Merge branch 'lukas/erts/non-blocking-shell'

* lukas/erts/non-blocking-shell:
  Fix io:columns/0 timeout when invoked via user
  kernel,ssh: Add synchronous user_drv protocol
  erts: Make writing to non-tty fds non-blocking
  erts: Make tty driver non-blocking

7 files changed, 614 insertions, 153 deletions

diff --git a/erts/emulator/beam/erl_alloc.types b/erts/emulator/beam/erl_alloc.types
index df33bbc2f7..6163e4129a 100644
--- a/erts/emulator/beam/erl_alloc.types
+++ b/erts/emulator/beam/erl_alloc.types
@@ -414,6 +414,8 @@ type	CS_PROG_PATH	LONG_LIVED	SYSTEM		cs_prog_path
 type	ENVIRONMENT	TEMPORARY	SYSTEM		environment
 type	PUTENV_STR	SYSTEM		SYSTEM		putenv_string
 type	PRT_REP_EXIT	STANDARD	SYSTEM		port_report_exit
+type	SYS_BLOCKING	STANDARD	SYSTEM		sys_blocking
+type	SYS_WRITE_BUF	TEMPORARY	SYSTEM		sys_write_buf
 
 +endif
 
diff --git a/erts/emulator/drivers/unix/ttsl_drv.c b/erts/emulator/drivers/unix/ttsl_drv.c
index 491e0a090e..be2fee1f25 100644
--- a/erts/emulator/drivers/unix/ttsl_drv.c
+++ b/erts/emulator/drivers/unix/ttsl_drv.c
@@ -32,6 +32,10 @@ static ErlDrvData ttysl_start(ErlDrvPort, char*);
 
 #ifdef HAVE_TERMCAP  /* else make an empty driver that can not be opened */
 
+#ifndef WANT_NONBLOCKING
+#define WANT_NONBLOCKING
+#endif
+
 #include "sys.h"
 #include <ctype.h>
 #include <stdlib.h>
@@ -39,6 +43,7 @@ static ErlDrvData ttysl_start(ErlDrvPort, char*);
 #include <string.h>
 #include <signal.h>
 #include <fcntl.h>
+#include <limits.h>
 #include <locale.h>
 #include <unistd.h>
 #include <termios.h>
@@ -57,6 +62,14 @@ static ErlDrvData ttysl_start(ErlDrvPort, char*);
 #include <langinfo.h>
 #endif
 
+#if defined IOV_MAX
+#define MAXIOV IOV_MAX
+#elif defined UIO_MAXIOV
+#define MAXIOV UIO_MAXIOV
+#else
+#define MAXIOV 16
+#endif
+
 #define TRUE 1
 #define FALSE 0
 
@@ -80,12 +93,15 @@ static volatile int cols_needs_update = FALSE;
 #define OP_INSC 2
 #define OP_DELC 3
 #define OP_BEEP 4
+#define OP_PUTC_SYNC 5
 /* Control op */
 #define CTRL_OP_GET_WINSIZE 100
 #define CTRL_OP_GET_UNICODE_STATE 101
 #define CTRL_OP_SET_UNICODE_STATE 102
 
-
+/* We use 1024 as the buf size as that was the default buf size of FILE streams
+   on all platforms that I checked. */
+#define TTY_BUFFSIZE 1024
 
 static int lbuf_size = BUFSIZ;
 static Uint32 *lbuf;		/* The current line buffer */
@@ -113,13 +129,19 @@ static int lpos;                /* The current "cursor position" in the line buf
 /* Main interface functions. */
 static void ttysl_stop(ErlDrvData);
 static void ttysl_from_erlang(ErlDrvData, char*, ErlDrvSizeT);
+static void ttysl_to_tty(ErlDrvData, ErlDrvEvent);
+static void ttysl_flush_tty(ErlDrvData);
 static void ttysl_from_tty(ErlDrvData, ErlDrvEvent);
 static void ttysl_stop_select(ErlDrvEvent, void*);
 static Sint16 get_sint16(char*);
 
 static ErlDrvPort ttysl_port;
 static int ttysl_fd;
-static FILE *ttysl_out;
+static int ttysl_terminate = 0;
+static int ttysl_send_ok = 0;
+static ErlDrvBinary *putcbuf;
+static int putcpos;
+static int putclen;
 
 /* Functions that work on the line buffer. */
 static int start_lbuf(void);
@@ -201,22 +223,22 @@ struct erl_drv_entry ttsl_driver_entry = {
     IF_IMPL(ttysl_stop),
     IF_IMPL(ttysl_from_erlang),
     IF_IMPL(ttysl_from_tty),
-    NULL,
-    "tty_sl",
-    NULL,
-    NULL,
+    IF_IMPL(ttysl_to_tty),
+    "tty_sl", /* driver_name */
+    NULL, /* finish */
+    NULL, /* handle */
     IF_IMPL(ttysl_control),
     NULL, /* timeout */
     NULL, /* outputv */
     NULL, /* ready_async */
-    NULL, /* flush */
+    IF_IMPL(ttysl_flush_tty),
     NULL, /* call */
     NULL, /* event */
     ERL_DRV_EXTENDED_MARKER,
     ERL_DRV_EXTENDED_MAJOR_VERSION,
     ERL_DRV_EXTENDED_MINOR_VERSION,
     0, /* ERL_DRV_FLAGs */
-    NULL,
+    NULL, /* handle2 */
     NULL, /* process_exit */
     IF_IMPL(ttysl_stop_select)
 };
@@ -296,8 +318,7 @@ static ErlDrvData ttysl_start(ErlDrvPort port, char* buf)
 	return ERL_DRV_ERROR_GENERAL;
     }
 
-    /* Open the terminal and set the terminal */
-    ttysl_out = fdopen(ttysl_fd, "w");
+    SET_NONBLOCKING(ttysl_fd);
 
 #ifdef PRIMITIVE_UTF8_CHECK
     setlocale(LC_CTYPE, "");  /* Set international environment, 
@@ -400,12 +421,14 @@ static void ttysl_stop(ErlDrvData ttysl_data)
 	stop_lbuf();
 	stop_termcap();
 	tty_reset(ttysl_fd);
-	driver_select(ttysl_port, (ErlDrvEvent)(UWord)ttysl_fd, ERL_DRV_READ|ERL_DRV_USE, 0);
+	driver_select(ttysl_port, (ErlDrvEvent)(UWord)ttysl_fd,
+                      ERL_DRV_WRITE|ERL_DRV_READ|ERL_DRV_USE, 0);
 	sys_sigset(SIGCONT, SIG_DFL);
 	sys_sigset(SIGWINCH, SIG_DFL);
     }
     ttysl_port = (ErlDrvPort)-1;
     ttysl_fd = -1;
+    ttysl_terminate = 0;
     /* return TRUE; */
 }
 
@@ -650,10 +673,26 @@ static int check_buf_size(byte *s, int n)
 
 static void ttysl_from_erlang(ErlDrvData ttysl_data, char* buf, ErlDrvSizeT count)
 {
+    ErlDrvSizeT sz;
+
+    sz = driver_sizeq(ttysl_port);
+
+    putclen = count > TTY_BUFFSIZE ? TTY_BUFFSIZE : count;
+    putcbuf = driver_alloc_binary(putclen);
+    putcpos = 0;
+
     if (lpos > MAXSIZE) 
 	put_chars((byte*)"\n", 1);
 
     switch (buf[0]) {
+    case OP_PUTC_SYNC:
+        /* Using sync means that we have to send an ok to the
+           controlling process for each command call. We delay
+           sending ok if the driver queue exceeds a certain size.
+           We do not set ourselves as a busy port, as this
+           could be very bad for user_drv, if it gets blocked on
+           the port_command. */
+        /* fall through */
     case OP_PUTC:
 	DEBUGLOG(("OP: Putc(%lu)",(unsigned long) count-1));
 	if (check_buf_size((byte*)buf+1, count-1) == 0)
@@ -678,10 +717,104 @@ static void ttysl_from_erlang(ErlDrvData ttysl_data, char* buf, ErlDrvSizeT coun
 	/* Unknown op, just ignore. */
 	break;
     }
-    fflush(ttysl_out);
+
+    driver_enq_bin(ttysl_port,putcbuf,0,putcpos);
+
+    if (sz == 0) {
+        for (;;) {
+            int written, qlen;
+            SysIOVec *iov;
+
+            iov = driver_peekq(ttysl_port,&qlen);
+            if (iov)
+                written = writev(ttysl_fd, iov, qlen > MAXIOV ? MAXIOV : qlen);
+            else
+                written = 0;
+            if (written < 0) {
+                if (errno == EAGAIN) {
+                    driver_select(ttysl_port,(ErlDrvEvent)(long)ttysl_fd,
+                                  ERL_DRV_USE|ERL_DRV_WRITE,1);
+                    break;
+                } else {
+                    /* we ignore all other errors */
+                    break;
+                }
+            } else {
+                if (driver_deq(ttysl_port, written) == 0)
+                    break;
+            }
+        }
+    }
+
+    if (buf[0] == OP_PUTC_SYNC) {
+        if (driver_sizeq(ttysl_port) > TTY_BUFFSIZE && !ttysl_terminate) {
+            /* We delay sending the ack until the buffer has been consumed */
+            ttysl_send_ok = 1;
+        } else {
+            ErlDrvTermData spec[] = {
+                ERL_DRV_PORT, driver_mk_port(ttysl_port),
+                ERL_DRV_ATOM, driver_mk_atom("ok"),
+                ERL_DRV_TUPLE, 2
+            };
+            ASSERT(ttysl_send_ok == 0);
+            erl_drv_output_term(driver_mk_port(ttysl_port), spec,
+                                sizeof(spec) / sizeof(spec[0]));
+        }
+    }
+
     return; /* TRUE; */
 }
 
+static void ttysl_to_tty(ErlDrvData ttysl_data, ErlDrvEvent fd) {
+    for (;;) {
+        int written, qlen;
+        SysIOVec *iov;
+        ErlDrvSizeT sz;
+
+        iov = driver_peekq(ttysl_port,&qlen);
+        if (iov)
+            written = writev(ttysl_fd, iov, qlen > MAXIOV ? MAXIOV : qlen);
+        else
+            written = 0;
+        if (written < 0) {
+            if (errno == EAGAIN) {
+                break;
+            } else {
+                /* we ignore all other errors */
+            }
+        } else {
+            sz = driver_deq(ttysl_port, written);
+            if (sz < TTY_BUFFSIZE && ttysl_send_ok) {
+                ErlDrvTermData spec[] = {
+                    ERL_DRV_PORT, driver_mk_port(ttysl_port),
+                    ERL_DRV_ATOM, driver_mk_atom("ok"),
+                    ERL_DRV_TUPLE, 2
+                };
+                ttysl_send_ok = 0;
+                erl_drv_output_term(driver_mk_port(ttysl_port), spec,
+                                    sizeof(spec) / sizeof(spec[0]));
+            }
+            if (sz == 0) {
+                driver_select(ttysl_port,(ErlDrvEvent)(long)ttysl_fd,
+                              ERL_DRV_WRITE,0);
+                if (ttysl_terminate)
+                    /* flush has been called, which means we should terminate
+                       when queue is empty. This will not send any exit
+                       message */
+                    driver_failure_atom(ttysl_port, "normal");
+                break;
+            }
+        }
+    }
+
+    return;
+}
+
+static void ttysl_flush_tty(ErlDrvData ttysl_data) {
+    ttysl_terminate = 1;
+    return;
+}
+
 static void ttysl_from_tty(ErlDrvData ttysl_data, ErlDrvEvent fd)
 {
     byte b[1024];
@@ -1070,7 +1203,14 @@ static int write_buf(Uint32 *s, int n)
 /* The basic procedure for outputting one character. */
 static int outc(int c)
 {
-    return (int)putc(c, ttysl_out);
+    putcbuf->orig_bytes[putcpos++] = c;
+    if (putcpos == putclen) {
+        driver_enq_bin(ttysl_port,putcbuf,0,putclen);
+        putcpos = 0;
+        putclen = TTY_BUFFSIZE;
+        putcbuf = driver_alloc_binary(BUFSIZ);
+    }
+    return 1;
 }
 
 static int move_cursor(int from, int to)
diff --git a/erts/emulator/drivers/win32/ttsl_drv.c b/erts/emulator/drivers/win32/ttsl_drv.c
index 502cb58dfa..851c336a11 100644
--- a/erts/emulator/drivers/win32/ttsl_drv.c
+++ b/erts/emulator/drivers/win32/ttsl_drv.c
@@ -46,6 +46,7 @@ static int rows;		/* Number of rows available. */
 #define OP_INSC 2
 #define OP_DELC 3
 #define OP_BEEP 4
+#define OP_PUTC_SYNC 5
 
 /* Control op */
 #define CTRL_OP_GET_WINSIZE 100
@@ -458,6 +459,7 @@ static void ttysl_from_erlang(ErlDrvData ttysl_data, char* buf, ErlDrvSizeT coun
 
     switch (buf[0]) {
     case OP_PUTC:
+    case OP_PUTC_SYNC:
 	DEBUGLOG(("OP: Putc(%I64u)",(unsigned long long)count-1));
 	if (check_buf_size((byte*)buf+1, count-1) == 0)
 	    return; 
@@ -481,6 +483,20 @@ static void ttysl_from_erlang(ErlDrvData ttysl_data, char* buf, ErlDrvSizeT coun
 	/* Unknown op, just ignore. */
 	break;
     }
+
+    if (buf[0] == OP_PUTC_SYNC) {
+        /* On windows we do a blocking write to the tty so we just
+           send the ack immidiately. If at some point in the future
+           someone has a problem with tty output being blocking
+           this has to be changed. */
+        ErlDrvTermData spec[] = {
+            ERL_DRV_PORT, driver_mk_port(ttysl_port),
+            ERL_DRV_ATOM, driver_mk_atom("ok"),
+            ERL_DRV_TUPLE, 2
+        };
+        erl_drv_output_term(driver_mk_port(ttysl_port), spec,
+                            sizeof(spec) / sizeof(spec[0]));
+    }
     return;
 }
 
diff --git a/erts/emulator/sys/unix/sys.c b/erts/emulator/sys/unix/sys.c
index c3d7440409..eed93cb2a0 100644
--- a/erts/emulator/sys/unix/sys.c
+++ b/erts/emulator/sys/unix/sys.c
@@ -34,6 +34,7 @@
 #include <termios.h>
 #include <ctype.h>
 #include <sys/utsname.h>
+#include <sys/select.h>
 
 #ifdef ISC32
 #include <sys/bsdtypes.h>
@@ -91,8 +92,10 @@ static erts_smp_rwmtx_t environ_rwmtx;
 #  else
 #    define CHLDWTHR 0
 #  endif
+#  define FDBLOCK 1
 #else
 #  define CHLDWTHR 0
+#  define FDBLOCK 0
 #endif
 /*
  * [OTP-3906]
@@ -121,6 +124,15 @@ struct ErtsSysReportExit_ {
 #endif
 };
 
+/* Used by the fd driver iff the fd could not be set to non-blocking */
+typedef struct ErtsSysBlocking_ {
+    ErlDrvPDL pdl;
+    int res;
+    int err;
+    unsigned int pkey;
+} ErtsSysBlocking;
+
+
 /* This data is shared by these drivers - initialized by spawn_init() */
 static struct driver_data {
     ErlDrvPort port_num;
@@ -129,6 +141,8 @@ static struct driver_data {
     int pid;
     int alive;
     int status;
+    int terminating;
+    ErtsSysBlocking *blocking;
 } *driver_data;			/* indexed by fd */
 
 static ErtsSysReportExit *report_exit_list;
@@ -1108,11 +1122,16 @@ void fini_getenv_state(GETENV_STATE *state)
 /* Driver interfaces */
 static ErlDrvData spawn_start(ErlDrvPort, char*, SysDriverOpts*);
 static ErlDrvData fd_start(ErlDrvPort, char*, SysDriverOpts*);
+#if FDBLOCK
+static void fd_async(void *);
+static void fd_ready_async(ErlDrvData drv_data, ErlDrvThreadData thread_data);
+#endif
 static ErlDrvSSizeT fd_control(ErlDrvData, unsigned int, char *, ErlDrvSizeT,
 			       char **, ErlDrvSizeT);
 static ErlDrvData vanilla_start(ErlDrvPort, char*, SysDriverOpts*);
 static int spawn_init(void);
 static void fd_stop(ErlDrvData);
+static void fd_flush(ErlDrvData);
 static void stop(ErlDrvData);
 static void ready_input(ErlDrvData, ErlDrvEvent);
 static void ready_output(ErlDrvData, ErlDrvEvent);
@@ -1157,8 +1176,12 @@ struct erl_drv_entry fd_driver_entry = {
     fd_control,
     NULL,
     outputv,
-    NULL, /* ready_async */
-    NULL, /* flush */
+#if FDBLOCK
+    fd_ready_async, /* ready_async */
+#else
+    NULL,
+#endif
+    fd_flush, /* flush */
     NULL, /* call */
     NULL, /* event */
     ERL_DRV_EXTENDED_MARKER,
@@ -1212,13 +1235,28 @@ static RETSIGTYPE onchld(int signum)
 #endif
 }
 
+static int set_blocking_data(struct driver_data *dd) {
+
+    dd->blocking = erts_alloc(ERTS_ALC_T_SYS_BLOCKING, sizeof(ErtsSysBlocking));
+
+    erts_smp_atomic_add_nob(&sys_misc_mem_sz, sizeof(ErtsSysBlocking));
+
+    dd->blocking->pdl = driver_pdl_create(dd->port_num);
+    dd->blocking->res = 0;
+    dd->blocking->err = 0;
+    dd->blocking->pkey = driver_async_port_key(dd->port_num);
+
+    return 1;
+}
+
 static int set_driver_data(ErlDrvPort port_num,
 			   int ifd,
 			   int ofd,
 			   int packet_bytes,
 			   int read_write,
 			   int exit_status,
-			   int pid)
+			   int pid,
+                           int is_blocking)
 {
     Port *prt;
     ErtsSysReportExit *report_exit;
@@ -1250,8 +1288,13 @@ static int set_driver_data(ErlDrvPort port_num,
 	driver_data[ifd].pid = pid;
 	driver_data[ifd].alive = 1;
 	driver_data[ifd].status = 0;
+        driver_data[ifd].terminating = 0;
+        driver_data[ifd].blocking = NULL;
 	if (read_write & DO_WRITE) {
 	    driver_data[ifd].ofd = ofd;
+            if (is_blocking && FDBLOCK)
+                if (!set_blocking_data(driver_data+ifd))
+                    return -1;
 	    if (ifd != ofd)
 		driver_data[ofd] = driver_data[ifd];  /* structure copy */
 	} else {		/* DO_READ only */
@@ -1267,6 +1310,11 @@ static int set_driver_data(ErlDrvPort port_num,
 	driver_data[ofd].pid = pid;
 	driver_data[ofd].alive = 1;
 	driver_data[ofd].status = 0;
+        driver_data[ofd].terminating = 0;
+        driver_data[ofd].blocking = NULL;
+        if (is_blocking && FDBLOCK)
+            if (!set_blocking_data(driver_data+ofd))
+                return -1;
 	return(ofd);
     }
 }
@@ -1276,6 +1324,7 @@ static int spawn_init()
    int i;
 #if CHLDWTHR
    erts_thr_opts_t thr_opts = ERTS_THR_OPTS_DEFAULT_INITER;
+
    thr_opts.detached = 0;
    thr_opts.suggested_stack_size = 0; /* Smallest possible */
 #endif
@@ -1755,7 +1804,7 @@ static ErlDrvData spawn_start(ErlDrvPort port_num, char* name, SysDriverOpts* op
     }
 
     res = set_driver_data(port_num, ifd[0], ofd[1], opts->packet_bytes,
-			  opts->read_write, opts->exit_status, pid);
+			  opts->read_write, opts->exit_status, pid, 0);
     /* Don't unblock SIGCHLD until now, since the call above must
        first complete putting away the info about our new subprocess. */
     unblock_signals();
@@ -1840,6 +1889,7 @@ static ErlDrvData fd_start(ErlDrvPort port_num, char* name,
 			   SysDriverOpts* opts)
 {
     ErlDrvData res;
+    int non_blocking = 0;
 
     if (((opts->read_write & DO_READ) && opts->ifd >= max_files) ||
 	((opts->read_write & DO_WRITE) && opts->ofd >= max_files))
@@ -1912,6 +1962,20 @@ static ErlDrvData fd_start(ErlDrvPort port_num, char* name,
      * case - it can be called with any old pre-existing file descriptors,
      * the relations between which (if they're even two) we can only guess
      * at - still, we try our best...
+     *
+     * Added note OTP 18: Some systems seem to use stdout/stderr to log data
+     * using unix pipes, so we cannot allow the system to block on a write.
+     * Therefore we use an async thread to write the data to fd's that could
+     * not be set to non-blocking. When no async threads are available we
+     * fall back on the old behaviour.
+     *
+     * Also the guarantee about what is delivered to the OS has changed.
+     * Pre 18 the fd driver did no flushing of data before terminating.
+     * Now it does. This is because we want to be able to guarantee that things
+     * such as escripts and friends really have outputted all data before
+     * terminating. This could potentially block the termination of the system
+     * for a very long time, but if the user wants to terminate fast she should
+     * use erlang:halt with flush=false.
      */
 
     if (opts->read_write & DO_READ) {
@@ -1934,6 +1998,7 @@ static ErlDrvData fd_start(ErlDrvPort port_num, char* name,
 		       imagine a scenario where setting non-blocking mode
 		       here would cause problems - go ahead and do it. */
 
+                    non_blocking = 1;
 		    SET_NONBLOCKING(opts->ofd);
 
 		} else {	/* output fd is a tty, input fd isn't */
@@ -1976,6 +2041,7 @@ static ErlDrvData fd_start(ErlDrvPort port_num, char* name,
 			(nfd = open(tty, O_WRONLY)) != -1) {
 			dup2(nfd, opts->ofd);
 			close(nfd);
+                        non_blocking = 1;
 			SET_NONBLOCKING(opts->ofd);
 		    }
 		}
@@ -1984,8 +2050,9 @@ static ErlDrvData fd_start(ErlDrvPort port_num, char* name,
     }
     CHLD_STAT_LOCK;
     res = (ErlDrvData)(long)set_driver_data(port_num, opts->ifd, opts->ofd,
-				      opts->packet_bytes,
-				      opts->read_write, 0, -1);
+                                            opts->packet_bytes,
+                                            opts->read_write, 0, -1,
+                                            !non_blocking);
     CHLD_STAT_UNLOCK;
     return res;
 }
@@ -2011,14 +2078,30 @@ static void nbio_stop_fd(ErlDrvPort prt, int fd)
     SET_BLOCKING(fd);
 }
 
-static void fd_stop(ErlDrvData fd)  /* Does not close the fds */
+static void fd_stop(ErlDrvData ev)  /* Does not close the fds */
 {
     int ofd;
+    int fd = (int)(long)ev;
+    ErlDrvPort prt = driver_data[fd].port_num;
     
-    nbio_stop_fd(driver_data[(int)(long)fd].port_num, (int)(long)fd);
-    ofd = driver_data[(int)(long)fd].ofd;
-    if (ofd != (int)(long)fd && ofd != -1) 
-	nbio_stop_fd(driver_data[(int)(long)fd].port_num, (int)(long)ofd);
+#if FDBLOCK
+    if (driver_data[fd].blocking) {
+        erts_free(ERTS_ALC_T_SYS_BLOCKING,driver_data[fd].blocking);
+        driver_data[fd].blocking = NULL;
+        erts_smp_atomic_add_nob(&sys_misc_mem_sz, -1*sizeof(ErtsSysBlocking));
+    }
+#endif
+
+    nbio_stop_fd(prt, fd);
+    ofd = driver_data[fd].ofd;
+    if (ofd != fd && ofd != -1)
+	nbio_stop_fd(prt, ofd);
+}
+
+static void fd_flush(ErlDrvData fd)
+{
+    if (!driver_data[(int)(long)fd].terminating)
+        driver_data[(int)(long)fd].terminating = 1;
 }
 
 static ErlDrvData vanilla_start(ErlDrvPort port_num, char* name,
@@ -2041,8 +2124,8 @@ static ErlDrvData vanilla_start(ErlDrvPort port_num, char* name,
 
     CHLD_STAT_LOCK;
     res = (ErlDrvData)(long)set_driver_data(port_num, fd, fd,
-				      opts->packet_bytes,
-				      opts->read_write, 0, -1);
+                                            opts->packet_bytes,
+                                            opts->read_write, 0, -1, 0);
     CHLD_STAT_UNLOCK;
     return res;
 }
@@ -2079,6 +2162,7 @@ static void stop(ErlDrvData fd)
     }
 }
 
+/* used by fd_driver */
 static void outputv(ErlDrvData e, ErlIOVec* ev)
 {
     int fd = (int)(long)e;
@@ -2104,12 +2188,21 @@ static void outputv(ErlDrvData e, ErlIOVec* ev)
     ev->iov[0].iov_base = lbp;
     ev->iov[0].iov_len = pb;
     ev->size += pb;
+
+    if (driver_data[fd].blocking && FDBLOCK)
+        driver_pdl_lock(driver_data[fd].blocking->pdl);
+
     if ((sz = driver_sizeq(ix)) > 0) {
 	driver_enqv(ix, ev, 0);
+
+        if (driver_data[fd].blocking && FDBLOCK)
+            driver_pdl_unlock(driver_data[fd].blocking->pdl);
+
 	if (sz + ev->size >= (1 << 13))
 	    set_busy_port(ix, 1);
     }
-    else {
+    else if (!driver_data[fd].blocking || !FDBLOCK) {
+        /* We try to write directly if the fd in non-blocking */
 	int vsize = ev->vsize > MAX_VSIZE ? MAX_VSIZE : ev->vsize;
 
 	n = writev(ofd, (const void *) (ev->iov), vsize);
@@ -2125,10 +2218,22 @@ static void outputv(ErlDrvData e, ErlIOVec* ev)
 	driver_enqv(ix, ev, n);  /* n is the skip value */
 	driver_select(ix, ofd, ERL_DRV_WRITE|ERL_DRV_USE, 1);
     }
+#if FDBLOCK
+    else {
+        if (ev->size != 0) {
+            driver_enqv(ix, ev, 0);
+            driver_pdl_unlock(driver_data[fd].blocking->pdl);
+            driver_async(ix, &driver_data[fd].blocking->pkey,
+                         fd_async, driver_data+fd, NULL);
+        } else {
+            driver_pdl_unlock(driver_data[fd].blocking->pdl);
+        }
+    }
+#endif
     /* return 0;*/
 }
 
-
+/* Used by spawn_driver and vanilla driver */
 static void output(ErlDrvData e, char* buf, ErlDrvSizeT len)
 {
     int fd = (int)(long)e;
@@ -2191,6 +2296,23 @@ static int port_inp_failure(ErlDrvPort port_num, int ready_fd, int res)
     ASSERT(res <= 0);
     (void) driver_select(port_num, ready_fd, ERL_DRV_READ|ERL_DRV_WRITE, 0); 
     clear_fd_data(ready_fd);
+
+    if (driver_data[ready_fd].blocking && FDBLOCK) {
+        driver_pdl_lock(driver_data[ready_fd].blocking->pdl);
+        if (driver_sizeq(driver_data[ready_fd].port_num) > 0) {
+            driver_pdl_unlock(driver_data[ready_fd].blocking->pdl);
+            /* We have stuff in the output queue, so we just
+               set the state to terminating and wait for fd_async_ready
+               to terminate the port */
+            if (res == 0)
+                driver_data[ready_fd].terminating = 2;
+            else
+                driver_data[ready_fd].terminating = -err;
+            return 0;
+        }
+        driver_pdl_unlock(driver_data[ready_fd].blocking->pdl);
+    }
+
     if (res == 0) {
 	if (driver_data[ready_fd].report_exit) {
 	    CHLD_STAT_LOCK;
@@ -2241,6 +2363,7 @@ static void ready_input(ErlDrvData e, ErlDrvEvent ready_fd)
     port_num = driver_data[fd].port_num;
     packet_bytes = driver_data[fd].packet_bytes;
 
+
     if (packet_bytes == 0) {
 	byte *read_buf = (byte *) erts_alloc(ERTS_ALC_T_SYS_READ_BUF,
 					     ERTS_SYS_READ_BUF_SZ);
@@ -2364,6 +2487,8 @@ static void ready_output(ErlDrvData e, ErlDrvEvent ready_fd)
 
     if ((iv = (struct iovec*) driver_peekq(ix, &vsize)) == NULL) {
 	driver_select(ix, ready_fd, ERL_DRV_WRITE, 0);
+        if (driver_data[fd].terminating)
+            driver_failure_atom(driver_data[fd].port_num,"normal");
 	return; /* 0; */
     }
     vsize = vsize > MAX_VSIZE ? MAX_VSIZE : vsize;
@@ -2389,6 +2514,82 @@ static void stop_select(ErlDrvEvent fd, void* _)
     close((int)fd);
 }
 
+#if FDBLOCK
+
+static void
+fd_async(void *async_data)
+{
+    int res;
+    struct driver_data *dd = (struct driver_data*)async_data;
+    SysIOVec      *iov0;
+    SysIOVec      *iov;
+    int            iovlen;
+    int            iovcnt;
+    int            p;
+    /* much of this code is stolen from efile_drv:invoke_writev */
+    driver_pdl_lock(dd->blocking->pdl);
+    iov0 = driver_peekq(dd->port_num, &iovlen);
+    /* Calculate iovcnt */
+    for (p = 0, iovcnt = 0; iovcnt < iovlen;
+         p += iov0[iovcnt++].iov_len)
+        ;
+    iov = erts_alloc_fnf(ERTS_ALC_T_SYS_WRITE_BUF,
+                         sizeof(SysIOVec)*iovcnt);
+    if (!iov) {
+        res = -1;
+        errno = ENOMEM;
+        erts_free(ERTS_ALC_T_SYS_WRITE_BUF, iov);
+        driver_pdl_unlock(dd->blocking->pdl);
+    } else {
+        memcpy(iov,iov0,iovcnt*sizeof(SysIOVec));
+        driver_pdl_unlock(dd->blocking->pdl);
+
+        res = writev(dd->ofd, iov, iovlen);
+
+        erts_free(ERTS_ALC_T_SYS_WRITE_BUF, iov);
+    }
+    dd->blocking->res = res;
+    dd->blocking->err = errno;
+}
+
+void fd_ready_async(ErlDrvData drv_data,
+                    ErlDrvThreadData thread_data) {
+    struct driver_data *dd = (struct driver_data *)thread_data;
+    ErlDrvPort port_num = dd->port_num;
+
+    ASSERT(dd->blocking);
+    ASSERT(dd == (driver_data + (int)(long)drv_data));
+
+    if (dd->blocking->res > 0) {
+        driver_pdl_lock(dd->blocking->pdl);
+        if (driver_deq(port_num, dd->blocking->res) == 0) {
+            driver_pdl_unlock(dd->blocking->pdl);
+            set_busy_port(port_num, 0);
+            if (dd->terminating) {
+                /* The port is has been ordered to terminate
+                   from either fd_flush or port_inp_failure */
+                if (dd->terminating == 1)
+                    driver_failure_atom(port_num, "normal");
+                else if (dd->terminating == 2)
+                    driver_failure_eof(port_num);
+                else if (dd->terminating < 0)
+                    driver_failure_posix(port_num, -dd->terminating);
+                return; /* -1; */
+            }
+        } else {
+            driver_pdl_unlock(dd->blocking->pdl);
+            /* still data left to write in queue */
+            driver_async(port_num, &dd->blocking->pkey, fd_async, dd, NULL);
+            return /* 0; */;
+        }
+    } else if (dd->blocking->res < 0) {
+        driver_failure_posix(port_num, dd->blocking->err);
+        return; /* -1; */
+    }
+    return; /* 0; */
+}
+
+#endif
 
 void erts_do_break_handling(void)
 {
@@ -2658,18 +2859,30 @@ void sys_preload_end(Preload* p)
     /* Nothing */
 }
 
-/* Read a key from console (?) */
-
+/* Read a key from console, used by break.c
+   Here we assume that all schedulers are stopped so that erl_poll
+   does not interfere with the select below.
+*/
 int sys_get_key(fd)
 int fd;
 {
-    int c;
+    int c, ret;
     unsigned char rbuf[64];
+    fd_set fds;
 
     fflush(stdout);		/* Flush query ??? */
 
-    if ((c = read(fd,rbuf,64)) <= 0) {
-      return c; 
+    FD_ZERO(&fds);
+    FD_SET(fd,&fds);
+
+    ret = select(fd+1, &fds, NULL, NULL, NULL);
+
+    if (ret == 1) {
+        do {
+            c = read(fd,rbuf,64);
+        } while (c < 0 && errno == EAGAIN);
+        if (c <= 0)
+            return c;
     }
 
     return rbuf[0]; 
diff --git a/lib/kernel/src/group.erl b/lib/kernel/src/group.erl
index b36dbf33dd..046885f885 100644
--- a/lib/kernel/src/group.erl
+++ b/lib/kernel/src/group.erl
@@ -111,8 +111,13 @@ start_shell1(Fun) ->
 server_loop(Drv, Shell, Buf0) ->
     receive
 	{io_request,From,ReplyAs,Req} when is_pid(From) ->
-	    Buf = io_request(Req, From, ReplyAs, Drv, Buf0),
-	    server_loop(Drv, Shell, Buf);
+            %% This io_request may cause a transition to a couple of
+            %% selective receive loops elsewhere in this module.
+            Buf = io_request(Req, From, ReplyAs, Drv, Buf0),
+            server_loop(Drv, Shell, Buf);
+        {reply,{{From,ReplyAs},Reply}} ->
+            io_reply(From, ReplyAs, Reply),
+	    server_loop(Drv, Shell, Buf0);
 	{driver_id,ReplyTo} ->
 	    ReplyTo ! {self(),driver_id,Drv},
 	    server_loop(Drv, Shell, Buf0);
@@ -172,10 +177,13 @@ set_unicode_state(Drv,Bool) ->
 			   
 
 io_request(Req, From, ReplyAs, Drv, Buf0) ->
-    case io_request(Req, Drv, Buf0) of
+    case io_request(Req, Drv, {From,ReplyAs}, Buf0) of
 	{ok,Reply,Buf} ->
 	    io_reply(From, ReplyAs, Reply),
 	    Buf;
+        {noreply,Buf} ->
+            %% We expect a {reply,_} message from the Drv when request is done
+            Buf;
 	{error,Reply,Buf} ->
 	    io_reply(From, ReplyAs, Reply),
 	    Buf;
@@ -196,78 +204,85 @@ io_request(Req, From, ReplyAs, Drv, Buf0) ->
 %% io_request({put_chars,unicode,Binary}, Drv, Buf) when is_binary(Binary) -> 
 %%     send_drv(Drv, {put_chars,Binary}),
 %%     {ok,ok,Buf};
-io_request({put_chars,unicode,Chars}, Drv, Buf) ->
+%%
+%% These put requests have to be synchronous to the driver as otherwise
+%% there is no guarantee that the data has actually been printed.
+io_request({put_chars,unicode,Chars}, Drv, From, Buf) ->
     case catch unicode:characters_to_binary(Chars,utf8) of
 	Binary when is_binary(Binary) ->
-	    send_drv(Drv, {put_chars, unicode, Binary}),
-	    {ok,ok,Buf};
+	    send_drv(Drv, {put_chars_sync, unicode, Binary, {From,ok}}),
+	    {noreply,Buf};
 	_ ->
 	    {error,{error,{put_chars, unicode,Chars}},Buf}
     end;
-io_request({put_chars,unicode,M,F,As}, Drv, Buf) ->
+io_request({put_chars,unicode,M,F,As}, Drv, From, Buf) ->
     case catch apply(M, F, As) of
 	Binary when is_binary(Binary) ->
-	    send_drv(Drv, {put_chars, unicode,Binary}),
-	    {ok,ok,Buf};
+	    send_drv(Drv, {put_chars_sync, unicode, Binary, {From,ok}}),
+	    {noreply,Buf};
 	Chars ->
 	    case catch unicode:characters_to_binary(Chars,utf8) of
 		B when is_binary(B) ->
-		    send_drv(Drv, {put_chars, unicode,B}),
-		    {ok,ok,Buf};
+		    send_drv(Drv, {put_chars_sync, unicode, B, {From,ok}}),
+		    {noreply,Buf};
 		_ ->
 		    {error,{error,F},Buf}
 	    end
     end;
-io_request({put_chars,latin1,Binary}, Drv, Buf) when is_binary(Binary) -> 
-    send_drv(Drv, {put_chars, unicode,unicode:characters_to_binary(Binary,latin1)}),
-    {ok,ok,Buf};
-io_request({put_chars,latin1,Chars}, Drv, Buf) ->
+io_request({put_chars,latin1,Binary}, Drv, From, Buf) when is_binary(Binary) -> 
+    send_drv(Drv, {put_chars_sync, unicode,
+                   unicode:characters_to_binary(Binary,latin1),
+                   {From,ok}}),
+    {noreply,Buf};
+io_request({put_chars,latin1,Chars}, Drv, From, Buf) ->
     case catch unicode:characters_to_binary(Chars,latin1) of
 	Binary when is_binary(Binary) ->
-	    send_drv(Drv, {put_chars, unicode,Binary}),
-	    {ok,ok,Buf};
+	    send_drv(Drv, {put_chars_sync, unicode, Binary, {From,ok}}),
+	    {noreply,Buf};
 	_ ->
 	    {error,{error,{put_chars,latin1,Chars}},Buf}
     end;
-io_request({put_chars,latin1,M,F,As}, Drv, Buf) ->
+io_request({put_chars,latin1,M,F,As}, Drv, From, Buf) ->
     case catch apply(M, F, As) of
 	Binary when is_binary(Binary) ->
-	    send_drv(Drv, {put_chars, unicode,unicode:characters_to_binary(Binary,latin1)}),
-	    {ok,ok,Buf};
+	    send_drv(Drv, {put_chars_sync, unicode,
+                           unicode:characters_to_binary(Binary,latin1),
+                           {From,ok}}),
+	    {noreply,Buf};
 	Chars ->
 	    case catch unicode:characters_to_binary(Chars,latin1) of
 		B when is_binary(B) ->
-		    send_drv(Drv, {put_chars, unicode,B}),
-		    {ok,ok,Buf};
+		    send_drv(Drv, {put_chars_sync, unicode, B, {From,ok}}),
+		    {noreply,Buf};
 		_ ->
 		    {error,{error,F},Buf}
 	    end
     end;
 
-io_request({get_chars,Encoding,Prompt,N}, Drv, Buf) ->
+io_request({get_chars,Encoding,Prompt,N}, Drv, _From, Buf) ->
     get_chars(Prompt, io_lib, collect_chars, N, Drv, Buf, Encoding);
-io_request({get_line,Encoding,Prompt}, Drv, Buf) ->
+io_request({get_line,Encoding,Prompt}, Drv, _From, Buf) ->
     get_chars(Prompt, io_lib, collect_line, [], Drv, Buf, Encoding);
-io_request({get_until,Encoding, Prompt,M,F,As}, Drv, Buf) ->
+io_request({get_until,Encoding, Prompt,M,F,As}, Drv, _From, Buf) ->
     get_chars(Prompt, io_lib, get_until, {M,F,As}, Drv, Buf, Encoding);
-io_request({get_password,_Encoding},Drv,Buf) ->
+io_request({get_password,_Encoding},Drv,_From,Buf) ->
     get_password_chars(Drv, Buf);
-io_request({setopts,Opts}, Drv, Buf) when is_list(Opts) ->
+io_request({setopts,Opts}, Drv, _From, Buf) when is_list(Opts) ->
     setopts(Opts, Drv, Buf);
-io_request(getopts, Drv, Buf) ->
+io_request(getopts, Drv, _From, Buf) ->
     getopts(Drv, Buf);
-io_request({requests,Reqs}, Drv, Buf) ->
-    io_requests(Reqs, {ok,ok,Buf}, Drv);
+io_request({requests,Reqs}, Drv, From, Buf) ->
+    io_requests(Reqs, {ok,ok,Buf}, From, Drv);
 
 %% New in R12
-io_request({get_geometry,columns},Drv,Buf) ->
+io_request({get_geometry,columns},Drv,_From,Buf) ->
     case get_tty_geometry(Drv) of
 	{W,_H} ->
 	    {ok,W,Buf};
 	_ ->
 	    {error,{error,enotsup},Buf}
     end;
-io_request({get_geometry,rows},Drv,Buf) ->
+io_request({get_geometry,rows},Drv,_From,Buf) ->
     case get_tty_geometry(Drv) of
 	{_W,H} ->
 	    {ok,H,Buf};
@@ -276,38 +291,49 @@ io_request({get_geometry,rows},Drv,Buf) ->
     end;
 
 %% BC with pre-R13
-io_request({put_chars,Chars}, Drv, Buf) ->
-    io_request({put_chars,latin1,Chars}, Drv, Buf);
-io_request({put_chars,M,F,As}, Drv, Buf) ->
-    io_request({put_chars,latin1,M,F,As}, Drv, Buf);
-io_request({get_chars,Prompt,N}, Drv, Buf) ->
-    io_request({get_chars,latin1,Prompt,N}, Drv, Buf);
-io_request({get_line,Prompt}, Drv, Buf) ->
-    io_request({get_line,latin1,Prompt}, Drv, Buf);
-io_request({get_until, Prompt,M,F,As}, Drv, Buf) ->
-    io_request({get_until,latin1, Prompt,M,F,As}, Drv, Buf);
-io_request(get_password,Drv,Buf) ->
-    io_request({get_password,latin1},Drv,Buf);
-
-
-
-io_request(_, _Drv, Buf) ->
+io_request({put_chars,Chars}, Drv, From, Buf) ->
+    io_request({put_chars,latin1,Chars}, Drv, From, Buf);
+io_request({put_chars,M,F,As}, Drv, From, Buf) ->
+    io_request({put_chars,latin1,M,F,As}, Drv, From, Buf);
+io_request({get_chars,Prompt,N}, Drv, From, Buf) ->
+    io_request({get_chars,latin1,Prompt,N}, Drv, From, Buf);
+io_request({get_line,Prompt}, Drv, From, Buf) ->
+    io_request({get_line,latin1,Prompt}, Drv, From, Buf);
+io_request({get_until, Prompt,M,F,As}, Drv, From, Buf) ->
+    io_request({get_until,latin1, Prompt,M,F,As}, Drv, From, Buf);
+io_request(get_password,Drv,From,Buf) ->
+    io_request({get_password,latin1},Drv,From,Buf);
+
+
+
+io_request(_, _Drv, _From, Buf) ->
     {error,{error,request},Buf}.
 
-%% Status = io_requests(RequestList, PrevStat, Drv)
-%%  Process a list of output requests as long as the previous status is 'ok'.
-
-io_requests([R|Rs], {ok,ok,Buf}, Drv) ->
-    io_requests(Rs, io_request(R, Drv, Buf), Drv);
-io_requests([_|_], Error, _Drv) ->
+%% Status = io_requests(RequestList, PrevStat, From, Drv)
+%%  Process a list of output requests as long as
+%%  the previous status is 'ok' or noreply.
+%%
+%%  We use undefined as the From for all but the last request
+%%  in order to discards acknowledgements from those requests.
+%%
+io_requests([R|Rs], {noreply,Buf}, From, Drv) ->
+    ReqFrom = if Rs =:= [] -> From; true -> undefined end,
+    io_requests(Rs, io_request(R, Drv, ReqFrom, Buf), From, Drv);
+io_requests([R|Rs], {ok,ok,Buf}, From, Drv) ->
+    ReqFrom = if Rs =:= [] -> From; true -> undefined end,
+    io_requests(Rs, io_request(R, Drv, ReqFrom, Buf), From, Drv);
+io_requests([_|_], Error, _From, _Drv) ->
     Error;
-io_requests([], Stat, _) ->
+io_requests([], Stat, _From, _) ->
     Stat.
 
 %% io_reply(From, ReplyAs, Reply)
 %%  The function for sending i/o command acknowledgement.
 %%  The ACK contains the return value.
 
+io_reply(undefined, _ReplyAs, _Reply) ->
+    %% Ignore these replies as they are generated from io_requests/4.
+    ok;
 io_reply(From, ReplyAs, Reply) ->
     From ! {io_reply,ReplyAs,Reply},
     ok.
@@ -619,6 +645,10 @@ more_data(What, Cont0, Drv, Ls, Encoding) ->
 	    io_request(Req, From, ReplyAs, Drv, []), %WRONG!!!
 	    send_drv_reqs(Drv, edlin:redraw_line(Cont)),
 	    get_line1({more_chars,Cont,[]}, Drv, Ls, Encoding);
+        {reply,{{From,ReplyAs},Reply}} ->
+            %% We take care of replies from puts here as well
+            io_reply(From, ReplyAs, Reply),
+            more_data(What, Cont0, Drv, Ls, Encoding);
 	{'EXIT',Drv,interrupt} ->
 	    interrupted;
 	{'EXIT',Drv,_} ->
@@ -641,6 +671,10 @@ get_line_echo_off1({Chars,[]}, Drv) ->
 	{io_request,From,ReplyAs,Req} when is_pid(From) ->
 	    io_request(Req, From, ReplyAs, Drv, []),
 	    get_line_echo_off1({Chars,[]}, Drv);
+        {reply,{{From,ReplyAs},Reply}} when From =/= undefined ->
+            %% We take care of replies from puts here as well
+            io_reply(From, ReplyAs, Reply),
+            get_line_echo_off1({Chars,[]},Drv);
 	{'EXIT',Drv,interrupt} ->
 	    interrupted;
 	{'EXIT',Drv,_} ->
@@ -790,6 +824,10 @@ get_password1({Chars,[]}, Drv) ->
 	    %% set to []. But do we expect anything but plain output?
 
 	    get_password1({Chars, []}, Drv);
+        {reply,{{From,ReplyAs},Reply}} ->
+            %% We take care of replies from puts here as well
+            io_reply(From, ReplyAs, Reply),
+            get_password1({Chars, []},Drv);
 	{'EXIT',Drv,interrupt} ->
 	    interrupted;
 	{'EXIT',Drv,_} ->
diff --git a/lib/kernel/src/user_drv.erl b/lib/kernel/src/user_drv.erl
index a91c23539d..e6ce85c379 100644
--- a/lib/kernel/src/user_drv.erl
+++ b/lib/kernel/src/user_drv.erl
@@ -29,6 +29,7 @@
 -define(OP_INSC,2).
 -define(OP_DELC,3).
 -define(OP_BEEP,4).
+-define(OP_PUTC_SYNC,5).
 % Control op
 -define(CTRL_OP_GET_WINSIZE,100).
 -define(CTRL_OP_GET_UNICODE_STATE,101).
@@ -133,7 +134,7 @@ server1(Iport, Oport, Shell) ->
 				      [erlang:system_info(system_version)]))},
 	       Iport, Oport),
     %% Enter the server loop.
-    server_loop(Iport, Oport, Curr, User, Gr).
+    server_loop(Iport, Oport, Curr, User, Gr, queue:new()).
 
 rem_sh_opts(Node) ->
     [{expand_fun,fun(B)-> rpc:call(Node,edlin_expand,expand,[B]) end}].
@@ -158,42 +159,41 @@ start_user() ->
 	    User
     end.
    
-server_loop(Iport, Oport, User, Gr) ->
+server_loop(Iport, Oport, User, Gr, IOQueue) ->
     Curr = gr_cur_pid(Gr),
     put(current_group, Curr),
-    server_loop(Iport, Oport, Curr, User, Gr).
+    server_loop(Iport, Oport, Curr, User, Gr, IOQueue).
 
-server_loop(Iport, Oport, Curr, User, Gr) ->
+server_loop(Iport, Oport, Curr, User, Gr, IOQueue) ->
     receive
 	{Iport,{data,Bs}} ->
 	    BsBin = list_to_binary(Bs),
 	    Unicode = unicode:characters_to_list(BsBin,utf8),
-	    port_bytes(Unicode, Iport, Oport, Curr, User, Gr);
+	    port_bytes(Unicode, Iport, Oport, Curr, User, Gr, IOQueue);
 	{Iport,eof} ->
 	    Curr ! {self(),eof},
-	    server_loop(Iport, Oport, Curr, User, Gr);
-	{User,Req} ->	% never block from user!
-	    io_request(Req, Iport, Oport),
-	    server_loop(Iport, Oport, Curr, User, Gr);
-	{Curr,tty_geometry} ->
-	    Curr ! {self(),tty_geometry,get_tty_geometry(Iport)},
-	    server_loop(Iport, Oport, Curr, User, Gr);
-	{Curr,get_unicode_state} ->
-	    Curr ! {self(),get_unicode_state,get_unicode_state(Iport)},
-	    server_loop(Iport, Oport, Curr, User, Gr);
-	{Curr,set_unicode_state, Bool} ->
-	    Curr ! {self(),set_unicode_state,set_unicode_state(Iport,Bool)},
-	    server_loop(Iport, Oport, Curr, User, Gr);
-	{Curr,Req} ->
-	    io_request(Req, Iport, Oport),
-	    server_loop(Iport, Oport, Curr, User, Gr);
+	    server_loop(Iport, Oport, Curr, User, Gr, IOQueue);
+        Req when element(1,Req) =:= User orelse element(1,Req) =:= Curr,
+                 tuple_size(Req) =:= 2 orelse tuple_size(Req) =:= 3 ->
+            %% We match {User|Curr,_}|{User|Curr,_,_}
+            NewQ = handle_req(Req, Iport, Oport, IOQueue),
+            server_loop(Iport, Oport, Curr, User, Gr, NewQ);
+        {Oport,ok} ->
+            %% We get this ok from the port, in io_request we store
+            %% info about where to send reply at head of queue
+            {{value,{Origin,Reply}},ReplyQ} = queue:out(IOQueue),
+            Origin ! {reply,Reply},
+            NewQ = handle_req(next, Iport, Oport, ReplyQ),
+            server_loop(Iport, Oport, Curr, User, Gr, NewQ);
 	{'EXIT',Iport,_R} ->
-	    server_loop(Iport, Oport, Curr, User, Gr);
+	    server_loop(Iport, Oport, Curr, User, Gr, IOQueue);
 	{'EXIT',Oport,_R} ->
-	    server_loop(Iport, Oport, Curr, User, Gr);
+	    server_loop(Iport, Oport, Curr, User, Gr, IOQueue);
+        {'EXIT',User,shutdown} ->               % force data to port
+	    server_loop(Iport, Oport, Curr, User, Gr, IOQueue);
 	{'EXIT',User,_R} ->			% keep 'user' alive
 	    NewU = start_user(),
-	    server_loop(Iport, Oport, Curr, NewU, gr_set_num(Gr, 1, NewU, {}));
+	    server_loop(Iport, Oport, Curr, NewU, gr_set_num(Gr, 1, NewU, {}), IOQueue);
 	{'EXIT',Pid,R} ->			% shell and group leader exit
 	    case gr_cur_pid(Gr) of
 		Pid when R =/= die ,
@@ -213,18 +213,51 @@ server_loop(Iport, Oport, Curr, User, Gr) ->
 			    {ok,Gr2} = gr_set_cur(gr_set_num(Gr1, Ix, Pid1, 
 							     {shell,start,Params}), Ix),
 			    put(current_group, Pid1),
-			    server_loop(Iport, Oport, Pid1, User, Gr2);
+			    server_loop(Iport, Oport, Pid1, User, Gr2, IOQueue);
 			_ -> % remote shell
 			    io_requests([{put_chars,unicode,"(^G to start new job) ***\n"}],
 					Iport, Oport),	    
-			    server_loop(Iport, Oport, Curr, User, Gr1)
+			    server_loop(Iport, Oport, Curr, User, Gr1, IOQueue)
 		    end;
 		_ ->				% not current, just remove it
-		    server_loop(Iport, Oport, Curr, User, gr_del_pid(Gr, Pid))
+		    server_loop(Iport, Oport, Curr, User, gr_del_pid(Gr, Pid), IOQueue)
 	    end;
 	_X ->
 	    %% Ignore unknown messages.
-	    server_loop(Iport, Oport, Curr, User, Gr)
+	    server_loop(Iport, Oport, Curr, User, Gr, IOQueue)
+    end.
+
+%% We always handle geometry and unicode requests
+handle_req({Curr,tty_geometry},Iport,_Oport,IOQueue) ->
+    Curr ! {self(),tty_geometry,get_tty_geometry(Iport)},
+    IOQueue;
+handle_req({Curr,get_unicode_state},Iport,_Oport,IOQueue) ->
+    Curr ! {self(),get_unicode_state,get_unicode_state(Iport)},
+    IOQueue;
+handle_req({Curr,set_unicode_state, Bool},Iport,_Oport,IOQueue) ->
+    Curr ! {self(),set_unicode_state,set_unicode_state(Iport,Bool)},
+    IOQueue;
+handle_req(next,Iport,Oport,IOQueue) ->
+    case queue:out(IOQueue) of
+        {{value,Next},ExecQ} ->
+            NewQ = handle_req(Next,Iport,Oport,queue:new()),
+            queue:join(NewQ,ExecQ);
+        {empty,_} ->
+            IOQueue
+    end;
+handle_req(Msg,Iport,Oport,IOQueue) ->
+    case queue:peek(IOQueue) of
+        empty ->
+            {Origin,Req} = Msg,
+            case io_request(Req, Iport, Oport) of
+                ok -> IOQueue;
+                Reply ->
+                    %% Push reply info to front of queue
+                    queue:in_r({Origin,Reply},IOQueue)
+            end;
+        _Else ->
+            %% All requests are queued when we have outstanding sync put_chars
+            queue:in(Msg,IOQueue)
     end.
 
 %% port_bytes(Bytes, InPort, OutPort, CurrentProcess, UserProcess, Group)
@@ -232,34 +265,34 @@ server_loop(Iport, Oport, Curr, User, Gr) ->
 %%  either escape to switch_loop or restart the shell. Otherwise send 
 %%  the bytes to Curr.
 
-port_bytes([$\^G|_Bs], Iport, Oport, _Curr, User, Gr) ->
-    handle_escape(Iport, Oport, User, Gr);
+port_bytes([$\^G|_Bs], Iport, Oport, _Curr, User, Gr, IOQueue) ->
+    handle_escape(Iport, Oport, User, Gr, IOQueue);
 
-port_bytes([$\^C|_Bs], Iport, Oport, Curr, User, Gr) ->
-    interrupt_shell(Iport, Oport, Curr, User, Gr);
+port_bytes([$\^C|_Bs], Iport, Oport, Curr, User, Gr, IOQueue) ->
+    interrupt_shell(Iport, Oport, Curr, User, Gr, IOQueue);
 
-port_bytes([B], Iport, Oport, Curr, User, Gr) ->
+port_bytes([B], Iport, Oport, Curr, User, Gr, IOQueue) ->
     Curr ! {self(),{data,[B]}},
-    server_loop(Iport, Oport, Curr, User, Gr);
-port_bytes(Bs, Iport, Oport, Curr, User, Gr) ->
+    server_loop(Iport, Oport, Curr, User, Gr, IOQueue);
+port_bytes(Bs, Iport, Oport, Curr, User, Gr, IOQueue) ->
     case member($\^G, Bs) of
 	true ->
-	    handle_escape(Iport, Oport, User, Gr);
+	    handle_escape(Iport, Oport, User, Gr, IOQueue);
 	false ->
 	    Curr ! {self(),{data,Bs}},
-	    server_loop(Iport, Oport, Curr, User, Gr)
+	    server_loop(Iport, Oport, Curr, User, Gr, IOQueue)
     end.
 
-interrupt_shell(Iport, Oport, Curr, User, Gr) ->
+interrupt_shell(Iport, Oport, Curr, User, Gr, IOQueue) ->
     case gr_get_info(Gr, Curr) of
 	undefined -> 
 	    ok;					% unknown
 	_ ->
 	    exit(Curr, interrupt)
     end,
-    server_loop(Iport, Oport, Curr, User, Gr).
+    server_loop(Iport, Oport, Curr, User, Gr, IOQueue).
 
-handle_escape(Iport, Oport, User, Gr) ->
+handle_escape(Iport, Oport, User, Gr, IOQueue) ->
     case application:get_env(stdlib, shell_esc) of
 	{ok,abort} ->
 	    Pid = gr_cur_pid(Gr),
@@ -278,11 +311,11 @@ handle_escape(Iport, Oport, User, Gr) ->
 	    Pid1 = group:start(self(), {shell,start,[]}),
 	    io_request({put_chars,unicode,"\n"}, Iport, Oport),
 	    server_loop(Iport, Oport, User, 
-			gr_add_cur(Gr1, Pid1, {shell,start,[]}));
+			gr_add_cur(Gr1, Pid1, {shell,start,[]}), IOQueue);
 
 	_ ->					% {ok,jcl} | undefined
 	    io_request({put_chars,unicode,"\nUser switch command\n"}, Iport, Oport),
-	    server_loop(Iport, Oport, User, switch_loop(Iport, Oport, Gr))
+	    server_loop(Iport, Oport, User, switch_loop(Iport, Oport, Gr), IOQueue)
     end.
 
 switch_loop(Iport, Oport, Gr) ->
@@ -492,9 +525,12 @@ set_unicode_state(Iport, Bool) ->
 
 io_request(Request, Iport, Oport) ->
     try io_command(Request) of
-        Command ->
+        {command,_} = Command ->
             Oport ! {self(),Command},
-            ok
+            ok;
+        {Command,Reply} ->
+            Oport ! {self(),Command},
+            Reply
     catch
         {requests,Rs} ->
             io_requests(Rs, Iport, Oport);
@@ -511,6 +547,13 @@ io_requests([], _Iport, _Oport) ->
 put_int16(N, Tail) ->
     [(N bsr 8)band 255,N band 255|Tail].
 
+%% When a put_chars_sync command is used, user_drv guarantees that
+%% the bytes have been put in the buffer of the port before an acknowledgement
+%% is sent back to the process sending the request. This command was added in
+%% OTP 18 to make sure that data sent from io:format is actually printed
+%% to the console before the vm stops when calling erlang:halt(integer()).
+io_command({put_chars_sync, unicode,Cs,Reply}) ->
+    {{command,[?OP_PUTC_SYNC|unicode:characters_to_binary(Cs,utf8)]},Reply};
 io_command({put_chars, unicode,Cs}) ->
     {command,[?OP_PUTC|unicode:characters_to_binary(Cs,utf8)]};
 io_command({move_rel,N}) ->
diff --git a/lib/ssh/src/ssh_cli.erl b/lib/ssh/src/ssh_cli.erl
index 18841e3d2d..de6d246403 100644
--- a/lib/ssh/src/ssh_cli.erl
+++ b/lib/ssh/src/ssh_cli.erl
@@ -98,7 +98,7 @@ handle_ssh_msg({ssh_cm, ConnectionHandler,
     Pty = Pty0#ssh_pty{width = Width, height = Height,
 		       pixel_width = PixWidth,
 		       pixel_height = PixHeight},
-    {Chars, NewBuf} = io_request({window_change, Pty0}, Buf, Pty),
+    {Chars, NewBuf} = io_request({window_change, Pty0}, Buf, Pty, undefined),
     write_chars(ConnectionHandler, ChannelId, Chars),
     {ok, State#state{pty = Pty, buf = NewBuf}};
 
@@ -188,7 +188,7 @@ handle_msg({Group, tty_geometry}, #state{group = Group,
 handle_msg({Group, Req}, #state{group = Group, buf = Buf, pty = Pty,
 				 cm = ConnectionHandler,
 				 channel = ChannelId} = State) ->
-    {Chars, NewBuf} = io_request(Req, Buf, Pty),
+    {Chars, NewBuf} = io_request(Req, Buf, Pty, Group),
     write_chars(ConnectionHandler, ChannelId, Chars),
     {ok, State#state{buf = NewBuf}};
 
@@ -263,40 +263,49 @@ eval(Error) ->
 %%% displaying device...
 %%% We are *not* really unicode aware yet, we just filter away characters 
 %%% beyond the latin1 range. We however handle the unicode binaries...
-io_request({window_change, OldTty}, Buf, Tty) ->
+io_request({window_change, OldTty}, Buf, Tty, _Group) ->
     window_change(Tty, OldTty, Buf);
-io_request({put_chars, Cs}, Buf, Tty) ->
+io_request({put_chars, Cs}, Buf, Tty, _Group) ->
     put_chars(bin_to_list(Cs), Buf, Tty);
-io_request({put_chars, unicode, Cs}, Buf, Tty) ->
+io_request({put_chars, unicode, Cs}, Buf, Tty, _Group) ->
     put_chars(unicode:characters_to_list(Cs,unicode), Buf, Tty);
-io_request({insert_chars, Cs}, Buf, Tty) ->
+io_request({insert_chars, Cs}, Buf, Tty, _Group) ->
     insert_chars(bin_to_list(Cs), Buf, Tty);
-io_request({insert_chars, unicode, Cs}, Buf, Tty) ->
+io_request({insert_chars, unicode, Cs}, Buf, Tty, _Group) ->
     insert_chars(unicode:characters_to_list(Cs,unicode), Buf, Tty);
-io_request({move_rel, N}, Buf, Tty) ->
+io_request({move_rel, N}, Buf, Tty, _Group) ->
     move_rel(N, Buf, Tty);
-io_request({delete_chars,N}, Buf, Tty) ->
+io_request({delete_chars,N}, Buf, Tty, _Group) ->
     delete_chars(N, Buf, Tty);
-io_request(beep, Buf, _Tty) ->
+io_request(beep, Buf, _Tty, _Group) ->
     {[7], Buf};
 
 %% New in R12
-io_request({get_geometry,columns},Buf,Tty) ->
+io_request({get_geometry,columns},Buf,Tty, _Group) ->
     {ok, Tty#ssh_pty.width, Buf};
-io_request({get_geometry,rows},Buf,Tty) ->
+io_request({get_geometry,rows},Buf,Tty, _Group) ->
     {ok, Tty#ssh_pty.height, Buf};
-io_request({requests,Rs}, Buf, Tty) ->
-    io_requests(Rs, Buf, Tty, []);
-io_request(tty_geometry, Buf, Tty) ->
-    io_requests([{move_rel, 0}, {put_chars, unicode, [10]}], Buf, Tty, []);
+io_request({requests,Rs}, Buf, Tty, Group) ->
+    io_requests(Rs, Buf, Tty, [], Group);
+io_request(tty_geometry, Buf, Tty, Group) ->
+    io_requests([{move_rel, 0}, {put_chars, unicode, [10]}],
+                Buf, Tty, [], Group);
      %{[], Buf};
-io_request(_R, Buf, _Tty) ->
+
+%% New in 18
+io_request({put_chars_sync, Class, Cs, Reply}, Buf, Tty, Group) ->
+    %% We handle these asynchronous for now, if we need output guarantees
+    %% we have to handle these synchronously
+    Group ! {reply, Reply},
+    io_request({put_chars, Class, Cs}, Buf, Tty, Group);
+
+io_request(_R, Buf, _Tty, _Group) ->
     {[], Buf}.
 
-io_requests([R|Rs], Buf, Tty, Acc) ->
-    {Chars, NewBuf} = io_request(R, Buf, Tty),
-    io_requests(Rs, NewBuf, Tty, [Acc|Chars]);
-io_requests([], Buf, _Tty, Acc) ->
+io_requests([R|Rs], Buf, Tty, Acc, Group) ->
+    {Chars, NewBuf} = io_request(R, Buf, Tty, Group),
+    io_requests(Rs, NewBuf, Tty, [Acc|Chars], Group);
+io_requests([], Buf, _Tty, Acc, _Group) ->
     {Acc, Buf}.
 
 %%% return commands for cursor navigation, assume everything is ansi