[Intel-gfx] [PATCH 3/4] drm/i915: Restore engine->submit_request before unwedging

[Chris Wilson]

When we wedge the device, we override engine->submit_request with a nop
to ensure that all in-flight requests are marked in error. However, igt
would like to unwedge the device to test -EIO handling. This requires us
to flush those in-flight requests and restore the original
engine->submit_request.

v2: Use a vfunc to unify enabling request submission to engines
v3: Split new vfunc to a separate patch.
v4: Make the wait interruptible -- the third party fences we wait upon
may be indefinitely broken, so allow the reset to be aborted.

Fixes: 821ed7df6e2a ("drm/i915: Update reset path to fix incomplete requests")
Testcase: igt/gem_eio
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Cc: Mika Kuoppala <mika.kuoppala at intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com> #v3
---
 drivers/gpu/drm/i915/i915_drv.c |  9 ++++---
 drivers/gpu/drm/i915/i915_drv.h |  1 +
 drivers/gpu/drm/i915/i915_gem.c | 59 +++++++++++++++++++++++++++++++++++++++++
 3 files changed, 66 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index be3c81221d11..03d9e45694c9 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -1821,7 +1821,9 @@ void i915_reset(struct drm_i915_private *dev_priv)
 		return;
 
 	/* Clear any previous failed attempts at recovery. Time to try again. */
-	__clear_bit(I915_WEDGED, &error->flags);
+	if (!i915_gem_unset_wedged(dev_priv))
+		goto wakeup;
+
 	error->reset_count++;
 
 	pr_notice("drm/i915: Resetting chip after gpu hang\n");
@@ -1867,17 +1869,18 @@ void i915_reset(struct drm_i915_private *dev_priv)
 
 	i915_queue_hangcheck(dev_priv);
 
-wakeup:
+finish:
 	i915_gem_reset_finish(dev_priv);
 	enable_irq(dev_priv->drm.irq);
 
+wakeup:
 	clear_bit(I915_RESET_HANDOFF, &error->flags);
 	wake_up_bit(&error->flags, I915_RESET_HANDOFF);
 	return;
 
 error:
 	i915_gem_set_wedged(dev_priv);
-	goto wakeup;
+	goto finish;
 }
 
 static int i915_pm_suspend(struct device *kdev)
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 0b02a6d710e1..9981a216342e 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -3441,6 +3441,7 @@ int i915_gem_reset_prepare(struct drm_i915_private *dev_priv);
 void i915_gem_reset(struct drm_i915_private *dev_priv);
 void i915_gem_reset_finish(struct drm_i915_private *dev_priv);
 void i915_gem_set_wedged(struct drm_i915_private *dev_priv);
+bool i915_gem_unset_wedged(struct drm_i915_private *dev_priv);
 
 void i915_gem_init_mmio(struct drm_i915_private *i915);
 int __must_check i915_gem_init(struct drm_i915_private *dev_priv);
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index ecd1b038318a..fd611b4c1a2c 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2997,6 +2997,65 @@ void i915_gem_set_wedged(struct drm_i915_private *dev_priv)
 	mod_delayed_work(dev_priv->wq, &dev_priv->gt.idle_work, 0);
 }
 
+bool i915_gem_unset_wedged(struct drm_i915_private *i915)
+{
+	struct i915_gem_timeline *tl;
+	int i;
+
+	lockdep_assert_held(&i915->drm.struct_mutex);
+	if (!test_bit(I915_WEDGED, &i915->gpu_error.flags))
+		return true;
+
+	/* Before unwedging, make sure that all pending operations
+	 * are flushed and errored out - we may have requests waiting upon
+	 * third party fences. We marked all inflight requests as EIO, and
+	 * every execbuf since returned EIO, for consistency we want all
+	 * the currently pending requests to also be marked as EIO, which
+	 * is done inside our nop_submit_request - and so we must wait.
+	 *
+	 * No more can be submitted until we reset the wedged bit.
+	 */
+	list_for_each_entry(tl, &i915->gt.timelines, link) {
+		for (i = 0; i < ARRAY_SIZE(tl->engine); i++) {
+			struct drm_i915_gem_request *rq;
+
+			rq = i915_gem_active_peek(&tl->engine[i].last_request,
+						  &i915->drm.struct_mutex);
+			if (!rq)
+				continue;
+
+			/* We can't use our normal waiter as we want to
+			 * avoid recursively trying to handle the current
+			 * reset. The basic dma_fence_default_wait() installs
+			 * a callback for dma_fence_signal(), which is
+			 * triggered by our nop handler (indirectly, the
+			 * callback enables the signaler thread which is
+			 * woken by the nop_submit_request() advancing the seqno
+			 * and when the seqno passes the fence, the signaler
+			 * then signals the fence waking us up).
+			 */
+			if (dma_fence_default_wait(&rq->fence, true,
+						   MAX_SCHEDULE_TIMEOUT) < 0)
+				return false;
+		}
+	}
+
+	/* Undo nop_submit_request. We prevent all new i915 requests from
+	 * being queued (by disallowing execbuf whilst wedged) so having
+	 * waited for all active requests above, we know the system is idle
+	 * and do not have to worry about a thread being inside
+	 * engine->submit_request() as we swap over. So unlike installing
+	 * the nop_submit_request on reset, we can do this from normal
+	 * context and do not require stop_machine().
+	 */
+	intel_engines_reset_default_submission(i915);
+
+	smp_mb__before_atomic(); /* complete takeover before enabling execbuf */
+	clear_bit(I915_WEDGED, &i915->gpu_error.flags);
+
+	return true;
+}
+
 static void
 i915_gem_retire_work_handler(struct work_struct *work)
 {
-- 
2.11.0