On Mon, 2008-08-04 at 20:43 +1000, Nick Piggin wrote:

quoted text
> True. It would make it possible for the userspace program to pass in
> anonymous pages, but maybe not a big deal if you're using files and
> shmem based management.

We considered using anonymous pages, but as the user-mapping is not a
feature, it seemed like it wasn't the right model. Plus, many of these
objects need to be shared across multiple processes, so anonymous pages
would be a pain there.

quoted text
> pagecache_write_begin/pagecache_write_end should be reasonable, but you
> have to be careful of the semantics of it. For example, you can't really
> read anything from the page inside the calls because the filesystem may
> not bring it up to date.

Ok, that's useful information which isn't clear from the docs.

quoted text
> read_mapping_page might help there.

That does look a lot more like what I want, as it returns an unlocked
page. And, makes my code look cleaner to boot:

        inode =3D obj->filp->f_path.dentry->d_inode;
        mapping =3D inode->i_mapping;
        for (i =3D 0; i < page_count; i++) {
                page =3D read_mapping_page(mapping, i, NULL);
                if (IS_ERR(page)) {
                        ret =3D PTR_ERR(page);
                        DRM_ERROR("read_mapping_page failed: %d\n", ret);
                        i915_gem_object_free_page_list(obj);
                        return ret;
                }
                obj_priv->page_list[i] =3D page;
        }

Does this look like it conforms to the vfs api? It appears to work when
using shmem at least.

quoted text
> Right, that's your specific implementation, but for some cases the
> memory management can map or be implemented using generic primitives.
> Using pagecache for your memory for example should work nicely. Making
> it shmem specific and using internal APIs seems like a negative step
> until you really have a good reason to.

Yup, I'm liking the general file mechanism, I used shmem only because
that seemed like the obvious file system you'd want underneath these
objects.

--=20
keith.packard@intel.com