Messages in this thread | | | Subject | Re: [RFC PATCH] net: phy: Don't miss phy_suspend() on PHY_HALTED for PHYs with interrupts | From | Florian Fainelli <> | Date | Mon, 20 Mar 2017 09:41:52 -0700 |
| |
On 03/16/2017 12:46 AM, Roger Quadros wrote: > On 15/03/17 17:49, Andrew Lunn wrote: >> On Wed, Mar 15, 2017 at 05:00:08PM +0200, Roger Quadros wrote: >>> Andrew, >>> >>> On 15/03/17 16:08, Andrew Lunn wrote: >>>> On Wed, Mar 15, 2017 at 03:51:27PM +0200, Roger Quadros wrote: >>>>> Since commit 3c293f4e08b5 ("net: phy: Trigger state machine on state change and not polling.") >>>>> phy_suspend() doesn't get called as part of phy_stop() for PHYs using >>>>> interrupts because the phy state machine is never triggered after a phy_stop(). >>>>> >>>>> Explicitly trigger the PHY state machine so that it can >>>>> see the new PHY state (HALTED) and suspend the PHY. >>>>> >>>>> Signed-off-by: Roger Quadros <rogerq@ti.com> >>>> >>>> Hi Roger >>>> >>>> This seems sensible. It mirrors what phy_start() does. >>>> >>>> Reviewed-by: Andrew Lunn <andrew@lunn.ch> >>> >>> The reason for this being an RFC was the following comment just before >>> where I add the phy_trigger_machine() >>> >>> /* Cannot call flush_scheduled_work() here as desired because >>> * of rtnl_lock(), but PHY_HALTED shall guarantee phy_change() >>> * will not reenable interrupts. >>> */ >>> >>> Is this comment still applicable? If yes, is it OK to call >>> phy_trigger_machine() there? >> >> Humm, good question. >> >> My _guess_ would be, calling it with sync=True could >> deadlock. sync=False is probably safe. But lets see what Florian says. > > I agree that we should use phy_trigger_machine() with sync=False. > >> >>> >>>> >>>> It does however lead to a follow up question. Are there other places >>>> phydev->state is changed and it is missing a phy_trigger_machine()? >>>> >>> >>> One other place I think we should add phy_trigger_machine() is phy_start_aneg(). >> >> Humm, that might get us into a tight loop. >> >> phy_start_aneg() kicks the phy driver to start autoneg and sets >> phydev->state = PHY_AN. >> >> phy_trigger_machine() triggers the state machine immediately. >> >> In state PHY_AN, we check if aneg is done. If not, it sets needs_aneg >> = true. At the end of the state machine, this then calls >> phy_start_aneg(), and it all starts again. >> >> We are missing the 1s delay we have with polling. So for >> phy_start_aneg(), we might need a phy_delayed_trigger_machine(), which >> waits a second before doing anything? > > I think that should do the trick. > > How about this?
This sounds like a possible fix indeed, however I would like to better assess the impact on non interrupt driven PHYs before rolling such a change.
> > diff --git a/drivers/net/phy/phy.c b/drivers/net/phy/phy.c > index 8fef03b..162061c 100644 > --- a/drivers/net/phy/phy.c > +++ b/drivers/net/phy/phy.c > @@ -630,6 +630,10 @@ int phy_start_aneg(struct phy_device *phydev) > > out_unlock: > mutex_unlock(&phydev->lock); > + > + if (!err) > + queue_delayed_work(system_power_efficient_wq, &phydev->state_queue, HZ); > + > return err; > } > EXPORT_SYMBOL(phy_start_aneg); >
-- Florian
| |