Re: [PATCH net] igb: fix deadlock caused by taking RTNL in RPM resume path

[Heiner Kallweit <hkallweit1@gmail.com>]

On 30.11.2021 18:12, Jakub Kicinski wrote:
> On Tue, 30 Nov 2021 07:46:22 +0100 Heiner Kallweit wrote:
>> On 30.11.2021 02:17, Jakub Kicinski wrote:
>>> On Mon, 29 Nov 2021 22:14:06 +0100 Heiner Kallweit wrote:  
>>>> -	rtnl_lock();
>>>> +	if (!rpm)
>>>> +		rtnl_lock();  
>>>
>>> Is there an ASSERT_RTNL() hidden in any of the below? Can we add one?
>>> Unless we're 100% confident nobody will RPM resume without rtnl held..
>>>   
>>
>> Not sure whether igb uses RPM the same way as r8169. There the device
>> is runtime-suspended (D3hot) w/o link. Once cable is plugged in the PHY
>> triggers a PME, and PCI core runtime-resumes the device (MAC).
>> In this case RTNL isn't held by the caller. Therefore I don't think
>> it's safe to assume that all callers hold RTNL.
> 
> No, no - I meant to leave the locking in but add ASSERT_RTNL() to catch
> if rpm == true && rtnl_held() == false.
> 
This is a valid case. Maybe it's not my day today, I still don't get
how we would benefit from adding an ASSERT_RTNL().

Based on the following I think that RPM resume and device open()
can't collide, because RPM resume is finished before open()
starts its actual work.

static int __igb_open(struct net_device *netdev, bool resuming)
{
...
if (!resuming)
		pm_runtime_get_sync(&pdev->dev);