Re: Odroid XU4 deadlock

From: Javier Martinez Canillas
Date: Mon Apr 11 2016 - 11:34:08 EST


Hello Krzysztof,

On 04/09/2016 07:00 AM, Krzysztof Kozlowski wrote:
> On Fri, Apr 8, 2016 at 5:36 PM, Anand Moon <linux.amoon@xxxxxxxxx> wrote:
>> hi Javier,
>>
>> On 8 April 2016 at 20:25, Javier Martinez Canillas
>> <javier@xxxxxxxxxxxxxxx> wrote:
>>> Hello Anand,
>>>
>>> On 04/08/2016 09:51 AM, Anand Moon wrote:
>>>> Hi All,
>>>>
>>>> I am observing this deadlock or warning on my Odroid XU4.
>>>> I thinks the is related to clk from exynos5_i2c_xfer .... clk_prepare_enable
>>>> but I am not able to figure out which clk.
>>>>
>>>
>>> By reading your logs, it seems the problem is a possible ABBA deadlock since
>>> the S3C RTC driver grabs the prepare_lock and then the regmap->lock, and the
>>> CPUFreq driver leads to the locks being grabbed in the inverse order.
>
> First of all this is not a deadlock. Yet. :) This is just lockdep

Yes, that's why I said the problem seemed to be a possible deadlock :)

> report of possible deadlock. This is deadlock:
> http://www.krzk.eu/builders/boot-odroid-xu3-multi_v7/builds/13/steps/Boot%20odroid/logs/serial
>
> ...and I believe it is the same as the report from lockdep. Which
> means that this report truly a predicted deadlock. There are minor
> differences between deadlock on my board and report from Anand but
> still the same devices are involved. In Anands report the involved
> S2MPS11 regulator and S2MPS11 clock use the same regmap. In my case
> one path comes from S5M RTC which uses different regmap than S2MPS11
> clock. This is the difference which puzzles me.
>

Yeah, that is puzzling indeed...

> Anyway this is known issue. In case of my board mentioned deadlock
> happens only on multi_v7 on around 15% of boots. I noticed it around
> v4.4 and it happens still.
>
>>>
>>>> [ 11.640221] ======================================================
>>>> [ 11.646343] [ INFO: possible circular locking dependency detected ]
>>>> [ 11.652590] 4.6.0-rc2-xu4ml #35 Not tainted
>>>> [ 11.656749] -------------------------------------------------------
>>>> [ 11.662992] kworker/1:1/85 is trying to acquire lock:
>>>> [ 11.668010] (prepare_lock){+.+...}, at: [<c05bfed8>]
>>>> clk_prepare_lock+0x50/0xf8
>>>> [ 11.675375]
>>>> [ 11.675375] but task is already holding lock:
>>>> [ 11.681185] (sec_core:428:(regmap)->lock){+.+...}, at:
>>>> [<c0481aec>] regmap_read+0x2c/0x5c
>>>> [ 11.689417]
>>>> [ 11.689417] which lock already depends on the new lock.
>>>> [ 11.689417]
>>>> [ 11.697561]
>>>> [ 11.697561] the existing dependency chain (in reverse order) is:
>>>> [ 11.705008]
>>>> [ 11.705008] -> #1 (sec_core:428:(regmap)->lock){+.+...}:
>>>> [ 11.710467] [<c0482d30>] regmap_update_bits_base+0x30/0x70
>>>> [ 11.716545] [<c05c617c>] s2mps11_clk_prepare+0x30/0x38
>>>> [ 11.722262] [<c05c01f0>] clk_core_prepare+0x98/0xbc
>>>> [ 11.727711] [<c05c12bc>] clk_prepare+0x1c/0x2c
>>>> [ 11.732734] [<c0556fe4>] s3c_rtc_probe+0x2d0/0x43c
>>>
>>> the S3C RTC driver tries to prepare the RTC source clock S2MPS11_CLK_AP
>>> and that leads to the prepare_lock to be held and then the regmap->lock
>>> since s2mps11_clk_prepare() calls regmap_update_bits().
>>>
>>>> [ 11.738108] [<c046b9fc>] platform_drv_probe+0x4c/0xb0
>>>> [ 11.743749] [<c046a1fc>] driver_probe_device+0x20c/0x2b8
>>>> [ 11.749641] [<c0468594>] bus_for_each_drv+0x60/0x94
>>>> [ 11.755103] [<c0469f10>] __device_attach+0xb4/0x118
>>>> [ 11.760558] [<c04693b4>] bus_probe_device+0x88/0x90
>>>> [ 11.766020] [<c0469850>] deferred_probe_work_func+0x6c/0x9c
>>>> [ 11.772169] [<c013b52c>] process_one_work+0x1a8/0x514
>>>> [ 11.777806] [<c013b8d0>] worker_thread+0x38/0x56c
>>>> [ 11.783082] [<c0141d14>] kthread+0xf4/0x10c
>>>> [ 11.787847] [<c0107950>] ret_from_fork+0x14/0x24
>>>> [ 11.793050]
>>>> [ 11.793050] -> #0 (prepare_lock){+.+...}:
>>>> [ 11.797212] [<c0768dd8>] mutex_lock_nested+0x78/0x4dc
>>>> [ 11.802845] [<c05bfed8>] clk_prepare_lock+0x50/0xf8
>>>> [ 11.808304] [<c05c19c4>] clk_unprepare+0x1c/0x2c
>>>> [ 11.813499] [<c055d720>] exynos5_i2c_xfer+0x1dc/0x304
>>>
>>> and here the locks are grabbed in the inverse order, since the regulator
>>> driver uses regmap read (grabbing the regmap->lock) and then a clock is
>>> prepared in exynos5_i2c_xfer.
>>>
>>>> [ 11.819129] [<c0559620>] __i2c_transfer+0x13c/0x278
>>>> [ 11.824589] [<c05597f0>] i2c_transfer+0x94/0xc4
>>>> [ 11.829701] [<c04868f0>] regmap_i2c_read+0x48/0x64
>>>> [ 11.835074] [<c0482048>] _regmap_raw_read+0xb4/0x114
>>>> [ 11.840630] [<c04820cc>] _regmap_bus_read+0x24/0x58
>>>> [ 11.846084] [<c0481a68>] _regmap_read+0x60/0xb8
>>>> [ 11.851201] [<c0481afc>] regmap_read+0x3c/0x5c
>>>> [ 11.856224] [<c03e2354>] regulator_get_voltage_sel_regmap+0x20/0x54
>>>> [ 11.863109] [<c03dc5d4>] _regulator_get_voltage+0x20/0xb8
>>>> [ 11.863141] [<c03df058>] _regulator_do_set_voltage+0x240/0x370
>>>> [ 11.863173] [<c03df254>] regulator_set_voltage_unlocked+0xcc/0x230
>>>> [ 11.863205] [<c03df3e0>] regulator_set_voltage+0x28/0x54
>>>> [ 11.863234] [<c0478a94>] _set_opp_voltage+0x30/0x98
>>>> [ 11.863264] [<c0479ba8>] dev_pm_opp_set_rate+0x1e0/0x540
>>>> [ 11.863337] [<c0580824>] __cpufreq_driver_target+0x168/0x290
>>>> [ 11.863375] [<c0583b48>] od_dbs_timer+0xdc/0x164
>>>> [ 11.863407] [<c0584114>] dbs_work_handler+0x30/0x58
>>>> [ 11.863437] [<c013b52c>] process_one_work+0x1a8/0x514
>>>> [ 11.863465] [<c013b8d0>] worker_thread+0x38/0x56c
>>>> [ 11.863496] [<c0141d14>] kthread+0xf4/0x10c
>>>> [ 11.863528] [<c0107950>] ret_from_fork+0x14/0x24
>>>> [ 11.863538]
>>>> [ 11.863538] other info that might help us debug this:
>>>> [ 11.863538]
>>>> [ 11.863548] Possible unsafe locking scenario:
>>>> [ 11.863548]
>>>> [ 11.863557] CPU0 CPU1
>>>> [ 11.863565] ---- ----
>>>> [ 11.863589] lock(sec_core:428:(regmap)->lock);
>>>> [ 11.863612] lock(prepare_lock);
>>>> [ 11.863634] lock(sec_core:428:(regmap)->lock);
>>>> [ 11.863655] lock(prepare_lock);
>>>> [ 11.863664]
>>>> [ 11.863664] *** DEADLOCK ***
>>>> [ 11.863664]
>>>
>>> I should take a look in deep to the regmap and clock subsystems to better
>>> understand the issue on how to solve it. But AFAICT the problem is that
>>> the prepare_lock is a global mutex and the same regmap is used for both
>>> the s2mps11 clocks and regulators, so the ABBA deadlock described below
>>> can happen.
>
> Exactly. One of ideas I remember was to make clocks mutex more
> granular but I did not see any patches doing that. Another interesting
> point is (in my case) deadlock itself (not lockdep report) happens
> only on multi_v7, not on exynos defconfig.
>

I see that RTC_DRV_S5M is enabled as a module in multi_v7 but built-in
in exynos_defconfig. Maybe that could cause the behavior difference?

> BTW, Anand, if you report please provide more details for
> reproduction: kernel config and reproduction steps.
>

I see that Anand provided his defconfig, so I'll try to reproduce the
lockdep report and look at the issue deeper this week.

> Best regards,
> Krzysztof
>

Best regards,
--
Javier Martinez Canillas
Open Source Group
Samsung Research America