The restart timer was racy and didn't allow for immediate restarts
from limits. Now the protocols stores the last restart time and in case
of too frequent autorestarts caused by exceeded limits, the protocol
gets disabled with an error message.
Also now there is a configuration knob for this.
There were some nasty problems with deferred protocol state updates and
race conditions on BGP startup, shutdown, and also with referencing the
cached states.
Now it looks fixed.
To allow reading of protocol states from other protocols or completely
different routines, we have to export these states to data structures
not requiring to lock the protocol loops.
On one hand, this doesn't give the reader the actual state "right now",
on the other hand, getting that is impossible in a properly
multithreaded environment and you will always get the information with
some (little but noteworthy) delay.
This implementation handles only the basic state information of the
protocols, common for all the protocols. Adding protocol-specific state
information should be done by implementing the protocol hook init_state().
Channel information is stored but not announced, as we don't need the
announcements for now.
For the upcoming rework of protocol state information propagation,
we need some more eattr types to be defined.
These types are probably not defined completely and before using
them for route attributes, you should check that they don't lack
some crucial methods.
Before this commit, on kernel shutdown, the routes were re-exported by
the regular export but treated as withdraw. This was too hairy and
caused unnecessary complexity of the protocol's state machine.
Instead of that, we found out that it makes more sense to just refeed
the routes synchronously and convert to withdraw. This is done by the
direct export access instead of the channel.
It would (maybe) make more sense to run export filters on this in case
the export filter updates the krt_metric attribute, but as this doesn't
work on regular withdraw anyway, it's better for now to just let it be
and maybe somebody in the future fixes this issue.
When VRFs are used, BIRD correctly binds listening (and connecting)
sockets to their VRFs but also re-binds accepted sockets to the same VRF.
This is not needed as the interface bind is inherited in this case, and
indeed this redundant bind causes an -EPERM if BIRD is running as
non-root making BIRD close the connection and reject the peer.
Thanks to Christian Svensson for the original patch and Alexander Zubkov
for suggestions.
This allows to have one main socket for the heavy operations
very restricted just for the appropriate users, whereas the
looking glass socket may be more open.
Implemented an idea originally submitted and requested by Akamai.
