Fix couple of issues with 'ethtool' get/set API's

PATCH In review Arun Parameswaran 3 participants 10 messages May 22, 2015 – Jun 1, 2015

Subsystems (2)

Networking Network drivers

lore

Share & tools

lore

Arun Parameswaran

author

2015-05-22 22:43 UTC #0 lore raw

cover Fix couple of issues with 'ethtool' get/set API's

Hi,
The patch fixes 2 issues with 'ethtool' getting/setting parametres in the do_gset() do_sset() API's.

I have pushed a patch to the Kernel to fix an issue in the handling of the 'ethtool' commands which got accepted.
This Kernel patch was based on Linux v4.1-rc4 and is available in:
https://github.com/Broadcom/cygnus-linux/tree/net-core-ethtool-fix-v1

The Kernel was always clearing the command from the 'ethtool' resulting in all operations to deal with PHY0. This prevents querying/setting PHY 1's settings.

The first patch is required due to the fact that Kernel will not be clearing the command data structure. Without this the query will fail as the command will get to the Kernel with arbitary phyad info.

The second patch is to fix the issue in which the 'ethtool' gets the settings of the wrong phyad before applying the new settings.

These changes will still work with older Kernel's (as the Kernel clears the command data structure).

Thanks

Arun Parameswaran (2):
ethtool: Clear the command data structure before sending requests
ethtool: Fix an issue with handling 'phyad' while updating settings

 ethtool.c |    9 +++++++++
 1 file changed, 9 insertions(+)

Signature · 1 line

1.7.9.5

Arun Parameswaran

author

2015-05-22 22:43 UTC #1 lore raw

PATCH 1/2 ethtool: Clear the command data structure before sending requests

The 'ethtool' is not clearing the command data structure in the do_gset() & do_sset() API's while sending commands to get/set parametres. This used to work since the Kernel clears the data structure (which was a bug and is being fixed).

This patch adds a 'memset' to the do_gset() & do_sset() to clear the command data structure. This will prevent the 'ethtool' from passing wrong parametres, for example incorrect 'phyad'.

Signed-off-by Arun Parameswaran <[email protected]>

Reviewed-by JD (Jiandong) Zheng <[email protected]>

---
 ethtool.c |    3 +++
 1 file changed, 3 insertions(+)

ethtool.c +3 -0

index 01b13a6..668c189 100644
@@ -2275,6 +2275,7 @@ static int do_gset(struct cmd_context *ctx)
 
 	fprintf(stdout, "Settings for %s:\n", ctx->devname);
 
+	memset(&ecmd, 0, sizeof(struct ethtool_cmd));
 	ecmd.cmd = ETHTOOL_GSET;
 	err = send_ioctl(ctx, &ecmd);
 	if (err == 0) {

@@ -2520,6 +2521,8 @@ static int do_sset(struct cmd_context *ctx)
 	if (gset_changed) {
 		struct ethtool_cmd ecmd;
 
+		memset(&ecmd, 0, sizeof(struct ethtool_cmd));
+
 		ecmd.cmd = ETHTOOL_GSET;
 		err = send_ioctl(ctx, &ecmd);
 		if (err < 0) {

Signature · 1 line

1.7.9.5

Arun Parameswaran

author

2015-05-22 22:43 UTC #2 lore raw

PATCH 2/2 ethtool: Fix an issue with handling 'phyad' while updating settings

In reply to Arun Parameswaran

When using the 'ethtool' to set PHY1's settings, the 'ethtool' seems to do the 'get' settings with incorrect PHY ID and then apply the new settings on it.

For example when using the command, to change PHY1's speed, 'ethtool -s eth0 phyad 1 speed 100', the 'ethtool' will end up querying PHY0's settings and then apply the new settings specified in the command on top of the PHY0's settings. So the PHY1 will end up with worng settings (partial settings of PHY0 for example).

This patch fixes the issue by specifying the 'phyad' or 'port' requested in the command to query for the settings, before applying the new ones.

Signed-off-by Arun Parameswaran <[email protected]>

Reviewed-by JD (Jiandong) Zheng <[email protected]>

---
 ethtool.c |    6 ++++++
 1 file changed, 6 insertions(+)

ethtool.c +6 -0

index 668c189..6c57297 100644
@@ -2523,6 +2523,12 @@ static int do_sset(struct cmd_context *ctx)
 
 		memset(&ecmd, 0, sizeof(struct ethtool_cmd));
 
+		if (phyad_wanted != -1)
+			ecmd.phy_address = phyad_wanted;
+
+		if (port_wanted != -1)
+			ecmd.port = port_wanted;
+
 		ecmd.cmd = ETHTOOL_GSET;
 		err = send_ioctl(ctx, &ecmd);
 		if (err < 0) {

Signature · 1 line

1.7.9.5

Ben Hutchings

2015-05-31 19:59 UTC #3 lore raw

In reply to Arun Parameswaran

Quoted 4 lines · Arun Parameswaran

Hi,
The patch fixes 2 issues with 'ethtool' getting/setting parametres in the do_gset() do_sset() API's.

Quoted 8 lines

I have pushed a patch to the Kernel to fix an issue in the handling of the 'ethtool' commands which got accepted.
This Kernel patch was based on Linux v4.1-rc4 and is available in:
https://github.com/Broadcom/cygnus-linux/tree/net-core-ethtool-fix-v1
The Kernel was always clearing the command from the 'ethtool' resulting in all operations to deal with PHY0. This prevents querying/setting PHY 1's settings.

[...]

Each net device can be associated with a single PHY at a time, and the ETHTOOL_GSET implementation should fill in the PHY address in the ethtool_cmd::phy_address field. Where there are multiple PHYs that can be connected to the net device's MAC, an ETHTOOL_SSET operation can be used to change that PHY address.

The ethtool API is not meant for controlling other PHYs that aren't connected to the MAC; if you want to do that then create more net devices for them or use the MDIO ioctls.

Ben.

Signature · 2 lines

Ben Hutchings
Reality is just a crutch for people who can't handle science fiction.

Arun Parameswaran

author

2015-06-01 17:14 UTC #4 lore raw

Quoted 13 lines · Ben Hutchings

Quoted 4 lines · Arun Parameswaran
Hi,
The patch fixes 2 issues with 'ethtool' getting/setting parametres in the do_gset() do_sset() API's.
Quoted 8 lines
I have pushed a patch to the Kernel to fix an issue in the handling of the 'ethtool' commands which got accepted.
This Kernel patch was based on Linux v4.1-rc4 and is available in:
https://github.com/Broadcom/cygnus-linux/tree/net-core-ethtool-fix-v1
The Kernel was always clearing the command from the 'ethtool' resulting in all operations to deal with PHY0. This prevents querying/setting PHY 1's settings.

Quoted 8 lines

[...]
Each net device can be associated with a single PHY at a time, and the ETHTOOL_GSET implementation should fill in the PHY address in the ethtool_cmd::phy_address field. Where there are multiple PHYs that can be connected to the net device's MAC, an ETHTOOL_SSET operation can be used to change that PHY address.

The above can be done by the driver when there is one PHY per MAC. In our case we have multiple PHYs controlled by the same MAC. I should have clarified this earlier, I apologize.

When we specify the 'phyad', in the command line, we were expecting the 'ethtool' to fetch/set data for that 'phyad'. This is the intend of the patch.

With the patch (in 'ethtool' and Kernel), if 'phyad' is not specified, it will still function as you described above, it will be up to the driver to return the proper 'phyad' and related settings.

Quoted 4 lines

The ethtool API is not meant for controlling other PHYs that aren't connected to the MAC; if you want to do that then create more net devices for them or use the MDIO ioctls.

In the SoC, there are multiple PHYs (in our case there are 2) controlled by the same MAC. We are trying to use 'ethtool' to control both the PHYs connected to the same MAC.

Quoted 3 lines

Ben.

Thanks
Arun

Ben Hutchings

2015-06-01 18:07 UTC #5 lore raw

Quoted 22 lines · Arun Parameswaran

Quoted 13 lines · Ben Hutchings
Quoted 4 lines · Arun Parameswaran
Hi,
The patch fixes 2 issues with 'ethtool' getting/setting parametres in the do_gset() do_sset() API's.
Quoted 8 lines
I have pushed a patch to the Kernel to fix an issue in the handling of the 'ethtool' commands which got accepted.
This Kernel patch was based on Linux v4.1-rc4 and is available in:
https://github.com/Broadcom/cygnus-linux/tree/net-core-ethtool-fix-v1
The Kernel was always clearing the command from the 'ethtool' resulting in all operations to deal with PHY0. This prevents querying/setting PHY 1's settings.
Quoted 8 lines
[...]
Each net device can be associated with a single PHY at a time, and the ETHTOOL_GSET implementation should fill in the PHY address in the ethtool_cmd::phy_address field. Where there are multiple PHYs that can be connected to the net device's MAC, an ETHTOOL_SSET operation can be used to change that PHY address.

Quoted 3 lines

The above can be done by the driver when there is one PHY per MAC. In our case we have multiple PHYs controlled by the same MAC. I should have clarified this earlier, I apologize.

I understand that you can have multiple PHYs on the same MDIO bus, but not how the MAC can use them at the same time. Is this hardware level bonding? Or are multiple PHYs needed for a single link?

Quoted 7 lines

When we specify the 'phyad', in the command line, we were expecting the 'ethtool' to fetch/set data for that 'phyad'. This is the intend of the patch.
With the patch (in 'ethtool' and Kernel), if 'phyad' is not specified, it will still function as you described above, it will be up to the driver to return the proper 'phyad' and related settings.

[...]

But without the patch in ethtool and other programs calling this API (it's not just the ethtool command!), you get random junk as the phy_address. How will you tell whether it's valid or not?

Ben.

Signature · 3 lines

Ben Hutchings
Power corrupts.  Absolute power is kind of neat.
                           - John Lehman, Secretary of the US Navy 1981-1987

Arun Parameswaran

author

2015-06-01 19:12 UTC #6 lore raw

Quoted 23 lines · Ben Hutchings

Quoted 22 lines · Arun Parameswaran
Quoted 21 lines · Ben Hutchings
Quoted 4 lines · Arun Parameswaran
Hi,
The patch fixes 2 issues with 'ethtool' getting/setting parametres in the do_gset() do_sset() API's.
Quoted 8 lines
I have pushed a patch to the Kernel to fix an issue in the handling of the 'ethtool' commands which got accepted.
This Kernel patch was based on Linux v4.1-rc4 and is available in:
https://github.com/Broadcom/cygnus-linux/tree/net-core-ethtool-fix-v1
The Kernel was always clearing the command from the 'ethtool' resulting in all operations to deal with PHY0. This prevents querying/setting PHY 1's settings.
[...]
Each net device can be associated with a single PHY at a time, and the ETHTOOL_GSET implementation should fill in the PHY address in the ethtool_cmd::phy_address field. Where there are multiple PHYs that can be connected to the net device's MAC, an ETHTOOL_SSET operation can be used to change that PHY address.

Quoted 8 lines

Quoted 3 lines
The above can be done by the driver when there is one PHY per MAC. In our case we have multiple PHYs controlled by the same MAC. I should have clarified this earlier, I apologize.
I understand that you can have multiple PHYs on the same MDIO bus, but not how the MAC can use them at the same time. Is this hardware level bonding? Or are multiple PHYs needed for a single link?

We have an internal switch which manages the traffic to the PHY's (ports). There is 1 PHY per external port.
The MAC is connected to the internal port of the switch.

Quoted 4 lines

Quoted 4 lines
When we specify the 'phyad', in the command line, we were expecting the 'ethtool' to fetch/set data for that 'phyad'. This is the intend of the patch.

Quoted 8 lines

Quoted 3 lines
With the patch (in 'ethtool' and Kernel), if 'phyad' is not specified, it will still function as you described above, it will be up to the driver to return the proper 'phyad' and related settings.
[...]
But without the patch in ethtool and other programs calling this API (it's not just the ethtool command!), you get random junk as the phy_address. How will you tell whether it's valid or not?

Yes, if other programs make use of the 'ethtool' commands, they will have to clear the data structure. I think that the clearing of the command data structure should be the responsibility of the programs calling it. I understand that the programs didn't have to do it as Kernel was clearing the command structure for them.

But this prevents the 'ethtool' from being used to get/set data of specific PHY's.

Quoted 3 lines

Ben.

Thanks
Arun

Ben Hutchings

2015-06-01 19:29 UTC #7 lore raw

Quoted 28 lines · Arun Parameswaran

Quoted 27 lines · Ben Hutchings
Quoted 25 lines · Arun Parameswaran
Quoted 13 lines · Ben Hutchings
Quoted 4 lines · Arun Parameswaran
Hi,
The patch fixes 2 issues with 'ethtool' getting/setting parametres in the do_gset() do_sset() API's.
Quoted 8 lines
I have pushed a patch to the Kernel to fix an issue in the handling of the 'ethtool' commands which got accepted.
This Kernel patch was based on Linux v4.1-rc4 and is available in:
https://github.com/Broadcom/cygnus-linux/tree/net-core-ethtool-fix-v1
The Kernel was always clearing the command from the 'ethtool' resulting in all operations to deal with PHY0. This prevents querying/setting PHY 1's settings.
Quoted 8 lines
[...]
Each net device can be associated with a single PHY at a time, and the ETHTOOL_GSET implementation should fill in the PHY address in the ethtool_cmd::phy_address field. Where there are multiple PHYs that can be connected to the net device's MAC, an ETHTOOL_SSET operation can be used to change that PHY address.
The above can be done by the driver when there is one PHY per MAC. In our case we have multiple PHYs controlled by the same MAC. I should have clarified this earlier, I apologize.

Quoted 7 lines

Quoted 4 lines
I understand that you can have multiple PHYs on the same MDIO bus, but not how the MAC can use them at the same time. Is this hardware level bonding? Or are multiple PHYs needed for a single link?
We have an internal switch which manages the traffic to the PHY's (ports). There is 1 PHY per external port.
The MAC is connected to the internal port of the switch.

Then you should create net devices for those external ports as well as the internal port.

If I understand the switchdev API rightly, the external port devices should implement the ethtool {get,set}_settings operations and the ndo_switch_parent_id_get operation. The existing net device should expose only the internal link to the switch (which presumably isn't configurable at all).

[...]

Quoted 2 lines

But this prevents the 'ethtool' from being used to get/set data of specific PHY's.

That is fine because it is meant to manage the net device's own link (in this case, the internal port), not other switch ports.

Ben.

Signature · 3 lines

Ben Hutchings
Power corrupts.  Absolute power is kind of neat.
                           - John Lehman, Secretary of the US Navy 1981-1987

Arun Parameswaran

author

2015-06-01 21:00 UTC #8 lore raw

Quoted 33 lines · Ben Hutchings

Quoted 32 lines · Arun Parameswaran
Quoted 31 lines · Ben Hutchings
Quoted 22 lines · Arun Parameswaran
Quoted 13 lines · Ben Hutchings
Quoted 4 lines · Arun Parameswaran
Hi,
The patch fixes 2 issues with 'ethtool' getting/setting parametres in the do_gset() do_sset() API's.
Quoted 8 lines
I have pushed a patch to the Kernel to fix an issue in the handling of the 'ethtool' commands which got accepted.
This Kernel patch was based on Linux v4.1-rc4 and is available in:
https://github.com/Broadcom/cygnus-linux/tree/net-core-ethtool-fix-v1
The Kernel was always clearing the command from the 'ethtool' resulting in all operations to deal with PHY0. This prevents querying/setting PHY 1's settings.
Quoted 8 lines
[...]
Each net device can be associated with a single PHY at a time, and the ETHTOOL_GSET implementation should fill in the PHY address in the ethtool_cmd::phy_address field. Where there are multiple PHYs that can be connected to the net device's MAC, an ETHTOOL_SSET operation can be used to change that PHY address.
Quoted 3 lines
The above can be done by the driver when there is one PHY per MAC. In our case we have multiple PHYs controlled by the same MAC. I should have clarified this earlier, I apologize.
I understand that you can have multiple PHYs on the same MDIO bus, but not how the MAC can use them at the same time. Is this hardware level bonding? Or are multiple PHYs needed for a single link?

Quoted 6 lines

Quoted 3 lines
We have an internal switch which manages the traffic to the PHY's (ports). There is 1 PHY per external port.
The MAC is connected to the internal port of the switch.
Then you should create net devices for those external ports as well as the internal port.

We can only have one MAC address as we have only one MAC. So we prefer to have a single net device which reflects the hardware better.

Quoted 6 lines

If I understand the switchdev API rightly, the external port devices should implement the ethtool {get,set}_settings operations and the ndo_switch_parent_id_get operation. The existing net device should expose only the internal link to the switch (which presumably isn't configurable at all).

I agree, there can be an implementation like the above.
The switch can be handled in multiple ways, in our case we use one net device as we are limited by the MAC.

Since the 'ethtool' provides the 'phyad' parameter to specifically address a particular PHY, I think the 'ethtool' should allow for this, irrespective of the net implementation. Wouldn't this make the 'ethtool' flexible and detached from the network driver implementation.

Quoted 10 lines

[...]
Quoted 2 lines
But this prevents the 'ethtool' from being used to get/set data of specific PHY's.
That is fine because it is meant to manage the net device's own link (in this case, the internal port), not other switch ports.
Ben.

David Miller

maintainer

2015-06-01 21:39 UTC #9 lore raw

In reply to Ben Hutchings

From: Ben Hutchings <[email protected]>
Date: Mon, 01 Jun 2015 20:29:23 +0100

Quoted 6 lines · Arun Parameswaran

Quoted 2 lines
But this prevents the 'ethtool' from being used to get/set data of specific PHY's.
That is fine because it is meant to manage the net device's own link (in this case, the internal port), not other switch ports.

Ok I'm convinced, I will revert the change in question.

Thanks Ben.