[PATCH RFC net-next 03/14] net: phylink: add support for PCS link change notifications

Sean Anderson sean.anderson at seco.com
Tue Jan 23 13:09:52 PST 2024


On 1/23/24 16:05, Russell King (Oracle) wrote:
> On Tue, Jan 23, 2024 at 03:33:57PM -0500, Sean Anderson wrote:
>> On 1/23/24 15:07, Russell King (Oracle) wrote:
>>> On Tue, Jan 23, 2024 at 02:46:15PM -0500, Sean Anderson wrote:
>>>> Hi Russell,
>>>> 
>>>> Does there need to be any locking when calling
>>>> phylink_pcs_change? I noticed that you call it from threaded
>>>> IRQ context in [1]. Can that race with phylink_major_config?
>>> 
>>> What kind of scenario are you thinking may require locking?
>> 
>> Can't we at least get a spurious bounce? E.g.
>> 
>> pcs_major_config() pcs_disable(old_pcs) /* masks IRQ */ 
>> old_pcs->phylink = NULL; new_pcs->phylink = pl; ... 
>> pcs_enable(new_pcs) /* unmasks IRQ */ ...
>> 
>> pcs_handle_irq(new_pcs) /* Link up IRQ */ 
>> phylink_pcs_change(new_pcs, true) phylink_run_resolve(pl)
>> 
>> phylink_resolve(pl) /* Link up */
> 
> By this time, old_pcs->phylink has been set to NULL as you mentioned 
> above.
> 
>> pcs_handle_irq(old_pcs) /* Link down IRQ (pending from before
>> pcs_disable) */ phylink_pcs_change(old_pcs, false) 
>> phylink_run_resolve(pl) /* Doesn't see the NULL */
> 
> So here, phylink_pcs_change(old_pcs, ...) will read old_pcs->phylink
> and find that it's NULL, and do nothing.

This can happen on another CPU. There are no memory barriers on the read
side (until queue_work), so there's no guarantee that other CPUs will
see the write.

--Sean

>>> I guess the possibility would be if pcs->phylink changes and the 
>>> compiler reads it multiple times - READ_ONCE() should solve
>>> that.
>>> 
>>> However, in terms of the mechanics, there's no race.
>>> 
>>> During the initial bringup, the resolve worker isn't started
>>> until after phylink_major_config() has completed (it's started
>>> at phylink_enable_and_run_resolve().) So, if
>>> phylink_pcs_change() gets called while in phylink_major_config()
>>> there, it'll see that pl->phylink_disable_state is non-zero, and
>>> won't queue the work.
>>> 
>>> The next one is within the worker itself - and there can only be
>>> one instance of the worker running in totality. So, if 
>>> phylink_pcs_change() gets called while phylink_major_config() is 
>>> running from this path, the only thing it'll do is re-schedule 
>>> the resolve worker to run another iteration which is harmless 
>>> (whether or not the PCS is still current.)
>>> 
>>> The last case is phylink_ethtool_ksettings_set(). This runs
>>> under the state_mutex, which locks out the resolve worker (since
>>> it also takes that mutex).
>>> 
>>> So calling phylink_pcs_change() should be pretty harmless
>>> _unless_ the compiler re-reads pcs->phylink multiple times
>>> inside phylink_pcs_change(), which I suppose with modern
>>> compilers is possible. Hence my suggestion above about
>>> READ_ONCE() for that.
>>> 
>>> Have you encountered an OOPS because pcs->phylink has become
>>> NULL? Or have you spotted another issue?
>> 
>> I was looking at extending this code, and I was wondering if I
>> needed to e.g. take RTNL first. Thanks for the quick response.
> 
> Note that phylink_mac_change() gets called in irq context, so this 
> stuff can't take any mutexes or the rtnl. It is also intended that 
> phylink_pcs_change() is similarly callable in irq context.
> 




More information about the linux-arm-kernel mailing list