io-mapping: ensure io_mapping_map_atomic _is_ atomic
Daniel Vetter [Wed, 28 Sep 2011 09:57:23 +0000 (11:57 +0200)]
commit 24dd85ff723f142093f44244764b9b5c152235b8 upstream.

For the !HAVE_ATOMIC_IOMAP case the stub functions did not call
pagefault_disable/_enable. The i915 driver relies on the map
actually being atomic, otherwise it can deadlock with it's own
pagefault handler in the gtt pwrite fastpath.

This is exercised by gem_mmap_gtt from the intel-gpu-toosl gem
testsuite.

v2: Chris Wilson noted the lack of an include.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38115
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>

include/linux/io-mapping.h

index 8cdcc2a..1feeb52 100644 (file)
@@ -117,6 +117,8 @@ io_mapping_unmap(void __iomem *vaddr)
 
 #else
 
+#include <linux/uaccess.h>
+
 /* this struct isn't actually defined anywhere */
 struct io_mapping;
 
@@ -138,12 +140,14 @@ static inline void __iomem *
 io_mapping_map_atomic_wc(struct io_mapping *mapping,
                         unsigned long offset)
 {
+       pagefault_disable();
        return ((char __force __iomem *) mapping) + offset;
 }
 
 static inline void
 io_mapping_unmap_atomic(void __iomem *vaddr)
 {
+       pagefault_enable();
 }
 
 /* Non-atomic map/unmap */