rktio: better behavior for opening write and of fifo

When opening the write end of a fifo that doesn't have a reader
already, the old implementation could allow writing bytes that are
discarded. This new implementation uses a blocking `open` in a
`pthread`, and that way the write routines know whether the stream is
ready for writing or not.

The difference is visible in the Racket API in a two places:
`subprocess` needs to wait until a fifo writer is connected before
attempting to dup the corresponding file descriotor, and more
generally a use of `unsafe-port->file-descriptor` needs to wait. The
former blocking operation is now build into `subprocess` (and
documented), but the burden is place on callers of
`unsafe-port->file-descriptor` to wait is necessary.

The new `port-waiting-peer?` predicate exposes the waiting state,
while `sync` is sufficient to wait for a peer.

Closes #2784

pkgs/base/info.rkt

@@ -12,7 +12,7 @@
 
 (define collection 'multi)
 
-(define version "7.4.0.4")
+(define version "7.4.0.5")
 
 (define deps `("racket-lib"
                ["racket" #:version ,version]))

pkgs/racket-doc/scribblings/foreign/port.scrbl

@@ -63,7 +63,17 @@ previously created with @racket[unsafe-file-descriptor->semaphore] or
 )]{
 
 Returns a file descriptor (which is a @tt{HANDLE} value on Windows) of
-a socket for @racket[port] if it has one, @racket[#f] otherwise.}
+a socket for @racket[port] if it has one, @racket[#f] otherwise.
+
+On Unix and Mac OS, the result of
+@racket[unsafe-port->file-descriptor] can be @racket[#f] if it
+corresponds to a port that is waiting for its peer as reported by
+@racket[port-waiting-peer?], such as the write end of a fifo where no
+reader is connected. Wait until such is ready by using @racket[sync]).
+
+@history[#:changed "7.4.0.5" @elem{Accommodate a fifo write
+                                   end blocked on a reader by
+                                   returning @racket[#f].}]}
 
 
 @deftogether[(

pkgs/racket-doc/scribblings/reference/file-ports.scrbl

@@ -187,7 +187,10 @@ The file specified by @racket[path] need not be a regular file. It
 might be a device that is connected through the filesystem, such as
 @filepath{aux} on Windows or @filepath{/dev/null} on Unix. The output
 port is block-buffered by default, unless the file corresponds to a
-terminal, in which case it is line-buffered by default.
+terminal, in which case it is line-buffered by default. On Unix and
+Mac OS, if the file is a fifo, then the port will block for writing
+until a reader for the fifo is available; see also
+@racket[port-waiting-peer?].
 
 The port produced by @racket[open-output-file] should be explicitly
 closed, either though @racket[close-output-port] or indirectly via
@@ -212,7 +215,10 @@ then @exnraise[exn:fail:filesystem:errno].
 @history[#:changed "6.9.0.6" @elem{On Unix and Mac OS, make @racket['truncate/replace]
                                    replace on a permission error. On Windows, make
                                    @racket['replace] always replace instead truncating
-                                   like @racket['truncate/replace].}]}
+                                   like @racket['truncate/replace].}
+         #:changed "7.4.0.5" @elem{Changed handling of a fifo on Unix and Mac OS to
+                                   make the port block for output until the fifo has a
+                                   reader.}]}
 
 @defproc[(open-input-output-file [path path-string?]
                            [#:mode mode-flag (or/c 'binary 'text) 'binary]

pkgs/racket-doc/scribblings/reference/port-procs.scrbl

@@ -63,6 +63,24 @@ interactive terminal, @racket[#f] otherwise.
                                    to any value, instead of resticting
                                    the domain to ports}]}
 
+
+@defproc[(port-waiting-peer? [port port?]) boolean?]{
+
+Returns @racket[#t] if @racket[port] is not ready for reading or
+writing because it is waiting for a peer process to complete a stream
+construction, @racket[#f] otherwise.
+
+On Unix and Mac OS, opening a fifo for output creates a peer-waiting
+port if no reader for the same fifo is already opened. In that case,
+the output port is not ready for writing until a reader is opened;
+that is, write opertaions will block. Use @racket[sync] if necessary
+to wait until writing will not block---that is, until the read end of
+the fifo is opened.
+
+@history[#:added "7.4.0.5"]}
+
+
+
 @defthing[eof eof-object?]{A value (distinct from all other values)
 that represents an end-of-file.}
 

pkgs/racket-doc/scribblings/reference/subprocess.scrbl

@@ -71,6 +71,10 @@ returned by @racket[subprocess]. The @racket[stderr] argument can be
 that is supplied as standard output is also used for standard error.
 For each port or @racket['stdout] that is provided, no
 pipe is created and the corresponding returned value is @racket[#f].
+If @racket[stdout] or @racket[stderr] is a port for which
+@racket[port-waiting-peer?] returns true, then @racket[subprocess]
+waits for the port to become ready for writing before proceeding with
+the subprocess creation.
 
 If @racket[group] is @racket['new], then the new process is created as
 a new OS-level process group. In that case, @racket[subprocess-kill]
@@ -132,7 +136,9 @@ A subprocess can be used as a @tech{synchronizable event} (see @secref["sync"]).
 A subprocess value is @tech{ready for synchronization} when
 @racket[subprocess-wait] would not block; @resultItself{subprocess value}.
 
-@history[#:changed "6.11.0.1" @elem{Added the @racket[group] argument.}]}
+@history[#:changed "6.11.0.1" @elem{Added the @racket[group] argument.}
+         #:changed "7.4.0.5" @elem{Added waiting for a fifo without a reader
+                                   as @racket[stdout] and/or @racket[stderr].}]}
 
 
 @defproc[(subprocess-wait [subproc subprocess?]) void?]{

pkgs/racket-test-core/tests/racket/file.rktl

@@ -1724,6 +1724,52 @@
 (for ([f '("tmp1" "tmp2" "tmp3")] #:when (file-exists? f)) (delete-file f))
 
 (current-directory original-dir)
+
+(unless (eq? 'windows (system-type))
+  (define fifo (build-path work-dir "ff"))
+  (system* (find-executable-path "mkfifo") fifo)
+
+  (define i1 (open-input-file fifo))
+  (define o1 (open-output-file fifo #:exists 'update))
+  (write-bytes #"abc" o1)
+  (flush-output o1)
+  (test #"abc" read-bytes 3 i1)
+  (close-input-port i1)
+  (close-output-port o1)
+
+  (define (check-output-blocking do-write-abc)
+    ;; Make sure an output fifo blocks until there's a reader
+    (define t1
+      (thread
+       (lambda ()
+         (define o2 (open-output-file fifo #:exists 'update))
+         (test #t port-waiting-peer? o2)
+         (do-write-abc o2)
+         (close-output-port o2))))
+    (define t2
+      (thread
+       (lambda ()
+         (sync (system-idle-evt))
+         (define i2 (open-input-file fifo))
+         (test #"abc" read-bytes 3 i2)
+         (close-input-port i2))))
+    (sync t1)
+    (sync t2))
+
+  (check-output-blocking (lambda (o2) (write-bytes #"abc" o2)))
+  (check-output-blocking (lambda (o2)
+                           (parameterize ([current-output-port o2])
+                             (system* (find-executable-path "echo")
+                                      "-n"
+                                      "abc"))))
+
+  (delete-file fifo))
+
+(test #f port-waiting-peer? (current-input-port))
+(test #f port-waiting-peer? (current-output-port))
+(test #f port-waiting-peer? (open-input-bytes #""))
+(err/rt-test (port-waiting-peer? 10))
+
 (delete-directory work-dir)
 
 ;; Network - - - - - - - - - - - - - - - - - - - - - -

racket/src/cs/primitive/kernel.ss

@@ -622,6 +622,7 @@
   [port-provides-progress-evts? (known-procedure 2)]
   [port-read-handler (known-procedure 6)]
   [port-try-file-lock? (known-procedure 4)]
+  [port-waiting-peer? (known-procedure 2)]
   [port-write-handler (known-procedure 6)]
   [port-writes-atomic? (known-procedure 2)]
   [port-writes-special? (known-procedure 2)]

racket/src/io/host/rktio.rkt

@@ -11,7 +11,7 @@
          rktio-errstep
          racket-error?
          rktio-place-init!)
-;; More `provide`s added by macros below
+;; More `provide`s are added by macros below
 
 (define rktio-table
   (or (primitive-table '#%rktio)

racket/src/io/port/fd-port.rkt

@@ -1,6 +1,7 @@
 #lang racket/base
 (require racket/fixnum
          "../common/class.rkt"
+         "../common/check.rkt"
          "../host/rktio.rkt"
          "../host/error.rkt"
          "../host/thread.rkt"
@@ -26,6 +27,7 @@
          open-output-fd
          finish-fd-output-port
          terminal-port?
+         port-waiting-peer?
          fd-port-fd
          prop:fd-place-message-opener)
 
@@ -348,6 +350,19 @@
      (fd-output-port-fd cp)]
     [else #f]))
 
+(define/who (port-waiting-peer? p)
+  (define cp (->core-output-port p #:default #f))
+  (cond
+    [cp
+     (cond
+       [(fd-output-port? cp)
+        (define fd (fd-port-fd cp))
+        (rktio_fd_is_pending_open rktio fd)]
+       [else #f])]
+    [(input-port? p) #f]
+    [else
+     (raise-argument-error who "port?" p)]))
+
 ;; ----------------------------------------
 
 ;; in atomic mode

racket/src/io/port/main.rkt

@@ -16,7 +16,8 @@
          "file-port.rkt"
          "file-stream.rkt"
          (only-in "fd-port.rkt"
-                  terminal-port?)
+                  terminal-port?
+                  port-waiting-peer?)
          "file-identity.rkt"
          "file-lock.rkt"
          "bytes-port.rkt"
@@ -112,6 +113,7 @@
 
          file-stream-port?
          terminal-port?
+         port-waiting-peer?
 
          open-input-bytes
          open-output-bytes

racket/src/io/subprocess/main.rkt

@@ -116,6 +116,14 @@
         
         (define command-bstr (->host (->path command) who '(execute)))
 
+        ;; If `stdout` or `stderr` is a fifo with no read end open, wait for it:
+        (define (maybe-wait fd)
+          (when (and fd (rktio_fd_is_pending_open rktio (fd-port-fd fd)))
+            (sync fd)))
+        (maybe-wait stdout)
+        (unless (eq? stderr 'stdout)
+          (maybe-wait stderr))
+
         (start-atomic)
         (poll-subprocess-finalizations)
         (check-current-custodian who)

racket/src/io/unsafe/port.rkt

@@ -42,6 +42,7 @@
 (define (unsafe-port->file-descriptor p)
   (define fd (fd-port-fd p))
   (and fd
+       (not (rktio_fd_is_pending_open rktio fd))
        (rktio_fd_system_fd rktio fd)))
 
 (define (unsafe-port->socket p)

racket/src/racket/src/port.c

@@ -3296,6 +3296,30 @@ scheme_file_stream_port_p (int argc, Scheme_Object *argv[])
   return scheme_false;
 }
 
+Scheme_Object *scheme_port_waiting_peer_p(int argc, Scheme_Object *argv[])
+{
+  Scheme_Object *p = argv[0];
+
+  if (SCHEME_OUTPUT_PORTP(p)) {
+    Scheme_Output_Port *op;
+
+    op = scheme_output_port_record(p);
+
+    if (SAME_OBJ(op->sub_type, fd_output_port_type)) {
+      rktio_fd_t *rfd = ((Scheme_FD *)op->port_data)->fd;
+      if (rktio_fd_is_pending_open(scheme_rktio, rfd))
+        return scheme_true;
+    }
+  } else if (SCHEME_INPUT_PORTP(p)) {
+    /* ok */
+  } else {
+    scheme_wrong_contract("port-waiting-peer?", "port?", 0, argc, argv);
+    ESCAPED_BEFORE_HERE;
+  }
+
+  return scheme_false;
+}
+
 int scheme_get_port_file_descriptor(Scheme_Object *p, intptr_t *_fd)
 {
   intptr_t fd = 0;
@@ -3325,11 +3349,14 @@ int scheme_get_port_file_descriptor(Scheme_Object *p, intptr_t *_fd)
 	fd = MSC_IZE (fileno)((FILE *)((Scheme_Output_File *)op->port_data)->f);
 	fd_ok = 1;
       } else if (SAME_OBJ(op->sub_type, fd_output_port_type))  {
-	fd = rktio_fd_system_fd(scheme_rktio, ((Scheme_FD *)op->port_data)->fd);
+        rktio_fd_t *rfd = ((Scheme_FD *)op->port_data)->fd;
+        if (!rktio_fd_is_pending_open(scheme_rktio, rfd)) {
+          fd = rktio_fd_system_fd(scheme_rktio, rfd);
           fd_ok = 1;
         }
       }
     }
+  }
 
   if (!fd_ok)
     return 0;
@@ -6234,6 +6261,17 @@ static Scheme_Object *subprocess(int c, Scheme_Object *args[])
   if (!stdin_fd || !stdout_fd || !stderr_fd)
     scheme_custodian_check_available(NULL, name, "file-stream");
 
+  /* In case `stdout_fd` or `stderr_fd` is a fifo with no read end
+     open, wait for it. */
+  if (stdout_fd && rktio_fd_is_pending_open(scheme_rktio, stdout_fd)) {
+    a[0] = args[0];
+    scheme_sync(1, a);
+  }
+  if (stderr_fd && rktio_fd_is_pending_open(scheme_rktio, stderr_fd)) {
+    a[0] = args[2];
+    scheme_sync(1, a);
+  }
+
   /*--------------------------------------*/
   /*        Create subprocess             */
   /*--------------------------------------*/

racket/src/racket/src/portfun.c

@@ -8,8 +8,9 @@
 static Scheme_Object *input_port_p (int, Scheme_Object *[]);
 static Scheme_Object *output_port_p (int, Scheme_Object *[]);
 static Scheme_Object *port_closed_p (int, Scheme_Object *[]);
-static Scheme_Object *current_input_port (int, Scheme_Object *[]);
 static Scheme_Object *string_port_p(int, Scheme_Object *[]);
+static Scheme_Object *waiting_peer_port_p(int, Scheme_Object *[]);
+static Scheme_Object *current_input_port (int, Scheme_Object *[]);
 static Scheme_Object *current_output_port (int, Scheme_Object *[]);
 static Scheme_Object *current_error_port (int, Scheme_Object *[]);
 static Scheme_Object *make_input_port (int, Scheme_Object *[]);
@@ -221,6 +222,7 @@ scheme_init_port_fun(Scheme_Startup_Env *env)
   ADD_FOLDING_PRIM("file-stream-port?",      scheme_file_stream_port_p,  1, 1, 1, env);
   ADD_FOLDING_PRIM("string-port?",           string_port_p,              1, 1, 1, env);
   ADD_FOLDING_PRIM("terminal-port?",         scheme_terminal_port_p,     1, 1, 1, env);
+  ADD_FOLDING_PRIM("port-waiting-peer?",     scheme_port_waiting_peer_p, 1, 1, 1, env);
 
   ADD_NONCM_PRIM("port-closed?",             port_closed_p,          1, 1, env); 
   ADD_NONCM_PRIM("open-input-file",          open_input_file,        1, 3, env);

racket/src/racket/src/schminc.h

@@ -14,7 +14,7 @@
 
 #define USE_COMPILED_STARTUP 1
 
-#define EXPECTED_PRIM_COUNT 1454
+#define EXPECTED_PRIM_COUNT 1455
 
 #ifdef MZSCHEME_SOMETHING_OMITTED
 # undef USE_COMPILED_STARTUP

racket/src/racket/src/schpriv.h

@@ -3606,6 +3606,7 @@ void scheme_flush_orig_outputs(void);
 void scheme_flush_if_output_fds(Scheme_Object *o);
 Scheme_Object *scheme_file_stream_port_p(int, Scheme_Object *[]);
 Scheme_Object *scheme_terminal_port_p(int, Scheme_Object *[]);
+Scheme_Object *scheme_port_waiting_peer_p(int, Scheme_Object *[]);
 Scheme_Object *scheme_do_open_input_file(char *name, int offset, int argc, Scheme_Object *argv[], 
                                          int internal, int for_module);
 Scheme_Object *scheme_do_open_output_file(char *name, int offset, int argc, Scheme_Object *argv[], int and_read, 

racket/src/racket/src/schvers.h

@@ -16,7 +16,7 @@
 #define MZSCHEME_VERSION_X 7
 #define MZSCHEME_VERSION_Y 4
 #define MZSCHEME_VERSION_Z 0
-#define MZSCHEME_VERSION_W 4
+#define MZSCHEME_VERSION_W 5
 
 /* A level of indirection makes `#` work as needed: */
 #define AS_a_STR_HELPER(x) #x

racket/src/rktio/rktio.def

@@ -12,6 +12,7 @@ rktio_fd_is_socket
 rktio_fd_is_udp
 rktio_fd_is_terminal
 rktio_fd_is_text_converted
+rktio_fd_is_pending_open
 rktio_fd_modes
 rktio_open
 rktio_close

racket/src/rktio/rktio.h

@@ -193,7 +193,8 @@ RKTIO_EXTERN rktio_fd_t *rktio_system_fd(rktio_t *rktio, intptr_t system_fd, int
    use `RKTIO_OPEN_INIT`. */
 
 RKTIO_EXTERN_NOERR intptr_t rktio_fd_system_fd(rktio_t *rktio, rktio_fd_t *rfd);
-/* Extracts a native file descriptor or socket. */
+/* Extracts a native file descriptor or socket. A file descriptor must
+   not be in pending-open mode as reported by `rktio_fd_is_pending_open`. */
 
 RKTIO_EXTERN rktio_bool_t rktio_fd_is_regular_file(rktio_t *rktio, rktio_fd_t *rfd);
 RKTIO_EXTERN rktio_bool_t rktio_fd_is_directory(rktio_t *rktio, rktio_fd_t *rfd);
@@ -203,8 +204,14 @@ RKTIO_EXTERN rktio_bool_t rktio_fd_is_terminal(rktio_t *rktio, rktio_fd_t *rfd);
 /* The functions mostly report values of recorded mode flags. */
 
 RKTIO_EXTERN rktio_bool_t rktio_fd_is_text_converted(rktio_t *rktio, rktio_fd_t *rfd);
-/* Reports whether RKTIO_OPEN_TEXT was use and has an effect. The
-   RKTIO_OPEN_TEXT flag has an effect only on Windows. */
+/* Reports whether `RKTIO_OPEN_TEXT` was use and has an effect. The
+   `RKTIO_OPEN_TEXT` flag has an effect only on Windows. */
+
+RKTIO_EXTERN rktio_bool_t rktio_fd_is_pending_open(rktio_t *rktio, rktio_fd_t *rfd);
+/* Reports whether `rfd` will block on writing because it corresponds
+   to the write end of a fifo that has no open reader. In that case,
+   `rktio_fd_system_fd` cannot report a file descriptor and `rktio_ltps_add`
+   will error with `RKTIO_ERROR_UNSUPPORTED`. */
 
 RKTIO_EXTERN_NOERR int rktio_fd_modes(rktio_t *rktio, rktio_fd_t *rfd);
 /* Returns all of the recorded mode flags, including those provided to
@@ -216,7 +223,9 @@ RKTIO_EXTERN rktio_fd_t *rktio_open(rktio_t *rktio, rktio_const_string_t src, in
    in read mode, and can report `RKTIO_ERROR_IS_A_DIRECTORY`,
    `RKTIO_ERROR_EXISTS`, or `RKTIO_ERROR_ACCESS_DENIED` in place of a
    system error in write mode. On Windows, can report
-   `RKTIO_ERROR_UNSUPPORTED_TEXT_MODE`. */
+   `RKTIO_ERROR_UNSUPPORTED_TEXT_MODE`. If `modes` has `RKTIO_OPEN_WRITE`
+   without `RKTIO_OPEN_READ`, then the result may be a file descriptor
+   in pending-open mode until the read end is opened. */
 
 RKTIO_EXTERN rktio_ok_t rktio_close(rktio_t *rktio, rktio_fd_t *fd);
 /* Can report `RKTIO_ERROR_EXISTS` in place of system error,
@@ -582,7 +591,8 @@ RKTIO_EXTERN rktio_process_result_t *rktio_process(rktio_t *rktio,
                                                    rktio_const_string_t current_directory,
                                                    rktio_envvars_t *envvars,
                                                    int flags);
-/* `flags` flags: */
+/* The output file descriptors `stdin_fd` must not be a pending-open
+   descriptor. The `flags` are: */
 #define RKTIO_PROCESS_NEW_GROUP                 (1<<0)
 #define RKTIO_PROCESS_STDOUT_AS_STDERR          (1<<1)
 #define RKTIO_PROCESS_WINDOWS_EXACT_CMDLINE     (1<<2)

racket/src/rktio/rktio.inc

@@ -12,6 +12,7 @@ Sforeign_symbol("rktio_fd_is_socket", (void *)rktio_fd_is_socket);
 Sforeign_symbol("rktio_fd_is_udp", (void *)rktio_fd_is_udp);
 Sforeign_symbol("rktio_fd_is_terminal", (void *)rktio_fd_is_terminal);
 Sforeign_symbol("rktio_fd_is_text_converted", (void *)rktio_fd_is_text_converted);
+Sforeign_symbol("rktio_fd_is_pending_open", (void *)rktio_fd_is_pending_open);
 Sforeign_symbol("rktio_fd_modes", (void *)rktio_fd_modes);
 Sforeign_symbol("rktio_open", (void *)rktio_open);
 Sforeign_symbol("rktio_close", (void *)rktio_close);

racket/src/rktio/rktio.rktl

@@ -255,6 +255,11 @@
  rktio_bool_t
  rktio_fd_is_text_converted
  (((ref rktio_t) rktio) ((ref rktio_fd_t) rfd)))
+(define-function
+ ()
+ rktio_bool_t
+ rktio_fd_is_pending_open
+ (((ref rktio_t) rktio) ((ref rktio_fd_t) rfd)))
 (define-function
  ()
  int

racket/src/rktio/rktio_fd.c

@@ -29,6 +29,9 @@ struct rktio_fd_t {
   int bufcount;
   char buffer[1];
 # endif
+# ifdef RKTIO_USE_PENDING_OPEN
+  struct open_in_thread_t *pending;
+# endif
 #endif
   
 #ifdef RKTIO_SYSTEM_WINDOWS
@@ -201,9 +204,46 @@ rktio_fd_t *rktio_system_fd(rktio_t *rktio, intptr_t system_fd, int modes)
   return rfd;
 }
 
+int rktio_fd_is_pending_open(rktio_t *rktio, rktio_fd_t *rfd)
+{
+#ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending)
+    rktio_pending_open_poll(rktio, rfd, rfd->pending);
+  return !!rfd->pending;
+#else
+  return 0;
+#endif
+}
+
+#ifdef RKTIO_USE_PENDING_OPEN
+rktio_fd_t *rktio_pending_system_fd(rktio_t *rktio, struct open_in_thread_t *oit, int modes)
+{
+  rktio_fd_t *rfd;
+
+  rfd = calloc(1, sizeof(rktio_fd_t));
+  rfd->modes = (modes - (modes & RKTIO_OPEN_INIT));
+  rfd->pending = oit;
+
+  return rfd;
+}
+
+void rktio_update_system_fd(rktio_t *rktio, rktio_fd_t *rfd, int fd, int modes)
+/* Convert an open-pending rfd to a normal rfd. The given `modes` adds
+   to (and completes) the modes provided to `rktio_pending_system_fd`. */
+{
+  rfd->fd = fd;
+  rfd->modes |= modes;
+  rfd->pending = NULL;
+}
+#endif
+
+
 intptr_t rktio_internal_fd_system_fd(rktio_fd_t *rfd)
 {
 #ifdef RKTIO_SYSTEM_UNIX
+# ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending) return -1;
+# endif
   return rfd->fd;
 #endif
 #ifdef RKTIO_SYSTEM_WINDOWS
@@ -272,6 +312,14 @@ int rktio_fd_is_udp(rktio_t *rktio, rktio_fd_t *rfd)
   return ((rfd->modes & RKTIO_OPEN_UDP) ? 1 : 0);
 }
 
+int rktio_fd_is_open_pending(rktio_t *rktio, rktio_fd_t *rfd)
+{
+#ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending) return 1;
+#endif
+  return 0;
+}
+
 int rktio_fd_modes(rktio_t *rktio, rktio_fd_t *rfd)
 {
   return rfd->modes;
@@ -298,6 +346,9 @@ int rktio_system_fd_is_terminal(rktio_t *rktio, intptr_t fd)
 
 int rktio_fd_is_terminal(rktio_t *rktio, rktio_fd_t *rfd)
 {
+#ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending) return 0;
+#endif
   return rktio_system_fd_is_terminal(rktio, (intptr_t)rfd->fd);
 }
 
@@ -315,6 +366,13 @@ rktio_fd_t *rktio_dup(rktio_t *rktio, rktio_fd_t *rfd)
 #ifdef RKTIO_SYSTEM_UNIX
   intptr_t nfd;
 
+# ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending) {
+    rktio_pending_open_retain(rktio, rfd->pending);
+    return rktio_pending_system_fd(rktio, rfd->pending, rfd->modes);
+  }
+# endif
+
   do {
     nfd = dup(rfd->fd);
   } while (nfd == -1 && errno == EINTR);
@@ -375,16 +433,22 @@ void rktio_reliably_close(intptr_t s)
 #endif
 
 static rktio_ok_t do_close(rktio_t *rktio /* maybe NULL */, rktio_fd_t *rfd, int set_error)
+/* The `rktio` argument can be NULL for a detached file descriptor */
 {
   int ok;
 
 #ifdef RKTIO_SYSTEM_UNIX
   int cr;
 
+# ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending)
+    cr = rktio_pending_open_release(rktio, rfd->pending);
+  else
+# endif
     cr = rktio_reliably_close_err(rfd->fd);
 
 # ifdef RKTIO_USE_FCNTL_AND_FORK_FOR_FILE_LOCKS
-  if (rktio && !(rfd->modes & RKTIO_OPEN_SOCKET))
+  if (rktio && !(rfd->modes & RKTIO_OPEN_SOCKET) && (!cr || !set_error))
     rktio_release_lockf(rktio, rfd->fd);
 # endif
 
@@ -446,12 +510,24 @@ void rktio_forget(rktio_t *rktio, rktio_fd_t *rfd)
 
 rktio_fd_transfer_t *rktio_fd_detach(rktio_t *rktio, rktio_fd_t *rfd)
 {
+#ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending)
+    rktio_pending_open_detach(rktio, rfd->pending);
+#endif
+
   return (rktio_fd_transfer_t *)rfd;
 }
 
 rktio_fd_t *rktio_fd_attach(rktio_t *rktio, rktio_fd_transfer_t *rfdt)
 {
-  return (rktio_fd_t *)rfdt;
+  rktio_fd_t *rfd = (rktio_fd_t *)rfdt;
+
+#ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending)
+    rktio_pending_open_attach(rktio, rfd->pending);
+#endif
+
+  return rfd;
 }
 
 void rktio_fd_close_transfer(rktio_fd_transfer_t *rfdt)
@@ -592,6 +668,19 @@ int poll_write_ready_or_flushed(rktio_t *rktio, rktio_fd_t *rfd, int check_flush
     return RKTIO_POLL_READY;
   else {
     int sr;
+
+# ifdef RKTIO_USE_PENDING_OPEN
+    if (rfd->pending) {
+      int errval = rktio_pending_open_poll(rktio, rfd, rfd->pending);
+      if (errval) {
+        errno = errval;
+        get_posix_error();
+        return RKTIO_POLL_ERROR;
+      } else if (rfd->pending)
+        return RKTIO_POLL_NOT_READY;
+    }
+# endif
+
 # ifdef HAVE_POLL_SYSCALL
     struct pollfd pfd[1];
     pfd[0].fd = rfd->fd;
@@ -616,7 +705,7 @@ int poll_write_ready_or_flushed(rktio_t *rktio, rktio_fd_t *rfd, int check_flush
       /* Mac OS X 10.8 and 10.9: select() seems to claim that a pipe
          is always ready for output. To work around that problem,
          kqueue() support might be used for pipes, but that has different
-         problems. The poll() code above should be used, instead. */
+         problems. The pol() code above should be used, instead. */
       sr = select(rfd->fd + 1, NULL, RKTIO_FDS(writefds), RKTIO_FDS(exnfds), &time);
     } while ((sr == -1) && (errno == EINTR));
 # endif
@@ -693,6 +782,13 @@ void rktio_poll_add(rktio_t *rktio, rktio_fd_t *rfd, rktio_poll_set_t *fds, int
 #ifdef RKTIO_SYSTEM_UNIX
   rktio_poll_set_t *fds2;
 
+# ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending) {
+    rktio_poll_add_pending_open(rktio, rfd, rfd->pending, fds);
+    return;
+  }
+# endif
+
   if (modes & RKTIO_POLL_READ) {
     RKTIO_FD_SET(rfd->fd, fds);
   }
@@ -1233,6 +1329,18 @@ intptr_t rktio_write(rktio_t *rktio, rktio_fd_t *rfd, const char *buffer, intptr
   if (rfd->modes & RKTIO_OPEN_SOCKET)
     return rktio_socket_write(rktio, rfd, buffer, len);
 
+# ifdef RKTIO_USE_PENDING_OPEN
+  if (rfd->pending) {
+    int errval = rktio_pending_open_poll(rktio, rfd, rfd->pending);
+    if (errval) {
+      errno = errval;
+      get_posix_error();
+      return RKTIO_WRITE_ERROR;
+    } else if (rfd->pending)
+      return 0;
+  }
+# endif
+
   flags = fcntl(rfd->fd, F_GETFL, 0);
   if (!(flags & RKTIO_NONBLOCKING))
     fcntl(rfd->fd, F_SETFL, flags | RKTIO_NONBLOCKING);

racket/src/rktio/rktio_file.c

@@ -11,6 +11,18 @@
 #ifdef RKTIO_SYSTEM_WINDOWS
 # include <windows.h>
 #endif
+#ifdef RKTIO_USE_PENDING_OPEN
+#include <string.h>
+#endif
+
+#ifdef RKTIO_SYSTEM_UNIX
+static rktio_fd_t *finish_unix_fd_creation(rktio_t *rktio, int fd, int modes, rktio_fd_t *existing_rfd);
+#endif
+
+#ifdef RKTIO_USE_PENDING_OPEN
+static rktio_fd_t *open_via_thread(rktio_t *rktio, const char *filename, int modes, int flags);
+static int do_pending_open_release(rktio_t *rktio, struct open_in_thread_t *data, int close_fd);
+#endif
 
 /*========================================================================*/
 /* Opening a file                                                         */
@@ -106,8 +118,6 @@ static rktio_fd_t *open_write(rktio_t *rktio, const char *filename, int modes)
 #ifdef RKTIO_SYSTEM_UNIX
   int fd;
   int flags;
-  struct stat buf;
-  int cr;
 
   flags = (((modes & RKTIO_OPEN_READ) ? O_RDWR : O_WRONLY)
            | ((modes & RKTIO_OPEN_MUST_EXIST ? 0 : O_CREAT)));
@@ -124,48 +134,20 @@ static rktio_fd_t *open_write(rktio_t *rktio, const char *filename, int modes)
   } while ((fd == -1) && (errno == EINTR));
 
   if (errno == ENXIO) {
-    /* FIFO with no reader? Try opening in RW mode: */
+    /* FIFO with no reader? */
+#ifdef RKTIO_USE_PENDING_OPEN
+    return open_via_thread(rktio, filename, modes, flags | RKTIO_BINARY);
+#else
+    /* Try opening in RW mode: */
     flags -= O_WRONLY;
     flags |= O_RDWR;
     do {
       fd = open(filename, flags | RKTIO_NONBLOCKING | RKTIO_BINARY, 0666);
     } while ((fd == -1) && (errno == EINTR));
+#endif
   }
 
-  if (fd == -1) {
-    if (errno == EISDIR) {
-      set_racket_error(RKTIO_ERROR_IS_A_DIRECTORY);
-      return NULL;
-    } else if (errno == EEXIST) {
-      set_racket_error(RKTIO_ERROR_EXISTS);
-      return NULL;
-    } else if (errno == EACCES) {
-      set_racket_error(RKTIO_ERROR_ACCESS_DENIED);
-      return NULL;
-    }
-
-    if (fd == -1) {
-      get_posix_error();
-      return NULL;
-    }
-  }
-
-  do {
-    cr = fstat(fd, &buf);
-  } while ((cr == -1) && (errno == EINTR));
-
-  if (cr) {
-    get_posix_error();
-    do {
-      cr = close(fd);
-    } while ((cr == -1) && (errno == EINTR));
-    return NULL;
-  }
-
-  return rktio_system_fd(rktio, fd, (modes
-                                     | (S_ISREG(buf.st_mode)
-                                        ? RKTIO_OPEN_REGFILE
-                                        : RKTIO_OPEN_NOT_REGFILE)));
+  return finish_unix_fd_creation(rktio, fd, modes, NULL);
 #endif
 #ifdef RKTIO_SYSTEM_WINDOWS
   HANDLE fd;
@@ -236,6 +218,54 @@ static rktio_fd_t *open_write(rktio_t *rktio, const char *filename, int modes)
 #endif
 }
 
+#ifdef RKTIO_SYSTEM_UNIX
+static rktio_fd_t *finish_unix_fd_creation(rktio_t *rktio, int fd, int modes, rktio_fd_t *existing_rfd)
+{
+  struct stat buf;
+  int cr;
+
+  if (fd == -1) {
+    if (errno == EISDIR) {
+      set_racket_error(RKTIO_ERROR_IS_A_DIRECTORY);
+      return NULL;
+    } else if (errno == EEXIST) {
+      set_racket_error(RKTIO_ERROR_EXISTS);
+      return NULL;
+    } else if (errno == EACCES) {
+      set_racket_error(RKTIO_ERROR_ACCESS_DENIED);
+      return NULL;
+    }
+
+    if (fd == -1) {
+      get_posix_error();
+      return NULL;
+    }
+  }
+
+  do {
+    cr = fstat(fd, &buf);
+  } while ((cr == -1) && (errno == EINTR));
+
+  if (cr) {
+    get_posix_error();
+    do {
+      cr = close(fd);
+    } while ((cr == -1) && (errno == EINTR));
+    return NULL;
+  }
+
+  modes |= (S_ISREG(buf.st_mode)
+            ? RKTIO_OPEN_REGFILE
+            : RKTIO_OPEN_NOT_REGFILE);
+
+  if (existing_rfd) {
+    rktio_update_system_fd(rktio, existing_rfd, fd, modes);
+    return existing_rfd;
+  } else
+    return rktio_system_fd(rktio, fd, modes);
+}
+#endif
+
 rktio_fd_t *rktio_open(rktio_t *rktio, const char *filename, int modes)
 {
   if (modes & RKTIO_OPEN_WRITE)
@@ -373,3 +403,260 @@ rktio_ok_t rktio_set_file_size(rktio_t *rktio, rktio_fd_t *rfd, rktio_filesize_t
   }
 #endif
 }
+
+/*========================================================================*/
+/* Thread for blocking open                                               */
+/*========================================================================*/
+
+/* When opening a fifo for writing, then there's no way to open in
+   non-blocking mode and then poll for whether reader is ready.
+   Instead, we have to open in a separate thread. When the open
+   succeeds, post to a waiting rktio_t's signal handle.
+
+   To allow a blocked-on-opening file descriptor to be transferred
+   across rktio_t domains, the record that represents the extra thread
+   must be sharable among multiple OS threads, and we must in general
+   keep an array of handles.
+
+   If the blocked-omn-opening file descriptor is closed, then we have
+   to be able to cancel the thread. This isn't so bad, since we want
+   to support canceling only while the `open` system call is
+   blocked. */
+
+#ifdef RKTIO_USE_PENDING_OPEN
+
+/* An instance of `open_in_thread_t` can be shared by multiple threads
+   (i.e., multiple `rktio_t` instances) */
+typedef struct open_in_thread_t {
+  pthread_mutex_t lock;
+  int ready;
+  pthread_cond_t ready_cond; /* wait until helper thread is ready (including cleanup) */
+  char *filename;
+  int flags;
+  int done;
+  int fd;
+  int errval;
+  int refcount;
+  pthread_t th;
+  int num_handles;
+  rktio_signal_handle_t **handles;
+} open_in_thread_t;
+
+
+static void free_open_in_thread(open_in_thread_t *data)
+{
+  pthread_detach(data->th);
+  if (data->handles)
+    free(data->handles);
+  free(data->filename);
+  free(data);
+}
+
+static void cleanup_open_thread(void *_data)
+{
+  int i, refcount;
+  open_in_thread_t *data = (open_in_thread_t *)_data;
+
+  pthread_mutex_lock(&data->lock);
+  for (i = 0; i < data->num_handles; i++)
+    if (data->handles[i])
+      rktio_signal_received_at(data->handles[i]);
+  refcount = data->refcount;
+  data->done = 1;
+  pthread_mutex_unlock(&data->lock);
+
+  if (!refcount) {
+    if (data->fd != -1)
+      rktio_reliably_close(data->fd);
+    free_open_in_thread(data);
+  }
+}
+
+static void *do_open_in_thread(void *_data)
+{
+  open_in_thread_t *data = (open_in_thread_t *)_data;
+  int fd;
+  int old_type;
+
+  /* To be on the safe side, disable cancelation except 
+     just around the call to `open` */
+  pthread_setcanceltype(PTHREAD_CANCEL_DISABLE, &old_type);
+  pthread_cleanup_push(cleanup_open_thread, data);
+
+  pthread_mutex_lock(&data->lock);
+  data->ready = 1;
+  pthread_cond_signal(&data->ready_cond);
+  pthread_mutex_unlock(&data->lock);
+
+  data->fd = -1;
+
+  pthread_setcanceltype(PTHREAD_CANCEL_DEFERRED, NULL);
+
+  /* Cancelation only possible during `open`, a which point it's ok
+     and sufficient to run `cleanup_open_thread`; we're assuming that
+     either `open` returns with a file descriptor or the thread is
+     canceled before `open` returns, but not both (otherwise there
+     could be a space leak) */
+  do {
+    fd = open(data->filename, data->flags, 0666);
+  } while ((fd == -1) && (errno == EINTR));
+
+  pthread_setcanceltype(PTHREAD_CANCEL_DISABLE, NULL);
+
+  data->fd = fd;
+  if (fd == -1)
+    data->errval = errno;
+
+  /* Runs `cleanup_open_thread` while popping it: */
+  pthread_cleanup_pop(1);
+
+  /* In general, a function tha changes the cancelation type should
+     restore it before returning */
+  pthread_setcanceltype(old_type, NULL);
+
+  return NULL;
+}
+
+static rktio_fd_t *open_via_thread(rktio_t *rktio, const char *filename, int modes, int flags)
+{
+  open_in_thread_t *data;
+
+  data = calloc(1, sizeof(open_in_thread_t));
+
+  data->refcount = 1;
+
+  data->filename = strdup(filename);
+  data->flags = flags;
+  pthread_mutex_init(&data->lock, NULL);
+  pthread_cond_init(&data->ready_cond, NULL);
+
+  data->num_handles = 1;
+  data->handles = malloc(sizeof(rktio_signal_handle_t*));
+  data->handles[0] = rktio_get_signal_handle(rktio);
+
+  (void)pthread_create(&data->th, NULL, do_open_in_thread, data);
+
+  pthread_mutex_lock(&data->lock);
+  if (!data->ready)
+    pthread_cond_wait(&data->ready_cond, &data->lock);
+  pthread_mutex_unlock(&data->lock);
+
+  return rktio_pending_system_fd(rktio, data, modes);
+}
+
+int rktio_pending_open_poll(rktio_t *rktio, rktio_fd_t *existing_rfd, struct open_in_thread_t *data)
+/* non-zero result is an errno value */
+{
+  int done;
+
+  pthread_mutex_lock(&data->lock);
+  done = data->done;
+  pthread_mutex_unlock(&data->lock);
+
+  if (done) {
+    if (data->fd == -1)
+      return data->errval;
+    else {
+      int fd = data->fd;
+      (void)do_pending_open_release(rktio, data, 0);
+      if (!finish_unix_fd_creation(rktio, fd, 0, existing_rfd)) {
+        /* Posix error must be saved in `rktio` */
+        return rktio->errid;
+      }
+      return 0;
+    }
+  } else
+    return 0;
+}
+
+void rktio_poll_add_pending_open(rktio_t *rktio, rktio_fd_t *rfd, struct open_in_thread_t *data, rktio_poll_set_t *fds)
+{
+  int done;
+
+  pthread_mutex_lock(&data->lock);
+  done = data->done;
+  pthread_mutex_unlock(&data->lock);
+
+  if (done)
+    rktio_poll_set_add_nosleep(rktio, fds);
+}
+
+void rktio_pending_open_attach(rktio_t *rktio, struct open_in_thread_t *data)
+{
+  int i;
+  rktio_signal_handle_t *h = rktio_get_signal_handle(rktio);
+
+  pthread_mutex_lock(&data->lock);
+  if (!data->done) {
+    for (i = 0; i < data->num_handles; i++)
+      if (!data->handles[i]) {
+        data->handles[i] = h;
+        break;
+      }
+    if (i >= data->num_handles) {
+      rktio_signal_handle_t **old = data->handles;
+      int n = (2 * data->num_handles);
+      data->handles = calloc(n, sizeof(rktio_signal_handle_t*));
+      memcpy(data->handles, old, data->num_handles * sizeof(rktio_signal_handle_t*));
+      data->handles[data->num_handles] = h;
+      data->num_handles = n;
+    }
+  }
+  pthread_mutex_unlock(&data->lock);
+}
+
+static void do_detach(rktio_t *rktio, struct open_in_thread_t *data)
+{
+  int i;
+  rktio_signal_handle_t *h = rktio_get_signal_handle(rktio);
+
+  for (i = 0; i < data->num_handles; i++)
+    if (data->handles[i] == h)
+      data->handles[i] = NULL;
+}
+
+void rktio_pending_open_detach(rktio_t *rktio, struct open_in_thread_t *data)
+{
+  pthread_mutex_lock(&data->lock);
+  do_detach(rktio, data);
+  pthread_mutex_unlock(&data->lock);
+}
+
+
+void rktio_pending_open_retain(rktio_t *rktio, struct open_in_thread_t *data)
+{
+  pthread_mutex_lock(&data->lock);
+  data->refcount++;
+  pthread_mutex_unlock(&data->lock);
+}
+
+int do_pending_open_release(rktio_t *rktio, struct open_in_thread_t *data, int close_fd)
+/* The rktio argument can be NULL for a detached use */
+{
+  int bye;
+
+  pthread_mutex_lock(&data->lock);
+  --data->refcount;
+  bye = (data->done && !data->refcount);
+  if (!bye) {
+    do_detach(rktio, data);
+    if (!data->refcount)
+      pthread_cancel(data->th);
+  }
+  pthread_mutex_unlock(&data->lock);
+
+  if (bye) {
+    if (close_fd && (data->fd != -1))
+      rktio_reliably_close(data->fd);
+    free_open_in_thread(data);
+  }
+
+  return 0;
+}
+
+int rktio_pending_open_release(rktio_t *rktio, struct open_in_thread_t *data)
+{
+  return do_pending_open_release(rktio, data, 1);
+}
+
+#endif

racket/src/rktio/rktio_ltps.c

@@ -208,7 +208,16 @@ rktio_ltps_handle_t *rktio_ltps_add(rktio_t *rktio, rktio_ltps_t *lt, rktio_fd_t
 # endif
   rktio_ltps_handle_pair_t *v;
   rktio_ltps_handle_t *s;
-  intptr_t fd = rktio_fd_system_fd(rktio, rfd);
+  intptr_t fd;
+
+#ifdef RKTIO_USE_PENDING_OPEN
+  if (rktio_fd_is_pending_open(rktio, rfd)) {
+    set_racket_error(RKTIO_ERROR_UNSUPPORTED);
+    return NULL;
+  }
+#endif
+
+  fd = rktio_fd_system_fd(rktio, rfd);
 
 # ifdef HAVE_KQUEUE_SYSCALL
   if (!rktio_fd_is_socket(rktio, rfd)) {

racket/src/rktio/rktio_private.h

@@ -339,6 +339,23 @@ void rktio_close_fds_after_fork(int len, int skip1, int skip2, int skip3);
 
 int rktio_system_fd_is_terminal(rktio_t *rktio, intptr_t fd);
 
+#ifdef RKTIO_USE_PTHREADS
+# define RKTIO_USE_PENDING_OPEN
+#endif
+
+#ifdef RKTIO_USE_PENDING_OPEN
+struct open_in_thread_t;
+rktio_fd_t *rktio_pending_system_fd(rktio_t *rktio, struct open_in_thread_t *oit, int modes);
+void rktio_update_system_fd(rktio_t *rktio, rktio_fd_t *rfd, int fd, int modes);
+int rktio_fd_is_pending_open(rktio_t *rktio, rktio_fd_t *rfd);
+int rktio_pending_open_poll(rktio_t *rktio, rktio_fd_t *rfd, struct open_in_thread_t *oit);
+void rktio_poll_add_pending_open(rktio_t *rktio, rktio_fd_t *rfd, struct open_in_thread_t *oit, rktio_poll_set_t *fds);
+void rktio_pending_open_detach(rktio_t *rktio, struct open_in_thread_t *oit);
+void rktio_pending_open_attach(rktio_t *rktio, struct open_in_thread_t *oit);
+void rktio_pending_open_retain(rktio_t *rktio, struct open_in_thread_t *oit);
+int rktio_pending_open_release(rktio_t *rktio, struct open_in_thread_t *oit);
+#endif
+
 void *rktio_envvars_to_block(rktio_t *rktio, rktio_envvars_t *envvars);
 
 void rktio_stop_fs_change(rktio_t *rktio);