diff mbox series

[RFC,071/162] drm/i915: Keep userpointer bindings if seqcount is unchanged, v2.

Message ID 20201127120718.454037-72-matthew.auld@intel.com (mailing list archive)
State New, archived
Headers show
Series DG1 + LMEM enabling | expand

Commit Message

Matthew Auld Nov. 27, 2020, 12:05 p.m. UTC 
From: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>

Instead of force unbinding and rebinding every time, we try to check
if our notifier seqcount is still correct when pages are bound. This
way we only rebind userptr when we need to, and prevent stalls.

Reported-by: kernel test robot <lkp@intel.com>
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
---
 drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 27 ++++++++++++++++++---
 1 file changed, 24 insertions(+), 3 deletions(-)
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
index fb4bc30fbd9a..d1ecc31b5e90 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c
@@ -275,12 +275,33 @@  int i915_gem_object_userptr_submit_init(struct drm_i915_gem_object *obj)
 	if (ret)
 		return ret;
 
-	/* Make sure userptr is unbound for next attempt, so we don't use stale pages. */
-	ret = i915_gem_object_userptr_unbind(obj, false);
+	/* optimistically try to preserve current pages while unlocked */
+	if (i915_gem_object_has_pages(obj) &&
+	    !mmu_interval_check_retry(&obj->userptr.notifier,
+				      obj->userptr.notifier_seq)) {
+		spin_lock(&i915->mm.notifier_lock);
+		if (obj->userptr.pvec &&
+		    !mmu_interval_read_retry(&obj->userptr.notifier,
+					     obj->userptr.notifier_seq)) {
+			obj->userptr.page_ref++;
+
+			/* We can keep using the current binding, this is the fastpath */
+			ret = 1;
+		}
+		spin_unlock(&i915->mm.notifier_lock);
+	}
+
+	if (!ret) {
+		/* Make sure userptr is unbound for next attempt, so we don't use stale pages. */
+		ret = i915_gem_object_userptr_unbind(obj, false);
+	}
 	i915_gem_object_unlock(obj);
-	if (ret)
+	if (ret < 0)
 		return ret;
 
+	if (ret > 0)
+		return 0;
+
 	notifier_seq = mmu_interval_read_begin(&obj->userptr.notifier);
 
 	pvec = kvmalloc_array(num_pages, sizeof(struct page *), GFP_KERNEL);