Re: [PATCH v3 17/18] x86/kvm/tsc: Drop extra barrier and use rdtsc_ordered in kvmclock

From: Paolo Bonzini
Date: Wed Jun 17 2015 - 03:47:35 EST




On 17/06/2015 02:36, Andy Lutomirski wrote:
> __pvclock_read_cycles had an unnecessary barrier. Get rid of that
> barrier and clean up the code by just using rdtsc_ordered().
>
> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> Cc: Radim Krcmar <rkrcmar@xxxxxxxxxx>
> Cc: Marcelo Tosatti <mtosatti@xxxxxxxxxx>
> Cc: kvm@xxxxxxxxxxxxxxx
> Signed-off-by: Andy Lutomirski <luto@xxxxxxxxxx>
> ---
>
> I'm hoping to get an ack for this to go in through -tip. (Arguably
> I'm the maintainer of this code given how it's used, but I should
> still ask for an ack.)
>
> arch/x86/include/asm/pvclock.h | 21 ++++++++++++---------
> 1 file changed, 12 insertions(+), 9 deletions(-)

Can you send a URL to the rest of the series? I've never even seen v1
or v2 so I have no idea of what this is about.

> diff --git a/arch/x86/include/asm/pvclock.h b/arch/x86/include/asm/pvclock.h
> index 6084bce345fc..cf2329ca4812 100644
> --- a/arch/x86/include/asm/pvclock.h
> +++ b/arch/x86/include/asm/pvclock.h
> @@ -62,7 +62,18 @@ static inline u64 pvclock_scale_delta(u64 delta, u32 mul_frac, int shift)
> static __always_inline
> u64 pvclock_get_nsec_offset(const struct pvclock_vcpu_time_info *src)
> {
> - u64 delta = rdtsc() - src->tsc_timestamp;
> + /*
> + * Note: emulated platforms which do not advertise SSE2 support
> + * break rdtsc_ordered, resulting in kvmclock not using the
> + * necessary RDTSC barriers. Without barriers, it is possible
> + * that RDTSC instruction is executed before prior loads,
> + * resulting in violation of monotonicity.
> + *
> + * On an SMP guest without SSE2, it's unclear how anything is
> + * supposed to work correctly, though -- memory fences
> + * (e.g. smp_mb) are important for more than just timing.
> + */

On an SMP guest without SSE2, memory fences are obtained with e.g. "lock
addb $0, (%esp)".

> + u64 delta = rdtsc_ordered() - src->tsc_timestamp;
> return pvclock_scale_delta(delta, src->tsc_to_system_mul,
> src->tsc_shift);
> }
> @@ -76,17 +87,9 @@ unsigned __pvclock_read_cycles(const struct pvclock_vcpu_time_info *src,
> u8 ret_flags;
>
> version = src->version;
> - /* Note: emulated platforms which do not advertise SSE2 support
> - * result in kvmclock not using the necessary RDTSC barriers.
> - * Without barriers, it is possible that RDTSC instruction reads from
> - * the time stamp counter outside rdtsc_barrier protected section
> - * below, resulting in violation of monotonicity.
> - */
> - rdtsc_barrier();
> offset = pvclock_get_nsec_offset(src);
> ret = src->system_time + offset;
> ret_flags = src->flags;
> - rdtsc_barrier();
>
> *cycles = ret;
> *flags = ret_flags;
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/