Re: [PATCH] fbdefio: add set_page_dirty handler to deferred IO FB

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Mon, 2008-08-18 at 23:38 -0700, Andrew Morton wrote:
> On Tue, 19 Aug 2008 07:02:45 +0100 Ian Campbell <ijc@hellion.org.uk> wrote:
> 
> > Fixes kernel BUG at lib/radix-tree.c:473.
> > 
> > Previously the handler was incidentally provided by tmpfs but this was
> > removed with:
> > 
> >   commit 14fcc23fdc78e9d32372553ccf21758a9bd56fa1
> >   Author: Hugh Dickins <hugh@veritas.com>
> >   Date:   Mon Jul 28 15:46:19 2008 -0700
> > 
> >     tmpfs: fix kernel BUG in shmem_delete_inode
> > 
> > relying on this behaviour was incorrect in any case and the BUG
> > also appeared when the device node was on an ext3 filesystem.
> > 
> > Signed-off-by: Ian Campbell <ijc@hellion.org.uk>
> > Acked-by: Jaya Kumar <jayakumar.lkml@gmail.com>
> > Acked-by: Nick Piggin <npiggin@suse.de>
> > Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> > Cc: Jaya Kumar <jayakumar.lkml@gmail.com>
> > Cc: Nick Piggin <npiggin@suse.de>
> > Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
> > Cc: Hugh Dickins <hugh@veritas.com>
> > Cc: Johannes Weiner <hannes@saeurebad.de>
> > Cc: Jeremy Fitzhardinge <jeremy@goop.org>
> > Cc: Kel Modderman <kel@otaku42.de>
> > Cc: Markus Armbruster <armbru@redhat.com>
> > Cc: stable@vger.kernel.org [14fcc23fd is in 2.6.25.14 and 2.6.26.1]
> > ---
> >  drivers/video/fb_defio.c |   12 ++++++++++++
> >  1 files changed, 12 insertions(+), 0 deletions(-)
> > 
> > diff --git a/drivers/video/fb_defio.c b/drivers/video/fb_defio.c
> > index 59df132..214bb7c 100644
> > --- a/drivers/video/fb_defio.c
> > +++ b/drivers/video/fb_defio.c
> > @@ -114,11 +114,23 @@ static struct vm_operations_struct fb_deferred_io_vm_ops = {
> >  	.page_mkwrite	= fb_deferred_io_mkwrite,
> >  };
> >  
> > +static int fb_deferred_io_set_page_dirty(struct page *page)
> > +{
> > +	if (!PageDirty(page))
> > +		SetPageDirty(page);
> > +	return 0;
> > +}
> 
> <searches, finds the thread.  "kernel BUG at lib/radix-tree.c:473!">
> 
> Is there actually any benefit in setting these pages dirty?  Or should
> this be an empty function?  I see claims in the above thread that this
> driver uses PG_dirty for optimising writeback but I can't immediately
> locate any code which actually does that.
> 
> > +static const struct address_space_operations fb_deferred_io_aops = {
> > +	.set_page_dirty = fb_deferred_io_set_page_dirty,
> > +};
> > +
> >  static int fb_deferred_io_mmap(struct fb_info *info, struct vm_area_struct *vma)
> >  {
> >  	vma->vm_ops = &fb_deferred_io_vm_ops;
> >  	vma->vm_flags |= ( VM_IO | VM_RESERVED | VM_DONTEXPAND );
> >  	vma->vm_private_data = info;
> > +	vma->vm_file->f_mapping->a_ops = &fb_deferred_io_aops;
> >  	return 0;
> >  }
> 
> Seems a bit odd to rewrite the address_space_operations at mmap()-time.
> A filesystem will usually do this on the inode when it is first being
> set up, when no other thread of control can be looking at it.
> 
> 	grep 'if .*[-]>a_ops' */*.c
> 
> and you'll see that lots of code assumes that a_ops doesn't get
> swizzled around (to contain a bunch of NULL pointers!) under its feet. 
> Maybe none of those code paths are applicable to the /dev/fb0 inode,
> but it would be painful to work out which paths _are_ applicable and to
> verify that they're all safe wrt a_ops getting whisked away.
> 
> Rewriting page->mapping within the vm_operations_struct.fault handler
> seems a bit suspect for similar reasons.
> 
> I suspect that we just shouldn't be pretending that this is a regular
> anon/pagecache page to this extent.  Maybe a suitable fix would be to
> teach fb_deferred_io_fault() to instantiate the pte itself
> (vm_insert_page()) and then return VM_FAULT_NOPAGE?
> 
> I assume there's a reason why we aren't VM_IO/VM_RESERVED/PG_reserved in
> here.

IIRC there is the requirement to mmap the framebuffer from multiple
processes, so we need a shared mapping - this can be done using
vm_insertpage() if I understand the thing correctly.

We also need the fault on write, vma_want_writenotify() will generally
fail on VM_INSERTPAGE pages, except when you set a page_mkwrite()
handler, so here we're good too.

However, we also need page_mkclean() to work, and that requires rmap -
which if memory serves me, vm_insertpage() does not do.