Re: [question] About triggering a region snapshot through the devlink cmd

Re: [question] About triggering a region snapshot through the devlink cmd

[Keller, Jacob E]

On 2019/9/27 16:13, Jiri Pirko wrote:
> Fri, Sep 27, 2019 at 09:40:47AM CEST, linyunsheng@huawei.com wrote:
>> Hi, Jiri & Alex
>>
>>    It seems that a region' snapshot is only created through the
>> driver when some error is detected, for example:
>> mlx4_crdump_collect_fw_health() -> devlink_region_snapshot_create()
>>
>>    We want to trigger a region' snapshot creation through devlink
>> cmd, maybe by adding the "devlink region triger", because we want
>> to check some hardware register/state when the driver or the hardware
>> does not detect the error sometimes.
>>
>> Does about "devlink region triger" make sense?
>>
>> If yes, is there plan to implement it? or any suggestion to implement
>> it?
> 
> Actually, the plan is co convert mlx4 to "devlink health" api. Mlx5
> already uses that.
> 
> You should look into "devlink health".

Hi Jiri,

I agree, the mlx4 firmware dump stuff should use the devlink health API, as it's for error conditions.

I've recently been investigating using devlink regions to expose non-error regions where it really does make sense to "trigger" a snapshot.

The original region commit message indicates that there were plans to support requesting snapshots, and even potentially writing to the region directly.

I wanted to use a region to expose some set of flash/NVM contents. Being able to request a snapshot is essential in this case. I saw that netdevsim does this through a debugfs hook, but that really doesn't make sense to me for a production driver.

It makes sense to me to be able to trigger a region snapshot directly through devlink. Should I spin up some patches to implement it?

Thanks,
Jake

Re: FW: [question] About triggering a region snapshot through the devlink cmd

[Jacob Keller]

On 1/8/2020 4:15 AM, Alex Vesker wrote:
> I am a biased here but, devlink trigger can be useful... I am not aware
> of other alternatives,
> devlink health has it`s benefits but it is not devlink region. If you
> will decide to implement I can
> review the design, if Jiri is ok with the idea.
> 

Sure. I am not quite sure how long it will be till patches are on the
list, as I'm currently in the process of implementing devlink support
for one of the Intel drivers, and would be using that driver as an example.

Actually, come to think of it, I may just implement the region trigger
and use netdevsim as the example. That should enable those patches to
hit the list sooner than the patches for implementing devlink for the
ice driver.

I will Cc you for review when I send those patches.

Thanks,
Jake

Re: [question] About triggering a region snapshot through the devlink cmd

[Jakub Kicinski]

On Wed, 8 Jan 2020 11:14:38 -0800, Jacob Keller wrote:
> On 1/8/2020 4:15 AM, Alex Vesker wrote:
> > I am a biased here but, devlink trigger can be useful... I am not aware
> > of other alternatives,
> > devlink health has it`s benefits but it is not devlink region. If you
> > will decide to implement I can
> > review the design, if Jiri is ok with the idea.
> >   
> 
> Sure. I am not quite sure how long it will be till patches are on the
> list, as I'm currently in the process of implementing devlink support
> for one of the Intel drivers, and would be using that driver as an example.
> 
> Actually, come to think of it, I may just implement the region trigger
> and use netdevsim as the example. That should enable those patches to
> hit the list sooner than the patches for implementing devlink for the
> ice driver.

Just to be clear - you mean implement triggering the dump over netlink?
FWIW that seems fairly reasonable to me, but please do explain the use
case clearly in the patches.

netdevsim is a driver mock up, and debugfs is its control interface.
The trigger there is to simulate an async device error, this trigger
should not be used as example or justification for anything real driver
may need.

Re: [question] About triggering a region snapshot through the devlink cmd

[Jacob Keller]

On 1/9/2020 11:41 AM, Jakub Kicinski wrote:
> On Wed, 8 Jan 2020 11:14:38 -0800, Jacob Keller wrote:
>> On 1/8/2020 4:15 AM, Alex Vesker wrote:
>>> I am a biased here but, devlink trigger can be useful... I am not aware
>>> of other alternatives,
>>> devlink health has it`s benefits but it is not devlink region. If you
>>> will decide to implement I can
>>> review the design, if Jiri is ok with the idea.
>>>   
>>
>> Sure. I am not quite sure how long it will be till patches are on the
>> list, as I'm currently in the process of implementing devlink support
>> for one of the Intel drivers, and would be using that driver as an example.
>>
>> Actually, come to think of it, I may just implement the region trigger
>> and use netdevsim as the example. That should enable those patches to
>> hit the list sooner than the patches for implementing devlink for the
>> ice driver.
> 
> Just to be clear - you mean implement triggering the dump over netlink?
> FWIW that seems fairly reasonable to me, but please do explain the use
> case clearly in the patches.
> 

Yes. See the patches I just sent and let me know if the explanation
isn't totally clear.

> netdevsim is a driver mock up, and debugfs is its control interface.
> The trigger there is to simulate an async device error, this trigger
> should not be used as example or justification for anything real driver
> may need.
> 

The only two drivers that currently use regions at all are the mlx4 and
the netdevsim driver. In order to have some example of how the new
.trigger_snapshot and DEVLINK_CMD_REGION_TRIGGER work, (and so I could
test them out easily) I chose to implement it in netdevsim.

I'm working on patches for the ice driver that will implement devlink
support, but that is still in progress and isn't really yet at a point
where I can send it to the list. I wanted to get feedback on the
implementation of the triggering earlier.

Thanks,
Jake