drm/i915: Always run hangcheck while the GPU is busy
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 29 Jan 2018 14:41:04 +0000 (14:41 +0000)
committerChris Wilson <chris@chris-wilson.co.uk>
Wed, 31 Jan 2018 10:10:43 +0000 (10:10 +0000)
Previously, we relied on only running the hangcheck while somebody was
waiting on the GPU, in order to minimise the amount of time hangcheck
had to run. (If nobody was watching the GPU, nobody would notice if the
GPU wasn't responding -- eventually somebody would care and so kick
hangcheck into action.) However, this falls apart from around commit
4680816be336 ("drm/i915: Wait first for submission, before waiting for
request completion"), as not all waiters declare themselves to hangcheck
and so we could switch off hangcheck and miss GPU hangs even when
waiting under the struct_mutex.

If we enable hangcheck from the first request submission, and let it run
until the GPU is idle again, we forgo all the complexity involved with
only enabling around waiters. We just have to remember to be careful that
we do not declare a GPU hang when idly waiting for the next request to
be come ready, as we will run hangcheck continuously even when the
engines are stalled waiting for external events. This should be true
already as we should only be tracking requests submitted to hardware for
execution as an indicator that the engine is busy.

Fixes: 4680816be336 ("drm/i915: Wait first for submission, before waiting for request completion"
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104840
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180129144104.3921-1-chris@chris-wilson.co.uk
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_gem_request.c
drivers/gpu/drm/i915/intel_breadcrumbs.c
drivers/gpu/drm/i915/intel_hangcheck.c

index 062b21408698dc2a6725a56da5e0c76ebe858d52..63308ec016a3777259a6b56ed411182c74b098be 100644 (file)
@@ -3322,16 +3322,15 @@ i915_gem_retire_work_handler(struct work_struct *work)
                mutex_unlock(&dev->struct_mutex);
        }
 
-       /* Keep the retire handler running until we are finally idle.
+       /*
+        * Keep the retire handler running until we are finally idle.
         * We do not need to do this test under locking as in the worst-case
         * we queue the retire worker once too often.
         */
-       if (READ_ONCE(dev_priv->gt.awake)) {
-               i915_queue_hangcheck(dev_priv);
+       if (READ_ONCE(dev_priv->gt.awake))
                queue_delayed_work(dev_priv->wq,
                                   &dev_priv->gt.retire_work,
                                   round_jiffies_up_relative(HZ));
-       }
 }
 
 static void shrink_caches(struct drm_i915_private *i915)
index 0a890ef4c420e7b7fe459dd78df963ab4473b2c8..cf2f3895b70e4db61e8c82e269bfbafd7a466688 100644 (file)
@@ -285,6 +285,8 @@ static void mark_busy(struct drm_i915_private *i915)
 
        intel_engines_unpark(i915);
 
+       i915_queue_hangcheck(i915);
+
        queue_delayed_work(i915->wq,
                           &i915->gt.retire_work,
                           round_jiffies_up_relative(HZ));
index 86acac010bb8c747d7aeee49aac489dccf348d5b..62b2a20bc24ee34379b06b88175f34041081f915 100644 (file)
@@ -149,17 +149,6 @@ static void intel_breadcrumbs_fake_irq(struct timer_list *t)
                return;
 
        mod_timer(&b->fake_irq, jiffies + 1);
-
-       /* Ensure that even if the GPU hangs, we get woken up.
-        *
-        * However, note that if no one is waiting, we never notice
-        * a gpu hang. Eventually, we will have to wait for a resource
-        * held by the GPU and so trigger a hangcheck. In the most
-        * pathological case, this will be upon memory starvation! To
-        * prevent this, we also queue the hangcheck from the retire
-        * worker.
-        */
-       i915_queue_hangcheck(engine->i915);
 }
 
 static void irq_enable(struct intel_engine_cs *engine)
index 31f01d64c0212cee75f40b8f6f462c7b7f500147..348a4f7ffb674b435bdb77089f9c229abe171ab5 100644 (file)
@@ -411,7 +411,6 @@ static void i915_hangcheck_elapsed(struct work_struct *work)
        struct intel_engine_cs *engine;
        enum intel_engine_id id;
        unsigned int hung = 0, stuck = 0;
-       int busy_count = 0;
 
        if (!i915_modparams.enable_hangcheck)
                return;
@@ -429,7 +428,6 @@ static void i915_hangcheck_elapsed(struct work_struct *work)
        intel_uncore_arm_unclaimed_mmio_detection(dev_priv);
 
        for_each_engine(engine, dev_priv, id) {
-               const bool busy = intel_engine_has_waiter(engine);
                struct intel_engine_hangcheck hc;
 
                semaphore_clear_deadlocks(dev_priv);
@@ -443,16 +441,13 @@ static void i915_hangcheck_elapsed(struct work_struct *work)
                        if (hc.action != ENGINE_DEAD)
                                stuck |= intel_engine_flag(engine);
                }
-
-               busy_count += busy;
        }
 
        if (hung)
                hangcheck_declare_hang(dev_priv, hung, stuck);
 
        /* Reset timer in case GPU hangs without another request being added */
-       if (busy_count)
-               i915_queue_hangcheck(dev_priv);
+       i915_queue_hangcheck(dev_priv);
 }
 
 void intel_engine_init_hangcheck(struct intel_engine_cs *engine)