MTU fixes for DSA
Subsystems (3)
Vladimir Oltean
maintainercover MTU fixes for DSA
From: Vladimir Oltean <[email protected]>
This small series aims to address the complaints about a regression in DSA that were formulated here:
Vladimir Oltean (2):
net: dsa: be compatible with DSA masters with max_mtu of 1500 or less
net: dsa: don't fail to probe if we couldn't set the MTU
net/dsa/slave.c | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-)
Signature · 1 line
2.17.1
Vladimir Oltean
maintainerPATCH 1/2 net: dsa: be compatible with DSA masters with max_mtu of 1500 or less
From: Vladimir Oltean <[email protected]>
It would be ideal if the DSA switch ports would support an MTU of 1500 bytes by default, same as any other net device. But there are 2 cases of issues with trying to do that:
- Drivers that are legitimately MTU-challenged and don't support anything larger than ETH_DATA_LEN. A very quick search shows that sungem.c is one such example - there may be many others.
- Drivers that simply don't populate netdev->max_mtu. In that case, it seems that the ether_setup function sets dev->max_mtu to a default value of ETH_DATA_LEN. And due to the above cases which really are MTU-challenged, we can't really make any guesses.
So for these cases, if the max_mtu of the master net_device is lower than 1500, use that (minus the tagger overhead) as the max MTU of the switch ports.
--- net/dsa/slave.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
net/dsa/slave.c +4 -1
@@ -1723,6 +1723,7 @@ int dsa_slave_create(struct dsa_port *port)const char *name = port->name;struct net_device *slave_dev;struct dsa_slave_priv *p; +int mtu;int ret;if (!ds->num_tx_queues)
@@ -1767,8 +1768,10 @@ int dsa_slave_create(struct dsa_port *port) p->xmit = cpu_dp->tag_ops->xmit; port->slave = slave_dev; + mtu =min_t (int, master->max_mtu - cpu_dp->tag_ops->overhead, +ETH_DATA_LEN );rtnl_lock (); - ret =dsa_slave_change_mtu (slave_dev,ETH_DATA_LEN ); + ret =dsa_slave_change_mtu (slave_dev, mtu);rtnl_unlock ();if (ret && ret != -EOPNOTSUPP ) { dev_err(ds->dev, "error %d setting MTU on port %d\n",
Signature · 1 line
2.17.1
Andrew Lunn
maintainerRe: [PATCH net 1/2] net: dsa: be compatible with DSA masters with max_mtu of 1500 or less
Quoted 10 lines · Vladimir Oltean
From: Vladimir Oltean <[email protected]>
It would be ideal if the DSA switch ports would support an MTU of 1500 bytes by default, same as any other net device. But there are 2 cases of issues with trying to do that:
- Drivers that are legitimately MTU-challenged and don't support anything larger than
ETH_DATA_LEN. A very quick search shows thatsungem.cis one such example - there may be many others.
Quoted 8 lines
- Drivers that simply don't populate
netdev->max_mtu. In that case, it seems that theether_setupfunction setsdev->max_mtuto a default value ofETH_DATA_LEN. And due to the above cases which really are MTU-challenged, we can't really make any guesses.So for these cases, if the
max_mtuof the masternet_deviceis lower than 1500, use that (minus the tagger overhead) as the max MTU of the switch ports.
I don't like this. I suspect this will also break in subtle ways.
Please go back to the original behaviour. Make the call to request the minimum needed for DSA. And don't care at all if it fails. For jumbo frames then you can error out.
Andrew
Vladimir Oltean
maintainerRe: [PATCH net 1/2] net: dsa: be compatible with DSA masters with max_mtu of 1500 or less
Hi Andrew,
Quoted 17 lines · Andrew Lunn
Quoted 15 lines · Vladimir Oltean
From: Vladimir Oltean <[email protected]>
It would be ideal if the DSA switch ports would support an MTU of 1500 bytes by default, same as any other net device. But there are 2 cases of issues with trying to do that:
- Drivers that are legitimately MTU-challenged and don't support anything larger than
ETH_DATA_LEN. A very quick search shows thatsungem.cis one such example - there may be many others.- Drivers that simply don't populate
netdev->max_mtu. In that case, it seems that theether_setupfunction setsdev->max_mtuto a default value ofETH_DATA_LEN. And due to the above cases which really are MTU-challenged, we can't really make any guesses.
Quoted 8 lines
Quoted 3 lines
So for these cases, if the
max_mtuof the masternet_deviceis lower than 1500, use that (minus the tagger overhead) as the max MTU of the switch ports.I don't like this. I suspect this will also break in subtle ways.
Please go back to the original behaviour. Make the call to request the minimum needed for DSA.
In what sense "minimum needed"? It is minimum needed. If master->max_mtu is 1500, the MTU will be set to 1496.
Quoted 2 lines
And don't care at all if it fails. For jumbo
frames then you can error out.
Yes, that is patch 2/2.
Quoted 2 lines
Andrew
Thanks,
-Vladimir
Andrew Lunn
maintainerRe: [PATCH net 1/2] net: dsa: be compatible with DSA masters with max_mtu of 1500 or less
Quoted 24 lines · Vladimir Oltean
Hi Andrew,
Quoted 21 lines · Andrew Lunn
Quoted 18 lines · Vladimir Oltean
From: Vladimir Oltean <[email protected]>
It would be ideal if the DSA switch ports would support an MTU of 1500 bytes by default, same as any other net device. But there are 2 cases of issues with trying to do that:
- Drivers that are legitimately MTU-challenged and don't support anything larger than
ETH_DATA_LEN. A very quick search shows thatsungem.cis one such example - there may be many others.- Drivers that simply don't populate
netdev->max_mtu. In that case, it seems that theether_setupfunction setsdev->max_mtuto a default value ofETH_DATA_LEN. And due to the above cases which really are MTU-challenged, we can't really make any guesses.So for these cases, if the
max_mtuof the masternet_deviceis lower than 1500, use that (minus the tagger overhead) as the max MTU of the switch ports.
Quoted 7 lines
Quoted 4 lines
I don't like this. I suspect this will also break in subtle ways.
Please go back to the original behaviour. Make the call to request the minimum needed for DSA.
In what sense "minimum needed"? It is minimum needed. If
master->max_mtuis 1500, the MTU will be set to 1496.
Ah, sorry. This is the slave. I was thinking it was the master.
We have always assumed the slave can send normal sized frames, independent of what the master supports. This is just a follow on from the fact we ignore errors setting the MTU on the master for the DSA overhead for normal size frames. So don't set the MTU to 1496, leave it at 1500. For all working boards out in the wild, 1500 will work.
Andrew
Vladimir Oltean
maintainerRe: [PATCH net 1/2] net: dsa: be compatible with DSA masters with max_mtu of 1500 or less
Quoted 36 lines · Andrew Lunn
Quoted 24 lines · Vladimir Oltean
Hi Andrew,
Quoted 21 lines · Andrew Lunn
Quoted 18 lines · Vladimir Oltean
From: Vladimir Oltean <[email protected]>
It would be ideal if the DSA switch ports would support an MTU of 1500 bytes by default, same as any other net device. But there are 2 cases of issues with trying to do that:
- Drivers that are legitimately MTU-challenged and don't support anything larger than
ETH_DATA_LEN. A very quick search shows thatsungem.cis one such example - there may be many others.- Drivers that simply don't populate
netdev->max_mtu. In that case, it seems that theether_setupfunction setsdev->max_mtuto a default value ofETH_DATA_LEN. And due to the above cases which really are MTU-challenged, we can't really make any guesses.So for these cases, if the
max_mtuof the masternet_deviceis lower than 1500, use that (minus the tagger overhead) as the max MTU of the switch ports.Quoted 7 lines
Quoted 4 lines
I don't like this. I suspect this will also break in subtle ways.
Please go back to the original behaviour. Make the call to request the minimum needed for DSA.
In what sense "minimum needed"? It is minimum needed. If
master->max_mtuis 1500, the MTU will be set to 1496.Ah, sorry. This is the slave. I was thinking it was the master.
Quoted 7 lines
We have always assumed the slave can send normal sized frames, independent of what the master supports. This is just a follow on from the fact we ignore errors setting the MTU on the master for the DSA overhead for normal size frames. So don't set the MTU to 1496, leave it at 1500. For all working boards out in the wild, 1500 will work.
Andrew
Does iperf3 TCP work on your Vybrid board with the master MTU equal to the slave MTU equal to 1500 (without IP fragmentation, that is)? If it does, ok, this patch can maybe be dropped.
qca7000 doesn't support packets larger than 1500 MTU either, neither does broadcom b44, and neither do probably more adapters which I didn't find now.
Why would I not limit the MTU to 1496 if the master's max_mtu is 1500? Just to provoke more warnings in people's logs? I just don't want to have further issue reports, but I am basically reading your reply as "don't try to be correct". If you don't like this behavior for FEC, you can set its max_mtu since it doesn't do that by itself.
Thanks,
-Vladimir
Andrew Lunn
maintainerRe: [PATCH net 1/2] net: dsa: be compatible with DSA masters with max_mtu of 1500 or less
Quoted 11 lines
Quoted 7 lines
We have always assumed the slave can send normal sized frames, independent of what the master supports. This is just a follow on from the fact we ignore errors setting the MTU on the master for the DSA overhead for normal size frames. So don't set the MTU to 1496, leave it at 1500. For all working boards out in the wild, 1500 will work.
Andrew
Does iperf3 TCP work on your Vybrid board with the master MTU equal to the slave MTU equal to 1500 (without IP fragmentation, that is)? If it does, ok, this patch can maybe be dropped.
Yes it does.
Quoted 3 lines
qca7000 doesn't support packets larger than 1500 MTU either, neither does broadcom b44, and neither do probably more adapters which I didn't find now.
And unfortunately, there are probably a similar number of devices which do work, and don't have correct MTU settings. So long as the IP stack does MTU correctly, it does not matter what the network device does with frames bigger than the MTU. And so network devices whos MTU handling is wrong never causes issues, and so never get detected and fixed.
We have to live in a world where trying to be correct is going to cause regressions because of poor decisions in the past. But we can do jumbo correctly, since it is new, it cannot regress. And doing jumbo correct will help find some of these network drivers which do MTU settings wrong.
Andrew
Vladimir Oltean
maintainerPATCH 2/2 net: dsa: don't fail to probe if we couldn't set the MTU
In reply to Vladimir OlteanFrom: Vladimir Oltean <[email protected]>
There is no reason to fail the probing of the switch if the MTU couldn't be configured correctly (either the switch port itself, or the host port) for whatever reason. MTU-sized traffic probably won't work, sure, but we can still probably limp on and support some form of communication anyway, which the users would probably appreciate more.
--- net/dsa/slave.c | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)
net/dsa/slave.c +1 -3
@@ -1773,11 +1773,9 @@ int dsa_slave_create(struct dsa_port *port)rtnl_lock (); ret =dsa_slave_change_mtu (slave_dev, mtu);rtnl_unlock (); -if (ret && ret != -EOPNOTSUPP ) { +if (ret)dev_err (ds->dev,"error %d setting MTU on port %d\n" , ret, port->index); - goto out_free; - }netif_carrier_off (slave_dev);
Signature · 1 line
2.17.1