net: calxeda: fix race condition in xgmac_remove due to unfinshed work

[PATCH] net: calxeda: fix race condition in xgmac_remove due to unfinshed work

Posted by Zheng Wang 2 years, 6 months ago

In xgmac_probe, the priv->tx_timeout_work is bound with 
xgmac_tx_timeout_work. In xgmac_remove, if there is an 
unfinished work, there might be a race condition that 
priv->base was written byte after iounmap it.

Fix it by finishing the work before cleanup.

Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
---
 drivers/net/ethernet/calxeda/xgmac.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/net/ethernet/calxeda/xgmac.c b/drivers/net/ethernet/calxeda/xgmac.c
index f4f87dfa9687..94c3804001e3 100644
--- a/drivers/net/ethernet/calxeda/xgmac.c
+++ b/drivers/net/ethernet/calxeda/xgmac.c
@@ -1831,6 +1831,7 @@ static int xgmac_remove(struct platform_device *pdev)
 	/* Free the IRQ lines */
 	free_irq(ndev->irq, ndev);
 	free_irq(priv->pmt_irq, ndev);
+	cancel_work_sync(&priv->tx_timeout_work);
 
 	unregister_netdev(ndev);
 	netif_napi_del(&priv->napi);
-- 
2.25.1

Re: [PATCH] net: calxeda: fix race condition in xgmac_remove due to unfinshed work

Posted by Yunsheng Lin 2 years, 6 months ago

On 2023/3/9 11:56, Zheng Wang wrote:
> In xgmac_probe, the priv->tx_timeout_work is bound with 
> xgmac_tx_timeout_work. In xgmac_remove, if there is an 
> unfinished work, there might be a race condition that 
> priv->base was written byte after iounmap it.
> 
> Fix it by finishing the work before cleanup.

This should go to net branch, so title should be:

 [PATCH net] net: calxeda: fix race condition in xgmac_remove due to unfinshed work

From history commit, it seems more common to use "net: calxedaxgmac" instead of
"net: calxeda", I am not sure which one is better.

Also there should be a Fixes tag for net branch, maybe:

Fixes: 8746f671ef04 ("net: calxedaxgmac: fix race between xgmac_tx_complete and xgmac_tx_err")


> 
> Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
> ---
>  drivers/net/ethernet/calxeda/xgmac.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/drivers/net/ethernet/calxeda/xgmac.c b/drivers/net/ethernet/calxeda/xgmac.c
> index f4f87dfa9687..94c3804001e3 100644
> --- a/drivers/net/ethernet/calxeda/xgmac.c
> +++ b/drivers/net/ethernet/calxeda/xgmac.c
> @@ -1831,6 +1831,7 @@ static int xgmac_remove(struct platform_device *pdev)
>  	/* Free the IRQ lines */
>  	free_irq(ndev->irq, ndev);
>  	free_irq(priv->pmt_irq, ndev);
> +	cancel_work_sync(&priv->tx_timeout_work);

It seems the blow function need to stop the dev_watchdog() from
calling dev->netdev_ops->ndo_tx_timeout before calling
cancel_work_sync(&priv->tx_timeout_work), otherwise the
dev_watchdog() may trigger the priv->tx_timeout_work to run again.

	netif_carrier_off(ndev);
	netif_tx_disable(ndev);

>  
>  	unregister_netdev(ndev);
>  	netif_napi_del(&priv->napi);
>

Re: [PATCH] net: calxeda: fix race condition in xgmac_remove due to unfinshed work

Posted by Zheng Hacker 2 years, 6 months ago

Yunsheng Lin <linyunsheng@huawei.com> 于2023年3月9日周四 14:24写道：
>
> On 2023/3/9 11:56, Zheng Wang wrote:
> > In xgmac_probe, the priv->tx_timeout_work is bound with
> > xgmac_tx_timeout_work. In xgmac_remove, if there is an
> > unfinished work, there might be a race condition that
> > priv->base was written byte after iounmap it.
> >
> > Fix it by finishing the work before cleanup.
>
> This should go to net branch, so title should be:
>
>  [PATCH net] net: calxeda: fix race condition in xgmac_remove due to unfinshed work
>

Sorry for the confusion.

> From history commit, it seems more common to use "net: calxedaxgmac" instead of
> "net: calxeda", I am not sure which one is better.
>
> Also there should be a Fixes tag for net branch, maybe:
>
> Fixes: 8746f671ef04 ("net: calxedaxgmac: fix race between xgmac_tx_complete and xgmac_tx_err")
>
>

Yes, I was eager to report the fix and ignored that. Thanks for
pointing that out.

> >
> > Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
> > ---
> >  drivers/net/ethernet/calxeda/xgmac.c | 1 +
> >  1 file changed, 1 insertion(+)
> >
> > diff --git a/drivers/net/ethernet/calxeda/xgmac.c b/drivers/net/ethernet/calxeda/xgmac.c
> > index f4f87dfa9687..94c3804001e3 100644
> > --- a/drivers/net/ethernet/calxeda/xgmac.c
> > +++ b/drivers/net/ethernet/calxeda/xgmac.c
> > @@ -1831,6 +1831,7 @@ static int xgmac_remove(struct platform_device *pdev)
> >       /* Free the IRQ lines */
> >       free_irq(ndev->irq, ndev);
> >       free_irq(priv->pmt_irq, ndev);
> > +     cancel_work_sync(&priv->tx_timeout_work);
>
> It seems the blow function need to stop the dev_watchdog() from
> calling dev->netdev_ops->ndo_tx_timeout before calling
> cancel_work_sync(&priv->tx_timeout_work), otherwise the
> dev_watchdog() may trigger the priv->tx_timeout_work to run again.
>
>         netif_carrier_off(ndev);
>         netif_tx_disable(ndev);
>
> >
> >       unregister_netdev(ndev);
> >       netif_napi_del(&priv->napi);
> >

Yes, I agree with that. Thanks for your advice. I learned a lot from it.

Best regards,
Zheng

Re: [PATCH] net: calxeda: fix race condition in xgmac_remove due to unfinshed work

Posted by Yunsheng Lin 2 years, 6 months ago

On 2023/3/9 14:23, Yunsheng Lin wrote:
> On 2023/3/9 11:56, Zheng Wang wrote:
>> In xgmac_probe, the priv->tx_timeout_work is bound with 
>> xgmac_tx_timeout_work. In xgmac_remove, if there is an 
>> unfinished work, there might be a race condition that 
>> priv->base was written byte after iounmap it.
>>
>> Fix it by finishing the work before cleanup.
> 
> This should go to net branch, so title should be:
> 
>  [PATCH net] net: calxeda: fix race condition in xgmac_remove due to unfinshed work

typo error:
unfinshed -> unfinished

> 
>>From history commit, it seems more common to use "net: calxedaxgmac" instead of
> "net: calxeda", I am not sure which one is better.
> 
> Also there should be a Fixes tag for net branch, maybe:
> 
> Fixes: 8746f671ef04 ("net: calxedaxgmac: fix race between xgmac_tx_complete and xgmac_tx_err")
> 
> 
>>
>> Signed-off-by: Zheng Wang <zyytlz.wz@163.com>
>> ---
>>  drivers/net/ethernet/calxeda/xgmac.c | 1 +
>>  1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/net/ethernet/calxeda/xgmac.c b/drivers/net/ethernet/calxeda/xgmac.c
>> index f4f87dfa9687..94c3804001e3 100644
>> --- a/drivers/net/ethernet/calxeda/xgmac.c
>> +++ b/drivers/net/ethernet/calxeda/xgmac.c
>> @@ -1831,6 +1831,7 @@ static int xgmac_remove(struct platform_device *pdev)
>>  	/* Free the IRQ lines */
>>  	free_irq(ndev->irq, ndev);
>>  	free_irq(priv->pmt_irq, ndev);
>> +	cancel_work_sync(&priv->tx_timeout_work);
> 
> It seems the blow function need to stop the dev_watchdog() from
> calling dev->netdev_ops->ndo_tx_timeout before calling
> cancel_work_sync(&priv->tx_timeout_work), otherwise the
> dev_watchdog() may trigger the priv->tx_timeout_work to run again.
> 
> 	netif_carrier_off(ndev);
> 	netif_tx_disable(ndev);
> 
>>  
>>  	unregister_netdev(ndev);
>>  	netif_napi_del(&priv->napi);
>>
> .
>

Re: [PATCH] net: calxeda: fix race condition in xgmac_remove due to unfinshed work

Posted by Zheng Hacker 2 years, 6 months ago

Yunsheng Lin <linyunsheng@huawei.com> 于2023年3月9日周四 14:27写道：
>
> On 2023/3/9 14:23, Yunsheng Lin wrote:
> > On 2023/3/9 11:56, Zheng Wang wrote:
> >> In xgmac_probe, the priv->tx_timeout_work is bound with
> >> xgmac_tx_timeout_work. In xgmac_remove, if there is an
> >> unfinished work, there might be a race condition that
> >> priv->base was written byte after iounmap it.
> >>
> >> Fix it by finishing the work before cleanup.
> >
> > This should go to net branch, so title should be:
> >
> >  [PATCH net] net: calxeda: fix race condition in xgmac_remove due to unfinshed work
>
> typo error:
> unfinshed -> unfinished
>

Got it. Will correct it in the next version of patch.

Thanks,
Zheng