| Message ID | 5964fa47-2eff-4968-894c-0b7f487d820c@gmail.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | net: phy: improve phylib EEE handling | expand |
On Sat, Jan 11, 2025 at 10:06:02AM +0100, Heiner Kallweit wrote: > Link modes in phydev->eee_disabled_modes are filtered out by > genphy_c45_write_eee_adv() and won't be advertised. Therefore > don't accept such modes from userspace. Why do we need this? Surely if the MAC doesn't support modes, then they should be filtered out of phydev->supported_eee so that userspace knows that the mode is not supported by the network interface as a whole, just like we do for phydev->supported. That would give us the checking here.
On 11.01.2025 10:21, Russell King (Oracle) wrote: > On Sat, Jan 11, 2025 at 10:06:02AM +0100, Heiner Kallweit wrote: >> Link modes in phydev->eee_disabled_modes are filtered out by >> genphy_c45_write_eee_adv() and won't be advertised. Therefore >> don't accept such modes from userspace. > > Why do we need this? Surely if the MAC doesn't support modes, then they > should be filtered out of phydev->supported_eee so that userspace knows > that the mode is not supported by the network interface as a whole, just > like we do for phydev->supported. > > That would give us the checking here. > Removing EEE modes to be disabled from supported_eee is problematic because of how genphy_c45_write_eee_adv() works. Let's say we have a 2.5Gbps PHY and want to disable EEE at 2.5Gbps. If we remove 2.5Gbps from supported_eee, then the following check is false: if (linkmode_intersects(phydev->supported_eee, PHY_EEE_CAP2_FEATURES)) What would result in the 2.5Gbps mode not getting disabled.
On Sat, Jan 11, 2025 at 10:44:25AM +0100, Heiner Kallweit wrote: > On 11.01.2025 10:21, Russell King (Oracle) wrote: > > On Sat, Jan 11, 2025 at 10:06:02AM +0100, Heiner Kallweit wrote: > >> Link modes in phydev->eee_disabled_modes are filtered out by > >> genphy_c45_write_eee_adv() and won't be advertised. Therefore > >> don't accept such modes from userspace. > > > > Why do we need this? Surely if the MAC doesn't support modes, then they > > should be filtered out of phydev->supported_eee so that userspace knows > > that the mode is not supported by the network interface as a whole, just > > like we do for phydev->supported. > > > > That would give us the checking here. > > > Removing EEE modes to be disabled from supported_eee is problematic > because of how genphy_c45_write_eee_adv() works. > > Let's say we have a 2.5Gbps PHY and want to disable EEE at 2.5Gbps. If we > remove 2.5Gbps from supported_eee, then the following check is false: > if (linkmode_intersects(phydev->supported_eee, PHY_EEE_CAP2_FEATURES)) > What would result in the 2.5Gbps mode not getting disabled. Ok. Do we at least remove the broken modes from the supported mask reported to userspace?
On 11.01.2025 11:01, Russell King (Oracle) wrote: > On Sat, Jan 11, 2025 at 10:44:25AM +0100, Heiner Kallweit wrote: >> On 11.01.2025 10:21, Russell King (Oracle) wrote: >>> On Sat, Jan 11, 2025 at 10:06:02AM +0100, Heiner Kallweit wrote: >>>> Link modes in phydev->eee_disabled_modes are filtered out by >>>> genphy_c45_write_eee_adv() and won't be advertised. Therefore >>>> don't accept such modes from userspace. >>> >>> Why do we need this? Surely if the MAC doesn't support modes, then they >>> should be filtered out of phydev->supported_eee so that userspace knows >>> that the mode is not supported by the network interface as a whole, just >>> like we do for phydev->supported. >>> >>> That would give us the checking here. >>> >> Removing EEE modes to be disabled from supported_eee is problematic >> because of how genphy_c45_write_eee_adv() works. >> >> Let's say we have a 2.5Gbps PHY and want to disable EEE at 2.5Gbps. If we >> remove 2.5Gbps from supported_eee, then the following check is false: >> if (linkmode_intersects(phydev->supported_eee, PHY_EEE_CAP2_FEATURES)) >> What would result in the 2.5Gbps mode not getting disabled. > > Ok. Do we at least remove the broken modes from the supported mask > reported to userspace? > I think that's something we could do in addition, to provide a hint to the user about unavailable modes. It wouldn't remove the need for the check here. ethtool doesn't check the advertisement against the supported modes. And even if it would, we must not rely on input from user space being sane.
On Sat, Jan 11, 2025 at 02:19:04PM +0100, Heiner Kallweit wrote: > On 11.01.2025 11:01, Russell King (Oracle) wrote: > > On Sat, Jan 11, 2025 at 10:44:25AM +0100, Heiner Kallweit wrote: > >> On 11.01.2025 10:21, Russell King (Oracle) wrote: > >>> On Sat, Jan 11, 2025 at 10:06:02AM +0100, Heiner Kallweit wrote: > >>>> Link modes in phydev->eee_disabled_modes are filtered out by > >>>> genphy_c45_write_eee_adv() and won't be advertised. Therefore > >>>> don't accept such modes from userspace. > >>> > >>> Why do we need this? Surely if the MAC doesn't support modes, then they > >>> should be filtered out of phydev->supported_eee so that userspace knows > >>> that the mode is not supported by the network interface as a whole, just > >>> like we do for phydev->supported. > >>> > >>> That would give us the checking here. > >>> > >> Removing EEE modes to be disabled from supported_eee is problematic > >> because of how genphy_c45_write_eee_adv() works. > >> > >> Let's say we have a 2.5Gbps PHY and want to disable EEE at 2.5Gbps. If we > >> remove 2.5Gbps from supported_eee, then the following check is false: > >> if (linkmode_intersects(phydev->supported_eee, PHY_EEE_CAP2_FEATURES)) > >> What would result in the 2.5Gbps mode not getting disabled. > > > > Ok. Do we at least remove the broken modes from the supported mask > > reported to userspace? > > > I think that's something we could do in addition, to provide a hint to the > user about unavailable modes. It wouldn't remove the need for the check here. > ethtool doesn't check the advertisement against the supported modes. > And even if it would, we must not rely on input from user space being sane. I disagree with some of this. Userspace should expect: - read current settings - copy supported modes to advertised modes - write current settings to work. If it fails, then how does ethtool, or even the user, work out which link modes are actually supported or not. If we're introducing a failure on the "disabled" modes, then that is a user-breaking change, and we need to avoid that. The current code silently ignored the broken modes, your new code would error out on the above action - and that's a bug.
On 11.01.2025 16:13, Russell King (Oracle) wrote: > On Sat, Jan 11, 2025 at 02:19:04PM +0100, Heiner Kallweit wrote: >> On 11.01.2025 11:01, Russell King (Oracle) wrote: >>> On Sat, Jan 11, 2025 at 10:44:25AM +0100, Heiner Kallweit wrote: >>>> On 11.01.2025 10:21, Russell King (Oracle) wrote: >>>>> On Sat, Jan 11, 2025 at 10:06:02AM +0100, Heiner Kallweit wrote: >>>>>> Link modes in phydev->eee_disabled_modes are filtered out by >>>>>> genphy_c45_write_eee_adv() and won't be advertised. Therefore >>>>>> don't accept such modes from userspace. >>>>> >>>>> Why do we need this? Surely if the MAC doesn't support modes, then they >>>>> should be filtered out of phydev->supported_eee so that userspace knows >>>>> that the mode is not supported by the network interface as a whole, just >>>>> like we do for phydev->supported. >>>>> >>>>> That would give us the checking here. >>>>> >>>> Removing EEE modes to be disabled from supported_eee is problematic >>>> because of how genphy_c45_write_eee_adv() works. >>>> >>>> Let's say we have a 2.5Gbps PHY and want to disable EEE at 2.5Gbps. If we >>>> remove 2.5Gbps from supported_eee, then the following check is false: >>>> if (linkmode_intersects(phydev->supported_eee, PHY_EEE_CAP2_FEATURES)) >>>> What would result in the 2.5Gbps mode not getting disabled. >>> >>> Ok. Do we at least remove the broken modes from the supported mask >>> reported to userspace? >>> >> I think that's something we could do in addition, to provide a hint to the >> user about unavailable modes. It wouldn't remove the need for the check here. >> ethtool doesn't check the advertisement against the supported modes. >> And even if it would, we must not rely on input from user space being sane. > > I disagree with some of this. Userspace should expect: > > - read current settings > - copy supported modes to advertised modes > - write current settings > > to work. If it fails, then how does ethtool, or even the user, work out > which link modes are actually supported or not. > > If we're introducing a failure on the "disabled" modes, then that is > a user-breaking change, and we need to avoid that. The current code > silently ignored the broken modes, your new code would error out on > the above action - and that's a bug. > OK, then I think what we can/should do: - filter out disabled EEE modes when populating data->supported in genphy_c45_ethtool_get_eee - silently filter out disabled EEE modes from user space provided EEE advertisement in genphy_c45_ethtool_set_eee
> > I disagree with some of this. Userspace should expect: > > > > - read current settings > > - copy supported modes to advertised modes > > - write current settings > > > > to work. If it fails, then how does ethtool, or even the user, work out > > which link modes are actually supported or not. > > > > If we're introducing a failure on the "disabled" modes, then that is > > a user-breaking change, and we need to avoid that. The current code > > silently ignored the broken modes, your new code would error out on > > the above action - and that's a bug. > > > OK, then I think what we can/should do: > - filter out disabled EEE modes when populating data->supported in > genphy_c45_ethtool_get_eee > - silently filter out disabled EEE modes from user space provided > EEE advertisement in genphy_c45_ethtool_set_eee Ideally, the kAPI should work just the same as normal advertised modes. The read API returns what can actually be used, and write API expects a subset of that. But maybe we have too much history and cannot enforce the subset without regressions, we just silently ignore the extra modes? It might be too much plumbing, but it would be nice to include an extack saying some modes have been ignored? I _think_ extack can be used without an error. Andrew
On 11.01.2025 18:31, Andrew Lunn wrote: >>> I disagree with some of this. Userspace should expect: >>> >>> - read current settings >>> - copy supported modes to advertised modes >>> - write current settings >>> >>> to work. If it fails, then how does ethtool, or even the user, work out >>> which link modes are actually supported or not. >>> >>> If we're introducing a failure on the "disabled" modes, then that is >>> a user-breaking change, and we need to avoid that. The current code >>> silently ignored the broken modes, your new code would error out on >>> the above action - and that's a bug. >>> >> OK, then I think what we can/should do: >> - filter out disabled EEE modes when populating data->supported in >> genphy_c45_ethtool_get_eee >> - silently filter out disabled EEE modes from user space provided >> EEE advertisement in genphy_c45_ethtool_set_eee > > Ideally, the kAPI should work just the same as normal advertised > modes. The read API returns what can actually be used, and write API > expects a subset of that. > > But maybe we have too much history and cannot enforce the subset > without regressions, we just silently ignore the extra modes? > > It might be too much plumbing, but it would be nice to include an > extack saying some modes have been ignored? I _think_ extack can be > used without an error. > Nice idea, this could be done as a follow-up. I'll have a look into this. > Andrew > Heiner > > >
diff --git a/drivers/net/phy/phy-c45.c b/drivers/net/phy/phy-c45.c index 468d24611..b566faba9 100644 --- a/drivers/net/phy/phy-c45.c +++ b/drivers/net/phy/phy-c45.c @@ -1559,6 +1559,11 @@ int genphy_c45_ethtool_set_eee(struct phy_device *phydev, phydev_warn(phydev, "At least some EEE link modes are not supported.\n"); return -EINVAL; } + linkmode_and(tmp, adv, phydev->eee_disabled_modes); + if (!linkmode_empty(tmp)) { + phydev_warn(phydev, "At least some EEE link modes are disabled.\n"); + return -EINVAL; + } linkmode_copy(phydev->advertising_eee, adv); } else if (linkmode_empty(phydev->advertising_eee)) { phy_advertise_eee_all(phydev);
Link modes in phydev->eee_disabled_modes are filtered out by genphy_c45_write_eee_adv() and won't be advertised. Therefore don't accept such modes from userspace. Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com> --- drivers/net/phy/phy-c45.c | 5 +++++ 1 file changed, 5 insertions(+)