(************************************************************************) (* * The Rocq Prover / The Rocq Development Team *) (* v * Copyright INRIA, CNRS and contributors *) (* string array (* run by the master, on a thread *) val request_of_task : worker_status -> task -> request option val task_match : worker_status -> task -> bool val use_response : worker_status -> task -> response -> [ `Stay of competence * task list | `End ] val on_marshal_error : string -> task -> unit val on_task_cancellation_or_expiration_or_slave_death : task option -> unit val forward_feedback : Feedback.feedback -> unit (* run by the worker *) val perform : request -> response (* debugging *) val name_of_task : task -> string val name_of_request : request -> string end module Make(T : Task) () = struct exception Die type response = | Response of T.response * (NewProfile.MiniJson.t list * NewProfile.sums) option | RespFeedback of Feedback.feedback type request = { request : T.request; profiling : bool } let slave_respond { request = r; profiling } = if profiling then let events, sums, res = NewProfile.with_profiling (fun () -> T.perform r) in Response (res, Some (events, sums)) else let res = T.perform r in Response (res, None) exception MarshalError of string let marshal_to_channel oc data = Marshal.to_channel oc data []; flush oc let marshal_err s = raise (MarshalError s) let marshal_request oc (req : request) = try marshal_to_channel oc req with Failure s | Invalid_argument s | Sys_error s -> marshal_err ("marshal_request: "^s) (* [unmarshal_request] and [marshal_response] are used by the worker only, where a [Sys_error] means the channel to the master is gone. We let it through so that the main loop can tell a lost connection apart from a genuine marshalling failure, which alone becomes [MarshalError]. *) let unmarshal_request ic = try (CThread.thread_friendly_input_value ic : request) with Failure s | Invalid_argument s -> marshal_err ("unmarshal_request: "^s) let marshal_response oc (res : response) = try marshal_to_channel oc res with Failure s | Invalid_argument s -> marshal_err ("marshal_response: "^s) let unmarshal_response ic = try (CThread.thread_friendly_input_value ic : response) with Failure s | Invalid_argument s | Sys_error s -> marshal_err ("unmarshal_response: "^s) let report_status ?(id = !Flags.async_proofs_worker_id) s = let open Feedback in feedback ~id:Stateid.initial (WorkerStatus(id, s)) module Worker = Spawn.Sync () module Model = struct type process = Worker.process type extra = (T.task * cancel_switch) TQueue.t let uid = ref 0 let get_toplevel_path top = let dir = Findlib.package_directory "rocq-runtime" in let exe = if Sys.(os_type = "Win32" || os_type = "Cygwin") then ".exe" else "" in Filename.concat dir (top^exe) let spawn ~spawn_args id priority = let name = Printf.sprintf "%s:%d:%d" T.name id !uid in incr uid; let proc, ic, oc = (* Filter arguments for slaves. *) let args = let wselect = "--kind=" ^ T.name in let worker_opts = !async_proofs_flags_for_workers @ ["-worker-id"; name; "-async-proofs-worker-priority"; CoqworkmgrApi.(string_of_priority priority)] in Array.of_list (wselect :: spawn_args @ worker_opts) in let env = Array.append (T.extra_env ()) (Unix.environment ()) in let worker_name = get_toplevel_path "rocqworker" in Worker.spawn ~env worker_name args in name, proc, CThread.prepare_in_channel_for_thread_friendly_io ic, oc let manager cpanel (id, proc, ic, oc) = let { WorkerPool.extra = queue; exit; cancelled } = cpanel in let exit () = report_status ~id "Dead"; exit () in let last_task = ref None in let worker_age = ref T.Fresh in let got_token = ref false in let giveback_exec_token () = if !got_token then (CoqworkmgrApi.giveback 1; got_token := false) in let stop_waiting = ref false in let expiration_date = ref (ref false) in let pick_task () = stm_prerr_endline "waiting for a task"; let pick age (t, c) = not !c && T.task_match age t in let task, task_expiration = TQueue.pop ~picky:(pick !worker_age) ~destroy:stop_waiting queue in expiration_date := task_expiration; last_task := Some task; stm_prerr_endline ("got task: " ^ T.name_of_task task); task in let add_tasks l = List.iter (fun t -> TQueue.push queue (t,!expiration_date)) l in let get_exec_token () = ignore(CoqworkmgrApi.get 1); got_token := true; stm_prerr_endline ("got execution token") in let kill proc = Worker.kill proc; stm_prerr_endline ("Worker exited: " ^ match Worker.wait proc with | Unix.WEXITED 0x400 -> "exit code unavailable" | Unix.WEXITED i -> Printf.sprintf "exit(%d)" i | Unix.WSIGNALED sno -> Printf.sprintf "signalled(%d)" sno | Unix.WSTOPPED sno -> Printf.sprintf "stopped(%d)" sno) in let rec kill_if () = if not (Worker.is_alive proc) then () else if cancelled () || !(!expiration_date) then let () = stop_waiting := true in let () = TQueue.broadcast queue in Worker.kill proc else let () = Unix.sleep 1 in kill_if () in let kill_if () = try kill_if () with Sys.Break -> let () = stop_waiting := true in let () = TQueue.broadcast queue in Worker.kill proc in let _ = CThread.create kill_if () in try while true do report_status ~id "Idle"; let task = pick_task () in match T.request_of_task !worker_age task with | None -> stm_prerr_endline ("Task expired: " ^ T.name_of_task task) | Some req -> try get_exec_token (); marshal_request oc {request = req; profiling = NewProfile.is_profiling()}; let rec continue () = match unmarshal_response ic with | RespFeedback fbk -> T.forward_feedback fbk; continue () | Response (resp, prof) -> Option.iter (fun (events,sums) -> NewProfile.insert_results events sums) prof; match T.use_response !worker_age task resp with | `End -> raise Die | `Stay(competence, new_tasks) -> last_task := None; giveback_exec_token (); worker_age := T.Old competence; add_tasks new_tasks in continue () with | (Sys_error _|Invalid_argument _|End_of_file|Die) as e -> raise e (* we pass the exception to the external handler *) | MarshalError s -> T.on_marshal_error s task; raise Die | e -> stm_pr_err Pp.(seq [str "Uncaught exception in worker manager: "; print e]); flush_all (); raise Die done with | (Die | TQueue.BeingDestroyed) -> giveback_exec_token (); kill proc; exit () | Sys_error _ | Invalid_argument _ | End_of_file -> T.on_task_cancellation_or_expiration_or_slave_death !last_task; giveback_exec_token (); kill proc; exit () end module Pool = WorkerPool.Make(Model) type queue = { active : Pool.pool; queue : (T.task * cancel_switch) TQueue.t; cleaner : Thread.t option; } [@@warning "-unused-field"] (* cleaner unused, not sure if can be removed *) let create ~spawn_args size priority = let cleaner queue = while true do try ignore(TQueue.pop ~picky:(fun (_,cancelled) -> !cancelled) queue) with TQueue.BeingDestroyed -> (Thread.exit [@warning "-3"]) () done in let queue = TQueue.create () in { active = Pool.create ~spawn_args queue ~size priority; queue; cleaner = if size > 0 then Some (CThread.create cleaner queue) else None; } let destroy { active; queue } = Pool.destroy active; TQueue.destroy queue let broadcast { queue } = TQueue.broadcast queue let enqueue_task { queue; active } t ~cancel_switch = stm_prerr_endline ("Enqueue task "^T.name_of_task t); TQueue.push queue (t, cancel_switch) let cancel_worker { active } n = Pool.cancel n active let name_of_request {request = r} = T.name_of_request r let set_order { queue } cmp = TQueue.set_order queue (fun (t1,_) (t2,_) -> cmp t1 t2) let join { queue; active } = if not (Pool.is_empty active) then TQueue.wait_until_n_are_waiting_and_queue_empty (Pool.n_workers active + 1(*cleaner*)) queue let cancel_all { queue; active } = TQueue.clear queue; Pool.cancel_all active let slave_ic = ref None let slave_oc = ref None let init_stdout () = let ic, oc = Spawned.get_channels () in slave_oc := Some oc; slave_ic := Some ic let slave_handshake () = Pool.worker_handshake (Option.get !slave_ic) (Option.get !slave_oc) let pp_pid pp = Pp.(str (Spawned.process_id () ^ " ") ++ pp) let debug_with_pid = Feedback.(function | { contents = Message(Debug, loc, qf, pp) } as fb -> { fb with contents = Message(Debug,loc, qf, pp_pid pp) } | x -> x) let main_loop () = (* We pass feedback to master *) let slave_feeder oc fb = Control.protect_sigalrm (fun () -> Marshal.to_channel oc (RespFeedback (debug_with_pid fb)) []; flush oc) () in ignore (Feedback.add_feeder (fun x -> slave_feeder (Option.get !slave_oc) x)); let working = ref false in slave_handshake (); while true do try working := false; let request = unmarshal_request (Option.get !slave_ic) in working := true; report_status (name_of_request request); let response = slave_respond request in report_status "Idle"; marshal_response (Option.get !slave_oc) response; CEphemeron.clean () with | MarshalError s -> stm_pr_err Pp.(prlist str ["Fatal marshal error: "; s]); flush_all (); exit 2 (* The master going away shows up either as [End_of_file] or as a [Sys_error] (a connection reset on Windows, a broken pipe elsewhere). The worker is being torn down, so neither is worth reporting. *) | End_of_file | Sys_error _ -> stm_prerr_endline "connection lost"; flush_all (); exit 2 | e -> stm_pr_err Pp.(seq [str "Slave: critical exception: "; print e]); flush_all (); exit 1 done let clear { queue; active } = assert(Pool.is_empty active); (* We allow that only if no slaves *) TQueue.clear queue let snapshot { queue; active } = List.map fst (TQueue.wait_until_n_are_waiting_then_snapshot (Pool.n_workers active) queue) let with_n_workers ~spawn_args n priority f = let q = create ~spawn_args n priority in try let rc = f q in destroy q; rc with e -> let e = Exninfo.capture e in destroy q; Exninfo.iraise e let n_workers { active } = Pool.n_workers active end module MakeQueue(T : Task) () = struct include Make(T) () end module MakeWorker(T : Task) () = struct include Make(T) () end exception RemoteException of Pp.t let _ = CErrors.register_handler (function | RemoteException ppcmd -> Some ppcmd | _ -> None)