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.
closes#86
The uncork events are running from mainloop so these should just
dispatch the right event to the right loop. Doing anything long there
is bad for performance and latency as the uncork list may be huge.
the sync is actually needed when the pages get freed, not precisely
after every item cleanup, as the data technically stays intact until the
deferred free's are called.
This partially reverts commit d617801c31018926cccd75a64186ebb0597a6bcc.
The common lockfree doesn't work well for high-volume structures like
eattr cache because it expects the structure to be cleaned up by a
sweeper routine ... which is very ineffective for >1M records.
OTOH, we need the deferred ea_free in all cases ... so keeping that.
There is an IP table for every ROA table, holding special records
combining all known ROAs for every top-prefix.
The ROA digestor is now an IP digestor, running over the auxiliary
table.
The sending channel may be already gone when rte_free_deferred()
is finally called so we have to log about route freeing in the
synchronous call instead.
Channel is now just subscribing to yet another journal announcing
digested tries from the ROA table.
Creating tries in every channel on-the-fly was too slow to handle
and it ate obnoxious amounts of memory. Instead, the tries are
constructed directly in the table and the channels are notified
with the completed tries.
The delayed export-release mechanism is used to keep the tries allocated
until routes get reloaded.
Originally, this mechanism required to check whether there's enough time to work
and then to send an event. This macro combines all the logic and goes more straightforwardly
to the _end_ of the export processing loop.
One should note that there were two cases where the export processing loop
was deferred at the _beginning_, which led to ignoring some routes on
reimports. This wasn't easily noticeable in the tests until the one-task
limit got a ceiling on 300 ms to keep reasonable latency.
