[PATCH v2] clocksource: exynos-mct: Register the timer for stable udelay

[Doug Anderson]

Hi,

On Fri, Jun 20, 2014 at 4:19 AM, amit daniel kachhap
<amit.daniel at samsung.com> wrote:
> On Fri, Jun 20, 2014 at 1:45 PM, Will Deacon <will.deacon at arm.com> wrote:
>> On Thu, Jun 19, 2014 at 05:40:49PM +0100, Tomasz Figa wrote:
>>> On 19.06.2014 18:31, Doug Anderson wrote:
>>> >>> My personal vote would be to submit a patch to change "cycles_t" to
>>> >>> always be 32-bits.  Given that 32-bits was fine for udelay() for ARM
>>> >>> that seems sane and simple.  If someone later comes up with a super
>>> >>> compelling reason why we need 64-bit timers for udelay (really??) then
>>> >>> they can later add all the complexity needed.
>>> >>
>>> >> Yes, this could work. I'm not sure what else cycles_t is used for, though.
>>> >
>>> > True, it is a bit questionable to change this since it's a type that's
>>> > not obviously just for udelay().  Perhaps a better option would be to
>>> > make a new typedef for the result of read_current_timer().  ...or just
>>> > change it to return a u32?
>>> >
>>>
>>> Sounds good to me, but let's hear other opinions. I'm adding Will and
>>> Jonathan as they wrote the ARM delay timer code.
>>
>> I think cycles_t is only used for small interval calculations at the moment,
>> but I remember Ted mentioning something about using it (or something
>> similar) as a source of early entropy, in which case the more bits the
>> better.
>>
> Will,
> Thanks for the clarification that cycles_t is used for short
> intervals. So it is safe to return lower 32 bit
> counter for read_current_timer.

As I looked at it more, I realized that we have two types in Linux.
There's cycle_t and cycles_t.  Whoa, confusing!

I'd perhaps advocate a wholesale rename of cycles_t to avoid the
confusion.  I don't have a good name for it, though.  cycle32_t?  Or
we could just use u32 for the function...  :-/


> Tomasz, Doug,
> As of now let me send a minimal implementation of this read delay
> timer to fix the broken udelay for exynos platforms so that it goes to
> upstream in rc releases. I will also prepare a fix for all
> raw_readl/writel in mct to relaxed version to make it consistent.

I'm reworking my 32-bit conversion patches right now and it's getting
messy to intermingle this with yours.  I'm going to pick up your patch
and include it in my series.  I hope that's OK.

My plan is:

1. For 3.16 I think it's important to fix the udelay() problems and
trying to rework cycle_t there doesn't seem like it makes sense.  I'll
just use Amit's original code that uses exynos_frc_read().  It might
not be quite as optimal but it's good as a safe bugfix.

2. I'll post the cleanup patch moving away from the __raw_readl / __raw_writel

3. I'll post a patch moving to 32-bit, including moving Amit's code to
32-bit but with a compile time warning for now.  I'll add a KConfig
depends to keep it from compiling on ARM64.  We can improve this once
we change the delay timer to always request 32-bits.

-Doug