On 22/03/22 5:50 pm, Sondhauß, Jan wrote:
> Hi
>
> On 22/03/2022 11:34, Vignesh Raghavendra wrote:
>> Hi, Adding netdev list and maintainers Please cc netdev ML and net
>> maintainers ./scripts/get_maintainer.pl -f
>> drivers/net/ethernet/ti/cpsw_ethtool.c On 22/03/22 12:02 pm, Sondhauß,
>> Jan wrote: > cpsw_ethtool uses the power management in the
>> ZjQcmQRYFpfptBannerStart
>> This Message Is From an External Sender
>> Please use caution when clicking on links or opening attachments!
>> ZjQcmQRYFpfptBannerEnd
>>
>> Hi,
>>
>> Adding netdev list and maintainers
>>
>> Please cc netdev ML and net maintainers
>>
>> ./scripts/get_maintainer.pl -f drivers/net/ethernet/ti/cpsw_ethtool.c
>>
>> On 22/03/22 12:02 pm, Sondhauß, Jan wrote:
>>> cpsw_ethtool uses the power management in the begin and complete
>>> functions of the ethtool_ops. The result of pm_runtime_get_sync was
>>> returned unconditionally, which results in problems since the ethtool-
>>> interface relies on 0 for success and negativ values for errors.
>>> d43c65b05b84 (ethtool: runtime-resume netdev parent in ethnl_ops_begin)
>>> introduced power management to the netlink implementation for the
>>> ethtool interface and does not explicitly check for negative return
>>> values.
>>>
>>> As a result the pm_runtime_suspend function is called one-too-many
>>> times in ethnl_ops_begin and that leads to an access violation when
>>> the cpsw hardware is accessed after using
>>> 'ethtool -C eth-of-cpsw rx-usecs 1234'. To fix this the call to
>>> pm_runtime_get_sync in cpsw_ethtool_op_begin is replaced with a call
>>> to pm_runtime_resume_and_get as it provides a returnable error-code.
>>>
>>
>> pm_runtime_resume_and_get() is just wrapper around pm_runtime_get_sync()
>> + error handling (as done in the below code) and both return 0 on
>> success and -ve error code on failure
>
> pm_runtime_get_sync returns -ve error code on failure and 0 on success
> and also 1 is returned if nothing has to be done besides increment of
> the usage counter.
> So for active devices that don't need to be resumed a 1 is returned.
> pm_runtime_resume_and_get is a return-friendly wrapper that returns
> -error code on failure but returns 0 on both other cases.
>
I think this is a better explanation than the original commit message,
but see below
>>
>>
>>> Signed-off-by: Jan Sondhauss <jan.sondhauss@xxxxxxxx>
>>> ---
>>> drivers/net/ethernet/ti/cpsw_ethtool.c | 2 +-
>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/net/ethernet/ti/cpsw_ethtool.c b/drivers/net/ethernet/ti/cpsw_ethtool.c
>>> index 158c8d3793f4..5eda20039cc1 100644
>>> --- a/drivers/net/ethernet/ti/cpsw_ethtool.c
>>> +++ b/drivers/net/ethernet/ti/cpsw_ethtool.c
>>> @@ -364,7 +364,7 @@ int cpsw_ethtool_op_begin(struct net_device *ndev)
>>> struct cpsw_common *cpsw = priv->cpsw;
>>> int ret;
>>>
>>> - ret = pm_runtime_get_sync(cpsw->dev);
>>> + ret = pm_runtime_resume_and_get(cpsw->dev)> if (ret < 0) {
>>> cpsw_err(priv, drv, "ethtool begin failed %d\n", ret);
>>> pm_runtime_put_noidle(cpsw->dev);
>>
>>
>> In fact code now ends up calling pm_runtime_put_noidle() twice in case
>> of failure, once inside pm_runtime_resume_and_get() and again here?
>>
>> So something looks fishy?
>
> Sort of. There is no actual failure but pm_runtime_put is still called
> twice. That is due to
> 1. cpsw_ethtool_op_begin returning 1 when it should return 0
> 2. ethnl_ops_begin treating values not equal to 0 as failure
> 3. coalesce_prepare_data only treating negative values as failure
>
> The patch addresses 1.
>
> In net/ethtool/netlink.c:33 ethnl_ops_begin() the cpsw_ethtool_op_begin
> is called (returning 1) and in the error path of ethnl_ops_begin a
> pm_runtime_put is called. The function calling ethnl_ops_begin only
> checks for negative values: net/ethtool/coalesce.c:60
> coalesce_prepare_data and continues the sucess path calling
> ethnl_ops_complete. ethnl_ops_complete also calls pm_runtime_put. So the
> success path of coalesce_prepare_data and the error path of
> ethnl_ops_begin both end up calling pm_runtime_put when only one of them
> should.
>
Thanks for the explanation!
Sorry, But what about the error case (ie ret < 0) With this patch, don't
we end up calling pm_runtime_put_noidle() twice (once inside
pm_runtime_resume_and_get() and again in cpsw_ethtool_op_begin()). How
is that okay?
Regards
Vignesh
