The handling of BVI is awkward, with the autoderived interface name
"bviXX" based on the bridgedomain bd_id. Lots of special casing happens
on account of this decision, and to make matters worse there is poor
interaction (leading to VPP crashes) when BVIs and Loopbacks are used
at the same time: https://lists.fd.io/g/vpp-dev/message/21116
In this commit, I reintroduce the ability to set bridgedomain virtual
interfaces by means of the 'bvi' keyword. The 'bvi' must:
- be a Loopback interface
- must be used at most once (bvi_unique())
When pruning, I now need to prune bridgedomains before pruning
loopbacks, because any given loopback might be a BVI for a bridge. So,
I'll remove the loop/BVI from the bridge (by setting it to L3) and only
then removing the loopback from VPP.
In the reconciler, remove BVIs that have changed in prune_bridgedomains()
and set it in sync_bridgedomains().
The handling of BVI is awkward, with the autoderived interface name
"bviXX" based on the bridgedomain bd_id. Lots of special casing happens
on account of this decision, and to make matters worse there is poor
interaction (leading to VPP crashes) when BVIs and Loopbacks are used
at the same time: https://lists.fd.io/g/vpp-dev/message/21116
This is step one of a refactor of the logic. In this commit, I'm
removing all of the BVI logic from the codebase, rendering bridgedomains
unable to have IP interfaces. In the next commit, I will introduce new
behavior in the schema, allowing for 'bvi' to be a loopback
interfacename which will be used as BVI for a bridgedomain, restoring
the ability to use bridgedomains with IP interfaces (using a loop).
What this means is that the synchronization will run based on loops over
the configuration, rather than over the VPP state. The benefit of this
approach is that no additional API calls have to be made after the
initial VPP configuration is read, making pre-flight checks and diffing
more obvious.
Further, make sync_link_mtu() direction-aware; first we will raise the
link_mtu for interfaces that are growing, and after setting all the
children up/down with sync_mtu_direction() we'll finally shrink
sync_link_mtu() where needed.
We can now do an entire reconciliation run with only one API read at the
beginning!
Include special caveat on LCP MAC changes, for which I'll put in a
TODO for now with a VPP comment {} with the to be run command.
Also make the user aware of a quick in BondEthernets not being able
to have link_mtu != 9000 so if a packet MTU > 9000 is set, this will
work but is an undesirable configuration. Issue a warning in this
case.
This is rather straight forward: for each object (in correct order),
if the object exists in VPP, we can skip it. We know that it will exist
only if it was valid (ie correct encapsulation, tc). If it does not
exist in VPP, issue the correct creation request to VPP.
Implement the creation of all types, in the following order:
- create_loopbacks() and create_bvis()
- create_bondethernets()
- create_vxlan_tunnels()
- create_sub_interfaces() first 1-tag, then 2-tag
- create_bridgedomains()
- create_lcps() bottomsup: first 0-tag, then 1-tag, then 2-tag names.
Add a placeholder TODO to fix a bug with prune_sub_interfaces() which
should allow for TAPs belonging to LCPs; will fix in followup commit.
After pruning elements from the VPP configuration, also remove them
from the configuration. The purpose of this is to leave a reasonable
representation of the VPP state in the configuration, so that subsequent
creates and syncs do not have to query the VPP API repeatedly.
The goal of this change is to be able to plan a complete path from
prune, create and sync, with only reading the initial VPP configuration
once, not multiple times.
- Add a get_sub_interfaces() call to fetch all DPDK / Bond sub-ints
- In prune_bvis(), prune_loopbacks() and prune_sub_interfaces(),
use sub_number_of_tags to go out-to-in (qinx, dot1x, untagged)
- Prune bridgedomains before pruning BVIs. The reason for this is that
prune_bridgedomains() will set the BVI to L3 mode, and if the BVI is
removed before the bridge is pruned, this is an error.
- When pruning bridge members, use the VPP configuration as the member
may not exist in the config, upon which the call to interface.is_sub()
will return False even if it is actually a VPP sub-int.
Update README.md, also take into account the previous change which calls
prune_addresses() before object deletion.
1. sub-ints have to be removed before their parent, particularly
bondethernets, because destroying BondEthernet0 will cascade
destruction of all of its sub-ints. So:
- first prune_sub_interfaces()
- then prune_bondethernets()
- finally prune_phys()
2. for any interface (loop, bvi, sub, phy, bond), prune_addresses()
before destroying the interface
Update the README with this new flow.
