Re: [PATCH] Export shmem_file_setup and shmem_getpage for DRM-GEM

[Hugh Dickins <hugh@veritas.com>]

On Mon, 4 Aug 2008, Keith Packard wrote:
> On Mon, 2008-08-04 at 18:09 +0100, Hugh Dickins wrote:
> 
> > Whether such usage conforms to VFS API I'm not so sure: as I understand
> > it, it's really for internal use by a filesystem
> 
> Sure, but presumably it could even be used by a layered file system?

Could, yes, but should? I wouldn't presume to answer with any authority.

> >  - if it's going to be
> > used beyond that, we ought to add a check that the filesystem it's used
> > upon really has a ->readpage method (and I'd rather we add such a check
> > than you do it at your end, in case we change the implementation later
> > to use something other than a ->readpage method - Nick, you'll be
> > nauseated to hear I was looking to see if ->fault with a pseudo-vma
> > could do it).  But if the layering police are happy with this, I am.
> 
> It seems like I should put a check into my code that is kernel version
> dependent so that I can't oops if someone tries to use a filesystem that
> doesn't have ->readpage.

Well, I guess put the check on ->readpage into your code for now, and
by the time GEM gets into Linus's tree, we should have -EINVAL checks
on NULL filler() in __read_cache_page() and read_cache_page_async(),
so remove check at your end before final submission.

(You could leave it there, and strictly we ought to update GEM if we
make any change to our implementation; but it is the kind of detail
that gets overlooked - witness the way I failed to grasp the readahead
side-effects of adding ->readpage into tmpfs until recently.  I'm just
afraid we'd break you unwittingly: better not, though easily fixed.)

I'm not sending the patch right now, waiting to see if this direction
wins general favour.

> So, when I release the pages from the page cache, I'm currently calling
> mark_page_accessed for all pages, and set_page_dirty for pages which may
> have been written by the GPU. Are those calls still needed?

I think you should drop the mark_page_accessed(): that's done in
read_cache_page_async() as part of the initial read_mapping_page().
But do it again when releasing if you think there's a good chance
that object will be wanted again shortly.  set_page_dirty() if
modified by GPU, yes, that would still be needed.

For how long are these objects' pages pinned in memory like this?
I ask because Rik & Lee have patches in -mm, trying to avoid long
scans of LRUs cluttered with unevictable pages.  I've no idea
whether you're adding a lot or a little to that problem.

Hugh