OSDN Git Service

drm/i915: add infrastructure to hold off preemption on a request
authorLionel Landwerlin <lionel.g.landwerlin@intel.com>
Tue, 9 Jul 2019 16:42:27 +0000 (17:42 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Tue, 9 Jul 2019 20:26:40 +0000 (21:26 +0100)
We want to set this flag in the next commit on requests containing
perf queries so that the result of the perf query can just be a delta
of global counters, rather than doing post processing of the OA
buffer.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
[ickle: add basic selftest for nopreempt]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190709164227.25859-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gt/intel_lrc.c
drivers/gpu/drm/i915/gt/selftest_lrc.c
drivers/gpu/drm/i915/i915_priolist_types.h
drivers/gpu/drm/i915/i915_request.c
drivers/gpu/drm/i915/i915_request.h
drivers/gpu/drm/i915/intel_guc_submission.c
drivers/gpu/drm/i915/intel_pm.c

index dec7354..19ce8eb 100644 (file)
@@ -259,6 +259,17 @@ static int effective_prio(const struct i915_request *rq)
        int prio = rq_prio(rq);
 
        /*
+        * If this request is special and must not be interrupted at any
+        * cost, so be it. Note we are only checking the most recent request
+        * in the context and so may be masking an earlier vip request. It
+        * is hoped that under the conditions where nopreempt is used, this
+        * will not matter (i.e. all requests to that context will be
+        * nopreempt for as long as desired).
+        */
+       if (i915_request_has_nopreempt(rq))
+               prio = I915_PRIORITY_UNPREEMPTABLE;
+
+       /*
         * On unwinding the active request, we give it a priority bump
         * if it has completed waiting on any semaphore. If we know that
         * the request has already started, we can prevent an unwanted
index 672bdaa..b9b881a 100644 (file)
@@ -721,6 +721,114 @@ static void preempt_client_fini(struct preempt_client *c)
        kernel_context_close(c->ctx);
 }
 
+static int live_nopreempt(void *arg)
+{
+       struct drm_i915_private *i915 = arg;
+       struct intel_engine_cs *engine;
+       struct preempt_client a, b;
+       enum intel_engine_id id;
+       intel_wakeref_t wakeref;
+       int err = -ENOMEM;
+
+       /*
+        * Verify that we can disable preemption for an individual request
+        * that may be being observed and not want to be interrupted.
+        */
+
+       if (!HAS_LOGICAL_RING_PREEMPTION(i915))
+               return 0;
+
+       mutex_lock(&i915->drm.struct_mutex);
+       wakeref = intel_runtime_pm_get(&i915->runtime_pm);
+
+       if (preempt_client_init(i915, &a))
+               goto err_unlock;
+       if (preempt_client_init(i915, &b))
+               goto err_client_a;
+       b.ctx->sched.priority = I915_USER_PRIORITY(I915_PRIORITY_MAX);
+
+       for_each_engine(engine, i915, id) {
+               struct i915_request *rq_a, *rq_b;
+
+               if (!intel_engine_has_preemption(engine))
+                       continue;
+
+               engine->execlists.preempt_hang.count = 0;
+
+               rq_a = igt_spinner_create_request(&a.spin,
+                                                 a.ctx, engine,
+                                                 MI_ARB_CHECK);
+               if (IS_ERR(rq_a)) {
+                       err = PTR_ERR(rq_a);
+                       goto err_client_b;
+               }
+
+               /* Low priority client, but unpreemptable! */
+               rq_a->flags |= I915_REQUEST_NOPREEMPT;
+
+               i915_request_add(rq_a);
+               if (!igt_wait_for_spinner(&a.spin, rq_a)) {
+                       pr_err("First client failed to start\n");
+                       goto err_wedged;
+               }
+
+               rq_b = igt_spinner_create_request(&b.spin,
+                                                 b.ctx, engine,
+                                                 MI_ARB_CHECK);
+               if (IS_ERR(rq_b)) {
+                       err = PTR_ERR(rq_b);
+                       goto err_client_b;
+               }
+
+               i915_request_add(rq_b);
+
+               /* B is much more important than A! (But A is unpreemptable.) */
+               GEM_BUG_ON(rq_prio(rq_b) <= rq_prio(rq_a));
+
+               /* Wait long enough for preemption and timeslicing */
+               if (igt_wait_for_spinner(&b.spin, rq_b)) {
+                       pr_err("Second client started too early!\n");
+                       goto err_wedged;
+               }
+
+               igt_spinner_end(&a.spin);
+
+               if (!igt_wait_for_spinner(&b.spin, rq_b)) {
+                       pr_err("Second client failed to start\n");
+                       goto err_wedged;
+               }
+
+               igt_spinner_end(&b.spin);
+
+               if (engine->execlists.preempt_hang.count) {
+                       pr_err("Preemption recorded x%d; should have been suppressed!\n",
+                              engine->execlists.preempt_hang.count);
+                       err = -EINVAL;
+                       goto err_wedged;
+               }
+
+               if (igt_flush_test(i915, I915_WAIT_LOCKED))
+                       goto err_wedged;
+       }
+
+       err = 0;
+err_client_b:
+       preempt_client_fini(&b);
+err_client_a:
+       preempt_client_fini(&a);
+err_unlock:
+       intel_runtime_pm_put(&i915->runtime_pm, wakeref);
+       mutex_unlock(&i915->drm.struct_mutex);
+       return err;
+
+err_wedged:
+       igt_spinner_end(&b.spin);
+       igt_spinner_end(&a.spin);
+       i915_gem_set_wedged(i915);
+       err = -EIO;
+       goto err_client_b;
+}
+
 static int live_suppress_self_preempt(void *arg)
 {
        struct drm_i915_private *i915 = arg;
@@ -2028,6 +2136,7 @@ int intel_execlists_live_selftests(struct drm_i915_private *i915)
                SUBTEST(live_busywait_preempt),
                SUBTEST(live_preempt),
                SUBTEST(live_late_preempt),
+               SUBTEST(live_nopreempt),
                SUBTEST(live_suppress_self_preempt),
                SUBTEST(live_suppress_wait_preempt),
                SUBTEST(live_chain_preempt),
index 49709de..b02dea1 100644 (file)
@@ -17,6 +17,16 @@ enum {
        I915_PRIORITY_NORMAL = I915_CONTEXT_DEFAULT_PRIORITY,
        I915_PRIORITY_MAX = I915_CONTEXT_MAX_USER_PRIORITY + 1,
 
+       /*
+        * Requests containing performance queries must not be preempted by
+        * another context. They get scheduled with their default priority and
+        * once they reach the execlist ports we ensure that they stick on the
+        * HW until finished by pretending that they have maximum priority,
+        * i.e. nothing can have higher priority and force us to usurp the
+        * active request.
+        */
+       I915_PRIORITY_UNPREEMPTABLE = INT_MAX,
+
        I915_PRIORITY_INVALID = INT_MIN
 };
 
index 5ff87c4..222c9c5 100644 (file)
@@ -292,7 +292,7 @@ static bool i915_request_retire(struct i915_request *rq)
                dma_fence_signal_locked(&rq->fence);
        if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags))
                i915_request_cancel_breadcrumb(rq);
-       if (rq->waitboost) {
+       if (i915_request_has_waitboost(rq)) {
                GEM_BUG_ON(!atomic_read(&rq->i915->gt_pm.rps.num_waiters));
                atomic_dec(&rq->i915->gt_pm.rps.num_waiters);
        }
@@ -684,7 +684,7 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
        rq->file_priv = NULL;
        rq->batch = NULL;
        rq->capture_list = NULL;
-       rq->waitboost = false;
+       rq->flags = 0;
        rq->execution_mask = ALL_ENGINES;
 
        INIT_LIST_HEAD(&rq->active_list);
index b58ceef..313df3c 100644 (file)
@@ -216,7 +216,9 @@ struct i915_request {
        /** Time at which this request was emitted, in jiffies. */
        unsigned long emitted_jiffies;
 
-       bool waitboost;
+       unsigned long flags;
+#define I915_REQUEST_WAITBOOST BIT(0)
+#define I915_REQUEST_NOPREEMPT BIT(1)
 
        /** timeline->request entry for this request */
        struct list_head link;
@@ -430,6 +432,17 @@ static inline void i915_request_mark_complete(struct i915_request *rq)
        rq->hwsp_seqno = (u32 *)&rq->fence.seqno; /* decouple from HWSP */
 }
 
+static inline bool i915_request_has_waitboost(const struct i915_request *rq)
+{
+       return rq->flags & I915_REQUEST_WAITBOOST;
+}
+
+static inline bool i915_request_has_nopreempt(const struct i915_request *rq)
+{
+       /* Preemption should only be disabled very rarely */
+       return unlikely(rq->flags & I915_REQUEST_NOPREEMPT);
+}
+
 bool i915_retire_requests(struct drm_i915_private *i915);
 
 #endif /* I915_REQUEST_H */
index 12c2235..f104b94 100644 (file)
@@ -707,6 +707,16 @@ static inline int rq_prio(const struct i915_request *rq)
        return rq->sched.attr.priority | __NO_PREEMPTION;
 }
 
+static inline int effective_prio(const struct i915_request *rq)
+{
+       int prio = rq_prio(rq);
+
+       if (i915_request_has_nopreempt(rq))
+               prio = I915_PRIORITY_UNPREEMPTABLE;
+
+       return prio;
+}
+
 static struct i915_request *schedule_in(struct i915_request *rq, int idx)
 {
        trace_i915_request_in(rq, idx);
@@ -747,7 +757,8 @@ static void __guc_dequeue(struct intel_engine_cs *engine)
                                &engine->i915->guc.preempt_work[engine->id];
                        int prio = execlists->queue_priority_hint;
 
-                       if (i915_scheduler_need_preempt(prio, rq_prio(last))) {
+                       if (i915_scheduler_need_preempt(prio,
+                                                       effective_prio(last))) {
                                intel_write_status_page(engine,
                                                        I915_GEM_HWS_PREEMPT,
                                                        GUC_PREEMPT_INPROGRESS);
index 87244d8..0cecea2 100644 (file)
@@ -6876,9 +6876,10 @@ void gen6_rps_boost(struct i915_request *rq)
        /* Serializes with i915_request_retire() */
        boost = false;
        spin_lock_irqsave(&rq->lock, flags);
-       if (!rq->waitboost && !dma_fence_is_signaled_locked(&rq->fence)) {
+       if (!i915_request_has_waitboost(rq) &&
+           !dma_fence_is_signaled_locked(&rq->fence)) {
                boost = !atomic_fetch_inc(&rps->num_waiters);
-               rq->waitboost = true;
+               rq->flags |= I915_REQUEST_WAITBOOST;
        }
        spin_unlock_irqrestore(&rq->lock, flags);
        if (!boost)