Re: [RFC PATCH] net: ethtool: do runtime PM outside RTNL

[Heiner Kallweit <hkallweit1@gmail.com>]

On 06.12.2023 10:37, Johannes Berg wrote:
> On Wed, 2023-12-06 at 09:46 +0100, Przemek Kitszel wrote:
>>
>> That sounds right too; one could argue if your fix is orthogonal to that
>> or not. I would say that your fix makes core net code more robust
>> against drivers from past millennia. :)
>> igc folks are notified, no idea how much time it would take to propose
>> a fix.
> 
> Maybe it should be on whoever added runtime pm to ethtool ;-)
> 
> Heiner, the igc driver was already doing this when you added
> pm_runtime_get_sync() ops, was there a discussion at the time, or just
> missed?
> 
I think it went unnoticed at that time that igc is acquiring RTNL
in runtime-resume. I'm just astonished that this pops up only now,
considering that the change was done more than 2 yrs ago.

Note: In __dev_open() there's a similar scenario where the
runtime-resume callback may be executed under RTNL.

> I really don't know any of this ...
> 
>>> Well, according to the checks, the patch really should use
>>> netdev_get_by_name() and netdev_put()? But I don't know how to do that
>>> on short-term stack thing ... maybe it doesn't have to?
>>
>> Nice to have such checks :)
>> You need some &netdevice_tracker, perhaps one added into struct net
>> or other place that would allow to track it at ethtool level.
> 
> Yeah but that's dynamic? Seems weird to add something to allocations for
> something released in the same function ...
> 
>> "short term stack thing" does not relieve us from good coding practices,
>> but perhaps "you just replaced __dev_get_by_name() call by
>> dev_get_by_name()" to fix a bug would ;) - with transition to tracked
>> alloc as a next series to be promised :)
> 
> All I want is to know how ;)
> but I guess I can try to find examples.
> 
>> anyway, I'm fresh here, and would love to know what others think about
> 
> Not me, but me too ;-)
> 
> johannes