Garbage collect resolved promises

[FUDCo]

This issue summarizes subtasks related to the garbage collection of promises that resolve to data containing references to other promises. It is a tracking issue derived from the discussion in #1049 that produced a plan for doing this. There are various tickets related to retiring resolved promises that were collectively tracked by #675, but that although that issue both contains and points to a lot of relevant background, as a planning document it has drifted out of relevance as our understanding of the underlying problem has evolved and as various elements of the solution have been implemented and merged into the main trunk of development.

The current breakdown of subtasks is:

• Drop references that have no embedded promises. This work is largely complete. This was addressed in PRs Continued implementation of resolved promise pruning #930, Prune resolved promises (recipient side) #799, fix: Clear vat table entries for promises when they are resolved #737, which have landed, and by PR Implement kernel promise reference counting #1121 which as of this writing is awaiting review.
• Retire promises that reference only "new" other promises #1125 Drop references that have no already exported promises. (See discussion in Discussion of garbage collection of resolved kernel promises #1049 (comment)
• Periodic full garbage collection of resolved promises #1126 Periodic mark-and-sweep garbage collection of other cases.

Note that both the open issues above have a kernel portion and a liveslots portion, which, while they'll likely be implemented separately, will need to be coordinated so that there is no inconsistency between the kernel's and the vats' views of the world.

[warner]

@FUDCo 's current efforts are towards retiring the c-list entries for all resolved promises, performing the resolution in a batch so that any inter-promise references are available before their identifiers disappear. We have three phases of this work:

• kernel-to-vat (PR Finish kernel-side retirement of C-list entries for resolved promises #2110): the kernel needs to notify some vat about one promise being resolved, and actually generates a dispatch.notify that contains every already-resolved promise in the graph reachable from that one
• vat-to-kernel: a vat needs to notify the kernel about one promise being resolved, and generates a syscall.resolve with every already-resolved promise reachable from that one. This will require liveslots to track previously-resolved promises in a WeakMap so it can tell (synchronously) whether a referenced promise is resolved or not.
• comms: the comms vat needs to do the same, both when interacting with the kernel, and when interacting with a remote machine

Retiring c-list entries is the first step. Once all c-list references to a kernel promise table entry are gone, we can check to see if anything else within the kernel is still referencing the promise (run-queue entries, resolution data for other promises, messages within other promises, auxiliary kernel object data). If all other references are gone, we can deallocate the orphaned promise table entry (and save the kernel DB space it consumed).

There may be collectable cycles in the graph, which will require more sophisticated tools to collect safely. The simplest case is when some number of resolved promises reference each other, but nothing else references them.

[warner]

The remaining portion of this ticket is covered by #1126, so I'm closing this one.