[PATCH] net: amd: xgbe: fix possible sleep-in-atomic-context bugs in xgbe_powerdown()

[PATCH] net: amd: xgbe: fix possible sleep-in-atomic-context bugs in xgbe_powerdown()

[Jia-Ju Bai]

The driver may sleep while holding a spinlock.
The function call path (from bottom to top) in Linux 4.19 is:

drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1261: 
	flush_workqueue in xgbe_powerdown
drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1253: 
	_raw_spin_lock_irqsave in xgbe_powerdown

drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1055: 
	napi_disable in xgbe_napi_disable
drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1266: 
	xgbe_napi_disable in xgbe_powerdown
drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1253: 
	_raw_spin_lock_irqsave in xgbe_powerdown

drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1049: 
	napi_disable in xgbe_napi_disable
drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1266: 
	xgbe_napi_disable in xgbe_powerdown
drivers/net/ethernet/amd/xgbe/xgbe-drv.c, 1253: 
	_raw_spin_lock_irqsave in xgbe_powerdown

flush_workqueue() and napi_disable() can sleep at runtime.

To fix these bugs, flush_workqueue() and xgbe_napi_disable() are called
without holding the spinlock.

These bugs are found by a static analysis tool STCheck written by
myself.

Signed-off-by: Jia-Ju Bai <baijiaju1990@gmail.com>
---
 drivers/net/ethernet/amd/xgbe/xgbe-drv.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
index 98f8f2033154..328361d0e190 100644
--- a/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
+++ b/drivers/net/ethernet/amd/xgbe/xgbe-drv.c
@@ -1257,17 +1257,18 @@ int xgbe_powerdown(struct net_device *netdev, unsigned int caller)
 	netif_tx_stop_all_queues(netdev);
 
 	xgbe_stop_timers(pdata);
-	flush_workqueue(pdata->dev_workqueue);
 
 	hw_if->powerdown_tx(pdata);
 	hw_if->powerdown_rx(pdata);
 
-	xgbe_napi_disable(pdata, 0);
-
 	pdata->power_down = 1;
 
 	spin_unlock_irqrestore(&pdata->lock, flags);
 
+	flush_workqueue(pdata->dev_workqueue);
+
+	xgbe_napi_disable(pdata, 0);
+
 	DBGPR("<--xgbe_powerdown\n");
 
 	return 0;
-- 
2.17.1

Re: [PATCH] net: amd: xgbe: fix possible sleep-in-atomic-context bugs in xgbe_powerdown()

[Markus Elfring]

> The function call path (from bottom to top) in Linux 4.19 is:

Does this Linux version need more adjustments than the current
development version?


> These bugs are found by a static analysis tool STCheck written
> by myself.

Would you like to point any more background information out
for this software?

Regards,
Markus

Re: [PATCH] net: amd: xgbe: fix possible sleep-in-atomic-context bugs in xgbe_powerdown()

[David Miller]

From: Jia-Ju Bai <baijiaju1990@gmail.com>
Date: Wed, 18 Dec 2019 22:01:02 +0800

> @@ -1257,17 +1257,18 @@ int xgbe_powerdown(struct net_device *netdev, unsigned int caller)
>  	netif_tx_stop_all_queues(netdev);
>  
>  	xgbe_stop_timers(pdata);
> -	flush_workqueue(pdata->dev_workqueue);
>  
>  	hw_if->powerdown_tx(pdata);
>  	hw_if->powerdown_rx(pdata);
>  
> -	xgbe_napi_disable(pdata, 0);
> -
>  	pdata->power_down = 1;
>  
>  	spin_unlock_irqrestore(&pdata->lock, flags);
>  
> +	flush_workqueue(pdata->dev_workqueue);
> +
> +	xgbe_napi_disable(pdata, 0);
> +

Nope, this doesn't work at all.

You can't leave NAPI enabled, and thus packet processing, after the TX
and RX units of the chip have been powered down.

Re: [PATCH] net: amd: xgbe: fix possible sleep-in-atomic-context bugs in xgbe_powerdown()

[Jia-Ju Bai]

On 2019/12/19 5:26, David Miller wrote:
> From: Jia-Ju Bai <baijiaju1990@gmail.com>
> Date: Wed, 18 Dec 2019 22:01:02 +0800
>
>> @@ -1257,17 +1257,18 @@ int xgbe_powerdown(struct net_device *netdev, unsigned int caller)
>>   	netif_tx_stop_all_queues(netdev);
>>   
>>   	xgbe_stop_timers(pdata);
>> -	flush_workqueue(pdata->dev_workqueue);
>>   
>>   	hw_if->powerdown_tx(pdata);
>>   	hw_if->powerdown_rx(pdata);
>>   
>> -	xgbe_napi_disable(pdata, 0);
>> -
>>   	pdata->power_down = 1;
>>   
>>   	spin_unlock_irqrestore(&pdata->lock, flags);
>>   
>> +	flush_workqueue(pdata->dev_workqueue);
>> +
>> +	xgbe_napi_disable(pdata, 0);
>> +
> Nope, this doesn't work at all.
>
> You can't leave NAPI enabled, and thus packet processing, after the TX
> and RX units of the chip have been powered down.

Looking at the code, only xgbe_powerup() and xgbe_powerdown() use the 
spinlock "pdata->lock".
How about change the spinlock to a mutex?


Best wishes,
Jia-Ju Bai