ARM: 5687/1: fix an oops with highmem
commit13f96d8f4c5a3f6a6b5e578d08869d79d690e0b2
authorNicolas Pitre <nico@cam.org>
Tue, 1 Sep 2009 21:01:27 +0000 (1 22:01 +0100)
committerRussell King <rmk+kernel@arm.linux.org.uk>
Wed, 2 Sep 2009 10:33:24 +0000 (2 11:33 +0100)
treefe33eb0bce77b41d9e7a491b1241db6a70a93df2
parent37d0892c5a94e208cf863e3b7bac014edee4346d
ARM: 5687/1: fix an oops with highmem

In xdr_partial_copy_from_skb() there is that sequence:

kaddr = kmap_atomic(*ppage, KM_SKB_SUNRPC_DATA);
[...]
flush_dcache_page(*ppage);
kunmap_atomic(kaddr, KM_SKB_SUNRPC_DATA);

Mixing flush_dcache_page() and kmap_atomic() is a bit odd,
especially since kunmap_atomic() must deal with cache issues
already.  OTOH the non-highmem case must use flush_dcache_page()
as kunmap_atomic() becomes a no op with no cache maintenance.

Problem is that with highmem the implementation of kmap_atomic()
doesn't set page->virtual, and page_address(page) returns 0 in
that case. Here flush_dcache_page() calls __flush_dcache_page()
which calls __cpuc_flush_dcache_page(page_address(page)) resulting
in a kernel oops.

None of the kmap_atomic() implementations uses set_page_address().
Hence we can assume page_address() is always expected to return 0 in
that case. Let's conditionally call __cpuc_flush_dcache_page() only
when the page address is non zero, and perform that test only when
highmem is configured.

Signed-off-by: Nicolas Pitre <nico@marvell.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
arch/arm/mm/flush.c