[PATCH 43/46] drm/i915: Load balancing across a virtual engine

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Having allowed the user to define a set of engines that they will want
to only use, we go one step further and allow them to bind those engines
into a single virtual instance. Submitting a batch to the virtual engine
will then forward it to any one of the set in a manner as best to
distribute load.  The virtual engine has a single timeline across all
engines (it operates as a single queue), so it is not able to concurrently
run batches across multiple engines by itself; that is left up to the user
to submit multiple concurrent batches to multiple queues. Multiple users
will be load balanced across the system.

The mechanism used for load balancing in this patch is a late greedy
balancer. When a request is ready for execution, it is added to each
engine's queue, and when an engine is ready for its next request it
claims it from the virtual engine. The first engine to do so, wins, i.e.
the request is executed at the earliest opportunity (idle moment) in the
system.

As not all HW is created equal, the user is still able to skip the
virtual engine and execute the batch on a specific engine, all within the
same queue. It will then be executed in order on the correct engine,
with execution on other virtual engines being moved away due to the load
detection.

A couple of areas for potential improvement left!

- The virtual engine always take priority over equal-priority tasks.
Mostly broken up by applying FQ_CODEL rules for prioritising new clients,
and hopefully the virtual and real engines are not then congested (i.e.
all work is via virtual engines, or all work is to the real engine).

- We require the breadcrumb irq around every virtual engine request. For
normal engines, we eliminate the need for the slow round trip via
interrupt by using the submit fence and queueing in order. For virtual
engines, we have to allow any job to transfer to a new ring, and cannot
coalesce the submissions, so require the completion fence instead,
forcing the persistent use of interrupts.

- We only drip feed single requests through each virtual engine and onto
the physical engines, even if there was enough work to fill all ELSP,
leaving small stalls with an idle CS event at the end of every request.
Could we be greedy and fill both slots? Being lazy is virtuous for load
distribution on less-than-full workloads though.

Other areas of improvement are more general, such as reducing lock
contention, reducing dispatch overhead, looking at direct submission
rather than bouncing around tasklets etc.

Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
---
 drivers/gpu/drm/i915/i915_gem.h            |   5 +
 drivers/gpu/drm/i915/i915_gem_context.c    |  90 +++-
 drivers/gpu/drm/i915/i915_scheduler.c      |  20 +-
 drivers/gpu/drm/i915/i915_timeline_types.h |   1 +
 drivers/gpu/drm/i915/intel_engine_types.h  |   8 +
 drivers/gpu/drm/i915/intel_lrc.c           | 502 ++++++++++++++++++++-
 drivers/gpu/drm/i915/intel_lrc.h           |   6 +
 drivers/gpu/drm/i915/selftests/intel_lrc.c | 165 +++++++
 include/uapi/drm/i915_drm.h                |  27 ++
 9 files changed, 809 insertions(+), 15 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem.h b/drivers/gpu/drm/i915/i915_gem.h
index b0e4b976880c..9905fcdd33c8 100644
--- a/drivers/gpu/drm/i915/i915_gem.h
+++ b/drivers/gpu/drm/i915/i915_gem.h
@@ -89,4 +89,9 @@ static inline bool __tasklet_is_enabled(const struct tasklet_struct *t)
 	return !atomic_read(&t->count);
 }
 
+static inline bool __tasklet_is_scheduled(struct tasklet_struct *t)
+{
+	return test_bit(TASKLET_STATE_SCHED, &t->state);
+}
+
 #endif /* __I915_GEM_H__ */
diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c
index e75fad339ab8..2b474e72f9e2 100644
--- a/drivers/gpu/drm/i915/i915_gem_context.c
+++ b/drivers/gpu/drm/i915/i915_gem_context.c
@@ -91,6 +91,7 @@
 #include "i915_trace.h"
 #include "i915_user_extensions.h"
 #include "intel_lrc_reg.h"
+#include "intel_lrc.h"
 #include "intel_workarounds.h"
 
 #define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1
@@ -233,7 +234,10 @@ static void i915_gem_context_free(struct i915_gem_context *ctx)
 	release_hw_id(ctx);
 	i915_ppgtt_put(ctx->ppgtt);
 
-	kfree(ctx->engines);
+	if (ctx->engines) {
+		intel_virtual_engine_destroy(ctx->engines[0]);
+		kfree(ctx->engines);
+	}
 
 	rbtree_postorder_for_each_entry_safe(it, n, &ctx->hw_contexts, node) {
 		if (it->ops && it->ops->destroy)
@@ -1338,13 +1342,91 @@ static int set_sseu(struct i915_gem_context *ctx,
 	return 0;
 };
 
+static int check_user_mbz64(u64 __user *user)
+{
+	u64 mbz;
+
+	if (get_user(mbz, user))
+		return -EFAULT;
+
+	return mbz ? -EINVAL : 0;
+}
+
 struct set_engines {
 	struct i915_gem_context *ctx;
 	struct intel_engine_cs **engines;
 	unsigned int nengine;
 };
 
+static int
+set_engines__load_balance(struct i915_user_extension __user *base, void *data)
+{
+	struct i915_context_engines_load_balance __user *ext =
+		container_of_user(base, typeof(*ext), base);
+	const struct set_engines *set = data;
+	struct intel_engine_cs *ve;
+	unsigned int n;
+	u64 all_engines;
+	u64 mask;
+	int err;
+
+	/* We always use I915_EXEC_DEFAULT[0] for the load-balancer */
+	if (set->engines[0])
+		return -EEXIST;
+
+	if (!HAS_EXECLISTS(set->ctx->i915))
+		return -ENODEV;
+
+	if (USES_GUC_SUBMISSION(set->ctx->i915))
+		return -ENODEV; /* not implement yet */
+
+	err = check_user_mbz64(&ext->flags);
+	if (err)
+		return err;
+
+	for (n = 0; n < ARRAY_SIZE(ext->mbz); n++) {
+		err = check_user_mbz64(&ext->mbz[n]);
+		if (err)
+			return err;
+	}
+
+	if (get_user(mask, &ext->engines_mask))
+		return -EFAULT;
+
+	all_engines = GENMASK_ULL(set->nengine - 1, 1);
+	mask &= all_engines;
+
+	if (!mask) {
+		return -EINVAL;
+	} else if (is_power_of_2(mask)) {
+		ve = set->engines[__ffs64(mask)];
+	} else if (mask == all_engines) {
+		ve = intel_execlists_create_virtual(set->ctx,
+						    set->engines + 1,
+						    set->nengine - 1);
+	} else {
+		struct intel_engine_cs *stack[64];
+		int bit;
+
+		n = 0;
+		for_each_set_bit(bit, (unsigned long *)&mask, set->nengine)
+			stack[n++] = set->engines[bit];
+
+		ve = intel_execlists_create_virtual(set->ctx, stack, n);
+	}
+	if (IS_ERR(ve))
+		return PTR_ERR(ve);
+
+	if (cmpxchg(&set->engines[0], NULL, ve)) {
+		intel_virtual_engine_destroy(ve);
+		return -EEXIST;
+	}
+
+	return 0;
+}
+
 static const i915_user_extension_fn set_engines__extensions[] = {
+	[I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_engines__load_balance,
 };
 
 static int
@@ -1405,13 +1487,17 @@ set_engines(struct i915_gem_context *ctx,
 					   ARRAY_SIZE(set_engines__extensions),
 					   &set);
 	if (err) {
+		intel_virtual_engine_destroy(set.engines[0]);
 		kfree(set.engines);
 		return err;
 	}
 
 out:
 	mutex_lock(&ctx->i915->drm.struct_mutex);
-	kfree(ctx->engines);
+	if (ctx->engines) {
+		intel_virtual_engine_destroy(ctx->engines[0]);
+		kfree(ctx->engines);
+	}
 	ctx->engines = set.engines;
 	ctx->nengine = set.nengine;
 	mutex_unlock(&ctx->i915->drm.struct_mutex);
diff --git a/drivers/gpu/drm/i915/i915_scheduler.c b/drivers/gpu/drm/i915/i915_scheduler.c
index d579e21c0cf4..f34680543b08 100644
--- a/drivers/gpu/drm/i915/i915_scheduler.c
+++ b/drivers/gpu/drm/i915/i915_scheduler.c
@@ -236,18 +236,27 @@ i915_sched_lookup_priolist(struct intel_engine_cs *engine, int prio)
 }
 
 static struct intel_engine_cs *
-sched_lock_engine(struct i915_sched_node *node, struct intel_engine_cs *locked)
+sched_lock_engine(const struct i915_sched_node *node,
+		  struct intel_engine_cs *locked)
 {
-	struct intel_engine_cs *engine = node_to_request(node)->engine;
+	const struct i915_request *rq = node_to_request(node);
+	struct intel_engine_cs *engine;
 
 	GEM_BUG_ON(!locked);
 
-	if (engine != locked) {
+	/*
+	 * Virtual engines complicate acquiring the engine timeline lock,
+	 * as their rq->engine pointer is not stable until under that
+	 * engine lock. The simple ploy we use is to take the lock then
+	 * check that the rq still belongs to the newly locked engine.
+	 */
+	while (locked != (engine = READ_ONCE(rq->engine))) {
 		spin_unlock(&locked->timeline.lock);
 		spin_lock(&engine->timeline.lock);
+		locked = engine;
 	}
 
-	return engine;
+	return locked;
 }
 
 static bool inflight(const struct i915_request *rq,
@@ -357,8 +366,11 @@ static void __i915_schedule(struct i915_request *rq,
 		if (prio <= node->attr.priority || node_signaled(node))
 			continue;
 
+		GEM_BUG_ON(node_to_request(node)->engine != engine);
+
 		node->attr.priority = prio;
 		if (!list_empty(&node->link)) {
+			GEM_BUG_ON(intel_engine_is_virtual(engine));
 			if (last != engine) {
 				pl = i915_sched_lookup_priolist(engine, prio);
 				last = engine;
diff --git a/drivers/gpu/drm/i915/i915_timeline_types.h b/drivers/gpu/drm/i915/i915_timeline_types.h
index 398905e8888f..b02d4702dd5b 100644
--- a/drivers/gpu/drm/i915/i915_timeline_types.h
+++ b/drivers/gpu/drm/i915/i915_timeline_types.h
@@ -25,6 +25,7 @@ struct i915_timeline {
 	spinlock_t lock;
 #define TIMELINE_CLIENT 0 /* default subclass */
 #define TIMELINE_ENGINE 1
+#define TIMELINE_VIRTUAL 2
 
 	unsigned int pin_count;
 	const u32 *hwsp_seqno;
diff --git a/drivers/gpu/drm/i915/intel_engine_types.h b/drivers/gpu/drm/i915/intel_engine_types.h
index 8f0aedb2c2d8..1cf2740ca60e 100644
--- a/drivers/gpu/drm/i915/intel_engine_types.h
+++ b/drivers/gpu/drm/i915/intel_engine_types.h
@@ -215,6 +215,7 @@ struct intel_engine_execlists {
 	 * @queue: queue of requests, in priority lists
 	 */
 	struct rb_root_cached queue;
+	struct rb_root_cached virtual;
 
 	/**
 	 * @csb_write: control register for Context Switch buffer
@@ -423,6 +424,7 @@ struct intel_engine_cs {
 #define I915_ENGINE_SUPPORTS_STATS   BIT(1)
 #define I915_ENGINE_HAS_PREEMPTION   BIT(2)
 #define I915_ENGINE_HAS_SEMAPHORES   BIT(3)
+#define I915_ENGINE_IS_VIRTUAL       BIT(4)
 	unsigned int flags;
 
 	/*
@@ -506,6 +508,12 @@ intel_engine_has_semaphores(const struct intel_engine_cs *engine)
 	return engine->flags & I915_ENGINE_HAS_SEMAPHORES;
 }
 
+static inline bool
+intel_engine_is_virtual(const struct intel_engine_cs *engine)
+{
+	return engine->flags & I915_ENGINE_IS_VIRTUAL;
+}
+
 #define instdone_slice_mask(dev_priv__) \
 	(IS_GEN(dev_priv__, 7) ? \
 	 1 : RUNTIME_INFO(dev_priv__)->sseu.slice_mask)
diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
index 48797a1c0964..0b25fffdb8bc 100644
--- a/drivers/gpu/drm/i915/intel_lrc.c
+++ b/drivers/gpu/drm/i915/intel_lrc.c
@@ -166,6 +166,27 @@
 
 #define ACTIVE_PRIORITY (I915_PRIORITY_NEWCLIENT | I915_PRIORITY_NOSEMAPHORE)
 
+struct virtual_engine {
+	struct intel_engine_cs base;
+
+	struct intel_context context;
+	struct kref kref;
+
+	struct i915_request *request;
+	struct ve_node {
+		struct rb_node rb;
+		int prio;
+	} nodes[I915_NUM_ENGINES];
+
+	unsigned int count;
+	struct intel_engine_cs *siblings[0];
+};
+
+static struct virtual_engine *to_virtual_engine(struct intel_engine_cs *engine)
+{
+	return container_of(engine, struct virtual_engine, base);
+}
+
 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
 					    struct intel_engine_cs *engine,
 					    struct intel_context *ce);
@@ -236,7 +257,8 @@ static int queue_prio(const struct intel_engine_execlists *execlists)
 }
 
 static inline bool need_preempt(const struct intel_engine_cs *engine,
-				const struct i915_request *rq)
+				const struct i915_request *rq,
+				struct rb_node *rb)
 {
 	int last_prio;
 
@@ -271,6 +293,22 @@ static inline bool need_preempt(const struct intel_engine_cs *engine,
 	    rq_prio(list_next_entry(rq, link)) > last_prio)
 		return true;
 
+	if (rb) { /* XXX virtual precedence */
+		struct virtual_engine *ve =
+			rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
+		bool preempt = false;
+
+		if (engine == ve->siblings[0]) { /* only preempt one sibling */
+			spin_lock(&ve->base.timeline.lock);
+			if (ve->request)
+				preempt = rq_prio(ve->request) > last_prio;
+			spin_unlock(&ve->base.timeline.lock);
+		}
+
+		if (preempt)
+			return preempt;
+	}
+
 	/*
 	 * If the inflight context did not trigger the preemption, then maybe
 	 * it was the set of queued requests? Pick the highest priority in
@@ -390,6 +428,8 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
 	list_for_each_entry_safe_reverse(rq, rn,
 					 &engine->timeline.requests,
 					 link) {
+		struct intel_engine_cs *owner;
+
 		if (i915_request_completed(rq))
 			break;
 
@@ -398,14 +438,22 @@ __unwind_incomplete_requests(struct intel_engine_cs *engine)
 
 		GEM_BUG_ON(rq->hw_context->active);
 
-		GEM_BUG_ON(rq_prio(rq) == I915_PRIORITY_INVALID);
-		if (rq_prio(rq) != prio) {
-			prio = rq_prio(rq);
-			pl = i915_sched_lookup_priolist(engine, prio);
-		}
-		GEM_BUG_ON(RB_EMPTY_ROOT(&engine->execlists.queue.rb_root));
+		owner = rq->hw_context->engine;
+		if (likely(owner == engine)) {
+			GEM_BUG_ON(rq_prio(rq) == I915_PRIORITY_INVALID);
+			if (rq_prio(rq) != prio) {
+				prio = rq_prio(rq);
+				pl = i915_sched_lookup_priolist(engine, prio);
+			}
+			GEM_BUG_ON(RB_EMPTY_ROOT(&engine->execlists.queue.rb_root));
 
-		list_add(&rq->sched.link, pl);
+			list_add(&rq->sched.link, pl);
+		} else {
+			rq->engine = owner;
+			if (__i915_request_has_started(rq))
+				rq->sched.attr.priority |= ACTIVE_PRIORITY;
+			owner->submit_request(rq);
+		}
 
 		active = rq;
 	}
@@ -661,6 +709,50 @@ static void complete_preempt_context(struct intel_engine_execlists *execlists)
 						  execlists));
 }
 
+static void virtual_update_register_offsets(u32 *regs,
+					    struct intel_engine_cs *engine)
+{
+	u32 base = engine->mmio_base;
+
+	regs[CTX_CONTEXT_CONTROL] =
+		i915_mmio_reg_offset(RING_CONTEXT_CONTROL(engine));
+	regs[CTX_RING_HEAD] = i915_mmio_reg_offset(RING_HEAD(base));
+	regs[CTX_RING_TAIL] = i915_mmio_reg_offset(RING_TAIL(base));
+	regs[CTX_RING_BUFFER_START] = i915_mmio_reg_offset(RING_START(base));
+	regs[CTX_RING_BUFFER_CONTROL] = i915_mmio_reg_offset(RING_CTL(base));
+
+	regs[CTX_BB_HEAD_U] = i915_mmio_reg_offset(RING_BBADDR_UDW(base));
+	regs[CTX_BB_HEAD_L] = i915_mmio_reg_offset(RING_BBADDR(base));
+	regs[CTX_BB_STATE] = i915_mmio_reg_offset(RING_BBSTATE(base));
+	regs[CTX_SECOND_BB_HEAD_U] =
+		i915_mmio_reg_offset(RING_SBBADDR_UDW(base));
+	regs[CTX_SECOND_BB_HEAD_L] = i915_mmio_reg_offset(RING_SBBADDR(base));
+	regs[CTX_SECOND_BB_STATE] = i915_mmio_reg_offset(RING_SBBSTATE(base));
+
+	regs[CTX_CTX_TIMESTAMP] =
+		i915_mmio_reg_offset(RING_CTX_TIMESTAMP(base));
+	regs[CTX_PDP3_UDW] = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(engine, 3));
+	regs[CTX_PDP3_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(engine, 3));
+	regs[CTX_PDP2_UDW] = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(engine, 2));
+	regs[CTX_PDP2_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(engine, 2));
+	regs[CTX_PDP1_UDW] = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(engine, 1));
+	regs[CTX_PDP1_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(engine, 1));
+	regs[CTX_PDP0_UDW] = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(engine, 0));
+	regs[CTX_PDP0_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(engine, 0));
+
+	if (engine->class == RENDER_CLASS) {
+		regs[CTX_RCS_INDIRECT_CTX] =
+			i915_mmio_reg_offset(RING_INDIRECT_CTX(base));
+		regs[CTX_RCS_INDIRECT_CTX_OFFSET] =
+			i915_mmio_reg_offset(RING_INDIRECT_CTX_OFFSET(base));
+		regs[CTX_BB_PER_CTX_PTR] =
+			i915_mmio_reg_offset(RING_BB_PER_CTX_PTR(base));
+
+		regs[CTX_R_PWR_CLK_STATE] =
+			i915_mmio_reg_offset(GEN8_R_PWR_CLK_STATE);
+	}
+}
+
 static void execlists_dequeue(struct intel_engine_cs *engine)
 {
 	struct intel_engine_execlists * const execlists = &engine->execlists;
@@ -693,6 +785,28 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 	 * and context switches) submission.
 	 */
 
+	for (rb = rb_first_cached(&execlists->virtual); rb; ) {
+		struct virtual_engine *ve =
+			rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
+		struct i915_request *rq = READ_ONCE(ve->request);
+		struct intel_engine_cs *active;
+
+		if (!rq) {
+			rb_erase_cached(rb, &execlists->virtual);
+			RB_CLEAR_NODE(rb);
+			rb = rb_first_cached(&execlists->virtual);
+			continue;
+		}
+
+		active = READ_ONCE(ve->context.active);
+		if (active && active != engine) {
+			rb = rb_next(rb);
+			continue;
+		}
+
+		break;
+	}
+
 	if (last) {
 		/*
 		 * Don't resubmit or switch until all outstanding
@@ -714,7 +828,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 		if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_HWACK))
 			return;
 
-		if (need_preempt(engine, last)) {
+		if (need_preempt(engine, last, rb)) {
 			inject_preempt_context(engine);
 			return;
 		}
@@ -754,6 +868,72 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 		last->tail = last->wa_tail;
 	}
 
+	while (rb) { /* XXX virtual is always taking precedence */
+		struct virtual_engine *ve =
+			rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
+		struct i915_request *rq;
+
+		spin_lock(&ve->base.timeline.lock);
+
+		rq = ve->request;
+		if (unlikely(!rq)) { /* lost the race to a sibling */
+			spin_unlock(&ve->base.timeline.lock);
+			rb_erase_cached(rb, &execlists->virtual);
+			RB_CLEAR_NODE(rb);
+			rb = rb_first_cached(&execlists->virtual);
+			continue;
+		}
+
+		if (rq_prio(rq) >= queue_prio(execlists)) {
+			if (last && !can_merge_rq(last, rq)) {
+				spin_unlock(&ve->base.timeline.lock);
+				return; /* leave this rq for another engine */
+			}
+
+			GEM_BUG_ON(rq->engine != &ve->base);
+			ve->request = NULL;
+			ve->base.execlists.queue_priority_hint = INT_MIN;
+			rb_erase_cached(rb, &execlists->virtual);
+			RB_CLEAR_NODE(rb);
+
+			GEM_BUG_ON(rq->hw_context != &ve->context);
+			rq->engine = engine;
+
+			if (engine != ve->siblings[0]) {
+				u32 *regs = ve->context.lrc_reg_state;
+				unsigned int n;
+
+				GEM_BUG_ON(READ_ONCE(ve->context.active));
+				virtual_update_register_offsets(regs, engine);
+
+				/*
+				 * Move the bound engine to the top of the list
+				 * for future execution. We then kick this
+				 * tasklet first before checking others, so that
+				 * we preferentially reuse this set of bound
+				 * registers.
+				 */
+				for (n = 1; n < ve->count; n++) {
+					if (ve->siblings[n] == engine) {
+						swap(ve->siblings[n],
+						     ve->siblings[0]);
+						break;
+					}
+				}
+
+				GEM_BUG_ON(ve->siblings[0] != engine);
+			}
+
+			__i915_request_submit(rq);
+			trace_i915_request_in(rq, port_index(port, execlists));
+			submit = true;
+			last = rq;
+		}
+
+		spin_unlock(&ve->base.timeline.lock);
+		break;
+	}
+
 	while ((rb = rb_first_cached(&execlists->queue))) {
 		struct i915_priolist *p = to_priolist(rb);
 		struct i915_request *rq, *rn;
@@ -2919,6 +3099,287 @@ void intel_lr_context_resume(struct drm_i915_private *i915)
 	}
 }
 
+static void virtual_engine_free(struct kref *kref)
+{
+	struct virtual_engine *ve = container_of(kref, typeof(*ve), kref);
+	unsigned int n;
+
+	GEM_BUG_ON(ve->request);
+	GEM_BUG_ON(ve->context.active);
+
+	for (n = 0; n < ve->count; n++) {
+		struct intel_engine_cs *sibling = ve->siblings[n];
+		struct rb_node *node = &ve->nodes[sibling->id].rb;
+
+		if (RB_EMPTY_NODE(node))
+			continue;
+
+		spin_lock_irq(&sibling->timeline.lock);
+
+		if (!RB_EMPTY_NODE(node))
+			rb_erase_cached(node, &sibling->execlists.virtual);
+
+		spin_unlock_irq(&sibling->timeline.lock);
+	}
+	GEM_BUG_ON(__tasklet_is_scheduled(&ve->base.execlists.tasklet));
+
+	if (ve->context.state)
+		__execlists_context_fini(&ve->context);
+
+	i915_timeline_fini(&ve->base.timeline);
+	kfree(ve);
+}
+
+static void virtual_context_unpin(struct intel_context *ce)
+{
+	struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
+
+	execlists_context_unpin(ce);
+
+	kref_put(&ve->kref, virtual_engine_free);
+}
+
+static const struct intel_context_ops virtual_context_ops = {
+	.unpin = virtual_context_unpin,
+};
+
+static void virtual_initial_engine_hint(struct virtual_engine *ve)
+{
+	int swp;
+
+	/*
+	 * Pick a random sibling on starting to help spread the load around.
+	 *
+	 * New contexts are typically created with exactly the same order
+	 * of siblings, and often started in batches. Due to the way we iterate
+	 * the array of sibling when submitting requests, sibling[0] is
+	 * prioritised for dequeuing. If we make sure that sibling[0] is fairly
+	 * randomised across the system, we also help spread the load by the
+	 * first engine we inspect being different each time.
+	 *
+	 * NB This does not force us to execute on this engine, it will just
+	 * typically be the first we inspect for submission.
+	 */
+	swp = prandom_u32_max(ve->count);
+	if (!swp)
+		return;
+
+	swap(ve->siblings[swp], ve->siblings[0]);
+	virtual_update_register_offsets(ve->context.lrc_reg_state,
+					ve->siblings[0]);
+}
+
+static struct intel_context *
+virtual_context_pin(struct intel_engine_cs *engine,
+		    struct i915_gem_context *ctx)
+{
+	struct virtual_engine *ve = to_virtual_engine(engine);
+	struct intel_context *ce = &ve->context;
+
+	lockdep_assert_held(&ctx->i915->drm.struct_mutex);
+
+	if (likely(ce->pin_count++))
+		return ce;
+	GEM_BUG_ON(!ce->pin_count); /* no overflow please! */
+
+	kref_get(&ve->kref);
+	ce->ops = &virtual_context_ops;
+
+	/* Note: we must use a real engine class for setting up reg state */
+	ce = __execlists_context_pin(ve->siblings[0], ctx, ce);
+	if (!IS_ERR(ce))
+		virtual_initial_engine_hint(ve);
+
+	return ce;
+}
+
+static void virtual_submission_tasklet(unsigned long data)
+{
+	struct virtual_engine * const ve = (struct virtual_engine *)data;
+	unsigned int n;
+	int prio;
+
+	prio = READ_ONCE(ve->base.execlists.queue_priority_hint);
+	if (prio == INT_MIN)
+		return;
+
+	local_irq_disable();
+	for (n = 0; READ_ONCE(ve->request) && n < ve->count; n++) {
+		struct intel_engine_cs *sibling = ve->siblings[n];
+		struct ve_node * const node = &ve->nodes[sibling->id];
+		struct rb_node **parent, *rb;
+		bool first;
+
+		spin_lock(&sibling->timeline.lock);
+
+		if (!RB_EMPTY_NODE(&node->rb)) {
+			first = rb_first_cached(&sibling->execlists.virtual) == &node->rb;
+			if (prio == node->prio || (prio > node->prio && first))
+				goto submit_engine;
+
+			rb_erase_cached(&node->rb, &sibling->execlists.virtual);
+		}
+
+		rb = NULL;
+		first = true;
+		parent = &sibling->execlists.virtual.rb_root.rb_node;
+		while (*parent) {
+			struct ve_node *other;
+
+			rb = *parent;
+			other = rb_entry(rb, typeof(*other), rb);
+			if (prio > other->prio) {
+				parent = &rb->rb_left;
+			} else {
+				parent = &rb->rb_right;
+				first = false;
+			}
+		}
+
+		rb_link_node(&node->rb, rb, parent);
+		rb_insert_color_cached(&node->rb,
+				       &sibling->execlists.virtual,
+				       first);
+
+submit_engine:
+		GEM_BUG_ON(RB_EMPTY_NODE(&node->rb));
+		node->prio = prio;
+		if (first && prio > sibling->execlists.queue_priority_hint) {
+			sibling->execlists.queue_priority_hint = prio;
+			tasklet_hi_schedule(&sibling->execlists.tasklet);
+		}
+
+		spin_unlock(&sibling->timeline.lock);
+	}
+	local_irq_enable();
+}
+
+static void virtual_submit_request(struct i915_request *request)
+{
+	struct virtual_engine *ve = to_virtual_engine(request->engine);
+
+	GEM_BUG_ON(ve->base.submit_request != virtual_submit_request);
+
+	GEM_BUG_ON(ve->request);
+	ve->base.execlists.queue_priority_hint = rq_prio(request);
+	WRITE_ONCE(ve->request, request);
+
+	tasklet_schedule(&ve->base.execlists.tasklet);
+}
+
+struct intel_engine_cs *
+intel_execlists_create_virtual(struct i915_gem_context *ctx,
+			       struct intel_engine_cs **siblings,
+			       unsigned int count)
+{
+	struct virtual_engine *ve;
+	unsigned int n;
+	int err;
+
+	if (!count)
+		return ERR_PTR(-EINVAL);
+
+	ve = kzalloc(struct_size(ve, siblings, count), GFP_KERNEL);
+	if (!ve)
+		return ERR_PTR(-ENOMEM);
+
+	kref_init(&ve->kref);
+	ve->base.i915 = ctx->i915;
+	ve->base.id = -1;
+	ve->base.class = OTHER_CLASS;
+	ve->base.uabi_class = I915_ENGINE_CLASS_INVALID;
+	ve->base.instance = I915_ENGINE_CLASS_INVALID_VIRTUAL;
+	ve->base.flags = I915_ENGINE_IS_VIRTUAL;
+
+	snprintf(ve->base.name, sizeof(ve->base.name), "virtual");
+
+	err = i915_timeline_init(ctx->i915,
+				 &ve->base.timeline,
+				 ve->base.name,
+				 NULL);
+	if (err)
+		goto err_put;
+	i915_timeline_set_subclass(&ve->base.timeline, TIMELINE_VIRTUAL);
+
+	intel_context_init(&ve->context, ctx, &ve->base);
+
+	ve->base.context_pin = virtual_context_pin;
+	ve->base.request_alloc = execlists_request_alloc;
+
+	ve->base.schedule = i915_schedule;
+	ve->base.submit_request = virtual_submit_request;
+
+	ve->base.execlists.queue_priority_hint = INT_MIN;
+	tasklet_init(&ve->base.execlists.tasklet,
+		     virtual_submission_tasklet,
+		     (unsigned long)ve);
+
+	for (n = 0; n < count; n++) {
+		struct intel_engine_cs *sibling = siblings[n];
+
+		GEM_BUG_ON(!is_power_of_2(sibling->mask));
+		if (sibling->mask & ve->base.mask)
+			continue;
+
+		if (sibling->execlists.tasklet.func != execlists_submission_tasklet) {
+			err = -ENODEV;
+			goto err_put;
+		}
+
+		GEM_BUG_ON(RB_EMPTY_NODE(&ve->nodes[sibling->id].rb));
+		RB_CLEAR_NODE(&ve->nodes[sibling->id].rb);
+
+		ve->siblings[ve->count++] = sibling;
+		ve->base.mask |= sibling->mask;
+
+		if (ve->base.class != OTHER_CLASS) {
+			if (ve->base.class != sibling->class) {
+				err = -EINVAL;
+				goto err_put;
+			}
+			continue;
+		}
+
+		ve->base.class = sibling->class;
+		snprintf(ve->base.name, sizeof(ve->base.name),
+			 "v%dx%d", ve->base.class, count);
+		ve->base.context_size = sibling->context_size;
+
+		ve->base.emit_bb_start = sibling->emit_bb_start;
+		ve->base.emit_flush = sibling->emit_flush;
+		ve->base.emit_init_breadcrumb = sibling->emit_init_breadcrumb;
+		ve->base.emit_fini_breadcrumb = sibling->emit_fini_breadcrumb;
+		ve->base.emit_fini_breadcrumb_dw =
+			sibling->emit_fini_breadcrumb_dw;
+	}
+
+	/* gracefully replace a degenerate virtual engine */
+	if (is_power_of_2(ve->base.mask)) {
+		struct intel_engine_cs *actual = ve->siblings[0];
+		virtual_engine_free(&ve->kref);
+		return actual;
+	}
+
+	__intel_context_insert(ctx, &ve->base, &ve->context);
+	return &ve->base;
+
+err_put:
+	virtual_engine_free(&ve->kref);
+	return ERR_PTR(err);
+}
+
+void intel_virtual_engine_destroy(struct intel_engine_cs *engine)
+{
+	struct virtual_engine *ve = to_virtual_engine(engine);
+
+	if (!engine || !intel_engine_is_virtual(engine))
+		return;
+
+	__intel_context_remove(&ve->context);
+
+	kref_put(&ve->kref, virtual_engine_free);
+}
+
 void intel_execlists_show_requests(struct intel_engine_cs *engine,
 				   struct drm_printer *m,
 				   void (*show_request)(struct drm_printer *m,
@@ -2976,6 +3437,29 @@ void intel_execlists_show_requests(struct intel_engine_cs *engine,
 		show_request(m, last, "\t\tQ ");
 	}
 
+	last = NULL;
+	count = 0;
+	for (rb = rb_first_cached(&execlists->virtual); rb; rb = rb_next(rb)) {
+		struct virtual_engine *ve =
+			rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
+		struct i915_request *rq = READ_ONCE(ve->request);
+
+		if (rq) {
+			if (count++ < max - 1)
+				show_request(m, rq, "\t\tV ");
+			else
+				last = rq;
+		}
+	}
+	if (last) {
+		if (count > max) {
+			drm_printf(m,
+				   "\t\t...skipping %d virtual requests...\n",
+				   count - max);
+		}
+		show_request(m, last, "\t\tV ");
+	}
+
 	spin_unlock_irqrestore(&engine->timeline.lock, flags);
 }
 
diff --git a/drivers/gpu/drm/i915/intel_lrc.h b/drivers/gpu/drm/i915/intel_lrc.h
index f1aec8a6986f..c6f441137e3f 100644
--- a/drivers/gpu/drm/i915/intel_lrc.h
+++ b/drivers/gpu/drm/i915/intel_lrc.h
@@ -112,6 +112,12 @@ void intel_execlists_show_requests(struct intel_engine_cs *engine,
 							const char *prefix),
 				   unsigned int max);
 
+struct intel_engine_cs *
+intel_execlists_create_virtual(struct i915_gem_context *ctx,
+			       struct intel_engine_cs **siblings,
+			       unsigned int count);
+void intel_virtual_engine_destroy(struct intel_engine_cs *engine);
+
 u32 gen8_make_rpcs(struct drm_i915_private *i915, struct intel_sseu *ctx_sseu);
 
 #endif /* _INTEL_LRC_H_ */
diff --git a/drivers/gpu/drm/i915/selftests/intel_lrc.c b/drivers/gpu/drm/i915/selftests/intel_lrc.c
index 1a3af4b4107d..e85184e084e4 100644
--- a/drivers/gpu/drm/i915/selftests/intel_lrc.c
+++ b/drivers/gpu/drm/i915/selftests/intel_lrc.c
@@ -10,6 +10,7 @@
 
 #include "../i915_selftest.h"
 #include "igt_flush_test.h"
+#include "igt_live_test.h"
 #include "igt_spinner.h"
 #include "i915_random.h"
 
@@ -1040,6 +1041,169 @@ static int live_preempt_smoke(void *arg)
 	return err;
 }
 
+static int nop_virtual_engine(struct drm_i915_private *i915,
+			      struct intel_engine_cs **siblings,
+			      unsigned int nsibling,
+			      unsigned int nctx,
+			      unsigned int flags)
+#define CHAIN BIT(0)
+{
+	IGT_TIMEOUT(end_time);
+	struct i915_request *request[16];
+	struct i915_gem_context *ctx[16];
+	struct intel_engine_cs *ve[16];
+	unsigned long n, prime, nc;
+	struct igt_live_test t;
+	ktime_t times[2] = {};
+	int err;
+
+	GEM_BUG_ON(!nctx || nctx > ARRAY_SIZE(ctx));
+
+	for (n = 0; n < nctx; n++) {
+		ctx[n] = kernel_context(i915);
+		if (!ctx[n])
+			return -ENOMEM;
+
+		ve[n] = intel_execlists_create_virtual(ctx[n],
+						       siblings, nsibling);
+		if (IS_ERR(ve[n]))
+			return PTR_ERR(ve[n]);
+	}
+
+	err = igt_live_test_begin(&t, i915, __func__, ve[0]->name);
+	if (err)
+		goto out;
+
+	for_each_prime_number_from(prime, 1, 8192) {
+		times[1] = ktime_get_raw();
+
+		if (flags & CHAIN) {
+			for (nc = 0; nc < nctx; nc++) {
+				for (n = 0; n < prime; n++) {
+					request[nc] =
+						i915_request_alloc(ve[nc], ctx[nc]);
+					if (IS_ERR(request[nc])) {
+						err = PTR_ERR(request[nc]);
+						goto out;
+					}
+
+					i915_request_add(request[nc]);
+				}
+			}
+		} else {
+			for (n = 0; n < prime; n++) {
+				for (nc = 0; nc < nctx; nc++) {
+					request[nc] =
+						i915_request_alloc(ve[nc], ctx[nc]);
+					if (IS_ERR(request[nc])) {
+						err = PTR_ERR(request[nc]);
+						goto out;
+					}
+
+					i915_request_add(request[nc]);
+				}
+			}
+		}
+
+		for (nc = 0; nc < nctx; nc++) {
+			if (i915_request_wait(request[nc],
+					      I915_WAIT_LOCKED,
+					      HZ / 10) < 0) {
+				pr_err("%s(%s): wait for %llx:%lld timed out\n",
+				       __func__, ve[0]->name,
+				       request[nc]->fence.context,
+				       request[nc]->fence.seqno);
+
+				GEM_TRACE("%s(%s) failed at request %llx:%lld\n",
+					  __func__, ve[0]->name,
+					  request[nc]->fence.context,
+					  request[nc]->fence.seqno);
+				GEM_TRACE_DUMP();
+				i915_gem_set_wedged(i915);
+				break;
+			}
+		}
+
+		times[1] = ktime_sub(ktime_get_raw(), times[1]);
+		if (prime == 1)
+			times[0] = times[1];
+
+		if (__igt_timeout(end_time, NULL))
+			break;
+	}
+
+	err = igt_live_test_end(&t);
+	if (err)
+		goto out;
+
+	pr_info("Requestx%d latencies on %s: 1 = %lluns, %lu = %lluns\n",
+		nctx, ve[0]->name, ktime_to_ns(times[0]),
+		prime, div64_u64(ktime_to_ns(times[1]), prime));
+
+out:
+	if (igt_flush_test(i915, I915_WAIT_LOCKED))
+		err = -EIO;
+
+	for (nc = 0; nc < nctx; nc++) {
+		intel_virtual_engine_destroy(ve[nc]);
+		kernel_context_close(ctx[nc]);
+	}
+	return err;
+}
+
+static int live_virtual_engine(void *arg)
+{
+	struct drm_i915_private *i915 = arg;
+	struct intel_engine_cs *siblings[MAX_ENGINE_INSTANCE + 1];
+	struct intel_engine_cs *engine;
+	enum intel_engine_id id;
+	unsigned int class, inst;
+	int err = -ENODEV;
+
+	if (USES_GUC_SUBMISSION(i915))
+		return 0;
+
+	mutex_lock(&i915->drm.struct_mutex);
+
+	for_each_engine(engine, i915, id) {
+		err = nop_virtual_engine(i915, &engine, 1, 1, 0);
+		if (err) {
+			pr_err("Failed to wrap engine %s: err=%d\n",
+			       engine->name, err);
+			goto out_unlock;
+		}
+	}
+
+	for (class = 0; class <= MAX_ENGINE_CLASS; class++) {
+		int nsibling, n;
+
+		nsibling = 0;
+		for (inst = 0; inst <= MAX_ENGINE_INSTANCE; inst++) {
+			if (!i915->engine_class[class][inst])
+				break;
+
+			siblings[nsibling++] = i915->engine_class[class][inst];
+		}
+		if (nsibling < 2)
+			continue;
+
+		for (n = 1; n <= nsibling + 1; n++) {
+			err = nop_virtual_engine(i915, siblings, nsibling,
+						 n, 0);
+			if (err)
+				goto out_unlock;
+		}
+
+		err = nop_virtual_engine(i915, siblings, nsibling, n, CHAIN);
+		if (err)
+			goto out_unlock;
+	}
+
+out_unlock:
+	mutex_unlock(&i915->drm.struct_mutex);
+	return err;
+}
+
 int intel_execlists_live_selftests(struct drm_i915_private *i915)
 {
 	static const struct i915_subtest tests[] = {
@@ -1051,6 +1215,7 @@ int intel_execlists_live_selftests(struct drm_i915_private *i915)
 		SUBTEST(live_chain_preempt),
 		SUBTEST(live_preempt_hang),
 		SUBTEST(live_preempt_smoke),
+		SUBTEST(live_virtual_engine),
 	};
 
 	if (!HAS_EXECLISTS(i915))
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index 642af28ea6d3..47d2f56b6d90 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -123,6 +123,7 @@ enum drm_i915_gem_engine_class {
 };
 
 #define I915_ENGINE_CLASS_INVALID_NONE -1
+#define I915_ENGINE_CLASS_INVALID_VIRTUAL 0
 
 /**
  * DOC: perf_events exposed by i915 through /sys/bus/event_sources/drivers/i915
@@ -1567,8 +1568,34 @@ struct drm_i915_gem_context_param_sseu {
 	__u32 rsvd;
 };
 
+/*
+ * i915_context_engines_load_balance:
+ *
+ * Enable load balancing across this set of engines.
+ *
+ * Into the I915_EXEC_DEFAULT slot [0], a virtual engine is created that when
+ * used will proxy the execbuffer request onto one of the set of engines
+ * in such a way as to distribute the load evenly across the set.
+ *
+ * The set of engines must be compatible (e.g. the same HW class) as they
+ * will share the same logical GPU context and ring.
+ *
+ * To intermix rendering with the virtual engine and direct rendering onto
+ * the backing engines (bypassing the load balancing proxy), the context must
+ * be defined to use a single timeline for all engines.
+ */
+struct i915_context_engines_load_balance {
+	struct i915_user_extension base;
+
+	__u64 flags; /* all undefined flags must be zero */
+	__u64 engines_mask; /* selection mask of engines[] */
+
+	__u64 mbz[4]; /* reserved for future use; must be zero */
+};
+
 struct i915_context_param_engines {
 	__u64 extensions; /* linked chain of extension blocks, 0 terminates */
+#define I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE 0
 
 	struct {
 		__u16 engine_class; /* see enum drm_i915_gem_engine_class */
-- 
2.20.1

_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx




[Index of Archives]     [AMD Graphics]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux