drm/i915: Break modeset deadlocks on reset
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 22 Jun 2017 10:56:25 +0000 (11:56 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Fri, 23 Jun 2017 12:41:55 +0000 (13:41 +0100)
Trying to do a modeset from within a reset is fraught with danger. We
can fall into a cyclic deadlock where the modeset is waiting on a
previous modeset that is waiting on a request, and since the GPU hung
that request completion is waiting on the reset. As modesetting doesn't
allow its locks to be broken and restarted, or for its *own* reset
mechanism to take over the display, we have to do something very
evil instead. If we detect that we are stuck waiting to prepare the
display reset (by using a very simple timeout), resort to cancelling all
in-flight requests and throwing the user data into /dev/null, which is
marginally better than the driver locking up and keeping that data to
itself.

This is not a fix; this is just a workaround that unbreaks machines
until we can resolve the deadlock in a way that doesn't lose data!

v2: Move the retirement from set-wegded to the i915_reset() error path,
after which we no longer any delayed worker cleanup for
i915_handle_error()
v3: C abuse for syntactic sugar
v4: Cover all waits with the timeout to catch more driver breakage

References: https://bugs.freedesktop.org/show_bug.cgi?id=99093
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170622105625.16952-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
drivers/gpu/drm/i915/i915_drv.c
drivers/gpu/drm/i915/i915_gem.c
drivers/gpu/drm/i915/i915_irq.c

index 370429e2071fba9114f90e42d2adfefa786b083d..43e925933688742ae1496e5d43eafb58a6d77a60 100644 (file)
@@ -1919,6 +1919,7 @@ wakeup:
 
 error:
        i915_gem_set_wedged(dev_priv);
+       i915_gem_retire_requests(dev_priv);
        goto finish;
 }
 
index ae3ce1314bd11d0fcc25c8dd1b8621313cf643b8..36d838677982be24bc4e17b7321fd19c1125b91e 100644 (file)
@@ -3062,7 +3062,8 @@ static void engine_set_wedged(struct intel_engine_cs *engine)
        /* Mark all executing requests as skipped */
        spin_lock_irqsave(&engine->timeline->lock, flags);
        list_for_each_entry(request, &engine->timeline->requests, link)
-               dma_fence_set_error(&request->fence, -EIO);
+               if (!i915_gem_request_completed(request))
+                       dma_fence_set_error(&request->fence, -EIO);
        spin_unlock_irqrestore(&engine->timeline->lock, flags);
 
        /* Mark all pending requests as complete so that any concurrent
@@ -3108,6 +3109,7 @@ static int __i915_gem_set_wedged_BKL(void *data)
        struct intel_engine_cs *engine;
        enum intel_engine_id id;
 
+       set_bit(I915_WEDGED, &i915->gpu_error.flags);
        for_each_engine(engine, i915, id)
                engine_set_wedged(engine);
 
@@ -3116,20 +3118,7 @@ static int __i915_gem_set_wedged_BKL(void *data)
 
 void i915_gem_set_wedged(struct drm_i915_private *dev_priv)
 {
-       lockdep_assert_held(&dev_priv->drm.struct_mutex);
-       set_bit(I915_WEDGED, &dev_priv->gpu_error.flags);
-
-       /* Retire completed requests first so the list of inflight/incomplete
-        * requests is accurate and we don't try and mark successful requests
-        * as in error during __i915_gem_set_wedged_BKL().
-        */
-       i915_gem_retire_requests(dev_priv);
-
        stop_machine(__i915_gem_set_wedged_BKL, dev_priv, NULL);
-
-       i915_gem_contexts_lost(dev_priv);
-
-       mod_delayed_work(dev_priv->wq, &dev_priv->gt.idle_work, 0);
 }
 
 bool i915_gem_unset_wedged(struct drm_i915_private *i915)
@@ -3184,6 +3173,7 @@ bool i915_gem_unset_wedged(struct drm_i915_private *i915)
         * context and do not require stop_machine().
         */
        intel_engines_reset_default_submission(i915);
+       i915_gem_contexts_lost(i915);
 
        smp_mb__before_atomic(); /* complete takeover before enabling execbuf */
        clear_bit(I915_WEDGED, &i915->gpu_error.flags);
index f25e73fe567ce75b7e0c11cf328bef30bc5e6bac..e4934d5adc9e956232d7424b15a5f770dc97769d 100644 (file)
@@ -2599,6 +2599,46 @@ static irqreturn_t gen8_irq_handler(int irq, void *arg)
        return ret;
 }
 
+struct wedge_me {
+       struct delayed_work work;
+       struct drm_i915_private *i915;
+       const char *name;
+};
+
+static void wedge_me(struct work_struct *work)
+{
+       struct wedge_me *w = container_of(work, typeof(*w), work.work);
+
+       dev_err(w->i915->drm.dev,
+               "%s timed out, cancelling all in-flight rendering.\n",
+               w->name);
+       i915_gem_set_wedged(w->i915);
+}
+
+static void __init_wedge(struct wedge_me *w,
+                        struct drm_i915_private *i915,
+                        long timeout,
+                        const char *name)
+{
+       w->i915 = i915;
+       w->name = name;
+
+       INIT_DELAYED_WORK_ONSTACK(&w->work, wedge_me);
+       schedule_delayed_work(&w->work, timeout);
+}
+
+static void __fini_wedge(struct wedge_me *w)
+{
+       cancel_delayed_work_sync(&w->work);
+       destroy_delayed_work_on_stack(&w->work);
+       w->i915 = NULL;
+}
+
+#define i915_wedge_on_timeout(W, DEV, TIMEOUT)                         \
+       for (__init_wedge((W), (DEV), (TIMEOUT), __func__);             \
+            (W)->i915;                                                 \
+            __fini_wedge((W)))
+
 /**
  * i915_reset_device - do process context error handling work
  * @dev_priv: i915 device private
@@ -2612,36 +2652,36 @@ static void i915_reset_device(struct drm_i915_private *dev_priv)
        char *error_event[] = { I915_ERROR_UEVENT "=1", NULL };
        char *reset_event[] = { I915_RESET_UEVENT "=1", NULL };
        char *reset_done_event[] = { I915_ERROR_UEVENT "=0", NULL };
+       struct wedge_me w;
 
        kobject_uevent_env(kobj, KOBJ_CHANGE, error_event);
 
        DRM_DEBUG_DRIVER("resetting chip\n");
        kobject_uevent_env(kobj, KOBJ_CHANGE, reset_event);
 
-       intel_prepare_reset(dev_priv);
+       /* Use a watchdog to ensure that our reset completes */
+       i915_wedge_on_timeout(&w, dev_priv, 5*HZ) {
+               intel_prepare_reset(dev_priv);
 
-       set_bit(I915_RESET_HANDOFF, &dev_priv->gpu_error.flags);
-       wake_up_all(&dev_priv->gpu_error.wait_queue);
+               /* Signal that locked waiters should reset the GPU */
+               set_bit(I915_RESET_HANDOFF, &dev_priv->gpu_error.flags);
+               wake_up_all(&dev_priv->gpu_error.wait_queue);
 
-       do {
-               /*
-                * All state reset _must_ be completed before we update the
-                * reset counter, for otherwise waiters might miss the reset
-                * pending state and not properly drop locks, resulting in
-                * deadlocks with the reset work.
+               /* Wait for anyone holding the lock to wakeup, without
+                * blocking indefinitely on struct_mutex.
                 */
-               if (mutex_trylock(&dev_priv->drm.struct_mutex)) {
-                       i915_reset(dev_priv);
-                       mutex_unlock(&dev_priv->drm.struct_mutex);
-               }
-
-               /* We need to wait for anyone holding the lock to wakeup */
-       } while (wait_on_bit_timeout(&dev_priv->gpu_error.flags,
-                                    I915_RESET_HANDOFF,
-                                    TASK_UNINTERRUPTIBLE,
-                                    HZ));
+               do {
+                       if (mutex_trylock(&dev_priv->drm.struct_mutex)) {
+                               i915_reset(dev_priv);
+                               mutex_unlock(&dev_priv->drm.struct_mutex);
+                       }
+               } while (wait_on_bit_timeout(&dev_priv->gpu_error.flags,
+                                            I915_RESET_HANDOFF,
+                                            TASK_UNINTERRUPTIBLE,
+                                            1));
 
-       intel_finish_reset(dev_priv);
+               intel_finish_reset(dev_priv);
+       }
 
        if (!test_bit(I915_WEDGED, &dev_priv->gpu_error.flags))
                kobject_uevent_env(kobj,