x86: Do not try to sync identity map for non-mapped pages

kernel_map_sync_memtype() is called from a variety of contexts.  The
pat.c code that calls it seems to ensure that it is not called for
non-ram areas by checking via pat_pagerange_is_ram().  It is important
that it only be called on the actual identity map because there *IS*
no map to sync for highmem pages, or for memory holes.

The ioremap.c uses are not as careful as those from pat.c, and call
kernel_map_sync_memtype() on PCI space which is in the middle of the
kernel identity map _range_, but is not actually mapped.

This patch adds a check to kernel_map_sync_memtype() which probably
duplicates some of the checks already in pat.c.  But, it is necessary
for the ioremap.c uses and shouldn't hurt other callers.

I have reproduced this bug and this patch fixes it for me and the
original bug reporter:

	https://lkml.org/lkml/2013/2/5/396

Signed-off-by: Dave Hansen <dave@linux.vnet.ibm.com>
Link: http://lkml.kernel.org/r/20130307163151.D9B58C4E@kernel.stglabs.ibm.com
Signed-off-by: Dave Hansen <dave@sr71.net>
Tested-by: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>

diff --git a/arch/x86/mm/pat.c b/arch/x86/mm/pat.c
index 2610bd93c896871936caaa970abd414d039ff790..657438858e8358745484c828158634fba41a5b1c 100644 (file)
--- a/arch/x86/mm/pat.c
+++ b/arch/x86/mm/pat.c
@@ -563,6 +563,13 @@ int kernel_map_sync_memtype(u64 base, unsigned long size, unsigned long flags)
        if (base > __pa(high_memory-1))
                return 0;
 
+       /*
+        * some areas in the middle of the kernel identity range
+        * are not mapped, like the PCI space.
+        */
+       if (!page_is_ram(base >> PAGE_SHIFT))
+               return 0;
+
        id_sz = (__pa(high_memory-1) <= base + size) ?
                                __pa(high_memory) - base :
                                size;