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

[David S. Miller]

On 22 Aug 2003 13:56:06 -0500
James Bottomley <James.Bottomley@SteelEye.com> wrote:

> On Fri, 2003-08-22 at 13:31, David S. Miller wrote:
> > On Fri, 22 Aug 2003 19:34:41 +0100 (BST)
> > Hugh Dickins <hugh@veritas.com> wrote:
> > 
> > > And to me.  If VM_SHARED is set, then __vma_link_file puts the vma on
> > > on i_mmap_shared.  If VM_SHARED is not set, it puts the vma on i_mmap.
> > > flush_dcache_page treats i_mmap_shared and i_mmap lists equally.
> > 
> > But file system page cache writes only call flush_dache_page()
> > if the page has a non-empty i_mmap_shared list.
> 
> Hmm, but if it does that then the glibc bug test should show up on sparc
> because the i_mmap_shared list is empty if we only do MAP_SHARED of read
> only files.

Sparc64's alias'able caches are 1) write-through and 2) quite small.

I think I begin to see the issue clearly now.

But you cannot do the VM_SHARED change without an audit first.
Lots of code thinks that VM_SHARED means someone maybe wrote
to the page through a mmap().  For example look at how filemap
sync interprets this flag bit.