Currently the sh_cmt clocksource timer is disabled or enabled
unconditionally on clocksource suspend resp. resume, even if a
better clocksource is present (e.g. arch_sys_counter) and the
sh_cmt clocksource is not enabled.

As sh_cmt is a syscore device when its timer is enabled, this
may lead to a genpd.prepared_count imbalance in the presence of
PM Domains, which may cause a lock-up during reboot after s2ram.

During suspend:
  - pm_genpd_prepare() is called for all non-syscore devices (incl.
    sh_cmt), increasing genpd.prepared_count for each device,
  - clocksource.suspend() is called for all clocksource devices,
  - sh_cmt_clocksource_suspend() calls sh_cmt_stop(), which is a no-op
    as the clocksource was not enabled.

During resume:
  - clocksource.resume() is called for all clocksource devices,
  - sh_cmt_clocksource_resume() calls sh_cmt_start(), which enables the
    clocksource timer, and turns sh_cmt into a syscore device,
  - pm_genpd_complete() is called for all non-syscore devices (excl.
    sh_cmt now!), decreasing genpd.prepared_count for each device but
    sh_cmt.

Now genpd.prepared_count of the PM Domain containing sh_cmt is
still 1 instead of zero.  On subsequent suspend/resume cycles,
sh_cmt is still a syscore device, hence it's skipped for
pm_genpd_{prepare,complete}(), keeping the imbalance of
genpd.prepared_count at 1.

During reboot:

  - platform_drv_shutdown() is called for any platform device that has
    a driver with a .shutdown() method (only rcar-dmac on R-Car Gen2),

  - platform_drv_shutdown() calls dev_pm_domain_detach(), which
    calls genpd_dev_pm_detach(),

  - genpd_dev_pm_detach() keeps calling pm_genpd_remove_device() until
    it doesn't return -EAGAIN[*],

  - If the device is part of the same PM Domain as sh_cmt,
    pm_genpd_remove_device() always fails with -EAGAIN due to
    genpd.prepared_count > 0.

  - Infinite loop in genpd_dev_pm_detach()[*].

[*] Commit 93af5e93 ("PM / Domains: Avoid infinite loops in
    attach/detach code") already limited the number of loop iterations,
    avoiding the lock-up.

To fix this, only disable or enable the clocksource timer on
clocksource suspend resp. resume if the clocksource was enabled.

This was tested on r8a7791/koelsch with the CPG Clock Domain:

  - using arch_sys_counter as the clocksource, which is the default, and
    which showed the problem,

  - using sh_cmt as a clocksource ("echo ffca0000.timer > \
    /sys/devices/system/clocksource/clocksource0/current_clocksource"),
    which behaves the same as before.
Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be>
Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
Acked-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1438875126-12596-2-git-send-email-daniel.lezcano@linaro.orgSigned-off-by: Ingo Molnar <mingo@kernel.org>