drm/i915: Complete both freed-object passes before draining the workqueue

The workqueue code complains viciously if we try to queue more work onto
the queue while attampting to drain it. As we asynchronously free
objects and defer their enqueuing with RCU, it is quite tricky to
quiesce the system before attempting to drain the workqueue. Yet drain
we must to ensure that the worker is idle before unloading the module.

Give the freed object drain 3 whole passes with multiple rcu_barrier()
to give the defer freeing of several levels each protected by RCU and
needing a grace period before its parent can be freed, ultimately
resulting in a GEM object being freed after another RCU period.

A consequence is that it will make module unload even slower.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=110550
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190501135753.8711-1-chris@chris-wilson.co.uk

diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 13270e19eb87342a9bc3e0e12f034d14e8eb8b15..9a634ba57ff95e4713541f3f2ff9b9b7159ae7b2 100644 (file)
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -2824,15 +2824,15 @@ static inline void i915_gem_drain_workqueue(struct drm_i915_private *i915)
         * grace period so that we catch work queued via RCU from the first
         * pass. As neither drain_workqueue() nor flush_workqueue() report
         * a result, we make an assumption that we only don't require more
-        * than 2 passes to catch all recursive RCU delayed work.
+        * than 3 passes to catch all _recursive_ RCU delayed work.
         *
         */
-       int pass = 2;
+       int pass = 3;
        do {
                rcu_barrier();
                i915_gem_drain_freed_objects(i915);
-               drain_workqueue(i915->wq);
        } while (--pass);
+       drain_workqueue(i915->wq);
 }
 
 struct i915_vma * __must_check