drm/i915: Serialise with engine-pm around requests on the kernel_context

As the engine->kernel_context is used within the engine-pm barrier, we
have to be careful when emitting requests outside of the barrier, as the
strict timeline locking rules do not apply. Instead, we must ensure the
engine_park() cannot be entered as we build the request, which is
simplest by taking an explicit engine-pm wakeref around the request
construction.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191125105858.1718307-1-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/gem/i915_gem_context.c

@@ -70,6 +70,7 @@
 #include <drm/i915_drm.h>
 
 #include "gt/intel_engine_heartbeat.h"
+#include "gt/intel_engine_pm.h"
 #include "gt/intel_engine_user.h"
 #include "gt/intel_lrc_reg.h"
 #include "gt/intel_ring.h"
@@ -1265,7 +1266,7 @@ gen8_modify_rpcs(struct intel_context *ce, struct intel_sseu sseu)
 	if (!intel_context_is_pinned(ce))
 		return 0;
 
-	rq = i915_request_create(ce->engine->kernel_context);
+	rq = intel_engine_create_kernel_request(ce->engine);
 	if (IS_ERR(rq))
 		return PTR_ERR(rq);
 

drivers/gpu/drm/i915/gem/selftests/i915_gem_client_blt.c

@@ -24,6 +24,7 @@ static int __igt_client_fill(struct intel_engine_cs *engine)
 
 	prandom_seed_state(&prng, i915_selftest.random_seed);
 
+	intel_engine_pm_get(engine);
 	do {
 		const u32 max_block_size = S16_MAX * PAGE_SIZE;
 		u32 sz = min_t(u64, ce->vm->total >> 4, prandom_u32_state(&prng));
@@ -99,6 +100,7 @@ static int __igt_client_fill(struct intel_engine_cs *engine)
 err_flush:
 	if (err == -ENOMEM)
 		err = 0;
+	intel_engine_pm_put(engine);
 
 	return err;
 }

drivers/gpu/drm/i915/gem/selftests/i915_gem_coherency.c

@@ -6,6 +6,7 @@
 
 #include <linux/prime_numbers.h>
 
+#include "gt/intel_engine_pm.h"
 #include "gt/intel_gt.h"
 #include "gt/intel_gt_pm.h"
 #include "gt/intel_ring.h"
@@ -200,7 +201,7 @@ static int gpu_set(struct context *ctx, unsigned long offset, u32 v)
 	if (IS_ERR(vma))
 		return PTR_ERR(vma);
 
-	rq = i915_request_create(ctx->engine->kernel_context);
+	rq = intel_engine_create_kernel_request(ctx->engine);
 	if (IS_ERR(rq)) {
 		i915_vma_unpin(vma);
 		return PTR_ERR(rq);

drivers/gpu/drm/i915/gem/selftests/i915_gem_context.c

@@ -7,6 +7,7 @@
 #include <linux/prime_numbers.h>
 
 #include "gem/i915_gem_pm.h"
+#include "gt/intel_engine_pm.h"
 #include "gt/intel_gt.h"
 #include "gt/intel_gt_requests.h"
 #include "gt/intel_reset.h"
@@ -1190,9 +1191,11 @@ __sseu_test(const char *name,
 	struct igt_spinner *spin = NULL;
 	int ret;
 
+	intel_engine_pm_get(ce->engine);
+
 	ret = __sseu_prepare(name, flags, ce, &spin);
 	if (ret)
-		return ret;
+		goto out_pm;
 
 	ret = intel_context_reconfigure_sseu(ce, sseu);
 	if (ret)
@@ -1207,6 +1210,8 @@ __sseu_test(const char *name,
 		igt_spinner_fini(spin);
 		kfree(spin);
 	}
+out_pm:
+	intel_engine_pm_put(ce->engine);
 	return ret;
 }
 

drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c

@@ -6,6 +6,7 @@
 
 #include <linux/prime_numbers.h>
 
+#include "gt/intel_engine_pm.h"
 #include "gt/intel_gt.h"
 #include "gt/intel_gt_pm.h"
 #include "huge_gem_object.h"
@@ -536,7 +537,7 @@ static int make_obj_busy(struct drm_i915_gem_object *obj)
 		if (err)
 			return err;
 
-		rq = i915_request_create(engine->kernel_context);
+		rq = intel_engine_create_kernel_request(engine);
 		if (IS_ERR(rq)) {
 			i915_vma_unpin(vma);
 			return PTR_ERR(rq);

drivers/gpu/drm/i915/gem/selftests/i915_gem_object_blt.c

@@ -41,6 +41,7 @@ static int __perf_fill_blt(struct drm_i915_gem_object *obj)
 		if (!engine)
 			return 0;
 
+		intel_engine_pm_get(engine);
 		for (pass = 0; pass < ARRAY_SIZE(t); pass++) {
 			struct intel_context *ce = engine->kernel_context;
 			ktime_t t0, t1;
@@ -49,17 +50,20 @@ static int __perf_fill_blt(struct drm_i915_gem_object *obj)
 
 			err = i915_gem_object_fill_blt(obj, ce, 0);
 			if (err)
-				return err;
+				break;
 
 			err = i915_gem_object_wait(obj,
 						   I915_WAIT_ALL,
 						   MAX_SCHEDULE_TIMEOUT);
 			if (err)
-				return err;
+				break;
 
 			t1 = ktime_get();
 			t[pass] = ktime_sub(t1, t0);
 		}
+		intel_engine_pm_put(engine);
+		if (err)
+			return err;
 
 		sort(t, ARRAY_SIZE(t), sizeof(*t), wrap_ktime_compare, NULL);
 		pr_info("%s: blt %zd KiB fill: %lld MiB/s\n",
@@ -109,33 +113,37 @@ static int __perf_copy_blt(struct drm_i915_gem_object *src,
 		struct intel_engine_cs *engine;
 		ktime_t t[5];
 		int pass;
+		int err = 0;
 
 		engine = intel_engine_lookup_user(i915,
 						  I915_ENGINE_CLASS_COPY,
 						  inst++);
 		if (!engine)
 			return 0;
 
+		intel_engine_pm_get(engine);
 		for (pass = 0; pass < ARRAY_SIZE(t); pass++) {
 			struct intel_context *ce = engine->kernel_context;
 			ktime_t t0, t1;
-			int err;
 
 			t0 = ktime_get();
 
 			err = i915_gem_object_copy_blt(src, dst, ce);
 			if (err)
-				return err;
+				break;
 
 			err = i915_gem_object_wait(dst,
 						   I915_WAIT_ALL,
 						   MAX_SCHEDULE_TIMEOUT);
 			if (err)
-				return err;
+				break;
 
 			t1 = ktime_get();
 			t[pass] = ktime_sub(t1, t0);
 		}
+		intel_engine_pm_put(engine);
+		if (err)
+			return err;
 
 		sort(t, ARRAY_SIZE(t), sizeof(*t), wrap_ktime_compare, NULL);
 		pr_info("%s: blt %zd KiB copy: %lld MiB/s\n",

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

@@ -215,18 +215,26 @@ int intel_engine_pulse(struct intel_engine_cs *engine)
 int intel_engine_flush_barriers(struct intel_engine_cs *engine)
 {
 	struct i915_request *rq;
+	int err = 0;
 
 	if (llist_empty(&engine->barrier_tasks))
 		return 0;
 
+	if (!intel_engine_pm_get_if_awake(engine))
+		return 0;
+
 	rq = i915_request_create(engine->kernel_context);
-	if (IS_ERR(rq))
-		return PTR_ERR(rq);
+	if (IS_ERR(rq)) {
+		err = PTR_ERR(rq);
+		goto out_rpm;
+	}
 
 	idle_pulse(engine, rq);
 	i915_request_add(rq);
 
-	return 0;
+out_rpm:
+	intel_engine_pm_put(engine);
+	return err;
 }
 
 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)

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

@@ -7,6 +7,7 @@
 #ifndef INTEL_ENGINE_PM_H
 #define INTEL_ENGINE_PM_H
 
+#include "i915_request.h"
 #include "intel_engine_types.h"
 #include "intel_wakeref.h"
 
@@ -41,6 +42,26 @@ static inline void intel_engine_pm_flush(struct intel_engine_cs *engine)
 	intel_wakeref_unlock_wait(&engine->wakeref);
 }
 
+static inline struct i915_request *
+intel_engine_create_kernel_request(struct intel_engine_cs *engine)
+{
+	struct i915_request *rq;
+
+	/*
+	 * The engine->kernel_context is special as it is used inside
+	 * the engine-pm barrier (see __engine_park()), circumventing
+	 * the usual mutexes and relying on the engine-pm barrier
+	 * instead. So whenever we use the engine->kernel_context
+	 * outside of the barrier, we must manually handle the
+	 * engine wakeref to serialise with the use inside.
+	 */
+	intel_engine_pm_get(engine);
+	rq = i915_request_create(engine->kernel_context);
+	intel_engine_pm_put(engine);
+
+	return rq;
+}
+
 void intel_engine_init__pm(struct intel_engine_cs *engine);
 
 #endif /* INTEL_ENGINE_PM_H */

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

@@ -6,6 +6,7 @@
 
 #include "i915_drv.h"
 #include "intel_context.h"
+#include "intel_engine_pm.h"
 #include "intel_gt.h"
 #include "intel_ring.h"
 #include "intel_workarounds.h"
@@ -1582,7 +1583,9 @@ static int engine_wa_list_verify(struct intel_context *ce,
 	if (IS_ERR(vma))
 		return PTR_ERR(vma);
 
+	intel_engine_pm_get(ce->engine);
 	rq = intel_context_create_request(ce);
+	intel_engine_pm_put(ce->engine);
 	if (IS_ERR(rq)) {
 		err = PTR_ERR(rq);
 		goto err_vma;

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

@@ -121,7 +121,7 @@ static int __live_context_size(struct intel_engine_cs *engine,
 		goto err_unpin;
 
 	/* Force the context switch */
-	rq = i915_request_create(engine->kernel_context);
+	rq = intel_engine_create_kernel_request(engine);
 	if (IS_ERR(rq)) {
 		err = PTR_ERR(rq);
 		goto err_unpin;

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

@@ -132,14 +132,18 @@ static int perf_mi_bb_start(void *arg)
 		u32 cycles[COUNT];
 		int i;
 
+		intel_engine_pm_get(engine);
+
 		batch = create_empty_batch(ce);
 		if (IS_ERR(batch)) {
 			err = PTR_ERR(batch);
+			intel_engine_pm_put(engine);
 			break;
 		}
 
 		err = i915_vma_sync(batch);
 		if (err) {
+			intel_engine_pm_put(engine);
 			i915_vma_put(batch);
 			break;
 		}
@@ -180,6 +184,7 @@ static int perf_mi_bb_start(void *arg)
 			cycles[i] = rq->hwsp_seqno[3] - rq->hwsp_seqno[2];
 		}
 		i915_vma_put(batch);
+		intel_engine_pm_put(engine);
 		if (err)
 			break;
 
@@ -251,29 +256,35 @@ static int perf_mi_noop(void *arg)
 		u32 cycles[COUNT];
 		int i;
 
+		intel_engine_pm_get(engine);
+
 		base = create_empty_batch(ce);
 		if (IS_ERR(base)) {
 			err = PTR_ERR(base);
+			intel_engine_pm_put(engine);
 			break;
 		}
 
 		err = i915_vma_sync(base);
 		if (err) {
 			i915_vma_put(base);
+			intel_engine_pm_put(engine);
 			break;
 		}
 
 		nop = create_nop_batch(ce);
 		if (IS_ERR(nop)) {
 			err = PTR_ERR(nop);
 			i915_vma_put(base);
+			intel_engine_pm_put(engine);
 			break;
 		}
 
 		err = i915_vma_sync(nop);
 		if (err) {
 			i915_vma_put(nop);
 			i915_vma_put(base);
+			intel_engine_pm_put(engine);
 			break;
 		}
 
@@ -327,6 +338,7 @@ static int perf_mi_noop(void *arg)
 		}
 		i915_vma_put(nop);
 		i915_vma_put(base);
+		intel_engine_pm_put(engine);
 		if (err)
 			break;
 

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

@@ -348,7 +348,7 @@ release_queue(struct intel_engine_cs *engine,
 	struct i915_request *rq;
 	u32 *cs;
 
-	rq = i915_request_create(engine->kernel_context);
+	rq = intel_engine_create_kernel_request(engine);
 	if (IS_ERR(rq))
 		return PTR_ERR(rq);
 
@@ -497,7 +497,7 @@ static struct i915_request *nop_request(struct intel_engine_cs *engine)
 {
 	struct i915_request *rq;
 
-	rq = i915_request_create(engine->kernel_context);
+	rq = intel_engine_create_kernel_request(engine);
 	if (IS_ERR(rq))
 		return rq;
 
@@ -3698,7 +3698,7 @@ static int gpr_make_dirty(struct intel_engine_cs *engine)
 	u32 *cs;
 	int n;
 
-	rq = i915_request_create(engine->kernel_context);
+	rq = intel_engine_create_kernel_request(engine);
 	if (IS_ERR(rq))
 		return PTR_ERR(rq);
 

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

@@ -261,7 +261,9 @@ static int live_mocs_kernel(void *arg)
 		return err;
 
 	for_each_engine(engine, gt, id) {
+		intel_engine_pm_get(engine);
 		err = check_mocs_engine(&mocs, engine->kernel_context);
+		intel_engine_pm_put(engine);
 		if (err)
 			break;
 	}

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

@@ -458,7 +458,7 @@ tl_write(struct intel_timeline *tl, struct intel_engine_cs *engine, u32 value)
 		goto out;
 	}
 
-	rq = i915_request_create(engine->kernel_context);
+	rq = intel_engine_create_kernel_request(engine);
 	if (IS_ERR(rq))
 		goto out_unpin;
 
@@ -675,9 +675,7 @@ static int live_hwsp_wrap(void *arg)
 		if (!intel_engine_can_store_dword(engine))
 			continue;
 
-		intel_engine_pm_get(engine);
-		rq = i915_request_create(engine->kernel_context);
-		intel_engine_pm_put(engine);
+		rq = intel_engine_create_kernel_request(engine);
 		if (IS_ERR(rq)) {
 			err = PTR_ERR(rq);
 			goto out;