Re: 2.6.20-rc4: known regressions with patches (v3)

[David Chinner <dgc@sgi.com>]

On Thu, Jan 11, 2007 at 02:02:41PM -0800, Andrew Morton wrote:
> On Fri, 12 Jan 2007 08:39:16 +1100
> David Chinner <dgc@sgi.com> wrote:
> 
> > On Thu, Jan 11, 2007 at 06:13:29AM +0100, Adrian Bunk wrote:
> > > This email lists some known regressions in 2.6.20-rc4 compared to 2.6.19
> > > with patches available.
> > > 
> > > Subject    : BUG: at mm/truncate.c:60 cancel_dirty_page()  (XFS)
> > > References : http://lkml.org/lkml/2007/1/5/308
> > > Submitter  : Sami Farin <7atbggg02@sneakemail.com>
> > > Handled-By : David Chinner <dgc@sgi.com>
> > > Patch      : http://lkml.org/lkml/2007/1/7/201
> > > Status     : patch available
> > 
> > Patch is broken, do not merge. The original had an off-by-one bug in
> > it, and the fixed one I have has just shown a worse problem than
> > before - partial page truncation (i.e.  filesystem block size less
> > than page size) is busted because invalidate_complete_page2_range() can
> > only handle complete pages.
> > 
> > Andrew - looking at unmap_mapping_pages, it says it cannot handle
> > partial pages and must get rid of them whereas vmtrucate() handles
> > partial pages but changes file size so can't be used. I see that
> > vmtruncate handles this by not unmapping the first partial page.
> > 
> > I can use the vmtruncate mechanism (unmap_mapping_pages, then
> > truncate_inode_pages) but that seems racy to me because we are not
> > actually truncating the file so a mmap could remap a page between
> > the unmap and the truncate and hence we still get the warning.
> 
> Yes, truncate relies upon there being nothing outside i_size, and that
> i_mutex is held.
> 
> > So the question is - is there any generic function that handles
> > this case (i.e. don't unmap first partial page, unmap the rest,
> > partial truncate of first page, complete truncate of the rest)
> > without racing? Or do I need to write a variation of
> > invalidate_inode_pages2_range() to do this?
> 
> umm, nothing I can immediately think of.  Perhaps you can generalise
> vmtruncate_range() a bit?

I had a look at that - apart from being used for actually freeing disk
blocks as well (punching a hole in the file) - it requires locks that
we may or may not be able to grab and still has the problem of
separate calls to unmap_mapping_pages and truncate_inode_pages_range.

Unless I'm misunderstanding the purpose of vmtruncate_range() I
don't think it's the right API to be using because XFS only needs
to invalidate the page cache (hence my thoughts on a variant of
invalidate_inode_pages2_range being required).

Am I making sense, or do I need more coffe this morning?

Cheers,

Dave.
-- 
Dave Chinner
Principal Engineer
SGI Australian Software Group