Re: PATCH 1/5] scsi: megaraid_sas - Add Online Controller Reset to MegaRAID SAS drive

[Tomas Henzl <thenzl@redhat.com>]

On 10/11/2010 05:37 PM, Yang, Bo wrote:
> Tomas,
>
> The reason the driver does flush_scheduled_work() is driver did cancel_delayed_work().  I am not sure driver need to call flush_scheduled_work() if the scheduled work already done, but I will test this changes and submit another patch as soon as it works fine.  
>   
Driver needs to call flush_scheduled_work() because you are adding in this patch a new workqueue:
INIT_WORK(&instance->work_init, process_fw_state_change_wq); so calling the flush_scheduled_work()
only if a certain condition is met is the problem. If the condition comes true for both workqueues
then it would be fine, but I think that it isn't so.
With this patch we could have a started workqueue (with a long 30 sec sleep) and when the module
is removed then probably a memory corruption etc. 


> This change should not be the part of this patch, we would have the new patch submit after the verification. 
>   
I would prefer to have it in one patch, but if it is for you easier, then do it as you wish.

Tomas


> Regards,
>
> Bo Yang   
>  
> -----Original Message-----
> From: Tomas Henzl [mailto:thenzl@redhat.com] 
> Sent: Monday, October 11, 2010 9:21 AM
> To: Yang, Bo
> Cc: James Bottomley; bo yang; linux-scsi@vger.kernel.org; akpm@osdl.org; linux-kernel@vger.kernel.org
> Subject: Re: PATCH 1/5] scsi: megaraid_sas - Add Online Controller Reset to MegaRAID SAS drive
>
> On 10/11/2010 02:55 PM, Yang, Bo wrote:
>   
>> Tomas,
>>
>>   
>>     
>>> Another correction - flush_scheduled_work is already present in megass_detach_one
>>> it only should be moved away from the if statement.
>>>     
>>>       
>> The flush_scheduled_work is for schedule_delayed_work().  We need to flush it and remove.
>>   
>>     
> I'm not saying it should be removed, I think a move outside from the 'if' block
> makes it work for ev->hotplug_work and for the newly added instance->work_init
>
>
> diff --git a/drivers/scsi/megaraid/megaraid_sas.c b/drivers/scsi/megaraid/megaraid_sas.c
> index 55951f4..7773707 100644
> --- a/drivers/scsi/megaraid/megaraid_sas.c
> +++ b/drivers/scsi/megaraid/megaraid_sas.c
> @@ -4088,9 +4088,9 @@ static void __devexit megasas_detach_one(struct pci_dev *pdev)
>  		struct megasas_aen_event *ev = instance->ev;
>  		cancel_delayed_work(
>  			(struct delayed_work *)&ev->hotplug_work);
> -		flush_scheduled_work();
> -		instance->ev = NULL;
>  	}
> +	flush_scheduled_work();
> +	instance->ev = NULL;
>  
>  	tasklet_kill(&instance->isr_tasklet);
>  
> --
>
>   
>> -----Original Message-----
>> From: Tomas Henzl [mailto:thenzl@redhat.com] 
>> Sent: Saturday, October 09, 2010 4:38 PM
>> To: James Bottomley
>> Cc: bo yang; linux-scsi@vger.kernel.org; akpm@osdl.org; linux-kernel@vger.kernel.org; Yang, Bo
>> Subject: Re: PATCH 1/5] scsi: megaraid_sas - Add Online Controller Reset to MegaRAID SAS drive
>>
>> On 10/08/2010 09:28 PM, Tomas Henzl wrote:
>>   
>>     
>>> On 10/08/2010 06:39 PM, James Bottomley wrote:
>>>   
>>>     
>>>       
>>>> On Fri, 2010-10-08 at 17:51 +0200, Tomas Henzl wrote:
>>>>   
>>>>     
>>>>       
>>>>         
>>>>> On 09/23/2010 04:36 AM, bo yang wrote:
>>>>>     
>>>>>       
>>>>>         
>>>>>           
>>>>>> This patch is too big.  I am using attachment to submit.  Please
>>>>>> use attached file to apply.  Also let me know if it can't be accepted.
>>>>>>
>>>>>> To add the Online controller reset support, driver need to do:
>>>>>> a). reset the controller chips -- Xscale and Gen2 which will change
>>>>>> the function calls and add the reset function related to this two
>>>>>> chips.
>>>>>> b). during the reset, driver will store the pending cmds which not
>>>>>> returned by FW to driver's pending queue.  Driver will re-issue those
>>>>>> pending cmds again to FW after the OCR finished.
>>>>>> c). In driver's timeout routine, driver will report to OS as reset.
>>>>>> Also driver's queue routine will block the cmds until the OCR
>>>>>> finished.
>>>>>> d). in Driver's ISR routine, if driver get the FW state as state
>>>>>> change, FW in Failure status and FW support online controller
>>>>>> reset (OCR), driver will start to do the controller reset.
>>>>>> e). In driver's IOCTL routine, the application cmds will wait for the
>>>>>> OCR to finish, then issue the cmds to FW.
>>>>>>
>>>>>> Signed-off-by Bo Yang<bo.yang@lsi.com>
>>>>>>
>>>>>> ---
>>>>>>  drivers/scsi/megaraid/megaraid_sas.c |  756 ++++++++++++++++++++++++++++++++---
>>>>>>  drivers/scsi/megaraid/megaraid_sas.h |   88 +++-
>>>>>>  2 files changed, 787 insertions(+), 57 deletions(-)
>>>>>>       
>>>>>>         
>>>>>>           
>>>>>>             
>>>>> Hi Bo,
>>>>> in the workqueue function you sleep for 30s,
>>>>> it's scheduled here - schedule_work(&instance->work_init);
>>>>>
>>>>> +process_fw_state_change_wq(struct work_struct *work)
>>>>> +{
>>>>> ...
>>>>> +		/*waitting for about 20 second before start the second init*/
>>>>> +		for (wait = 0; wait < 30; wait++) {
>>>>> +			msleep(1000);
>>>>> +		}
>>>>>     
>>>>>       
>>>>>         
>>>>>           
>>>> this lot should be ssleep(20) if you want a 20 sec sleep.
>>>>   
>>>>     
>>>>       
>>>>         
>>> please do that on every place where you use the 
>>> "for (wait = 0; wait < n; wait++) msleep(1000);" construction
>>>
>>>   
>>>     
>>>       
>>>>> - this is not a good practice to sleep for a so long time I think
>>>>>     
>>>>>       
>>>>>         
>>>>>           
>>> this long sleep might might be ok, if the workqueue is used only rarely
>>> is it so?
>>>
>>>   
>>>     
>>>       
>>>>> - you should use in your exit function some synchronization 
>>>>>   for example 'cancel_work_sync', without that if someone rmmods your 
>>>>>   module, it could then lead to a memory corruption
>>>>>     
>>>>>       
>>>>>         
>>>>>           
>>>> Actually flush_scheduled_work() should be fine ... it will force the
>>>> module removal to wait for completion ... cancellation can be error
>>>> prone, so just forcing the wait sounds easier.
>>>>   
>>>>     
>>>>       
>>>>         
>> Another correction - flush_scheduled_work is already present in megass_detach_one
>> it only should be moved away from the if statement.
>>
>>
>>   
>>     
>>> someone told that cancel_work_sync is safer then flush_scheduled_work
>>> but I'm not an expert, so ok 
>>>
>>> Tomas
>>>
>>>   
>>>     
>>>       
>>>> James
>>>>
>>>>
>>>>