[PATCH v2] ARM: Don't use complete() during __cpu_die

[Krzysztof Kozlowski]

On czw, 2015-02-05 at 10:53 +0000, Mark Rutland wrote:
> On Thu, Feb 05, 2015 at 10:14:30AM +0000, Krzysztof Kozlowski wrote:
> > The complete() should not be used on offlined CPU. Rewrite the
> > wait-complete mechanism with wait_on_bit_timeout().
> > 
> > The CPU triggering hot unplug (e.g. CPU0) will loop until some bit is
> > cleared. In each iteration schedule_timeout() is used with initial sleep
> > time of 1 ms. Later it is increased to 10 ms.
> > 
> > The dying CPU will clear the bit which is safe in that context.
> > 
> > This fixes following RCU warning on ARMv8 (Exynos 4412, Trats2) during
> > suspend to RAM:
> 
> Nit: isn't Exynos4412 a quad-A9 (ARMv7 rather than ARMv8)?

Yes, it should be ARMv7. However still this should be fixed for both
architectures.

> 
> > [   31.113925] ===============================
> > [   31.113928] [ INFO: suspicious RCU usage. ]
> > [   31.113935] 3.19.0-rc7-next-20150203 #1914 Not tainted
> > [   31.113938] -------------------------------
> > [   31.113943] kernel/sched/fair.c:4740 suspicious rcu_dereference_check() usage!
> > [   31.113946]
> > [   31.113946] other info that might help us debug this:
> > [   31.113946]
> > [   31.113952]
> > [   31.113952] RCU used illegally from offline CPU!
> > [   31.113952] rcu_scheduler_active = 1, debug_locks = 0
> > [   31.113957] 3 locks held by swapper/1/0:
> > [   31.113988]  #0:  ((cpu_died).wait.lock){......}, at: [<c005a114>] complete+0x14/0x44
> > [   31.114012]  #1:  (&p->pi_lock){-.-.-.}, at: [<c004a790>] try_to_wake_up+0x28/0x300
> > [   31.114035]  #2:  (rcu_read_lock){......}, at: [<c004f1b8>] select_task_rq_fair+0x5c/0xa04
> > [   31.114038]
> > [   31.114038] stack backtrace:
> > [   31.114046] CPU: 1 PID: 0 Comm: swapper/1 Not tainted 3.19.0-rc7-next-20150203 #1914
> > [   31.114050] Hardware name: SAMSUNG EXYNOS (Flattened Device Tree)
> > [   31.114076] [<c0014ce4>] (unwind_backtrace) from [<c0011c30>] (show_stack+0x10/0x14)
> > [   31.114091] [<c0011c30>] (show_stack) from [<c04dc048>] (dump_stack+0x70/0xbc)
> > [   31.114105] [<c04dc048>] (dump_stack) from [<c004f83c>] (select_task_rq_fair+0x6e0/0xa04)
> > [   31.114118] [<c004f83c>] (select_task_rq_fair) from [<c004a83c>] (try_to_wake_up+0xd4/0x300)
> > [   31.114129] [<c004a83c>] (try_to_wake_up) from [<c00598a0>] (__wake_up_common+0x4c/0x80)
> > [   31.114140] [<c00598a0>] (__wake_up_common) from [<c00598e8>] (__wake_up_locked+0x14/0x1c)
> > [   31.114150] [<c00598e8>] (__wake_up_locked) from [<c005a134>] (complete+0x34/0x44)
> > [   31.114167] [<c005a134>] (complete) from [<c04d6ca4>] (cpu_die+0x24/0x84)
> > [   31.114179] [<c04d6ca4>] (cpu_die) from [<c005a508>] (cpu_startup_entry+0x328/0x358)
> > [   31.114189] [<c005a508>] (cpu_startup_entry) from [<40008784>] (0x40008784)
> > [   31.114226] CPU1: shutdown
> > 
> > Signed-off-by: Krzysztof Kozlowski <k.kozlowski at samsung.com>
> > 
> > ---
> > Changes since v1:
> > 1. Use adaptive sleep time when waiting for CPU die (idea and code
> >    from Paul E. McKenney). Paul also acked the patch but I made evem more
> >    changes.
> > 
> > 2. Add another bit (CPU_DIE_TIMEOUT_BIT) for synchronizing power down
> >    failure in case:
> >     CPU0 (killing)                           CPUx (killed)
> >      wait_for_cpu_die
> >      timeout
> >                                               cpu_die()
> >                                                clear_bit()
> >                                                self power down
> > 
> >    In this case the bit would be cleared and CPU would be powered down
> >    introducing wrong behavior in next power down sequence (CPU0 would
> >    see the bit cleared).
> >    I think that such race is still possible but was narrowed to very
> >    short time frame. Any CPU up will reset the bit to proper values.
> 
> In the case of shutting down 2 CPUs in quick succession (without an
> intervening boot of a CPU), surely this does not solve the potential
> race on the wait_cpu_die variable?

Right, the race is not fully fixed.

> 
> I think we instead need a percpu synchronisation variable, which would
> prevent racing on the value between CPUs, and a CPU would have to be
> brought up before we could decide to kill it again. With that I think we
> only need a single bit, too.

You mean a single bit-value per cpu?

Best regards,
Krzysztof
> 
> Thanks,
> Mark.