When preparing the migration from macvlan to veth for local-node, MAC
address conflicts occurred as some ports of br-client had the same address
as local-node. Reverting the roles of both interfaces fixes this.
By default, br-client is left as an interface without addresses and
firewall rules that drop everything, so the bridge is used to connect its
ports only. gluon-mesh-batman-adv-core changes this to the usual set
of addresses and firewall rules.
The batadv debugfs requires large memory blocks to write the text debug
tables. This is inefficient for large tables like the global translation
table or the originators table.
The memory requirement can be reduced by using netlink. It copies smaller
packets in a binary format to the userspace program. The respondd module of
gluon-mesh-batman-adv-core can therefore parse larger originator tables
without causing an OOM on systems which are tight on memory.
Signed-off-by: Sven Eckelmann <sven@narfation.org>
MAC and IP addresses are switched. This makes the gluon-client-bridge
package more useful for different routing protocols that don't need a
unique address on the client bridge.
As a side effect, gluon-radvd is now using the next-node address, which had
been considered before, but was dismissed to avoid having gluon-radvd
depend on gluon-next-node and gluon-mesh-batman-adv. This will be useful
for announcing default routes via gluon-radvd.
One downside is that this introduces a minor dependency on batman-adv in
gluon-respondd: the hotplug script that checked for the client interface
before will now check for local-node. This doesn't really matter: for mesh
protocols without a local-node interface, the check will do nothing (which
makes sense, as there is no interface to bind to for mesh-wide respondd).
Because we unconditionally appended `-i br-client` to the command line of
respondd, it wasn't restarted when br-client changed state. Now, we use a
jsonfilter expression on the network.interface dump data, similar to how the
other interface names are generated, and only add the interface to the argument
list if it is up.
Switch to:
1. WAN
2. LAN
3. Mesh VPN
As WAN and LAN are setup in gluon-mesh-batman-adv-core (and will be moved
to gluon-core), while the mesh VPN has its own package, giving WAN and LAN
the first indices is preferable.
A few issues with the bridge snooping were identified and fixed
upstream in OpenWRT:
* "firewall: Allow IGMP and MLD input on WAN" (r45613)
* "kernel: bridge: backport two snooping related patches" (r45783)
* netifd: "bridge: Fix multicast_to_unicast feature by hairpin+isolate"
(OW: "netifd: update to the latest version, adds multicast-to-unicast fixes" (r46719))
* "kernel: bridge, multicast-to-unicast: assign src after pskb_may_pull()" (r46721)
* "kernel: bridge, multicast-to-unicast: fix echoes on STA" (46765)
These have very likely caused issues with the bridge snooping before,
which led to disabling it in the past. Let's reenable the multicast
snooping now that they were fixed for reduced multicast overhead on the
wifi.
Advantages are the following:
This mildly reduces overhead on the mesh layer. And significantly reduces
overhead on the AP interface and therefore significantly increases
available airtime (the currently most significant scalability bottleneck).
Secondly removes an easy, often accidental node-local Denial-of-Service
vector based on multicast flooding / streaming.
Thirdly, makes node-local multicast streaming feasible.
Finally should noticably increase battery life of mobile devices.
Note: bridge querier is disabled for br-wan. We want to avoid becoming
too "bossy"/"noisy" on a foreign network.
Signed-off-by: Linus Lüssing <linus.luessing@c0d3.blue>
The mesh side has become fairly huge in many communities. Up to
a few thousand entries can currently be found in the forwarding
database (fdb) of a bridge for its bridge port bat0.
The bridge fdb is kind of redundant to the batman-adv global translation
table here. Therefore this patch tries to reduce memory footprint by
following an approach similar to the IGMP/MLD split patchset approach:
Make the bridge oblivious not only regarding multicast listeners towards
the mesh but with this patch unicast hosts on the mesh, too.
If the destination of an ethernet frame is known by the bridge to be a
local one, then the frame is forwarded to the according port. If it is
unknown, then the frame is forwarded to the wifi AP interface and bat0.
mac80211 and batman-adv then know whether to drop or forward a frame
further through their own book-keeping.
Note that unicast-flood is not disabled for the wifi AP bridge port, nor
is learning disabled on the wifi AP. This is mainly to keep the
configuration in UCI and according setup scripts simple ;). However, not
disalbling unicast-flood on the wifi AP interface might also give a
minor latency improvement for newly joining wifi clients.
Signed-off-by: Linus Lüssing <linus.luessing@c0d3.blue>
The arguments are now provided by gluon-mesh-batman-adv-core, so
gluon-radvd can be used with other mesh protocols.
[Matthias Schiffer: removed PROVIDES dependency]
Some drivers (mt76) don't support arbitrary MAC addresses. Use the
addresses provided by the driver (avoiding the primary address) by default,
but fall back to our has-based scheme when the driver doesn't provide
(enough) addresses.
Most doubles that are delivered via respondd have limited input
precision, but are converted with up to 17 digits of precision. That can
cause ugly blowups like 0.2800000000000001 in the output, which is
avoided by specifying better format strings (like "%.2f" in most cases).
Lua's tables are 1-based, so we must decrement the index by 1 to get the
desired MAC addresses. By not doing this, the second IBSS interface would
get the address with index 8, but only indices 0..7 are available.
Fixes: c73a12e0ea
There are a few devices which have more than one LAN interface (for example
some revision of the TL-WR941ND, which uses a DSA-based switch, so each
switch port has its own netdev.) On these devices we need a bridge for
mesh-on-lan (as the alternative of adding them to batman-adv individually
would need too many MAC addresses.)
While ath9k/ath10k devices can supprt VIFs with any combination of MAC addresses, there are also adapters which have a hardware MAC filter which only allows a few bits to differ. This commit changes the addresses of all VIFs to ony differ in the last 3 bits, which is required to support many Ralink/Mediatek based WLAN adapters.
Technically, the new addresses are generated by calculating an MD5 hash of the primary MAC address and using a part of this hash as a prefix for the MAC addresses.
The addresses (BSSIDs) of the AP VIFs are also reused for the LAN and WAN interfaces in mesh-on-LAN/WAN mode to reduce the number of needed addresses, and thus reduce the chance of collisions. This is not a problem as the MAC addresses of the AP VIFs are never used except as BSSID, and thus not seen by routing protocols like batman-adv.
Fixes#648
[Matthias Schiffer: rewrote commit message]
We are mostly dealing with the 2.4GHz crap-band here, so increasing the
IGMP/MLD robustness parameter to three to be able to compensate for up
to two consecutive instead of just one lost packet.
Signed-off-by: Linus Lüssing <linus.luessing@c0d3.blue>
There are some devices not acting properly to roaming events, in that
they do not timely reissue IGMP/MLD reports after reconnecting.
To compensate for that this commit reduces the query interval from 125
seconds to 20 and the query response interval from 20 seconds to 5.
This reduces a timeout to 20+5 seconds in the worst-case (12.5s average)
after a roaming event for such broken devices. This should be below the
30s "impatient user threshold" and below any connection timeout.
Until the bridge multicast snooping + querier gets re-enabled this is a
no-op.
Signed-off-by: Linus Lüssing <linus.luessing@c0d3.blue>
