locking/qspinlock: Pull in asm/byteorder.h to ensure correct endianness
authorDave Airlie <airlied@redhat.com>
Thu, 24 Jan 2019 18:54:15 +0000 (18:54 +0000)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sat, 26 Jan 2019 08:38:36 +0000 (09:38 +0100)
This commit is not required upstream, but is required for the 4.9.y
stable series.

Upstream commit 101110f6271c ("Kbuild: always define endianess in
kconfig.h") ensures that either __LITTLE_ENDIAN or __BIG_ENDIAN is
defined to reflect the endianness of the target CPU architecture
regardless of whether or not <asm/byteorder.h> has been #included. The
upstream definition of 'struct qspinlock' relies on this property.

Unfortunately, the 4.9.y stable series does not provide this guarantee,
so the 'spin_unlock()' routine can erroneously treat the underlying
lockword as big-endian on little-endian architectures using native
qspinlock (i.e. x86_64 without PV) if the caller has not included
<asm/byteorder.h>. This can lead to hangs such as the one in
'i915_gem_request()' reported via bugzilla:

  https://bugzilla.kernel.org/show_bug.cgi?id=202063

Fix the issue by ensuring that <asm/byteorder.h> is #included in
<asm/qspinlock_types.h>, where 'struct qspinlock' is defined.

Cc: <stable@vger.kernel.org> # 4.9
Signed-off-by: Dave Airlie <airlied@redhat.com>
[will: wrote commit message]
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
include/asm-generic/qspinlock_types.h

index d10f1e7d6ba8c37140ae9332b59399baae4ffdf1..6503e96710fa50629c01d08b77375801cff70e1c 100644 (file)
@@ -18,6 +18,8 @@
 #ifndef __ASM_GENERIC_QSPINLOCK_TYPES_H
 #define __ASM_GENERIC_QSPINLOCK_TYPES_H
 
+#include <asm/byteorder.h>
+
 /*
  * Including atomic.h with PARAVIRT on will cause compilation errors because
  * of recursive header file incluson via paravirt_types.h. So don't include