Message-ID: <561ED270.60003@gmail.com>
Date:	Wed, 14 Oct 2015 15:08:48 -0700
From:	Florian Fainelli <f.fainelli@...il.com>
To:	Vivien Didelot <vivien.didelot@...oirfairelinux.com>,
	Ido Schimmel <idosch@...lanox.com>
CC:	Scott Feldman <sfeldma@...il.com>,
	Nikolay Aleksandrov <razor@...ckwall.org>,
	Netdev <netdev@...r.kernel.org>,
	Jiří Pírko <jiri@...nulli.us>,
	"David S. Miller" <davem@...emloft.net>,
	Nikolay Aleksandrov <nikolay@...ulusnetworks.com>,
	Elad Raz <eladr@...lanox.com>
Subject: Re: [PATCH net-next] switchdev: enforce no pvid flag in vlan ranges

On 14/10/15 11:51, Vivien Didelot wrote:
> On Oct. Wednesday 14 (42) 08:42 PM, Ido Schimmel wrote:
>> Wed, Oct 14, 2015 at 08:14:24PM IDT, sfeldma@...il.com wrote:
>>> On Wed, Oct 14, 2015 at 8:25 AM, Vivien Didelot
>>> <vivien.didelot@...oirfairelinux.com> wrote:
>>>> On Oct. Wednesday 14 (42) 09:14 AM, Ido Schimmel wrote:
>>>>> Tue, Oct 13, 2015 at 05:32:26PM IDT, vivien.didelot@...oirfairelinux.com wrote:
>>>>>> On Oct. Tuesday 13 (42) 11:31 AM, Ido Schimmel wrote:
>>>>>>> Mon, Oct 12, 2015 at 08:36:25PM IDT, vivien.didelot@...oirfairelinux.com wrote:
>>>>>>>> Hi guys,
>>>>>>>>
>>>>>>>> On Oct. Monday 12 (42) 02:01 PM, Nikolay Aleksandrov wrote:
>>>>>>>>> From: Nikolay Aleksandrov <nikolay@...ulusnetworks.com>
>>>>>>>>>
>>>>>>>>> We shouldn't allow BRIDGE_VLAN_INFO_PVID flag in VLAN ranges.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Nikolay Aleksandrov <nikolay@...ulusnetworks.com>
>>>>>>>>> ---
>>>>>>>>>  net/switchdev/switchdev.c | 3 +++
>>>>>>>>>  1 file changed, 3 insertions(+)
>>>>>>>>>
>>>>>>>>> diff --git a/net/switchdev/switchdev.c b/net/switchdev/switchdev.c
>>>>>>>>> index 6e4a4f9ad927..256c596de896 100644
>>>>>>>>> --- a/net/switchdev/switchdev.c
>>>>>>>>> +++ b/net/switchdev/switchdev.c
>>>>>>>>> @@ -720,6 +720,9 @@ static int switchdev_port_br_afspec(struct net_device *dev,
>>>>>>>>>                         if (vlan.vid_begin)
>>>>>>>>>                                 return -EINVAL;
>>>>>>>>>                         vlan.vid_begin = vinfo->vid;
>>>>>>>>> +                       /* don't allow range of pvids */
>>>>>>>>> +                       if (vlan.flags & BRIDGE_VLAN_INFO_PVID)
>>>>>>>>> +                               return -EINVAL;
>>>>>>>>>                 } else if (vinfo->flags & BRIDGE_VLAN_INFO_RANGE_END) {
>>>>>>>>>                         if (!vlan.vid_begin)
>>>>>>>>>                                 return -EINVAL;
>>>>>>>>> --
>>>>>>>>> 2.4.3
>>>>>>>>>
>>>>>>>>
>>>>>>>> Yes the patch looks good, but it is a minor check though. I hope the
>>>>>>>> subject of this thread is making sense.
>>>>>>>>
>>>>>>>> VLAN ranges seem to have been included for an UX purpose (so commands
>>>>>>>> look like Cisco IOS). We don't want to change any existing interface, so
>>>>>>>> we pushed that down to drivers, with the only valid reason that, maybe
>>>>>>>> one day, an hardware can be capable of programming a range on a per-port
>>>>>>>> basis.
>>>>>>> Hi,
>>>>>>>
>>>>>>> That's actually what we are doing in mlxsw. We can do up to 256 entries in
>>>>>>> one go. We've yet to submit this part.
>>>>>>
>>>>>> Perfect Ido, thanks for pointing this out! I'm OK with the range then.
>>>>>>
>>>>>> So there is now a very last question in my head for this, which is more
>>>>>> a matter of kernel design. Should the user be aware of such underlying
>>>>>> support? In other words, would it make sense to do this in a driver:
>>>>>>
>>>>>>    foo_port_vlan_add(struct net_device *dev,
>>>>>>                      struct switchdev_obj_port_vlan *vlan)
>>>>>>    {
>>>>>>        if (vlan->vid_begin != vlan->vid_end)
>>>>>>            return -ENOTSUPP; /* or something more relevant for user */
>>>>>>
>>>>>>        return foo_port_single_vlan_add(dev, vlan->vid_begin);
>>>>>>    }
>>>>>>
>>>>>> So drivers keep being simple, and we can easily propagate the fact that
>>>>>> one-or-all VLAN is not supportable, vs. the VLAN feature itself is not
>>>>>> implemented and must be done in software.
>>>>> I think that if you want to keep it simple, then Scott's advice from the
>>>>> previous thread is the most appropriate one. I believe the hardware you
>>>>> are using is simply not meant to support multiple 802.1Q bridges.
>>>>
>>>> You mean allowing only one Linux bridge over an hardware switch?
>>>>
>>>> It would for sure simplify how, as developers and users, we represent a
>>>> physical switch. But I am not sure how to achieve that and I don't have
>>>> strong opinions on this TBH.
>>>
>>> Hi Vivien, I think it's possible to keep switch ports on just one
>>> bridge if we do a little bit of work on the NETDEV_CHANGEUPPER
>>> notifier.  This will give you the driver-level control you want.  Do
>>> you have time to investigate?  The idea is:
>>>
>>> 1) In your driver's handler for NETDEV_CHANGEUPPER, if switch port is
>>> being added to a second bridge,then return NOTIFY_BAD.  Your driver
>>> needs to track the bridge count.
>>>
>>> 2) In __netdev_upper_dev_link(), check the return code from the
>>> call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, ...) call, and if
>>> NOTIFY_BAD, abort the linking operation (goto rollback_xxx).
>>>
>> Hi,
>>
>> We are doing something similar in mlxsw (not upstream yet). Jiri
>> introduced PRE_CHANGEUPPER, which is called from the function you
>> mentioned, but before the linking operation (so that you don't need to
>> rollback).
>>
>> If the notification is about a linking operation and the master is a
>> bridge different than the current one, then NOTIFY_BAD is returned.
> 
> Great, I'll wait for this then.
> 
> Scott, this is another good reason why we definitely need a simple
> struct device per switch chip. In addition to the port net_device
> registration, the netdev notifier is another exact same piece of code
> that both Rocker and DSA implement.
> 
>> Vivien, regarding your WAN interface question, this is something we
>> currently don't do. We don't even flood traffic from bridged ports
>> to CPU (although we can), as it can saturate the bus. Only control
>> traffic is supposed to go there.
> 
> I kinda answered it myself: a Linux bridge needs to remain a user
> abstraction of a logical group of net_device. In other words, we must
> allow physical distinct ports under the same bridge.
> 
> Below is an example of a custom router with 2 chained switch chips sw0
> and sw1, and what usage I believe we expect:
> 
>     [ Linux soft bridge "br0" which can accelerate VLAN, STP, etc.    ]
>                                                        (CPU)    (WAN)
>     [ sw0p0 sw0p1 sw0p2 ] [ sw1p0 sw1p1 sw1p2 sw1p3 ] [ eth0 ] [ eth1 ]
>                      `--DSA--'                   `-------'

Your use case is something that is fairly common with PON/GPON devices,
but AFAIR they typically implement this by having two sets of bridges,
one which spans the WAN interface and a bunch of other ports, and
another bridge which is LAN only (few ports + Wi-Fi typically). Usually
this is under the same physical switch though, so this is all about
partitioning physical ports and physical interfaces under logical groups.

By definition the WAN and LAN domains are separate logical and broadcast
domains, with separate admission control rules, STP etc. I do not think
your "br0" example should span the WAN interface in this case. Also,
with eth0 being the conduit interface, it cannot be allowed to be a
bridge member, that's something I ended up fixing in the bridge layer,
otherwise untagging does not work, but that is nitpicking.

It still makes sense though allow the creation of a "br0" device which
spans the entire set of ports and interfaces (except eth0), but not name
eth1 "WAN", just treat it as an extension in that case.

Sorry if that seems like f'ing flies, but having concise example should
help make progress on these design issues ;)
-- 
Florian
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists