Re: [parisc-linux] Re: Problems with kernel mmap (failing tst-mmap-eofsync in glibc on parisc)

[James Bottomley <James.Bottomley@steeleye.com>]

On Sat, 2003-08-23 at 16:43, David S. Miller wrote:
> On 22 Aug 2003 20:09:30 -0500
> James Bottomley <James.Bottomley@SteelEye.com> wrote:
> 
> > What we were hoping is that we could rely on this little property of
> > mmap:
> > 
> >        MAP_PRIVATE
> >                   Create a private copy-on-write mapping.  Stores
> >                   to the region do not affect the original  file.
> >                   It  is  unspecified whether changes made to the
> >                   file after the mmap call  are  visible  in  the
> >                   mapped region.
> > 
> > To avoid having to flush the non-shared mappings (basically on parisc if
> > you write to a file backing a MAP_PRIVATE mapping then we don't
> > guarantee you see the update).
> > 
> > I suppose if we had a way of telling if any of the i_mmap list members
> > were really MAP_SHARED semantics mappings, then we could alter our
> > flush_dcache_page() implementation to work.
> 
> I thought about this very deeply last night and this morning.
> And what you're trying to optimize won't work.  Here is why.
> 
> If the first access to a MAP_PRIVATE mapping of a page is a read,
> we'll use the page-cache page.  This means that, with your
> optimization, during this time if another cpu write()`s into the
> page we'll lose the data update.
> 
> Sorry :(

Could you elaborate some more?  I agree that the MAP_PRIVATE mapping may
not see cpu1's write because of cache incoherencies (but that's what I
believe is covered by the `unspecified' bit of the MAP_PRIVATE
definition above).  MAP_PRIVATE is COW (i.e. the page is marked read
only while it's shared), so there can be no data to flush in the cache
for the page of the MAP_PRIVATE process...The only scenario I see where
we can get cache based data destruction is if two cache aliases both
contain dirty caches for the page (which can only happen for MAP_SHARED
RW, which we already have the correct semantics for...they're racing to
do a msync and the last one in wins).

James