drm/i915: Load balancing across a virtual engine
authorChris Wilson <chris@chris-wilson.co.uk>
Tue, 21 May 2019 21:11:30 +0000 (22:11 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 22 May 2019 07:40:38 +0000 (08:40 +0100)
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.

sseu: Lift the restriction to allow sseu to be reconfigured on virtual
engines composed of RENDER_CLASS (rcs).

v2: macroize check_user_mbz()
v3: Cancel virtual engines on wedging
v4: Commence commenting
v5: Replace 64b sibling_mask with a list of class:instance
v6: Drop the one-element array in the uabi
v7: Assert it is an virtual engine in to_virtual_engine()
v8: Skip over holes in [class][inst] so we can selftest with (vcs0, vcs2)

Link: https://github.com/intel/media-driver/pull/283
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/20190521211134.16117-6-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gt/intel_engine_types.h
drivers/gpu/drm/i915/gt/intel_lrc.c
drivers/gpu/drm/i915/gt/intel_lrc.h
drivers/gpu/drm/i915/gt/selftest_lrc.c
drivers/gpu/drm/i915/i915_gem.h
drivers/gpu/drm/i915/i915_gem_context.c
drivers/gpu/drm/i915/i915_scheduler.c
drivers/gpu/drm/i915/i915_timeline_types.h
include/uapi/drm/i915_drm.h

index e381c1c7390288fb786a055c6d895f785142cc8d..7b47e00fa082904f0fbc81f92ca974d772df3852 100644 (file)
@@ -227,6 +227,7 @@ 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
@@ -445,6 +446,7 @@ struct intel_engine_cs {
 #define I915_ENGINE_HAS_PREEMPTION   BIT(2)
 #define I915_ENGINE_HAS_SEMAPHORES   BIT(3)
 #define I915_ENGINE_NEEDS_BREADCRUMB_TASKLET BIT(4)
+#define I915_ENGINE_IS_VIRTUAL       BIT(5)
        unsigned int flags;
 
        /*
@@ -534,6 +536,12 @@ intel_engine_needs_breadcrumb_tasklet(const struct intel_engine_cs *engine)
        return engine->flags & I915_ENGINE_NEEDS_BREADCRUMB_TASKLET;
 }
 
+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)
index f263a8374273421733027531468ef9649e920462..affa5e2dfce1d8597c33462d703134b627f0bd82 100644 (file)
 #define WA_TAIL_DWORDS 2
 #define WA_TAIL_BYTES (sizeof(u32) * WA_TAIL_DWORDS)
 
+struct virtual_engine {
+       struct intel_engine_cs base;
+       struct intel_context context;
+
+       /*
+        * We allow only a single request through the virtual engine at a time
+        * (each request in the timeline waits for the completion fence of
+        * the previous before being submitted). By restricting ourselves to
+        * only submitting a single request, each request is placed on to a
+        * physical to maximise load spreading (by virtue of the late greedy
+        * scheduling -- each real engine takes the next available request
+        * upon idling).
+        */
+       struct i915_request *request;
+
+       /*
+        * We keep a rbtree of available virtual engines inside each physical
+        * engine, sorted by priority. Here we preallocate the nodes we need
+        * for the virtual engine, indexed by physical_engine->id.
+        */
+       struct ve_node {
+               struct rb_node rb;
+               int prio;
+       } nodes[I915_NUM_ENGINES];
+
+       /* And finally, which physical engines this virtual engine maps onto. */
+       unsigned int num_siblings;
+       struct intel_engine_cs *siblings[0];
+};
+
+static struct virtual_engine *to_virtual_engine(struct intel_engine_cs *engine)
+{
+       GEM_BUG_ON(!intel_engine_is_virtual(engine));
+       return container_of(engine, struct virtual_engine, base);
+}
+
 static int execlists_context_deferred_alloc(struct intel_context *ce,
                                            struct intel_engine_cs *engine);
 static void execlists_init_reg_state(u32 *reg_state,
@@ -216,7 +252,8 @@ 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;
 
@@ -251,6 +288,25 @@ static inline bool need_preempt(const struct intel_engine_cs *engine,
            rq_prio(list_next_entry(rq, link)) > last_prio)
                return true;
 
+       if (rb) {
+               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 */
+                       struct i915_request *next;
+
+                       rcu_read_lock();
+                       next = READ_ONCE(ve->request);
+                       if (next)
+                               preempt = rq_prio(next) > last_prio;
+                       rcu_read_unlock();
+               }
+
+               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
@@ -369,6 +425,8 @@ __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;
 
@@ -377,16 +435,29 @@ __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));
-
-               list_add(&rq->sched.link, pl);
+               /*
+                * Push the request back into the queue for later resubmission.
+                * If this request is not native to this physical engine (i.e.
+                * it came from a virtual source), push it back onto the virtual
+                * engine so that it can be moved across onto another physical
+                * engine as load dictates.
+                */
+               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));
 
-               active = rq;
+                       list_add(&rq->sched.link, pl);
+                       active = rq;
+               } else {
+                       rq->engine = owner;
+                       owner->submit_request(rq);
+                       active = NULL;
+               }
        }
 
        return active;
@@ -622,6 +693,90 @@ 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;
+
+       /* Must match execlists_init_reg_state()! */
+
+       regs[CTX_CONTEXT_CONTROL] =
+               i915_mmio_reg_offset(RING_CONTEXT_CONTROL(base));
+       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(base, 3));
+       regs[CTX_PDP3_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, 3));
+       regs[CTX_PDP2_UDW] = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, 2));
+       regs[CTX_PDP2_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, 2));
+       regs[CTX_PDP1_UDW] = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, 1));
+       regs[CTX_PDP1_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, 1));
+       regs[CTX_PDP0_UDW] = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, 0));
+       regs[CTX_PDP0_LDW] = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, 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 bool virtual_matches(const struct virtual_engine *ve,
+                           const struct i915_request *rq,
+                           const struct intel_engine_cs *engine)
+{
+       const struct intel_engine_cs *active;
+
+       /*
+        * We track when the HW has completed saving the context image
+        * (i.e. when we have seen the final CS event switching out of
+        * the context) and must not overwrite the context image before
+        * then. This restricts us to only using the active engine
+        * while the previous virtualized request is inflight (so
+        * we reuse the register offsets). This is a very small
+        * hystersis on the greedy seelction algorithm.
+        */
+       active = READ_ONCE(ve->context.active);
+       if (active && active != engine)
+               return false;
+
+       return true;
+}
+
+static void virtual_xfer_breadcrumbs(struct virtual_engine *ve,
+                                    struct intel_engine_cs *engine)
+{
+       struct intel_engine_cs *old = ve->siblings[0];
+
+       /* All unattached (rq->engine == old) must already be completed */
+
+       spin_lock(&old->breadcrumbs.irq_lock);
+       if (!list_empty(&ve->context.signal_link)) {
+               list_move_tail(&ve->context.signal_link,
+                              &engine->breadcrumbs.signalers);
+               intel_engine_queue_breadcrumbs(engine);
+       }
+       spin_unlock(&old->breadcrumbs.irq_lock);
+}
+
 static void execlists_dequeue(struct intel_engine_cs *engine)
 {
        struct intel_engine_execlists * const execlists = &engine->execlists;
@@ -654,6 +809,26 @@ 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);
+
+               if (!rq) { /* lazily cleanup after another engine handled rq */
+                       rb_erase_cached(rb, &execlists->virtual);
+                       RB_CLEAR_NODE(rb);
+                       rb = rb_first_cached(&execlists->virtual);
+                       continue;
+               }
+
+               if (!virtual_matches(ve, rq, engine)) {
+                       rb = rb_next(rb);
+                       continue;
+               }
+
+               break;
+       }
+
        if (last) {
                /*
                 * Don't resubmit or switch until all outstanding
@@ -675,7 +850,7 @@ 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;
                }
@@ -715,6 +890,92 @@ 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;
+               }
+
+               GEM_BUG_ON(rq != ve->request);
+               GEM_BUG_ON(rq->engine != &ve->base);
+               GEM_BUG_ON(rq->hw_context != &ve->context);
+
+               if (rq_prio(rq) >= queue_prio(execlists)) {
+                       if (!virtual_matches(ve, rq, engine)) {
+                               spin_unlock(&ve->base.timeline.lock);
+                               rb = rb_next(rb);
+                               continue;
+                       }
+
+                       if (last && !can_merge_rq(last, rq)) {
+                               spin_unlock(&ve->base.timeline.lock);
+                               return; /* leave this rq for another engine */
+                       }
+
+                       GEM_TRACE("%s: virtual rq=%llx:%lld%s, new engine? %s\n",
+                                 engine->name,
+                                 rq->fence.context,
+                                 rq->fence.seqno,
+                                 i915_request_completed(rq) ? "!" :
+                                 i915_request_started(rq) ? "*" :
+                                 "",
+                                 yesno(engine != ve->siblings[0]));
+
+                       ve->request = NULL;
+                       ve->base.execlists.queue_priority_hint = INT_MIN;
+                       rb_erase_cached(rb, &execlists->virtual);
+                       RB_CLEAR_NODE(rb);
+
+                       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);
+
+                               if (!list_empty(&ve->context.signals))
+                                       virtual_xfer_breadcrumbs(ve, 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->num_siblings; 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;
@@ -1838,6 +2099,25 @@ static void reset_csb_pointers(struct intel_engine_execlists *execlists)
                               &execlists->csb_status[reset_value]);
 }
 
+static struct i915_request *active_request(struct i915_request *rq)
+{
+       const struct list_head * const list = &rq->engine->timeline.requests;
+       const struct intel_context * const context = rq->hw_context;
+       struct i915_request *active = NULL;
+
+       list_for_each_entry_from_reverse(rq, list, link) {
+               if (i915_request_completed(rq))
+                       break;
+
+               if (rq->hw_context != context)
+                       break;
+
+               active = rq;
+       }
+
+       return active;
+}
+
 static void __execlists_reset(struct intel_engine_cs *engine, bool stalled)
 {
        struct intel_engine_execlists * const execlists = &engine->execlists;
@@ -1858,7 +2138,8 @@ static void __execlists_reset(struct intel_engine_cs *engine, bool stalled)
        if (!port_isset(execlists->port))
                goto out_clear;
 
-       ce = port_request(execlists->port)->hw_context;
+       rq = port_request(execlists->port);
+       ce = rq->hw_context;
 
        /*
         * Catch up with any missed context-switch interrupts.
@@ -1871,16 +2152,10 @@ static void __execlists_reset(struct intel_engine_cs *engine, bool stalled)
         */
        execlists_cancel_port_requests(execlists);
 
-       /* Push back any incomplete requests for replay after the reset. */
-       rq = __unwind_incomplete_requests(engine);
+       rq = active_request(rq);
        if (!rq)
                goto out_replay;
 
-       if (rq->hw_context != ce) { /* caught just before a CS event */
-               rq = NULL;
-               goto out_replay;
-       }
-
        /*
         * If this request hasn't started yet, e.g. it is waiting on a
         * semaphore, we need to avoid skipping the request or else we
@@ -1927,13 +2202,16 @@ static void __execlists_reset(struct intel_engine_cs *engine, bool stalled)
        }
        execlists_init_reg_state(regs, ce, engine, ce->ring);
 
-       /* Rerun the request; its payload has been neutered (if guilty). */
 out_replay:
+       /* Rerun the request; its payload has been neutered (if guilty). */
        ce->ring->head =
                rq ? intel_ring_wrap(ce->ring, rq->head) : ce->ring->tail;
        intel_ring_update_space(ce->ring);
        __execlists_update_reg_state(ce, engine);
 
+       /* Push back any incomplete requests for replay after the reset. */
+       __unwind_incomplete_requests(engine);
+
 out_clear:
        execlists_clear_all_active(execlists);
 }
@@ -2007,6 +2285,26 @@ static void execlists_cancel_requests(struct intel_engine_cs *engine)
                i915_priolist_free(p);
        }
 
+       /* Cancel all attached virtual engines */
+       while ((rb = rb_first_cached(&execlists->virtual))) {
+               struct virtual_engine *ve =
+                       rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
+
+               rb_erase_cached(rb, &execlists->virtual);
+               RB_CLEAR_NODE(rb);
+
+               spin_lock(&ve->base.timeline.lock);
+               if (ve->request) {
+                       ve->request->engine = engine;
+                       __i915_request_submit(ve->request);
+                       dma_fence_set_error(&ve->request->fence, -EIO);
+                       i915_request_mark_complete(ve->request);
+                       ve->base.execlists.queue_priority_hint = INT_MIN;
+                       ve->request = NULL;
+               }
+               spin_unlock(&ve->base.timeline.lock);
+       }
+
        /* Remaining _unready_ requests will be nop'ed when submitted */
 
        execlists->queue_priority_hint = INT_MIN;
@@ -2491,12 +2789,15 @@ static void execlists_init_reg_state(u32 *regs,
        bool rcs = engine->class == RENDER_CLASS;
        u32 base = engine->mmio_base;
 
-       /* A context is actually a big batch buffer with several
+       /*
+        * A context is actually a big batch buffer with several
         * MI_LOAD_REGISTER_IMM commands followed by (reg, value) pairs. The
         * values we are setting here are only for the first context restore:
         * on a subsequent save, the GPU will recreate this batchbuffer with new
         * values (including all the missing MI_LOAD_REGISTER_IMM commands that
         * we are not initializing here).
+        *
+        * Must keep consistent with virtual_update_register_offsets().
         */
        regs[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(rcs ? 14 : 11) |
                                 MI_LRI_FORCE_POSTED;
@@ -2715,6 +3016,320 @@ error_deref_obj:
        return ret;
 }
 
+static void virtual_context_destroy(struct kref *kref)
+{
+       struct virtual_engine *ve =
+               container_of(kref, typeof(*ve), context.ref);
+       unsigned int n;
+
+       GEM_BUG_ON(ve->request);
+       GEM_BUG_ON(ve->context.active);
+
+       for (n = 0; n < ve->num_siblings; 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);
+
+               /* Detachment is lazily performed in the execlists tasklet */
+               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_engine_initial_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->num_siblings);
+       if (!swp)
+               return;
+
+       swap(ve->siblings[swp], ve->siblings[0]);
+       virtual_update_register_offsets(ve->context.lrc_reg_state,
+                                       ve->siblings[0]);
+}
+
+static int virtual_context_pin(struct intel_context *ce)
+{
+       struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
+       int err;
+
+       /* Note: we must use a real engine class for setting up reg state */
+       err = __execlists_context_pin(ce, ve->siblings[0]);
+       if (err)
+               return err;
+
+       virtual_engine_initial_hint(ve);
+       return 0;
+}
+
+static void virtual_context_enter(struct intel_context *ce)
+{
+       struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
+       unsigned int n;
+
+       for (n = 0; n < ve->num_siblings; n++)
+               intel_engine_pm_get(ve->siblings[n]);
+}
+
+static void virtual_context_exit(struct intel_context *ce)
+{
+       struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
+       unsigned int n;
+
+       ce->saturated = 0;
+       for (n = 0; n < ve->num_siblings; n++)
+               intel_engine_pm_put(ve->siblings[n]);
+}
+
+static const struct intel_context_ops virtual_context_ops = {
+       .pin = virtual_context_pin,
+       .unpin = execlists_context_unpin,
+
+       .enter = virtual_context_enter,
+       .exit = virtual_context_exit,
+
+       .destroy = virtual_context_destroy,
+};
+
+static void virtual_submission_tasklet(unsigned long data)
+{
+       struct virtual_engine * const ve = (struct virtual_engine *)data;
+       const int prio = ve->base.execlists.queue_priority_hint;
+       unsigned int n;
+
+       local_irq_disable();
+       for (n = 0; READ_ONCE(ve->request) && n < ve->num_siblings; 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)) {
+                       /*
+                        * Cheat and avoid rebalancing the tree if we can
+                        * reuse this node in situ.
+                        */
+                       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 *rq)
+{
+       struct virtual_engine *ve = to_virtual_engine(rq->engine);
+
+       GEM_TRACE("%s: rq=%llx:%lld\n",
+                 ve->base.name,
+                 rq->fence.context,
+                 rq->fence.seqno);
+
+       GEM_BUG_ON(ve->base.submit_request != virtual_submit_request);
+
+       GEM_BUG_ON(ve->request);
+       ve->base.execlists.queue_priority_hint = rq_prio(rq);
+       WRITE_ONCE(ve->request, rq);
+
+       tasklet_schedule(&ve->base.execlists.tasklet);
+}
+
+struct intel_context *
+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 == 0)
+               return ERR_PTR(-EINVAL);
+
+       if (count == 1)
+               return intel_context_create(ctx, siblings[0]);
+
+       ve = kzalloc(struct_size(ve, siblings, count), GFP_KERNEL);
+       if (!ve)
+               return ERR_PTR(-ENOMEM);
+
+       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, NULL);
+       if (err)
+               goto err_put;
+       i915_timeline_set_subclass(&ve->base.timeline, TIMELINE_VIRTUAL);
+
+       intel_engine_init_execlists(&ve->base);
+
+       ve->base.cops = &virtual_context_ops;
+       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);
+
+       intel_context_init(&ve->context, ctx, &ve->base);
+
+       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) {
+                       DRM_DEBUG("duplicate %s entry in load balancer\n",
+                                 sibling->name);
+                       err = -EINVAL;
+                       goto err_put;
+               }
+
+               /*
+                * The virtual engine implementation is tightly coupled to
+                * the execlists backend -- we push out request directly
+                * into a tree inside each physical engine. We could support
+                * layering if we handle cloning of the requests and
+                * submitting a copy into each backend.
+                */
+               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->num_siblings++] = sibling;
+               ve->base.mask |= sibling->mask;
+
+               /*
+                * All physical engines must be compatible for their emission
+                * functions (as we build the instructions during request
+                * construction and do not alter them before submission
+                * on the physical engine). We use the engine class as a guide
+                * here, although that could be refined.
+                */
+               if (ve->base.class != OTHER_CLASS) {
+                       if (ve->base.class != sibling->class) {
+                               DRM_DEBUG("invalid mixing of engine class, sibling %d, already %d\n",
+                                         sibling->class, ve->base.class);
+                               err = -EINVAL;
+                               goto err_put;
+                       }
+                       continue;
+               }
+
+               ve->base.class = sibling->class;
+               ve->base.uabi_class = sibling->uabi_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;
+       }
+
+       return &ve->context;
+
+err_put:
+       intel_context_put(&ve->context);
+       return ERR_PTR(err);
+}
+
+struct intel_context *
+intel_execlists_clone_virtual(struct i915_gem_context *ctx,
+                             struct intel_engine_cs *src)
+{
+       struct virtual_engine *se = to_virtual_engine(src);
+       struct intel_context *dst;
+
+       dst = intel_execlists_create_virtual(ctx,
+                                            se->siblings,
+                                            se->num_siblings);
+       if (IS_ERR(dst))
+               return dst;
+
+       return dst;
+}
+
 void intel_execlists_show_requests(struct intel_engine_cs *engine,
                                   struct drm_printer *m,
                                   void (*show_request)(struct drm_printer *m,
@@ -2772,6 +3387,29 @@ 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);
 }
 
index a0dc907a7249753ba71aec2e553d95a0e737ee50..5530606052e5b5837b0a0aed6ca3dc7cd7e21316 100644 (file)
@@ -114,4 +114,13 @@ void intel_execlists_show_requests(struct intel_engine_cs *engine,
                                                        const char *prefix),
                                   unsigned int max);
 
+struct intel_context *
+intel_execlists_create_virtual(struct i915_gem_context *ctx,
+                              struct intel_engine_cs **siblings,
+                              unsigned int count);
+
+struct intel_context *
+intel_execlists_clone_virtual(struct i915_gem_context *ctx,
+                             struct intel_engine_cs *src);
+
 #endif /* _INTEL_LRC_H_ */
index 5b3d8e33f1cfdf6cbeb3a6fea47fa1c1e01d7b6a..f880271fb9bad47f1f9ef2bea5d3eea0fcae2184 100644 (file)
@@ -1310,6 +1310,185 @@ err_unlock:
        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_context *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]) {
+                       err = -ENOMEM;
+                       nctx = n;
+                       goto out;
+               }
+
+               ve[n] = intel_execlists_create_virtual(ctx[n],
+                                                      siblings, nsibling);
+               if (IS_ERR(ve[n])) {
+                       kernel_context_close(ctx[n]);
+                       err = PTR_ERR(ve[n]);
+                       nctx = n;
+                       goto out;
+               }
+
+               err = intel_context_pin(ve[n]);
+               if (err) {
+                       intel_context_put(ve[n]);
+                       kernel_context_close(ctx[n]);
+                       nctx = n;
+                       goto out;
+               }
+       }
+
+       err = igt_live_test_begin(&t, i915, __func__, ve[0]->engine->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_create(ve[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_create(ve[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]->engine->name,
+                                      request[nc]->fence.context,
+                                      request[nc]->fence.seqno);
+
+                               GEM_TRACE("%s(%s) failed at request %llx:%lld\n",
+                                         __func__, ve[0]->engine->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]->engine->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_context_unpin(ve[nc]);
+               intel_context_put(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])
+                               continue;
+
+                       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[] = {
@@ -1322,6 +1501,7 @@ 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))
index 67f8a4a807a011c423fe4e28c5e1d340067abda5..fe82d3571072b0bbdcba564f0af6e0ea53b4018f 100644 (file)
@@ -91,4 +91,9 @@ static inline bool __tasklet_enable(struct tasklet_struct *t)
        return atomic_dec_and_test(&t->count);
 }
 
+static inline bool __tasklet_is_scheduled(struct tasklet_struct *t)
+{
+       return test_bit(TASKLET_STATE_SCHED, &t->state);
+}
+
 #endif /* __I915_GEM_H__ */
index 24736fcd463d94816d2d72d718e3c8cf93ac9b66..68db6b28e6067923ea647f430ed074c89a740b20 100644 (file)
@@ -86,6 +86,7 @@
  */
 
 #include <linux/log2.h>
+#include <linux/nospec.h>
 
 #include <drm/i915_drm.h>
 
@@ -1218,7 +1219,6 @@ __intel_context_reconfigure_sseu(struct intel_context *ce,
        int ret;
 
        GEM_BUG_ON(INTEL_GEN(ce->gem_context->i915) < 8);
-       GEM_BUG_ON(ce->engine->id != RCS0);
 
        ret = intel_context_lock_pinned(ce);
        if (ret)
@@ -1412,7 +1412,100 @@ struct set_engines {
        struct i915_gem_engines *engines;
 };
 
+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 *stack[16];
+       struct intel_engine_cs **siblings;
+       struct intel_context *ce;
+       u16 num_siblings, idx;
+       unsigned int n;
+       int err;
+
+       if (!HAS_EXECLISTS(set->ctx->i915))
+               return -ENODEV;
+
+       if (USES_GUC_SUBMISSION(set->ctx->i915))
+               return -ENODEV; /* not implement yet */
+
+       if (get_user(idx, &ext->engine_index))
+               return -EFAULT;
+
+       if (idx >= set->engines->num_engines) {
+               DRM_DEBUG("Invalid placement value, %d >= %d\n",
+                         idx, set->engines->num_engines);
+               return -EINVAL;
+       }
+
+       idx = array_index_nospec(idx, set->engines->num_engines);
+       if (set->engines->engines[idx]) {
+               DRM_DEBUG("Invalid placement[%d], already occupied\n", idx);
+               return -EEXIST;
+       }
+
+       if (get_user(num_siblings, &ext->num_siblings))
+               return -EFAULT;
+
+       err = check_user_mbz(&ext->flags);
+       if (err)
+               return err;
+
+       err = check_user_mbz(&ext->mbz64);
+       if (err)
+               return err;
+
+       siblings = stack;
+       if (num_siblings > ARRAY_SIZE(stack)) {
+               siblings = kmalloc_array(num_siblings,
+                                        sizeof(*siblings),
+                                        GFP_KERNEL);
+               if (!siblings)
+                       return -ENOMEM;
+       }
+
+       for (n = 0; n < num_siblings; n++) {
+               struct i915_engine_class_instance ci;
+
+               if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) {
+                       err = -EFAULT;
+                       goto out_siblings;
+               }
+
+               siblings[n] = intel_engine_lookup_user(set->ctx->i915,
+                                                      ci.engine_class,
+                                                      ci.engine_instance);
+               if (!siblings[n]) {
+                       DRM_DEBUG("Invalid sibling[%d]: { class:%d, inst:%d }\n",
+                                 n, ci.engine_class, ci.engine_instance);
+                       err = -EINVAL;
+                       goto out_siblings;
+               }
+       }
+
+       ce = intel_execlists_create_virtual(set->ctx, siblings, n);
+       if (IS_ERR(ce)) {
+               err = PTR_ERR(ce);
+               goto out_siblings;
+       }
+
+       if (cmpxchg(&set->engines->engines[idx], NULL, ce)) {
+               intel_context_put(ce);
+               err = -EEXIST;
+               goto out_siblings;
+       }
+
+out_siblings:
+       if (siblings != stack)
+               kfree(siblings);
+
+       return err;
+}
+
 static const i915_user_extension_fn set_engines__extensions[] = {
+       [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_engines__load_balance,
 };
 
 static int
@@ -1737,14 +1830,29 @@ static int clone_engines(struct i915_gem_context *dst,
 
        clone->i915 = dst->i915;
        for (n = 0; n < e->num_engines; n++) {
+               struct intel_engine_cs *engine;
+
                if (!e->engines[n]) {
                        clone->engines[n] = NULL;
                        continue;
                }
+               engine = e->engines[n]->engine;
 
-               clone->engines[n] =
-                       intel_context_create(dst, e->engines[n]->engine);
-               if (!clone->engines[n]) {
+               /*
+                * Virtual engines are singletons; they can only exist
+                * inside a single context, because they embed their
+                * HW context... As each virtual context implies a single
+                * timeline (each engine can only dequeue a single request
+                * at any time), it would be surprising for two contexts
+                * to use the same engine. So let's create a copy of
+                * the virtual engine instead.
+                */
+               if (intel_engine_is_virtual(engine))
+                       clone->engines[n] =
+                               intel_execlists_clone_virtual(dst, engine);
+               else
+                       clone->engines[n] = intel_context_create(dst, engine);
+               if (IS_ERR_OR_NULL(clone->engines[n])) {
                        __free_engines(clone, n);
                        goto err_unlock;
                }
index d215dcdf0b1eac2a3a0ee5920b76b69611307594..78ceb56d78017fac77b6484608ff21a7e9bc46bf 100644 (file)
@@ -150,17 +150,26 @@ sched_lock_engine(const struct i915_sched_node *node,
                  struct intel_engine_cs *locked,
                  struct sched_cache *cache)
 {
-       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);
                memset(cache, 0, sizeof(*cache));
                spin_lock(&engine->timeline.lock);
+               locked = engine;
        }
 
-       return engine;
+       GEM_BUG_ON(locked != engine);
+       return locked;
 }
 
 static inline int rq_prio(const struct i915_request *rq)
@@ -272,6 +281,7 @@ static void __i915_schedule(struct i915_sched_node *node,
        spin_lock(&engine->timeline.lock);
 
        /* Fifo and depth-first replacement ensure our deps execute before us */
+       engine = sched_lock_engine(node, engine, &cache);
        list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
                INIT_LIST_HEAD(&dep->dfs_link);
 
@@ -283,8 +293,11 @@ static void __i915_schedule(struct i915_sched_node *node,
                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 (!cache.priolist)
                                cache.priolist =
                                        i915_sched_lookup_priolist(engine,
index 5256a0b5c5f77b005199f9d5a2b3e09100ab1070..1688705f4a2b7633db520f40f71c82cd10347ba6 100644 (file)
@@ -26,6 +26,7 @@ struct i915_timeline {
        spinlock_t lock;
 #define TIMELINE_CLIENT 0 /* default subclass */
 #define TIMELINE_ENGINE 1
+#define TIMELINE_VIRTUAL 2
        struct mutex mutex; /* protects the flow of requests */
 
        unsigned int pin_count;
index 62396d575e28a8e42d105e2ee0923178b2646469..f9770948161c39b1ec540ad13d0c26a06f3d15a0 100644 (file)
@@ -137,6 +137,7 @@ struct i915_engine_class_instance {
        __u16 engine_class; /* see enum drm_i915_gem_engine_class */
        __u16 engine_instance;
 #define I915_ENGINE_CLASS_INVALID_NONE -1
+#define I915_ENGINE_CLASS_INVALID_VIRTUAL -2
 };
 
 /**
@@ -1608,8 +1609,46 @@ 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;
+
+       __u16 engine_index;
+       __u16 num_siblings;
+       __u32 flags; /* all undefined flags must be zero */
+
+       __u64 mbz64; /* reserved for future use; must be zero */
+
+       struct i915_engine_class_instance engines[0];
+} __attribute__((packed));
+
+#define I915_DEFINE_CONTEXT_ENGINES_LOAD_BALANCE(name__, N__) struct { \
+       struct i915_user_extension base; \
+       __u16 engine_index; \
+       __u16 num_siblings; \
+       __u32 flags; \
+       __u64 mbz64; \
+       struct i915_engine_class_instance engines[N__]; \
+} __attribute__((packed)) name__
+
 struct i915_context_param_engines {
        __u64 extensions; /* linked chain of extension blocks, 0 terminates */
+#define I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE 0 /* see i915_context_engines_load_balance */
        struct i915_engine_class_instance engines[0];
 } __attribute__((packed));