1 files changed, 209 insertions, 1 deletions

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c
index 7b01f4605f21..de6e55af82cf 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_context.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c
@@ -69,8 +69,10 @@
 
 #include <drm/i915_drm.h>
 
-#include "gt/intel_lrc_reg.h"
+#include "gt/intel_engine_heartbeat.h"
 #include "gt/intel_engine_user.h"
+#include "gt/intel_lrc_reg.h"
+#include "gt/intel_ring.h"
 
 #include "i915_gem_context.h"
 #include "i915_globals.h"
@@ -276,6 +278,153 @@ void i915_gem_context_release(struct kref *ref)
 		schedule_work(&gc->free_work);
 }
 
+static inline struct i915_gem_engines *
+__context_engines_static(const struct i915_gem_context *ctx)
+{
+	return rcu_dereference_protected(ctx->engines, true);
+}
+
+static bool __reset_engine(struct intel_engine_cs *engine)
+{
+	struct intel_gt *gt = engine->gt;
+	bool success = false;
+
+	if (!intel_has_reset_engine(gt))
+		return false;
+
+	if (!test_and_set_bit(I915_RESET_ENGINE + engine->id,
+			      &gt->reset.flags)) {
+		success = intel_engine_reset(engine, NULL) == 0;
+		clear_and_wake_up_bit(I915_RESET_ENGINE + engine->id,
+				      &gt->reset.flags);
+	}
+
+	return success;
+}
+
+static void __reset_context(struct i915_gem_context *ctx,
+			    struct intel_engine_cs *engine)
+{
+	intel_gt_handle_error(engine->gt, engine->mask, 0,
+			      "context closure in %s", ctx->name);
+}
+
+static bool __cancel_engine(struct intel_engine_cs *engine)
+{
+	/*
+	 * Send a "high priority pulse" down the engine to cause the
+	 * current request to be momentarily preempted. (If it fails to
+	 * be preempted, it will be reset). As we have marked our context
+	 * as banned, any incomplete request, including any running, will
+	 * be skipped following the preemption.
+	 *
+	 * If there is no hangchecking (one of the reasons why we try to
+	 * cancel the context) and no forced preemption, there may be no
+	 * means by which we reset the GPU and evict the persistent hog.
+	 * Ergo if we are unable to inject a preemptive pulse that can
+	 * kill the banned context, we fallback to doing a local reset
+	 * instead.
+	 */
+	if (IS_ACTIVE(CONFIG_DRM_I915_PREEMPT_TIMEOUT) &&
+	    !intel_engine_pulse(engine))
+		return true;
+
+	/* If we are unable to send a pulse, try resetting this engine. */
+	return __reset_engine(engine);
+}
+
+static struct intel_engine_cs *__active_engine(struct i915_request *rq)
+{
+	struct intel_engine_cs *engine, *locked;
+
+	/*
+	 * Serialise with __i915_request_submit() so that it sees
+	 * is-banned?, or we know the request is already inflight.
+	 */
+	locked = READ_ONCE(rq->engine);
+	spin_lock_irq(&locked->active.lock);
+	while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) {
+		spin_unlock(&locked->active.lock);
+		spin_lock(&engine->active.lock);
+		locked = engine;
+	}
+
+	engine = NULL;
+	if (i915_request_is_active(rq) && !rq->fence.error)
+		engine = rq->engine;
+
+	spin_unlock_irq(&locked->active.lock);
+
+	return engine;
+}
+
+static struct intel_engine_cs *active_engine(struct intel_context *ce)
+{
+	struct intel_engine_cs *engine = NULL;
+	struct i915_request *rq;
+
+	if (!ce->timeline)
+		return NULL;
+
+	rcu_read_lock();
+	list_for_each_entry_reverse(rq, &ce->timeline->requests, link) {
+		if (i915_request_completed(rq))
+			break;
+
+		/* Check with the backend if the request is inflight */
+		engine = __active_engine(rq);
+		if (engine)
+			break;
+	}
+	rcu_read_unlock();
+
+	return engine;
+}
+
+static void kill_context(struct i915_gem_context *ctx)
+{
+	struct i915_gem_engines_iter it;
+	struct intel_context *ce;
+
+	/*
+	 * If we are already banned, it was due to a guilty request causing
+	 * a reset and the entire context being evicted from the GPU.
+	 */
+	if (i915_gem_context_is_banned(ctx))
+		return;
+
+	i915_gem_context_set_banned(ctx);
+
+	/*
+	 * Map the user's engine back to the actual engines; one virtual
+	 * engine will be mapped to multiple engines, and using ctx->engine[]
+	 * the same engine may be have multiple instances in the user's map.
+	 * However, we only care about pending requests, so only include
+	 * engines on which there are incomplete requests.
+	 */
+	for_each_gem_engine(ce, __context_engines_static(ctx), it) {
+		struct intel_engine_cs *engine;
+
+		/*
+		 * Check the current active state of this context; if we
+		 * are currently executing on the GPU we need to evict
+		 * ourselves. On the other hand, if we haven't yet been
+		 * submitted to the GPU or if everything is complete,
+		 * we have nothing to do.
+		 */
+		engine = active_engine(ce);
+
+		/* First attempt to gracefully cancel the context */
+		if (engine && !__cancel_engine(engine))
+			/*
+			 * If we are unable to send a preemptive pulse to bump
+			 * the context from the GPU, we have to resort to a full
+			 * reset. We hope the collateral damage is worth it.
+			 */
+			__reset_context(ctx, engine);
+	}
+}
+
 static void context_close(struct i915_gem_context *ctx)
 {
 	struct i915_address_space *vm;
@@ -298,9 +447,47 @@ static void context_close(struct i915_gem_context *ctx)
 	lut_close(ctx);
 
 	mutex_unlock(&ctx->mutex);
+
+	/*
+	 * If the user has disabled hangchecking, we can not be sure that
+	 * the batches will ever complete after the context is closed,
+	 * keeping the context and all resources pinned forever. So in this
+	 * case we opt to forcibly kill off all remaining requests on
+	 * context close.
+	 */
+	if (!i915_gem_context_is_persistent(ctx) ||
+	    !i915_modparams.enable_hangcheck)
+		kill_context(ctx);
+
 	i915_gem_context_put(ctx);
 }
 
+static int __context_set_persistence(struct i915_gem_context *ctx, bool state)
+{
+	if (i915_gem_context_is_persistent(ctx) == state)
+		return 0;
+
+	if (state) {
+		/*
+		 * Only contexts that are short-lived [that will expire or be
+		 * reset] are allowed to survive past termination. We require
+		 * hangcheck to ensure that the persistent requests are healthy.
+		 */
+		if (!i915_modparams.enable_hangcheck)
+			return -EINVAL;
+
+		i915_gem_context_set_persistence(ctx);
+	} else {
+		/* To cancel a context we use "preempt-to-idle" */
+		if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PREEMPTION))
+			return -ENODEV;
+
+		i915_gem_context_clear_persistence(ctx);
+	}
+
+	return 0;
+}
+
 static struct i915_gem_context *
 __create_context(struct drm_i915_private *i915)
 {
@@ -335,6 +522,7 @@ __create_context(struct drm_i915_private *i915)
 
 	i915_gem_context_set_bannable(ctx);
 	i915_gem_context_set_recoverable(ctx);
+	__context_set_persistence(ctx, true /* cgroup hook? */);
 
 	for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
 		ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
@@ -491,6 +679,7 @@ i915_gem_context_create_kernel(struct drm_i915_private *i915, int prio)
 		return ctx;
 
 	i915_gem_context_clear_bannable(ctx);
+	i915_gem_context_set_persistence(ctx);
 	ctx->sched.priority = I915_USER_PRIORITY(prio);
 
 	GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
@@ -1601,6 +1790,16 @@ err_free:
 	return err;
 }
 
+static int
+set_persistence(struct i915_gem_context *ctx,
+		const struct drm_i915_gem_context_param *args)
+{
+	if (args->size)
+		return -EINVAL;
+
+	return __context_set_persistence(ctx, args->value);
+}
+
 static int ctx_setparam(struct drm_i915_file_private *fpriv,
 			struct i915_gem_context *ctx,
 			struct drm_i915_gem_context_param *args)
@@ -1678,6 +1877,10 @@ static int ctx_setparam(struct drm_i915_file_private *fpriv,
 		ret = set_engines(ctx, args);
 		break;
 
+	case I915_CONTEXT_PARAM_PERSISTENCE:
+		ret = set_persistence(ctx, args);
+		break;
+
 	case I915_CONTEXT_PARAM_BAN_PERIOD:
 	default:
 		ret = -EINVAL;
@@ -2130,6 +2333,11 @@ int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
 		ret = get_engines(ctx, args);
 		break;
 
+	case I915_CONTEXT_PARAM_PERSISTENCE:
+		args->size = 0;
+		args->value = i915_gem_context_is_persistent(ctx);
+		break;
+
 	case I915_CONTEXT_PARAM_BAN_PERIOD:
 	default:
 		ret = -EINVAL;