Re: [Xen-devel] [RFC PATCH v2 10/12] cpufreq: add dom0-cpufreq driver

[Oleksandr Dmytryshyn <oleksandr.dmytryshyn@xxxxxxxxxxxxxxx>]

On Thu, Oct 23, 2014 at 1:56 PM, Ian Campbell <Ian.Campbell@xxxxxxxxxx> wrote:
> On Thu, 2014-10-23 at 12:49 +0300, Oleksandr Dmytryshyn wrote:
>> On Wed, Oct 22, 2014 at 5:00 PM, Ian Campbell <Ian.Campbell@xxxxxxxxxx> 
>> wrote:
>> > On Wed, 2014-10-22 at 16:24 +0300, Oleksandr Dmytryshyn wrote:
>> >> On Wed, Oct 22, 2014 at 2:00 PM, Ian Campbell <Ian.Campbell@xxxxxxxxxx> 
>> >> wrote:
>> >> > On Wed, 2014-10-22 at 10:30 +0100, Jan Beulich wrote:
>> >> >> >>> On 22.10.14 at 10:43, <oleksandr.dmytryshyn@xxxxxxxxxxxxxxx> wrote:
>> >> >> > On Fri, Oct 17, 2014 at 4:03 PM, Jan Beulich <JBeulich@xxxxxxxx> 
>> >> >> > wrote:
>> >> >> >>>>> On 16.10.14 at 13:27, <oleksandr.dmytryshyn@xxxxxxxxxxxxxxx> 
>> >> >> >>>>> wrote:
>> >> >> >>> +static void notify_cpufreq_domains(void)
>> >> >> >> Why "domains", not "hwdom"? You don't really want to send this
>> >> >> >> to other than the hardware domain I hope.
>> >> >> > All domains (not only hwdomain) should receive this interrupt.
>> >> >> > In case is hwdomain is Linux kernel it can automaticaly recalculate
>> >> >> > jiffies. But other domains should recalculate jiffies too. I'll
>> >> >> > implement this mechanism a bit later.
>> >> >>
>> >> >> This needs more explanation then: If all domains need to do
>> >> >> adjustments when the frequency changes, there's something
>> >> >> wrong here imo. Such changes are supposed to be visible to
>> >> >> the hypervisor only (with, in the case here, assistance - but
>> >> >> nothing else - by the hardware domain).
>> >> >
>> >> > Guests should be using the arch/generic timers as their time source.
>> >> > This is required by the hardware to increment at a fixed frequency, so
>> >> > guests shouldn't need to be aware of the change to the underlying
>> >> > processor clock, at least not for timekeeping/jiffies purposes.
>> >> >
>> >> > The timers are allowed to increment by larger amounts less frequency in
>> >> > lower power modes, i.e. +=4 every 4 ticks. Guests shouldn't need to
>> >> > adjust anything for this though IMHO.
>> >> There are many functions which are using jiffies. What about them?
>> >> All those functions will measure time not correctly if frequency
>> >> of the microprocessor will change.
>> >
>> > Jiffies is based on the timer tick, isn't it? Therefore it is ultimately
>> > based on the arch timer.
>> Even if jiffies is ticked from the timer and what about
>> 'loops_per_jiffy' variable(s)? This variable is declared for ARM and it
>> should be recalculated in every change of frequency.
>
> AFAICT it is unused if arch_timers are in use, via the call to
> register_current_timer_delay which takes the loops_per_jiffy based delay
> function out of the equation.
>
> Rather than making me run around pointing out why each potential use of
> LPJ is not relevant in practice one by one please instead point out an
> actual case where you see it being used in practice in a Xen guest using
> arch timers.
We are using Linux kernel 3.8 (as Dom0) and Android +  Linux kernel 3.8
and both kernels contain arch_timers (and function
register_current_timer_delay()).
So I will not recalculate LPJ in DomU even if some legacy code is present
which may use LPJ. I'll use VIRQ_CPUFREQ only to notify hwdom to
change frequency.

> Ian.
>

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel