[RFC] ath10k: change to do napi_enable and napi_disable when insmod and rmmod for sdio

[Ben Greear]

On 8/20/20 10:00 AM, Krishna Chaitanya wrote:
> On Thu, Aug 20, 2020 at 10:02 PM Ben Greear <greearb at candelatech.com> wrote:
>>
>> On 8/20/20 9:08 AM, Krishna Chaitanya wrote:
>>> On Thu, Aug 20, 2020 at 8:07 PM Wen Gong <wgong at codeaurora.org> wrote:
>>>>
>>>> On 2020-08-20 18:52, Krishna Chaitanya wrote:
>>>>> On Thu, Aug 20, 2020 at 3:45 PM Wen Gong <wgong at codeaurora.org> wrote:
>>>>>>
>>>>>> On 2020-08-20 17:19, Krishna Chaitanya wrote:
>>>> ...
>>>>>>>> I'm not really convinced that this is the right fix, but I'm no NAPI
>>>>>>>> expert. Can anyone else help?
>>>>>>> Calling napi_disable() twice can lead to hangs, but moving NAPI from
>>>>>>> start/stop to
>>>>>>> the probe isn't the right approach as the datapath is tied to
>>>>>>> start/stop.
>>>>>>>
>>>>>>> Maybe check the state of NAPI before disable?
>>>>>>>
>>>>>>>    if (test_bit(NAPI_STATE_SCHED, &ar->napi.napi.state))
>>>>>>>     napi_disable(&ar->napi)
>>>>>>> or maintain napi_state like this
>>>>>>> https://patchwork.kernel.org/patch/10249365/
>>>>>> it is better to use above link's patch.
>>>>>> napi.state is controlled by napi API, it is better ath10k not know it.
>>>>> Sure, but IMHO just canceling the async rx work should solve the issue.
>>>> Oh no, canceling the async rx work will not solve this issue, rx worker
>>>> ath10k_rx_indication_async_work call napi_schedule, after napi_complete,
>>>> the NAPI_STATE_SCHED will clear.
>>>> The issue of this patch is because 2 thread called to hif_stop and
>>>> NAPI_STATE_SCHED not clear.
>>> That fix is still valid and good to have.
>>>
>>> ndev_stop being called twice is typical scenarios (stop vs rmmod), so
>>>    just checking the netdev_flags for IFF_UP and returning from hif_Stop
>>> should suffice, no?
>>
>> My approach to fix this problem was to add a boolean in ath10k as to whether
>> it had napi enabled or not, and then check that before trying to enable/disable
>> it again.  Seems to work fine, and cleaner in my mind than checking internal
>> napi flags.
> A much simpler approach is just to check for IFF_UP and skip NAPI (and others)
> in the hif_stop no? (provided proper RTNL locking is done if hif_stop
> is being called
> internally as well).
> 

I'm not sure, but I think the driver should be internally consistent and not
spend a lot of time trying to guess about interactions with objects higher
in the stack.

Here is my original patch to fix this, it is not complex.

https://patchwork.kernel.org/patch/10249363/

Thanks,
Ben

-- 
Ben Greear <greearb at candelatech.com>
Candela Technologies Inc  http://www.candelatech.com