drm/i915: Move gt_revoke() slightly

We get a lockdep splat when the reset mutex is held, because it can be
taken from fence_wait. This conflicts with the mmu notifier we have,
because we recurse between reset mutex and mmap lock -> mmu notifier.

Remove this recursion by calling revoke_mmaps before taking the lock.

The reset code still needs fixing, as taking mmap locks during reset
is not allowed.

Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
[danvet: Add FIXME.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20210323155059.628690-64-maarten.lankhorst@linux.intel.com

drivers/gpu/drm/i915/gt/intel_reset.c

@@ -974,8 +974,6 @@ static int do_reset(struct intel_gt *gt, intel_engine_mask_t stalled_mask)
 {
 	int err, i;
 
-	gt_revoke(gt);
-
 	err = __intel_gt_reset(gt, ALL_ENGINES);
 	for (i = 0; err && i < RESET_MAX_RETRIES; i++) {
 		msleep(10 * (i + 1));
@@ -1030,6 +1028,13 @@ void intel_gt_reset(struct intel_gt *gt,
 
 	might_sleep();
 	GEM_BUG_ON(!test_bit(I915_RESET_BACKOFF, &gt->reset.flags));
+
+	/*
+	 * FIXME: Revoking cpu mmap ptes cannot be done from a dma_fence
+	 * critical section like gpu reset.
+	 */
+	gt_revoke(gt);
+
 	mutex_lock(&gt->reset.mutex);
 
 	/* Clear any previous failed attempts at recovery. Time to try again. */