drm/i915: Introduce intel_context.pin_mutex for pin management

Introduce a mutex to start locking the HW contexts independently of
struct_mutex, with a view to reducing the coarse struct_mutex. The
intel_context.pin_mutex is used to guard the transition to and from being
pinned on the gpu, and so is required before starting to build any
request. The intel_context will then remain pinned until the request
completes, but the mutex can be released immediately unpin completion of
pinning the context.

A slight variant of the above is used by per-context sseu that wants to
inspect the pinned status of the context, and requires that it remains
stable (either !pinned or pinned) across its operation. By using the
pin_mutex to serialise operations while pin_count==0, we can take that
pin_mutex for stabilise the boolean pin status.

v2: for Tvrtko!
* Improved commit message.
* Dropped _gpu suffix from gen8_modify_rpcs_gpu.
v3: Repair the locking for sseu selftests
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190308132522.21573-7-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/i915_gem.c

@@ -4667,7 +4667,7 @@ static int __intel_engines_record_defaults(struct drm_i915_private *i915)
 		if (!state)
 			continue;
 
-		GEM_BUG_ON(ce->pin_count);
+		GEM_BUG_ON(intel_context_is_pinned(ce));
 
 		/*
 		 * As we will hold a reference to the logical state, it will

drivers/gpu/drm/i915/i915_gem_context.c

@@ -810,7 +810,6 @@ static int get_sseu(struct i915_gem_context *ctx,
 	struct drm_i915_gem_context_param_sseu user_sseu;
 	struct intel_engine_cs *engine;
 	struct intel_context *ce;
-	int ret;
 
 	if (args->size == 0)
 		goto out;
@@ -830,21 +829,16 @@ static int get_sseu(struct i915_gem_context *ctx,
 	if (!engine)
 		return -EINVAL;
 
-	ce = intel_context_instance(ctx, engine);
+	ce = intel_context_pin_lock(ctx, engine); /* serialises with set_sseu */
 	if (IS_ERR(ce))
 		return PTR_ERR(ce);
 
-	/* Only use for mutex here is to serialize get_param and set_param. */
-	ret = mutex_lock_interruptible(&ctx->i915->drm.struct_mutex);
-	if (ret)
-		return ret;
-
 	user_sseu.slice_mask = ce->sseu.slice_mask;
 	user_sseu.subslice_mask = ce->sseu.subslice_mask;
 	user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice;
 	user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice;
 
-	mutex_unlock(&ctx->i915->drm.struct_mutex);
+	intel_context_pin_unlock(ce);
 
 	if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu,
 			 sizeof(user_sseu)))
@@ -940,23 +934,28 @@ static int gen8_emit_rpcs_config(struct i915_request *rq,
 }
 
 static int
-gen8_modify_rpcs_gpu(struct intel_context *ce,
-		     struct intel_engine_cs *engine,
-		     struct intel_sseu sseu)
+gen8_modify_rpcs(struct intel_context *ce, struct intel_sseu sseu)
 {
-	struct drm_i915_private *i915 = engine->i915;
+	struct drm_i915_private *i915 = ce->engine->i915;
 	struct i915_request *rq, *prev;
 	intel_wakeref_t wakeref;
 	int ret;
 
-	GEM_BUG_ON(!ce->pin_count);
+	lockdep_assert_held(&ce->pin_mutex);
 
-	lockdep_assert_held(&i915->drm.struct_mutex);
+	/*
+	 * If the context is not idle, we have to submit an ordered request to
+	 * modify its context image via the kernel context (writing to our own
+	 * image, or into the registers directory, does not stick). Pristine
+	 * and idle contexts will be configured on pinning.
+	 */
+	if (!intel_context_is_pinned(ce))
+		return 0;
 
 	/* Submitting requests etc needs the hw awake. */
 	wakeref = intel_runtime_pm_get(i915);
 
-	rq = i915_request_alloc(engine, i915->kernel_context);
+	rq = i915_request_alloc(ce->engine, i915->kernel_context);
 	if (IS_ERR(rq)) {
 		ret = PTR_ERR(rq);
 		goto out_put;
@@ -1010,25 +1009,20 @@ __i915_gem_context_reconfigure_sseu(struct i915_gem_context *ctx,
 	GEM_BUG_ON(INTEL_GEN(ctx->i915) < 8);
 	GEM_BUG_ON(engine->id != RCS0);
 
-	ce = intel_context_instance(ctx, engine);
+	ce = intel_context_pin_lock(ctx, engine);
 	if (IS_ERR(ce))
 		return PTR_ERR(ce);
 
 	/* Nothing to do if unmodified. */
 	if (!memcmp(&ce->sseu, &sseu, sizeof(sseu)))
-		return 0;
-
-	/*
-	 * If context is not idle we have to submit an ordered request to modify
-	 * its context image via the kernel context. Pristine and idle contexts
-	 * will be configured on pinning.
-	 */
-	if (ce->pin_count)
-		ret = gen8_modify_rpcs_gpu(ce, engine, sseu);
+		goto unlock;
 
+	ret = gen8_modify_rpcs(ce, sseu);
 	if (!ret)
 		ce->sseu = sseu;
 
+unlock:
+	intel_context_pin_unlock(ce);
 	return ret;
 }
 

drivers/gpu/drm/i915/intel_context.c

@@ -102,7 +102,7 @@ void __intel_context_remove(struct intel_context *ce)
 	spin_unlock(&ctx->hw_contexts_lock);
 }
 
-struct intel_context *
+static struct intel_context *
 intel_context_instance(struct i915_gem_context *ctx,
 		       struct intel_engine_cs *engine)
 {
@@ -126,6 +126,23 @@ intel_context_instance(struct i915_gem_context *ctx,
 	return pos;
 }
 
+struct intel_context *
+intel_context_pin_lock(struct i915_gem_context *ctx,
+		       struct intel_engine_cs *engine)
+	__acquires(ce->pin_mutex)
+{
+	struct intel_context *ce;
+
+	ce = intel_context_instance(ctx, engine);
+	if (IS_ERR(ce))
+		return ce;
+
+	if (mutex_lock_interruptible(&ce->pin_mutex))
+		return ERR_PTR(-EINTR);
+
+	return ce;
+}
+
 struct intel_context *
 intel_context_pin(struct i915_gem_context *ctx,
 		  struct intel_engine_cs *engine)
@@ -133,16 +150,20 @@ intel_context_pin(struct i915_gem_context *ctx,
 	struct intel_context *ce;
 	int err;
 
-	lockdep_assert_held(&ctx->i915->drm.struct_mutex);
-
 	ce = intel_context_instance(ctx, engine);
 	if (IS_ERR(ce))
 		return ce;
 
-	if (unlikely(!ce->pin_count++)) {
+	if (likely(atomic_inc_not_zero(&ce->pin_count)))
+		return ce;
+
+	if (mutex_lock_interruptible(&ce->pin_mutex))
+		return ERR_PTR(-EINTR);
+
+	if (likely(!atomic_read(&ce->pin_count))) {
 		err = ce->ops->pin(ce);
 		if (err)
-			goto err_unpin;
+			goto err;
 
 		mutex_lock(&ctx->mutex);
 		list_add(&ce->active_link, &ctx->active_engines);
@@ -150,16 +171,35 @@ intel_context_pin(struct i915_gem_context *ctx,
 
 		i915_gem_context_get(ctx);
 		GEM_BUG_ON(ce->gem_context != ctx);
+
+		smp_mb__before_atomic(); /* flush pin before it is visible */
 	}
-	GEM_BUG_ON(!ce->pin_count); /* no overflow! */
 
+	atomic_inc(&ce->pin_count);
+	GEM_BUG_ON(!intel_context_is_pinned(ce)); /* no overflow! */
+
+	mutex_unlock(&ce->pin_mutex);
 	return ce;
 
-err_unpin:
-	ce->pin_count = 0;
+err:
+	mutex_unlock(&ce->pin_mutex);
 	return ERR_PTR(err);
 }
 
+void intel_context_unpin(struct intel_context *ce)
+{
+	if (likely(atomic_add_unless(&ce->pin_count, -1, 1)))
+		return;
+
+	/* We may be called from inside intel_context_pin() to evict another */
+	mutex_lock_nested(&ce->pin_mutex, SINGLE_DEPTH_NESTING);
+
+	if (likely(atomic_dec_and_test(&ce->pin_count)))
+		ce->ops->unpin(ce);
+
+	mutex_unlock(&ce->pin_mutex);
+}
+
 static void intel_context_retire(struct i915_active_request *active,
 				 struct i915_request *rq)
 {
@@ -181,6 +221,8 @@ intel_context_init(struct intel_context *ce,
 	INIT_LIST_HEAD(&ce->signal_link);
 	INIT_LIST_HEAD(&ce->signals);
 
+	mutex_init(&ce->pin_mutex);
+
 	/* Use the whole device by default */
 	ce->sseu = intel_device_default_sseu(ctx->i915);
 

drivers/gpu/drm/i915/intel_context.h

@@ -7,6 +7,8 @@
 #ifndef __INTEL_CONTEXT_H__
 #define __INTEL_CONTEXT_H__
 
+#include <linux/lockdep.h>
+
 #include "intel_context_types.h"
 #include "intel_engine_types.h"
 
@@ -29,18 +31,30 @@ intel_context_lookup(struct i915_gem_context *ctx,
 		     struct intel_engine_cs *engine);
 
 /**
- * intel_context_instance - Lookup or allocate the HW context for (ctx, engine)
+ * intel_context_pin_lock - Stablises the 'pinned' status of the HW context
  * @ctx - the parent GEM context
  * @engine - the target HW engine
  *
- * Returns the existing HW context for this pair of (GEM context, engine), or
- * allocates and initialises a fresh context. Once allocated, the HW context
- * remains resident until the GEM context is destroyed.
+ * Acquire a lock on the pinned status of the HW context, such that the context
+ * can neither be bound to the GPU or unbound whilst the lock is held, i.e.
+ * intel_context_is_pinned() remains stable.
  */
 struct intel_context *
-intel_context_instance(struct i915_gem_context *ctx,
+intel_context_pin_lock(struct i915_gem_context *ctx,
 		       struct intel_engine_cs *engine);
 
+static inline bool
+intel_context_is_pinned(struct intel_context *ce)
+{
+	return atomic_read(&ce->pin_count);
+}
+
+static inline void intel_context_pin_unlock(struct intel_context *ce)
+__releases(ce->pin_mutex)
+{
+	mutex_unlock(&ce->pin_mutex);
+}
+
 struct intel_context *
 __intel_context_insert(struct i915_gem_context *ctx,
 		       struct intel_engine_cs *engine,
@@ -53,18 +67,10 @@ intel_context_pin(struct i915_gem_context *ctx, struct intel_engine_cs *engine);
 
 static inline void __intel_context_pin(struct intel_context *ce)
 {
-	GEM_BUG_ON(!ce->pin_count);
-	ce->pin_count++;
+	GEM_BUG_ON(!intel_context_is_pinned(ce));
+	atomic_inc(&ce->pin_count);
 }
 
-static inline void intel_context_unpin(struct intel_context *ce)
-{
-	GEM_BUG_ON(!ce->pin_count);
-	if (--ce->pin_count)
-		return;
-
-	GEM_BUG_ON(!ce->ops);
-	ce->ops->unpin(ce);
-}
+void intel_context_unpin(struct intel_context *ce);
 
 #endif /* __INTEL_CONTEXT_H__ */

drivers/gpu/drm/i915/intel_context_types.h

@@ -8,6 +8,7 @@
 #define __INTEL_CONTEXT_TYPES__
 
 #include <linux/list.h>
+#include <linux/mutex.h>
 #include <linux/rbtree.h>
 #include <linux/types.h>
 
@@ -38,14 +39,19 @@ struct intel_context {
 	struct i915_gem_context *gem_context;
 	struct intel_engine_cs *engine;
 	struct intel_engine_cs *active;
+
 	struct list_head active_link;
 	struct list_head signal_link;
 	struct list_head signals;
+
 	struct i915_vma *state;
 	struct intel_ring *ring;
+
 	u32 *lrc_reg_state;
 	u64 lrc_desc;
-	int pin_count;
+
+	atomic_t pin_count;
+	struct mutex pin_mutex; /* guards pinning and associated on-gpuing */
 
 	/**
 	 * active_tracker: Active tracker for the external rq activity

drivers/gpu/drm/i915/intel_lrc.c

@@ -1244,7 +1244,7 @@ static void __execlists_context_fini(struct intel_context *ce)
 
 static void execlists_context_destroy(struct intel_context *ce)
 {
-	GEM_BUG_ON(ce->pin_count);
+	GEM_BUG_ON(intel_context_is_pinned(ce));
 
 	if (ce->state)
 		__execlists_context_fini(ce);
@@ -1481,7 +1481,7 @@ static int execlists_request_alloc(struct i915_request *request)
 {
 	int ret;
 
-	GEM_BUG_ON(!request->hw_context->pin_count);
+	GEM_BUG_ON(!intel_context_is_pinned(request->hw_context));
 
 	/*
 	 * Flush enough space to reduce the likelihood of waiting after

drivers/gpu/drm/i915/intel_ringbuffer.c

@@ -1357,7 +1357,7 @@ static void __ring_context_fini(struct intel_context *ce)
 
 static void ring_context_destroy(struct intel_context *ce)
 {
-	GEM_BUG_ON(ce->pin_count);
+	GEM_BUG_ON(intel_context_is_pinned(ce));
 
 	if (ce->state)
 		__ring_context_fini(ce);
@@ -1918,7 +1918,7 @@ static int ring_request_alloc(struct i915_request *request)
 {
 	int ret;
 
-	GEM_BUG_ON(!request->hw_context->pin_count);
+	GEM_BUG_ON(!intel_context_is_pinned(request->hw_context));
 	GEM_BUG_ON(request->timeline->has_initial_breadcrumb);
 
 	/*

drivers/gpu/drm/i915/selftests/mock_engine.c

@@ -131,7 +131,7 @@ static void mock_context_unpin(struct intel_context *ce)
 
 static void mock_context_destroy(struct intel_context *ce)
 {
-	GEM_BUG_ON(ce->pin_count);
+	GEM_BUG_ON(intel_context_is_pinned(ce));
 
 	if (ce->ring)
 		mock_ring_free(ce->ring);