Ease Butterfly upgrade

[jerome-jutteau]

Currently, the way to upgrade a Butterfly is to:

• stop old Butterfly
• start new Butterfly
• push all configuration to the new Butterfly
• connect on qemu monitors and recreate each vnic

Here are some possible scenarios to ease upgrade:

1. Easy scenario

• query a configuration dump from the old Butterfly
• stop old Butterfly
• start new Butterfly
• load configuration dump
• connect on qemu monitors and recreate each vnic

Pros:

• Simple procedure => less bugs
• In case of crash of the new Butterfly, the old one can be re-launched using configuration dump

Cons:

• Network downtime during migration

2. Easy scenario with qemu as a server

• query a configuration dump from the old Butterfly
• stop old Butterfly
• start new Butterfly
• load configuration dump
• Butterfly reconnects to existing vnics managed by Qemus

Pros:

• simple procedure
• no need to connect on qemu console and recreate vnics
• in case of crash of the new Butterfly, the old one can be re-launched using configuration dump

Cons:

• network downtime during migration
• should invert vhost-user client and server (qemu become the vhost server)

3. Stream configuration

• launch the new Butterfly and provide him location of the old Butterfly
• new Butterfly use an other available DPDK port
• new Butterfly query old one to switch it in Read Only mode for 1 minute (no configuration change can now be made)
• new Butterfly query the old one, get all it's security groups and apply them
• for each vnic, the new Butterfly will :
  • get vnic configuration from the old one
  • ask the old Butterfly to disconnect from this particular vhost-user
  • create nic and connect
  • Send a Read Only mode query to the old Butterfly (reset counter)
• In case of success for all vnics:
  • shutdown old butterfly
  • new Butterfly listen to API port
• In case of failure of one of the vnics (or the new Butterfly crashed):
  • Migration fails, new butterfly shutdown by itself
  • On the old Butterfly, the Read Only query timeouts
  • All disabled nics from the old Butterfly reconnects to vhost-user, DPDK port is re-enabled again

Pros:

• vnic network downtime is minimal IF
• no operation to make on qemu nics
• fallback in case of error/crash of the new Butterfly

Cons:

• complex and more risky. Must manage all possible case of error
• heavy modifications
• should invert vhost-user client and server (qemu become the vhost server)
• must run 2 Butterfly at the same time: we may have to fix some problems related to hugepage usage and DPDK lock.

[jerome-jutteau]

We will go for solution 2. Check #275 and #276