From mboxrd@z Thu Jan  1 00:00:00 1970
From: robherring2@gmail.com (Rob Herring)
Date: Wed, 17 Apr 2013 18:18:23 -0500
Subject: [GIT PULL] Timer clean-ups for 3.10, Part 2
In-Reply-To: <516EE86A.9070003@codeaurora.org>
References: <516720B4.3000703@gmail.com> <516DEC68.8030101@codeaurora.org>
 <516EA638.1060800@gmail.com> <516EE86A.9070003@codeaurora.org>
Message-ID: <516F2DBF.30401@gmail.com>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On 04/17/2013 01:22 PM, Stephen Boyd wrote:
> On 04/17/13 06:40, Rob Herring wrote:
>> On 04/16/2013 07:27 PM, Stephen Boyd wrote:
>>> On 04/11/13 13:44, Rob Herring wrote:
>>>> Rob Herring (13):
>>>>       ARM: sched_clock: allow changing to higher frequency counter
>>>>       ARM: make sched_clock just call a function pointer
>>>>       ARM: arch_timer: use full 64-bit counter for sched_clock
>>> If I leave my system in the bootloader for a while this seems to cause
>>> my sched_clock timestamps to jump once the sched_clock is setup. It also
>>> sets up a sched_clock twice because read_sched_clock ==
>>> jiffy_sched_clock_read.
>>>
>>> [    0.000000] Switching to timer-based delay loop
>>> [    0.000000] sched_clock: ARM arch timer >56 bits at 19200kHz,
>>> resolution 52ns
>>> [    0.000000] sched_clock: 32 bits at 100 Hz, resolution 10000000ns,
>>> wraps every 4294967286ms
>>> [    0.000000] Console: colour dummy device 80x30
>>> [16645.193054] Calibrating delay loop (skipped), value calculated using
>>> timer frequency.. 38.40 BogoMIPS (lpj=192000)
>>>
>>> I suspect it's because we don't do any cyc_to_sched_clock() stuff in the
>>> arm architected timer case. Instead we just return the value from the
>>> counter when we really should do some sort of subtraction from the first
>>> value we read.
>>>
>>> I'm also curious how this is going to work for suspend/resume because it
>>> doesn't look like we're going to stop sched_clock on arm architected
>>> timer systems. See 6a4dae5e138a3 (ARM: 7565/1: sched: stop sched_clock()
>>> during suspend, 2012-10-23) for why we need to do this.
>>
>> Well, I think arm64 is broken in both ways too. So we should fix this
>> for both.
>>
>> I think this can be handled in a much more simple way than the 32-bit
>> code since we don't need to deal with wrapping.
>>
>> Maintain a cycle offset that starts as the cycle count at init time.
>> This offset can be subtracted from the current count. On suspend and
>> resume, we need to calculate the cycle count delta while in suspend and
>> then add this to the cycle offset.
> 
> Agreed. It looks like we're missing out on the irq time accounting stuff
> because enable_sched_clock_irqtime() is never called too.
> 

Yes, but until ARM selects HAVE_IRQ_TIME_ACCOUNTING that doesn't matter.
Looks like adding that was forgotten.

Rob