x86, tsc, sched: Recompute cyc2ns_offset's during resume from sleep states
authorSuresh Siddha <suresh.b.siddha@intel.com>
Fri, 20 Aug 2010 00:03:38 +0000 (17:03 -0700)
committerIngo Molnar <mingo@elte.hu>
Fri, 20 Aug 2010 12:59:02 +0000 (14:59 +0200)
commitcd7240c0b900eb6d690ccee088a6c9b46dae815a
tree0a1ed10298a2bb2c9d6010c4d03a7f9508bdcba6
parent861d034ee814917a83bd5de4b26e3b8336ddeeb8
x86, tsc, sched: Recompute cyc2ns_offset's during resume from sleep states

TSC's get reset after suspend/resume (even on cpu's with invariant TSC
which runs at a constant rate across ACPI P-, C- and T-states). And in
some systems BIOS seem to reinit TSC to arbitrary large value (still
sync'd across cpu's) during resume.

This leads to a scenario of scheduler rq->clock (sched_clock_cpu()) less
than rq->age_stamp (introduced in 2.6.32). This leads to a big value
returned by scale_rt_power() and the resulting big group power set by the
update_group_power() is causing improper load balancing between busy and
idle cpu's after suspend/resume.

This resulted in multi-threaded workloads (like kernel-compilation) go
slower after suspend/resume cycle on core i5 laptops.

Fix this by recomputing cyc2ns_offset's during resume, so that
sched_clock() continues from the point where it was left off during
suspend.

Reported-by: Florian Pritz <flo@xssn.at>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Cc: <stable@kernel.org> # [v2.6.32+]
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1282262618.2675.24.camel@sbsiddha-MOBL3.sc.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86/include/asm/tsc.h
arch/x86/kernel/tsc.c
arch/x86/power/cpu.c