Re: [PATCH v10 54/55] Input: atmel_mxt_ts: Implement synchronization during various operation

[Dmitry Osipenko]

02.04.2020 16:24, Dmitry Osipenko ÐÐÑÐÑ:
> 02.04.2020 14:50, Wang, Jiada ÐÐÑÐÑ:
>> Hi Dmitry
>>
>> On 2020/04/02 1:04, Dmitry Osipenko wrote:
>>> 31.03.2020 13:50, Jiada Wang ÐÐÑÐÑ:
>>>> From: Sanjeev Chugh <sanjeev_chugh@xxxxxxxxxx>
>>>>
>>>> There could be scope of race conditions when sysfs is being handled
>>>> and at the same time, device removal is occurring. For example,
>>>> we don't want the device removal to begin if the Atmel device
>>>> cfg update is going on or firmware update is going on. In such
>>>> cases, wait for device update to be completed before the removal
>>>> continues.
>>>>
>>>> ÂÂÂÂ ThreadÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ Thread 2:
>>>> =========================ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ
>>>> =========================
>>>> mxt_update_fw_store()ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ mxt_remove()
>>>> mutex_lock(&data->lock)ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ ...
>>>> mxt_initialize()ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ //Tries to acquire lock
>>>> ÂÂ request_firmware_nowait()ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ mutex_lock(&data->lock)
>>>> ...ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ ==>waits for lock()
>>>> ...ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ .
>>>> ...ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ .
>>>> mutex_unlock(&data->lock)ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ .
>>>> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ //Gets lock and
>>>> proceeds
>>>> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ
>>>> mxt_free_input_device();
>>>> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ ...
>>>> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ
>>>> mutex_unlock(&data->lock)
>>>> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ //Frees atmel driver
>>>> data
>>>> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ kfree(data)
>>>>
>>>> If the request_firmware_nowait() completes after the driver removal,
>>>> and callback is triggered. But kernel crashes since the module is
>>>> already removed.
>>>>
>>>> This commit adds state machine to serialize such scenarios.
>>>
>>> Won't it be easier to bump driver's module use-count by __module_get()
>>> while firmware is updating? Or remove sysfs during of mxt_remove()? >
>>
>> thanks for your inspiration, I will replace state machine with module
>> use-count.
> 
> I'm actually now thinking that the suggestion about the module-count
> wasn't very correct because this won't really help in regards to
> mxt_update_fw_store() / mxt_remove() racing.
> 
> I see that mxt_remove() already invokes the mxt_sysfs_remove(), which
> should block until mxt_update_fw_store() is completed, shouldn't it?
> 
> I guess the kfree(data) isn't the real cause of the problem and
> something like this should help:
> 
> diff --git a/drivers/input/touchscreen/atmel_mxt_ts.c
> b/drivers/input/touchscreen/atmel_mxt_ts.c
> index b2edf51e1595..4e66106feeb9 100644
> --- a/drivers/input/touchscreen/atmel_mxt_ts.c
> +++ b/drivers/input/touchscreen/atmel_mxt_ts.c
> @@ -4254,6 +4254,7 @@ static void mxt_sysfs_remove(struct mxt_data *data)
>  	struct i2c_client *client = data->client;
> 
>  	sysfs_remove_group(&client->dev.kobj, &mxt_attr_group);
> +	sysfs_remove_group(&client->dev.kobj, &mxt_fw_attr_group);
>  }
> 
>  static void mxt_reset_slots(struct mxt_data *data)
> @@ -4649,31 +4650,19 @@ static int mxt_remove(struct i2c_client *client)
>  {
>  	struct mxt_data *data = i2c_get_clientdata(client);
> 
> -	mutex_lock(&data->lock);
> -	if (data->e_state == MXT_STATE_UPDATING_CONFIG_ASYNC ||
> -	    data->e_state == MXT_STATE_UPDATING_CONFIG) {
> -		data->e_state = MXT_STATE_GOING_AWAY;
> -		mutex_unlock(&data->lock);
> -		mxt_wait_for_completion(data, &data->update_cfg_completion,
> -					MXT_CONFIG_TIMEOUT);
> -	} else {
> -		data->e_state = MXT_STATE_GOING_AWAY;
> -		mutex_unlock(&data->lock);
> -	}
> +	mxt_sysfs_remove(data);
> 
> -	disable_irq(data->irq);
> -	sysfs_remove_group(&client->dev.kobj, &mxt_fw_attr_group);
>  	if (data->reset_gpio) {
>  		sysfs_remove_link(&client->dev.kobj, "reset");
>  		gpiod_unexport(data->reset_gpio);
>  	}
> +
>  	mxt_debug_msg_remove(data);
> -	mxt_sysfs_remove(data);
>  	mxt_free_input_device(data);
>  	mxt_free_object_table(data);
> 
> 	if (debug_state)
> 		cancel_delayed_work_sync(&data->watchdog_work);
> +	disable_irq(data->irq);
> 
>  	return 0;
>  }
> 

I'm looking at this again and the original tear-down order of the
mxt_remove() looks okay, so no need to change it.

Reading the commit message, it says that request_firmware_nowait() races
with kfree(data), but that can't happen because the data is
resource-managed and request_firmware_nowait() bumps device's use-count.

https://elixir.bootlin.com/linux/v5.6.2/source/drivers/base/firmware_loader/main.c#L1043

I think this patch was ported from some very old kernel version and it's
simply not applicable to the upstream anymore.