[Xen-devel] RE: Xen c/s 18938

["Jiang, Yunhong" <yunhong.jiang@xxxxxxxxx>]

Jan Beulich <mailto:jbeulich@xxxxxxxxxx> wrote:
> I am observing some odd boot behavior resulting from the changes done
> in this c/s. In particular am I seeing banks being initialized
> inconsistently on my dual quad-core system:
> - after a cold boot, CPU0 has all its MCi_CTL non-zero (apparently BIOS-
>  initialized), while CPUs 1, 4, and 5 have MC1_CTL non-zero
> but the others
>  zero, and the remaining CPUs have MC1_CTL and MC3_CTL non-zero
>  (MC3 presumably is non-zero here indeed because of it being shared
>  across two of the four cores)
> - after a warm boot, all of the CPUs have all their MCi_CTL registers except
>  for MC3_CTL non-zero, so only bank 3 actually gets initialized (and
>  reported - again not for CPUs 0, 2, 3, 6, and 7).
> 
> I'm pretty certain that there should neither be a difference
> between cold
> and warm boot here,
> nor should CPU0's BIOS settings (and any
> other CPUs'
> if the BIOS cares to touch them) survive (specifically also resulting in

We thought it is ok because according to IA32 SDM , software can only write all 
0 or 1 to the MCi_CTL register, and then it will be ok either BIOS or OS to 
setup the value, and there will be no difference for cold/warm boot situation. 
(Or maybe buggy BIOS will not working like this way? )

But seems this feature is only for P6 family, we will change that code to 
consider other situation.

> MCi_STATUS not being cleared). The old code simply wrote all

The MCi_STATUS will be cleared in machine_check_poll() if it is valid (maybe we 
can always clear it).

> ones and all zeros
> respectively to MCi_CTL and MCi_STATUS without checking their original

In system with CMCI, we can't simply clear MCi_STATUS, considering when CPU 
offline/online situation and especially the CMCI ownership requirement.

> values. 
> 
> Jan
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel