We use the global device inode, shared amongst all files, and not the
user's device filp to provide the backing storage for the mmap. The
vma->vm_file provides a redundant reference that breaks existing
expected behaviour that closing the user's device fd will release the
resources bound to it, if a mmap persists. (Even without the
vma->vm_file, the mmap will persist past the user's fd as the storage is
bound to the device, i.e. our reference is on the object not file.)

Fixes: cc662126b413 ("drm/i915: Introduce DRM_I915_GEM_MMAP_OFFSET")
Signed-off-by: Chris Wilson <[email protected]>
---
 drivers/gpu/drm/i915/gem/i915_gem_mman.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_mman.c 
b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
index 4caf7d750c34..3c5695a2858a 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
@@ -751,6 +751,10 @@ int i915_gem_mmap(struct file *filp, struct vm_area_struct 
*vma)
 
        vma->vm_flags |= VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
        vma->vm_private_data = mmo;
+       if (vma->vm_file) { /* we keep mmo->obj ref, not vm_file */
+               fput(vma->vm_file);
+               vma->vm_file = NULL;
+       }
 
        switch (mmo->mmap_type) {
        case I915_MMAP_TYPE_WC:
-- 
2.25.0.rc0

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

Reply via email to