[RESEND PATCH] ARM: Handle user space mapped pages in flush_kernel_dcache_page

On Sat, Jul 28, 2012 at 10:41:54AM +0200, Simon Baatz wrote:

a while ago I sent the patch above to fix a data corruption problem
on VIVT architectures (and probably VIPT aliasing).  There has been a
bit of discussion with Catalin, but there was no real conclusion on
how to proceed.  (See
http://www.spinics.net/lists/arm-kernel/msg176913.html for the
original post)

Going back to that post:

 However, this assumption is not true for direct IO or SG_IO (see
 e.g. [1] and [2]). This is why vgchange fails with mv_cesa as reported
 by me in [3]. (Btw., flush_kernel_dcache_page is also called from
 "copy_strings" in fs/exec.c when copying env/arg strings between
 processes. Thus, its use is not restricted to device drivers.)

The calls from copy_strings is not a problem - because the newly allocated
pages will have PG_arch_1 clear, and when the page is passed to set_pte(),
it will be flushed.

We _certainly_ do not want to make flush_kernel_dcache_page() do what you're
doing, because it will mean we're over-flushing *all* pages on VIVT caches.
Not only will we be flushing them for DMA, but we'll then do it again
when flush_kernel_dcache_page() is invoked, and then possibly again when
the page eventually ends up being visible to userspace.

The case is not hit too often apparently; the ingredients are PIO
(like) driver, use of flush_kernel_dcache_page(), and direct I/O. 

Right, so we need to analyse the direct IO paths and work out whether
they're using the flush_* interfaces correctly, and even whether they're
using the correct interfaces.

Note that flush_*dcache_page() are supposed to only be concerned with
page cache pages, not anonymous pages.  flush_anon_page() is all about
anonymous pages only.