Re: [PATCH net v3 2/2] net: ethernet: ti: am65-cpsw: avoid devm_alloc_etherdev, fix module removal

[Roger Quadros]

On 04/10/2024 18:37, Nicolas Pitre wrote:
> On Fri, 4 Oct 2024, Roger Quadros wrote:
> 
>> Hi Nicolas,
>>
>> On 04/10/2024 07:10, Nicolas Pitre wrote:
>>> From: Nicolas Pitre <npitre@xxxxxxxxxxxx>
>>>
>>> Usage of devm_alloc_etherdev_mqs() conflicts with
>>> am65_cpsw_nuss_cleanup_ndev() as the same struct net_device instances
>>> get unregistered twice. Switch to alloc_etherdev_mqs() and make sure
>>
>> Do we know why the same net device gets unregistered twice?
> 
> When using devm_alloc_etherdev_mqs() every successful allocation is put 
> in a resource list tied to the device. When the driver is removed, 
> there's a net device unregister from am65_cpsw_nuss_cleanup_ndev() and 
> another one from devm_free_netdev().

I couldn't find out where devm_free_netdev() calls unregister_netdev().
Also we didn't use devm_register_netdev() so resource manager will not
call unregister_netdev().

> 
> We established in patch #1 that net devices must be unregistered before 
> devlink_port_unregister() is invoked, meaning we can't rely on the 
> implicit devm_free_netdev() as it happens too late, hence the explicit 
> am65_cpsw_nuss_cleanup_ndev().
> 
>>> am65_cpsw_nuss_cleanup_ndev() unregisters and frees those net_device
>>> instances properly.
>>>
>>> With this, it is finally possible to rmmod the driver without oopsing
>>> the kernel.
>>>
>>> Fixes: 93a76530316a ("net: ethernet: ti: introduce am65x/j721e gigabit eth subsystem driver")
>>> Signed-off-by: Nicolas Pitre <npitre@xxxxxxxxxxxx>
>>> ---
>>>  drivers/net/ethernet/ti/am65-cpsw-nuss.c | 20 ++++++++++++--------
>>>  1 file changed, 12 insertions(+), 8 deletions(-)
>>>
>>> diff --git a/drivers/net/ethernet/ti/am65-cpsw-nuss.c b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
>>> index f6bc8a4dc6..e95457c988 100644
>>> --- a/drivers/net/ethernet/ti/am65-cpsw-nuss.c
>>> +++ b/drivers/net/ethernet/ti/am65-cpsw-nuss.c
>>> @@ -2744,10 +2744,9 @@ am65_cpsw_nuss_init_port_ndev(struct am65_cpsw_common *common, u32 port_idx)
>>>  		return 0;
>>>  
>>>  	/* alloc netdev */
>>> -	port->ndev = devm_alloc_etherdev_mqs(common->dev,
>>> -					     sizeof(struct am65_cpsw_ndev_priv),
>>> -					     AM65_CPSW_MAX_QUEUES,
>>> -					     AM65_CPSW_MAX_QUEUES);
>>> +	port->ndev = alloc_etherdev_mqs(sizeof(struct am65_cpsw_ndev_priv),
>>> +					AM65_CPSW_MAX_QUEUES,
>>> +					AM65_CPSW_MAX_QUEUES);
>>
>> Can we solve this issue without doing this change as
>> there are many error cases relying on devm managed freeing of netdev.
> 
> If you know of a way to do this differently I'm all ears.

I sent another approach already. please check.
https://lore.kernel.org/all/67c9ede4-9751-4255-b752-27dd60495ff3@xxxxxxxxxx/

> 
> About the many error cases needing the freeing of net devices, as far as 
> I know they're all covered with this patch.

No they are not. you now have to explicitly call free_netdev() in error paths of am65_cpsw_nuss_init_port_ndev().
I see 3 places directly returning error code.
i.e.
        default:
                dev_err(dev, "selected phy-mode is not supported\n");
                return -EOPNOTSUPP;
        }
...
        if (IS_ERR(phylink))
                return PTR_ERR(phylink);
...
        ndev_priv->stats = netdev_alloc_pcpu_stats(struct am65_cpsw_ndev_stats);
        if (!ndev_priv->stats)
                return -ENOMEM;

> 
>> I still can't see what we are doing wrong in existing code.
> 
> Did you try to rmmod this driver lately?

Yes and it throws an oops, so we do need a fix.
> 
> 
> Nicolas

-- 
cheers,
-roger