Re: [MIPS] Fix aliasing bug in copy_user_highpage, take 2.

[Ralf Baechle <ralf@linux-mips.org>]

On Wed, Oct 17, 2007 at 01:13:49AM +0900, Atsushi Nemoto wrote:

> On Tue, 16 Oct 2007 16:38:06 +0100, Ralf Baechle <ralf@linux-mips.org> wrote:
> > > Hmm, setting SetPageDcacheDirty() will not make sure the modified data
> > > flushed before reading via the kernel mapping.  The flush_dcache_page()
> > > should be used for such case, shouldn't it?
> > 
> > You're right - and the intent is to _not_ flush the page.  But we're
> > bringing it into the cache, so we better flush it before it will be mapped
> > to userspace.  We want to delay the flush operation.
> 
> I see, but I'm afraid of unnecessary flushing might hide another
> bug...
> 
> Setting SetPageDcacheDirty() for non-modified page looks a bit
> overkill for me.  For example, in copy_user_highpage() we do not flush
> the source page if the page was not mapped to userspace.

kunmap_atomic() is supposed to deal with that.

It doesn't quite on the MIPS implementation so the kernel refuses highmem
with aliasing caches.

  Ralf