We need to use the worker key from the lookback tipset state. We can't
reasonably maintain this cache, unfortunately (unless we kept some form of
"per-height" cache).
1. Allow duplicate blocks from bitswap. This shouldn't happen, but there's no
reason to bail (just log loudly).
2. Simplify logic to very explicitly check to make sure we're fetching every
block.
1. Explicitly use a session when fetching messages for a block. Technically,
GetBlocks would use an internal session so we'd only end up with two, but one
per block is even better.
2. Actually cancel sessions after a timeout (threading through the context).
NOTE: We should seriously consider having a single session for all blocks, but
we can do that in a followup as it may have unintended consequences (e.g., leaks).
This patch starts adding support for network upgrades.
* It adds an actors abstraction layer for loading abstract (cross-version) actors.
* It starts switching over to a shared deadline type.
* It adds an abstraction for ADTs (hamt/amt).
* It removes the callback-based API in the StateManager (difficult to abstract
across actor versions).
* It _does not_ actually add support for actors v2. We can do that in a followup
patch but that should be relatively easy.
This patch is heavily WIP and does not compile. Feel free to push changes
directly to this branch.
Notes:
* State tree access now needs a network version, because the HAMT type will change.
* I haven't figured out a nice way to abstract over changes to the _message_
types. However, many of them will be type aliased to actors v0 in actors v2 so
we can likely continue using the v0 versions (or use the v2 versions
everywhere). I've been renaming imports to `v0*` to make it clear that we're
importing types from a _specific_ actors version.
TODO:
* Consider merging incremental improvements? We'd have to get this compiling
again first but we could merge in the new abstractions, and slowly switch over.
* Finish migrating to the new abstractions.
* Remove all actor state types from the public API. See `miner.State.Info()` for
the planned approach here.
* Fix the tests. This is likely going to be a massive pain.
We were ignoring quite a few error cases, and had one case where we weren't
actually updating state where we wanted to. Unfortunately, if the linter doesn't
pass, nobody has any reason to actually check lint failures in CI.
There are three remaining XXXs marked in the code for lint.
