Re: [PATCHv2 4/6] sched_clock: Add support for >32 bit sched_clock

[Catalin Marinas <catalin.marinas@arm.com>]

On Mon, Jun 10, 2013 at 05:12:08AM +0100, Rob Herring wrote:
> On 06/04/2013 05:21 AM, Russell King - ARM Linux wrote:
> > On Mon, Jun 03, 2013 at 06:51:59PM -0700, Stephen Boyd wrote:
> >> On 06/03/13 15:12, Russell King - ARM Linux wrote:
> >>> If you have a 56-bit clock which ticks at a period of 1ns, then
> >>> cd.rate = 1, and your sched_clock() values will be truncated to 56-bits.
> >>> The scheduler always _requires_ 64-bits from sched_clock.  That's why we
> >>> have the complicated code to extend the 32-bits-or-less to a _full_
> >>> 64-bit value.
> >>>
> >>> Let me make this clearer: sched_clock() return values _must_ without
> >>> exception monotonically increment from zero to 2^64-1 and then wrap
> >>> back to zero.  No other behaviour is acceptable for sched_clock().
> >>
> >> Ok so you're saying if we have less than 64 bits of useable information
> >> we _must_ do something to find where the wraparound will occur and
> >> adjust for it so that epoch_ns is always incrementing until 2^64-1. Fair
> >> enough. I was trying to avoid more work because on arm architected timer
> >> platforms it takes many years for that to happen.
> >>
> >> I'll see what I can do.
> > 
> > Well, 56 bits at 1ns intervals is 833 days (2^56 / (1000000000*60*60*24)).
> > We used to say that 497 days was enough several years ago, and that got
> > fixed.  We used to say 640K was enough memory for anything, and that
> > got fixed.
> 
> The ARM ARM states a minimum resolution of 40 years AND at least 56-bits
> of resolution. So a 1Gz counter would have to have more that 56 bits.

At a quick calculation, with a full 64-bit counter and 40-year roll-over
we can have maximum 14.6GHz clock. So we shouldn't just mask the top
8-bit of the counter as the bottom 56 could roll over in much less time.

-- 
Catalin