Gradual removal of excessive barriers in pvclock reading functions (commits 502dfeff239e8313bfbe906ca0a1a6827ac8481b, a3eb97bd80134ba07864ca00747466c02118aca1) ended up removing too much: although rdtsc is now orderd WRT other loads, there's no protection against the compiler reordering the loads of ->version with the loads of other fields.
E.g. on my system gcc-5.3.1 generates code which loads ->system_time and ->flags outside of the ->version test loop. (Re)introduce the compiler barriers around accesses to the contents of pvclock. While at this, make the function a bit more compact by removing unnecessary local variables. Signed-off-by: Roman Kagan <rka...@virtuozzo.com> Cc: Thomas Gleixner <t...@linutronix.de> Cc: Ingo Molnar <mi...@redhat.com> Cc: "H. Peter Anvin" <h...@zytor.com> Cc: x...@kernel.org Cc: Andy Lutomirski <l...@kernel.org> Cc: Borislav Petkov <b...@suse.de> Cc: Paolo Bonzini <pbonz...@redhat.com> Cc: sta...@vger.kernel.org --- arch/x86/include/asm/pvclock.h | 17 +++++------------ 1 file changed, 5 insertions(+), 12 deletions(-) diff --git a/arch/x86/include/asm/pvclock.h b/arch/x86/include/asm/pvclock.h index fdcc040..65c4de2 100644 --- a/arch/x86/include/asm/pvclock.h +++ b/arch/x86/include/asm/pvclock.h @@ -80,18 +80,11 @@ static __always_inline unsigned __pvclock_read_cycles(const struct pvclock_vcpu_time_info *src, cycle_t *cycles, u8 *flags) { - unsigned version; - cycle_t ret, offset; - u8 ret_flags; - - version = src->version; - - offset = pvclock_get_nsec_offset(src); - ret = src->system_time + offset; - ret_flags = src->flags; - - *cycles = ret; - *flags = ret_flags; + unsigned version = src->version; + barrier(); + *cycles = src->system_time + pvclock_get_nsec_offset(src); + *flags = src->flags; + barrier(); return version; } -- 2.5.5