drm/i915: Allow sharing the idle-barrier from other kernel requests

By placing our idle-barriers in the i915_active fence tree, we expose
those for reuse by other components that are issuing requests along the
kernel_context. Reusing the proto-barrier active_node is perfectly fine
as the new request implies a context-switch, and so an opportune point
to run the idle-barrier. However, the proto-barrier is not equivalent
to a normal active_node and care must be taken to avoid dereferencing the
ERR_PTR used as its request marker.

v2: Comment the more egregious cheek
v3: A glossary!
Reported-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Fixes: ce476c80 ("drm/i915: Keep contexts pinned until after the next kernel context switch")
Fixes: a9877da2 ("drm/i915/oa: Reconfigure contexts on the fly")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190802100015.1281-1-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/gt/intel_context.c

@@ -162,23 +162,41 @@ static int __intel_context_active(struct i915_active *active)
 	if (err)
 		goto err_ring;
 
+	return 0;
+
+err_ring:
+	intel_ring_unpin(ce->ring);
+err_put:
+	intel_context_put(ce);
+	return err;
+}
+
+int intel_context_active_acquire(struct intel_context *ce)
+{
+	int err;
+
+	err = i915_active_acquire(&ce->active);
+	if (err)
+		return err;
+
 	/* Preallocate tracking nodes */
 	if (!i915_gem_context_is_kernel(ce->gem_context)) {
 		err = i915_active_acquire_preallocate_barrier(&ce->active,
 							      ce->engine);
-		if (err)
-			goto err_state;
+		if (err) {
+			i915_active_release(&ce->active);
+			return err;
+		}
 	}
 
 	return 0;
+}
 
-err_state:
-	__context_unpin_state(ce->state);
-err_ring:
-	intel_ring_unpin(ce->ring);
-err_put:
-	intel_context_put(ce);
-	return err;
+void intel_context_active_release(struct intel_context *ce)
+{
+	/* Nodes preallocated in intel_context_active() */
+	i915_active_acquire_barrier(&ce->active);
+	i915_active_release(&ce->active);
 }
 
 void
@@ -301,3 +319,7 @@ struct i915_request *intel_context_create_request(struct intel_context *ce)
 
 	return rq;
 }
+
+#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
+#include "selftest_context.c"
+#endif

drivers/gpu/drm/i915/gt/intel_context.h

@@ -104,17 +104,8 @@ static inline void intel_context_exit(struct intel_context *ce)
 		ce->ops->exit(ce);
 }
 
-static inline int intel_context_active_acquire(struct intel_context *ce)
-{
-	return i915_active_acquire(&ce->active);
-}
-
-static inline void intel_context_active_release(struct intel_context *ce)
-{
-	/* Nodes preallocated in intel_context_active() */
-	i915_active_acquire_barrier(&ce->active);
-	i915_active_release(&ce->active);
-}
+int intel_context_active_acquire(struct intel_context *ce);
+void intel_context_active_release(struct intel_context *ce);
 
 static inline struct intel_context *intel_context_get(struct intel_context *ce)
 {

drivers/gpu/drm/i915/gt/intel_engine_pm.c

@@ -90,7 +90,7 @@ static bool switch_to_kernel_context(struct intel_engine_cs *engine)
 	/* Check again on the next retirement. */
 	engine->wakeref_serial = engine->serial + 1;
 
-	i915_request_add_barriers(rq);
+	i915_request_add_active_barriers(rq);
 	__i915_request_commit(rq);
 
 	return false;

drivers/gpu/drm/i915/gt/selftest_context.c

+/*
+ * SPDX-License-Identifier: GPL-2.0
+ *
+ * Copyright © 2019 Intel Corporation
+ */
+
+#include "i915_selftest.h"
+#include "intel_gt.h"
+
+#include "gem/selftests/mock_context.h"
+#include "selftests/igt_flush_test.h"
+#include "selftests/mock_drm.h"
+
+static int request_sync(struct i915_request *rq)
+{
+	long timeout;
+	int err = 0;
+
+	i915_request_get(rq);
+
+	i915_request_add(rq);
+	timeout = i915_request_wait(rq, 0, HZ / 10);
+	if (timeout < 0)
+		err = timeout;
+	else
+		i915_request_retire_upto(rq);
+
+	i915_request_put(rq);
+
+	return err;
+}
+
+static int context_sync(struct intel_context *ce)
+{
+	struct intel_timeline *tl = ce->ring->timeline;
+	int err = 0;
+
+	do {
+		struct i915_request *rq;
+		long timeout;
+
+		rcu_read_lock();
+		rq = rcu_dereference(tl->last_request.request);
+		if (rq)
+			rq = i915_request_get_rcu(rq);
+		rcu_read_unlock();
+		if (!rq)
+			break;
+
+		timeout = i915_request_wait(rq, 0, HZ / 10);
+		if (timeout < 0)
+			err = timeout;
+		else
+			i915_request_retire_upto(rq);
+
+		i915_request_put(rq);
+	} while (!err);
+
+	return err;
+}
+
+static int __live_active_context(struct intel_engine_cs *engine,
+				 struct i915_gem_context *fixme)
+{
+	struct intel_context *ce;
+	int pass;
+	int err;
+
+	/*
+	 * We keep active contexts alive until after a subsequent context
+	 * switch as the final write from the context-save will be after
+	 * we retire the final request. We track when we unpin the context,
+	 * under the presumption that the final pin is from the last request,
+	 * and instead of immediately unpinning the context, we add a task
+	 * to unpin the context from the next idle-barrier.
+	 *
+	 * This test makes sure that the context is kept alive until a
+	 * subsequent idle-barrier (emitted when the engine wakeref hits 0
+	 * with no more outstanding requests).
+	 */
+
+	if (intel_engine_pm_is_awake(engine)) {
+		pr_err("%s is awake before starting %s!\n",
+		       engine->name, __func__);
+		return -EINVAL;
+	}
+
+	ce = intel_context_create(fixme, engine);
+	if (!ce)
+		return -ENOMEM;
+
+	for (pass = 0; pass <= 2; pass++) {
+		struct i915_request *rq;
+
+		rq = intel_context_create_request(ce);
+		if (IS_ERR(rq)) {
+			err = PTR_ERR(rq);
+			goto err;
+		}
+
+		err = request_sync(rq);
+		if (err)
+			goto err;
+
+		/* Context will be kept active until after an idle-barrier. */
+		if (i915_active_is_idle(&ce->active)) {
+			pr_err("context is not active; expected idle-barrier (%s pass %d)\n",
+			       engine->name, pass);
+			err = -EINVAL;
+			goto err;
+		}
+
+		if (!intel_engine_pm_is_awake(engine)) {
+			pr_err("%s is asleep before idle-barrier\n",
+			       engine->name);
+			err = -EINVAL;
+			goto err;
+		}
+	}
+
+	/* Now make sure our idle-barriers are flushed */
+	err = context_sync(engine->kernel_context);
+	if (err)
+		goto err;
+
+	if (!i915_active_is_idle(&ce->active)) {
+		pr_err("context is still active!");
+		err = -EINVAL;
+	}
+
+	if (intel_engine_pm_is_awake(engine)) {
+		struct drm_printer p = drm_debug_printer(__func__);
+
+		intel_engine_dump(engine, &p,
+				  "%s is still awake after idle-barriers\n",
+				  engine->name);
+		GEM_TRACE_DUMP();
+
+		err = -EINVAL;
+		goto err;
+	}
+
+err:
+	intel_context_put(ce);
+	return err;
+}
+
+static int live_active_context(void *arg)
+{
+	struct intel_gt *gt = arg;
+	struct intel_engine_cs *engine;
+	struct i915_gem_context *fixme;
+	enum intel_engine_id id;
+	struct drm_file *file;
+	int err = 0;
+
+	file = mock_file(gt->i915);
+	if (IS_ERR(file))
+		return PTR_ERR(file);
+
+	mutex_lock(&gt->i915->drm.struct_mutex);
+
+	fixme = live_context(gt->i915, file);
+	if (!fixme) {
+		err = -ENOMEM;
+		goto unlock;
+	}
+
+	for_each_engine(engine, gt->i915, id) {
+		err = __live_active_context(engine, fixme);
+		if (err)
+			break;
+
+		err = igt_flush_test(gt->i915, I915_WAIT_LOCKED);
+		if (err)
+			break;
+	}
+
+unlock:
+	mutex_unlock(&gt->i915->drm.struct_mutex);
+	mock_file_free(gt->i915, file);
+	return err;
+}
+
+static int __remote_sync(struct intel_context *ce, struct intel_context *remote)
+{
+	struct i915_request *rq;
+	int err;
+
+	err = intel_context_pin(remote);
+	if (err)
+		return err;
+
+	rq = intel_context_create_request(ce);
+	if (IS_ERR(rq)) {
+		err = PTR_ERR(rq);
+		goto unpin;
+	}
+
+	err = intel_context_prepare_remote_request(remote, rq);
+	if (err) {
+		i915_request_add(rq);
+		goto unpin;
+	}
+
+	err = request_sync(rq);
+
+unpin:
+	intel_context_unpin(remote);
+	return err;
+}
+
+static int __live_remote_context(struct intel_engine_cs *engine,
+				 struct i915_gem_context *fixme)
+{
+	struct intel_context *local, *remote;
+	int pass;
+	int err;
+
+	/*
+	 * Check that our idle barriers do not interfere with normal
+	 * activity tracking. In particular, check that operating
+	 * on the context image remotely (intel_context_prepare_remote_request),
+	 * which inserts foreign fences into intel_context.active, does not
+	 * clobber the idle-barrier.
+	 */
+
+	remote = intel_context_create(fixme, engine);
+	if (!remote)
+		return -ENOMEM;
+
+	local = intel_context_create(fixme, engine);
+	if (!local) {
+		err = -ENOMEM;
+		goto err_remote;
+	}
+
+	for (pass = 0; pass <= 2; pass++) {
+		err = __remote_sync(local, remote);
+		if (err)
+			break;
+
+		err = __remote_sync(engine->kernel_context, remote);
+		if (err)
+			break;
+
+		if (i915_active_is_idle(&remote->active)) {
+			pr_err("remote context is not active; expected idle-barrier (%s pass %d)\n",
+			       engine->name, pass);
+			err = -EINVAL;
+			break;
+		}
+	}
+
+	intel_context_put(local);
+err_remote:
+	intel_context_put(remote);
+	return err;
+}
+
+static int live_remote_context(void *arg)
+{
+	struct intel_gt *gt = arg;
+	struct intel_engine_cs *engine;
+	struct i915_gem_context *fixme;
+	enum intel_engine_id id;
+	struct drm_file *file;
+	int err = 0;
+
+	file = mock_file(gt->i915);
+	if (IS_ERR(file))
+		return PTR_ERR(file);
+
+	mutex_lock(&gt->i915->drm.struct_mutex);
+
+	fixme = live_context(gt->i915, file);
+	if (!fixme) {
+		err = -ENOMEM;
+		goto unlock;
+	}
+
+	for_each_engine(engine, gt->i915, id) {
+		err = __live_remote_context(engine, fixme);
+		if (err)
+			break;
+
+		err = igt_flush_test(gt->i915, I915_WAIT_LOCKED);
+		if (err)
+			break;
+	}
+
+unlock:
+	mutex_unlock(&gt->i915->drm.struct_mutex);
+	mock_file_free(gt->i915, file);
+	return err;
+}
+
+int intel_context_live_selftests(struct drm_i915_private *i915)
+{
+	static const struct i915_subtest tests[] = {
+		SUBTEST(live_active_context),
+		SUBTEST(live_remote_context),
+	};
+	struct intel_gt *gt = &i915->gt;
+
+	if (intel_gt_is_wedged(gt))
+		return 0;
+
+	return intel_gt_live_subtests(tests, gt);
+}

drivers/gpu/drm/i915/i915_active.h

@@ -413,6 +413,6 @@ static inline void i915_active_fini(struct i915_active *ref) { }
 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 					    struct intel_engine_cs *engine);
 void i915_active_acquire_barrier(struct i915_active *ref);
-void i915_request_add_barriers(struct i915_request *rq);
+void i915_request_add_active_barriers(struct i915_request *rq);
 
 #endif /* _I915_ACTIVE_H_ */

drivers/gpu/drm/i915/i915_active_types.h

@@ -42,7 +42,7 @@ struct i915_active {
 	int (*active)(struct i915_active *ref);
 	void (*retire)(struct i915_active *ref);
 
-	struct llist_head barriers;
+	struct llist_head preallocated_barriers;
 };
 
 #endif /* _I915_ACTIVE_TYPES_H_ */

drivers/gpu/drm/i915/selftests/i915_live_selftests.h

@@ -15,6 +15,7 @@ selftest(workarounds, intel_workarounds_live_selftests)
 selftest(timelines, intel_timeline_live_selftests)
 selftest(requests, i915_request_live_selftests)
 selftest(active, i915_active_live_selftests)
+selftest(gt_contexts, intel_context_live_selftests)
 selftest(objects, i915_gem_object_live_selftests)
 selftest(mman, i915_gem_mman_live_selftests)
 selftest(dmabuf, i915_gem_dmabuf_live_selftests)
@@ -24,7 +25,7 @@ selftest(gtt, i915_gem_gtt_live_selftests)
 selftest(gem, i915_gem_live_selftests)
 selftest(evict, i915_gem_evict_live_selftests)
 selftest(hugepages, i915_gem_huge_page_live_selftests)
-selftest(contexts, i915_gem_context_live_selftests)
+selftest(gem_contexts, i915_gem_context_live_selftests)
 selftest(blt, i915_gem_object_blt_live_selftests)
 selftest(client, i915_gem_client_blt_live_selftests)
 selftest(reset, intel_reset_live_selftests)