f6a4035e6..16d6bb13c2bf 100644
--- a/drivers/gpu/drm/i915/i915_guc_submission.c
+++ b/drivers/gpu/drm/i915/i915_guc_submission.c
@@ -305,7 +305,7 @@  static void guc_proc_desc_init(struct intel_guc *guc,
 	desc->db_base_addr = 0;
 
 	desc->stage_id = client->stage_id;
-	desc->wq_size_bytes = client->wq_size;
+	desc->wq_size_bytes = GUC_WQ_SIZE;
 	desc->wq_status = WQ_STATUS_ACTIVE;
 	desc->priority = client->priority;
 }
@@ -390,8 +390,8 @@  static void guc_stage_desc_init(struct intel_guc *guc,
 	desc->db_trigger_cpu = (uintptr_t)__get_doorbell(client);
 	desc->db_trigger_uk = gfx_addr + client->doorbell_offset;
 	desc->process_desc = gfx_addr + client->proc_desc_offset;
-	desc->wq_addr = gfx_addr + client->wq_offset;
-	desc->wq_size = client->wq_size;
+	desc->wq_addr = gfx_addr + GUC_DB_SIZE;
+	desc->wq_size = GUC_WQ_SIZE;
 
 	desc->desc_private = (uintptr_t)client;
 }
@@ -415,15 +415,11 @@  static void guc_wq_item_append(struct i915_guc_client *client,
 	struct intel_engine_cs *engine = rq->engine;
 	struct guc_process_desc *desc = __get_process_desc(client);
 	struct guc_wq_item *wqi;
-	u32 freespace, tail, wq_off;
-
-	/* Free space is guaranteed */
-	freespace = CIRC_SPACE(client->wq_tail, desc->head, client->wq_size);
-	GEM_BUG_ON(freespace < wqi_size);
+	u32 ring_tail, wq_off, wq_next;
 
 	/* The GuC firmware wants the tail index in QWords, not bytes */
-	tail = intel_ring_set_tail(rq->ring, rq->tail) >> 3;
-	GEM_BUG_ON(tail > WQ_RING_TAIL_MAX);
+	ring_tail = intel_ring_set_tail(rq->ring, rq->tail) >> 3;
+	GEM_BUG_ON(ring_tail > WQ_RING_TAIL_MAX);
 
 	/* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we
 	 * should not have the case where structure wqi is across page, neither
@@ -434,11 +430,16 @@  static void guc_wq_item_append(struct i915_guc_client *client,
 	 */
 	BUILD_BUG_ON(wqi_size != 16);
 
-	/* postincrement WQ tail for next time */
-	wq_off = client->wq_tail;
+	/* Find our offset and postincrement WQ tail for next time, free space
+	 * is guaranteed.
+	 */
+	do {
+		wq_off = READ_ONCE(desc->tail);
+		wq_next = (wq_off + wqi_size) & (GUC_WQ_SIZE - 1);
+		GEM_BUG_ON(CIRC_SPACE(wq_off, READ_ONCE(desc->head),
+				      GUC_WQ_SIZE) < wqi_size);
+	} while (cmpxchg(&desc->tail, wq_off, wq_next) != wq_off);
 	GEM_BUG_ON(wq_off & (wqi_size - 1));
-	client->wq_tail += wqi_size;
-	client->wq_tail &= client->wq_size - 1;
 
 	/* WQ starts from the page after doorbell / process_desc */
 	wqi = client->vaddr + wq_off + GUC_DB_SIZE;
@@ -452,7 +453,7 @@  static void guc_wq_item_append(struct i915_guc_client *client,
 	/* The GuC wants only the low-order word of the context descriptor */
 	wqi->context_desc = (u32)intel_lr_context_descriptor(rq->ctx, engine);
 
-	wqi->submit_element_info = tail << WQ_RING_TAIL_SHIFT;
+	wqi->submit_element_info = ring_tail << WQ_RING_TAIL_SHIFT;
 	wqi->fence_id = rq->global_seqno;
 }
 
@@ -462,8 +463,6 @@  static void guc_reset_wq(struct i915_guc_client *client)
 
 	desc->head = 0;
 	desc->tail = 0;
-
-	client->wq_tail = 0;
 }
 
 static void guc_ring_doorbell(struct i915_guc_client *client)
@@ -497,7 +496,6 @@  static void i915_guc_submit(struct intel_engine_cs *engine)
 	struct execlist_port *port = engine->execlist_port;
 	unsigned int engine_id = engine->id;
 	unsigned int n;
-	unsigned long flags;
 
 	for (n = 0; n < ARRAY_SIZE(engine->execlist_port); n++) {
 		struct drm_i915_gem_request *rq;
@@ -510,14 +508,10 @@  static void i915_guc_submit(struct intel_engine_cs *engine)
 			if (i915_vma_is_map_and_fenceable(rq->ring->vma))
 				POSTING_READ_FW(GUC_STATUS);
 
-			spin_lock_irqsave(&client->wq_lock, flags);
-
 			guc_wq_item_append(client, rq);
 			guc_ring_doorbell(client);
 
 			client->submissions[engine_id] += 1;
-
-			spin_unlock_irqrestore(&client->wq_lock, flags);
 		}
 	}
 }
@@ -810,9 +804,6 @@  guc_client_alloc(struct drm_i915_private *dev_priv,
 	client->engines = engines;
 	client->priority = priority;
 	client->doorbell_id = GUC_DOORBELL_INVALID;
-	client->wq_offset = GUC_DB_SIZE;
-	client->wq_size = GUC_WQ_SIZE;
-	spin_lock_init(&client->wq_lock);
 
 	ret = ida_simple_get(&guc->stage_ids, 0, GUC_MAX_STAGE_DESCRIPTORS,
 				GFP_KERNEL);
diff --git a/drivers/gpu/drm/i915/intel_uc.h b/drivers/gpu/drm/i915/intel_uc.h
index f505dcccd613..386e0bcd0cc3 100644
--- a/drivers/gpu/drm/i915/intel_uc.h
+++ b/drivers/gpu/drm/i915/intel_uc.h
@@ -67,11 +67,6 @@  struct i915_guc_client {
 	u16 doorbell_id;
 	unsigned long doorbell_offset;
 
-	spinlock_t wq_lock;
-	uint32_t wq_offset;
-	uint32_t wq_size;
-	uint32_t wq_tail;
-
 	/* Per-engine counts of GuC submissions */
 	uint64_t submissions[I915_NUM_ENGINES];
 };