sh: Flush executable pages in copy_user_highpage

This resolves a problem seen when using the Android dynamic linker.
Sometimes the dynamic linker would seg-fault at start up and this
was eventually traced to the handling of a COW fault for a page which
was being modified by the linker. If there was no cache aliasing between
the kernel and the user page, the page was not flushed, leaving the
newly copied data in the D-cache. However when executing instructions
from that page, the I-cache is filled directly from external memory,
rather than the D-cache, and causing garbage to be executed.

Signed-off-by: Stuart Menefy <stuart.menefy@st.com>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>

diff --git a/arch/sh/mm/cache.c b/arch/sh/mm/cache.c
index 88d3dc3d30d50aabcbdfcff04ac197cf19470bb7..5a580ea04429801eef9bda7d3301c8311492989e 100644 (file)
--- a/arch/sh/mm/cache.c
+++ b/arch/sh/mm/cache.c
@@ -108,7 +108,8 @@ void copy_user_highpage(struct page *to, struct page *from,
                kunmap_atomic(vfrom, KM_USER0);
        }
 
-       if (pages_do_alias((unsigned long)vto, vaddr & PAGE_MASK))
+       if (pages_do_alias((unsigned long)vto, vaddr & PAGE_MASK) ||
+           (vma->vm_flags & VM_EXEC))
                __flush_purge_region(vto, PAGE_SIZE);
 
        kunmap_atomic(vto, KM_USER1);