| Message ID | 20190124185415.29830-1-will.deacon@arm.com |
|---|---|
| State | New |
| Headers | show |
| Series | [stable-4.9.y] locking/qspinlock: Pull in asm/byteorder.h to ensure correct endianness | expand |
On Thu, Jan 24, 2019 at 06:54:15PM +0000, Will Deacon wrote: > From: Dave Airlie <airlied@redhat.com> > > 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. That is crazy... Thanks for the patch, now queued up. greg k-h
On Fri, Jan 25, 2019 at 5:04 AM Greg KH <gregkh@linuxfoundation.org> wrote: > > On Thu, Jan 24, 2019 at 06:54:15PM +0000, Will Deacon wrote: > > From: Dave Airlie <airlied@redhat.com> > > > > 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. > > That is crazy... You've no idea, we've spent a couple of days at LCA figuring it all out :-) Dave.
diff --git a/include/asm-generic/qspinlock_types.h b/include/asm-generic/qspinlock_types.h index d10f1e7d6ba8..6503e96710fa 100644 --- a/include/asm-generic/qspinlock_types.h +++ b/include/asm-generic/qspinlock_types.h @@ -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