[Intel-gfx] [PATCH v5] drm/i915/userptr: Probe vma range before gup

[Chris Wilson]

We want to exclude any GGTT objects from being present on our internal
lists to avoid the deadlock we may run into with our requirement for
struct_mutex during invalidate. However, if the gup_fast fails, we put
the userptr onto the workqueue and mark it as active, so that we
remember to serialise the worker upon mmu_invalidate.

v2: Hold mmap_sem to prevent modifications to the mm while we probe and
add ourselves to the interval-tree for notificiation.
v3: Rely on mmap_sem for a simpler patch.
v4: Mark up the mmap_sem nesting
v5: Don't deactivate on -EAGAIN as that means the worker is queued

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104209
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Cc: Michał Winiarski <michal.winiarski at intel.com>
---
 drivers/gpu/drm/i915/i915_gem_userptr.c | 55 +++++++++++++++++++++++++++------
 1 file changed, 45 insertions(+), 10 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_gem_userptr.c b/drivers/gpu/drm/i915/i915_gem_userptr.c
index 382a77a1097e..b16b8e226e40 100644
--- a/drivers/gpu/drm/i915/i915_gem_userptr.c
+++ b/drivers/gpu/drm/i915/i915_gem_userptr.c
@@ -598,13 +598,43 @@ __i915_gem_userptr_get_pages_schedule(struct drm_i915_gem_object *obj)
 	return ERR_PTR(-EAGAIN);
 }
 
+static int
+probe_range(struct mm_struct *mm, unsigned long addr, unsigned long len)
+{
+	const unsigned long end = addr + len;
+	struct vm_area_struct *vma;
+	int ret = -EFAULT;
+
+	for (vma = find_vma(mm, addr); vma; vma = vma->vm_next) {
+		if (vma->vm_start > addr)
+			break;
+
+		/*
+		 * Exclude any VMA that is backed only by struct_page, i.e.
+		 * IO regions that include our own GGTT mmaps. We cannot handle
+		 * such ranges, as we may encounter deadlocks around our
+		 * struct_mutex on mmu_invalidate_range.
+		 */
+		if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
+			break;
+
+		if (vma->vm_end >= end) {
+			ret = 0;
+			break;
+		}
+
+		addr = vma->vm_end;
+	}
+
+	return ret;
+}
+
 static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
 {
 	const int num_pages = obj->base.size >> PAGE_SHIFT;
 	struct mm_struct *mm = obj->userptr.mm->mm;
 	struct page **pvec;
 	struct sg_table *pages;
-	bool active;
 	int pinned;
 
 	/* If userspace should engineer that these pages are replaced in
@@ -634,7 +664,6 @@ static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
 
 	pvec = NULL;
 	pinned = 0;
-
 	if (mm == current->mm) {
 		pvec = kvmalloc_array(num_pages, sizeof(struct page *),
 				      GFP_KERNEL |
@@ -647,19 +676,25 @@ static int i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
 						       pvec);
 	}
 
-	active = false;
+	down_read_nested(&mm->mmap_sem, SINGLE_DEPTH_NESTING);
 	if (pinned < 0) {
 		pages = ERR_PTR(pinned);
 		pinned = 0;
-	} else if (pinned < num_pages) {
-		pages = __i915_gem_userptr_get_pages_schedule(obj);
-		active = pages == ERR_PTR(-EAGAIN);
+	} else if (pinned < num_pages &&
+		   probe_range(mm, obj->userptr.ptr, obj->base.size)) {
+		pages = ERR_PTR(-EFAULT);
+	} else  if (__i915_gem_userptr_set_active(obj, true)) {
+		pages = ERR_PTR(-EAGAIN);
 	} else {
-		pages = __i915_gem_userptr_alloc_pages(obj, pvec, num_pages);
-		active = !IS_ERR(pages);
+		if (pinned < num_pages)
+			pages = __i915_gem_userptr_get_pages_schedule(obj);
+		else
+			pages = __i915_gem_userptr_alloc_pages(obj,
+							       pvec, num_pages);
+		if (IS_ERR(pages) && pages != ERR_PTR(-EAGAIN))
+			__i915_gem_userptr_set_active(obj, false);
 	}
-	if (active)
-		__i915_gem_userptr_set_active(obj, true);
+	up_read(&mm->mmap_sem);
 
 	if (IS_ERR(pages))
 		release_pages(pvec, pinned);
-- 
2.15.1