[patch] mm: fix XIP file writes

[Nick Piggin <npiggin@suse.de>]

On Tue, Dec 04, 2007 at 12:35:49PM +0100, Nick Piggin wrote:
> On Tue, Dec 04, 2007 at 03:26:20AM -0800, Andrew Morton wrote:
> > On Tue, 4 Dec 2007 12:21:00 +0100 Nick Piggin <npiggin@suse.de> wrote:
> > 
> > > +	 *
> > > +	 * Cannot support XIP and highmem, because our ->direct_access
> > > +	 * routine for XIP must return memory that is always addressable.
> > > +	 * If XIP was reworked to use pfns and kmap throughout, this
> > > +	 * restriction might be able to be lifted.
> > >  	 */
> > > +	gfp_flags = GFP_NOIO | __GFP_ZERO;
> > > +#ifndef CONFIG_BLK_DEV_XIP
> > > +	gfp_flags |= __GFP_HIGHMEM;
> > > +#endif
> > 
> > A dubious tradeoff?
> 
> On big highmem machines certainly. It may be somewhat useful on small
> memory systems... but having the config option there is nice for a VM
> developer without an s390 easily available ;)
> 
> But don't apply these XIP patches yet -- after a bit more testing I'm
> seeing some data corruption, so I'll have to work out what's going
> wrong with that first.

Here we go. See, brd already found a bug ;)

Can you apply the ext2 XIP patch too? And I'll resend the brd XIP patch.

---

Writing to XIP files at a non-page-aligned offset results in data corruption
because the writes were always sent to the start of the page.

Signed-off-by: Nick Piggin <npiggin@suse.de>
---
Index: linux-2.6/mm/filemap_xip.c
===================================================================
--- linux-2.6.orig/mm/filemap_xip.c
+++ linux-2.6/mm/filemap_xip.c
@@ -314,7 +314,7 @@ __xip_file_write(struct file *filp, cons
 		fault_in_pages_readable(buf, bytes);
 		kaddr = kmap_atomic(page, KM_USER0);
 		copied = bytes -
-			__copy_from_user_inatomic_nocache(kaddr, buf, bytes);
+			__copy_from_user_inatomic_nocache(kaddr + offset, buf, bytes);
 		kunmap_atomic(kaddr, KM_USER0);
 		flush_dcache_page(page);