Related issues

• caml_process_pending_actions_exn is missing ocaml-multicore/ocaml-multicore#791
• Unify GC interrupt and signals triggering mechanisms ocaml-multicore/ocaml-multicore#806
• This series of review comments: Multicore OCaml #10831 (review)

Action points

First in any order:

• Review the diff of finaliser.c between multicore and old trunk to make sure nothing is missing; reimplement caml_final_do_calls_exn.
• Separate signalling from payload for inter-domain interrupts. Record the request for an interrupt separately from young_limit to distinguish the various asynchronous actions to perform when an allocation fails.
• Discuss atomicity of young_limit (Multicore OCaml #10831 (comment)).

Then (in order):

• caml_alloc_small_dispatch is a common entry point to the GC (restore its subtle logic)
• Reimplement caml_do_pending_actions_exn/caml_process_pending_actions_exn
• Minimise the diff with old trunk (check that nothing misses; preserve relevant bits of memprof implementation to avoid to forget things later on); make sure that there is no unjustified change to caml/signals.h and caml/alloc.h.

I think those are a pre-requisite for 5.0 (public API in actual use and the documented way to poll in long-running C functions). If not possible, at the very least, caml_process_pending_actions_exn should be reintroduced with a partial implementation.

Checklist for expected behaviours

The following are needed for 5.0:

• caml_check_urgent_gc runs pending GC actions including inter-domain interrupts; no asynchronous callbacks. Idem when handling an inter-domain interrupt.
• A function update_young_limits is used to locally set young_limit to the right value according to various pending action flags (signals, finalisers, inter-domain interrupts...), to make sure young_limit is always set to the right value.
• caml_process_pending_actions_exn does not raise directly into OCaml.
• caml_do_pending_actions_exn informs of possible future actions if interrupted by an exception.
• finalisers do not run immediately when allocating from C; restore the guarantee at the top of caml/alloc.h.
• Reintroduce caml_something_to_do and its logic?
• Events are properly stopped in case of an asynchronous exception inside runtime/gc_ctrl.c.
• caml_leave_blocking_section checks for signals like it used to.
• Gc.alarms work as intended.

The following are not strictly needed for 5.0 (but most are likely fixed along with the rest).

• OCaml signal handlers are guaranteed to be run as soon as possible.
• No need for caml_handle_gc_interrupts_no_async.
• caml_process_pending_actions* is the single entry point to run all pending actions (no separate caml_process_pending_signals*).
• When recording a backtrace, no need to assume that finalisers could run.
• The VM polls by comparing young_limit to young_ptr; investigate Use a bit-vector to record pending signals #10880 (comment).

(cc @sadiqj, @jhjourdan)