x86, vdso: Use asm volatile in __getcpu

In Linux 3.18 and below, GCC hoists the lsl instructions in the
pvclock code all the way to the beginning of __vdso_clock_gettime,
slowing the non-paravirt case significantly.  For unknown reasons,
presumably related to the removal of a branch, the performance issue
is gone as of

e76b027e6408 x86,vdso: Use LSL unconditionally for vgetcpu

but I don't trust GCC enough to expect the problem to stay fixed.

There should be no correctness issue, because the __getcpu calls in
__vdso_vlock_gettime were never necessary in the first place.

Note to stable maintainers: In 3.18 and below, depending on
configuration, gcc 4.9.2 generates code like this:

     9c3:       44 0f 03 e8             lsl    %ax,%r13d
     9c7:       45 89 eb                mov    %r13d,%r11d
     9ca:       0f 03 d8                lsl    %ax,%ebx

This patch won't apply as is to any released kernel, but I'll send a
trivial backported version if needed.

Fixes: 51c19b4f5927 x86: vdso: pvclock gettime support
Cc: stable@vger.kernel.org # 3.8+
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Acked-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Andy Lutomirski <luto@amacapital.net>

1 files changed, 4 insertions, 2 deletions

diff --git a/arch/x86/include/asm/vgtod.h b/arch/x86/include/asm/vgtod.h
index e7e9682a33e9..f556c4843aa1 100644
--- a/arch/x86/include/asm/vgtod.h
+++ b/arch/x86/include/asm/vgtod.h
@@ -80,9 +80,11 @@ static inline unsigned int __getcpu(void)
 
 	/*
 	 * Load per CPU data from GDT.  LSL is faster than RDTSCP and
-	 * works on all CPUs.
+	 * works on all CPUs.  This is volatile so that it orders
+	 * correctly wrt barrier() and to keep gcc from cleverly
+	 * hoisting it out of the calling function.
 	 */
-	asm("lsl %1,%0" : "=r" (p) : "r" (__PER_CPU_SEG));
+	asm volatile ("lsl %1,%0" : "=r" (p) : "r" (__PER_CPU_SEG));
 
 	return p;
 }