Date:   Mon, 20 Jul 2020 13:00:37 +0300
From:   Vladimir Oltean <olteanv@...il.com>
To:     Ido Schimmel <idosch@...sch.org>
Cc:     Andrew Lunn <andrew@...n.ch>,
        Florian Fainelli <f.fainelli@...il.com>,
        Vivien Didelot <vivien.didelot@...il.com>,
        "David S. Miller" <davem@...emloft.net>,
        Jiri Pirko <jiri@...nulli.us>,
        Jakub Kicinski <kuba@...nel.org>,
        Ivan Vecera <ivecera@...hat.com>, vyasevich@...il.com,
        netdev <netdev@...r.kernel.org>, UNGLinuxDriver@...rochip.com,
        Nikolay Aleksandrov <nikolay@...ulusnetworks.com>,
        Roopa Prabhu <roopa@...ulusnetworks.com>
Subject: Re: [PATCH RFC net-next 00/13] RX filtering for DSA switches

On Thu, May 28, 2020 at 05:37:18PM +0300, Ido Schimmel wrote:
> On Wed, May 27, 2020 at 02:36:53PM +0300, Vladimir Oltean wrote:
> > On Tue, 26 May 2020 at 17:02, Ido Schimmel <idosch@...sch.org> wrote:
> > >
> > > On Mon, May 25, 2020 at 11:23:34PM +0300, Vladimir Oltean wrote:
> > > > Hi Ido,
> > > >
> > > > On Mon, 25 May 2020 at 22:48, Ido Schimmel <idosch@...sch.org> wrote:
> > > > >
> > > > > On Sun, May 24, 2020 at 07:24:27PM +0300, Vladimir Oltean wrote:
> > > > > > On Sun, 24 May 2020 at 17:07, Ido Schimmel <idosch@...sch.org> wrote:
> > > > > > >
> > > > > > > On Fri, May 22, 2020 at 12:10:23AM +0300, Vladimir Oltean wrote:
> > > > > > > > From: Vladimir Oltean <vladimir.oltean@....com>
> > > > > > > >
> > > > > > > > This is a WIP series whose stated goal is to allow DSA and switchdev
> > > > > > > > drivers to flood less traffic to the CPU while keeping the same level of
> > > > > > > > functionality.
> > > > > > > >
> > > > > > > > The strategy is to whitelist towards the CPU only the {DMAC, VLAN} pairs
> > > > > > > > that the operating system has expressed its interest in, either due to
> > > > > > > > those being the MAC addresses of one of the switch ports, or addresses
> > > > > > > > added to our device's RX filter via calls to dev_uc_add/dev_mc_add.
> > > > > > > > Then, the traffic which is not explicitly whitelisted is not sent by the
> > > > > > > > hardware to the CPU, under the assumption that the CPU didn't ask for it
> > > > > > > > and would have dropped it anyway.
> > > > > > > >
> > > > > > > > The ground for these patches were the discussions surrounding RX
> > > > > > > > filtering with switchdev in general, as well as with DSA in particular:
> > > > > > > >
> > > > > > > > "[PATCH net-next 0/4] DSA: promisc on master, generic flow dissector code":
> > > > > > > > https://www.spinics.net/lists/netdev/msg651922.html
> > > > > > > > "[PATCH v3 net-next 2/2] net: dsa: felix: Allow unknown unicast traffic towards the CPU port module":
> > > > > > > > https://www.spinics.net/lists/netdev/msg634859.html
> > > > > > > > "[PATCH v3 0/2] net: core: Notify on changes to dev->promiscuity":
> > > > > > > > https://lkml.org/lkml/2019/8/29/255
> > > > > > > > LPC2019 - SwitchDev offload optimizations:
> > > > > > > > https://www.youtube.com/watch?v=B1HhxEcU7Jg
> > > > > > > >
> > > > > > > > Unicast filtering comes to me as most important, and this includes
> > > > > > > > termination of MAC addresses corresponding to the network interfaces in
> > > > > > > > the system (DSA switch ports, VLAN sub-interfaces, bridge interface).
> > > > > > > > The first 4 patches use Ivan Khoronzhuk's IVDF framework for extending
> > > > > > > > network interface addresses with a Virtual ID (typically VLAN ID). This
> > > > > > > > matches DSA switches perfectly because their FDB already contains keys
> > > > > > > > of the {DMAC, VID} form.
> > > > > > >
> > > > > > > Hi,
> > > > > > >
> > > > > > > I read through the series and I'm not sure how unicast filtering works.
> > > > > > > Instead of writing a very long mail I just created a script with
> > > > > > > comments. I think it's clearer that way. Note that this is not a made up
> > > > > > > configuration. It is used in setups involving VRRP / VXLAN, for example.
> > > > > > >
> > > > > > > ```
> > > > > > > #!/bin/bash
> > > > > > >
> > > > > > > ip netns add ns1
> > > > > > >
> > > > > > > ip -n ns1 link add name br0 type bridge vlan_filtering 1
> > > > > > > ip -n ns1 link add name dummy10 up type dummy
> > > > > > >
> > > > > > > ip -n ns1 link set dev dummy10 master br0
> > > > > > > ip -n ns1 link set dev br0 up
> > > > > > >
> > > > > > > ip -n ns1 link add link br0 name vlan10 up type vlan id 10
> > > > > > > bridge -n ns1 vlan add vid 10 dev br0 self
> > > > > > >
> > > > > > > echo "Before adding macvlan:"
> > > > > > > echo "======================"
> > > > > > >
> > > > > > > echo -n "Promiscuous mode: "
> > > > > > > ip -n ns1 -j -p -d link show dev br0 | jq .[][\"promiscuity\"]
> > > > > > >
> > > > > > > echo -e "\nvlan10's MAC is in br0's FDB:"
> > > > > > > bridge -n ns1 fdb show br0 vlan 10
> > > > > > >
> > > > > > > echo
> > > > > > > echo "After adding macvlan:"
> > > > > > > echo "====================="
> > > > > > >
> > > > > > > ip -n ns1 link add link vlan10 name vlan10-v up address 00:00:5e:00:01:01 \
> > > > > > >         type macvlan mode private
> > > > > > >
> > > > > > > echo -n "Promiscuous mode: "
> > > > > > > ip -n ns1 -j -p -d link show dev br0 | jq .[][\"promiscuity\"]
> > > > > > >
> > > > > > > echo -e "\nvlan10-v's MAC is not in br0's FDB:"
> > > > > > > bridge -n ns1 fdb show br0 | grep master | grep 00:00:5e:00:01:01
> > > > > > > ```
> > > > > > >
> > > > > > > This is the output on my laptop (kernel 5.6.8):
> > > > > > >
> > > > > > > ```
> > > > > > > Before adding macvlan:
> > > > > > > ======================
> > > > > > > Promiscuous mode: 0
> > > > > > >
> > > > > > > vlan10's MAC is in br0's FDB:
> > > > > > > 42:bd:b1:cc:67:15 dev br0 vlan 10 master br0 permanent
> > > > > > >
> > > > > > > After adding macvlan:
> > > > > > > =====================
> > > > > > > Promiscuous mode: 1
> > > > > > >
> > > > > > > vlan10-v's MAC is not in br0's FDB:
> > > > > > > ```
> > > > > > >
> > > > > > > Basically, if the MAC of the VLAN device is not inherited from the
> > > > > > > bridge or you stack macvlans on top, then the bridge will go into
> > > > > > > promiscuous mode and it will locally receive all frames passing through
> > > > > > > it. It's not ideal, but it's a very old and simple behavior. It does not
> > > > > > > require you to track the VLAN associated with the MAC addresses, for
> > > > > > > example.
> > > > > > >
> > > > > >
> > > > > > This is a good point. I wasn't aware that the bridge 'gives up' with
> > > > > > macvlan upper devices, but if I understand correctly, we do have the
> > > > > > necessary tools to improve that.
> > > > > > But actually, I'm wondering if this simple behavior from the bridge is
> > > > > > correct.
> > > > >
> > > > > Why would it be incorrect?
> > > > >
> > > > > > As you, Jiri and Ivan pointed out in last summer's email
> > > > > > thread about the Linux bridge and promiscuous mode, putting the
> > > > > > interface in IFF_PROMISC is only going to guarantee acceptance through
> > > > > > the net device's RX filter, but not that the packets will go to the
> > > > > > CPU.
> > > > >
> > > > > IFF_PROMISC has no bearing on whether a packet should go to the CPU or
> > > > > not. It only influences the device's RX filter, like you said. If you
> > > > > only look at the software data path, the bridge being in promiscuous
> > > > > mode means that all received packets will be injected to the kernel's Rx
> > > > > path as if they were received through the bridge device. This includes,
> > > > > for example, an IPv4 packet with an unknown unicast MAC (does not
> > > > > correspond to your MAC). Such a packet will be later dropped by the IPv4
> > > > > code since it's not addressed to you:
> > > > >
> > > > > vi net/ipv4/ip_input.c +443
> > > > >
> > > > > We maintain the same behavior in the hardware data path. We don't have
> > > > > MAC filtering in the router like the software data path, so we only send
> > > > > to the router unicast MACs that correspond to the bridge's MAC and its
> > > > > uppers. If such packets later hit a local route (for example), then they
> > > > > will be trapped to the CPU, but the more common case is to simply route
> > > > > them through a different device due to a prefix / gateway route. These
> > > > > never reach the CPU.
> > > > >
> > > > > > So from that perspective, the current series would break things, so we
> > > > > > should definitely fix that and keep the {MAC, VLAN} pairs in the
> > > > > > bridge's local FDB.
> > > > >
> > > > > Not sure I follow. Can you explain what will break and why?
> > > > >
> > > >
> > > > I haven't done any further testing since yesterday, so my level of
> > > > (mis)understanding is the same. Let's hope at least I can explain
> > > > better this time.
> > > >
> > > > I guess what I didn't understand from your "macvlan upper whose MAC
> > > > address isn't inherited from bridge" is why does the bridge go in
> > > > promiscuous mode.
> > >
> > > Packets received from bridge slaves with DMAC equal to an active bridge
> > > upper (e.g., macvlan) should be received by this upper. When a packet is
> > > received from a bridge slave it performs FDB lookup. Since {VID, MAC}
> > > entries are not programmed for bridge uppers, packets addressed to such
> > > addresses will incur an FDB miss and be flooded. If the bridge is not in
> > > promiscuous mode, these packets will not be received via the bridge
> > > interface and will not reach the relevant upper device.
> > >
> > > > You said that it's so that the slave ports won't drop packets with
> > > > that DMAC,
> > >
> > > I did not say that. I explained above that if promiscuous mode is not
> > > enabled on the bridge interface itself (a soft device), the packet will
> > > not be received via the bridge interface and will not reach the upper
> > > device.
> > >
> > > > I said ok, yes the packets would get dropped without promisc, but also
> > > > promisc still doesn't mean the packets will land on the CPU. This is
> > > > one of the cases where the bridge puts an interface in promisc mode
> > > > with the intention of making the CPU see some frames,
> > >
> > > The statement "the bridge puts an interface in promisc mode with the
> > > intention of making the CPU see some frames" is incorrect. The bridge
> > > puts an interface in promiscuous mode so that the bridge will see all
> > > the frames received by this interface. If the bridge is offloaded,
> > > bridging happens in hardware and there is no reason to send all the
> > > frames to the CPU.
> > >
> > > > something which has been argued, in the context of switchdev, that was
> > > > never the case. You said that's all true, and that in mlxsw you're
> > > > giving the bridge a helping hand, by tracking the bridge's uppers in
> > > > order to keep something that works by accident in software working
> > > > with switchdev too.
> > >
> > > I never said that the software bridge works by accident. I explained
> > > why, to my understanding, the bridge works the way it's working and what
> > > can be done in order to prevent the bridge from going into promiscuous
> > > mode. It involves very careful (and error-prone?) tracking of the upper
> > > devices and their VLANs.
> > >
> > > Also, please differentiate between the bridge interface itself going
> > > into promiscuous mode and bridge slaves going into promiscuous mode.
> > >
> > > > I said that this is a
> > > > weird layering violation, because the bridge's job is to notify the
> > > > driver of addresses it needs to see, not for the driver to fish for
> > > > them.
> > > > As for "what will break and why". My current patch proposal is to only
> > > > send to the CPU the addresses added via dev_uc_add and dev_mc_add,
> > > > basically. The macvlan upper of the bridge would not be part of that
> > > > list. My rhetorical question then becomes: whose fault is it that
> > > > macvlan breaks? Mine for not tracking the bridge upper, or the bridge
> > > > for not notifying me and just pretending that 'promisc' means 'the CPU
> > > > will see all packets, including the ones I need'? Of course I think
> > > > it's the bridge.
> > > >
> > 
> > Ok, bridge promisc vs slave promisc is not a difference I explicitly
> > made, but my point is actually beyond that.
> > The bridge going in promisc will only help if the packets are sent to
> > the CPU in the first place. And it does nothing to ensure that that
> > will happen. So the bridge code works by accident.
> 
> It's not beyond your point and the bridge code does not work by
> accident. When the bridge interface is in promiscuous mode every packet
> is injected to the kernel's Rx path as if it was accepted by the bridge
> interface. Packets then reach the protocol handlers. In the case of
> IPv4/IPv6, the packets go to ipv_rcv() / ip6_rcv() and perform routing.
> Packets with a unicast destination MAC that does not correspond to that
> of the receiving interface are dropped.
> 

Hi Ido,

I still maintain that the bridge and the network stack in general don't
have a proper design for managing a switchdev's filter of what goes to
the CPU, let me explain.

The whole purpose of my patch series is to remove the CPU port from the
flood domain of all switchdev net_devices. That means, when an unknown
unicast packet ingresses, it will be flooded but not to the CPU. For
frames that the CPU wants to see, there should be a universal mechanism
for it to whitelist them, by {DMAC, VID}. Otherwise, things don't scale.

There is one such mechanism already, and that is dev_uc_add(). It used
to install an address into a device's RX filter using DMAC only, and
Ivan Khoronzhuk's patches have added a new dev_vid_uc_add() that allow
additional filtering by VLAN.

That mechanism used to have a meaning. Its meaning was: for a
non-promisc net_device, don't drop unicast frames having a MAC address
equal to the argument passed to dev_uc_add().

For a promiscuous net_device, that is not needed, because promisc means
that no frames should be dropped due to MAC/VID match.

For a switchdev, promisc vs non-promisc doesn't mean a thing. A switch
is a switch, it's promiscuous by definition. It is _supposed_ to accept
traffic regardless of destination. There isn't even a mechanism in the
switchdevs that I know of to install an RX filter in the ingress MAC (or
if there is, it filters by source MAC address, and it's done for
different reasons). This is fundamentally because the destination MAC
address is parsed by a network card for _termination_ purposes. And
because a switch doesn't do _termination_, there is no reason to filter
by destination MAC (ignore ACL and such). But a Linux switchdev is
capable of termination. In the case of switchdev, termination means
sending to the CPU.

My interpretation of the meaning of dev_uc_add() for switchdev (and
therefore, of its opposite - promiscuous mode) is at odds with previous
work done for non-switchdev. Take Vlad Yasevich's work "[Bridge] [PATCH
net-next 0/8] Non-promisc bidge ports support" for example:

https://lists.linuxfoundation.org/pipermail/bridge/2014-May/008940.html

He is arguing that a bridge port without flood&learn doesn't need
promiscuous mode, because all addresses can be statically known, and
therefore, he added code to the bridge that does the following:

- syncs the bridge MAC address to all non-promisc bridge slaves, via
  dev_uc_add()
- syncs the MAC addresses of all static FDB entries on all ingress
  non-promisc bridge slave ports, via dev_uc_add()

with the obvious goal that "the bridge slave shouldn't drop these
packets".

In my interpretation of dev_uc_add(), I would have expected that:
- the bridge MAC address, as well as any other secondary unicast
  addresses that the bridge has, by means of its uppers (like macvlan,
  802.1q, etc) calling dev_uc_add() on it, would be synced to the bridge
  slaves anyway, regardless of whether they're promisc or not
- the static FDB entries are synced to the bridge ports only in the
  non-switchdev case. This is because for switchdev, I am treating a
  dev_uc_add() as a FDB entry towards the CPU, and therefore this would
  overwrite the FDB entry towards the external port.

In my interpretation, things would have worked neatly for the most part,
not only for unicast but also for multicast. For example, an application
wants to see a multicast stream, so it calls setsockopt(SOL_SOCKET,
PACKET_ADD_MEMBERSHIP, PACKET_MR_MULTICAST) with the multicast address
it wants to see. This is translated by the kernel into a dev_mc_add()
and sent to the network device. For a non-switchdev, this would have
been enough. For a switchdev, if I also installed the address in the
CPU's filter, it would have also been enough. Things 'just work' and
everybody's happy.

> When you look at it from hardware offload perspective, not every packet
> received by the bridge interface should reach the CPU. Actually, most
> should not reach it. Otherwise it would mean that every routed packet
> would need to go to the CPU, which is not feasible. If you can't perform
> routing in hardware, then yes, you need to send such packets to the CPU.
> 
> In mlxsw we can't perform MAC filtering in the router like in the
> software data path, so in order not to route packets we should not, we
> only send to the router packets with destination MACs that correspond to
> that of the bridge or one of its uppers. We don't flood all unknown
> unicast packets there.
> 
> In the case of hardware offload it's relatively easy to do this sort of
> tracking because only a limited set of upper devices topologies are
> actually supported. I'm not sure how feasible it is with every
> combination of upper devices supported by the kernel. It seems easiest
> to just put the bridge interface in promiscuous mode and let upper
> layers perform the filtering. Like it is today.
> 

Are you suggesting that tracking the uppers is the only way to do what I
want? (I didn't even find that piece of code in mlxsw, btw).

I am a bit reluctant to do such management at driver level. It is not a
driver problem, it is a switchdev design question. I shouldn't need to
care if there's a macvlan or an 802.1q or a bridge upper or whatnot, and
how many addresses those are listening to, as long as those network
interfaces can tell me what addresses they want to see, and as long as I
can interpret that information as the list of addresses I should be
delivering to the CPU. It is clear that right now, some of the uses of
dev_uc_add() are simply there to prevent drops, and not because the
bridge has a particular interest in seeing those frames. So either the
meaning of dev_uc_add() changes, and a meaning is standardized for
promisc on a switchdev port, or we add a new set of functions, such as
dev_cpu_filter_uc_add() specifically for switchdev and spray them
throughout the network stack, mostly in the places where dev_uc_add() is
currently also used.

Bridging with a non-switchdev interface is also a situation that should
be dealt with generically, as in that case, the CPU filter should
obviously become larger, as termination is no longer done just on this
CPU, and plain unicast filtering is no longer enough.

I hope it's a bit clearer now what is the problem I'm trying to address.
Needless to say, I would prefer that a new API is not introduced,
because upper layers shouldn't necessarily care about switchdev, unless
they can leverage it for offloading.

> > 
> > > > > >
> > > > > > > When you are offloading the Linux data path to hardware this behavior is
> > > > > > > not ideal as your hardware can handle much higher packet rates than the
> > > > > > > CPU.
> > > > > > >
> > > > > > > In mlxsw we handle this by tracking the upper devices of the bridge. I
> > > > > > > was hoping that with Ivan's patches we could add support for unicast
> > > > > > > filtering in the bridge driver and program the MAC addresses to its FDB
> > > > > > > with 'local' flag. Then the FDB entries would be notified via switchdev
> > > > > > > to device drivers.
> > > > > > >
> > > > > >
> > > > > > Yes, it should be possible to do that. I'll try and see how far I get.
> > > > > >
> > > > > > > >
> > > > > > > > Multicast filtering was taken and reworked from Florian Fainelli's
> > > > > > > > previous attempts, according to my own understanding of multicast
> > > > > > > > forwarding requirements of an IGMP snooping switch. This is the part
> > > > > > > > that needs the most extra work, not only in the DSA core but also in
> > > > > > > > drivers. For this reason, I've left out of this patchset anything that
> > > > > > > > has to do with driver-level configuration (since the audience is a bit
> > > > > > > > larger than usual), as I'm trying to focus more on policy for now, and
> > > > > > > > the series is already pretty huge.
> > > > > > >
> > > > > > > From what I remember, this is the logic in the Linux bridge:
> > > > > > >
> > > > > > > * Broadcast is always locally received
> > > > > > > * Multicast is locally received if:
> > > > > > >         * Snooping disabled
> > > > > > >         * Snooping enabled:
> > > > > > >                 * Bridge netdev is mrouter port
> > > > > > >                 or
> > > > > > >                 * Matches MDB entry with 'host_joined' indication
> > > > > > >
> > > > > > > >
> > > > > > > > Florian Fainelli (3):
> > > > > > > >   net: bridge: multicast: propagate br_mc_disabled_update() return
> > > > > > > >   net: dsa: add ability to program unicast and multicast filters for CPU
> > > > > > > >     port
> > > > > > > >   net: dsa: wire up multicast IGMP snooping attribute notification
> > > > > > > >
> > > > > > > > Ivan Khoronzhuk (4):
> > > > > > > >   net: core: dev_addr_lists: add VID to device address
> > > > > > > >   net: 8021q: vlan_dev: add vid tag to addresses of uc and mc lists
> > > > > > > >   net: 8021q: vlan_dev: add vid tag for vlan device own address
> > > > > > > >   ethernet: eth: add default vid len for all ethernet kind devices
> > > > > > > >
> > > > > > > > Vladimir Oltean (6):
> > > > > > > >   net: core: dev_addr_lists: export some raw __hw_addr helpers
> > > > > > > >   net: dsa: don't use switchdev_notifier_fdb_info in
> > > > > > > >     dsa_switchdev_event_work
> > > > > > > >   net: dsa: mroute: don't panic the kernel if called without the prepare
> > > > > > > >     phase
> > > > > > > >   net: bridge: add port flags for host flooding
> > > > > > > >   net: dsa: deal with new flooding port attributes from bridge
> > > > > > > >   net: dsa: treat switchdev notifications for multicast router connected
> > > > > > > >     to port
> > > > > > > >
> > > > > > > >  include/linux/if_bridge.h |   3 +
> > > > > > > >  include/linux/if_vlan.h   |   2 +
> > > > > > > >  include/linux/netdevice.h |  11 ++
> > > > > > > >  include/net/dsa.h         |  17 +++
> > > > > > > >  net/8021q/Kconfig         |  12 ++
> > > > > > > >  net/8021q/vlan.c          |   3 +
> > > > > > > >  net/8021q/vlan.h          |   2 +
> > > > > > > >  net/8021q/vlan_core.c     |  25 ++++
> > > > > > > >  net/8021q/vlan_dev.c      | 102 +++++++++++---
> > > > > > > >  net/bridge/br_if.c        |  40 ++++++
> > > > > > > >  net/bridge/br_multicast.c |  21 ++-
> > > > > > > >  net/bridge/br_switchdev.c |   4 +-
> > > > > > > >  net/core/dev_addr_lists.c | 144 +++++++++++++++----
> > > > > > > >  net/dsa/Kconfig           |   1 +
> > > > > > > >  net/dsa/dsa2.c            |   6 +
> > > > > > > >  net/dsa/dsa_priv.h        |  27 +++-
> > > > > > > >  net/dsa/port.c            | 155 ++++++++++++++++----
> > > > > > > >  net/dsa/slave.c           | 288 +++++++++++++++++++++++++++++++-------
> > > > > > > >  net/dsa/switch.c          |  36 +++++
> > > > > > > >  net/ethernet/eth.c        |  12 +-
> > > > > > > >  20 files changed, 780 insertions(+), 131 deletions(-)
> > > > > > > >
> > > > > > > > --
> > > > > > > > 2.25.1
> > > > > > > >
> > > > > >
> > > > > > Thanks,
> > > > > > -Vladimir
> > > >
> > > > -Vladimir
> > 
> > Thanks,
> > -Vladimir

Thanks,
-Vladimir

Powered by blists - more mailing lists