Re: [PATCH] net: phy: Don't disable irqs on shutdown if WoL is enabled

[Vladimir Oltean]

Hi Uwe,

(I hope the threading won't be broken)

On Fri, Aug 04, 2023 at 09:17:57AM +0200, Uwe Kleine-König wrote:
> Most PHYs signal WoL using an interrupt. So disabling interrupts breaks
> WoL at least on PHYs covered by the marvell driver. So skip disabling
> irqs on shutdown if WoL is enabled.
> 
> While at it also explain the motivation that irqs are disabled at all.
> 
> Fixes: e2f016cf7751 ("net: phy: add a shutdown procedure")
> Signed-off-by: Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxxxx>
> ---
> Hello,
> 
> while I'm not sure that disabling interrupts is a good idea in general,
> this change at least should fix the WoL case. Note that this change is
> only compile tested as next doesn't boot on my test machine (because of
> https://git.kernel.org/linus/b3574f579ece24439c90e9a179742c61205fbcfa)
> and 6.1 (which is the other kernel I have running) doesn't know about
> .wol_enabled. I don't want to delay this fix until I bisected this new
> issue.
> 
> Assuming this patch is eligible for backporting to stable, maybe point
> out that it depends on v6.5-rc1~163^2~286^2~2 ("net: phy: Allow drivers
> to always call into ->suspend()"). Didn't try to backport that.
> 
> Best regards
> Uwe
> 
>  drivers/net/phy/phy_device.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/drivers/net/phy/phy_device.c b/drivers/net/phy/phy_device.c
> index 61921d4dbb13..6d1526bdd1d7 100644
> --- a/drivers/net/phy/phy_device.c
> +++ b/drivers/net/phy/phy_device.c
> @@ -3340,6 +3340,15 @@ static void phy_shutdown(struct device *dev)
>  	if (phydev->state == PHY_READY || !phydev->attached_dev)
>  		return;
>  
> +	/* Most phys signal WoL via the irq line. So for these irqs shouldn't be
> +	 * disabled.
> +	 */
> +	if (phydev->wol_enabled)
> +		return;
> +
> +	/* On shutdown disable irqs to prevent an irq storm on systems where the
> +	 * irq line is shared by several devices.
> +	 */
>  	phy_disable_interrupts(phydev);
>  }
>  
> -- 
> 2.40.1
> 
> 

I think the idea is not bad and something along these lines might be the
way to go, but I don't think it works (as currently implemented, and
tested by me, prints below).

Upon a quick search, phydev->wol_enabled is only set from phy_suspend(),
and phy_suspend() isn't invoked from the ethnl_set_wol() call stack.

Confirmed this way:

$ ethtool -s end0 wol g # &enet0 in the device tree
$ reboot -f
Rebooting.
[  288.682444] Qualcomm Atheros AR8031/AR8033 mdio@2d24000:02: phy_shutdown: wol_enabled 0
[  288.690935] Qualcomm Atheros AR8031/AR8033 mdio@2d24000:01: phy_shutdown: wol_enabled 0
[  288.736145] reboot: Restarting system

This is on the arch/arm/boot/dts/nxp/ls/ls1021a-tsn.dts, the same board
as the one which Ioana worked on (with the shared AR8031 PHY interrupts).

Sure, it needs to be mentioned that WoL + shared PHY interrupts is not
by any means an impossible combination, but it still won't work reliably.
I guess that's ok temporarily, since WoL requires user opt-in, so in the
default configuration, the LS1021A-TSN is not broken by this change (in
its functional variant).

pw-bot: cr