drm/i915/guc: Preemption! With GuC

Pretty similar to what we have on execlists.
We're reusing most of the GEM code, however, due to GuC quirks we need a
couple of extra bits.
Preemption is implemented as GuC action, and actions can be pretty slow.
Because of that, we're using a mutex to serialize them. Since we're
requesting preemption from the tasklet, the task of creating a workitem
and wrapping it in GuC action is delegated to a worker.

To distinguish that preemption has finished, we're using additional
piece of HWSP, and since we're not getting context switch interrupts,
we're also adding a user interrupt.

The fact that our special preempt context has completed unfortunately
doesn't mean that we're ready to submit new work. We also need to wait
for GuC to finish its own processing.

v2: Don't compile out the wait for GuC, handle workqueue flush on reset,
no need for ordered workqueue, put on a reviewer hat when looking at my own
patches (Chris)
Move struct work around in intel_guc, move user interruput outside of
conditional (Michał)
Keep ring around rather than chase though intel_context

v3: Extract WA for flushing ggtt writes to a helper (Chris)
Keep work_struct in intel_guc rather than engine (Michał)
Use ordered workqueue for inject_preempt worker to avoid GuC quirks.

v4: Drop now unused INTEL_GUC_PREEMPT_OPTION_IMMEDIATE (Daniele)
Drop stray newlines, use container_of for intel_guc in worker,
check for presence of workqueue when flushing it, rather than
enable_guc_submission modparam, reorder preempt postprocessing (Chris)

v5: Make wq NULL after destroying it

v6: Swap struct guc_preempt_work members (Michał)

Signed-off-by: Michał Winiarski <michal.winiarski@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jeff McGee <jeff.mcgee@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Oscar Mateo <oscar.mateo@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20171026133558.19580-1-michal.winiarski@intel.com

drivers/gpu/drm/i915/i915_drv.c

@@ -373,8 +373,7 @@ static int i915_getparam(struct drm_device *dev, void *data,
 			value |= I915_SCHEDULER_CAP_PRIORITY;
 
 			if (HAS_LOGICAL_RING_PREEMPTION(dev_priv) &&
-			    i915_modparams.enable_execlists &&
-			    !i915_modparams.enable_guc_submission)
+			    i915_modparams.enable_execlists)
 				value |= I915_SCHEDULER_CAP_PREEMPTION;
 		}
 		break;

drivers/gpu/drm/i915/i915_gem.c

@@ -2921,6 +2921,16 @@ i915_gem_reset_prepare_engine(struct intel_engine_cs *engine)
 	tasklet_kill(&engine->execlists.irq_tasklet);
 	tasklet_disable(&engine->execlists.irq_tasklet);
 
+	/*
+	 * We're using worker to queue preemption requests from the tasklet in
+	 * GuC submission mode.
+	 * Even though tasklet was disabled, we may still have a worker queued.
+	 * Let's make sure that all workers scheduled before disabling the
+	 * tasklet are completed before continuing with the reset.
+	 */
+	if (engine->i915->guc.preempt_wq)
+		flush_workqueue(engine->i915->guc.preempt_wq);
+
 	if (engine->irq_seqno_barrier)
 		engine->irq_seqno_barrier(engine);
 

drivers/gpu/drm/i915/i915_guc_submission.c

@@ -565,6 +565,108 @@ static void guc_add_request(struct intel_guc *guc,
 	spin_unlock(&client->wq_lock);
 }
 
+/*
+ * When we're doing submissions using regular execlists backend, writing to
+ * ELSP from CPU side is enough to make sure that writes to ringbuffer pages
+ * pinned in mappable aperture portion of GGTT are visible to command streamer.
+ * Writes done by GuC on our behalf are not guaranteeing such ordering,
+ * therefore, to ensure the flush, we're issuing a POSTING READ.
+ */
+static void flush_ggtt_writes(struct i915_vma *vma)
+{
+	struct drm_i915_private *dev_priv = to_i915(vma->obj->base.dev);
+
+	if (i915_vma_is_map_and_fenceable(vma))
+		POSTING_READ_FW(GUC_STATUS);
+}
+
+#define GUC_PREEMPT_FINISHED 0x1
+#define GUC_PREEMPT_BREADCRUMB_DWORDS 0x8
+static void inject_preempt_context(struct work_struct *work)
+{
+	struct guc_preempt_work *preempt_work =
+		container_of(work, typeof(*preempt_work), work);
+	struct intel_engine_cs *engine = preempt_work->engine;
+	struct intel_guc *guc = container_of(preempt_work, typeof(*guc),
+					     preempt_work[engine->id]);
+	struct i915_guc_client *client = guc->preempt_client;
+	struct intel_ring *ring = client->owner->engine[engine->id].ring;
+	u32 ctx_desc = lower_32_bits(intel_lr_context_descriptor(client->owner,
+								 engine));
+	u32 *cs = ring->vaddr + ring->tail;
+	u32 data[7];
+
+	if (engine->id == RCS) {
+		cs = gen8_emit_ggtt_write_rcs(cs, GUC_PREEMPT_FINISHED,
+				intel_hws_preempt_done_address(engine));
+	} else {
+		cs = gen8_emit_ggtt_write(cs, GUC_PREEMPT_FINISHED,
+				intel_hws_preempt_done_address(engine));
+		*cs++ = MI_NOOP;
+		*cs++ = MI_NOOP;
+	}
+	*cs++ = MI_USER_INTERRUPT;
+	*cs++ = MI_NOOP;
+
+	GEM_BUG_ON(!IS_ALIGNED(ring->size,
+			       GUC_PREEMPT_BREADCRUMB_DWORDS * sizeof(u32)));
+	GEM_BUG_ON((void *)cs - (ring->vaddr + ring->tail) !=
+		   GUC_PREEMPT_BREADCRUMB_DWORDS * sizeof(u32));
+
+	ring->tail += GUC_PREEMPT_BREADCRUMB_DWORDS * sizeof(u32);
+	ring->tail &= (ring->size - 1);
+
+	flush_ggtt_writes(ring->vma);
+
+	spin_lock_irq(&client->wq_lock);
+	guc_wq_item_append(client, engine->guc_id, ctx_desc,
+			   ring->tail / sizeof(u64), 0);
+	spin_unlock_irq(&client->wq_lock);
+
+	data[0] = INTEL_GUC_ACTION_REQUEST_PREEMPTION;
+	data[1] = client->stage_id;
+	data[2] = INTEL_GUC_PREEMPT_OPTION_DROP_WORK_Q |
+		  INTEL_GUC_PREEMPT_OPTION_DROP_SUBMIT_Q;
+	data[3] = engine->guc_id;
+	data[4] = guc->execbuf_client->priority;
+	data[5] = guc->execbuf_client->stage_id;
+	data[6] = guc_ggtt_offset(guc->shared_data);
+
+	if (WARN_ON(intel_guc_send(guc, data, ARRAY_SIZE(data)))) {
+		execlists_clear_active(&engine->execlists,
+				       EXECLISTS_ACTIVE_PREEMPT);
+		tasklet_schedule(&engine->execlists.irq_tasklet);
+	}
+}
+
+/*
+ * We're using user interrupt and HWSP value to mark that preemption has
+ * finished and GPU is idle. Normally, we could unwind and continue similar to
+ * execlists submission path. Unfortunately, with GuC we also need to wait for
+ * it to finish its own postprocessing, before attempting to submit. Otherwise
+ * GuC may silently ignore our submissions, and thus we risk losing request at
+ * best, executing out-of-order and causing kernel panic at worst.
+ */
+#define GUC_PREEMPT_POSTPROCESS_DELAY_MS 10
+static void wait_for_guc_preempt_report(struct intel_engine_cs *engine)
+{
+	struct intel_guc *guc = &engine->i915->guc;
+	struct guc_shared_ctx_data *data = guc->shared_data_vaddr;
+	struct guc_ctx_report *report =
+		&data->preempt_ctx_report[engine->guc_id];
+
+	WARN_ON(wait_for_atomic(report->report_return_status ==
+				INTEL_GUC_REPORT_STATUS_COMPLETE,
+				GUC_PREEMPT_POSTPROCESS_DELAY_MS));
+	/*
+	 * GuC is expecting that we're also going to clear the affected context
+	 * counter, let's also reset the return status to not depend on GuC
+	 * resetting it after recieving another preempt action
+	 */
+	report->affected_count = 0;
+	report->report_return_status = INTEL_GUC_REPORT_STATUS_UNKNOWN;
+}
+
 /**
  * i915_guc_submit() - Submit commands through GuC
  * @engine: engine associated with the commands
@@ -574,8 +676,7 @@ static void guc_add_request(struct intel_guc *guc,
  */
 static void i915_guc_submit(struct intel_engine_cs *engine)
 {
-	struct drm_i915_private *dev_priv = engine->i915;
-	struct intel_guc *guc = &dev_priv->guc;
+	struct intel_guc *guc = &engine->i915->guc;
 	struct intel_engine_execlists * const execlists = &engine->execlists;
 	struct execlist_port *port = execlists->port;
 	unsigned int n;
@@ -588,8 +689,7 @@ static void i915_guc_submit(struct intel_engine_cs *engine)
 		if (rq && count == 0) {
 			port_set(&port[n], port_pack(rq, ++count));
 
-			if (i915_vma_is_map_and_fenceable(rq->ring->vma))
-				POSTING_READ_FW(GUC_STATUS);
+			flush_ggtt_writes(rq->ring->vma);
 
 			guc_add_request(guc, rq);
 		}
@@ -617,13 +717,32 @@ static void i915_guc_dequeue(struct intel_engine_cs *engine)
 	bool submit = false;
 	struct rb_node *rb;
 
-	if (port_isset(port))
-		port++;
-
 	spin_lock_irq(&engine->timeline->lock);
 	rb = execlists->first;
 	GEM_BUG_ON(rb_first(&execlists->queue) != rb);
-	while (rb) {
+
+	if (!rb)
+		goto unlock;
+
+	if (HAS_LOGICAL_RING_PREEMPTION(engine->i915) && port_isset(port)) {
+		struct guc_preempt_work *preempt_work =
+			&engine->i915->guc.preempt_work[engine->id];
+
+		if (rb_entry(rb, struct i915_priolist, node)->priority >
+		    max(port_request(port)->priotree.priority, 0)) {
+			execlists_set_active(execlists,
+					     EXECLISTS_ACTIVE_PREEMPT);
+			queue_work(engine->i915->guc.preempt_wq,
+				   &preempt_work->work);
+			goto unlock;
+		} else if (port_isset(last_port)) {
+			goto unlock;
+		}
+
+		port++;
+	}
+
+	do {
 		struct i915_priolist *p = rb_entry(rb, typeof(*p), node);
 		struct drm_i915_gem_request *rq, *rn;
 
@@ -653,7 +772,7 @@ static void i915_guc_dequeue(struct intel_engine_cs *engine)
 		INIT_LIST_HEAD(&p->requests);
 		if (p->priority != I915_PRIORITY_NORMAL)
 			kmem_cache_free(engine->i915->priorities, p);
-	}
+	} while (rb);
 done:
 	execlists->first = rb;
 	if (submit) {
@@ -661,6 +780,7 @@ done:
 		execlists_set_active(execlists, EXECLISTS_ACTIVE_USER);
 		i915_guc_submit(engine);
 	}
+unlock:
 	spin_unlock_irq(&engine->timeline->lock);
 }
 
@@ -669,8 +789,6 @@ static void i915_guc_irq_handler(unsigned long data)
 	struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
 	struct intel_engine_execlists * const execlists = &engine->execlists;
 	struct execlist_port *port = execlists->port;
-	const struct execlist_port * const last_port =
-		&execlists->port[execlists->port_mask];
 	struct drm_i915_gem_request *rq;
 
 	rq = port_request(&port[0]);
@@ -685,7 +803,19 @@ static void i915_guc_irq_handler(unsigned long data)
 	if (!rq)
 		execlists_clear_active(execlists, EXECLISTS_ACTIVE_USER);
 
-	if (!port_isset(last_port))
+	if (execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT) &&
+	    intel_read_status_page(engine, I915_GEM_HWS_PREEMPT_INDEX) ==
+	    GUC_PREEMPT_FINISHED) {
+		execlists_cancel_port_requests(&engine->execlists);
+		execlists_unwind_incomplete_requests(execlists);
+
+		wait_for_guc_preempt_report(engine);
+
+		execlists_clear_active(execlists, EXECLISTS_ACTIVE_PREEMPT);
+		intel_write_status_page(engine, I915_GEM_HWS_PREEMPT_INDEX, 0);
+	}
+
+	if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT))
 		i915_guc_dequeue(engine);
 }
 
@@ -1059,6 +1189,51 @@ static void guc_ads_destroy(struct intel_guc *guc)
 	i915_vma_unpin_and_release(&guc->ads_vma);
 }
 
+static int guc_preempt_work_create(struct intel_guc *guc)
+{
+	struct drm_i915_private *dev_priv = guc_to_i915(guc);
+	struct intel_engine_cs *engine;
+	enum intel_engine_id id;
+
+	/*
+	 * Even though both sending GuC action, and adding a new workitem to
+	 * GuC workqueue are serialized (each with its own locking), since
+	 * we're using mutliple engines, it's possible that we're going to
+	 * issue a preempt request with two (or more - each for different
+	 * engine) workitems in GuC queue. In this situation, GuC may submit
+	 * all of them, which will make us very confused.
+	 * Our preemption contexts may even already be complete - before we
+	 * even had the chance to sent the preempt action to GuC!. Rather
+	 * than introducing yet another lock, we can just use ordered workqueue
+	 * to make sure we're always sending a single preemption request with a
+	 * single workitem.
+	 */
+	guc->preempt_wq = alloc_ordered_workqueue("i915-guc_preempt",
+						  WQ_HIGHPRI);
+	if (!guc->preempt_wq)
+		return -ENOMEM;
+
+	for_each_engine(engine, dev_priv, id) {
+		guc->preempt_work[id].engine = engine;
+		INIT_WORK(&guc->preempt_work[id].work, inject_preempt_context);
+	}
+
+	return 0;
+}
+
+static void guc_preempt_work_destroy(struct intel_guc *guc)
+{
+	struct drm_i915_private *dev_priv = guc_to_i915(guc);
+	struct intel_engine_cs *engine;
+	enum intel_engine_id id;
+
+	for_each_engine(engine, dev_priv, id)
+		cancel_work_sync(&guc->preempt_work[id].work);
+
+	destroy_workqueue(guc->preempt_wq);
+	guc->preempt_wq = NULL;
+}
+
 /*
  * Set up the memory resources to be shared with the GuC (via the GGTT)
  * at firmware loading time.
@@ -1083,12 +1258,18 @@ int i915_guc_submission_init(struct drm_i915_private *dev_priv)
 	if (ret < 0)
 		goto err_shared_data;
 
+	ret = guc_preempt_work_create(guc);
+	if (ret)
+		goto err_log;
+
 	ret = guc_ads_create(guc);
 	if (ret < 0)
-		goto err_log;
+		goto err_wq;
 
 	return 0;
 
+err_wq:
+	guc_preempt_work_destroy(guc);
 err_log:
 	intel_guc_log_destroy(guc);
 err_shared_data:
@@ -1103,6 +1284,7 @@ void i915_guc_submission_fini(struct drm_i915_private *dev_priv)
 	struct intel_guc *guc = &dev_priv->guc;
 
 	guc_ads_destroy(guc);
+	guc_preempt_work_destroy(guc);
 	intel_guc_log_destroy(guc);
 	guc_shared_data_destroy(guc);
 	guc_stage_desc_pool_destroy(guc);

drivers/gpu/drm/i915/intel_guc.h

@@ -34,6 +34,11 @@
 #include "i915_guc_reg.h"
 #include "i915_vma.h"
 
+struct guc_preempt_work {
+	struct work_struct work;
+	struct intel_engine_cs *engine;
+};
+
 /*
  * Top level structure of GuC. It handles firmware loading and manages client
  * pool and doorbells. intel_guc owns a i915_guc_client to replace the legacy
@@ -60,6 +65,9 @@ struct intel_guc {
 	struct i915_guc_client *execbuf_client;
 	struct i915_guc_client *preempt_client;
 
+	struct guc_preempt_work preempt_work[I915_NUM_ENGINES];
+	struct workqueue_struct *preempt_wq;
+
 	DECLARE_BITMAP(doorbell_bitmap, GUC_NUM_DOORBELLS);
 	/* Cyclic counter mod pagesize	*/
 	u32 db_cacheline;

drivers/gpu/drm/i915/intel_lrc.c

@@ -385,7 +385,7 @@ static void __unwind_incomplete_requests(struct intel_engine_cs *engine)
 	}
 }
 
-static void
+void
 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists)
 {
 	struct intel_engine_cs *engine =
@@ -696,7 +696,7 @@ unlock:
 	}
 }
 
-static void
+void
 execlists_cancel_port_requests(struct intel_engine_execlists * const execlists)
 {
 	struct execlist_port *port = execlists->port;

drivers/gpu/drm/i915/intel_lrc.h

@@ -107,7 +107,6 @@ intel_lr_context_descriptor(struct i915_gem_context *ctx,
 	return ctx->engine[engine->id].lrc_desc;
 }
 
-
 /* Execlists */
 int intel_sanitize_enable_execlists(struct drm_i915_private *dev_priv,
 				    int enable_execlists);

drivers/gpu/drm/i915/intel_ringbuffer.h

@@ -557,6 +557,12 @@ execlists_is_active(const struct intel_engine_execlists *execlists,
 	return test_bit(bit, (unsigned long *)&execlists->active);
 }
 
+void
+execlists_cancel_port_requests(struct intel_engine_execlists * const execlists);
+
+void
+execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists);
+
 static inline unsigned int
 execlists_num_ports(const struct intel_engine_execlists * const execlists)
 {