[RFC v2] ARM: sched_clock: update epoch_cyc on resume

On Fri, Jul 27, 2012 at 8:30 PM, Colin Cross [off-list ref] wrote:

On Fri, Jul 27, 2012 at 6:15 PM, Colin Cross [off-list ref] wrote:

quoted

That patch was merged in 3.4, and my patch is on top of it.  Your
patch updates epoch_cyc and epoch_ns in suspend, but if the first call
to cyc_to_sched_clock after resume gets cyc = 0, cyc - epoch_cyc can
be negative, although it will be cast back to a large positive number.

With my patch, epoch_cyc is updated in resume to whatever
read_sched_clock() returns, and epoch_ns is still set to the suspend
value, so it appears that sched_clock has not changed between
sched_clock_suspend and sched_clock_resume.  It will work with any
timer behavior (reset to 0, reset to random value, or continuing
counting).  The old setup_sched_clock function maintains the old
behavior to appease those who like their 32kHz sched clock to continue
ticking in suspend, although I think it is more correct for all sched
clocks to be faster than 32 kHz (when possible) and stop in suspend.

I think the existing code can cause sched_clock to go backwards if the
register read by the read function resets to 0 in suspend:

In sched_clock_suspend epoch_cyc is updated to 0x00002000.

The first sched_clock call after resume gets cyc = 0, returning
epoch_ns + cyc_to_ns(0xffffe000)

Later, sched_clock gets cyc = 0x5000, returning epoch_ns +
cyc_to_ns(0x3000), which will be much smaller than the previous
sched_clock value.

Russell, any update on this?  Should sched_clock.c handle read
functions that go backwards in suspend, or should I modify the read
function to track an offset in suspend and always return a
monotonically increasing value across suspend?