svn commit: r278183 - in head/sys: amd64/include conf dev/xen/timer i386/include i386/xen x86/include x86/x86

[Konstantin Belousov]

On Wed, Mar 04, 2015 at 01:24:52PM -0500, Ryan Stone wrote:
> On Wed, Feb 4, 2015 at 3:26 AM, Bryan Venteicher <bryanv at freebsd.org> wrote:
> > +static void
> > +pvclock_read_time_info(struct pvclock_vcpu_time_info *ti,
> > +    uint64_t *cycles, uint8_t *flags)
> > +{
> > +       uint32_t version;
> > +
> > +       do {
> > +               version = ti->version;
> > +               rmb();
> > +               *cycles = ti->system_time + pvclock_get_nsec_offset(ti);
> > +               *flags = ti->flags;
> > +               rmb();
> > +       } while ((ti->version & 1) != 0 || ti->version != version);
> > +}
> 
> The Linux implementation of this does a "rdtsc_barrier();" in place of
> the rmb(), and in some situations that appears to result in a full
> mfence, not just an lfence (the code is macro soup so it's difficult
> to understand what's happening).  Are you sure that the rmb() is
> enough?  I believe that the concern is whether the rdtsc() itself is
> allowed to be re-ordered around an lfence.
> 
> This stack overflow answer seems to indicate the a full mfence is
> necessary on AMD processors:
> 
> http://stackoverflow.com/questions/12631856/difference-between-rdtscp-rdtsc-memory-and-cpuid-rdtsc

Yes, Intel declares that lfence is enough, but for AMD the mfence should
be used.  See r238973/r238975 and sys/x86/x86/tsc.c.