Concurrent updates to qp tries

Knot DNS uses qp tries and a couple of years ago I added support for concurrent updates. It supports lock-free multiple reader / single writer updates in RCU (read/copy/update) style.

It is designed to fit the way authoritative DNS servers work, where updates to a zone happen one at a time, not too frequently, and should not affect queries.

The cache in recursive servers is more difficult. When there is a cache miss, the server needs to resolve the query, and there can be a lot of resolver jobs in progress that may need to add multiple names to the trie (e.g. for zones and nameservers they discovered while resolving the query). And the server needs to clean expired records from the cache.

It is a much more concurrent setting than an authoritative server.

RCU for qp-trie caches

It would be nice if resolver jobs could update the cache without having to pass all updates to the single writer that is able to update the qp trie, and without having to wait for a full RCU epoch.

It might be worth using a second data structure that's optimized for concurrent writes, rather than for storing a lot of data with high read speed. (A concurrent hash map of some kind, perhaps.)

The idea is that when there is a cache miss on the main qp trie, check the hash map, and if that lookup also fails, start a resolver job. Resolver jobs only update the hashmap.

Periodically a cache cleaning job runs, which folds the contents of the hashmap into the qp trie, and deletes expired records or records that need to be purged to remain within memory limits. This cleaning job uses the copy-on-write concurrent qp trie update code.

One question I'm not sure about is how much the leaf objects - the DNS records structures that are application data from the qp trie point of view - might need to be updated after they are created, and whether it matters if they are updated after being moved from the hashmap to the main trie. For example,

The qp trie is about name lookups, but in the DNS a a name can have multiple RRsets with different TTLs, so the per-name cache structure is relatively complicated and needs to support concurrent updates. Since that is the case, it is probably OK if the per-name record is moved from the hashmap to the qp trie by the cleaner, even if resolver jobs are in progress.

Written by Tony Finch; You may do anything with this. It has no warranty.