Skip to contentSkip to conversation

8139too: send NETDEV_CHANGE manually when autoneg is disabled

PATCH In review Veaceslav Falico 4 participants 12 messages Mar 6, 2013 – Mar 11, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 8139too: send NETDEV_CHANGE manually when autoneg is disabled

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

---
 drivers/net/ethernet/realtek/8139too.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
drivers/net/ethernet/realtek/8139too.c +5 -0
index 1276ac7..96ee18c 100644
@@ -2393,6 +2393,11 @@ static int rtl8139_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
 	spin_lock_irq(&tp->lock);
 	rc = mii_ethtool_sset(&tp->mii, cmd);
 	spin_unlock_irq(&tp->lock);
+	/*
+	 * we don't restart on autoneg off, so notify manually
+	 */
+	if (tp->mii.force_media)
+		call_netdevice_notifiers(NETDEV_CHANGE, dev);
 	return rc;
 }
 
Signature · 1 line
1.7.1
#1 lore raw
Quoted 8 lines · Veaceslav Falico

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

Ben.

Quoted 21 lines
---
 drivers/net/ethernet/realtek/8139too.c |    5 +++++
 1 files changed, 5 insertions(+), 0 deletions(-)
drivers/net/ethernet/realtek/8139too.c +5 -0
index 1276ac7..96ee18c 100644
@@ -2393,6 +2393,11 @@ static int rtl8139_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
 	spin_lock_irq(&tp->lock);
 	rc = mii_ethtool_sset(&tp->mii, cmd);
 	spin_unlock_irq(&tp->lock);
+	/*
+	 * we don't restart on autoneg off, so notify manually
+	 */
+	if (tp->mii.force_media)
+		call_netdevice_notifiers(NETDEV_CHANGE, dev);
 	return rc;
 }
 
Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
maintainer
#2 lore raw

From: Ben Hutchings <[email protected]>
Date: Wed, 6 Mar 2013 20:22:52 +0000

Quoted 6 lines · Veaceslav Falico
Quoted 5 lines

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Quoted 6 lines
Quoted 3 lines

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

I think mii_check_media() is responsible for handling this too.

#3 lore raw

David Miller <[email protected]> :

Quoted 1 line

From: Ben Hutchings <[email protected]>

[...]

Quoted 4 lines
Quoted 2 lines

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

I think mii_check_media() is responsible for handling this too.

There are a few drivers that use mii_ethtool_sset as 8139too does while not relying upon mii_check_media.

What ought to be done for those ?

Signature · 1 line
Ueimor
#4 lore raw
In reply to David Miller
Quoted 9 lines · David Miller

From: Ben Hutchings <[email protected]>
Date: Wed, 6 Mar 2013 20:22:52 +0000

Quoted 6 lines · Veaceslav Falico
Quoted 5 lines

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

I think mii_check_media() is responsible for handling this too.

The mii_check_media() doesn't get called, AFAIK. The problem here is that, after we call ethtool -s eth0 autoneg off speed X, with eth0 being 8139too, the speed/autoneg options are changed via mii_ethtool_sset(), however the interface itself isn't down'ed/up'ed, and thus no NETDEV_ notifications are sent.

Other drivers either explicitly reset the interface after ethtool.set_settings() call (like netxen_nic using ndo_close()/ndo_open()), do it on the logic level (tg3) without involving mii_ethtool_sset(), or just reset on their own (e100 iirc), so that most of them are responsible for somehow triggering these events.

Silently changing speed can break things a bit - bonding relies on interface speeds for 802.3ad/alb/tlb/active-backup iirc, bridge relies on stp port cost etc. and they all get it via NETDEV_ notifications. So without them, they would end up with outdated data, per example (eth2 being 8139too):

darkmag:~#grep 'Interface\|Speed' /proc/net/bonding/bond0 Slave Interface: eth0
Speed: 100 Mbps
Slave Interface: eth2
Speed: 100 Mbps
darkmag:~#ethtool -s eth2 autoneg off speed 10 darkmag:~#cat /sys/class/net/eth2/speed
10
darkmag:~#grep 'Interface\|Speed' /proc/net/bonding/bond0 Slave Interface: eth0
Speed: 100 Mbps
Slave Interface: eth2
Speed: 100 Mbps

However, I think that mii_check_media() is also wrong :), though I didn't really dig into it. I'll check it when I'll have time.

#5 lore raw
Quoted 17 lines · Veaceslav Falico
Quoted 16 lines · David Miller

From: Ben Hutchings <[email protected]>
Date: Wed, 6 Mar 2013 20:22:52 +0000

Quoted 12 lines · Veaceslav Falico
Quoted 8 lines

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

Quoted 7 lines
Quoted 1 line

I think mii_check_media() is responsible for handling this too.

The mii_check_media() doesn't get called, AFAIK. The problem here is that, after we call ethtool -s eth0 autoneg off speed X, with eth0 being 8139too, the speed/autoneg options are changed via mii_ethtool_sset(), however the interface itself isn't down'ed/up'ed, and thus no NETDEV_ notifications are sent.

Does the hardware not send link interrupts if autoneg is disabled?

Quoted 11 lines

Other drivers either explicitly reset the interface after ethtool.set_settings() call (like netxen_nic using ndo_close()/ndo_open()), do it on the logic level (tg3) without involving mii_ethtool_sset(), or just reset on their own (e100 iirc), so that most of them are responsible for somehow triggering these events.

Silently changing speed can break things a bit - bonding relies on interface speeds for 802.3ad/alb/tlb/active-backup iirc, bridge relies on stp port cost etc. and they all get it via NETDEV_ notifications. So without them, they would end up with outdated data, per example (eth2 being 8139too):

[...]

Yes, I get it. But on real hardware, changing speed/duplex is always going to break the link if it's up. The link change notification should result in kernel and userland notifications via linkwatch_do_dev().

(If you're testing against QEMU rather than real hardware, there could be a bug in the emulation of link interrupts.)

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#6 lore raw
Quoted 20 lines · Ben Hutchings
Quoted 19 lines · Veaceslav Falico
Quoted 17 lines · David Miller

From: Ben Hutchings <[email protected]>
Date: Wed, 6 Mar 2013 20:22:52 +0000

Quoted 6 lines · Veaceslav Falico
Quoted 5 lines

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Quoted 6 lines
Quoted 3 lines

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

I think mii_check_media() is responsible for handling this too.

Quoted 7 lines
Quoted 5 lines

The mii_check_media() doesn't get called, AFAIK. The problem here is that, after we call ethtool -s eth0 autoneg off speed X, with eth0 being 8139too, the speed/autoneg options are changed via mii_ethtool_sset(), however the interface itself isn't down'ed/up'ed, and thus no NETDEV_ notifications are sent.

Does the hardware not send link interrupts if autoneg is disabled?

Yes, it doesn't send them when the autoneg off is set.

Quoted 12 lines
Quoted 7 lines

Other drivers either explicitly reset the interface after ethtool.set_settings() call (like netxen_nic using ndo_close()/ndo_open()), do it on the logic level (tg3) without involving mii_ethtool_sset(), or just reset on their own (e100 iirc), so that most of them are responsible for somehow triggering these events.

Quoted 5 lines

Silently changing speed can break things a bit - bonding relies on interface speeds for 802.3ad/alb/tlb/active-backup iirc, bridge relies on stp port cost etc. and they all get it via NETDEV_ notifications. So without them, they would end up with outdated data, per example (eth2 being 8139too):

Quoted 8 lines

[...]

Yes, I get it. But on real hardware, changing speed/duplex is always going to break the link if it's up. The link change notification should result in kernel and userland notifications via linkwatch_do_dev().

(If you're testing against QEMU rather than real hardware, there could be a bug in the emulation of link interrupts.)

It's real hardware:

[    4.339413] 8139cp 0000:10:09.0: This (id 10ec:8139 rev 10) is not an 8139C+ compatible chip, use 8139too
[    4.348210] 8139too: 8139too Fast Ethernet driver 0.9.28
[    4.350017] 8139too 0000:10:09.0 eth2: RealTek RTL8139 at 0xffffc90004574100, 00:14:d1:1f:b9:49, IRQ 21

darkmag:~#ethtool eth2
Settings for eth2:
<snip>
Supports auto-negotiation: Yes
Advertised link modes: 10baseT/Half 10baseT/Full
100baseT/Half 100baseT/Full
Advertised pause frame use: No
Advertised auto-negotiation: Yes
Link partner advertised link modes: 10baseT/Half 10baseT/Full
100baseT/Half 100baseT/Full
Link partner advertised pause frame use: Symmetric Link partner advertised auto-negotiation: Yes
Speed: 100Mb/s
Duplex: Full
<snip>
darkmag:~#dmesg | tail -n 10 | grep 8139too

[  164.019622] 8139too 0000:10:09.0 eth2: link up, 100Mbps, full-duplex, lpa 0x45E1
darkmag:~#ethtool -s eth2 autoneg off speed 10
darkmag:~#dmesg | tail -n 10 | grep 8139too
[  164.019622] 8139too 0000:10:09.0 eth2: link up, 100Mbps, full-duplex, lpa 0x45E1

darkmag:~#ethtool eth2
Settings for eth2:
<snip>
Supports auto-negotiation: Yes
Advertised link modes: Not reported
Advertised pause frame use: No
Advertised auto-negotiation: No
Speed: 10Mb/s
Duplex: Full

And, obviously, bonding's not updated:

darkmag:~#grep 'Interface\|Speed' /proc/net/bonding/bond0 Slave Interface: eth0
Speed: 100 Mbps
Slave Interface: eth2
Speed: 100 Mbps

Quoted 8 lines

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#7 lore raw
Quoted 49 lines · Veaceslav Falico
Quoted 20 lines · Ben Hutchings
Quoted 19 lines · Veaceslav Falico
Quoted 17 lines · David Miller

From: Ben Hutchings <[email protected]>
Date: Wed, 6 Mar 2013 20:22:52 +0000

Quoted 6 lines · Veaceslav Falico
Quoted 5 lines

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Quoted 6 lines
Quoted 3 lines

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

I think mii_check_media() is responsible for handling this too.

Quoted 7 lines
Quoted 5 lines

The mii_check_media() doesn't get called, AFAIK. The problem here is that, after we call ethtool -s eth0 autoneg off speed X, with eth0 being 8139too, the speed/autoneg options are changed via mii_ethtool_sset(), however the interface itself isn't down'ed/up'ed, and thus no NETDEV_ notifications are sent.

Does the hardware not send link interrupts if autoneg is disabled?

Yes, it doesn't send them when the autoneg off is set.

Quoted 18 lines
Quoted 7 lines

Other drivers either explicitly reset the interface after ethtool.set_settings() call (like netxen_nic using ndo_close()/ndo_open()), do it on the logic level (tg3) without involving mii_ethtool_sset(), or just reset on their own (e100 iirc), so that most of them are responsible for somehow triggering these events.

Quoted 5 lines

Silently changing speed can break things a bit - bonding relies on interface speeds for 802.3ad/alb/tlb/active-backup iirc, bridge relies on stp port cost etc. and they all get it via NETDEV_ notifications. So without them, they would end up with outdated data, per example (eth2 being 8139too):

[...]

Yes, I get it. But on real hardware, changing speed/duplex is always going to break the link if it's up. The link change notification should result in kernel and userland notifications via linkwatch_do_dev().

Quoted 8 lines
Quoted 2 lines

(If you're testing against QEMU rather than real hardware, there could be a bug in the emulation of link interrupts.)

It's real hardware:

[    4.339413] 8139cp 0000:10:09.0: This (id 10ec:8139 rev 10) is not an 8139C+ compatible chip, use 8139too
[    4.348210] 8139too: 8139too Fast Ethernet driver 0.9.28
[    4.350017] 8139too 0000:10:09.0 eth2: RealTek RTL8139 at 0xffffc90004574100, 00:14:d1:1f:b9:49, IRQ 21

[...]

OK. But it's generally not enough to send a 'something changed' notification; the driver actually has to update the link state. MAybe in your test you reconfigure the other end of the link too, or it's smart enough to do parallel detect. But in general, changing the link speed is going to change the link state, and the TX scheduler needs to enabled or disabled accordingly.

Perhaps this driver needs to run a link polling timer while autoneg is off.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#8 lore raw
Quoted 58 lines · Ben Hutchings
Quoted 49 lines · Veaceslav Falico
Quoted 20 lines · Ben Hutchings
Quoted 19 lines · Veaceslav Falico
Quoted 17 lines · David Miller

From: Ben Hutchings <[email protected]>
Date: Wed, 6 Mar 2013 20:22:52 +0000

Quoted 6 lines · Veaceslav Falico
Quoted 5 lines

When setting autoneg off (with any additional parameters, like speed/duplex), 8139too doesn't do an interface reset, and thus doesn't notify anyone that its speed/duplex might have changed (bonding and bridge will not see the speed changes, per example).

Quoted 6 lines
Quoted 3 lines

Verify if we've force_media and send notification manually, so that the listeners have a chance to see the changes. It's quite ugly, however I don't see anything better.

Isn't this really a bug in mii_check_media()? It shouldn't shortcut the calls to netif_carrier_{off,on}() just because mii->force_media is set.

I think mii_check_media() is responsible for handling this too.

Quoted 7 lines
Quoted 5 lines

The mii_check_media() doesn't get called, AFAIK. The problem here is that, after we call ethtool -s eth0 autoneg off speed X, with eth0 being 8139too, the speed/autoneg options are changed via mii_ethtool_sset(), however the interface itself isn't down'ed/up'ed, and thus no NETDEV_ notifications are sent.

Does the hardware not send link interrupts if autoneg is disabled?

Yes, it doesn't send them when the autoneg off is set.

Quoted 18 lines
Quoted 7 lines

Other drivers either explicitly reset the interface after ethtool.set_settings() call (like netxen_nic using ndo_close()/ndo_open()), do it on the logic level (tg3) without involving mii_ethtool_sset(), or just reset on their own (e100 iirc), so that most of them are responsible for somehow triggering these events.

Quoted 5 lines

Silently changing speed can break things a bit - bonding relies on interface speeds for 802.3ad/alb/tlb/active-backup iirc, bridge relies on stp port cost etc. and they all get it via NETDEV_ notifications. So without them, they would end up with outdated data, per example (eth2 being 8139too):

[...]

Yes, I get it. But on real hardware, changing speed/duplex is always going to break the link if it's up. The link change notification should result in kernel and userland notifications via linkwatch_do_dev().

Quoted 8 lines
Quoted 2 lines

(If you're testing against QEMU rather than real hardware, there could be a bug in the emulation of link interrupts.)

It's real hardware:

[    4.339413] 8139cp 0000:10:09.0: This (id 10ec:8139 rev 10) is not an 8139C+ compatible chip, use 8139too
[    4.348210] 8139too: 8139too Fast Ethernet driver 0.9.28
[    4.350017] 8139too 0000:10:09.0 eth2: RealTek RTL8139 at 0xffffc90004574100, 00:14:d1:1f:b9:49, IRQ 21
Quoted 8 lines

[...]

OK. But it's generally not enough to send a 'something changed' notification; the driver actually has to update the link state. MAybe in your test you reconfigure the other end of the link too, or it's smart enough to do parallel detect. But in general, changing the link speed is going to change the link state, and the TX scheduler needs to enabled or disabled accordingly.

Sorry, I think I didn't explain it clearly. The driver *does* notify about link changes even when it has autonegotiation disabled.

The only thing that doesn't work is the notification when someone changes it via ethtool. In the same moment.

i.e. when the other peer ifdowns the port - the driver does see it and notifies about it.

The only scenario that doesn't work is when 8139too is manually (locally) changed, by ethtool, to autoneg off. Then it doesn't notify anything. So, effectively, if the administrator issues an 'ethtool -s eth0 autoneg off speed X', bonding/bridge/etc. won't be notified about that. However, if the cable is pulled afterwards - it will notify about link down.

Basically, when changing the interface manually to autoneg off and whatever else, the driver doesn't issue the NETDEV_* event. That's what I've added.

Other drivers, after setting autoneg off, usually ifdown/ifup the interface manually, and thus everything gets updated. That is another way to fix this driver - to issue ->ndo_stop(); ->ndo_open(); (as netxen_nic/qlcnic/etc do), but I think that just issuing a notification that we've changed the state, without flapping, is better.

Quoted 11 lines

Perhaps this driver needs to run a link polling timer while autoneg is off.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#9 lore raw
Quoted 4 lines · Veaceslav Falico
Quoted 3 lines · Ben Hutchings
Quoted 2 lines · Veaceslav Falico
Quoted 1 line · Ben Hutchings

On Thu, 2013-03-07 at 11:27 +0100, Veaceslav Falico wrote:

[...]

Quoted 28 lines
Quoted 19 lines
Quoted 11 lines
Quoted 11 lines
Quoted 5 lines

Silently changing speed can break things a bit - bonding relies on interface speeds for 802.3ad/alb/tlb/active-backup iirc, bridge relies on stp port cost etc. and they all get it via NETDEV_ notifications. So without them, they would end up with outdated data, per example (eth2 being 8139too):

[...]

Yes, I get it. But on real hardware, changing speed/duplex is always going to break the link if it's up. The link change notification should result in kernel and userland notifications via linkwatch_do_dev().

Quoted 8 lines
Quoted 2 lines

(If you're testing against QEMU rather than real hardware, there could be a bug in the emulation of link interrupts.)

It's real hardware:

[    4.339413] 8139cp 0000:10:09.0: This (id 10ec:8139 rev 10) is not an 8139C+ compatible chip, use 8139too
[    4.348210] 8139too: 8139too Fast Ethernet driver 0.9.28
[    4.350017] 8139too 0000:10:09.0 eth2: RealTek RTL8139 at 0xffffc90004574100, 00:14:d1:1f:b9:49, IRQ 21
Quoted 8 lines

[...]

OK. But it's generally not enough to send a 'something changed' notification; the driver actually has to update the link state. MAybe in your test you reconfigure the other end of the link too, or it's smart enough to do parallel detect. But in general, changing the link speed is going to change the link state, and the TX scheduler needs to enabled or disabled accordingly.

Quoted 5 lines

Sorry, I think I didn't explain it clearly. The driver *does* notify about link changes even when it has autonegotiation disabled.

The only thing that doesn't work is the notification when someone changes it via ethtool. In the same moment.

The link *should* go down (assuming it was up before) because you likely have a speed mismatch and won't be able to pass traffic. I don't know whether this is possible for the hardware to detect immediately; maybe a PHY reset will ensure that it is detected.

[...]

Quoted 5 lines

Other drivers, after setting autoneg off, usually ifdown/ifup the interface manually, and thus everything gets updated. That is another way to fix this driver - to issue ->ndo_stop(); ->ndo_open(); (as netxen_nic/qlcnic/etc do), but I think that just issuing a notification that we've changed the state, without flapping, is better.

The link *should* flap, though doing a full stop/start is likely unnecessary.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#10 lore raw
Quoted 37 lines · Ben Hutchings
Quoted 4 lines · Veaceslav Falico
Quoted 3 lines · Ben Hutchings
Quoted 2 lines · Veaceslav Falico
Quoted 1 line · Ben Hutchings

On Thu, 2013-03-07 at 11:27 +0100, Veaceslav Falico wrote:

[...]

Quoted 31 lines
Quoted 19 lines
Quoted 11 lines
Quoted 11 lines
Quoted 5 lines

Silently changing speed can break things a bit - bonding relies on interface speeds for 802.3ad/alb/tlb/active-backup iirc, bridge relies on stp port cost etc. and they all get it via NETDEV_ notifications. So without them, they would end up with outdated data, per example (eth2 being 8139too):

[...]

Yes, I get it. But on real hardware, changing speed/duplex is always going to break the link if it's up. The link change notification should result in kernel and userland notifications via linkwatch_do_dev().

Quoted 8 lines
Quoted 2 lines

(If you're testing against QEMU rather than real hardware, there could be a bug in the emulation of link interrupts.)

It's real hardware:

[    4.339413] 8139cp 0000:10:09.0: This (id 10ec:8139 rev 10) is not an 8139C+ compatible chip, use 8139too
[    4.348210] 8139too: 8139too Fast Ethernet driver 0.9.28
[    4.350017] 8139too 0000:10:09.0 eth2: RealTek RTL8139 at 0xffffc90004574100, 00:14:d1:1f:b9:49, IRQ 21
Quoted 8 lines

[...]

OK. But it's generally not enough to send a 'something changed' notification; the driver actually has to update the link state. MAybe in your test you reconfigure the other end of the link too, or it's smart enough to do parallel detect. But in general, changing the link speed is going to change the link state, and the TX scheduler needs to enabled or disabled accordingly.

Sorry, I think I didn't explain it clearly. The driver *does* notify about link changes even when it has autonegotiation disabled.

Quoted 7 lines
Quoted 2 lines

The only thing that doesn't work is the notification when someone changes it via ethtool. In the same moment.

The link *should* go down (assuming it was up before) because you likely have a speed mismatch and won't be able to pass traffic. I don't know whether this is possible for the hardware to detect immediately; maybe a PHY reset will ensure that it is detected.

Sorry for late response - I've had a feeling that I was writing something utterly dumb. My feeling was correct. I've looked through it and came up with the next patch, though I have a feeling... :)

Anyway, you were right - the hardware does everything needed, and the driver also. The normal way (with autoneg on) would be for the driver to call mii_check_media() when it got the interrupt, which would verify if state/duplex has changed and report it.

However, with autoneg off, and thus mii->force_duplex == 1, mii_check_media() just returns 'nothing changed' back, without doing anything else. This way, any media change while in autoneg off (like up/down notifications) are completely hidden to the rest of the kernel.

I've modified the mii_check_media() to not verify for ->force_media and to get all the media settings from mii_ethtool_gset(), which would take care of whether autoneg is on or off and return the correct values. Tested, works ok - it sees when the cable is dis/connected, notifies bonding of speed/duplex changes, when enslaved.

Subject: [PATCH] mii: make mii_check_media() work with ->force_media == 1

mii_check_media() does nothing when ->force_media is set, and thus might miss media change events reported by the drivers if the autonegotiation is off.

Make mii_check_media() use mii_ethtool_gset() for getting media settings, which cleans the code a lot and works with both autoneg off and on. This allows us to catch link notifications even ->force_media on.

---
 drivers/net/mii.c |   45 ++++++++++++---------------------------------
 1 files changed, 12 insertions(+), 33 deletions(-)
drivers/net/mii.c +11 -33
index 4a99c39..2447487 100644
@@ -308,19 +308,14 @@ void mii_check_link (struct mii_if_info *mii)
  * @init_media: OK to save duplex mode in @mii
  *
  * Returns 1 if the duplex mode changed, 0 if not.
- * If the media type is forced, always returns 0.
  */
 unsigned int mii_check_media (struct mii_if_info *mii,
 			      unsigned int ok_to_print,
 			      unsigned int init_media)
 {
 	unsigned int old_carrier, new_carrier;
-	int advertise, lpa, media, duplex;
-	int lpa2 = 0;
-
-	/* if forced media, go no further */
-	if (mii->force_media)
-		return 0; /* duplex did not change */
+	int duplex;
+	struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET };
 
 	/* check current and old link status */
 	old_carrier = netif_carrier_ok(mii->dev) ? 1 : 0;
@@ -345,37 +340,21 @@ unsigned int mii_check_media (struct mii_if_info *mii,
 	 */
 	netif_carrier_on(mii->dev);
 
-	/* get MII advertise and LPA values */
-	if ((!init_media) && (mii->advertising))
-		advertise = mii->advertising;
-	else {
-		advertise = mii->mdio_read(mii->dev, mii->phy_id, MII_ADVERTISE);
-		mii->advertising = advertise;
-	}
-	lpa = mii->mdio_read(mii->dev, mii->phy_id, MII_LPA);
-	if (mii->supports_gmii)
-		lpa2 = mii->mdio_read(mii->dev, mii->phy_id, MII_STAT1000);
+	/*
+	 * save the previous state of the duplex, mii_ethtool_gset()
+	 * modifies it
+	 */
+	duplex = mii->full_duplex;
 
-	/* figure out media and duplex from advertise and LPA values */
-	media = mii_nway_result(lpa & advertise);
-	duplex = (media & ADVERTISE_FULL) ? 1 : 0;
-	if (lpa2 & LPA_1000FULL)
-		duplex = 1;
+	mii_ethtool_gset(mii, &ecmd);
 
 	if (ok_to_print)
 		netdev_info(mii->dev, "link up, %uMbps, %s-duplex, lpa 0x%04X\n",
-			    lpa2 & (LPA_1000FULL | LPA_1000HALF) ? 1000 :
-			    media & (ADVERTISE_100FULL | ADVERTISE_100HALF) ?
-			    100 : 10,
-			    duplex ? "full" : "half",
-			    lpa);
-
-	if ((init_media) || (mii->full_duplex != duplex)) {
-		mii->full_duplex = duplex;
-		return 1; /* duplex changed */
-	}
+			    ethtool_cmd_speed(&ecmd), +			    ecmd.duplex == DUPLEX_FULL ? "full" : "half",
+			    ecmd.lp_advertising);
 
-	return 0; /* duplex did not change */
+	return (init_media || (mii->full_duplex != duplex)) ? 1 : 0;
 }
 
 /**
Signature · 1 line
1.7.1
Quoted 18 lines

[...]

Quoted 5 lines

Other drivers, after setting autoneg off, usually ifdown/ifup the interface manually, and thus everything gets updated. That is another way to fix this driver - to issue ->ndo_stop(); ->ndo_open(); (as netxen_nic/qlcnic/etc do), but I think that just issuing a notification that we've changed the state, without flapping, is better.

The link *should* flap, though doing a full stop/start is likely unnecessary.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
#11 lore raw

On Mon, 2013-03-11 at 18:58 +0100, Veaceslav Falico wrote:
[...]

Quoted 36 lines

Anyway, you were right - the hardware does everything needed, and the driver also. The normal way (with autoneg on) would be for the driver to call mii_check_media() when it got the interrupt, which would verify if state/duplex has changed and report it.

However, with autoneg off, and thus mii->force_duplex == 1, mii_check_media() just returns 'nothing changed' back, without doing anything else. This way, any media change while in autoneg off (like up/down notifications) are completely hidden to the rest of the kernel.

I've modified the mii_check_media() to not verify for ->force_media and to get all the media settings from mii_ethtool_gset(), which would take care of whether autoneg is on or off and return the correct values. Tested, works ok - it sees when the cable is dis/connected, notifies bonding of speed/duplex changes, when enslaved.

Subject: [PATCH] mii: make mii_check_media() work with ->force_media == 1

mii_check_media() does nothing when ->force_media is set, and thus might miss media change events reported by the drivers if the autonegotiation is off.

Make mii_check_media() use mii_ethtool_gset() for getting media settings, which cleans the code a lot and works with both autoneg off and on. This allows us to catch link notifications even ->force_media on.

---
  drivers/net/mii.c |   45 ++++++++++++---------------------------------
  1 files changed, 12 insertions(+), 33 deletions(-)
drivers/net/mii.c +0 -0
index 4a99c39..2447487 100644

drivers/net/mii.c void mii_check_link (struct mii_if_info *mii) · line 311

Quoted 5 lines
drivers/net/mii.c +0 -1
@@ -308,19 +308,14 @@ void mii_check_link (struct mii_if_info *mii)
   * @init_media: OK to save duplex mode in @mii
   *
   * Returns 1 if the duplex mode changed, 0 if not.
- * If the media type is forced, always returns 0.
   */

Perhaps you could also correct the summary line and the description of @init_media while you're doing this.

Quoted 17 lines
drivers/net/mii.c +2 -6
  unsigned int mii_check_media (struct mii_if_info *mii,
  			      unsigned int ok_to_print,
  			      unsigned int init_media)
  {
  	unsigned int old_carrier, new_carrier;
-	int advertise, lpa, media, duplex;
-	int lpa2 = 0;
-
-	/* if forced media, go no further */
-	if (mii->force_media)
-		return 0; /* duplex did not change */
+	int duplex;
+	struct ethtool_cmd ecmd = { .cmd = ETHTOOL_GSET };
 
  	/* check current and old link status */
  	old_carrier = netif_carrier_ok(mii->dev) ? 1 : 0;

drivers/net/mii.c unsigned int mii_check_media (struct mii_if_info *mii, · line 353

Quoted 9 lines
drivers/net/mii.c +0 -6
@@ -345,37 +340,21 @@ unsigned int mii_check_media (struct mii_if_info *mii,
  	 */
  	netif_carrier_on(mii->dev);
 
-	/* get MII advertise and LPA values */
-	if ((!init_media) && (mii->advertising))
-		advertise = mii->advertising;
-	else {
-		advertise = mii->mdio_read(mii->dev, mii->phy_id, MII_ADVERTISE);
-		mii->advertising = advertise;

Now mii->advertising won't be initialised properly.

Quoted 22 lines
drivers/net/mii.c +6 -12
-	}
-	lpa = mii->mdio_read(mii->dev, mii->phy_id, MII_LPA);
-	if (mii->supports_gmii)
-		lpa2 = mii->mdio_read(mii->dev, mii->phy_id, MII_STAT1000);
+	/*
+	 * save the previous state of the duplex, mii_ethtool_gset()
+	 * modifies it
+	 */
+	duplex = mii->full_duplex;
 
-	/* figure out media and duplex from advertise and LPA values */
-	media = mii_nway_result(lpa & advertise);
-	duplex = (media & ADVERTISE_FULL) ? 1 : 0;
-	if (lpa2 & LPA_1000FULL)
-		duplex = 1;
+	mii_ethtool_gset(mii, &ecmd);
 
  	if (ok_to_print)
  		netdev_info(mii->dev, "link up, %uMbps, %s-duplex, lpa 0x%04X\n",
-			    lpa2 & (LPA_1000FULL | LPA_1000HALF) ? 1000 :
-			    media & (ADVERTISE_100FULL | ADVERTISE_100HALF) ?
-			    100 : 10,

drivers/net/mii.c unsigned int mii_check_media (struct mii_if_info *mii, · line 354

Quoted 10 lines
drivers/net/mii.c +3 -7
-			    duplex ? "full" : "half",
-			    lpa);
-
-	if ((init_media) || (mii->full_duplex != duplex)) {
-		mii->full_duplex = duplex;
-		return 1; /* duplex changed */
-	}
+			    ethtool_cmd_speed(&ecmd), 
+			    ecmd.duplex == DUPLEX_FULL ? "full" : "half",
+			    ecmd.lp_advertising);

This log message used to include the raw LPA register value, but you're changing it to use ecmd.lp_advertising which has entirely different flag values. In order to get the raw LPA value out, I think you'll need to delete mii_get_an() and split up mii_ethtool_gset() so you end up with:

static void __mii_check_media(struct mii_if_info *mii, struct ethtool_cmd *ecmd,

			      u16 *raw_lpa)
{
	*raw_lpa = 0;
	...
		mii->advertising = mii->mdio_read(mii, mii->phy_id, MII_ADVERTISE);
		ecmd->advertising |= mii_lpa_to_ethtool_lpa_t(mii->advertising);
	...
			*raw_lpa = mii->mdio_read(mii, mii->phy_id, MII_LPA);
			ecmd->lp_advertising =
				mii_lpa_to_ethtool_lpa_t(*raw_lpa);
	...
}

int mii_ethtool_gset(struct mii_if_info *mii, struct ethtool_cmd *ecmd)

{
	u16 raw_lpa;
	__mii_check_media(mii, ecmd, &raw_lpa);
	return 0;
}

unsigned int mii_check_media (struct mii_if_info *mii,
unsigned int ok_to_print, unsigned int init_media)

{
	u16 raw_lpa;
	...
	__mii_check_media(mii, ecmd, &raw_lpa);
	...
	if (ok_to_print)
		netdev_info(mii->dev, "link up, %uMbps, %s-duplex, lpa 0x%04X\n",
			    ethtool_cmd_speed(&ecmd), 
			    ecmd.duplex == DUPLEX_FULL ? "full" : "half",
			    raw_lpa);
	...
}

drivers/net/mii.c unsigned int mii_check_media (struct mii_if_info *mii, · line 359

Quoted 5 lines
drivers/net/mii.c +1 -1
-	return 0; /* duplex did not change */
+	return (init_media || (mii->full_duplex != duplex)) ? 1 : 0;
  }
 
  /**

No need for parentheses or the ?: operator.

Ben.

Signature · 3 lines
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.