RE: [PATCH 3/5] x86/pvclock: add vsyscall implementation

[Dan Magenheimer <dan.magenheimer@oracle.com>]

> From: Jeremy Fitzhardinge [mailto:jeremy.fitzhardinge@citrix.com]
> With this in place, I can do a gettimeofday in about 100ns on a 2.4GHz
> Q6600.  I'm sure this could be tuned a bit more, but it is 
> already much better than a syscall.

To evaluate the goodness of this, we really need a full
set of measurements for:

a) cost of rdtsc (and rdtscp if different)
b) cost of vsyscall+pvclock
c) cost of rdtsc emulated
d) cost of a hypercall that returns "hypervisor system time"

On a E6850 (3Ghz but let's use cycles), I measured;

a == 72 cycles
c == 1080 cycles
d == 780 cycles

It may be partly apples and oranges, but it looks
like a good guess for b on my machine is

b == 240 cycles

Not bad, but is there any additional context switch
cost to support it?

> From: Avi Kivity [mailto:avi@redhat.com]
> Instead of using vgetcpu() and rdtsc() independently, you can 
> use rdtscp 
> to read both atomically.  This removes the need for the 
> preempt notifier.

Xen does not currently expose rdtscp and so does not emulate
(or context switch) TSC_AUX.  Context switching TSC_AUX
is certainly possible, but will likely be expensive.
If the primary reason for vsyscall+pvclock is to maximize
performance for gettimeofday/clock_gettime, this cost
would need to be added to the mix.

> preempt notifiers are per-thread, not global, and will upset 
> the cycle 
> counters.  I'd drop them and use rdtscp instead (and give up if the 
> processor doesn't support it).

Even if rdtscp is used, in the Intel processor lineup
only the very latest (Nehalem) supports rdtscp, so
"give up" doesn't seem like a very good option, at least
in the near future.