[PATCH v4 05/17] drm: add mmap() to drm_gem_object_funcs

Gerd Hoffmann kraxel at redhat.com
Mon Sep 9 10:06:53 UTC 2019


> > +               vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
> > +               vma->vm_page_prot = pgprot_writecombine(vm_get_page_prot(vma->vm_flags));
> > +               vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot);
> > +       }
> Totally unrelated discussion around HMM lead me to a bit a chase, and
> realizing that we probably want a
>     WARN_ON(!(vma->vm_flags & VM_SPECIAL));
> here, to make sure drivers set at least one of the "this is a special
> vma, don't try to treat it like pagecache/anon memory". Just to be on
> the safe side. Maybe we also want to keep the entire vma->vm_flags
> assignment in the common code, but at least the WARN_ON would be a
> good check I think. Maybe also check for VM_DONTEXPAND while at it

Hmm.  VM_SPECIAL is this:


Requiring VM_DONTEXPAND makes sense for sure.  Dunno about the other
ones.  For drm_gem_vram_helper VM_IO + VM_PFNMAP are needed.

But we also have drm_gem_shmem_helper which backs objects with normal
pages.  In fact drm_gem_shmem_mmap does this:

	/* VM_PFNMAP was set by drm_gem_mmap() */
	vma->vm_flags &= ~VM_PFNMAP;
	vma->vm_flags |= VM_MIXEDMAP;

include/linux/mm.h isn't very helpful in explaining how VM_MIXEDMAP
should be used, only saying can be both "struct page" and pfnmap, so I'm
not sure why VM_MIXEDMAP is set here, it should always be "struct page"
for shmem, no?


More information about the dri-devel mailing list