Xen devel - Dec 2011 - Re: [PATCH] remove blocked time accounting from xen "clockchip"

>>> On 21.12.11 at 14:53, Laszlo Ersek <lersek@redhat.com> wrote:
> On 12/21/11 09:32, Jan Beulich wrote:
>> I currently have no idea what the remain 1% could be attributed to,
>> so thoughts from others would certainly be welcome.
> 
> What about irqtime_account_process_tick() calling account_idle_time() 
> with "cputime_one_jiffy" -- it could more frequently trigger the 
> underflow in _cputime_to_cputime64(). In that case we might want to 
> decrease idle time (ie. account "negative" cputime against idle),
but can''t.
Despite my thinking of this supposedly being impossible, I put in
accounting of what is getting lost here. With an unexpected result:
There''s far more stolen ticks that can''t get subtracted from
one of the
other counters than such that can. With the result that if I kept a
record of how many still didn''t get used for adjustment, there would
then be about 10% under-accounting (instead on the 1% over-
accounting). I understand this as meaning that the majority of stolen
ticks don''t get mis-accounted in any of the other counters, and
it''s
just very few that need to be corrected for. But I''m not entirely
convinced of this explanation.

Jan