drm/i915: Drive request submission through fence callbacks

Drive final request submission from a callback from the fence. This way
the request is queued until all dependencies are resolved, at which
point it is handed to the backend for queueing to hardware. At this
point, no dependencies are set on the request, so the callback is
immediate.

A side-effect of imposing a heavier-irqsafe spinlock for execlist
submission is that we lose the softirq enabling after scheduling the
execlists tasklet. To compensate, we manually kickstart the softirq by
disabling and enabling the bh around the fence signaling.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: John Harrison <john.c.harrison@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20160909131201.16673-14-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/i915_gem.c

@@ -2549,6 +2549,9 @@ i915_gem_find_active_request(struct intel_engine_cs *engine)
 		if (i915_gem_request_completed(request))
 			continue;
 
+		if (!i915_sw_fence_done(&request->submit))
+			break;
+
 		return request;
 	}
 

drivers/gpu/drm/i915/i915_gem_request.c

@@ -318,6 +318,26 @@ static int i915_gem_get_seqno(struct drm_i915_private *dev_priv, u32 *seqno)
 	return 0;
 }
 
+static int __i915_sw_fence_call
+submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
+{
+	struct drm_i915_gem_request *request =
+		container_of(fence, typeof(*request), submit);
+
+	/* Will be called from irq-context when using foreign DMA fences */
+
+	switch (state) {
+	case FENCE_COMPLETE:
+		request->engine->submit_request(request);
+		break;
+
+	case FENCE_FREE:
+		break;
+	}
+
+	return NOTIFY_DONE;
+}
+
 /**
  * i915_gem_request_alloc - allocate a request structure
  *
@@ -396,6 +416,8 @@ i915_gem_request_alloc(struct intel_engine_cs *engine,
 		   engine->fence_context,
 		   seqno);
 
+	i915_sw_fence_init(&req->submit, submit_notify);
+
 	INIT_LIST_HEAD(&req->active_list);
 	req->i915 = dev_priv;
 	req->engine = engine;
@@ -530,7 +552,10 @@ void __i915_add_request(struct drm_i915_gem_request *request, bool flush_caches)
 		  reserved_tail, ret);
 
 	i915_gem_mark_busy(engine);
-	engine->submit_request(request);
+
+	local_bh_disable();
+	i915_sw_fence_commit(&request->submit);
+	local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
 }
 
 static void reset_wait_queue(wait_queue_head_t *q, wait_queue_t *wait)

drivers/gpu/drm/i915/i915_gem_request.h

@@ -28,6 +28,7 @@
 #include <linux/fence.h>
 
 #include "i915_gem.h"
+#include "i915_sw_fence.h"
 
 struct intel_wait {
 	struct rb_node node;
@@ -82,6 +83,8 @@ struct drm_i915_gem_request {
 	struct intel_ring *ring;
 	struct intel_signal_node signaling;
 
+	struct i915_sw_fence submit;
+
 	/** GEM sequence number associated with the previous request,
 	 * when the HWS breadcrumb is equal to this the GPU is processing
 	 * this request.

drivers/gpu/drm/i915/i915_guc_submission.c

@@ -1016,7 +1016,8 @@ int i915_guc_submission_enable(struct drm_i915_private *dev_priv)
 
 		/* Replay the current set of previously submitted requests */
 		list_for_each_entry(request, &engine->request_list, link)
-			i915_guc_submit(request);
+			if (i915_sw_fence_done(&request->submit))
+				i915_guc_submit(request);
 	}
 
 	return 0;

drivers/gpu/drm/i915/intel_breadcrumbs.c

@@ -462,7 +462,10 @@ static int intel_breadcrumbs_signaler(void *arg)
 			 */
 			intel_engine_remove_wait(engine,
 						 &request->signaling.wait);
+
+			local_bh_disable();
 			fence_signal(&request->fence);
+			local_bh_enable(); /* kick start the tasklets */
 
 			/* Find the next oldest signal. Note that as we have
 			 * not been holding the lock, another client may

drivers/gpu/drm/i915/intel_lrc.c

@@ -590,14 +590,15 @@ static void intel_lrc_irq_handler(unsigned long data)
 static void execlists_submit_request(struct drm_i915_gem_request *request)
 {
 	struct intel_engine_cs *engine = request->engine;
+	unsigned long flags;
 
-	spin_lock_bh(&engine->execlist_lock);
+	spin_lock_irqsave(&engine->execlist_lock, flags);
 
 	list_add_tail(&request->execlist_link, &engine->execlist_queue);
 	if (execlists_elsp_idle(engine))
 		tasklet_hi_schedule(&engine->irq_tasklet);
 
-	spin_unlock_bh(&engine->execlist_lock);
+	spin_unlock_irqrestore(&engine->execlist_lock, flags);
 }
 
 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request)

drivers/gpu/drm/i915/intel_ringbuffer.h

@@ -226,7 +226,15 @@ struct intel_engine_cs {
 #define I915_DISPATCH_PINNED BIT(1)
 #define I915_DISPATCH_RS     BIT(2)
 	int		(*emit_request)(struct drm_i915_gem_request *req);
+
+	/* Pass the request to the hardware queue (e.g. directly into
+	 * the legacy ringbuffer or to the end of an execlist).
+	 *
+	 * This is called from an atomic context with irqs disabled; must
+	 * be irq safe.
+	 */
 	void		(*submit_request)(struct drm_i915_gem_request *req);
+
 	/* Some chipsets are not quite as coherent as advertised and need
 	 * an expensive kick to force a true read of the up-to-date seqno.
 	 * However, the up-to-date seqno is not always required and the last