On 14/04/2022 15:05, Thomas Hellström wrote:
On Tue, 2022-04-12 at 15:18 +0000, Robert Beckett wrote:
stolen/kernel buffers should not be mmapable by userland.
do not provide callbacks to facilitate this for these buffers.
Signed-off-by: Robert Beckett <bob.beck...@collabora.com>
---
drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 32 +++++++++++++++++++++--
--
1 file changed, 27 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
index a878910a563c..b20f81836c54 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
@@ -1092,8 +1092,8 @@ static void i915_ttm_unmap_virtual(struct
drm_i915_gem_object *obj)
ttm_bo_unmap_virtual(i915_gem_to_ttm(obj));
}
-static const struct drm_i915_gem_object_ops i915_gem_ttm_obj_ops = {
- .name = "i915_gem_object_ttm",
+static const struct drm_i915_gem_object_ops
i915_gem_ttm_user_obj_ops = {
+ .name = "i915_gem_object_ttm_user",
.flags = I915_GEM_OBJECT_IS_SHRINKABLE |
I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST,
@@ -1111,6 +1111,21 @@ static const struct drm_i915_gem_object_ops
i915_gem_ttm_obj_ops = {
.mmap_ops = &vm_ops_ttm,
};
+static const struct drm_i915_gem_object_ops
i915_gem_ttm_kern_obj_ops = {
+ .name = "i915_gem_object_ttm_kern",
+ .flags = I915_GEM_OBJECT_IS_SHRINKABLE |
+ I915_GEM_OBJECT_SELF_MANAGED_SHRINK_LIST,
+
+ .get_pages = i915_ttm_get_pages,
+ .put_pages = i915_ttm_put_pages,
+ .truncate = i915_ttm_truncate,
+ .shrink = i915_ttm_shrink,
+
+ .adjust_lru = i915_ttm_adjust_lru,
+ .delayed_free = i915_ttm_delayed_free,
+ .migrate = i915_ttm_migrate,
+};
Do we really need two different ops here?
Since if we don't have mmap ops, basically that tells GEM it should do
the mmapping rather than TTM.
That might of course come in handy for the shmem backend, but I don't
fully follow why we need this for stolen.
the main rationale for doing this was to avoid
drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c:can_mmap() presuming
that is can use I915_MMAP_TYPE_FIXED
As the original backend also did not have mmap_offset handlers for
stolen, this seemed like a reasonable design.
If desired, we could add a special case for the testing logic, but those
special cases have tendency to multiply.
Also for the framebuffer handed over from BIOS to fbdev, Does that need
mmapping and if so, how do we handle that?
I'm not sure of the usecase there. Do you know of any igt test that
tests this? I can investigate further if you do not.
/Thomas
+
void i915_ttm_bo_destroy(struct ttm_buffer_object *bo)
{
struct drm_i915_gem_object *obj = i915_ttm_to_gem(bo);
@@ -1165,10 +1180,19 @@ int __i915_gem_ttm_object_init(struct
intel_memory_region *mem,
.no_wait_gpu = false,
};
enum ttm_bo_type bo_type;
+ const struct drm_i915_gem_object_ops *ops;
int ret;
drm_gem_private_object_init(&i915->drm, &obj->base, size);
- i915_gem_object_init(obj, &i915_gem_ttm_obj_ops, &lock_class,
flags);
+
+ if (flags & I915_BO_ALLOC_USER &&
intel_region_to_ttm_type(mem) != I915_PL_STOLEN) {
+ bo_type = ttm_bo_type_device;
+ ops = &i915_gem_ttm_user_obj_ops;
+ } else {
+ bo_type = ttm_bo_type_kernel;
+ ops = &i915_gem_ttm_kern_obj_ops;
+ }
+ i915_gem_object_init(obj, ops, &lock_class, flags);
obj->bo_offset = offset;
@@ -1178,8 +1202,6 @@ int __i915_gem_ttm_object_init(struct
intel_memory_region *mem,
INIT_RADIX_TREE(&obj->ttm.get_io_page.radix, GFP_KERNEL |
__GFP_NOWARN);
mutex_init(&obj->ttm.get_io_page.lock);
- bo_type = (obj->flags & I915_BO_ALLOC_USER) ?
ttm_bo_type_device :
- ttm_bo_type_kernel;
obj->base.vma_node.driver_private = i915_gem_to_ttm(obj);
