[PATCH] tick: fix tick_broadcast_pending_mask not cleared

[Joseph Lo]

On Tue, 2013-06-18 at 00:15 +0800, Daniel Lezcano wrote:
> The recent modification in the cpuidle framework consolidated the timer
> broadcast code across the different drivers by setting a new flag in the idle
> state. It tells the cpuidle core code to enter/exit to the broadcast mode for
> the cpu when entering a deep idle state. The broadcast timer enter/exit is no
> longer handled by the back-end driver.
> 
> This change made the local interrupt to be enabled *before* calling
> CLOCK_EVENT_NOTIFY_EXIT. bad or not (see below) ?
> 
> On a tegra114, a four cores system, when the flag has been introduced in the
> driver, the following warning appeared:
> 
> [   25.629559] WARNING: CPU: 2 PID: 0 at
> kernel/time/tick-broadcast.c:578 tick_broadcast_oneshot_control
> +0x1a4/0x1d0()
> [   25.629565] Modules linked in:
> [   25.629574] CPU: 2 PID: 0 Comm: swapper/2 Not tainted
> 3.10.0-rc3-next-20130529+ #15
> [   25.629601] [<c00148f4>] (unwind_backtrace+0x0/0xf8) from
> [<c0011040>] (show_stack+0x10/0x14)
> [   25.629616] [<c0011040>] (show_stack+0x10/0x14) from [<c0504248>]
> (dump_stack+0x80/0xc4)
> [   25.629633] [<c0504248>] (dump_stack+0x80/0xc4) from [<c00231ac>]
> (warn_slowpath_common+0x64/0x88)
> [   25.629646] [<c00231ac>] (warn_slowpath_common+0x64/0x88) from
> [<c00231ec>] (warn_slowpath_null+0x1c/0x24)
> [   25.629657] [<c00231ec>] (warn_slowpath_null+0x1c/0x24) from
> [<c00667f8>] (tick_broadcast_oneshot_control+0x1a4/0x1d0)
> [   25.629670] [<c00667f8>] (tick_broadcast_oneshot_control+0x1a4/0x1d0)
> from [<c0065cd0>] (tick_notify+0x240/0x40c)
> [   25.629685] [<c0065cd0>] (tick_notify+0x240/0x40c) from [<c0044724>]
> (notifier_call_chain+0x44/0x84)
> [   25.629699] [<c0044724>] (notifier_call_chain+0x44/0x84) from
> [<c0044828>] (raw_notifier_call_chain+0x18/0x20)
> [   25.629712] [<c0044828>] (raw_notifier_call_chain+0x18/0x20) from
> [<c00650cc>] (clockevents_notify+0x28/0x170)
> [   25.629726] [<c00650cc>] (clockevents_notify+0x28/0x170) from
> [<c033f1f0>] (cpuidle_idle_call+0x11c/0x168)
> [   25.629739] [<c033f1f0>] (cpuidle_idle_call+0x11c/0x168) from
> [<c000ea94>] (arch_cpu_idle+0x8/0x38)
> [   25.629755] [<c000ea94>] (arch_cpu_idle+0x8/0x38) from [<c005ea80>]
> (cpu_startup_entry+0x60/0x134)
> [   25.629767] [<c005ea80>] (cpu_startup_entry+0x60/0x134) from
> [<804fe9a4>] (0x804fe9a4)
> [   25.629772] ---[ end trace 5484e77e2531bccc ]---
> 
> I don't have the hardware, so I wasn't able to reproduce the warning but after
> looking a while in the code, I deduced the following:
> 
>  1. the CPU2 enters a deep idle state and sets the broadcast timer
>  2. the timer expires, the tick_handle_oneshot_broadcast function is called,
>     setting the tick_broadcast_pending_mask and waking up the idle cpu CPU2
>  3. the CPU2 exits idle and invokes tick_broadcast_oneshot_control with
>     CLOCK_EVENT_NOTIFY_EXIT with the following code:
>     [...]
>     if (dev->next_event.tv64 == KTIME_MAX)
>             goto out;
> 
>     if (cpumask_test_and_clear_cpu(cpu,
>                                  tick_broadcast_pending_mask))
>             goto out;
>     [...]
> 
>  4. if there is no next event planned for CPU2, we fulfil the first condition and
>     we jump to the 'out' section without clearing the tick_broadcast_pending_mask
> 
>  5. CPU2 goes to deep idle again and calls tick_broadcast_oneshot_control with
>     CLOCK_NOTIFY_EVENT_ENTER but with the tick_broadcast_pending_mask set for
>     CPU2, leading to the WARNING.
> 
> Above, it is mentionned the change moved the CLOCK_EVENT_NOTIFY_EXIT after the
> local interrupt were enabled. If it is before, this warning does not occur. My
> hypothesis is the code path described before does not happen because when a
> broadcast timer expires, dev->next_event.tv64 is always different from KTIME_MAX
> because the timer handler did not set the value yet (local interrupt are still
> disabled).
> 
> I don't see anywhere in the code, a clockevents_notify(ENTER/EXIT) block must be
> done with the local interrupt disabled in between, furthermore the function uses
> 'raw_spin_lock_irqsave' which make me think, we don't care about that.
> 
> Invert the conditions and make the tick broadcast code immune from the local
> interrupts context.
> 
> Signed-off-by: Daniel Lezcano <daniel.lezcano at linaro.org>
> Reported-by: Joseph Lo <josephl at nvidia.com>
> ---

Daniel,

Thanks, This patch fixes the problem indeed. I can't reproduce it
anymore. Verified on Tegra30 and Tegra114 platform. Looks good to me.

Tested-by: Joseph Lo <josephl at nvidia.com>

Joseph