Re: [Qemu-devel] TSC in qem[-kvm] 1.1+ and in-kernel irqchip

[Marcelo Tosatti]

On Thu, Aug 09, 2012 at 10:27:43PM +0400, Michael Tokarev wrote:
> As a follow-up to the patch "tsc: use kvmclock for
> calibration".
> 
> There's another problem reported by several users.
> The sympthom is that grub does not show boot menu,
> it boots default entry right away without any pause.
> 
> After quite some debugging it turned out to be
> TSC issue.  Grub uses tsc for its timeout handling.
> When setting timeout to some very large value
> (10000), I can see the counter is ticking backwards
> at a very high speed, ticking from 10000 to 0 in
> about 5 seconds.
> 
> Running kvm -cpu host,-tsc forces grub to use
> rtc clocksource, and the problem goes away.
> 
> The most interesting thing is that this is a
> problem new for qemu-kvm 1.1 (and is still
> present in current git), 1.0 version had no
> such issue.  And it only happens when in-kernel
> irqchip is enabled -- running with -no-kvm-irqchip
> also fixes the grub problem, so that tsc starts
> counting "correctly" for grub again.
> 
> Gerd mentioned mis-calibration of bios timer
> when host is heavily loaded.  I tested grub on
> my workstation today which was completely idle,
> no other processes running.
> 
> It smells like a bug in kvm somewhere.  And it
> happens when I explicitly pin kvm to a single
> core, so tsc should tick correctly even if its
> syncronization is broken between cores.
> 
> Current qemu also has this issue (since 1.1),
> since it also has in-kernel irqchip support now.
> 
> FWIW, here's the TSC calibration routine from
> grub:
> 
> /* Calibrate the TSC based on the RTC.  */
> static void
> calibrate_tsc (void)
> {
>   /* First calibrate the TSC rate (relative, not absolute time). */
>   grub_uint64_t start_tsc;
>   grub_uint64_t end_tsc;
> 
>   start_tsc = grub_get_tsc ();
>   grub_pit_wait (0xffff);
>   end_tsc = grub_get_tsc ();
> 
>   tsc_ticks_per_ms = grub_divmod64 (end_tsc - start_tsc, 55, 0);
> }

Emulation of grub_pit_wait sequence by in-kernel PIT is probably broken.

QEMU PIT emulation is also affected by miscalibration.

Please provide steps to reproduce.