fix: rpc bind failure fatal (#14330)

Report a proper error when we're unable to initialize the RPC server

Signed-off-by: Rudi Grinberg <me@rgrinberg.com>

Author: rgrinberg

src/dune_rpc_impl/dune

@@ -9,6 +9,7 @@
   source
   memo
   dune_util
+  dune_trace
   fiber
   stdune
   unix)

src/dune_rpc_impl/server.ml

@@ -36,7 +36,7 @@ module Run = struct
     ; root : string
     ; where : Dune_rpc.Where.t
     ; server : Csexp_rpc.Server.t Lazy.t
-    ; server_ivar : Csexp_rpc.Server.t Fiber.Ivar.t
+    ; startup_ivar : (Csexp_rpc.Server.t, Exn_with_backtrace.t) result Fiber.Ivar.t
     ; registry : [ `Add | `Skip ]
     }
 
@@ -61,8 +61,18 @@ module Run = struct
     in
     let run () =
       let open Fiber.O in
-      let server = Lazy.force t.server in
-      let* () = Fiber.Ivar.fill t.server_ivar server in
+      let* server =
+        match Exn_with_backtrace.try_with (fun () -> Lazy.force t.server) with
+        | Ok server ->
+          let+ () = Fiber.Ivar.fill t.startup_ivar (Ok server) in
+          server
+        | Error exn ->
+          let () =
+            Dune_trace.emit Rpc (fun () -> Dune_trace.Event.Rpc.startup_failure exn)
+          in
+          let* () = Fiber.Ivar.fill t.startup_ivar (Error exn) in
+          Exn_with_backtrace.reraise exn
+      in
       Fiber.fork_and_join_unit
         (fun () ->
            let* sessions = Csexp_rpc.Server.serve server in
@@ -196,15 +206,18 @@ type 'build_arg t =
   }
 
 let ready (t : _ t) =
-  let* server = Fiber.Ivar.read t.config.server_ivar in
-  Csexp_rpc.Server.ready server
+  Fiber.Ivar.read t.config.startup_ivar
+  >>= function
+  | Ok server -> Csexp_rpc.Server.ready server
+  | Error _exn -> raise Dune_util.Report_error.Already_reported
 ;;
 
 let stop (t : _ t) =
-  let* server = Fiber.Ivar.peek t.config.server_ivar in
-  match server with
+  Fiber.Ivar.peek t.config.startup_ivar
+  >>= function
   | None -> Fiber.return ()
-  | Some server -> Csexp_rpc.Server.stop server
+  | Some (Error _) -> Fiber.return ()
+  | Some (Ok server) -> Csexp_rpc.Server.stop server
 ;;
 
 let get_current_diagnostic_errors () =
@@ -494,7 +507,7 @@ let create ~lock_timeout ~registry ~root =
     ; where
     ; server
     ; registry
-    ; server_ivar = Fiber.Ivar.create ()
+    ; startup_ivar = Fiber.Ivar.create ()
     }
   in
   let res = { config; pending_jobs; clients = Clients.empty } in

src/dune_trace/dune_trace.mli

@@ -188,6 +188,7 @@ module Event : sig
       -> t
 
     val shutdown : id:int -> stage -> t
+    val startup_failure : Exn_with_backtrace.t -> t
     val close : id:int -> t
     val dropped_write_client_disconnect : Exn.t -> t
   end

src/dune_trace/event.ml

@@ -507,6 +507,12 @@ module Rpc = struct
 
   let shutdown ~id stage = server ~id ~name:"shutdown" stage
 
+  let startup_failure exn =
+    let now = Time.now () in
+    let args = [ "error", Arg.dyn (Exn_with_backtrace.to_dyn exn) ] in
+    Event.instant ~args ~name:"startup-failure" now Rpc
+  ;;
+
   let close ~id =
     let now = Time.now () in
     let args = [ "id", Arg.int id ] in

test/blackbox-tests/setup-script.sh

@@ -186,6 +186,8 @@ summarize_rpc_trace () {
              and .args.stage == "stop"
              and (.args | has("error"))
         then "accept stop error"
+        elif .name == "startup-failure"
+        then "startup failure"
         elif .name == "request" and .args.meth == "build"
         then "build \(.args.stage)"
         elif .name == "shutdown"

test/blackbox-tests/test-cases/watching/dune

@@ -18,6 +18,11 @@
   (= "macosx" %{ocaml-config:system}))
  (applies_to path-pwd))
 
+(cram
+ (applies_to rpc-bind-failure)
+ (enabled_if
+  (<> %{ocaml-config:system} win)))
+
 ;; Disabled due to timeouts
 
 (cram

test/blackbox-tests/test-cases/watching/rpc-bind-failure.t

@@ -0,0 +1,43 @@
+Startup RPC bind failures should be reported immediately and terminate dune.
+
+  $ export DUNE_TRACE=rpc
+
+  $ echo "(lang dune 3.23)" > dune-project
+
+  $ cat > dune <<EOF
+  > (rule
+  >  (target x)
+  >  (action (write-file %{target} ok)))
+  > EOF
+
+Poison the parent of the Unix-domain RPC socket so that binding
+_build/.rpc/dune fails during startup.
+
+  $ mkdir -p _build
+  $ : > _build/.rpc
+
+  $ OUTPUT=$(mktemp)
+  $ ((dune build --passive-watch-mode >"$OUTPUT" 2>&1) || (echo exit $? >>"$OUTPUT")) &
+  $ DUNE_PID=$!
+
+  $ wait_for_dune_exit_with_timeout
+
+  $ grep "Uncaught RPC Error" "$OUTPUT"
+  Uncaught RPC Error
+
+  $ grep "bind" "$OUTPUT"
+  Unix.Unix_error(Unix.ENOTDIR, "bind", "")
+  Raised by primitive operation at Rpc__Csexp_rpc.Socket.U.bind in file
+  Called from Rpc__Csexp_rpc.Socket.bind in file "src/rpc/csexp_rpc.ml"
+  Error: bind(): Not a directory
+
+  $ grep '^exit 1$' "$OUTPUT"
+  exit 1
+
+  $ dune trace cat | jq -c '
+  >    select(.cat == "rpc" and .name == "startup-failure")
+  > | { name, error: .args.error.exn }
+  > '
+  {"name":"startup-failure","error":"Unix.Unix_error(Unix.ENOTDIR, \"bind\", \"\")"}
+
+  $ ! with_timeout_quiet dune rpc ping >/dev/null 2>&1