[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] Proof(?) for system-wide TSC invariance

There's been a lot of debate as to whether the Invariant
TSC flag in recent Intel and AMD processors should be
"trusted" by Xen, especially on larger multi-socket or
multi-enclosure systems.

I recently learned an interesting fact that reinforces
this trust:  On Intel QPI and AMD HT systems, the clock
signal that is used to drive the Invariant TSC is the
same one that is used for the system-wide cache-coherency
protocol! IOW on any ccNUMA system where the TSC invariant
bit is set on any processor, TSC invariance applies
to all processors, or the system is broken.

This means that if:
1) The Invariant TSC bit is set; and
2) The BIOS has sync'ed the TSC's (which can be
   tested once at boot time by Xen using the
   newly added check_tsc_warp() routine); and
3) System software (Xen) never writes to any
   TSC on any processor (which is true in Xen
   when the Invariant TSC bit is set after
   a recent patch);

then the TSC is "reliable" across all processors
and will remain reliable and any difference (e.g. due
to signal propagation delays) will be unobservable
by software.... that is "time" will never go backwards.

While my personal main concern involves TSC in userland,
TSC reliability has other (minor) implications inside
Xen itself:

1) Performance: Xen's 1 Hz clock synchronization mechanism
   is (I think) entirely unneeded.
2) Power: An idle system can (I think) now be truly

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.