drm/i915/selftests: Exercise long preemption chains
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 14 Nov 2019 22:57:35 +0000 (22:57 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 15 Nov 2019 16:46:18 +0000 (16:46 +0000)
Verify that we can execute a long chain of dependent requests from
userspace, each one slightly more important than the last.

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/20191114225736.616885-4-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/gt/selftest_lrc.c

index d1ed3c0..16ebe4d 100644 (file)
@@ -1915,6 +1915,201 @@ err_wedged:
        goto err_client_lo;
 }
 
+static int create_gang(struct intel_engine_cs *engine,
+                      struct i915_request **prev)
+{
+       struct drm_i915_gem_object *obj;
+       struct intel_context *ce;
+       struct i915_request *rq;
+       struct i915_vma *vma;
+       u32 *cs;
+       int err;
+
+       ce = intel_context_create(engine->kernel_context->gem_context, engine);
+       if (IS_ERR(ce))
+               return PTR_ERR(ce);
+
+       obj = i915_gem_object_create_internal(engine->i915, 4096);
+       if (IS_ERR(obj)) {
+               err = PTR_ERR(obj);
+               goto err_ce;
+       }
+
+       vma = i915_vma_instance(obj, ce->vm, NULL);
+       if (IS_ERR(vma)) {
+               err = PTR_ERR(vma);
+               goto err_obj;
+       }
+
+       err = i915_vma_pin(vma, 0, 0, PIN_USER);
+       if (err)
+               goto err_obj;
+
+       cs = i915_gem_object_pin_map(obj, I915_MAP_WC);
+       if (IS_ERR(cs))
+               goto err_obj;
+
+       /* Semaphore target: spin until zero */
+       *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
+
+       *cs++ = MI_SEMAPHORE_WAIT |
+               MI_SEMAPHORE_POLL |
+               MI_SEMAPHORE_SAD_EQ_SDD;
+       *cs++ = 0;
+       *cs++ = lower_32_bits(vma->node.start);
+       *cs++ = upper_32_bits(vma->node.start);
+
+       if (*prev) {
+               u64 offset = (*prev)->batch->node.start;
+
+               /* Terminate the spinner in the next lower priority batch. */
+               *cs++ = MI_STORE_DWORD_IMM_GEN4;
+               *cs++ = lower_32_bits(offset);
+               *cs++ = upper_32_bits(offset);
+               *cs++ = 0;
+       }
+
+       *cs++ = MI_BATCH_BUFFER_END;
+       i915_gem_object_flush_map(obj);
+       i915_gem_object_unpin_map(obj);
+
+       rq = intel_context_create_request(ce);
+       if (IS_ERR(rq))
+               goto err_obj;
+
+       rq->batch = vma;
+       i915_request_get(rq);
+
+       i915_vma_lock(vma);
+       err = i915_request_await_object(rq, vma->obj, false);
+       if (!err)
+               err = i915_vma_move_to_active(vma, rq, 0);
+       if (!err)
+               err = rq->engine->emit_bb_start(rq,
+                                               vma->node.start,
+                                               PAGE_SIZE, 0);
+       i915_vma_unlock(vma);
+       i915_request_add(rq);
+       if (err)
+               goto err_rq;
+
+       i915_gem_object_put(obj);
+       intel_context_put(ce);
+
+       rq->client_link.next = &(*prev)->client_link;
+       *prev = rq;
+       return 0;
+
+err_rq:
+       i915_request_put(rq);
+err_obj:
+       i915_gem_object_put(obj);
+err_ce:
+       intel_context_put(ce);
+       return err;
+}
+
+static int live_preempt_gang(void *arg)
+{
+       struct intel_gt *gt = arg;
+       struct intel_engine_cs *engine;
+       enum intel_engine_id id;
+
+       if (!HAS_LOGICAL_RING_PREEMPTION(gt->i915))
+               return 0;
+
+       /*
+        * Build as long a chain of preempters as we can, with each
+        * request higher priority than the last. Once we are ready, we release
+        * the last batch which then precolates down the chain, each releasing
+        * the next oldest in turn. The intent is to simply push as hard as we
+        * can with the number of preemptions, trying to exceed narrow HW
+        * limits. At a minimum, we insist that we can sort all the user
+        * high priority levels into execution order.
+        */
+
+       for_each_engine(engine, gt, id) {
+               struct i915_request *rq = NULL;
+               struct igt_live_test t;
+               IGT_TIMEOUT(end_time);
+               int prio = 0;
+               int err = 0;
+               u32 *cs;
+
+               if (!intel_engine_has_preemption(engine))
+                       continue;
+
+               if (igt_live_test_begin(&t, gt->i915, __func__, engine->name))
+                       return -EIO;
+
+               do {
+                       struct i915_sched_attr attr = {
+                               .priority = I915_USER_PRIORITY(prio++),
+                       };
+
+                       err = create_gang(engine, &rq);
+                       if (err)
+                               break;
+
+                       /* Submit each spinner at increasing priority */
+                       engine->schedule(rq, &attr);
+
+                       if (prio <= I915_PRIORITY_MAX)
+                               continue;
+
+                       if (prio > (INT_MAX >> I915_USER_PRIORITY_SHIFT))
+                               break;
+
+                       if (__igt_timeout(end_time, NULL))
+                               break;
+               } while (1);
+               pr_debug("%s: Preempt chain of %d requests\n",
+                        engine->name, prio);
+
+               /*
+                * Such that the last spinner is the highest priority and
+                * should execute first. When that spinner completes,
+                * it will terminate the next lowest spinner until there
+                * are no more spinners and the gang is complete.
+                */
+               cs = i915_gem_object_pin_map(rq->batch->obj, I915_MAP_WC);
+               if (!IS_ERR(cs)) {
+                       *cs = 0;
+                       i915_gem_object_unpin_map(rq->batch->obj);
+               } else {
+                       err = PTR_ERR(cs);
+                       intel_gt_set_wedged(gt);
+               }
+
+               while (rq) { /* wait for each rq from highest to lowest prio */
+                       struct i915_request *n =
+                               list_next_entry(rq, client_link);
+
+                       if (err == 0 && i915_request_wait(rq, 0, HZ / 5) < 0) {
+                               struct drm_printer p =
+                                       drm_info_printer(engine->i915->drm.dev);
+
+                               pr_err("Failed to flush chain of %d requests, at %d\n",
+                                      prio, rq_prio(rq) >> I915_USER_PRIORITY_SHIFT);
+                               intel_engine_dump(engine, &p,
+                                                 "%s\n", engine->name);
+
+                               err = -ETIME;
+                       }
+
+                       i915_request_put(rq);
+                       rq = n;
+               }
+
+               if (igt_live_test_end(&t))
+                       err = -EIO;
+               if (err)
+                       return err;
+       }
+
+       return 0;
+}
+
 static int live_preempt_hang(void *arg)
 {
        struct intel_gt *gt = arg;
@@ -3028,6 +3223,7 @@ int intel_execlists_live_selftests(struct drm_i915_private *i915)
                SUBTEST(live_suppress_self_preempt),
                SUBTEST(live_suppress_wait_preempt),
                SUBTEST(live_chain_preempt),
+               SUBTEST(live_preempt_gang),
                SUBTEST(live_preempt_hang),
                SUBTEST(live_preempt_timeout),
                SUBTEST(live_preempt_smoke),