Discussion:
[RFC PATCH v0 1/2] net: bridge: propagate FDB table into hardware
John Fastabend
2012-02-09 03:22:06 UTC
Permalink
Propagate software FDB table into hardware uc, mc lists when
the NETIF_F_HW_FDB is set.

This resolves the case below where an embedded switch is used
in hardware to do inter-VF or VF-PF switching. This patch
pushes the FDB entry (specifically the MAC address) into the
embedded switch with dev_add_uc and dev_add_mc so the switch
"learns" about the software bridge.


veth0 veth2
| |
------------
| bridge0 | <---- software bridging
------------
/
/
ethx.y ethx
VF PF
\ \ <---- propagate FDB entries to HW
\ \
--------------------
| Embedded Bridge | <---- hardware offloaded switching
--------------------

This is only an RFC couple more changes are needed.

(1) Optimize HW FDB set/del to only walk list if an FDB offloaded
device is attached. Or decide it doesn't matter from unlikely()
path.

(2) Is it good enough to just call dev_uc_{add|del} or
dev_mc_{add|del}? Or do some devices really need a new netdev
callback to do this operation correctly. I think it should be
good enough as is.

(3) wrapped list walk in rcu_read_lock() just in case maybe every
case is already inside rcu_read_lock()/unlock().

Also this is in response to this thread regarding the macvlan and
exposing rx filters posting now to see if folks think this is the
right idea and if it will resolve at least the bridge case.

http://lists.openwall.net/netdev/2011/11/08/135

Signed-off-by: John Fastabend <***@intel.com>
---

include/linux/netdev_features.h | 2 ++
net/bridge/br_fdb.c | 34 ++++++++++++++++++++++++++++++++++
2 files changed, 36 insertions(+), 0 deletions(-)

diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
index 77f5202..5936fae 100644
--- a/include/linux/netdev_features.h
+++ b/include/linux/netdev_features.h
@@ -55,6 +55,8 @@ enum {
NETIF_F_NOCACHE_COPY_BIT, /* Use no-cache copyfromuser */
NETIF_F_LOOPBACK_BIT, /* Enable loopback */

+ NETIF_F_HW_FDB, /* Hardware supports switching */
+
/*
* Add your fresh new feature above and remember to update
* netdev_features_strings[] in net/core/ethtool.c and maybe
diff --git a/net/bridge/br_fdb.c b/net/bridge/br_fdb.c
index 5ba0c84..4cc545b 100644
--- a/net/bridge/br_fdb.c
+++ b/net/bridge/br_fdb.c
@@ -81,9 +81,26 @@ static void fdb_rcu_free(struct rcu_head *head)
kmem_cache_free(br_fdb_cache, ent);
}

+static void fdb_hw_delete(struct net_bridge *br,
+ struct net_bridge_fdb_entry *fdb)
+{
+ struct net_bridge_port *op;
+
+ rcu_read_lock();
+ list_for_each_entry_rcu(op, &br->port_list, list) {
+ struct net_device *dev = op->dev;
+
+ if ((dev->features & NETIF_F_HW_FDB) &&
+ dev != fdb->dst->dev)
+ dev_uc_del(dev, fdb->addr.addr);
+ }
+ rcu_read_unlock();
+}
+
static void fdb_delete(struct net_bridge *br, struct net_bridge_fdb_entry *f)
{
hlist_del_rcu(&f->hlist);
+ fdb_hw_delete(br, f);
fdb_notify(br, f, RTM_DELNEIGH);
call_rcu(&f->rcu, fdb_rcu_free);
}
@@ -350,6 +367,22 @@ static struct net_bridge_fdb_entry *fdb_find_rcu(struct hlist_head *head,
return NULL;
}

+static void fdb_hw_create(struct net_bridge *br,
+ struct net_bridge_fdb_entry *fdb)
+{
+ struct net_bridge_port *op;
+
+ rcu_read_lock();
+ list_for_each_entry_rcu(op, &br->port_list, list) {
+ struct net_device *dev = op->dev;
+
+ if ((dev->features & NETIF_F_HW_FDB) &&
+ dev != fdb->dst->dev)
+ dev_uc_add(dev, fdb->addr.addr);
+ }
+ rcu_read_unlock();
+}
+
static struct net_bridge_fdb_entry *fdb_create(struct hlist_head *head,
struct net_bridge_port *source,
const unsigned char *addr)
@@ -363,6 +396,7 @@ static struct net_bridge_fdb_entry *fdb_create(struct hlist_head *head,
fdb->is_local = 0;
fdb->is_static = 0;
fdb->updated = fdb->used = jiffies;
+ fdb_hw_create(source->br, fdb);
hlist_add_head_rcu(&fdb->hlist, head);
}
return fdb;
John Fastabend
2012-02-09 03:22:11 UTC
Permalink
Add support for NETIF_F_HW_FDB flag when SR-IOV is enabled. This
allows the bridge to push fdb entries into the hardware so the
VF can communicate with virtual devices attached to the bridge.


veth0 veth2
| |
------------
| bridge0 | <---- software bridging
------------
/
/
ethx.y ethx
VF PF
\ \ <---- propagate FDB entries to HW
\ \
--------------------
| Embedded Bridge | <---- hardware offloaded switching
--------------------


Signed-off-by: John Fastabend <***@intel.com>
---

drivers/net/ethernet/intel/ixgbe/ixgbe_main.c | 35 +++++++++++++++++--------
1 files changed, 24 insertions(+), 11 deletions(-)

diff --git a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
index ecc46ce..66261fa 100644
--- a/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
+++ b/drivers/net/ethernet/intel/ixgbe/ixgbe_main.c
@@ -3207,6 +3207,11 @@ static int ixgbe_write_uc_addr_list(struct net_device *netdev)
netdev_for_each_uc_addr(ha, netdev) {
if (!rar_entries)
break;
+
+ netif_printk(adapter, hw, KERN_DEBUG, adapter->netdev,
+ "%s %s: write vfn %i %pM\n",
+ __func__, netdev->name, vfn, ha->addr);
+
hw->mac.ops.set_rar(hw, rar_entries--, ha->addr,
vfn, IXGBE_RAH_AV);
count++;
@@ -3268,16 +3273,17 @@ void ixgbe_set_rx_mode(struct net_device *netdev)
}
ixgbe_vlan_filter_enable(adapter);
hw->addr_ctrl.user_set_promisc = false;
- /*
- * Write addresses to available RAR registers, if there is not
- * sufficient space to store all the addresses then enable
- * unicast promiscuous mode
- */
- count = ixgbe_write_uc_addr_list(netdev);
- if (count < 0) {
- fctrl |= IXGBE_FCTRL_UPE;
- vmolr |= IXGBE_VMOLR_ROPE;
- }
+ }
+
+ /*
+ * Write addresses to available RAR registers, if there is not
+ * sufficient space to store all the addresses then enable
+ * unicast promiscuous mode
+ */
+ count = ixgbe_write_uc_addr_list(netdev);
+ if (count < 0) {
+ fctrl |= IXGBE_FCTRL_UPE;
+ vmolr |= IXGBE_VMOLR_ROPE;
}

if (adapter->num_vfs) {
@@ -7214,6 +7220,10 @@ static netdev_features_t ixgbe_fix_features(struct net_device *netdev,
e_info(probe, "rx-usecs set too low, not enabling RSC\n");
}

+ /* Only use offloaded FDB if SR-IOV is enabled */
+ if (!(adapter->flags & IXGBE_FLAG_SRIOV_ENABLED))
+ data &= ~NETIF_F_HW_FDB;
+
return data;
}

@@ -7549,9 +7559,12 @@ static int __devinit ixgbe_probe(struct pci_dev *pdev,

netdev->priv_flags |= IFF_UNICAST_FLT;

- if (adapter->flags & IXGBE_FLAG_SRIOV_ENABLED)
+ if (adapter->flags & IXGBE_FLAG_SRIOV_ENABLED) {
adapter->flags &= ~(IXGBE_FLAG_RSS_ENABLED |
IXGBE_FLAG_DCB_ENABLED);
+ netdev->hw_features |= NETIF_F_HW_FDB;
+ netdev->features |= NETIF_F_HW_FDB;
+ }

#ifdef CONFIG_IXGBE_DCB
netdev->dcbnl_ops = &dcbnl_ops;

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Stephen Hemminger
2012-02-09 04:36:27 UTC
Permalink
On Wed, 08 Feb 2012 19:22:06 -0800
John Fastabend <***@intel.com> wrote:

> Propagate software FDB table into hardware uc, mc lists when
> the NETIF_F_HW_FDB is set.
>
> This resolves the case below where an embedded switch is used
> in hardware to do inter-VF or VF-PF switching. This patch
> pushes the FDB entry (specifically the MAC address) into the
> embedded switch with dev_add_uc and dev_add_mc so the switch
> "learns" about the software bridge.
>
>
> veth0 veth2
> | |
> ------------
> | bridge0 | <---- software bridging
> ------------
> /
> /
> ethx.y ethx
> VF PF
> \ \ <---- propagate FDB entries to HW
> \ \
> --------------------
> | Embedded Bridge | <---- hardware offloaded switching
> --------------------
>
> This is only an RFC couple more changes are needed.
>
> (1) Optimize HW FDB set/del to only walk list if an FDB offloaded
> device is attached. Or decide it doesn't matter from unlikely()
> path.
>
> (2) Is it good enough to just call dev_uc_{add|del} or
> dev_mc_{add|del}? Or do some devices really need a new netdev
> callback to do this operation correctly. I think it should be
> good enough as is.
>
> (3) wrapped list walk in rcu_read_lock() just in case maybe every
> case is already inside rcu_read_lock()/unlock().
>
> Also this is in response to this thread regarding the macvlan and
> exposing rx filters posting now to see if folks think this is the
> right idea and if it will resolve at least the bridge case.
>
> http://lists.openwall.net/netdev/2011/11/08/135
>
> Signed-off-by: John Fastabend <***@intel.com>
> ---
>
> include/linux/netdev_features.h | 2 ++
> net/bridge/br_fdb.c | 34 ++++++++++++++++++++++++++++++++++
> 2 files changed, 36 insertions(+), 0 deletions(-)
>
> diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
> index 77f5202..5936fae 100644

Rather than yet another device feature, I would rather use netlink_notifier
callback. The notifier is more general and generic without messing with internals
of bridge.

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-09 17:36:47 UTC
Permalink
On 2/8/2012 8:36 PM, Stephen Hemminger wrote:
> On Wed, 08 Feb 2012 19:22:06 -0800
> John Fastabend <***@intel.com> wrote:
>
>> Propagate software FDB table into hardware uc, mc lists when
>> the NETIF_F_HW_FDB is set.
>>
>> This resolves the case below where an embedded switch is used
>> in hardware to do inter-VF or VF-PF switching. This patch
>> pushes the FDB entry (specifically the MAC address) into the
>> embedded switch with dev_add_uc and dev_add_mc so the switch
>> "learns" about the software bridge.
>>
>>
>> veth0 veth2
>> | |
>> ------------
>> | bridge0 | <---- software bridging
>> ------------
>> /
>> /
>> ethx.y ethx
>> VF PF
>> \ \ <---- propagate FDB entries to HW
>> \ \
>> --------------------
>> | Embedded Bridge | <---- hardware offloaded switching
>> --------------------
>>
>> This is only an RFC couple more changes are needed.
>>
>> (1) Optimize HW FDB set/del to only walk list if an FDB offloaded
>> device is attached. Or decide it doesn't matter from unlikely()
>> path.
>>
>> (2) Is it good enough to just call dev_uc_{add|del} or
>> dev_mc_{add|del}? Or do some devices really need a new netdev
>> callback to do this operation correctly. I think it should be
>> good enough as is.
>>
>> (3) wrapped list walk in rcu_read_lock() just in case maybe every
>> case is already inside rcu_read_lock()/unlock().
>>
>> Also this is in response to this thread regarding the macvlan and
>> exposing rx filters posting now to see if folks think this is the
>> right idea and if it will resolve at least the bridge case.
>>
>> http://lists.openwall.net/netdev/2011/11/08/135
>>
>> Signed-off-by: John Fastabend <***@intel.com>
>> ---
>>
>> include/linux/netdev_features.h | 2 ++
>> net/bridge/br_fdb.c | 34 ++++++++++++++++++++++++++++++++++
>> 2 files changed, 36 insertions(+), 0 deletions(-)
>>
>> diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
>> index 77f5202..5936fae 100644
>
> Rather than yet another device feature, I would rather use netlink_notifier
> callback. The notifier is more general and generic without messing with internals
> of bridge.
>

But the device features makes it easy for user space to learn that the device
supports this sort of offload. Now if all SR-IOV devices support this then it
doesn't matter but I thought there were SR-IOV devices that didn't do any
switching? I'll dig through the SR-IOV drivers to check there are not too
many of them.

By netlink_notifier do you mean adding a notifier_block and using atomic_notifier_call_chain()
probably in rtnl_notify()? Then drivers could register with the notifier chain with
atomic_notifier_chain_register() and receive the events correctly. Or did I miss
some notifier chain that already exists?

Thanks,
John
Stephen Hemminger
2012-02-09 17:40:47 UTC
Permalink
On Thu, 09 Feb 2012 09:36:47 -0800
John Fastabend <***@intel.com> wrote:

> But the device features makes it easy for user space to learn that the device
> supports this sort of offload. Now if all SR-IOV devices support this then it
> doesn't matter but I thought there were SR-IOV devices that didn't do any
> switching? I'll dig through the SR-IOV drivers to check there are not too
> many of them.

If user space needs to know then the OS is not designed properly.
The purpose of the network device is to abstract all those details, and more and more
of them are bleeding through. This makes writing management applications harder and makes
things dependent on features that may or may not be present. The best design is when
the change is invisible.

> By netlink_notifier do you mean adding a notifier_block and using atomic_notifier_call_chain()
> probably in rtnl_notify()? Then drivers could register with the notifier chain with
> atomic_notifier_chain_register() and receive the events correctly. Or did I miss
> some notifier chain that already exists?

Yes. that is what I mean. The callbacks you need may or may not already be present.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-09 17:52:55 UTC
Permalink
On 2/9/2012 9:40 AM, Stephen Hemminger wrote:
> On Thu, 09 Feb 2012 09:36:47 -0800
> John Fastabend <***@intel.com> wrote:
>
>> But the device features makes it easy for user space to learn that the device
>> supports this sort of offload. Now if all SR-IOV devices support this then it
>> doesn't matter but I thought there were SR-IOV devices that didn't do any
>> switching? I'll dig through the SR-IOV drivers to check there are not too
>> many of them.
>
> If user space needs to know then the OS is not designed properly.
> The purpose of the network device is to abstract all those details, and more and more
> of them are bleeding through. This makes writing management applications harder and makes
> things dependent on features that may or may not be present. The best design is when
> the change is invisible.
>

Agreed.

>> By netlink_notifier do you mean adding a notifier_block and using atomic_notifier_call_chain()
>> probably in rtnl_notify()? Then drivers could register with the notifier chain with
>> atomic_notifier_chain_register() and receive the events correctly. Or did I miss
>> some notifier chain that already exists?
>
> Yes. that is what I mean. The callbacks you need may or may not already be present.


OK thanks I'll put together an update here shortly.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
jamal
2012-02-09 21:11:34 UTC
Permalink
On Thu, 2012-02-09 at 09:52 -0800, John Fastabend wrote:

> >> By netlink_notifier do you mean adding a notifier_block and using atomic_notifier_call_chain()
> >> probably in rtnl_notify()? Then drivers could register with the notifier chain with
> >> atomic_notifier_chain_register() and receive the events correctly. Or did I miss
> >> some notifier chain that already exists?
> >
> > Yes. that is what I mean. The callbacks you need may or may not already be present.

I'll go one step further.
This stuff shouldnt be in the kernel at all.
The disadvantage is you need a user space app to update the hardware.
i.e, the same mechanism should be usable for either a switch embedded
in a NIC or a standalone hardware switch (with/out the s/ware bridge
presence)

cheers,
jamal
John Fastabend
2012-02-10 02:14:46 UTC
Permalink
On 2/9/2012 1:11 PM, jamal wrote:
> On Thu, 2012-02-09 at 09:52 -0800, John Fastabend wrote:
>
>>>> By netlink_notifier do you mean adding a notifier_block and using atomic_notifier_call_chain()
>>>> probably in rtnl_notify()? Then drivers could register with the notifier chain with
>>>> atomic_notifier_chain_register() and receive the events correctly. Or did I miss
>>>> some notifier chain that already exists?
>>>
>>> Yes. that is what I mean. The callbacks you need may or may not already be present.
>
> I'll go one step further.
> This stuff shouldnt be in the kernel at all.
> The disadvantage is you need a user space app to update the hardware.
> i.e, the same mechanism should be usable for either a switch embedded
> in a NIC or a standalone hardware switch (with/out the s/ware bridge
> presence)
>
> cheers,
> jamal
>

Hi Jamal,

The user space app in this case would listen for FDB updates to the SW
bridge and then mirror them at the embedded NIC. In this case it seems
easier to just add a notifier chain and let the kernel keep these in
sync. Otherwise we need a daemon in user space to replicate these.

On the other hand if you could make the same RTM_NEWNEIGH, RTM_DELNEIGH,
and RTM_GETNEIGH work for the bridge, embedded bridge, and macvlan you
would have one common interface to drive these. But the bridge already
has this protocol/msgtype so that would require either some demux or
new protocol/msgtype pairs to be created.

Let me think on it. I'm tempted by the simplicity of adding notifier
hooks though.

.John


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-10 04:14:52 UTC
Permalink
On 2/9/2012 6:14 PM, John Fastabend wrote:
> On 2/9/2012 1:11 PM, jamal wrote:
>> On Thu, 2012-02-09 at 09:52 -0800, John Fastabend wrote:
>>
>>>>> By netlink_notifier do you mean adding a notifier_block and using atomic_notifier_call_chain()
>>>>> probably in rtnl_notify()? Then drivers could register with the notifier chain with
>>>>> atomic_notifier_chain_register() and receive the events correctly. Or did I miss
>>>>> some notifier chain that already exists?
>>>>
>>>> Yes. that is what I mean. The callbacks you need may or may not already be present.
>>
>> I'll go one step further.
>> This stuff shouldnt be in the kernel at all.
>> The disadvantage is you need a user space app to update the hardware.
>> i.e, the same mechanism should be usable for either a switch embedded
>> in a NIC or a standalone hardware switch (with/out the s/ware bridge
>> presence)
>>
>> cheers,
>> jamal
>>
>
> Hi Jamal,
>
> The user space app in this case would listen for FDB updates to the SW
> bridge and then mirror them at the embedded NIC. In this case it seems
> easier to just add a notifier chain and let the kernel keep these in
> sync. Otherwise we need a daemon in user space to replicate these.
>
> On the other hand if you could make the same RTM_NEWNEIGH, RTM_DELNEIGH,
> and RTM_GETNEIGH work for the bridge, embedded bridge, and macvlan you
> would have one common interface to drive these. But the bridge already
> has this protocol/msgtype so that would require either some demux or
> new protocol/msgtype pairs to be created.
>
> Let me think on it. I'm tempted by the simplicity of adding notifier
> hooks though.
>
> .John
>

Actually because the bridge is adding/removing fdb entries dynamically
maybe its best this gets done in kernel. Here's the example case,


---------- ---------
| ethx.y | <---- E | veth0 | <--- A
---------- ---------
| |
| |
| |
| --------------
| | SW Bridge | <--- B
| --------------
| |
| |
| ---------
| | eth0 | <--- C
| ---------
| |
-----------------------------------
| embedded switch | <--- D
-----------------------------------
|
|
G

With the flow by letters above hope this is not too difficult to follow.

(A) veth0 a virtual device transmits packet destined for ethx.y
(B) SW bridge receives frames and updates FDB flooding to C
(C) eth0 the PF in this case sends the frame to the HW backed by the
embedded bridge
(D) The HW embedded switch has a static entry for ethx.y and forwards
the frame to the VF or if its a broadcast frame also floods it to
the wire and ethx.y
(E) ethx.y receives the frame and generates a response to the dest mac of
veth0

Now here is the potential issue,

(G) The frame transmitted from ethx.y with the destination address of
veth0 but the embedded switch is not a learning switch. If the FDB
update is done in user space its possible (likely?) that the FDB
entry for veth0 has not been added to the embedded switch yet. Now
we either have to flood the frame which is not horrible but not
ideal or worse if the embedded switch does not support flooding send
it to the wire and veth0 never receives it. If the SW bridge pushes
the FDB update down into the embedded switch the address is for
sure in the embedded switches forwarding tables and the switching
works as expected.

So to handle this case correctly its probably best IMHO to use a notifier
hook. Having a RTM_GETNEIGH for the embedded switch implemented though
would be nice for dumping the FDB of the embedded switch and SET/DEL
could be used to configure the FDB when its not being driven by the SW
switch. Of course we should try to be minimalists here.

.John
jamal
2012-02-10 15:18:31 UTC
Permalink
Hi John,

I went backwards to summarize at the top after going through your email.

TL;DR version 0.1:
you provide a good use case where it makes sense to do things in the
kernel. IMO, you could make the same arguement if your embedded switch
could do ACLs, IPv4 forwarding etc. And the kernel bloats.
I am always bigoted to move all policy control to user space instead of
bloating in the kernel.


On Thu, 2012-02-09 at 20:14 -0800, John Fastabend wrote:

> >
> > Hi Jamal,
> >
> > The user space app in this case would listen for FDB updates to the SW
> > bridge and then mirror them at the embedded NIC. In this case it seems
> > easier to just add a notifier chain and let the kernel keep these in
> > sync. Otherwise we need a daemon in user space to replicate these.
> >

A user space daemon if you need to ensure synchronization. Thats what i
meant when i said there was a "disadvantage" over the simple case when
the goal is always to synchronize.

> > On the other hand if you could make the same RTM_NEWNEIGH, RTM_DELNEIGH,
> > and RTM_GETNEIGH work for the bridge, embedded bridge, and macvlan you
> > would have one common interface to drive these. But the bridge already
> > has this protocol/msgtype so that would require either some demux or
> > new protocol/msgtype pairs to be created.
> >

The bridge is very netlink friendly these days. Given the rest of the
network stack (*NEIGH* you mention above) talks netlink to user space
it should be workable.

> > Let me think on it. I'm tempted by the simplicity of adding notifier
> > hooks though.

If something is missing bridge-side it may need to be added (as Per
Stephen's comment) - i just took it one further indicating those
notifiers need to also netlink-speak


> Actually because the bridge is adding/removing fdb entries dynamically
> maybe its best this gets done in kernel. Here's the example case,

[..]

>
> With the flow by letters above hope this is not too difficult to follow.

> (A) veth0 a virtual device transmits packet destined for ethx.y
> (B) SW bridge receives frames and updates FDB flooding to C
> (C) eth0 the PF in this case sends the frame to the HW backed by the
> embedded bridge

Following so far.
Can you have more than one PF per embedded switch? Or is the intent here
purely to do VMs/VF separation?

> (D) The HW embedded switch has a static entry for ethx.y and forwards
> the frame to the VF or if its a broadcast frame also floods it to
> the wire and ethx.y

nod.

> (E) ethx.y receives the frame and generates a response to the dest mac of
> veth0

nod.
Since you said in #D the entries in the switch are static, I am assuming
at this point neither ethx.y nor veth0 exist in the embedded FDB.

> Now here is the potential issue,
>
> (G) The frame transmitted from ethx.y with the destination address of
> veth0 but the embedded switch is not a learning switch. If the FDB
> update is done in user space its possible (likely?) that the FDB
> entry for veth0 has not been added to the embedded switch yet.

Ok, got it - so the catch here is the switch is not capable of learning.
I think this depends on where learning is done. Your intent is to
use the S/W bridge as something that does the learning for you i.e in
the kernel. This makes the s/w bridge part of MUST-have-for-this-to-run.
And that maybe the case for your use case.

What if I dont wanna run the S/W bridge at all?
Ive been making a point that with a simple knob(Stephen doesn like to
add such a knob), the SW bridge could defer learning to user space.
[This way you can add a lot of richness e.g on ACLs such as restricting
what MAC addresses etc are allowed to talk to which ones etc.].
But if bypass the s/w bridge all together and learn in user space
or have a static config in which i populate the embedded switch, i dont
see the issue.

> Now
> we either have to flood the frame which is not horrible but not
> ideal or worse if the embedded switch does not support flooding send
> it to the wire and veth0 never receives it.

If it is a switch it has to flood, no? Otherwise it sounds broken.

> If the SW bridge pushes
> the FDB update down into the embedded switch the address is for
> sure in the embedded switches forwarding tables and the switching
> works as expected.

Yes, there is a small gap between the s/w bridge learning and the
synchronization happening to the embedded nic switch. That gap gets
larger if you defer learning to user space. But like you said earlier,
during that gap packets are flooded - and do you care if the
synchronization doesnt happen immediately?

> So to handle this case correctly its probably best IMHO to use a notifier
> hook. Having a RTM_GETNEIGH for the embedded switch implemented though
> would be nice for dumping the FDB of the embedded switch and SET/DEL
> could be used to configure the FDB when its not being driven by the SW
> switch. Of course we should try to be minimalists here.

Do you need to have a different *NEIGH* than what we already have
really?

The problem with putting policies in the kernel is you are gonna keep
adding more. Bloat user space instead.

cheers,
jamal


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Stephen Hemminger
2012-02-10 16:39:17 UTC
Permalink
On Fri, 10 Feb 2012 10:18:31 -0500
jamal <***@cyberus.ca> wrote:

> Hi John,
>
> I went backwards to summarize at the top after going through your email.
>
> TL;DR version 0.1:
> you provide a good use case where it makes sense to do things in the
> kernel. IMO, you could make the same arguement if your embedded switch
> could do ACLs, IPv4 forwarding etc. And the kernel bloats.
> I am always bigoted to move all policy control to user space instead of
> bloating in the kernel.
>
>
> On Thu, 2012-02-09 at 20:14 -0800, John Fastabend wrote:
>
> > >
> > > Hi Jamal,
> > >
> > > The user space app in this case would listen for FDB updates to the SW
> > > bridge and then mirror them at the embedded NIC. In this case it seems
> > > easier to just add a notifier chain and let the kernel keep these in
> > > sync. Otherwise we need a daemon in user space to replicate these.
> > >
>
> A user space daemon if you need to ensure synchronization. Thats what i
> meant when i said there was a "disadvantage" over the simple case when
> the goal is always to synchronize.
>
> > > On the other hand if you could make the same RTM_NEWNEIGH, RTM_DELNEIGH,
> > > and RTM_GETNEIGH work for the bridge, embedded bridge, and macvlan you
> > > would have one common interface to drive these. But the bridge already
> > > has this protocol/msgtype so that would require either some demux or
> > > new protocol/msgtype pairs to be created.
> > >
>
> The bridge is very netlink friendly these days. Given the rest of the
> network stack (*NEIGH* you mention above) talks netlink to user space
> it should be workable.
>
> > > Let me think on it. I'm tempted by the simplicity of adding notifier
> > > hooks though.
>
> If something is missing bridge-side it may need to be added (as Per
> Stephen's comment) - i just took it one further indicating those
> notifiers need to also netlink-speak
>
>
> > Actually because the bridge is adding/removing fdb entries dynamically
> > maybe its best this gets done in kernel. Here's the example case,
>
> [..]
>
> >
> > With the flow by letters above hope this is not too difficult to follow.
>
> > (A) veth0 a virtual device transmits packet destined for ethx.y
> > (B) SW bridge receives frames and updates FDB flooding to C
> > (C) eth0 the PF in this case sends the frame to the HW backed by the
> > embedded bridge
>
> Following so far.
> Can you have more than one PF per embedded switch? Or is the intent here
> purely to do VMs/VF separation?
>
> > (D) The HW embedded switch has a static entry for ethx.y and forwards
> > the frame to the VF or if its a broadcast frame also floods it to
> > the wire and ethx.y
>
> nod.
>
> > (E) ethx.y receives the frame and generates a response to the dest mac of
> > veth0
>
> nod.
> Since you said in #D the entries in the switch are static, I am assuming
> at this point neither ethx.y nor veth0 exist in the embedded FDB.
>
> > Now here is the potential issue,
> >
> > (G) The frame transmitted from ethx.y with the destination address of
> > veth0 but the embedded switch is not a learning switch. If the FDB
> > update is done in user space its possible (likely?) that the FDB
> > entry for veth0 has not been added to the embedded switch yet.
>
> Ok, got it - so the catch here is the switch is not capable of learning.
> I think this depends on where learning is done. Your intent is to
> use the S/W bridge as something that does the learning for you i.e in
> the kernel. This makes the s/w bridge part of MUST-have-for-this-to-run.
> And that maybe the case for your use case.
>
> What if I dont wanna run the S/W bridge at all?
> Ive been making a point that with a simple knob(Stephen doesn like to
> add such a knob), the SW bridge could defer learning to user space.
> [This way you can add a lot of richness e.g on ACLs such as restricting
> what MAC addresses etc are allowed to talk to which ones etc.].
> But if bypass the s/w bridge all together and learn in user space
> or have a static config in which i populate the embedded switch, i dont
> see the issue.
>
> > Now
> > we either have to flood the frame which is not horrible but not
> > ideal or worse if the embedded switch does not support flooding send
> > it to the wire and veth0 never receives it.
>
> If it is a switch it has to flood, no? Otherwise it sounds broken.
>
> > If the SW bridge pushes
> > the FDB update down into the embedded switch the address is for
> > sure in the embedded switches forwarding tables and the switching
> > works as expected.
>
> Yes, there is a small gap between the s/w bridge learning and the
> synchronization happening to the embedded nic switch. That gap gets
> larger if you defer learning to user space. But like you said earlier,
> during that gap packets are flooded - and do you care if the
> synchronization doesnt happen immediately?
>
> > So to handle this case correctly its probably best IMHO to use a notifier
> > hook. Having a RTM_GETNEIGH for the embedded switch implemented though
> > would be nice for dumping the FDB of the embedded switch and SET/DEL
> > could be used to configure the FDB when its not being driven by the SW
> > switch. Of course we should try to be minimalists here.
>
> Do you need to have a different *NEIGH* than what we already have
> really?
>
> The problem with putting policies in the kernel is you are gonna keep
> adding more. Bloat user space instead.

Some related discussion points:
* the bridge needs to support control from both userspace (MSTP, TRILL, ...)
and kernel space (offload etc)
* the bridge forwarding database is simpler and different than the existing
neighbor table, don't remember the details but last time I checked it
using neighbor table in bridge would be putting square peg in round hole.


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
jamal
2012-02-13 13:54:31 UTC
Permalink
On Fri, 2012-02-10 at 08:39 -0800, Stephen Hemminger wrote:

> Some related discussion points:
> * the bridge needs to support control from both userspace (MSTP, TRILL, ...)
> and kernel space (offload etc)

I think all are pretty much covered if you let some controler (I prefer
user space) ADD/DEL/GET/Event on the fdb
TRILL really is outside the scope of this; from an encap/decap it
probably needs to be YAND (Yet another netdev) and from a control side
of things you need to just provide the above netlink ops(ADD, etC) on
the fdb and let the controller worry about things (Actually you _may_
need to have learning done outside of the kernel for TRILL)

> * the bridge forwarding database is simpler and different than the existing
> neighbor table, don't remember the details but last time I checked it
> using neighbor table in bridge would be putting square peg in round hole.

Agreed.

cheers,
jamal
John Fastabend
2012-02-13 15:13:03 UTC
Permalink
On 2/10/2012 7:18 AM, jamal wrote:
> Hi John,
>
> I went backwards to summarize at the top after going through your email.
>
> TL;DR version 0.1:
> you provide a good use case where it makes sense to do things in the
> kernel. IMO, you could make the same arguement if your embedded switch
> could do ACLs, IPv4 forwarding etc. And the kernel bloats.
> I am always bigoted to move all policy control to user space instead of
> bloating in the kernel.
>
>
> On Thu, 2012-02-09 at 20:14 -0800, John Fastabend wrote:
>
>>>
>>> Hi Jamal,
>>>
>>> The user space app in this case would listen for FDB updates to the SW
>>> bridge and then mirror them at the embedded NIC. In this case it seems
>>> easier to just add a notifier chain and let the kernel keep these in
>>> sync. Otherwise we need a daemon in user space to replicate these.
>>>
>
> A user space daemon if you need to ensure synchronization. Thats what i
> meant when i said there was a "disadvantage" over the simple case when
> the goal is always to synchronize.
>
>>> On the other hand if you could make the same RTM_NEWNEIGH, RTM_DELNEIGH,
>>> and RTM_GETNEIGH work for the bridge, embedded bridge, and macvlan you
>>> would have one common interface to drive these. But the bridge already
>>> has this protocol/msgtype so that would require either some demux or
>>> new protocol/msgtype pairs to be created.
>>>
>
> The bridge is very netlink friendly these days. Given the rest of the
> network stack (*NEIGH* you mention above) talks netlink to user space
> it should be workable.
>
>>> Let me think on it. I'm tempted by the simplicity of adding notifier
>>> hooks though.
>
> If something is missing bridge-side it may need to be added (as Per
> Stephen's comment) - i just took it one further indicating those
> notifiers need to also netlink-speak
>

Sure.

>
>> Actually because the bridge is adding/removing fdb entries dynamically
>> maybe its best this gets done in kernel. Here's the example case,
>
> [..]
>
>>
>> With the flow by letters above hope this is not too difficult to follow.
>
>> (A) veth0 a virtual device transmits packet destined for ethx.y
>> (B) SW bridge receives frames and updates FDB flooding to C
>> (C) eth0 the PF in this case sends the frame to the HW backed by the
>> embedded bridge
>
> Following so far.
> Can you have more than one PF per embedded switch? Or is the intent here
> purely to do VMs/VF separation?
>

The use case here is multiple VFs but the same solution should work with
multiple PFs as well. FDB controls should be independent of how the ports
are exposed VFs, PFs, VMDQ/queue pairs, macvlan, etc.

>> (D) The HW embedded switch has a static entry for ethx.y and forwards
>> the frame to the VF or if its a broadcast frame also floods it to
>> the wire and ethx.y
>
> nod.
>
>> (E) ethx.y receives the frame and generates a response to the dest mac of
>> veth0
>
> nod.
> Since you said in #D the entries in the switch are static, I am assuming
> at this point neither ethx.y nor veth0 exist in the embedded FDB.
>
>> Now here is the potential issue,
>>
>> (G) The frame transmitted from ethx.y with the destination address of
>> veth0 but the embedded switch is not a learning switch. If the FDB
>> update is done in user space its possible (likely?) that the FDB
>> entry for veth0 has not been added to the embedded switch yet.
>
> Ok, got it - so the catch here is the switch is not capable of learning.
> I think this depends on where learning is done. Your intent is to
> use the S/W bridge as something that does the learning for you i.e in
> the kernel. This makes the s/w bridge part of MUST-have-for-this-to-run.
> And that maybe the case for your use case.
>

This is _my_ use case today.

> What if I dont wanna run the S/W bridge at all?
> Ive been making a point that with a simple knob(Stephen doesn like to
> add such a knob), the SW bridge could defer learning to user space.
> [This way you can add a lot of richness e.g on ACLs such as restricting
> what MAC addresses etc are allowed to talk to which ones etc.].
> But if bypass the s/w bridge all together and learn in user space
> or have a static config in which i populate the embedded switch, i dont
> see the issue.

With events and ADD/DEL/GET FDB controls we can solve both cases. This also
solves Roopa's case with macvlan where he wants to add additional addresses
to macvlan ports.

>
>> Now
>> we either have to flood the frame which is not horrible but not
>> ideal or worse if the embedded switch does not support flooding send
>> it to the wire and veth0 never receives it.
>
> If it is a switch it has to flood, no? Otherwise it sounds broken.
>

Yes it should flood here, unless its acting as a 802.1Qbg VEB or VEPA.

>> If the SW bridge pushes
>> the FDB update down into the embedded switch the address is for
>> sure in the embedded switches forwarding tables and the switching
>> works as expected.
>
> Yes, there is a small gap between the s/w bridge learning and the
> synchronization happening to the embedded nic switch. That gap gets
> larger if you defer learning to user space. But like you said earlier,
> during that gap packets are flooded - and do you care if the
> synchronization doesnt happen immediately?
>

Maybe not. But the kernel already has the needed signals with one extra
hook we can save running a daemon in user space. Maybe that's not a great
argument to add kernel code though.

>> So to handle this case correctly its probably best IMHO to use a notifier
>> hook. Having a RTM_GETNEIGH for the embedded switch implemented though
>> would be nice for dumping the FDB of the embedded switch and SET/DEL
>> could be used to configure the FDB when its not being driven by the SW
>> switch. Of course we should try to be minimalists here.
>
> Do you need to have a different *NEIGH* than what we already have
> really?
>

The PF_BRIDGE:RTM_GETNEIGH,RTM_NEWNEIGH,RTM_DELNEIGH are registered in the
br_netlink_init() path. Adding notifier hooks here might be possible but
I'm wondering if its better to add new message types or tear apart the
existing bridging events. I'll play with the code some today and see what
works out better. As Stephen noted the PF_UNSPEC:RTM_XXX events in the
neighbor code are not really for bridging.

> The problem with putting policies in the kernel is you are gonna keep
> adding more. Bloat user space instead.
>

Agree policy is best left for user space.

Thanks,
John

> cheers,
> jamal
>
>

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
jamal
2012-02-14 13:18:46 UTC
Permalink
On Mon, 2012-02-13 at 07:13 -0800, John Fastabend wrote:

> The use case here is multiple VFs but the same solution should work with
> multiple PFs as well. FDB controls should be independent of how the ports
> are exposed VFs, PFs, VMDQ/queue pairs, macvlan, etc.

Makes sense.

> With events and ADD/DEL/GET FDB controls we can solve both cases. This also
> solves Roopa's case with macvlan where he wants to add additional addresses
> to macvlan ports.

Not familiar with that issue - I'll prowl the list.

> Yes it should flood here, unless its acting as a 802.1Qbg VEB or VEPA.

Ok. So there is a toggle somewhere which controls how flooding should
happen.

>
> Maybe not. But the kernel already has the needed signals with one extra
> hook we can save running a daemon in user space. Maybe that's not a great
> argument to add kernel code though.

You make a reasonable arguement to have it in the kernel but i think we
win more if we separate the control. So while i empathize, I am hoping
that youd go with the path that is hard to travel ;->

> The PF_BRIDGE:RTM_GETNEIGH,RTM_NEWNEIGH,RTM_DELNEIGH are registered in the
> br_netlink_init() path.

Hrm - hadnt paid attention to that before. Nasty.
The bridge seems to be hard-coding policy on station movement, no?
This is a good example of the qualms i have on adding things to the
kernel;->
I may not want to auto update a MAC address moving ports as part of
some policy i have. I can go and add YAK (Yet Another Knob) - but where
is the line drawn?

cheers,
jamal
John Fastabend
2012-02-14 18:57:04 UTC
Permalink
On 2/14/2012 5:18 AM, jamal wrote:
> On Mon, 2012-02-13 at 07:13 -0800, John Fastabend wrote:
>
>> The use case here is multiple VFs but the same solution should work with
>> multiple PFs as well. FDB controls should be independent of how the ports
>> are exposed VFs, PFs, VMDQ/queue pairs, macvlan, etc.
>
> Makes sense.
>
>> With events and ADD/DEL/GET FDB controls we can solve both cases. This also
>> solves Roopa's case with macvlan where she wants to add additional addresses
>> to macvlan ports.
>
> Not familiar with that issue - I'll prowl the list.

Roopa was likely on the right track here,

http://patchwork.ozlabs.org/patch/123064/

But I think the proper syntax is to use the existing PF_BRIDGE:RTM_XXX
netlink messages. And if possible drive this without extending ndo_ops.

An ideal user space interaction IMHO would look like,

[***@jf-dev1-dcblab iproute2]# ./br/br fdb add 52:e5:62:7b:57:88 dev veth10
[***@jf-dev1-dcblab iproute2]# ./br/br fdb
port mac addr flags
veth2 36:a6:35:9b:96:c4 local
veth4 aa:54:b0:7b:42:ef local
veth0 2a:e8:5c:95:6c:1b local
veth6 6e:26:d5:43:a3:36 local
veth0 f2:c1:39:76:6a:fb
veth8 4e:35:16:af:87:13 local
veth10 52:e5:62:7b:57:88 static
veth10 aa:a9:35:21:15:c4 local
[***@jf-dev1-dcblab iproute2]# ./br/br fdb add dev eth3 to 52:e5:62:7b:57:88
RTNETLINK answers: Invalid argument

Using Stephen's br tool. First command adds FDB entry to SW bridge and
if the same tool could be used to add entries to embedded bridge I think
that would be the best case. So no RTNETLINK error on the second cmd. Then
embedded FDB entries could be dumped this way also so I get a complete view
of my FDB setup across multiple sw bridges and embedded bridges.

I don't think br is part of iproute2 yet I just pulled it out of some RFC
but it works reasonably well and is intuitive enough.

>
>> Yes it should flood here, unless its acting as a 802.1Qbg VEB or VEPA.
>
> Ok. So there is a toggle somewhere which controls how flooding should
> happen.
>

Yes. The hardware has a bit to support this which is currently not exposed
to user space. That's a case where we have 'yet another knob' that needs
a clean solution. This causes real bugs today when users try to use the
macvlan devices in VEPA mode on top of SR-IOV. By the way these modes are
all part of the 802.1Qbg spec which people actually want to use with Linux
so a good clean solution is probably needed.

>>
>> Maybe not. But the kernel already has the needed signals with one extra
>> hook we can save running a daemon in user space. Maybe that's not a great
>> argument to add kernel code though.
>
> You make a reasonable arguement to have it in the kernel but i think we
> win more if we separate the control. So while i empathize, I am hoping
> that youd go with the path that is hard to travel ;->
>
>> The PF_BRIDGE:RTM_GETNEIGH,RTM_NEWNEIGH,RTM_DELNEIGH are registered in the
>> br_netlink_init() path.
>
> Hrm - hadnt paid attention to that before. Nasty.
> The bridge seems to be hard-coding policy on station movement, no?
> This is a good example of the qualms i have on adding things to the
> kernel;->
> I may not want to auto update a MAC address moving ports as part of
> some policy i have. I can go and add YAK (Yet Another Knob) - but where
> is the line drawn?
>

I have no problem with drawing the line here and trying to implement something
over PF_BRIDGE:RTM_xxx nlmsgs. I'll work with Roopa and see if we can come
up with something in the next couple days.

w.r.t. VEPA/VEB and flooding behavior we could probably have a bit to indicate
if the port is a flooding port or not. Then users could build any sort of forwarding
table they wanted OR we could just drive it through a notifier (ndo_ops?) in the
macvlan path which does VEPA today.

OK I'll try to write some actual code now that can be critiqued.

> cheers,
> jamal
>
>

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Stephen Hemminger
2012-02-14 19:05:04 UTC
Permalink
On Tue, 14 Feb 2012 10:57:04 -0800
John Fastabend <***@intel.com> wrote:

> On 2/14/2012 5:18 AM, jamal wrote:
> > On Mon, 2012-02-13 at 07:13 -0800, John Fastabend wrote:
> >
> >> The use case here is multiple VFs but the same solution should work with
> >> multiple PFs as well. FDB controls should be independent of how the ports
> >> are exposed VFs, PFs, VMDQ/queue pairs, macvlan, etc.
> >
> > Makes sense.
> >
> >> With events and ADD/DEL/GET FDB controls we can solve both cases. This also
> >> solves Roopa's case with macvlan where she wants to add additional addresses
> >> to macvlan ports.
> >
> > Not familiar with that issue - I'll prowl the list.
>
> Roopa was likely on the right track here,
>
> http://patchwork.ozlabs.org/patch/123064/
>
> But I think the proper syntax is to use the existing PF_BRIDGE:RTM_XXX
> netlink messages. And if possible drive this without extending ndo_ops.
>
> An ideal user space interaction IMHO would look like,
>
> [***@jf-dev1-dcblab iproute2]# ./br/br fdb add 52:e5:62:7b:57:88 dev veth10
> [***@jf-dev1-dcblab iproute2]# ./br/br fdb
> port mac addr flags
> veth2 36:a6:35:9b:96:c4 local
> veth4 aa:54:b0:7b:42:ef local
> veth0 2a:e8:5c:95:6c:1b local
> veth6 6e:26:d5:43:a3:36 local
> veth0 f2:c1:39:76:6a:fb
> veth8 4e:35:16:af:87:13 local
> veth10 52:e5:62:7b:57:88 static
> veth10 aa:a9:35:21:15:c4 local
> [***@jf-dev1-dcblab iproute2]# ./br/br fdb add dev eth3 to 52:e5:62:7b:57:88
> RTNETLINK answers: Invalid argument

I am going to put bridge (nameclash with br) tool into iproute2 (soon).
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-14 19:08:21 UTC
Permalink
On 2/14/2012 11:05 AM, Stephen Hemminger wrote:
> On Tue, 14 Feb 2012 10:57:04 -0800
> John Fastabend <***@intel.com> wrote:
>
>> On 2/14/2012 5:18 AM, jamal wrote:
>>> On Mon, 2012-02-13 at 07:13 -0800, John Fastabend wrote:
>>>
>>>> The use case here is multiple VFs but the same solution should work with
>>>> multiple PFs as well. FDB controls should be independent of how the ports
>>>> are exposed VFs, PFs, VMDQ/queue pairs, macvlan, etc.
>>>
>>> Makes sense.
>>>
>>>> With events and ADD/DEL/GET FDB controls we can solve both cases. This also
>>>> solves Roopa's case with macvlan where she wants to add additional addresses
>>>> to macvlan ports.
>>>
>>> Not familiar with that issue - I'll prowl the list.
>>
>> Roopa was likely on the right track here,
>>
>> http://patchwork.ozlabs.org/patch/123064/
>>
>> But I think the proper syntax is to use the existing PF_BRIDGE:RTM_XXX
>> netlink messages. And if possible drive this without extending ndo_ops.
>>
>> An ideal user space interaction IMHO would look like,
>>
>> [***@jf-dev1-dcblab iproute2]# ./br/br fdb add 52:e5:62:7b:57:88 dev veth10
>> [***@jf-dev1-dcblab iproute2]# ./br/br fdb
>> port mac addr flags
>> veth2 36:a6:35:9b:96:c4 local
>> veth4 aa:54:b0:7b:42:ef local
>> veth0 2a:e8:5c:95:6c:1b local
>> veth6 6e:26:d5:43:a3:36 local
>> veth0 f2:c1:39:76:6a:fb
>> veth8 4e:35:16:af:87:13 local
>> veth10 52:e5:62:7b:57:88 static
>> veth10 aa:a9:35:21:15:c4 local
>> [***@jf-dev1-dcblab iproute2]# ./br/br fdb add dev eth3 to 52:e5:62:7b:57:88
>> RTNETLINK answers: Invalid argument
>
> I am going to put bridge (nameclash with br) tool into iproute2 (soon).

I've been using it on my dev box for awhile now and it works well for
me.
Jamal Hadi Salim
2012-02-15 14:10:57 UTC
Permalink
On Tue, 2012-02-14 at 10:57 -0800, John Fastabend wrote:

> Roopa was likely on the right track here,
>
> http://patchwork.ozlabs.org/patch/123064/

Doesnt seem related to the bridging stuff - the modeling looks
reasonable however.

> But I think the proper syntax is to use the existing PF_BRIDGE:RTM_XXX
> netlink messages. And if possible drive this without extending ndo_ops.
>
> An ideal user space interaction IMHO would look like,
>
> [***@jf-dev1-dcblab iproute2]# ./br/br fdb add 52:e5:62:7b:57:88 dev veth10
> [***@jf-dev1-dcblab iproute2]# ./br/br fdb
> port mac addr flags
> veth2 36:a6:35:9b:96:c4 local
> veth4 aa:54:b0:7b:42:ef local
> veth0 2a:e8:5c:95:6c:1b local
> veth6 6e:26:d5:43:a3:36 local
> veth0 f2:c1:39:76:6a:fb
> veth8 4e:35:16:af:87:13 local
> veth10 52:e5:62:7b:57:88 static
> veth10 aa:a9:35:21:15:c4 local

Looks nice, where is the targeted bridge(eg br0) in that syntax?

> Using Stephen's br tool. First command adds FDB entry to SW bridge and
> if the same tool could be used to add entries to embedded bridge I think
> that would be the best case.

That would be nice (although adds dependency on the presence of the
s/ware bridge). Would be nicer to have either a knob in the kernel to
say "synchronize with h/w bridge foo" which can be turned off.

> So no RTNETLINK error on the second cmd. Then
> embedded FDB entries could be dumped this way also so I get a complete view
> of my FDB setup across multiple sw bridges and embedded bridges.

So if you had multiple h/ware bridges - which one is tied to br0?


> Yes. The hardware has a bit to support this which is currently not exposed
> to user space. That's a case where we have 'yet another knob' that needs
> a clean solution. This causes real bugs today when users try to use the
> macvlan devices in VEPA mode on top of SR-IOV. By the way these modes are
> all part of the 802.1Qbg spec which people actually want to use with Linux
> so a good clean solution is probably needed.


I think the knobs to "flood" and "learn" are important. The hardware
seems to have the "flood" but not the "learn/discover". I think the
s/ware bridge needs to have both. At the moment - as pointed out in that
*NEIGH* notification, s/w bridge assumes a policy that could be
considered a security flaw in some circles - just because you are my
neighbor does not mean i trust you to come into my house; i may trust
you partially and allow you only to come through the front door. Even in
Canada with a default policy of not locking your door we sometimes lock
our doors ;->


> I have no problem with drawing the line here and trying to implement something
> over PF_BRIDGE:RTM_xxx nlmsgs.


My comment/concern was in regard to the bridge built-in policy of
reading from the neighbor updates (refer to above comments)

cheers,
jamal


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-16 01:26:28 UTC
Permalink
On 2/15/2012 6:10 AM, Jamal Hadi Salim wrote:
> On Tue, 2012-02-14 at 10:57 -0800, John Fastabend wrote:
>
>> Roopa was likely on the right track here,
>>
>> http://patchwork.ozlabs.org/patch/123064/
>
> Doesnt seem related to the bridging stuff - the modeling looks
> reasonable however.
>

The operations are really the same ADD/DEL/GET additional MAC
addresses to a port, in this case a macvlan type port. The
difference is the macvlan port type drops any packet with an
address not in the FDB where the bridge type floods these.

>> But I think the proper syntax is to use the existing PF_BRIDGE:RTM_XXX
>> netlink messages. And if possible drive this without extending ndo_ops.
>>
>> An ideal user space interaction IMHO would look like,
>>
>> [***@jf-dev1-dcblab iproute2]# ./br/br fdb add 52:e5:62:7b:57:88 dev veth10
>> [***@jf-dev1-dcblab iproute2]# ./br/br fdb
>> port mac addr flags
>> veth2 36:a6:35:9b:96:c4 local
>> veth4 aa:54:b0:7b:42:ef local
>> veth0 2a:e8:5c:95:6c:1b local
>> veth6 6e:26:d5:43:a3:36 local
>> veth0 f2:c1:39:76:6a:fb
>> veth8 4e:35:16:af:87:13 local
>> veth10 52:e5:62:7b:57:88 static
>> veth10 aa:a9:35:21:15:c4 local
>
> Looks nice, where is the targeted bridge(eg br0) in that syntax?

[***@jf-dev1-dcblab src]# br fdb help
Usage: br fdb { add | del | replace } ADDR dev DEV
br fdb {show} [ dev DEV ]

In my example I just dumped all bridge devices,

#br fdb show dev bridge0

>
>> Using Stephen's br tool. First command adds FDB entry to SW bridge and
>> if the same tool could be used to add entries to embedded bridge I think
>> that would be the best case.
>
> That would be nice (although adds dependency on the presence of the
> s/ware bridge). Would be nicer to have either a knob in the kernel to
> say "synchronize with h/w bridge foo" which can be turned off.
>

Seems we need both a synchronize and a { add | del | replace } option.

>> So no RTNETLINK error on the second cmd. Then
>> embedded FDB entries could be dumped this way also so I get a complete view
>> of my FDB setup across multiple sw bridges and embedded bridges.
>
> So if you had multiple h/ware bridges - which one is tied to br0?
>

Not sure I follow but does the additional dev parameter above answer this?

>
>> Yes. The hardware has a bit to support this which is currently not exposed
>> to user space. That's a case where we have 'yet another knob' that needs
>> a clean solution. This causes real bugs today when users try to use the
>> macvlan devices in VEPA mode on top of SR-IOV. By the way these modes are
>> all part of the 802.1Qbg spec which people actually want to use with Linux
>> so a good clean solution is probably needed.
>
>
> I think the knobs to "flood" and "learn" are important. The hardware
> seems to have the "flood" but not the "learn/discover". I think the
> s/ware bridge needs to have both. At the moment - as pointed out in that
> *NEIGH* notification, s/w bridge assumes a policy that could be
> considered a security flaw in some circles - just because you are my
> neighbor does not mean i trust you to come into my house; i may trust
> you partially and allow you only to come through the front door. Even in
> Canada with a default policy of not locking your door we sometimes lock
> our doors ;->
>
>
>> I have no problem with drawing the line here and trying to implement something
>> over PF_BRIDGE:RTM_xxx nlmsgs.
>
>
> My comment/concern was in regard to the bridge built-in policy of
> reading from the neighbor updates (refer to above comments)
>

So I think what your saying is a per port bit to disable learning...

hmm but if you start tweaking it too much it looks less and less like a
802.1D bridge and more like something you would want to build with tc or
openvswitch or tc+bridge or tc+macvlan.

.John

> cheers,
> jamal
>
>

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
jamal
2012-02-17 14:28:52 UTC
Permalink
On Wed, 2012-02-15 at 17:26 -0800, John Fastabend wrote:
> On 2/15/2012 6:10 AM, Jamal Hadi Salim wrote:
> > On Tue, 2012-02-14 at 10:57 -0800, John Fastabend wrote:
> >
> >> Roopa was likely on the right track here,
> >>
> >> http://patchwork.ozlabs.org/patch/123064/
> >
> > Doesnt seem related to the bridging stuff - the modeling looks
> > reasonable however.
> >
>
> The operations are really the same ADD/DEL/GET additional MAC
> addresses to a port, in this case a macvlan type port. The
> difference is the macvlan port type drops any packet with an
> address not in the FDB where the bridge type floods these.

Ok.
[the vlan piece really should have been an integrated part of bridging;
in the early days this was the case]


> [***@jf-dev1-dcblab src]# br fdb help
> Usage: br fdb { add | del | replace } ADDR dev DEV
> br fdb {show} [ dev DEV ]
>
> In my example I just dumped all bridge devices,
>

Ok, makes sense.


> Seems we need both a synchronize and a { add | del | replace } option.

I am conflicted on this.
Not sure if that is a command line thing or something built into a user
space daemon. It may be useful to have the command line variant but i
feel having a daemon take care of things helps in faster
synchronization.
I think user space is a good spot to add such functionality (as opposed
to the kernel). That way user space can work with h/ware switching such
as yours as well as a standalone switching chips (from sillicon vendors
like Marvel etc).
IMO, the average user doesnt need to be aware of such low level stuff;
so the default should be for the user not to be responsible for
configuration of synchronization. IOW, I want to just run well
understood user interface tools things like ifconfig, ip link etc, the
new br tool and not even need to be aware that we are offloading.
So as long as s/w br0 is mapping to the bridge on ixgb-0 i dont need
to know ixgb0 h/w bridge exists.

One last comment:
With synchronization there are other challenges when the entry in the
hardware conflicts with the entry in software when you intend the
behavior to be the same. This is not such a big deal with bridging but
becomes more apparent when you start offloading ACLs etc.


> So I think what your saying is a per port bit to disable learning...
> hmm but if you start tweaking it too much it looks less and less like a
> 802.1D bridge and more like something you would want to build with tc or
> openvswitch or tc+bridge or tc+macvlan.

These are pretty commodity features in most silicon switching chips ive
come across. You have a knob to control learning and another to control
flooding.

cheers,
jamal

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-17 17:10:25 UTC
Permalink
On 2/17/2012 6:28 AM, jamal wrote:
> On Wed, 2012-02-15 at 17:26 -0800, John Fastabend wrote:
>> On 2/15/2012 6:10 AM, Jamal Hadi Salim wrote:
>>> On Tue, 2012-02-14 at 10:57 -0800, John Fastabend wrote:
>>>
>>>> Roopa was likely on the right track here,
>>>>
>>>> http://patchwork.ozlabs.org/patch/123064/
>>>
>>> Doesnt seem related to the bridging stuff - the modeling looks
>>> reasonable however.
>>>
>>
>> The operations are really the same ADD/DEL/GET additional MAC
>> addresses to a port, in this case a macvlan type port. The
>> difference is the macvlan port type drops any packet with an
>> address not in the FDB where the bridge type floods these.
>
> Ok.
> [the vlan piece really should have been an integrated part of bridging;
> in the early days this was the case]
>
>
>> [***@jf-dev1-dcblab src]# br fdb help
>> Usage: br fdb { add | del | replace } ADDR dev DEV
>> br fdb {show} [ dev DEV ]
>>
>> In my example I just dumped all bridge devices,
>>
>
> Ok, makes sense.
>
>
>> Seems we need both a synchronize and a { add | del | replace } option.
>
> I am conflicted on this.
> Not sure if that is a command line thing or something built into a user
> space daemon. It may be useful to have the command line variant but i
> feel having a daemon take care of things helps in faster
> synchronization.
> I think user space is a good spot to add such functionality (as opposed
> to the kernel). That way user space can work with h/ware switching such
> as yours as well as a standalone switching chips (from sillicon vendors
> like Marvel etc).
> IMO, the average user doesnt need to be aware of such low level stuff;
> so the default should be for the user not to be responsible for
> configuration of synchronization. IOW, I want to just run well
> understood user interface tools things like ifconfig, ip link etc, the
> new br tool and not even need to be aware that we are offloading.
> So as long as s/w br0 is mapping to the bridge on ixgb-0 i dont need
> to know ixgb0 h/w bridge exists.
>

Yes I agree that is the goal.

> One last comment:
> With synchronization there are other challenges when the entry in the
> hardware conflicts with the entry in software when you intend the
> behavior to be the same. This is not such a big deal with bridging but
> becomes more apparent when you start offloading ACLs etc.
>

OK and these sorts of conflicts certainly don't need to be resolved
by kernel code. So I think this is a reasonable reason to drive the
synchronization into a user space daemon.

>
>> So I think what your saying is a per port bit to disable learning...
>> hmm but if you start tweaking it too much it looks less and less like a
>> 802.1D bridge and more like something you would want to build with tc or
>> openvswitch or tc+bridge or tc+macvlan.
>
> These are pretty commodity features in most silicon switching chips ive
> come across. You have a knob to control learning and another to control
> flooding.
>

All right this looks like a follow up patch to me. First build the interface
to configure the HW FDB. Then a second series to add a flooding knob which
works for both embedded switches and software switches.

> cheers,
> jamal
>
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
jamal
2012-02-18 12:41:40 UTC
Permalink
On Fri, 2012-02-17 at 09:10 -0800, John Fastabend wrote:

> Yes I agree that is the goal.
>
> > One last comment:
> > With synchronization there are other challenges when the entry in the
> > hardware conflicts with the entry in software when you intend the
> > behavior to be the same. This is not such a big deal with bridging but
> > becomes more apparent when you start offloading ACLs etc.
> >
>
> OK and these sorts of conflicts certainly don't need to be resolved
> by kernel code. So I think this is a reasonable reason to drive the
> synchronization into a user space daemon.


Yep.
Thanks for listening John. Waiting to see them patches.

cheers,
jamal
John Fastabend
2012-02-29 04:40:06 UTC
Permalink
On 2/18/2012 4:41 AM, jamal wrote:
> On Fri, 2012-02-17 at 09:10 -0800, John Fastabend wrote:
>
>> Yes I agree that is the goal.
>>
>>> One last comment:
>>> With synchronization there are other challenges when the entry in the
>>> hardware conflicts with the entry in software when you intend the
>>> behavior to be the same. This is not such a big deal with bridging but
>>> becomes more apparent when you start offloading ACLs etc.
>>>
>>
>> OK and these sorts of conflicts certainly don't need to be resolved
>> by kernel code. So I think this is a reasonable reason to drive the
>> synchronization into a user space daemon.
>
>
> Yep.
> Thanks for listening John. Waiting to see them patches.
>
> cheers,
> jamal
>
>
>

+Lennert

OK back to this. The last piece is where to put these messages...
we could take PF_ROUTE:RTM_*NEIGH

PF_ROUTE:RTM_NEWNEIGH - Add a new FDB entry to an offloaded
switch.
PF_ROUTE:RTM_DELNEIGH - Delete a FDB entry from an offlaoded
switch.
PF_ROUTE:RTM_GETNEIGH - Dumps the embedded FDB table

The neighbor code is using the PF_UNSPEC protocol type so we won't
collide with these unless someone was using PF_ROUTE and relying on
falling back to PF_UNSPEC however I couldn't find any programs that
did this iproute2 certainly doesn't. And the bridge pieces are using
PF_BRIDGE so no collision there.

I briefly thought about trying to pull the PF_BRIDGE protocol out
and use this for both types but I think its better to leave the
bridge code alone and there is also the issue of disambiguating a msg
at a port which has both an embedded switch and has SW bridge for a
master.

Also if there are embedded switches with learning capabilities they
might want to trigger events to user space. In this case having
a protocol type makes user space a bit easier to manage. I've
added Lennert so maybe he can comment I think the Marvell chipsets
might support something along these lines. The SR-IOV chipsets I'm
aware of _today_ don't do learning. Learning makes the event model
more plausible.

The other mechanism would be to embed some more attributes into the
PF_UNSPEC:RTM_XXXLINK msg however I'm thinking that if we want to
support learning and triggering events then we likely also don't
want to send these events to every app with RTNLGRP_LINK set.

Plus there is already a proliferation of LINK attributes and dumping
the FDB out of this seems a bit much but could be done with some
bitmasks. Although the current ext_filter_mask u32 doesn't seem to
be sufficient for events to trigger this.

so much for a short note...

Thanks
.John




--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-29 05:14:40 UTC
Permalink
On 2/28/2012 8:40 PM, John Fastabend wrote:
> On 2/18/2012 4:41 AM, jamal wrote:
>> On Fri, 2012-02-17 at 09:10 -0800, John Fastabend wrote:
>>
>>> Yes I agree that is the goal.
>>>
>>>> One last comment:
>>>> With synchronization there are other challenges when the entry in the
>>>> hardware conflicts with the entry in software when you intend the
>>>> behavior to be the same. This is not such a big deal with bridging but
>>>> becomes more apparent when you start offloading ACLs etc.
>>>>
>>>
>>> OK and these sorts of conflicts certainly don't need to be resolved
>>> by kernel code. So I think this is a reasonable reason to drive the
>>> synchronization into a user space daemon.
>>
>>
>> Yep.
>> Thanks for listening John. Waiting to see them patches.
>>
>> cheers,
>> jamal
>>
>>
>>
>
> +Lennert
>
> OK back to this. The last piece is where to put these messages...
> we could take PF_ROUTE:RTM_*NEIGH
>
> PF_ROUTE:RTM_NEWNEIGH - Add a new FDB entry to an offloaded
> switch.
> PF_ROUTE:RTM_DELNEIGH - Delete a FDB entry from an offlaoded
> switch.
> PF_ROUTE:RTM_GETNEIGH - Dumps the embedded FDB table
>
> The neighbor code is using the PF_UNSPEC protocol type so we won't
> collide with these unless someone was using PF_ROUTE and relying on
> falling back to PF_UNSPEC however I couldn't find any programs that
> did this iproute2 certainly doesn't. And the bridge pieces are using
> PF_BRIDGE so no collision there.
>
> I briefly thought about trying to pull the PF_BRIDGE protocol out
> and use this for both types but I think its better to leave the
> bridge code alone and there is also the issue of disambiguating a msg
> at a port which has both an embedded switch and has SW bridge for a
> master.

Maybe I gave up too quickly here I could use a bit in the ndm_flags to
specify embedded or sw bridge. But would require having the bridge
module loaded.

>
> Also if there are embedded switches with learning capabilities they
> might want to trigger events to user space. In this case having
> a protocol type makes user space a bit easier to manage. I've
> added Lennert so maybe he can comment I think the Marvell chipsets
> might support something along these lines. The SR-IOV chipsets I'm
> aware of _today_ don't do learning. Learning makes the event model
> more plausible.
>

Just checked looks like the DSA infrastructure has commands to enable
STP so guess it is doing learning.

> The other mechanism would be to embed some more attributes into the
> PF_UNSPEC:RTM_XXXLINK msg however I'm thinking that if we want to
> support learning and triggering events then we likely also don't
> want to send these events to every app with RTNLGRP_LINK set.
>
> Plus there is already a proliferation of LINK attributes and dumping
> the FDB out of this seems a bit much but could be done with some
> bitmasks. Although the current ext_filter_mask u32 doesn't seem to
> be sufficient for events to trigger this.
>
> so much for a short note...
>
> Thanks
> .John
>
>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to ***@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Jamal Hadi Salim
2012-02-29 13:57:26 UTC
Permalink
On Tue, 2012-02-28 at 21:14 -0800, John Fastabend wrote:

> Just checked looks like the DSA infrastructure has commands to enable
> STP so guess it is doing learning.

IIRC, Lennert built some of this stuff tied to the kernel.

cheers,
jamal

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Jamal Hadi Salim
2012-02-29 13:56:19 UTC
Permalink
On Tue, 2012-02-28 at 20:40 -0800, John Fastabend wrote:

> OK back to this. The last piece is where to put these messages...
> we could take PF_ROUTE:RTM_*NEIGH
>
> PF_ROUTE:RTM_NEWNEIGH - Add a new FDB entry to an offloaded
> switch.
> PF_ROUTE:RTM_DELNEIGH - Delete a FDB entry from an offlaoded
> switch.
> PF_ROUTE:RTM_GETNEIGH - Dumps the embedded FDB table
>

Why RTM_*NEIGH? RTM tends to map to Route/L3 and NEIGH tends to map
to ndisc or ARP both tied to IP address resolution. While both ARP/Ndisc
may play a role in the user space app populating the FDB, i dont think
they are necessary players.
Learning could be via a table entry miss and packet redirect to user
space.
So my suggestion is to use FDB_*ENTRY for names

> The neighbor code is using the PF_UNSPEC protocol type so we won't
> collide with these unless someone was using PF_ROUTE and relying on
> falling back to PF_UNSPEC however I couldn't find any programs that
> did this iproute2 certainly doesn't. And the bridge pieces are using
> PF_BRIDGE so no collision there.

They have to be different calls from the calls that talk to the s/ware
bridge. In my opinion, as controversial as this may sound, you need to
be flexible enough that some vendor can replace these calls with
proprietary calls which are more efficient for their hardware. So a
"plugin" to replace these calls in the user space code would be a
good idea. Alternatively, you could make that something they do at
the driver level i.e from user space to kernel it is "hardware, please
addthistotheFDBtable()" call and the implementation of that could be
proprietary to the specific hardware.

[..]

> Also if there are embedded switches with learning capabilities they
> might want to trigger events to user space. In this case having
> a protocol type makes user space a bit easier to manage. I've
> added Lennert so maybe he can comment I think the Marvell chipsets
> might support something along these lines. The SR-IOV chipsets I'm
> aware of _today_ don't do learning. Learning makes the event model
> more plausible.

The other events to consider is aging of hardware entries.

> The other mechanism would be to embed some more attributes into the
> PF_UNSPEC:RTM_XXXLINK msg however I'm thinking that if we want to
> support learning and triggering events then we likely also don't
> want to send these events to every app with RTNLGRP_LINK set.

I think this needs to be a different event message.
FDB_TABLEMISS? FDB_EXCEPTION?

> Plus there is already a proliferation of LINK attributes and dumping
> the FDB out of this seems a bit much but could be done with some
> bitmasks. Although the current ext_filter_mask u32 doesn't seem to
> be sufficient for events to trigger this.

Dumping the FDB table should be something along the lines of FDB_GET
with the dump flag. It shouldnt tie to the LINK side of things.

cheers,
jamal

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-29 17:25:56 UTC
Permalink
On 2/29/2012 5:56 AM, Jamal Hadi Salim wrote:
> On Tue, 2012-02-28 at 20:40 -0800, John Fastabend wrote:
>
>> OK back to this. The last piece is where to put these messages...
>> we could take PF_ROUTE:RTM_*NEIGH
>>
>> PF_ROUTE:RTM_NEWNEIGH - Add a new FDB entry to an offloaded
>> switch.
>> PF_ROUTE:RTM_DELNEIGH - Delete a FDB entry from an offlaoded
>> switch.
>> PF_ROUTE:RTM_GETNEIGH - Dumps the embedded FDB table
>>
>
> Why RTM_*NEIGH? RTM tends to map to Route/L3 and NEIGH tends to map
> to ndisc or ARP both tied to IP address resolution. While both ARP/Ndisc
> may play a role in the user space app populating the FDB, i dont think
> they are necessary players.
> Learning could be via a table entry miss and packet redirect to user
> space.
> So my suggestion is to use FDB_*ENTRY for names
>

Well I think NETLINK_ROUTE is the most correct type to use in this
case. Per netlink.h its for routing and device hooks.

#define NETLINK_ROUTE 0 /* Routing/device hook */

And NETLINK_ROUTE msg_types use the RTM_* prefix. The _*NEIGH postfix
were merely a copy from the SW BRIDGE code paths. How about,

PF_BRIDGE:RTM_FDB_NEWENTRY
PF_BRIDGE:RTM_FDB_DELENTRY
PF_BRIDGE:RTM_FDB_GETENTRY

And a new group RTNLGRP_FDB. Also using NETLINK_ROUTE gives the correct
rtnl locking semantics for free.

>> The neighbor code is using the PF_UNSPEC protocol type so we won't
>> collide with these unless someone was using PF_ROUTE and relying on
>> falling back to PF_UNSPEC however I couldn't find any programs that
>> did this iproute2 certainly doesn't. And the bridge pieces are using
>> PF_BRIDGE so no collision there.
>
> They have to be different calls from the calls that talk to the s/ware
> bridge. In my opinion, as controversial as this may sound, you need to
> be flexible enough that some vendor can replace these calls with
> proprietary calls which are more efficient for their hardware. So a
> "plugin" to replace these calls in the user space code would be a
> good idea. Alternatively, you could make that something they do at
> the driver level i.e from user space to kernel it is "hardware, please
> addthistotheFDBtable()" call and the implementation of that could be
> proprietary to the specific hardware.
>

Agreed. I think adding some ndo_ops for bridging offloads here would
work. For example the DSA infrastructure and/or macvlan devices might
need this. Along the lines of extending this RFC,

[RFC] hardware bridging support for DSA switches
http://patchwork.ozlabs.org/patch/16578/


.John

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Stephen Hemminger
2012-02-29 17:52:04 UTC
Permalink
On Wed, 29 Feb 2012 09:25:56 -0800
John Fastabend <***@intel.com> wrote:

> On 2/29/2012 5:56 AM, Jamal Hadi Salim wrote:
> > On Tue, 2012-02-28 at 20:40 -0800, John Fastabend wrote:
> >
> >> OK back to this. The last piece is where to put these messages...
> >> we could take PF_ROUTE:RTM_*NEIGH
> >>
> >> PF_ROUTE:RTM_NEWNEIGH - Add a new FDB entry to an offloaded
> >> switch.
> >> PF_ROUTE:RTM_DELNEIGH - Delete a FDB entry from an offlaoded
> >> switch.
> >> PF_ROUTE:RTM_GETNEIGH - Dumps the embedded FDB table
> >>
> >
> > Why RTM_*NEIGH? RTM tends to map to Route/L3 and NEIGH tends to map
> > to ndisc or ARP both tied to IP address resolution. While both ARP/Ndisc
> > may play a role in the user space app populating the FDB, i dont think
> > they are necessary players.
> > Learning could be via a table entry miss and packet redirect to user
> > space.
> > So my suggestion is to use FDB_*ENTRY for names
> >
>
> Well I think NETLINK_ROUTE is the most correct type to use in this
> case. Per netlink.h its for routing and device hooks.
>
> #define NETLINK_ROUTE 0 /* Routing/device hook */
>
> And NETLINK_ROUTE msg_types use the RTM_* prefix. The _*NEIGH postfix
> were merely a copy from the SW BRIDGE code paths. How about,
>
> PF_BRIDGE:RTM_FDB_NEWENTRY
> PF_BRIDGE:RTM_FDB_DELENTRY
> PF_BRIDGE:RTM_FDB_GETENTRY
>
> And a new group RTNLGRP_FDB. Also using NETLINK_ROUTE gives the correct
> rtnl locking semantics for free.
>
> >> The neighbor code is using the PF_UNSPEC protocol type so we won't
> >> collide with these unless someone was using PF_ROUTE and relying on
> >> falling back to PF_UNSPEC however I couldn't find any programs that
> >> did this iproute2 certainly doesn't. And the bridge pieces are using
> >> PF_BRIDGE so no collision there.
> >
> > They have to be different calls from the calls that talk to the s/ware
> > bridge. In my opinion, as controversial as this may sound, you need to
> > be flexible enough that some vendor can replace these calls with
> > proprietary calls which are more efficient for their hardware. So a
> > "plugin" to replace these calls in the user space code would be a
> > good idea. Alternatively, you could make that something they do at
> > the driver level i.e from user space to kernel it is "hardware, please
> > addthistotheFDBtable()" call and the implementation of that could be
> > proprietary to the specific hardware.
> >
>
> Agreed. I think adding some ndo_ops for bridging offloads here would
> work. For example the DSA infrastructure and/or macvlan devices might
> need this. Along the lines of extending this RFC,
>
> [RFC] hardware bridging support for DSA switches
> http://patchwork.ozlabs.org/patch/16578/

I want to see a unified API so that user space control applications (RSTP, TRILL?)
can use one set of netlink calls for both software bridge and hardware offloaded
bridges. Does this proposal meet that requirement?
John Fastabend
2012-02-29 18:19:48 UTC
Permalink
On 2/29/2012 9:52 AM, Stephen Hemminger wrote:
> On Wed, 29 Feb 2012 09:25:56 -0800
> John Fastabend <***@intel.com> wrote:
>
>> On 2/29/2012 5:56 AM, Jamal Hadi Salim wrote:
>>> On Tue, 2012-02-28 at 20:40 -0800, John Fastabend wrote:
>>>
>>>> OK back to this. The last piece is where to put these messages...
>>>> we could take PF_ROUTE:RTM_*NEIGH
>>>>
>>>> PF_ROUTE:RTM_NEWNEIGH - Add a new FDB entry to an offloaded
>>>> switch.
>>>> PF_ROUTE:RTM_DELNEIGH - Delete a FDB entry from an offlaoded
>>>> switch.
>>>> PF_ROUTE:RTM_GETNEIGH - Dumps the embedded FDB table
>>>>
>>>
>>> Why RTM_*NEIGH? RTM tends to map to Route/L3 and NEIGH tends to map
>>> to ndisc or ARP both tied to IP address resolution. While both ARP/Ndisc
>>> may play a role in the user space app populating the FDB, i dont think
>>> they are necessary players.
>>> Learning could be via a table entry miss and packet redirect to user
>>> space.
>>> So my suggestion is to use FDB_*ENTRY for names
>>>
>>
>> Well I think NETLINK_ROUTE is the most correct type to use in this
>> case. Per netlink.h its for routing and device hooks.
>>
>> #define NETLINK_ROUTE 0 /* Routing/device hook */
>>
>> And NETLINK_ROUTE msg_types use the RTM_* prefix. The _*NEIGH postfix
>> were merely a copy from the SW BRIDGE code paths. How about,
>>
>> PF_BRIDGE:RTM_FDB_NEWENTRY
>> PF_BRIDGE:RTM_FDB_DELENTRY
>> PF_BRIDGE:RTM_FDB_GETENTRY
>>
>> And a new group RTNLGRP_FDB. Also using NETLINK_ROUTE gives the correct
>> rtnl locking semantics for free.
>>
>>>> The neighbor code is using the PF_UNSPEC protocol type so we won't
>>>> collide with these unless someone was using PF_ROUTE and relying on
>>>> falling back to PF_UNSPEC however I couldn't find any programs that
>>>> did this iproute2 certainly doesn't. And the bridge pieces are using
>>>> PF_BRIDGE so no collision there.
>>>
>>> They have to be different calls from the calls that talk to the s/ware
>>> bridge. In my opinion, as controversial as this may sound, you need to
>>> be flexible enough that some vendor can replace these calls with
>>> proprietary calls which are more efficient for their hardware. So a
>>> "plugin" to replace these calls in the user space code would be a
>>> good idea. Alternatively, you could make that something they do at
>>> the driver level i.e from user space to kernel it is "hardware, please
>>> addthistotheFDBtable()" call and the implementation of that could be
>>> proprietary to the specific hardware.
>>>
>>
>> Agreed. I think adding some ndo_ops for bridging offloads here would
>> work. For example the DSA infrastructure and/or macvlan devices might
>> need this. Along the lines of extending this RFC,
>>
>> [RFC] hardware bridging support for DSA switches
>> http://patchwork.ozlabs.org/patch/16578/
>
> I want to see a unified API so that user space control applications (RSTP, TRILL?)
> can use one set of netlink calls for both software bridge and hardware offloaded
> bridges. Does this proposal meet that requirement?
>

With the patches I sent out last night the same netlink calls are used
for both SW and HW with a flag set in ndm_flags to indicate it is a hardware
entry. The flag is needed when a port has offload support and is also
a slave of a SW bridge. Another option would be to apply the command to both
hardware and software tables. This might be good enough and user space would
not have to make distinctions between HW and SW bridges. Also helps with my
original use case where I want the SW and HW bridge FDBs to be in sync.

In response to Jamal's comment I proposed changing the type to RTM_FDB_XXXENTRY
but the message contents are the same in both cases.

Jamal, so why do "They have to be different calls"? I'm not so sure anymore...
moving to RTM_FDB_XXXENTRY saved some refactoring in the bridge module but that
is just cosmetic.

Thanks,
John
Jamal Hadi Salim
2012-03-01 13:36:20 UTC
Permalink
On Wed, 2012-02-29 at 10:19 -0800, John Fastabend wrote:

> >
> > I want to see a unified API so that user space control applications (RSTP, TRILL?)
> > can use one set of netlink calls for both software bridge and hardware offloaded
> > bridges. Does this proposal meet that requirement?
> >

I dont see any issues with those requirements being met.

> Jamal, so why do "They have to be different calls"? I'm not so sure anymore...
> moving to RTM_FDB_XXXENTRY saved some refactoring in the bridge module but that
> is just cosmetic.

I may not want to use the s/ware bridge i.e I may want to use h/ware
bridge. I may want to use both. So there are 3 variations there. You
need at least 1.5 bits to represent them if you are going to use the
same interface. There may be features in either h/ware but not in
s/ware and vice-versa.
A single interface with flags which say this applies to hware:sware:both
would be good, but it may be harder to achieve - thats why i suggested
they be different.

cheers,
jamal
John Fastabend
2012-03-01 22:17:28 UTC
Permalink
On 3/1/2012 5:36 AM, Jamal Hadi Salim wrote:
> On Wed, 2012-02-29 at 10:19 -0800, John Fastabend wrote:
>
>>>
>>> I want to see a unified API so that user space control applications (RSTP, TRILL?)
>>> can use one set of netlink calls for both software bridge and hardware offloaded
>>> bridges. Does this proposal meet that requirement?
>>>
>
> I dont see any issues with those requirements being met.
>
>> Jamal, so why do "They have to be different calls"? I'm not so sure anymore...
>> moving to RTM_FDB_XXXENTRY saved some refactoring in the bridge module but that
>> is just cosmetic.
>
> I may not want to use the s/ware bridge i.e I may want to use h/ware
> bridge. I may want to use both. So there are 3 variations there. You
> need at least 1.5 bits to represent them if you are going to use the
> same interface. There may be features in either h/ware but not in
> s/ware and vice-versa.
> A single interface with flags which say this applies to hware:sware:both
> would be good, but it may be harder to achieve - thats why i suggested
> they be different.
>
> cheers,
> jamal
>

Hmm so I think what I'll do is this...

both: ndm_flags = 0
sw : ndm_flags = NTF_SW_FDB
hw : ndm_flags = NTF_HW_FDB

Then current tools will work with embedded bridges and software bridges
with the interesting case being when a port supporting an offloaded FDB
is attached to a SW bridge. Doing both in this case seems to be a reasonable
default to me.

The tricky bit will be pulling the message handlers out of the ./net/bridge
code so that we don't have to always load the bridge module to add entries
to a macvlan for example. I need to look at a few other things today but
I'll code up a patch for this tomorrow.

.John
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
jamal
2012-03-02 13:20:45 UTC
Permalink
On Thu, 2012-03-01 at 14:17 -0800, John Fastabend wrote:
> Hmm so I think what I'll do is this...
>
> both: ndm_flags = 0
> sw : ndm_flags = NTF_SW_FDB
> hw : ndm_flags = NTF_HW_FDB
>
> Then current tools will work with embedded bridges and software
> bridges
> with the interesting case being when a port supporting an offloaded
> FDB is attached to a SW bridge. Doing both in this case seems to be a
> reasonable default to me.

Looks good, although it seems like no backward compat is broken, it
feels like the default should be whats goin on today i.e s/ware only.
IOW, I would make that the 0.

cheers,
jamal

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Lennert Buytenhek
2012-03-05 17:00:44 UTC
Permalink
On Thu, Mar 01, 2012 at 08:36:20AM -0500, Jamal Hadi Salim wrote:

> > > I want to see a unified API so that user space control applications (RSTP, TRILL?)
> > > can use one set of netlink calls for both software bridge and hardware offloaded
> > > bridges. Does this proposal meet that requirement?
> > >
>
> I dont see any issues with those requirements being met.
>
> > Jamal, so why do "They have to be different calls"? I'm not so sure anymore...
> > moving to RTM_FDB_XXXENTRY saved some refactoring in the bridge module but that
> > is just cosmetic.
>
> I may not want to use the s/ware bridge i.e I may want to use h/ware
> bridge. I may want to use both.

This is a rather common case in embedded wireless routers/access points,
where you want to have the 4 LAN ports bridged together with the wlan0
interface.

In this scenario, the bridging between the LAN ports is typically done
in hardware, and the bridging between the LAN ports and wlan0 in
software, but here you have to be careful when you send the packet from
the switch chip up the stack to be forwarded to the wlan0 interface to
not re-send it to the hardware switch chip ports other than the one
that the packet came from.

net/dsa currently solves this by not having the hardware handle
broadcast packets at all, which circumvents the problem, but for
multicast traffic you would still like to be able to do at least the
forwarding that can be done in hardware in hardware. (Unicast doesn't
have this problem as long as the kernel and the switch chip agree on
their view of the FDB.)
Jamal Hadi Salim
2012-03-01 13:24:22 UTC
Permalink
On Wed, 2012-02-29 at 09:25 -0800, John Fastabend wrote:

> Well I think NETLINK_ROUTE is the most correct type to use in this
> case. Per netlink.h its for routing and device hooks.
>
> #define NETLINK_ROUTE 0 /* Routing/device hook */
>
> And NETLINK_ROUTE msg_types use the RTM_* prefix. The _*NEIGH postfix
> were merely a copy from the SW BRIDGE code paths. How about,
>
> PF_BRIDGE:RTM_FDB_NEWENTRY
> PF_BRIDGE:RTM_FDB_DELENTRY
> PF_BRIDGE:RTM_FDB_GETENTRY

OK, I guess ;->

> And a new group RTNLGRP_FDB.

Nod.

> Also using NETLINK_ROUTE gives the correct
> rtnl locking semantics for free.

makes sense.


> Agreed. I think adding some ndo_ops for bridging offloads here would
> work. For example the DSA infrastructure and/or macvlan devices might
> need this. Along the lines of extending this RFC,
>
> [RFC] hardware bridging support for DSA switches
> http://patchwork.ozlabs.org/patch/16578/

Certainly - thats one approach that is reasonable.
Where is Lennert? ;-> I changed his email address to one that i am
familiar with.

cheers,
jamal


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Michael S. Tsirkin
2012-03-01 14:14:14 UTC
Permalink
On Wed, Feb 29, 2012 at 09:25:56AM -0800, John Fastabend wrote:
> Agreed. I think adding some ndo_ops for bridging offloads here would
> work. For example the DSA infrastructure and/or macvlan devices might
> need this. Along the lines of extending this RFC,
>
> [RFC] hardware bridging support for DSA switches
> http://patchwork.ozlabs.org/patch/16578/
>
>
> .John

One place where this might not work well would be
macvtap which is not a network device so it doesn't have
its own address, instead it inherits one from macvlan.

--
MST
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-03-01 22:10:44 UTC
Permalink
On 3/1/2012 6:14 AM, Michael S. Tsirkin wrote:
> On Wed, Feb 29, 2012 at 09:25:56AM -0800, John Fastabend wrote:
>> Agreed. I think adding some ndo_ops for bridging offloads here would
>> work. For example the DSA infrastructure and/or macvlan devices might
>> need this. Along the lines of extending this RFC,
>>
>> [RFC] hardware bridging support for DSA switches
>> http://patchwork.ozlabs.org/patch/16578/
>>
>>
>> .John
>
> One place where this might not work well would be
> macvtap which is not a network device so it doesn't have
> its own address, instead it inherits one from macvlan.
>

But is macvtap really doing any forwarding or implementing any
RX filters? Took a quick scan and it looks like the forwarding
logic is all in the macvlan code paths. In this case I suspect
if we enable macvlan then any device built on top of it would
work.

.John
Lennert Buytenhek
2012-03-05 16:53:39 UTC
Permalink
On Tue, Feb 28, 2012 at 08:40:06PM -0800, John Fastabend wrote:

> Also if there are embedded switches with learning capabilities they
> might want to trigger events to user space. In this case having
> a protocol type makes user space a bit easier to manage. I've
> added Lennert so maybe he can comment I think the Marvell chipsets
> might support something along these lines. The SR-IOV chipsets I'm
> aware of _today_ don't do learning. Learning makes the event model
> more plausible.

net/dsa currently configures any switch chips in the system to do
auto-learning. However, I would much prefer to disable that, and have
the switch chip just pass up packets for new source addresses, have
Linux do the learning, and then mirror the Linux software FDB into
the hardware instead -- that avoids having to manually flush the
hardware FDB on certain STP state transitions or having to configure
the hardware to use a shorter address learning timeout when we're in
the middle of an STP topology change, which are problems we are
running into in practice.

Just curious -- while your patches allow propagating FDB entries
into the hardware, do you also have hooks to tell the hardware which
ports are to share address databases?

For net/dsa, we currently have:

http://patchwork.ozlabs.org/patch/16578/

While I think this is conceptually sound, the implementation is hacky,
and I wonder how you've solved it for your setup, and if DSA can
piggy-back off that.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-03-06 03:45:22 UTC
Permalink
On 3/5/2012 8:53 AM, Lennert Buytenhek wrote:
> On Tue, Feb 28, 2012 at 08:40:06PM -0800, John Fastabend wrote:
>
>> Also if there are embedded switches with learning capabilities they
>> might want to trigger events to user space. In this case having
>> a protocol type makes user space a bit easier to manage. I've
>> added Lennert so maybe he can comment I think the Marvell chipsets
>> might support something along these lines. The SR-IOV chipsets I'm
>> aware of _today_ don't do learning. Learning makes the event model
>> more plausible.
>
> net/dsa currently configures any switch chips in the system to do
> auto-learning. However, I would much prefer to disable that, and have
> the switch chip just pass up packets for new source addresses, have
> Linux do the learning, and then mirror the Linux software FDB into
> the hardware instead -- that avoids having to manually flush the
> hardware FDB on certain STP state transitions or having to configure
> the hardware to use a shorter address learning timeout when we're in
> the middle of an STP topology change, which are problems we are
> running into in practice.
>

Great. And the plan is we should be able to use the same daemon with
minimal changes (currently a flag) to control both sw and hw bridges.

> Just curious -- while your patches allow propagating FDB entries
> into the hardware, do you also have hooks to tell the hardware which
> ports are to share address databases?
>

Not in the current patches. I don't have hardware right now
that can instantiate multiple bridges. When I get some I was hoping
to do something similar to this patch and use netlink commands
to create/delete bridges and add/remove ports to them. This would
be modifying the existing commands to work for both software and
hardware bridges.

By a bridge instantiation I mean a shared address database in this case.

> For net/dsa, we currently have:
>
> http://patchwork.ozlabs.org/patch/16578/
>
> While I think this is conceptually sound, the implementation is hacky,
> and I wonder how you've solved it for your setup, and if DSA can
> piggy-back off that.

Yep anything we come up with should work in both cases.
Lennert Buytenhek
2012-03-06 14:15:51 UTC
Permalink
On Mon, Mar 05, 2012 at 07:45:22PM -0800, John Fastabend wrote:

> >> Also if there are embedded switches with learning capabilities they
> >> might want to trigger events to user space. In this case having
> >> a protocol type makes user space a bit easier to manage. I've
> >> added Lennert so maybe he can comment I think the Marvell chipsets
> >> might support something along these lines. The SR-IOV chipsets I'm
> >> aware of _today_ don't do learning. Learning makes the event model
> >> more plausible.
> >
> > net/dsa currently configures any switch chips in the system to do
> > auto-learning. However, I would much prefer to disable that, and have
> > the switch chip just pass up packets for new source addresses, have
> > Linux do the learning, and then mirror the Linux software FDB into
> > the hardware instead -- that avoids having to manually flush the
> > hardware FDB on certain STP state transitions or having to configure
> > the hardware to use a shorter address learning timeout when we're in
> > the middle of an STP topology change, which are problems we are
> > running into in practice.
>
> Great. And the plan is we should be able to use the same daemon with
> minimal changes (currently a flag) to control both sw and hw bridges.

Why should userspace care at all whether there is a hw bridge present?


> > Just curious -- while your patches allow propagating FDB entries
> > into the hardware, do you also have hooks to tell the hardware which
> > ports are to share address databases?
>
> Not in the current patches. I don't have hardware right now
> that can instantiate multiple bridges. When I get some I was hoping
> to do something similar to this patch and use netlink commands
> to create/delete bridges and add/remove ports to them. This would
> be modifying the existing commands to work for both software and
> hardware bridges.

In the DSA h/w bridging patch, a hardware address database is only
allocated on addif to an existing bridge, not on addbr.

For one, at addbr time, you have no idea yet whether there will
be any switch chip ports in the bridge port group for this bridge.

Also, the h/w has some restrictions on the assignment of address
database identifiers (e.g. if you want to bridge between lan1.123
and lan2.123, you have to use address database identifier '123').


> By a bridge instantiation I mean a shared address database in this case.

Makes sense. (I'd add STP port states to this.)
jamal
2012-03-06 13:42:26 UTC
Permalink
On Mon, 2012-03-05 at 17:53 +0100, Lennert Buytenhek wrote:

> net/dsa currently configures any switch chips in the system to do
> auto-learning.

So we clearly need the (user configurable) knob to turn on/off learning.
I think it should also be upto the admin to decide whether the learning
happens in the kernel or user space.

> However, I would much prefer to disable that, and have
> the switch chip just pass up packets for new source addresses, have
> Linux do the learning, and then mirror the Linux software FDB into
> the hardware instead -- that avoids having to manually flush the
> hardware FDB on certain STP state transitions or having to configure
> the hardware to use a shorter address learning timeout when we're in
> the middle of an STP topology change, which are problems we are
> running into in practice.

So in the scenario you are describing then it seems the h/ware has
no stp state toggles, correct? In other ASICs i have seen, there is
influence from stp state on behavior.

> Just curious -- while your patches allow propagating FDB entries
> into the hardware, do you also have hooks to tell the hardware which
> ports are to share address databases?

I think those are missing in this discussion and makes a lot of sense to
be part of the interface.

> net/dsa currently solves this by not having the hardware handle
> broadcast packets at all, which circumvents the problem, but for
> multicast traffic you would still like to be able to do at least the
> forwarding that can be done in hardware in hardware. (Unicast doesn't
> have this problem as long as the kernel and the switch chip agree on
> their view of the FDB.)

Of course this could represent an interesting opportunity for a DOS.
Even at 4 port switch at 100Mbs, hitting 500Kpps to the CPU (I am
thinking these tiny switches end up in some tiny MIPS/ARM cpu) could
be devastating. How do you deal with that?

cheers,
jamal


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Lennert Buytenhek
2012-03-06 14:09:57 UTC
Permalink
On Tue, Mar 06, 2012 at 08:42:26AM -0500, jamal wrote:

> > net/dsa currently configures any switch chips in the system to do
> > auto-learning.
>
> So we clearly need the (user configurable) knob to turn on/off learning.

Why so? (I think the switch chips should just never do learning at
all..)


> I think it should also be upto the admin to decide whether the learning
> happens in the kernel or user space.

I can't see any point in doing it in userspace. What would be the
advantage of that? And based on what would the admin make the decision?


> > However, I would much prefer to disable that, and have
> > the switch chip just pass up packets for new source addresses, have
> > Linux do the learning, and then mirror the Linux software FDB into
> > the hardware instead -- that avoids having to manually flush the
> > hardware FDB on certain STP state transitions or having to configure
> > the hardware to use a shorter address learning timeout when we're in
> > the middle of an STP topology change, which are problems we are
> > running into in practice.
>
> So in the scenario you are describing then it seems the h/ware has
> no stp state toggles, correct? In other ASICs i have seen, there is
> influence from stp state on behavior.

It does, there is an STP state field per port in the switch chip,
which controls whether learning takes place on this port (in
Learning and Forwarding states) and whether packets are forwarded
(in the Forwarding state).

But e.g. it doesn't automatically flush this port's FDB entries if
you move a port from Forwarding to Listening -- the STP state field
only controls direct learning and forwarding for received packets.

And when you receive a BPDU with the topology change notification
bit set, the switch won't automatically shorten the FDB entry
timeout for you until the topology change is over, either.


> > Just curious -- while your patches allow propagating FDB entries
> > into the hardware, do you also have hooks to tell the hardware which
> > ports are to share address databases?
>
> I think those are missing in this discussion and makes a lot of
> sense to be part of the interface.

Keep in mind that these chips also do VLAN tagging in hardware, and
so a scenario like:

# brctl addbr br123
# brctl addif br123 lan1.123
# brctl addif br123 lan2.123

is also one that can be handled in hardware (which the current
patchwork patch doesn't handle yet).


> > net/dsa currently solves this by not having the hardware handle
> > broadcast packets at all, which circumvents the problem, but for
> > multicast traffic you would still like to be able to do at least the
> > forwarding that can be done in hardware in hardware. (Unicast doesn't
> > have this problem as long as the kernel and the switch chip agree on
> > their view of the FDB.)
>
> Of course this could represent an interesting opportunity for a DOS.
> Even at 4 port switch at 100Mbs, hitting 500Kpps to the CPU (I am
> thinking these tiny switches end up in some tiny MIPS/ARM cpu) could
> be devastating. How do you deal with that?

You can let the switch rate limit the number of packets passed up to
the CPU. 500 kp/s broadcast traffic seems somewhat excessive in any
case, and I'm not sure if this deserves handling apart from QoSing
those streams to manageable levels.


thanks,
Lennert
Jamal Hadi Salim
2012-03-07 14:11:40 UTC
Permalink
On Tue, 2012-03-06 at 15:09 +0100, Lennert Buytenhek wrote:

> Why so? (I think the switch chips should just never do learning at
> all..)

I agree that learning in software gives you more flexibility; however,
I am for providing interface flexibility as well - switches have
learning features. I think i should be able to use them when it makes
sense to.

> > I think it should also be upto the admin to decide whether the learning
> > happens in the kernel or user space.
>
> I can't see any point in doing it in userspace. What would be the
> advantage of that? And based on what would the admin make the decision?
>

If i wanted to do some funky access control based on some new MAC
address showing up - best place to do it is user space.

> It does, there is an STP state field per port in the switch chip,
> which controls whether learning takes place on this port (in
> Learning and Forwarding states) and whether packets are forwarded
> (in the Forwarding state).

ok, makes sense.

> But e.g. it doesn't automatically flush this port's FDB entries if
> you move a port from Forwarding to Listening -- the STP state field
> only controls direct learning and forwarding for received packets.
>
> And when you receive a BPDU with the topology change notification
> bit set, the switch won't automatically shorten the FDB entry
> timeout for you until the topology change is over, either.

I have to go back and look at some manuals i have - but iirc, the
ones ive played with behaved similarly. As long as we provide knobs
to set/unset those different attributes, I think the handling of all
that should be from software (likely some daemon in user space);
then it shouldnt matter whether we are working with STP BPDUs or TRILL
or thenewprotocolTM etc.

> Keep in mind that these chips also do VLAN tagging in hardware, and
> so a scenario like:
>
> # brctl addbr br123
> # brctl addif br123 lan1.123
> # brctl addif br123 lan2.123
>
> is also one that can be handled in hardware (which the current
> patchwork patch doesn't handle yet).
>

We would need to work with offloading VLANs, no? Do the current
VLAN offloads used for NICs suffice for switching chips as well?
i.e typically most chips have a table associated with some port in
which the Vlan is partof or is the lookup key.

> You can let the switch rate limit the number of packets passed up to
> the CPU. 500 kp/s broadcast traffic seems somewhat excessive in any
> case, and I'm not sure if this deserves handling apart from QoSing
> those streams to manageable levels.

Yes, that would provide a solution.
I havent seen anything where you can rate limit the learning(SA lookup
failure).

cheers,
jamal
Lennert Buytenhek
2012-03-12 08:48:50 UTC
Permalink
On Wed, Mar 07, 2012 at 09:11:40AM -0500, Jamal Hadi Salim wrote:

> > Why so? (I think the switch chips should just never do learning at
> > all..)
>
> I agree that learning in software gives you more flexibility;
> however, I am for providing interface flexibility as well - switches
> have learning features. I think i should be able to use them when it
> makes sense to.

Since it can lead to problems (address database mismatches, doesn't
correctly handle STP transitions or topology changes automatically),
I think it should be avoided whenever possible. I don't see any
advantages of hardware based learning over software based learning
anyway ('flexibility' doesn't seem like a very good argument).


> > > I think it should also be upto the admin to decide whether the
> > > learning happens in the kernel or user space.
> >
> > I can't see any point in doing it in userspace. What would be the
> > advantage of that? And based on what would the admin make the decision?
>
> If i wanted to do some funky access control based on some new MAC
> address showing up - best place to do it is user space.

Alright, that sounds fair.


> > Keep in mind that these chips also do VLAN tagging in hardware, and
> > so a scenario like:
> >
> > # brctl addbr br123
> > # brctl addif br123 lan1.123
> > # brctl addif br123 lan2.123
> >
> > is also one that can be handled in hardware (which the current
> > patchwork patch doesn't handle yet).
>
> We would need to work with offloading VLANs, no?

Yes.


> Do the current VLAN offloads used for NICs suffice for switching
> chips as well? i.e typically most chips have a table associated
> with some port in which the Vlan is partof or is the lookup key.

It should be doable along the lines of the current DSA patch --
add a VLAN ID argument to the interface add/remove callbacks, and
when a VLAN virtual interface is added to the bridge, call the
relevant callbacks with the parent interface + VLAN ID instead.
(This doesn't work for stacked VLANs, but the current net/dsa
supported chips don't handle those anyway.)
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Jamal Hadi Salim
2012-03-13 13:52:07 UTC
Permalink
On Mon, 2012-03-12 at 09:48 +0100, Lennert Buytenhek wrote:

> Since it can lead to problems (address database mismatches, doesn't
> correctly handle STP transitions or topology changes automatically),
> I think it should be avoided whenever possible. I don't see any
> advantages of hardware based learning over software based learning
> anyway ('flexibility' doesn't seem like a very good argument).

Indeed address mismatches may happen if you have two databases.
You have two choices then:
Do learning in user space or be able to tolerate some transient
inconsistency (if you have some software that lazily looks at the
database). But there is a case where the database sits only in hardware.
In such a case, you cant have mismatches.
I think the STP problem can be handled by user space regardless of
whether address mismatch happens or not.

> It should be doable along the lines of the current DSA patch --
> add a VLAN ID argument to the interface add/remove callbacks, and
> when a VLAN virtual interface is added to the bridge, call the
> relevant callbacks with the parent interface + VLAN ID instead.
> (This doesn't work for stacked VLANs, but the current net/dsa
> supported chips don't handle those anyway.)

Sounds like a good start - we could have a different interface for
stacked variants. I think you should push in the patch.

cheers,
jamal
Ben Hutchings
2012-02-16 03:58:48 UTC
Permalink
[I'm just catching up with this after getting my own driver changes into
shape.]

On Fri, 2012-02-10 at 10:18 -0500, jamal wrote:
> Hi John,
>
> I went backwards to summarize at the top after going through your email.
>
> TL;DR version 0.1:
> you provide a good use case where it makes sense to do things in the
> kernel. IMO, you could make the same arguement if your embedded switch
> could do ACLs, IPv4 forwarding etc. And the kernel bloats.
> I am always bigoted to move all policy control to user space instead of
> bloating in the kernel.
[...]
> > Now here is the potential issue,
> >
> > (G) The frame transmitted from ethx.y with the destination address of
> > veth0 but the embedded switch is not a learning switch. If the FDB
> > update is done in user space its possible (likely?) that the FDB
> > entry for veth0 has not been added to the embedded switch yet.
>
> Ok, got it - so the catch here is the switch is not capable of learning.
> I think this depends on where learning is done. Your intent is to
> use the S/W bridge as something that does the learning for you i.e in
> the kernel. This makes the s/w bridge part of MUST-have-for-this-to-run.
> And that maybe the case for your use case.
[...]

Well, in addition, there are SR-IOV network adapters that don't have any
bridge. For these, the software bridge is necessary to handle
multicast, broadcast and forwarding between local ports, not only to do
learning.

Solarflare's implementation of accelerated guest networking (which
Shradha and I are gradually sending upstream) builds on libvirt's
existing support for software bridges and assigns VFs to guests as a
means to offload some of the forwarding.

If and when we implement a hardware bridge, we would probably still want
to keep the software bridge as a fallback. If a guest is dependent on a
VF that's connected to a hardware bridge, it becomes impossible or at
least very disruptive to migrate it to another host that doesn't have a
compatible VF available.

Ben.

--
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.
Shradha Shah
2012-02-16 19:18:57 UTC
Permalink
Hello,

Please find my comments inline.

Regards,
Shradha Shah

On 02/16/2012 03:58 AM, Ben Hutchings wrote:
> [I'm just catching up with this after getting my own driver changes into
> shape.]
>
> On Fri, 2012-02-10 at 10:18 -0500, jamal wrote:
>> Hi John,
>>
>> I went backwards to summarize at the top after going through your email.
>>
>> TL;DR version 0.1:
>> you provide a good use case where it makes sense to do things in the
>> kernel. IMO, you could make the same arguement if your embedded switch
>> could do ACLs, IPv4 forwarding etc. And the kernel bloats.
>> I am always bigoted to move all policy control to user space instead of
>> bloating in the kernel.
> [...]
>>> Now here is the potential issue,
>>>
>>> (G) The frame transmitted from ethx.y with the destination address of
>>> veth0 but the embedded switch is not a learning switch. If the FDB
>>> update is done in user space its possible (likely?) that the FDB
>>> entry for veth0 has not been added to the embedded switch yet.
>>
>> Ok, got it - so the catch here is the switch is not capable of learning.
>> I think this depends on where learning is done. Your intent is to
>> use the S/W bridge as something that does the learning for you i.e in
>> the kernel. This makes the s/w bridge part of MUST-have-for-this-to-run.
>> And that maybe the case for your use case.
> [...]
>
> Well, in addition, there are SR-IOV network adapters that don't have any
> bridge. For these, the software bridge is necessary to handle
> multicast, broadcast and forwarding between local ports, not only to do
> learning.
>
> Solarflare's implementation of accelerated guest networking (which
> Shradha and I are gradually sending upstream) builds on libvirt's
> existing support for software bridges and assigns VFs to guests as a
> means to offload some of the forwarding.

I am also trying to work with bridging using macvtap. Libvirt supports
macvtap in four modes; vepa, bridge, private and passthrough mode.
Macvtap used in the bridge mode will work similar to a software bridge and
will improve performance.

>
> If and when we implement a hardware bridge, we would probably still want
> to keep the software bridge as a fallback. If a guest is dependent on a
> VF that's connected to a hardware bridge, it becomes impossible or at
> least very disruptive to migrate it to another host that doesn't have a
> compatible VF available.
>
> Ben.
>


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
jamal
2012-02-17 14:37:50 UTC
Permalink
On Thu, 2012-02-16 at 03:58 +0000, Ben Hutchings wrote:

>
> Well, in addition, there are SR-IOV network adapters that don't have any
> bridge. For these, the software bridge is necessary to handle
> multicast, broadcast and forwarding between local ports, not only to do
> learning.

For the scenario where there is no h/w bridge - the s/ware bridge should
be usable. There's no way working around that.
My contention is only with the case where there is a h/w bridge and
there being two FDB tables; one in hardware and another in s/w.
And both the h/w and s/w bridges doing flooding and learning.
It is desirable to have options to use one or other or both with
some synchronization.

> Solarflare's implementation of accelerated guest networking (which
> Shradha and I are gradually sending upstream) builds on libvirt's
> existing support for software bridges and assigns VFs to guests as a
> means to offload some of the forwarding.
> If and when we implement a hardware bridge, we would probably still want
> to keep the software bridge as a fallback. If a guest is dependent on a
> VF that's connected to a hardware bridge, it becomes impossible or at
> least very disruptive to migrate it to another host that doesn't have a
> compatible VF available.

In the scheme i described to John in last email, libvirt needs not be aware of
existence of hardware offloading (and migration should be transparent of whether
h/w bridge exists or not)...

cheers,
jamal

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Roopa Prabhu
2012-02-10 13:45:29 UTC
Permalink
On 2/9/12 9:36 AM, "John Fastabend" <***@intel.com> wrote:

> On 2/8/2012 8:36 PM, Stephen Hemminger wrote:
>> On Wed, 08 Feb 2012 19:22:06 -0800
>> John Fastabend <***@intel.com> wrote:
>>
>>> Propagate software FDB table into hardware uc, mc lists when
>>> the NETIF_F_HW_FDB is set.
>>>
>>> This resolves the case below where an embedded switch is used
>>> in hardware to do inter-VF or VF-PF switching. This patch
>>> pushes the FDB entry (specifically the MAC address) into the
>>> embedded switch with dev_add_uc and dev_add_mc so the switch
>>> "learns" about the software bridge.
>>>
>>>
>>> veth0 veth2
>>> | |
>>> ------------
>>> | bridge0 | <---- software bridging
>>> ------------
>>> /
>>> /
>>> ethx.y ethx
>>> VF PF
>>> \ \ <---- propagate FDB entries to HW
>>> \ \
>>> --------------------
>>> | Embedded Bridge | <---- hardware offloaded switching
>>> --------------------
>>>
>>> This is only an RFC couple more changes are needed.
>>>
>>> (1) Optimize HW FDB set/del to only walk list if an FDB offloaded
>>> device is attached. Or decide it doesn't matter from unlikely()
>>> path.
>>>
>>> (2) Is it good enough to just call dev_uc_{add|del} or
>>> dev_mc_{add|del}? Or do some devices really need a new netdev
>>> callback to do this operation correctly. I think it should be
>>> good enough as is.
>>>
>>> (3) wrapped list walk in rcu_read_lock() just in case maybe every
>>> case is already inside rcu_read_lock()/unlock().
>>>
>>> Also this is in response to this thread regarding the macvlan and
>>> exposing rx filters posting now to see if folks think this is the
>>> right idea and if it will resolve at least the bridge case.
>>>
>>> http://lists.openwall.net/netdev/2011/11/08/135
>>>
>>> Signed-off-by: John Fastabend <***@intel.com>
>>> ---
>>>
>>> include/linux/netdev_features.h | 2 ++
>>> net/bridge/br_fdb.c | 34 ++++++++++++++++++++++++++++++++++
>>> 2 files changed, 36 insertions(+), 0 deletions(-)
>>>
>>> diff --git a/include/linux/netdev_features.h
>>> b/include/linux/netdev_features.h
>>> index 77f5202..5936fae 100644
>>
>> Rather than yet another device feature, I would rather use netlink_notifier
>> callback. The notifier is more general and generic without messing with
>> internals
>> of bridge.
>>
>
> But the device features makes it easy for user space to learn that the device
> supports this sort of offload. Now if all SR-IOV devices support this then it
> doesn't matter but I thought there were SR-IOV devices that didn't do any
> switching? I'll dig through the SR-IOV drivers to check there are not too
> many of them.

Correct. Our 802.1Qbh sriov device (enic) does not do local switching.

>
> By netlink_notifier do you mean adding a notifier_block and using
> atomic_notifier_call_chain()
> probably in rtnl_notify()? Then drivers could register with the notifier chain
> with
> atomic_notifier_chain_register() and receive the events correctly. Or did I
> miss
> some notifier chain that already exists?
>
> Thanks,
> John
>

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Sridhar Samudrala
2012-02-09 18:14:47 UTC
Permalink
On Wed, 2012-02-08 at 19:22 -0800, John Fastabend wrote:
> Propagate software FDB table into hardware uc, mc lists when
> the NETIF_F_HW_FDB is set.
>
> This resolves the case below where an embedded switch is used
> in hardware to do inter-VF or VF-PF switching. This patch
> pushes the FDB entry (specifically the MAC address) into the
> embedded switch with dev_add_uc and dev_add_mc so the switch
> "learns" about the software bridge.
>
>
> veth0 veth2
> | |
> ------------
> | bridge0 | <---- software bridging
> ------------
> /
> /
> ethx.y ethx
> VF PF
> \ \ <---- propagate FDB entries to HW
> \ \
> --------------------
> | Embedded Bridge | <---- hardware offloaded switching
> --------------------
>

This scenario works now as adding an interface to a bridge puts it in
promiscuous mode. So adding a PF to a software bridge should not be
a problem as it supports promiscuous mode. But adding a VF will not
work.

Are you trying to avoid the requirement of having to put the interface
in promiscuous mode when adding to a bridge?

Thanks
Sridhar



> This is only an RFC couple more changes are needed.
>
> (1) Optimize HW FDB set/del to only walk list if an FDB offloaded
> device is attached. Or decide it doesn't matter from unlikely()
> path.
>
> (2) Is it good enough to just call dev_uc_{add|del} or
> dev_mc_{add|del}? Or do some devices really need a new netdev
> callback to do this operation correctly. I think it should be
> good enough as is.
>
> (3) wrapped list walk in rcu_read_lock() just in case maybe every
> case is already inside rcu_read_lock()/unlock().
>
> Also this is in response to this thread regarding the macvlan and
> exposing rx filters posting now to see if folks think this is the
> right idea and if it will resolve at least the bridge case.
>
> http://lists.openwall.net/netdev/2011/11/08/135
>
> Signed-off-by: John Fastabend <***@intel.com>
> ---
>
> include/linux/netdev_features.h | 2 ++
> net/bridge/br_fdb.c | 34 ++++++++++++++++++++++++++++++++++
> 2 files changed, 36 insertions(+), 0 deletions(-)
>
> diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h
> index 77f5202..5936fae 100644
> --- a/include/linux/netdev_features.h
> +++ b/include/linux/netdev_features.h
> @@ -55,6 +55,8 @@ enum {
> NETIF_F_NOCACHE_COPY_BIT, /* Use no-cache copyfromuser */
> NETIF_F_LOOPBACK_BIT, /* Enable loopback */
>
> + NETIF_F_HW_FDB, /* Hardware supports switching */
> +
> /*
> * Add your fresh new feature above and remember to update
> * netdev_features_strings[] in net/core/ethtool.c and maybe
> diff --git a/net/bridge/br_fdb.c b/net/bridge/br_fdb.c
> index 5ba0c84..4cc545b 100644
> --- a/net/bridge/br_fdb.c
> +++ b/net/bridge/br_fdb.c
> @@ -81,9 +81,26 @@ static void fdb_rcu_free(struct rcu_head *head)
> kmem_cache_free(br_fdb_cache, ent);
> }
>
> +static void fdb_hw_delete(struct net_bridge *br,
> + struct net_bridge_fdb_entry *fdb)
> +{
> + struct net_bridge_port *op;
> +
> + rcu_read_lock();
> + list_for_each_entry_rcu(op, &br->port_list, list) {
> + struct net_device *dev = op->dev;
> +
> + if ((dev->features & NETIF_F_HW_FDB) &&
> + dev != fdb->dst->dev)
> + dev_uc_del(dev, fdb->addr.addr);
> + }
> + rcu_read_unlock();
> +}
> +
> static void fdb_delete(struct net_bridge *br, struct net_bridge_fdb_entry *f)
> {
> hlist_del_rcu(&f->hlist);
> + fdb_hw_delete(br, f);
> fdb_notify(br, f, RTM_DELNEIGH);
> call_rcu(&f->rcu, fdb_rcu_free);
> }
> @@ -350,6 +367,22 @@ static struct net_bridge_fdb_entry *fdb_find_rcu(struct hlist_head *head,
> return NULL;
> }
>
> +static void fdb_hw_create(struct net_bridge *br,
> + struct net_bridge_fdb_entry *fdb)
> +{
> + struct net_bridge_port *op;
> +
> + rcu_read_lock();
> + list_for_each_entry_rcu(op, &br->port_list, list) {
> + struct net_device *dev = op->dev;
> +
> + if ((dev->features & NETIF_F_HW_FDB) &&
> + dev != fdb->dst->dev)
> + dev_uc_add(dev, fdb->addr.addr);
> + }
> + rcu_read_unlock();
> +}
> +
> static struct net_bridge_fdb_entry *fdb_create(struct hlist_head *head,
> struct net_bridge_port *source,
> const unsigned char *addr)
> @@ -363,6 +396,7 @@ static struct net_bridge_fdb_entry *fdb_create(struct hlist_head *head,
> fdb->is_local = 0;
> fdb->is_static = 0;
> fdb->updated = fdb->used = jiffies;
> + fdb_hw_create(source->br, fdb);
> hlist_add_head_rcu(&fdb->hlist, head);
> }
> return fdb;
>


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-09 20:30:36 UTC
Permalink
On 2/9/2012 10:14 AM, Sridhar Samudrala wrote:
> On Wed, 2012-02-08 at 19:22 -0800, John Fastabend wrote:
>> Propagate software FDB table into hardware uc, mc lists when
>> the NETIF_F_HW_FDB is set.
>>
>> This resolves the case below where an embedded switch is used
>> in hardware to do inter-VF or VF-PF switching. This patch
>> pushes the FDB entry (specifically the MAC address) into the
>> embedded switch with dev_add_uc and dev_add_mc so the switch
>> "learns" about the software bridge.
>>
>>
>> veth0 veth2
>> | |
>> ------------
>> | bridge0 | <---- software bridging
>> ------------
>> /
>> /
>> ethx.y ethx
>> VF PF
>> \ \ <---- propagate FDB entries to HW
>> \ \
>> --------------------
>> | Embedded Bridge | <---- hardware offloaded switching
>> --------------------
>>
>
> This scenario works now as adding an interface to a bridge puts it in
> promiscuous mode. So adding a PF to a software bridge should not be
> a problem as it supports promiscuous mode. But adding a VF will not
> work.

It shouldn't work because the embedded bridge will lookup the address
in its FDB and when it doesn't match any unicast filters it will forward
the packet onto the wire. Because the veth0 and veth2 above never get
inserted into the embedded brdige's FDB the packets will _never_ get
routed there.

That said the current 'ixgbe' driver is doing something broken in that
it is always setting the unicast hash table and mirroring bits to 1. So
if you think this is working your seeing a bug where packets are being
sent onto the wire AND upto the PF. Packets with destination addresses
matching veth1 should not end up on the wire and vice versa. This is
specific to ixgbe and is not the case for other SR-IOV devices.

This causes some issues (a) has some very real performance implications,
(b) at this point you have some strange behavior from my point of view.
The embedded bridge is not a learning bridge nor is it acting like an
802.1Q VEB or VEPA.

>
> Are you trying to avoid the requirement of having to put the interface
> in promiscuous mode when adding to a bridge?
>

I think the bridge being in promiscuous mode is correct.

Hope that helps sorry its a bit long winded.
John



--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Sridhar Samudrala
2012-02-10 00:39:09 UTC
Permalink
On Thu, 2012-02-09 at 12:30 -0800, John Fastabend wrote:
> On 2/9/2012 10:14 AM, Sridhar Samudrala wrote:
> > On Wed, 2012-02-08 at 19:22 -0800, John Fastabend wrote:
> >> Propagate software FDB table into hardware uc, mc lists when
> >> the NETIF_F_HW_FDB is set.
> >>
> >> This resolves the case below where an embedded switch is used
> >> in hardware to do inter-VF or VF-PF switching. This patch
> >> pushes the FDB entry (specifically the MAC address) into the
> >> embedded switch with dev_add_uc and dev_add_mc so the switch
> >> "learns" about the software bridge.
> >>
> >>
> >> veth0 veth2
> >> | |
> >> ------------
> >> | bridge0 | <---- software bridging
> >> ------------
> >> /
> >> /
> >> ethx.y ethx
> >> VF PF
> >> \ \ <---- propagate FDB entries to HW
> >> \ \
> >> --------------------
> >> | Embedded Bridge | <---- hardware offloaded switching
> >> --------------------
> >>
> >
> > This scenario works now as adding an interface to a bridge puts it in
> > promiscuous mode. So adding a PF to a software bridge should not be
> > a problem as it supports promiscuous mode. But adding a VF will not
> > work.
>
> It shouldn't work because the embedded bridge will lookup the address
> in its FDB and when it doesn't match any unicast filters it will forward
> the packet onto the wire. Because the veth0 and veth2 above never get
> inserted into the embedded brdige's FDB the packets will _never_ get
> routed there.
>
> That said the current 'ixgbe' driver is doing something broken in that
> it is always setting the unicast hash table and mirroring bits to 1. So
> if you think this is working your seeing a bug where packets are being
> sent onto the wire AND upto the PF. Packets with destination addresses
> matching veth1 should not end up on the wire and vice versa. This is
> specific to ixgbe and is not the case for other SR-IOV devices.

OK. Is this behavior going to be fixed.

>
> This causes some issues (a) has some very real performance implications,
> (b) at this point you have some strange behavior from my point of view.
> The embedded bridge is not a learning bridge nor is it acting like an
> 802.1Q VEB or VEPA.
>
> >
> > Are you trying to avoid the requirement of having to put the interface
> > in promiscuous mode when adding to a bridge?
> >
>
> I think the bridge being in promiscuous mode is correct.

The interface that is added to the bridge is put in promiscuous mode,
not the bridge itself. In this example, i assumed that setting
promiscuous on PF is putting the embedded bridge in learning mode.

Thanks
Sridhar

>
> Hope that helps sorry its a bit long winded.
> John
>
>
>


--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
John Fastabend
2012-02-10 00:51:18 UTC
Permalink
On 2/9/2012 4:39 PM, Sridhar Samudrala wrote:
> On Thu, 2012-02-09 at 12:30 -0800, John Fastabend wrote:
>> On 2/9/2012 10:14 AM, Sridhar Samudrala wrote:
>>> On Wed, 2012-02-08 at 19:22 -0800, John Fastabend wrote:
>>>> Propagate software FDB table into hardware uc, mc lists when
>>>> the NETIF_F_HW_FDB is set.
>>>>
>>>> This resolves the case below where an embedded switch is used
>>>> in hardware to do inter-VF or VF-PF switching. This patch
>>>> pushes the FDB entry (specifically the MAC address) into the
>>>> embedded switch with dev_add_uc and dev_add_mc so the switch
>>>> "learns" about the software bridge.
>>>>
>>>>
>>>> veth0 veth2
>>>> | |
>>>> ------------
>>>> | bridge0 | <---- software bridging
>>>> ------------
>>>> /
>>>> /
>>>> ethx.y ethx
>>>> VF PF
>>>> \ \ <---- propagate FDB entries to HW
>>>> \ \
>>>> --------------------
>>>> | Embedded Bridge | <---- hardware offloaded switching
>>>> --------------------
>>>>
>>>
>>> This scenario works now as adding an interface to a bridge puts it in
>>> promiscuous mode. So adding a PF to a software bridge should not be
>>> a problem as it supports promiscuous mode. But adding a VF will not
>>> work.
>>
>> It shouldn't work because the embedded bridge will lookup the address
>> in its FDB and when it doesn't match any unicast filters it will forward
>> the packet onto the wire. Because the veth0 and veth2 above never get
>> inserted into the embedded brdige's FDB the packets will _never_ get
>> routed there.
>>
>> That said the current 'ixgbe' driver is doing something broken in that
>> it is always setting the unicast hash table and mirroring bits to 1. So
>> if you think this is working your seeing a bug where packets are being
>> sent onto the wire AND upto the PF. Packets with destination addresses
>> matching veth1 should not end up on the wire and vice versa. This is
>> specific to ixgbe and is not the case for other SR-IOV devices.
>
> OK. Is this behavior going to be fixed.
>

Only after we have a mechanism to either configure the NIC FDB directly
or have it stay in sync with the SW switch. Flooding traffic seems better
than being unable to send traffic to the virtual device altogether. This
behavior is driver specific some devices just fail outright.

I'm thinking over Jamal's comment now.

>>
>> This causes some issues (a) has some very real performance implications,
>> (b) at this point you have some strange behavior from my point of view.
>> The embedded bridge is not a learning bridge nor is it acting like an
>> 802.1Q VEB or VEPA.
>>
>>>
>>> Are you trying to avoid the requirement of having to put the interface
>>> in promiscuous mode when adding to a bridge?
>>>
>>
>> I think the bridge being in promiscuous mode is correct.
>
> The interface that is added to the bridge is put in promiscuous mode,
> not the bridge itself. In this example, i assumed that setting
> promiscuous on PF is putting the embedded bridge in learning mode.


Yes I misspoke I mean the PF. The embedded bridge in this case does not
support learning. Also I'm not aware of any SR-IOV NICs that do support
learning.

>
> Thanks
> Sridhar
>
>>
>> Hope that helps sorry its a bit long winded.
>> John
>>
>>
>>
>
>
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to ***@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html

--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to ***@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Loading...