Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1502167 > unrolled thread

Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove usage

Back to article view | Back to linux.kernel

This discussion starts older than the indexed window; earlier articles aren't shown. The article labeled Started by below is the oldest one visible, not the original post.

Contents

  Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage Jakub Kicinski <kubakici@wp.pl> - 2016-10-17 18:30 +0200
    Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage David Miller <davem@davemloft.net> - 2016-10-17 19:00 +0200
      Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage Jakub Kicinski <kubakici@wp.pl> - 2016-10-17 19:10 +0200
        Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage Jakub Kicinski <kubakici@wp.pl> - 2016-10-17 19:10 +0200
        Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage David Miller <davem@davemloft.net> - 2016-10-17 19:20 +0200
          Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage David Miller <davem@davemloft.net> - 2016-10-17 19:30 +0200
            Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage Jarod Wilson <jarod@redhat.com> - 2016-10-17 22:10 +0200
              Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage Jarod Wilson <jarod@redhat.com> - 2016-10-17 22:30 +0200
          Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove  usage Jakub Kicinski <kubakici@wp.pl> - 2016-10-17 19:30 +0200

#1502167 — Re: [PATCH v3 net-next 2/2] net: deprecate eth_change_mtu, remove usage

On Fri,  7 Oct 2016 22:04:34 -0400, Jarod Wilson wrote:
> @@ -357,6 +356,8 @@ void ether_setup(struct net_device *dev)
>  	dev->type		= ARPHRD_ETHER;
>  	dev->hard_header_len 	= ETH_HLEN;
>  	dev->mtu		= ETH_DATA_LEN;
> +	dev->min_mtu		= ETH_MIN_MTU;
> +	dev->max_mtu		= ETH_DATA_LEN;
>  	dev->addr_len		= ETH_ALEN;
>  	dev->tx_queue_len	= 1000;	/* Ethernet wants good queues */
>  	dev->flags		= IFF_BROADCAST|IFF_MULTICAST;

This chunk seems to be breaking MTUs > 1500 for me.

On Fri,  7 Oct 2016 22:04:33 -0400, Jarod Wilson wrote:
> diff --git a/net/core/dev.c b/net/core/dev.c
> index f1fe26f..f376639 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -6499,9 +6499,18 @@ int dev_set_mtu(struct net_device *dev, int new_mtu)
>  	if (new_mtu == dev->mtu)
>  		return 0;
>  
> -	/*	MTU must be positive.	 */
> -	if (new_mtu < 0)
> +	/* MTU must be positive, and in range */
> +	if (new_mtu < 0 || new_mtu < dev->min_mtu) {
> +		net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
> +				    dev->name, new_mtu, dev->min_mtu);
>  		return -EINVAL;
> +	}
> +
> +	if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
> +		net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
> +				    dev->name, new_mtu, dev->min_mtu);
> +		return -EINVAL;
> +	}

Please correct me if I'm wrong but it seems like we are now limiting
_all_ ethernet drivers to ETH_DATA_LEN in net-next.

[toc] | [next] | [standalone]

#1502231

From: Jakub Kicinski <kubakici@wp.pl>
Date: Mon, 17 Oct 2016 17:20:06 +0100

> Please correct me if I'm wrong but it seems like we are now limiting
> _all_ ethernet drivers to ETH_DATA_LEN in net-next.

No, because the driver can increase the netdev->max_mtu value as needed.

[toc] | [prev] | [next] | [standalone]

#1502265

On Mon, 17 Oct 2016 12:49:54 -0400 (EDT), David Miller wrote:
> From: Jakub Kicinski <kubakici@wp.pl>
> Date: Mon, 17 Oct 2016 17:20:06 +0100
> 
> > Please correct me if I'm wrong but it seems like we are now limiting
> > _all_ ethernet drivers to ETH_DATA_LEN in net-next.  
> 
> No, because the driver can increase the netdev->max_mtu value as needed.

But since almost no driver is doing that, yet, right now in net-next
jumbo frames are not possible, no?  I thought the idea was the leave
the value at 0 so drivers can opt-in as needed but since setup_ether()
is initializing to 1500 now all ethernet driver get a default of
limiting to 1500.

IOW this patch made checks which were done only in eth_change_mtu()
mandatory for all drivers.

[toc] | [prev] | [next] | [standalone]

#1502275

On Mon, 17 Oct 2016 18:00:27 +0100, Jakub Kicinski wrote:
> On Mon, 17 Oct 2016 12:49:54 -0400 (EDT), David Miller wrote:
> > From: Jakub Kicinski <kubakici@wp.pl>
> > Date: Mon, 17 Oct 2016 17:20:06 +0100
> >   
> > > Please correct me if I'm wrong but it seems like we are now limiting
> > > _all_ ethernet drivers to ETH_DATA_LEN in net-next.    
> > 
> > No, because the driver can increase the netdev->max_mtu value as needed.  
> 
> But since almost no driver is doing that, yet, right now in net-next
> jumbo frames are not possible, no?  I thought the idea was the leave
> the value at 0 so drivers can opt-in as needed but since setup_ether()
> is initializing to 1500 now all ethernet driver get a default of
> limiting to 1500.
> 
> IOW this patch made checks which were done only in eth_change_mtu()
> mandatory for all drivers.

all ethernet drivers to be clear

[toc] | [prev] | [next] | [standalone]

#1502285

From: Jakub Kicinski <kubakici@wp.pl>
Date: Mon, 17 Oct 2016 18:00:27 +0100

> On Mon, 17 Oct 2016 12:49:54 -0400 (EDT), David Miller wrote:
>> From: Jakub Kicinski <kubakici@wp.pl>
>> Date: Mon, 17 Oct 2016 17:20:06 +0100
>> 
>> > Please correct me if I'm wrong but it seems like we are now limiting
>> > _all_ ethernet drivers to ETH_DATA_LEN in net-next.  
>> 
>> No, because the driver can increase the netdev->max_mtu value as needed.
> 
> But since almost no driver is doing that, yet, right now in net-next
> jumbo frames are not possible, no?  I thought the idea was the leave
> the value at 0 so drivers can opt-in as needed but since setup_ether()
> is initializing to 1500 now all ethernet driver get a default of
> limiting to 1500.
> 
> IOW this patch made checks which were done only in eth_change_mtu()
> mandatory for all drivers.

The conversions he made were in cases where the driver's method was doing
exactly the same thing eth_change_mtu() does not.

He strictly worked to keep the behavior identical compared to before his
changes, please read his patches carefully.

[toc] | [prev] | [next] | [standalone]

#1502292

From: Jakub Kicinski <kubakici@wp.pl>
Date: Mon, 17 Oct 2016 18:20:49 +0100

> Hm.  I must be missing something really obvious.  I just booted
> net-next an hour ago and couldn't set MTU to anything larger than 1500
> on either nfp or igb.  As far as I can read the code it will set the
> max_mtu to 1500 in setup_ether() but none of the jumbo-capable drivers
> had been touched by Jarod so far...

Indeed.

Jarod, this doesn't work.

I guess the idea was that if the driver overrides ndo_change_mtu and
enforeced it's limits there, the driver would still work after your
changes.

But that isn't what is happening, look at the IGB example.

It uses ether_setup(), which sets those new defaults, but now when
the MTU is changed you enforce those default min/max before the
driver's ->ndo_change_mtu() has a change to step in front and make
the decision on it's own.

This means your changes pretty much did indeed break a lot of
drivers's ability to set larger than a 1500 byte MTU.

[toc] | [prev] | [next] | [standalone]

#1502422

On Mon, Oct 17, 2016 at 01:25:53PM -0400, David Miller wrote:
> From: Jakub Kicinski <kubakici@wp.pl>
> Date: Mon, 17 Oct 2016 18:20:49 +0100
> 
> > Hm.  I must be missing something really obvious.  I just booted
> > net-next an hour ago and couldn't set MTU to anything larger than 1500
> > on either nfp or igb.  As far as I can read the code it will set the
> > max_mtu to 1500 in setup_ether() but none of the jumbo-capable drivers
> > had been touched by Jarod so far...
> 
> Indeed.
> 
> Jarod, this doesn't work.
> 
> I guess the idea was that if the driver overrides ndo_change_mtu and
> enforeced it's limits there, the driver would still work after your
> changes.
> 
> But that isn't what is happening, look at the IGB example.
> 
> It uses ether_setup(), which sets those new defaults, but now when
> the MTU is changed you enforce those default min/max before the
> driver's ->ndo_change_mtu() has a change to step in front and make
> the decision on it's own.
> 
> This means your changes pretty much did indeed break a lot of
> drivers's ability to set larger than a 1500 byte MTU.

Argh. Yeah, I see it now. I was primarily operating with the follow-on
patches also in play, which do touch all the ethernet drivers and set
max_mtu to match current behavior, didn't consider the max_mtu case where
only the initial patches were applied and the follow-on ones weren't. I've
sent that set, which should theoretically make this problem go away, but I
can also try to rework things if need be to restore intermediate jumbo
frames functionality. (And there are actually non-ethernet devices that
also call ether_setup and may or may not have larger than 1500 mtu that
aren't yet addressed by that follow-on set).

-- 
Jarod Wilson
jarod@redhat.com

[toc] | [prev] | [next] | [standalone]

#1502433

On Mon, Oct 17, 2016 at 04:07:12PM -0400, Jarod Wilson wrote:
> On Mon, Oct 17, 2016 at 01:25:53PM -0400, David Miller wrote:
> > From: Jakub Kicinski <kubakici@wp.pl>
> > Date: Mon, 17 Oct 2016 18:20:49 +0100
> > 
> > > Hm.  I must be missing something really obvious.  I just booted
> > > net-next an hour ago and couldn't set MTU to anything larger than 1500
> > > on either nfp or igb.  As far as I can read the code it will set the
> > > max_mtu to 1500 in setup_ether() but none of the jumbo-capable drivers
> > > had been touched by Jarod so far...
> > 
> > Indeed.
> > 
> > Jarod, this doesn't work.
> > 
> > I guess the idea was that if the driver overrides ndo_change_mtu and
> > enforeced it's limits there, the driver would still work after your
> > changes.
> > 
> > But that isn't what is happening, look at the IGB example.
> > 
> > It uses ether_setup(), which sets those new defaults, but now when
> > the MTU is changed you enforce those default min/max before the
> > driver's ->ndo_change_mtu() has a change to step in front and make
> > the decision on it's own.
> > 
> > This means your changes pretty much did indeed break a lot of
> > drivers's ability to set larger than a 1500 byte MTU.
> 
> Argh. Yeah, I see it now. I was primarily operating with the follow-on
> patches also in play, which do touch all the ethernet drivers and set
> max_mtu to match current behavior, didn't consider the max_mtu case where
> only the initial patches were applied and the follow-on ones weren't. I've
> sent that set, which should theoretically make this problem go away, but I
> can also try to rework things if need be to restore intermediate jumbo
> frames functionality. (And there are actually non-ethernet devices that
> also call ether_setup and may or may not have larger than 1500 mtu that
> aren't yet addressed by that follow-on set).

Looks like the simplest thing to do is going to be to revert a52ad514, and
only make that change after all callers of ether_setup() are setting
min/max_mtu themselves as needed, then it can be reintroduced.

-- 
Jarod Wilson
jarod@redhat.com

[toc] | [prev] | [next] | [standalone]

#1502297

On Mon, 17 Oct 2016 13:15:13 -0400 (EDT), David Miller wrote:
> From: Jakub Kicinski <kubakici@wp.pl>
> Date: Mon, 17 Oct 2016 18:00:27 +0100
> 
> > On Mon, 17 Oct 2016 12:49:54 -0400 (EDT), David Miller wrote:  
> >> From: Jakub Kicinski <kubakici@wp.pl>
> >> Date: Mon, 17 Oct 2016 17:20:06 +0100
> >>   
> >> > Please correct me if I'm wrong but it seems like we are now limiting
> >> > _all_ ethernet drivers to ETH_DATA_LEN in net-next.    
> >> 
> >> No, because the driver can increase the netdev->max_mtu value as needed.  
> > 
> > But since almost no driver is doing that, yet, right now in net-next
> > jumbo frames are not possible, no?  I thought the idea was the leave
> > the value at 0 so drivers can opt-in as needed but since setup_ether()
> > is initializing to 1500 now all ethernet driver get a default of
> > limiting to 1500.
> > 
> > IOW this patch made checks which were done only in eth_change_mtu()
> > mandatory for all drivers.  
> 
> The conversions he made were in cases where the driver's method was doing
> exactly the same thing eth_change_mtu() does not.
> 
> He strictly worked to keep the behavior identical compared to before his
> changes, please read his patches carefully.

Hm.  I must be missing something really obvious.  I just booted
net-next an hour ago and couldn't set MTU to anything larger than 1500
on either nfp or igb.  As far as I can read the code it will set the
max_mtu to 1500 in setup_ether() but none of the jumbo-capable drivers
had been touched by Jarod so far...

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web