Re: ast: cursor flashing softlockups

From: Peter Hurley
Date: Tue May 17 2016 - 10:29:15 EST


[ +to Scot Doyle ]

Scot, please take a look at this soft lockup.

Regards,
Peter Hurley


Hi Ming,

On 05/17/2016 02:12 AM, Ming Lei wrote:
> Hi,
>
> On Tue, May 17, 2016 at 4:07 AM, Dann Frazier
> <dann.frazier@xxxxxxxxxxxxx> wrote:
>> Hi,
>> I'm observing a soft lockup issue w/ the ASPEED controller on an
>> arm64 server platform. This was originally seen on Ubuntu's 4.4
>> kernel, but it is reproducible w/ vanilla 4.6-rc7 as well.
>>
>> [ 32.792656] NMI watchdog: BUG: soft lockup - CPU#38 stuck for 22s!
>> [swapper/38:0]
>>
>> I observe this just once each time I boot into debian-installer (I'm
>> using a serial console, but the ast module gets loaded during
>> startup).
>
> I have figured out that it is caused by 'mod_timer(timer, jiffies)' and
> 'ops->cur_blink_jiffies' is observed as zero in cursor_timer_handler()
> when the issue happened.

Thanks for tracking this down.

This softlockup looks to be caused by:

commit 27a4c827c34ac4256a190cc9d24607f953c1c459
Author: Scot Doyle <lkml14@xxxxxxxxxxxxx>
Date: Thu Mar 26 13:56:38 2015 +0000

fbcon: use the cursor blink interval provided by vt

vt now provides a cursor blink interval via vc_data. Use this
interval instead of the currently hardcoded 200 msecs. Store it in
fbcon_ops to avoid locking the console in cursor_timer_handler().

Signed-off-by: Scot Doyle <lkml14@xxxxxxxxxxxxx>
Acked-by: Pavel Machek <pavel@xxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>

and

commit bd63364caa8df38bad2b25b11b2a1b849475cce5
Author: Scot Doyle <lkml14@xxxxxxxxxxxxx>
Date: Thu Mar 26 13:54:39 2015 +0000

vt: add cursor blink interval escape sequence

Add an escape sequence to specify the current console's cursor blink
interval. The interval is specified as a number of milliseconds until
the next cursor display state toggle, from 50 to 65535. /proc/loadavg
did not show a difference with a one msec interval, but the lower
bound is set to 50 msecs since slower hardware wasn't tested.

Store the interval in the vc_data structure for later access by fbcon,
initializing the value to fbcon's current hardcoded value of 200 msecs.

Signed-off-by: Scot Doyle <lkml14@xxxxxxxxxxxxx>
Acked-by: Pavel Machek <pavel@xxxxxx>
Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>



> Looks it is a real fbcon/vt issue, see following:
>
> fbcon_init()
> <-.con_init
> <-visual_init()
>
> reset_terminal()
> <-vc_init()
>
> vc->vc_cur_blink_ms is just set in reset_terminal() from vc_init() path,
> and ops->cur_blink_jiffies is figured out from vc->vc_cur_blink_ms
> in fbcon_init().
>
> And visual_init() is always run before vc_init(), so ops->cur_blink_jiffies
> is initialized as zero and cause the soft lockup issue finally.
>
> Thanks,
> Ming
>
>>
>> perf shows that the CPU caught by the NMI is typically in code
>> updating the cursor timer:
>>
>> - 16.92% swapper [kernel.kallsyms] [k] _raw_spin_unlock_irqrestore
>> - _raw_spin_unlock_irqrestore
>> + 16.87% mod_timer
>> + 0.05% cursor_timer_handler
>> - 12.15% swapper [kernel.kallsyms] [k] queue_work_on
>> - queue_work_on
>> + 12.00% cursor_timer_handler
>> + 0.15% call_timer_fn
>> + 10.98% swapper [kernel.kallsyms] [k] run_timer_softirq
>> - 2.23% swapper [kernel.kallsyms] [k] mod_timer
>> - mod_timer
>> + 1.97% cursor_timer_handler
>> + 0.26% call_timer_fn
>>
>> During the same period, I can see that another CPU is actively
>> executing the timer function:
>>
>> - 42.18% kworker/u96:2 [kernel.kallsyms] [k] ww_mutex_unlock
>> - ww_mutex_unlock
>> - 40.70% ast_dirty_update
>> ast_imageblit
>> soft_cursor
>> bit_cursor
>> fb_flashcursor
>> process_one_work
>> worker_thread
>> kthread
>> ret_from_fork
>> + 1.48% ast_imageblit
>> - 40.15% kworker/u96:2 [kernel.kallsyms] [k] __memcpy_toio
>> - __memcpy_toio
>> + 31.54% ast_dirty_update
>> + 8.61% ast_imageblit
>>
>> Using the graph function tracer on fb_flashcursor(), I see that
>> ast_dirty_update usually takes around 60 us, in which it makes 16
>> calls to __memcpy_toio(). However, there is always one instance on
>> every boot of the installer where ast_dirty_update() takes ~98 *ms* to
>> complete, during which it makes 743 calls to __memcpy_toio(). While
>> that doesn't directly account for the full 22s, I do wonder if that
>> maybe a smoking gun.
>>
>> fyi, this is being tracked at: https://bugs.launchpad.net/bugs/1574814
>>
>> -dann