[PATCH 0/6][RFC] Rework vsyscall to avoid truncation/rounding issue in timekeeping core

[John Stultz]

This item has been on my todo list for a number of years.

One interesting bit of the timekeeping code is that we internally
keep sub-nanosecond precision. This allows for really fine
grained error accounting, which allows us to keep long term
accuracy, even if the clocksource can only make very coarse
adjustments.

Since sub-nanosecond precision isn't useful to userland, we
normally truncate this extra data off before handing it to
userland. So only the timekeeping core deals with this extra
resolution.

Brief background here:

Timekeeping roughly works as follows:
	time_ns = base_ns + cyc2ns(cycles)

With our sub-ns resolution we can internally do calculations
like:
	base_ns = 0.9
	cyc2ns(cycles) =  0.9
	Thus:
	time_ns = 0.9 + 0.9 = 1.8 (which we truncate down to 1)


Where we periodically accumulate the cyc2ns(cycles) portion into
the base_ns to avoid cycles getting to large where it might overflow.

So we might have a case where we accumulate 3 cycle "chunks", each
cycle being 10.3 ns long.

So before accumulation:
	base_ns = 0
	cyc2ns(4) = 41.2
	time_now = 41.2 (truncated to 41)

After accumulation:
	base_ns = 30.9
	cyc2ns(1) = 10.3
	time_now = 41.2 (truncated to 41)


One quirk is when we export timekeeping data to the vsyscall code,
we also truncate the extra resolution. This in the past has caused
problems, where single nanosecond inconsistencies could be detected.

So before accumulation:
	base_ns = 0
	cyc2ns(4) = 41.2 (truncated to 41)
	time_now = 41

After accumulation:
	base_ns = 30.9 (truncated to 30)
	cyc2ns(1) = 10.3 (truncated to 10)
	time_now = 40

And time looks like it went backwards!

In order to avoid this, we currently end round up to the next
nanosecond when we do accumulation. In order to keep this from
causing long term drift (as much as 1ns per tick), we add the
amount we rounded up to the error accounting, which will slow the
clocksource frequency appropriately to avoid the drift.

This works, but causes the clocksource adjustment code to do much
more work. Steven Rosdet pointed out that the unlikely() case in
timekeeping_adjust is ends up being true every time.

Further this, rounding up and slowing down adds more complexity to
the timekeeping core.

The better solution is to provide the full sub-nanosecond precision
data to the vsyscall code, so that we do the truncation on the final
data, in the exact same way the timekeeping core does, rather then
truncating some of the source data. This requires reworking the
vsyscall code paths (x86, ppc, s390, ia64) to be able to handle this
extra data.

This patch set provides an initial draft of how I'd like to solve it. 
1) Introducing first a way for the vsyscall data to access the entire
   timekeeper stat
2) Transitioning the existing update_vsyscall methods to
   update_vsyscall_old
3) Introduce the new full-resolution update_vsyscall method
4) Limit the problematic extra rounding to only systems using the
   old vsyscall method
5) Convert x86 to use the new vsyscall update and full resolution
   gettime calculation.

Powerpc, s390 and ia64 will also need to be converted, but this
allows for a slow transition.

Anyway, I'd greatly appreciate any thoughts or feedback on this
approach.

Thanks
-john

Cc: Tony Luck <tony.luck@xxxxxxxxx>
Cc: Paul Mackerras <paulus@xxxxxxxxx>
Cc: Benjamin Herrenschmidt <benh@xxxxxxxxxxxxxxxxxxx>
Cc: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
Cc: Martin Schwidefsky <schwidefsky@xxxxxxxxxx>
Cc: Paul Turner <pjt@xxxxxxxxxx>
Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
Cc: Richard Cochran <richardcochran@xxxxxxxxx>
Cc: Prarit Bhargava <prarit@xxxxxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>



John Stultz (6):
  time: Move timekeeper structure to timekeeper_internal.h for vsyscall
    changes
  time: Move update_vsyscall definitions to timekeeper_internal.h
  time: Convert CONFIG_GENERIC_TIME_VSYSCALL to
    CONFIG_GENERIC_TIME_VSYSCALL_OLD
  time: Introduce new GENERIC_TIME_VSYSCALL
  time: Only do nanosecond rounding on GENERIC_TIME_VSYSCALL_OLD
    systems
  time: Convert x86_64 to using new update_vsyscall

 arch/ia64/Kconfig                   |    2 +-
 arch/ia64/kernel/time.c             |    4 +-
 arch/powerpc/Kconfig                |    2 +-
 arch/powerpc/kernel/time.c          |    4 +-
 arch/s390/Kconfig                   |    2 +-
 arch/s390/kernel/time.c             |    4 +-
 arch/x86/include/asm/vgtod.h        |    4 +-
 arch/x86/kernel/vsyscall_64.c       |   49 +++++++++------
 arch/x86/vdso/vclock_gettime.c      |   22 ++++---
 include/linux/clocksource.h         |   16 -----
 include/linux/timekeeper_internal.h |  108 ++++++++++++++++++++++++++++++++
 kernel/time.c                       |    2 +-
 kernel/time/Kconfig                 |    4 ++
 kernel/time/timekeeping.c           |  115 ++++++++++-------------------------
 14 files changed, 200 insertions(+), 138 deletions(-)
 create mode 100644 include/linux/timekeeper_internal.h

-- 
1.7.9.5

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/