Merge branch 'master' into sverk/master/halt-INT_MIN

8 files changed, 2772 insertions, 2044 deletions

diff --git a/erts/emulator/sys/unix/erl_child_setup.c b/erts/emulator/sys/unix/erl_child_setup.c
index a3c5c20641..4e61530cf1 100644
--- a/erts/emulator/sys/unix/erl_child_setup.c
+++ b/erts/emulator/sys/unix/erl_child_setup.c
@@ -1,7 +1,7 @@
 /*
  * %CopyrightBegin%
  * 
- * Copyright Ericsson AB 2002-2009. All Rights Reserved.
+ * Copyright Ericsson AB 2002-2015. All Rights Reserved.
  * 
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -19,109 +19,233 @@
  */
 
 /*
- * After a vfork() (or fork()) the child exec()s to this program which
- * sets up the child and exec()s to the user program (see spawn_start()
- * in sys.c and ticket OTP-4389).
+ * This program is started at erts startup and all fork's that
+ * have to be done are done in here. This is done for a couple
+ * of reasons:
+ *  - Allow usage of fork without a memory explosion.
+ *  -- we do not want to use vfork, as it blocks the VM
+ *     until the execv is done, and if the program that
+ *     is to be executed is on an NFS that is unavailable,
+ *     the execv can block for a very long time.
+ *  -- we cannot do fork inside the VM as that would temporarily
+ *     duplicate the memory usage of the VM per parallel exec.
+ *
+ * Some implementation notes:
+ *  - A single Unix Domain Socket is setup in between the VM and
+ *    this program. Over that UDS the file descriptors that should
+ *    be used to talk to the child program are sent.
+ *    The actual command to execute, together with options and the
+ *    environment, is sent over the pipe represented by the
+ *    file descriptors mentioned above. We don't send the
+ *    command over the UDS as that would increase the likely hood
+ *    that it's buffer would be full.
+ *
+ *  - Since it is this program that execv's, it has to take care of
+ *    all the SIGCHLD signals that the child programs generate. The
+ *    signals are received and the pid+exit reason is sent as data
+ *    on the UDS to the VM. The VM is then able to map the pid to the
+ *    port of the child program that just exited and deliver the status
+ *    code if requested.
  */
 
 #ifdef HAVE_CONFIG_H
 #  include "config.h"
 #endif
 
-#define NEED_CHILD_SETUP_DEFINES
-#include "sys.h"
-#include "erl_misc_utils.h"
+#include <stdlib.h>
+#include <stdio.h>
+#include <sys/wait.h>
 
-#ifdef SIG_SIGSET		/* Old SysV */
-void sys_sigrelease(int sig)
+#define WANT_NONBLOCKING
+
+#include "erl_driver.h"
+#include "sys_uds.h"
+#include "hash.h"
+#include "erl_child_setup.h"
+
+#define SET_CLOEXEC(fd) fcntl(fd, F_SETFD, fcntl(fd, F_GETFD) | FD_CLOEXEC)
+
+#if defined(__ANDROID__)
+#define SHELL "/system/bin/sh"
+#else
+#define SHELL "/bin/sh"
+#endif /* __ANDROID__ */
+
+//#define HARD_DEBUG
+#ifdef HARD_DEBUG
+#define DEBUG_PRINT(fmt, ...) fprintf(stderr, fmt "\r\n", ##__VA_ARGS__)
+#else
+#define DEBUG_PRINT(fmt, ...)
+#endif
+
+#define ABORT(fmt, ...) do {                                            \
+        fprintf(stderr, "erl_child_setup: " fmt "\r\n", ##__VA_ARGS__); \
+        abort();                                                        \
+    } while(0)
+
+#ifdef DEBUG
+void
+erl_assert_error(const char* expr, const char* func, const char* file, int line)
 {
-    sigrelse(sig);
+    fflush(stdout);
+    fprintf(stderr, "%s:%d:%s() Assertion failed: %s\n",
+            file, line, func, expr);
+    fflush(stderr);
+    abort();
 }
-#else /* !SIG_SIGSET */
-#ifdef SIG_SIGNAL		/* Old BSD */
-sys_sigrelease(int sig)
+#endif
+
+void sys_sigblock(int sig)
 {
-    sigsetmask(sigblock(0) & ~sigmask(sig));
+    sigset_t mask;
+
+    sigemptyset(&mask);
+    sigaddset(&mask, sig);
+    sigprocmask(SIG_BLOCK, &mask, (sigset_t *)NULL);
 }
-#else /* !SIG_SIGNAL */	/* The True Way - POSIX!:-) */
+
 void sys_sigrelease(int sig)
 {
     sigset_t mask;
-
     sigemptyset(&mask);
     sigaddset(&mask, sig);
     sigprocmask(SIG_UNBLOCK, &mask, (sigset_t *)NULL);
 }
-#endif /* !SIG_SIGNAL */
-#endif /* !SIG_SIGSET */
-
-#if defined(__ANDROID__)
-static int system_properties_fd(void);
-#endif /* __ANDROID__ */
 
-#if defined(__ANDROID__)
-#define SHELL "/system/bin/sh"
-#else
-#define SHELL "/bin/sh"
-#endif /* __ANDROID__ */
+static void add_os_pid_to_port_id_mapping(Eterm, pid_t);
+static Eterm get_port_id(pid_t);
+static int forker_hash_init(void);
 
+static int max_files = -1;
+static int sigchld_pipe[2];
 
-int
-main(int argc, char *argv[])
+static int
+start_new_child(int pipes[])
 {
-    int i, from, to;
-    int erts_spawn_executable = 0;
+    int size, res, i, pos = 0;
+    char *buff, *o_buff;
+
+    char *cmd, *wd, **new_environ, **args = NULL;
+
+    Sint cnt, flags;
 
-    /* OBSERVE!
-     * Keep child setup after fork() (implemented in sys.c) up to date
-     * if changes are made here.
-     */
+    /* only child executes here */
 
-    if (argc != CS_ARGV_NO_OF_ARGS) {
-	if (argc < CS_ARGV_NO_OF_ARGS) {
-	    return 1;
-	} else {
-	    erts_spawn_executable = 1;
-	}
+    do {
+        res = read(pipes[0], (char*)&size, sizeof(size));
+    } while(res < 0 && (errno == EINTR || errno == ERRNO_BLOCK));
+
+    if (res <= 0) {
+        goto child_error;
     }
 
-    if (strcmp("false", argv[CS_ARGV_UNBIND_IX]) != 0)
-	if (erts_unbind_from_cpu_str(argv[CS_ARGV_UNBIND_IX]) != 0)
-	    return 1;
+    buff = malloc(size);
+
+    DEBUG_PRINT("size = %d", size);
+
+    do {
+        if ((res = read(pipes[0], buff + pos, size - pos)) < 0) {
+            if (errno == ERRNO_BLOCK || errno == EINTR)
+                continue;
+            goto child_error;
+        }
+        if (res == 0) {
+            errno = EPIPE;
+            goto child_error;
+        }
+        pos += res;
+    } while(size - pos != 0);
+
+    o_buff = buff;
+
+    flags = get_int32(buff);
+    buff += sizeof(Sint32);
 
-    for (i = 0; i < CS_ARGV_NO_OF_DUP2_OPS; i++) {
-	if (argv[CS_ARGV_DUP2_OP_IX(i)][0] == '-'
-	    && argv[CS_ARGV_DUP2_OP_IX(i)][1] == '\0')
-	    break;
-	if (sscanf(argv[CS_ARGV_DUP2_OP_IX(i)], "%d:%d", &from, &to) != 2)
-	    return 1;
-	if (dup2(from, to) < 0)
-	    return 1;
+    DEBUG_PRINT("flags = %d", flags);
+
+    cmd = buff;
+    buff += strlen(buff) + 1;
+    if (*buff == '\0') {
+        wd = NULL;
+    } else {
+        wd = buff;
+        buff += strlen(buff) + 1;
     }
+    buff++;
 
-    if (sscanf(argv[CS_ARGV_FD_CR_IX], "%d:%d", &from, &to) != 2)
-	return 1;
+    DEBUG_PRINT("wd = %s", wd);
 
-#if defined(HAVE_CLOSEFROM)
-    closefrom(from);
-#elif defined(__ANDROID__)
-    if (from <= to) {
-	int spfd = system_properties_fd();
-	for (i = from; i <= to; i++) {
-	    if (i != spfd) {
-		(void) close(i);
-            }
-	}
+    cnt = get_int32(buff);
+    buff += sizeof(Sint32);
+    new_environ = malloc(sizeof(char*)*(cnt + 1));
+
+    DEBUG_PRINT("env_len = %ld", cnt);
+    for (i = 0; i < cnt; i++, buff++) {
+        new_environ[i] = buff;
+        while(*buff != '\0') buff++;
     }
-#else  /* !__ANDROID__ */
-    for (i = from; i <= to; i++) {
-	(void) close(i);
+    new_environ[cnt] = NULL;
+
+    if (o_buff + size != buff) {
+        /* This is a spawn executable call */
+        cnt = get_int32(buff);
+        buff += sizeof(Sint32);
+        args = malloc(sizeof(char*)*(cnt + 1));
+        for (i = 0; i < cnt; i++, buff++) {
+            args[i] = buff;
+            while(*buff != '\0') buff++;
+        }
+        args[cnt] = NULL;
     }
-#endif /* HAVE_CLOSEFROM */
 
-    if (!(argv[CS_ARGV_WD_IX][0] == '.' && argv[CS_ARGV_WD_IX][1] == '\0')
-	&& chdir(argv[CS_ARGV_WD_IX]) < 0)
-	return 1;
+    if (o_buff + size != buff) {
+        errno = EINVAL;
+        goto child_error;
+    }
+
+    DEBUG_PRINT("read ack");
+    do {
+        ErtsSysForkerProto proto;
+        res = read(pipes[0], &proto, sizeof(proto));
+        if (res > 0) {
+            ASSERT(proto.action == ErtsSysForkerProtoAction_Ack);
+            ASSERT(res == sizeof(proto));
+        }
+    } while(res < 0 && (errno == EINTR || errno == ERRNO_BLOCK));
+    if (res < 1) {
+        errno = EPIPE;
+        goto child_error;
+    }
+
+    DEBUG_PRINT("Do that forking business: '%s'\n",cmd);
+
+    /* When the dup2'ing below is done, only
+       fd's 0, 1, 2 and maybe 3, 4 should survive the
+       exec. All other fds (i.e. the unix domain sockets
+       and stray pipe ends) should have CLOEXEC set on them
+       so they will be closed when the exec happens */
+    if (flags & FORKER_FLAG_USE_STDIO) {
+        /* stdin for process */
+        if (flags & FORKER_FLAG_DO_WRITE &&
+            dup2(pipes[0], 0) < 0)
+            goto child_error;
+        /* stdout for process */
+        if (flags & FORKER_FLAG_DO_READ &&
+            dup2(pipes[1], 1) < 0)
+            goto child_error;
+    }
+    else {	/* XXX will fail if pipes[0] == 4 (unlikely..) */
+        if (flags & FORKER_FLAG_DO_READ && dup2(pipes[1], 4) < 0)
+            goto child_error;
+        if (flags & FORKER_FLAG_DO_WRITE && dup2(pipes[0], 3) < 0)
+            goto child_error;
+    }
+
+    if (dup2(pipes[2], 2) < 0)
+        goto child_error;
+
+    if (wd && chdir(wd) < 0)
+        goto child_error;
 
 #if defined(USE_SETPGRP_NOARGS)		/* SysV */
     (void) setpgrp();
@@ -131,34 +255,301 @@ main(int argc, char *argv[])
     (void) setsid();
 #endif
 
+    close(pipes[0]);
+    close(pipes[1]);
+    close(pipes[2]);
+
     sys_sigrelease(SIGCHLD);
-    sys_sigrelease(SIGINT);
-    sys_sigrelease(SIGUSR1);
-
-    if (erts_spawn_executable) {
-	if (argv[CS_ARGV_NO_OF_ARGS + 1] == NULL) {
-	    execl(argv[CS_ARGV_NO_OF_ARGS],argv[CS_ARGV_NO_OF_ARGS],
-		  (char *) NULL);
-	} else {
-	    execv(argv[CS_ARGV_NO_OF_ARGS],&(argv[CS_ARGV_NO_OF_ARGS + 1]));
-	}
+
+    if (args) {
+        /* spawn_executable */
+        execve(cmd, args, new_environ);
     } else {
-        execl(SHELL, "sh", "-c", argv[CS_ARGV_CMD_IX], (char *) NULL);
+        execle(SHELL, "sh", "-c", cmd, (char *) NULL, new_environ);
     }
-    return 1;
+child_error:
+    DEBUG_PRINT("exec error: %d\r\n",errno);
+    _exit(128 + errno);
+}
+
+
+/*
+ * [OTP-3906]
+ * Solaris signal management gets confused when threads are used and a
+ * lot of child processes dies. The confusion results in that SIGCHLD
+ * signals aren't delivered to the emulator which in turn results in
+ * a lot of defunct processes in the system.
+ *
+ * The problem seems to appear when a signal is frequently
+ * blocked/unblocked at the same time as the signal is frequently
+ * propagated. The child waiter thread is a workaround for this problem.
+ * The SIGCHLD signal is always blocked (in all threads), and the child
+ * waiter thread fetches the signal by a call to sigwait(). See
+ * child_waiter().
+ *
+ * This should be a non-issue since the fork:ing was moved outside of
+ * the emulator into erl_child_setup. I'm leaving the comment here
+ * for posterity. */
+
+static void handle_sigchld(int sig) {
+    int buff[2], res;
+
+    sys_sigblock(SIGCHLD);
+
+    while ((buff[0] = waitpid((pid_t)(-1), buff+1, WNOHANG)) > 0) {
+        do {
+            res = write(sigchld_pipe[1], buff, sizeof(buff));
+        } while (res < 0 && errno == EINTR);
+        if (res <= 0)
+            ABORT("Failed to write to sigchld_pipe (%d): %d (%d)", sigchld_pipe[1], res, errno);
+        DEBUG_PRINT("Reap child %d (%d)", buff[0], buff[1]);
+    }
+
+    sys_sigrelease(SIGCHLD);
 }
 
 #if defined(__ANDROID__)
 static int system_properties_fd(void)
 {
-    int fd;
+    static int fd = -2;
     char *env;
 
+    if (fd != -2) return fd;
     env = getenv("ANDROID_PROPERTY_WORKSPACE");
     if (!env) {
+        fd = -1;
         return -1;
     }
     fd = atoi(env);
     return fd;
 }
 #endif /* __ANDROID__ */
+
+int
+main(int argc, char *argv[])
+{
+    /* This fd should be open from beam */
+    int uds_fd = 3, max_fd = 3;
+#ifndef HAVE_CLOSEFROM
+    int i;
+#endif
+    struct sigaction sa;
+
+    if (argc < 1 || sscanf(argv[1],"%d",&max_files) != 1) {
+        ABORT("Invalid arguments to child_setup");
+    }
+
+/* We close all fds except the uds from beam.
+   All other fds from now on will have the
+   CLOEXEC flags set on them. This means that we
+   only have to close a very limited number of fds
+   after we fork before the exec. */
+#if defined(HAVE_CLOSEFROM)
+    closefrom(4);
+#else
+    for (i = 4; i < max_files; i++)
+#if defined(__ANDROID__)
+        if (i != system_properties_fd())
+#endif
+        (void) close(i);
+#endif
+
+    if (pipe(sigchld_pipe) < 0) {
+        ABORT("Failed to setup sigchld pipe (%d)", errno);
+    }
+
+    SET_CLOEXEC(sigchld_pipe[0]);
+    SET_CLOEXEC(sigchld_pipe[1]);
+
+    max_fd = max_fd < sigchld_pipe[0] ? sigchld_pipe[0] : max_fd;
+
+    sa.sa_handler = &handle_sigchld;
+    sigemptyset(&sa.sa_mask);
+    sa.sa_flags = SA_RESTART | SA_NOCLDSTOP;
+    if (sigaction(SIGCHLD, &sa, 0) == -1) {
+        perror(0);
+        exit(1);
+    }
+
+    forker_hash_init();
+
+    SET_CLOEXEC(uds_fd);
+
+    DEBUG_PRINT("Starting forker %d", max_files);
+
+    while (1) {
+        fd_set read_fds;
+        int res;
+        FD_ZERO(&read_fds);
+        FD_SET(uds_fd, &read_fds);
+        FD_SET(sigchld_pipe[0], &read_fds);
+        DEBUG_PRINT("child_setup selecting on %d, %d (%d)",
+                uds_fd, sigchld_pipe[0], max_fd);
+        res = select(max_fd+1, &read_fds, NULL, NULL, NULL);
+
+        if (res < 0) {
+            if (errno == EINTR) continue;
+            ABORT("Select failed: %d (%d)",res, errno);
+        }
+
+        if (FD_ISSET(uds_fd, &read_fds)) {
+            int pipes[3], res, os_pid;
+            ErtsSysForkerProto proto;
+            errno = 0;
+            if ((res = sys_uds_read(uds_fd, (char*)&proto, sizeof(proto),
+                                    pipes, 3, MSG_DONTWAIT)) < 0) {
+                if (errno == EINTR)
+                    continue;
+                DEBUG_PRINT("erl_child_setup failed to read from uds: %d, %d", res, errno);
+                _exit(0);
+            }
+
+            if (res == 0) {
+                DEBUG_PRINT("uds was closed!");
+                _exit(0);
+            }
+            /* Since we use unix domain sockets and send the entire data in
+               one go we *should* get the entire payload at once. */
+            ASSERT(res == sizeof(proto));
+            ASSERT(proto.action == ErtsSysForkerProtoAction_Start);
+
+            sys_sigblock(SIGCHLD);
+
+            errno = 0;
+
+            os_pid = fork();
+            if (os_pid == 0)
+                start_new_child(pipes);
+
+            add_os_pid_to_port_id_mapping(proto.u.start.port_id, os_pid);
+
+            /* We write an ack here, but expect the reply on
+               the pipes[0] inside the fork */
+            proto.action = ErtsSysForkerProtoAction_Go;
+            proto.u.go.os_pid = os_pid;
+            proto.u.go.error_number = errno;
+            while (write(pipes[1], &proto, sizeof(proto)) < 0 && errno == EINTR)
+                ; /* remove gcc warning */
+
+#ifdef FORKER_PROTO_START_ACK
+            proto.action = ErtsSysForkerProtoAction_StartAck;
+            while (write(uds_fd, &proto, sizeof(proto)) < 0 && errno == EINTR)
+                ; /* remove gcc warning */
+#endif
+
+            sys_sigrelease(SIGCHLD);
+            close(pipes[0]);
+            close(pipes[1]);
+            close(pipes[2]);
+        }
+
+        if (FD_ISSET(sigchld_pipe[0], &read_fds)) {
+            int ibuff[2];
+            ErtsSysForkerProto proto;
+            res = read(sigchld_pipe[0], ibuff, sizeof(ibuff));
+            if (res <= 0) {
+                if (errno == EINTR)
+                    continue;
+                ABORT("Failed to read from sigchld pipe: %d (%d)", res, errno);
+            }
+
+            proto.u.sigchld.port_id = get_port_id((pid_t)(ibuff[0]));
+
+            if (proto.u.sigchld.port_id == THE_NON_VALUE)
+                continue; /* exit status report not requested */
+
+            proto.action = ErtsSysForkerProtoAction_SigChld;
+            proto.u.sigchld.error_number = ibuff[1];
+            DEBUG_PRINT("send %s to %d", buff, uds_fd);
+            if (write(uds_fd, &proto, sizeof(proto)) < 0) {
+                if (errno == EINTR)
+                    continue;
+                /* The uds was close, which most likely means that the VM
+                   has exited. This will be detected when we try to read
+                   from the uds_fd. */
+                DEBUG_PRINT("Failed to write to uds: %d (%d)", uds_fd, errno);
+            }
+        }
+    }
+    return 1;
+}
+
+typedef struct exit_status {
+    HashBucket hb;
+    pid_t os_pid;
+    Eterm port_id;
+} ErtsSysExitStatus;
+
+static Hash *forker_hash;
+
+static void add_os_pid_to_port_id_mapping(Eterm port_id, pid_t os_pid)
+{
+    if (port_id != THE_NON_VALUE) {
+        /* exit status report requested */
+        ErtsSysExitStatus es;
+        es.os_pid = os_pid;
+        es.port_id = port_id;
+        hash_put(forker_hash, &es);
+    }
+}
+
+static Eterm get_port_id(pid_t os_pid)
+{
+    ErtsSysExitStatus est, *es;
+    Eterm port_id;
+    est.os_pid = os_pid;
+    es = hash_remove(forker_hash, &est);
+    if (!es) return THE_NON_VALUE;
+    port_id = es->port_id;
+    free(es);
+    return port_id;
+}
+
+static int fcmp(void *a, void *b)
+{
+    ErtsSysExitStatus *sa = a;
+    ErtsSysExitStatus *sb = b;
+    return !(sa->os_pid == sb->os_pid);
+}
+
+static HashValue fhash(void *e)
+{
+    ErtsSysExitStatus *se = e;
+    Uint32 val = se->os_pid;
+    val = (val+0x7ed55d16) + (val<<12);
+    val = (val^0xc761c23c) ^ (val>>19);
+    val = (val+0x165667b1) + (val<<5);
+    val = (val+0xd3a2646c) ^ (val<<9);
+    val = (val+0xfd7046c5) + (val<<3);
+    val = (val^0xb55a4f09) ^ (val>>16);
+    return val;
+}
+
+static void *falloc(void *e)
+{
+    ErtsSysExitStatus *se = e;
+    ErtsSysExitStatus *ne = malloc(sizeof(ErtsSysExitStatus));
+    ne->os_pid = se->os_pid;
+    ne->port_id = se->port_id;
+    return ne;
+}
+
+static void *meta_alloc(int type, size_t size) { return malloc(size); }
+static void meta_free(int type, void *p)       { free(p); }
+
+static int forker_hash_init(void)
+{
+    HashFunctions forker_hash_functions;
+    forker_hash_functions.hash = fhash;
+    forker_hash_functions.cmp = fcmp;
+    forker_hash_functions.alloc = falloc;
+    forker_hash_functions.free = free;
+    forker_hash_functions.meta_alloc = meta_alloc;
+    forker_hash_functions.meta_free  = meta_free;
+    forker_hash_functions.meta_print = NULL;
+
+    forker_hash = hash_new(0, "forker_hash",
+                           16, forker_hash_functions);
+
+    return 1;
+}
diff --git a/erts/emulator/sys/unix/erl_child_setup.h b/erts/emulator/sys/unix/erl_child_setup.h
new file mode 100644
index 0000000000..a28b136bfc
--- /dev/null
+++ b/erts/emulator/sys/unix/erl_child_setup.h
@@ -0,0 +1,77 @@
+/*
+ * %CopyrightBegin%
+ * 
+ * Copyright Ericsson AB 2015-2015. All Rights Reserved.
+ * 
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * 
+ * %CopyrightEnd%
+ *
+ * This file defines the interface inbetween erts and child_setup.
+ */
+
+#ifndef _ERL_UNIX_FORKER_H
+#define _ERL_UNIX_FORKER_H
+
+#include "sys.h"
+
+#ifdef __FreeBSD__
+/* The freebsd sendmsg man page explicitly states that
+   you should not close fds before they are known
+   to have reached the other side, so this Ack protects
+   against that. */
+#define FORKER_PROTO_START_ACK 1
+#endif
+
+#define FORKER_ARGV_NO_OF_ARGS  3
+#define FORKER_ARGV_PROGNAME_IX	0    /* Program name                          */
+#define FORKER_ARGV_MAX_FILES	1    /* max_files                             */
+
+#define FORKER_FLAG_USE_STDIO   (1 << 0)    /* dup the pipe to stdin/stderr   */
+#define FORKER_FLAG_EXIT_STATUS (1 << 1)    /* send the exit status to parent */
+#define FORKER_FLAG_DO_READ     (1 << 2)    /* dup write fd */
+#define FORKER_FLAG_DO_WRITE    (1 << 3)    /* dup read  fd  */
+
+#if SIZEOF_VOID_P == SIZEOF_LONG
+typedef unsigned long ErtsSysPortId;
+#elif SIZEOF_VOID_P == SIZEOF_INT
+typedef unsigned int ErtsSysPortId;
+#elif SIZEOF_VOID_P == SIZEOF_LONG_LONG
+typedef unsigned long long ErtsSysPortId;
+#endif
+
+typedef struct ErtsSysForkerProto_ {
+    enum {
+        ErtsSysForkerProtoAction_Start,
+        ErtsSysForkerProtoAction_StartAck,
+        ErtsSysForkerProtoAction_Go,
+        ErtsSysForkerProtoAction_SigChld,
+        ErtsSysForkerProtoAction_Ack
+    } action;
+    union {
+        struct {
+            ErtsSysPortId port_id;
+            int fds[3];
+        } start;
+        struct {
+            pid_t os_pid;
+            int error_number;
+        } go;
+        struct {
+            ErtsSysPortId port_id;
+            int error_number;
+        } sigchld;
+    } u;
+} ErtsSysForkerProto;
+
+#endif /* #ifndef _ERL_UNIX_FORKER_H */
diff --git a/erts/emulator/sys/unix/erl_unix_sys.h b/erts/emulator/sys/unix/erl_unix_sys.h
index 8d4e98bf3a..8b1822ca9f 100644
--- a/erts/emulator/sys/unix/erl_unix_sys.h
+++ b/erts/emulator/sys/unix/erl_unix_sys.h
@@ -30,9 +30,7 @@
 #include <limits.h>
 #include <stdlib.h>
 #include <string.h>
-#ifndef QNX
 #include <memory.h>
-#endif
 
 #if defined(__sun__) && defined(__SVR4) && !defined(__EXTENSIONS__)
 #   define __EXTENSIONS__
@@ -92,11 +90,6 @@
 #include <ieeefp.h>
 #endif
 
-#ifdef QNX
-#include <process.h>
-#include <sys/qnx_glob.h>
-#endif
-
 #include <pwd.h>
 
 #ifndef HZ
@@ -108,6 +101,10 @@
 #endif
 #include <netdb.h>
 
+#ifdef HAVE_MACH_ABSOLUTE_TIME
+#include <mach/mach_time.h>
+#endif
+
 #ifdef HAVE_POSIX_MEMALIGN
 #  define ERTS_HAVE_ERTS_SYS_ALIGNED_ALLOC 1
 #endif
@@ -136,13 +133,6 @@
 #  define ERTS_POLL_NEED_ASYNC_INTERRUPT_SUPPORT
 #endif
 
-#ifndef ENABLE_CHILD_WAITER_THREAD
-#  ifdef ERTS_SMP
-#    define ERTS_SMP_SCHEDULERS_NEED_TO_CHECK_CHILDREN
-void erts_check_children(void);
-#  endif
-#endif
-
 typedef void *GETENV_STATE;
 
 /*
@@ -171,6 +161,7 @@ typedef long long ErtsSysHrTime;
 #endif
 
 typedef ErtsMonotonicTime ErtsSystemTime;
+typedef ErtsSysHrTime ErtsSysPerfCounter;
 
 #define ERTS_MONOTONIC_TIME_MIN (((ErtsMonotonicTime) 1) << 63)
 #define ERTS_MONOTONIC_TIME_MAX (~ERTS_MONOTONIC_TIME_MIN)
@@ -219,6 +210,7 @@ ErtsSystemTime erts_os_system_time(void);
  * It may or may not be monotonic.
  */
 ErtsSysHrTime erts_sys_hrtime(void);
+#define ERTS_HRTIME_UNIT (1000*1000*1000)
 
 struct erts_sys_time_read_only_data__ {
 #ifdef ERTS_OS_MONOTONIC_INLINE_FUNC_PTR_CALL__
@@ -227,6 +219,8 @@ struct erts_sys_time_read_only_data__ {
 #ifdef ERTS_OS_TIMES_INLINE_FUNC_PTR_CALL__
     void (*os_times)(ErtsMonotonicTime *, ErtsSystemTime *);
 #endif
+    ErtsSysPerfCounter (*perf_counter)(void);
+    ErtsSysPerfCounter perf_counter_unit;
     int ticks_per_sec;
 };
 
@@ -280,7 +274,24 @@ erts_os_times(ErtsMonotonicTime *mtimep, ErtsSystemTime *stimep)
 #endif /* ERTS_HAVE_OS_MONOTONIC_TIME_SUPPORT */
 
 /*
- *
+ * Functions for getting the performance counter
+ */
+
+ERTS_GLB_INLINE ErtsSysPerfCounter erts_sys_perf_counter(void);
+#define erts_sys_perf_counter_unit() erts_sys_time_data__.r.o.perf_counter_unit
+
+#if ERTS_GLB_INLINE_INCL_FUNC_DEF
+
+ERTS_GLB_INLINE ErtsSysPerfCounter
+erts_sys_perf_counter()
+{
+    return (*erts_sys_time_data__.r.o.perf_counter)();
+}
+
+#endif /* ERTS_GLB_INLINE_INCL_FUNC_DEF */
+
+/*
+ * Functions for measuring CPU time
  */
 
 #if (defined(HAVE_GETHRVTIME) || defined(HAVE_CLOCK_GETTIME_CPU_TIME))
@@ -310,7 +321,6 @@ typedef void (*SIGFUNC)(int);
 extern SIGFUNC sys_signal(int, SIGFUNC);
 extern void sys_sigrelease(int);
 extern void sys_sigblock(int);
-extern void sys_stop_cat(void);
 
 /*
  * Handling of floating point exceptions.
@@ -425,19 +435,6 @@ void erts_sys_unblock_fpe(int);
 #define ERTS_FP_ERROR_THOROUGH(p, f, A)	__ERTS_FP_ERROR_THOROUGH(&(p)->fp_exception, f, A)
 
 
-#ifdef NEED_CHILD_SETUP_DEFINES
-/* The child setup argv[] */
-#define CS_ARGV_PROGNAME_IX	0		/* Program name		*/
-#define CS_ARGV_UNBIND_IX	1		/* Unbind from cpu	*/
-#define CS_ARGV_WD_IX		2		/* Working directory	*/
-#define CS_ARGV_CMD_IX		3		/* Command		*/
-#define CS_ARGV_FD_CR_IX	4		/* Fd close range	*/
-#define CS_ARGV_DUP2_OP_IX(N)	((N) + 5)	/* dup2 operations	*/
-
-#define CS_ARGV_NO_OF_DUP2_OPS	3		/* Number of dup2 ops	*/
-#define CS_ARGV_NO_OF_ARGS	8		/* Number of arguments	*/
-#endif /* #ifdef NEED_CHILD_SETUP_DEFINES */
-
 /* Threads */
 #ifdef USE_THREADS
 extern int init_async(int);
diff --git a/erts/emulator/sys/unix/sys.c b/erts/emulator/sys/unix/sys.c
index d94b37430e..0a18e54389 100644
--- a/erts/emulator/sys/unix/sys.c
+++ b/erts/emulator/sys/unix/sys.c
@@ -49,7 +49,6 @@
 #include <sys/ioctl.h>
 #endif
 
-#define NEED_CHILD_SETUP_DEFINES
 #define ERTS_WANT_BREAK_HANDLING
 #define ERTS_WANT_GOT_SIGUSR1
 #define WANT_NONBLOCKING    /* must define this to pull in defs from sys.h */
@@ -67,7 +66,7 @@
 #include "erl_mseg.h"
 
 extern char **environ;
-static erts_smp_rwmtx_t environ_rwmtx;
+erts_smp_rwmtx_t environ_rwmtx;
 
 #define MAX_VSIZE 16		/* Max number of entries allowed in an I/O
 				 * vector sock_sendv().
@@ -76,89 +75,12 @@ static erts_smp_rwmtx_t environ_rwmtx;
  * Don't need global.h, but bif_table.h (included by bif.h),
  * won't compile otherwise
  */
-#include "global.h" 
+#include "global.h"
 #include "bif.h"
 
-#include "erl_sys_driver.h"
 #include "erl_check_io.h"
 #include "erl_cpu_topology.h"
 
-#ifndef DISABLE_VFORK
-#define DISABLE_VFORK 0
-#endif
-
-#if defined IOV_MAX
-#define MAXIOV IOV_MAX
-#elif defined UIO_MAXIOV
-#define MAXIOV UIO_MAXIOV
-#else
-#define MAXIOV 16
-#endif
-
-#ifdef USE_THREADS
-#  ifdef ENABLE_CHILD_WAITER_THREAD
-#    define CHLDWTHR ENABLE_CHILD_WAITER_THREAD
-#  else
-#    define CHLDWTHR 0
-#  endif
-#  define FDBLOCK 1
-#else
-#  define CHLDWTHR 0
-#  define FDBLOCK 0
-#endif
-/*
- * [OTP-3906]
- * Solaris signal management gets confused when threads are used and a
- * lot of child processes dies. The confusion results in that SIGCHLD
- * signals aren't delivered to the emulator which in turn results in
- * a lot of defunct processes in the system.
- *
- * The problem seems to appear when a signal is frequently
- * blocked/unblocked at the same time as the signal is frequently
- * propagated. The child waiter thread is a workaround for this problem.
- * The SIGCHLD signal is always blocked (in all threads), and the child
- * waiter thread fetches the signal by a call to sigwait(). See
- * child_waiter().
- */
-
-typedef struct ErtsSysReportExit_ ErtsSysReportExit;
-struct ErtsSysReportExit_ {
-    ErtsSysReportExit *next;
-    Eterm port;
-    int pid;
-    int ifd;
-    int ofd;
-#if CHLDWTHR && !defined(ERTS_SMP)
-    int status;
-#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;
-    int ofd, packet_bytes;
-    ErtsSysReportExit *report_exit;
-    int pid;
-    int alive;
-    int status;
-    int terminating;
-    ErtsSysBlocking *blocking;
-} *driver_data;			/* indexed by fd */
-
-static ErtsSysReportExit *report_exit_list;
-#if CHLDWTHR && !defined(ERTS_SMP)
-static ErtsSysReportExit *report_exit_transit_list;
-#endif
-
 extern int  driver_interrupt(int, int);
 extern void do_break(void);
 
@@ -170,33 +92,6 @@ extern void erts_sys_init_float(void);
 
 extern void erl_crash_dump(char* file, int line, char* fmt, ...);
 
-#define DIR_SEPARATOR_CHAR    '/'
-
-#if defined(__ANDROID__)
-#define SHELL "/system/bin/sh"
-#else
-#define SHELL "/bin/sh"
-#endif /* __ANDROID__ */
-
-
-#if defined(DEBUG)
-#define ERL_BUILD_TYPE_MARKER ".debug"
-#elif defined(PURIFY)
-#define ERL_BUILD_TYPE_MARKER ".purify"
-#elif defined(QUANTIFY)
-#define ERL_BUILD_TYPE_MARKER ".quantify"
-#elif defined(PURECOV)
-#define ERL_BUILD_TYPE_MARKER ".purecov"
-#elif defined(VALGRIND)
-#define ERL_BUILD_TYPE_MARKER ".valgrind"
-#else /* opt */
-#define ERL_BUILD_TYPE_MARKER
-#endif
-
-#define CHILD_SETUP_PROG_NAME	"child_setup" ERL_BUILD_TYPE_MARKER
-#if !DISABLE_VFORK
-static char *child_setup_prog;
-#endif
 
 #ifdef DEBUG
 static int debug_log = 0;
@@ -220,7 +115,7 @@ static volatile int have_prepared_crash_dump;
   (have_prepared_crash_dump++)
 #endif
 
-static erts_smp_atomic_t sys_misc_mem_sz;
+erts_smp_atomic_t sys_misc_mem_sz;
 
 #if defined(ERTS_SMP)
 static void smp_sig_notify(char c);
@@ -233,46 +128,6 @@ static int sig_suspend_fds[2] = {-1, -1};
 
 jmp_buf erts_sys_sigsegv_jmp;
 
-#if CHLDWTHR || defined(ERTS_SMP)
-erts_mtx_t chld_stat_mtx;
-#endif
-#if CHLDWTHR
-static erts_tid_t child_waiter_tid;
-/* chld_stat_mtx is used to protect against concurrent accesses
-   of the driver_data fields pid, alive, and status. */
-erts_cnd_t chld_stat_cnd;
-static long children_alive;
-#define CHLD_STAT_LOCK		erts_mtx_lock(&chld_stat_mtx)
-#define CHLD_STAT_UNLOCK	erts_mtx_unlock(&chld_stat_mtx)
-#define CHLD_STAT_WAIT		erts_cnd_wait(&chld_stat_cnd, &chld_stat_mtx)
-#define CHLD_STAT_SIGNAL	erts_cnd_signal(&chld_stat_cnd)
-#elif defined(ERTS_SMP) /* ------------------------------------------------- */
-#define CHLD_STAT_LOCK		erts_mtx_lock(&chld_stat_mtx)
-#define CHLD_STAT_UNLOCK	erts_mtx_unlock(&chld_stat_mtx)
-
-#else /* ------------------------------------------------------------------- */
-#define CHLD_STAT_LOCK
-#define CHLD_STAT_UNLOCK
-static volatile int children_died;
-#endif
-
-
-static struct fd_data {
-    char  pbuf[4];   /* hold partial packet bytes */
-    int   psz;       /* size of pbuf */
-    char  *buf;
-    char  *cpos;
-    int   sz;
-    int   remain;  /* for input on fd */
-} *fd_data;			/* indexed by fd */
-
-/* static FUNCTION(int, write_fill, (int, char*, int)); unused? */
-static void note_child_death(int, int);
-
-#if CHLDWTHR
-static void* child_waiter(void *);
-#endif
-
 static int crashdump_companion_cube_fd = -1;
 
 /********************* General functions ****************************/
@@ -453,9 +308,10 @@ MALLOC_USE_HASH(1);
 #ifdef USE_THREADS
 
 #ifdef ERTS_THR_HAVE_SIG_FUNCS
+
 /*
  * Child thread inherits parents signal mask at creation. In order to
- * guarantee that the main thread will receive all SIGINT, SIGCHLD, and
+ * guarantee that the main thread will receive all SIGINT, and
  * SIGUSR1 signals sent to the process, we block these signals in the
  * parent thread when creating a new thread.
  */
@@ -551,14 +407,11 @@ erts_sys_pre_init(void)
 #ifdef ERTS_THR_HAVE_SIG_FUNCS
     sigemptyset(&thr_create_sigmask);
     sigaddset(&thr_create_sigmask, SIGINT);   /* block interrupt */
-    sigaddset(&thr_create_sigmask, SIGCHLD);  /* block child signals */
     sigaddset(&thr_create_sigmask, SIGUSR1);  /* block user defined signal */
 #endif
 
     erts_thr_init(&eid);
 
-    report_exit_list = NULL;
-
 #ifdef ERTS_ENABLE_LOCK_COUNT
     erts_lcnt_init();
 #endif
@@ -569,17 +422,6 @@ erts_sys_pre_init(void)
 
 #ifdef USE_THREADS
 
-#if CHLDWTHR || defined(ERTS_SMP)
-    erts_mtx_init(&chld_stat_mtx, "child_status");
-#endif
-#if CHLDWTHR
-#ifndef ERTS_SMP
-    report_exit_transit_list = NULL;
-#endif
-    erts_cnd_init(&chld_stat_cnd);
-    children_alive = 0;
-#endif
-
 #ifdef ERTS_SMP
     erts_smp_atomic32_init_nob(&erts_break_requested, 0);
     erts_smp_atomic32_init_nob(&erts_got_sigusr1, 0);
@@ -589,9 +431,6 @@ erts_sys_pre_init(void)
     erts_got_sigusr1 = 0;
     have_prepared_crash_dump = 0;
 #endif
-#if !CHLDWTHR && !defined(ERTS_SMP)
-    children_died = 0;
-#endif
 
 #endif /* USE_THREADS */
 
@@ -628,39 +467,6 @@ erts_sys_pre_init(void)
 void
 erl_sys_init(void)
 {
-#if !DISABLE_VFORK
- {
-    int res;
-    char bindir[MAXPATHLEN];
-    size_t bindirsz = sizeof(bindir);
-    Uint csp_path_sz;
-
-    res = erts_sys_getenv_raw("BINDIR", bindir, &bindirsz);
-    if (res != 0) {
-	if (res < 0)
-	    erts_exit(1,
-		     "Environment variable BINDIR is not set\n");
-	if (res > 0)
-	    erts_exit(1,
-		     "Value of environment variable BINDIR is too large\n");
-    }
-    if (bindir[0] != DIR_SEPARATOR_CHAR)
-	erts_exit(1,
-		 "Environment variable BINDIR does not contain an"
-		 " absolute path\n");
-    csp_path_sz = (strlen(bindir)
-		   + 1 /* DIR_SEPARATOR_CHAR */
-		   + sizeof(CHILD_SETUP_PROG_NAME)
-		   + 1);
-    child_setup_prog = erts_alloc(ERTS_ALC_T_CS_PROG_PATH, csp_path_sz);
-    erts_smp_atomic_add_nob(&sys_misc_mem_sz, csp_path_sz);
-    erts_snprintf(child_setup_prog, csp_path_sz,
-            "%s%c%s",
-            bindir,
-            DIR_SEPARATOR_CHAR,
-            CHILD_SETUP_PROG_NAME);
- }
-#endif
 
 #ifdef USE_SETLINEBUF
     setlinebuf(stdout);
@@ -978,43 +784,6 @@ int sys_max_files(void)
    return(max_files);
 }
 
-static void block_signals(void)
-{
-#if !CHLDWTHR
-   sys_sigblock(SIGCHLD);
-#endif
-#ifndef ERTS_SMP
-   sys_sigblock(SIGINT);
-#ifndef ETHR_UNUSABLE_SIGUSRX
-   sys_sigblock(SIGUSR1);
-#endif /* #ifndef ETHR_UNUSABLE_SIGUSRX */
-#endif /* #ifndef ERTS_SMP */
-
-#if defined(ERTS_SMP) && !defined(ETHR_UNUSABLE_SIGUSRX)
-   sys_sigblock(ERTS_SYS_SUSPEND_SIGNAL);
-#endif
-
-}
-
-static void unblock_signals(void)
-{
-    /* Update erl_child_setup.c if changed */
-#if !CHLDWTHR
-   sys_sigrelease(SIGCHLD);
-#endif
-#ifndef ERTS_SMP
-   sys_sigrelease(SIGINT);
-#ifndef ETHR_UNUSABLE_SIGUSRX
-   sys_sigrelease(SIGUSR1);
-#endif /* #ifndef ETHR_UNUSABLE_SIGUSRX */
-#endif /* #ifndef ERTS_SMP */
-
-#if defined(ERTS_SMP) && !defined(ETHR_UNUSABLE_SIGUSRX)
-   sys_sigrelease(ERTS_SYS_SUSPEND_SIGNAL);
-#endif
-
-}
-
 /************************** OS info *******************************/
 
 /* Used by erlang:info/1. */
@@ -1102,1502 +871,6 @@ void fini_getenv_state(GETENV_STATE *state)
    erts_smp_rwmtx_runlock(&environ_rwmtx);
 }
 
-
-/************************** Port I/O *******************************/
-
-
-
-/* I. Common stuff */
-
-/*
- * Decreasing the size of it below 16384 is not allowed.
- */
-
-/* II. The spawn/fd/vanilla drivers */
-
-#define ERTS_SYS_READ_BUF_SZ (64*1024)
-
-/* 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);
-static void output(ErlDrvData, char*, ErlDrvSizeT);
-static void outputv(ErlDrvData, ErlIOVec*);
-static void stop_select(ErlDrvEvent, void*);
-
-struct erl_drv_entry spawn_driver_entry = {
-    spawn_init,
-    spawn_start,
-    stop,
-    output,
-    ready_input,
-    ready_output,
-    "spawn",
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    ERL_DRV_EXTENDED_MARKER,
-    ERL_DRV_EXTENDED_MAJOR_VERSION,
-    ERL_DRV_EXTENDED_MINOR_VERSION,
-    ERL_DRV_FLAG_USE_PORT_LOCKING,
-    NULL, NULL,
-    stop_select
-};
-struct erl_drv_entry fd_driver_entry = {
-    NULL,
-    fd_start,
-    fd_stop,
-    output,
-    ready_input,
-    ready_output, 
-    "fd",
-    NULL,
-    NULL,
-    fd_control,
-    NULL,
-    outputv,
-#if FDBLOCK
-    fd_ready_async, /* ready_async */
-#else
-    NULL,
-#endif
-    fd_flush, /* flush */
-    NULL, /* call */
-    NULL, /* event */
-    ERL_DRV_EXTENDED_MARKER,
-    ERL_DRV_EXTENDED_MAJOR_VERSION,
-    ERL_DRV_EXTENDED_MINOR_VERSION,
-    0, /* ERL_DRV_FLAGs */
-    NULL, /* handle2 */
-    NULL, /* process_exit */
-    stop_select
-};
-struct erl_drv_entry vanilla_driver_entry = {
-    NULL,
-    vanilla_start,
-    stop,
-    output,
-    ready_input,
-    ready_output,
-    "vanilla",
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL,
-    NULL, /* flush */
-    NULL, /* call */
-    NULL, /* event */
-    ERL_DRV_EXTENDED_MARKER,
-    ERL_DRV_EXTENDED_MAJOR_VERSION,
-    ERL_DRV_EXTENDED_MINOR_VERSION,
-    0, /* ERL_DRV_FLAGs */
-    NULL, /* handle2 */
-    NULL, /* process_exit */
-    stop_select
-};
-
-/* Handle SIGCHLD signals. */
-#if (defined(SIG_SIGSET) || defined(SIG_SIGNAL))
-static RETSIGTYPE onchld(void)
-#else
-static RETSIGTYPE onchld(int signum)
-#endif
-{
-#if CHLDWTHR
-    ASSERT(0); /* We should *never* catch a SIGCHLD signal */
-#elif defined(ERTS_SMP)
-    smp_sig_notify('C');
-#else
-    children_died = 1;
-    ERTS_CHK_IO_AS_INTR(); /* Make sure we don't sleep in poll */
-#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 is_blocking)
-{
-    Port *prt;
-    ErtsSysReportExit *report_exit;
-
-    if (!exit_status)
-	report_exit = NULL;
-    else {
-	report_exit = erts_alloc(ERTS_ALC_T_PRT_REP_EXIT,
-				 sizeof(ErtsSysReportExit));
-	report_exit->next = report_exit_list;
-	report_exit->port = erts_drvport2id(port_num);
-	report_exit->pid = pid;
-	report_exit->ifd = read_write & DO_READ ? ifd : -1;
-	report_exit->ofd = read_write & DO_WRITE ? ofd : -1;
-#if CHLDWTHR && !defined(ERTS_SMP)
-	report_exit->status = 0;
-#endif
-	report_exit_list = report_exit;
-    }
-
-    prt = erts_drvport2port(port_num);
-    if (prt != ERTS_INVALID_ERL_DRV_PORT)
-	prt->os_pid = pid;
-
-    if (read_write & DO_READ) {
-	driver_data[ifd].packet_bytes = packet_bytes;
-	driver_data[ifd].port_num = port_num;
-	driver_data[ifd].report_exit = report_exit;
-	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 */
-	    driver_data[ifd].ofd = -1;
-	}
-	(void) driver_select(port_num, ifd, (ERL_DRV_READ|ERL_DRV_USE), 1);
-	return(ifd);
-    } else {			/* DO_WRITE only */
-	driver_data[ofd].packet_bytes = packet_bytes;
-	driver_data[ofd].port_num = port_num;
-	driver_data[ofd].report_exit = report_exit;
-	driver_data[ofd].ofd = ofd;
-	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);
-    }
-}
-
-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 */
-   thr_opts.name = "child_waiter";
-#endif
-
-   sys_signal(SIGPIPE, SIG_IGN); /* Ignore - we'll handle the write failure */
-   driver_data = (struct driver_data *)
-       erts_alloc(ERTS_ALC_T_DRV_TAB, max_files * sizeof(struct driver_data));
-   erts_smp_atomic_add_nob(&sys_misc_mem_sz,
-			   max_files * sizeof(struct driver_data));
-
-   for (i = 0; i < max_files; i++)
-      driver_data[i].pid = -1;
-
-#if CHLDWTHR
-   sys_sigblock(SIGCHLD);
-#endif
-
-   sys_signal(SIGCHLD, onchld); /* Reap children */
-
-#if CHLDWTHR
-   erts_thr_create(&child_waiter_tid, child_waiter, NULL, &thr_opts);
-#endif
-
-   return 1;
-}
-
-static void close_pipes(int ifd[2], int ofd[2], int read_write)
-{
-    if (read_write & DO_READ) {
-	(void) close(ifd[0]);
-	(void) close(ifd[1]);
-    }
-    if (read_write & DO_WRITE) {
-	(void) close(ofd[0]);
-	(void) close(ofd[1]);
-    }
-}
-
-static void init_fd_data(int fd, ErlDrvPort port_num)
-{
-    fd_data[fd].buf = NULL;
-    fd_data[fd].cpos = NULL;
-    fd_data[fd].remain = 0;
-    fd_data[fd].sz = 0;
-    fd_data[fd].psz = 0;
-}
-
-static char **build_unix_environment(char *block)
-{
-    int i;
-    int j;
-    int len;
-    char *cp;
-    char **cpp;
-    char** old_env;
-    
-    ERTS_SMP_LC_ASSERT(erts_smp_lc_rwmtx_is_rlocked(&environ_rwmtx));
-
-    cp = block;
-    len = 0;
-    while (*cp != '\0') {
-	cp += strlen(cp) + 1;
-	len++;
-    }
-    old_env = environ;
-    while (*old_env++ != NULL) {
-	len++;
-    }
-    
-    cpp = (char **) erts_alloc_fnf(ERTS_ALC_T_ENVIRONMENT,
-				   sizeof(char *) * (len+1));
-    if (cpp == NULL) {
-	return NULL;
-    }
-
-    cp = block;
-    len = 0;
-    while (*cp != '\0') {
-	cpp[len] = cp;
-	cp += strlen(cp) + 1;
-	len++;
-    }
-    
-    i = len;
-    for (old_env = environ; *old_env; old_env++) {
-	char* old = *old_env;
-
-	for (j = 0; j < len; j++) {
-	    char *s, *t;
-
-	    s = cpp[j];
-	    t = old;
-	    while (*s == *t && *s != '=') {
-		s++, t++;
-	    }
-	    if (*s == '=' && *t == '=') {
-		break;
-	    }
-	}
-
-	if (j == len) {		/* New version not found */
-	    cpp[len++] = old;
-	}
-    }
-
-    for (j = 0; j < i; ) {
-        size_t last = strlen(cpp[j])-1;
-	if (cpp[j][last] == '=' && strchr(cpp[j], '=') == cpp[j]+last) {
-	    cpp[j] = cpp[--len];
-	    if (len < i) {
-		i--;
-	    } else {
-		j++;
-	    }
-	}
-	else {
-	    j++;
-	}
-    }
-
-    cpp[len] = NULL;
-    return cpp;
-}
-
-/*
-  [arndt] In most Unix systems, including Solaris 2.5, 'fork' allocates memory
-  in swap space for the child of a 'fork', whereas 'vfork' does not do this.
-  The natural call to use here is therefore 'vfork'. Due to a bug in
-  'vfork' in Solaris 2.5 (apparently fixed in 2.6), using 'vfork'
-  can be dangerous in what seems to be these circumstances:
-      If the child code under a vfork sets the signal action to SIG_DFL
-      (or SIG_IGN)
-      for any signal which was previously set to a signal handler, the
-      state of the parent is clobbered, so that the later arrival of
-      such a signal yields a sigsegv in the parent. If the signal was
-      not set to a signal handler, but ignored, all seems to work.
-  If you change the forking code below, beware of this.
- */
-
-static ErlDrvData spawn_start(ErlDrvPort port_num, char* name, SysDriverOpts* opts)
-{
-#define CMD_LINE_PREFIX_STR "exec "
-#define CMD_LINE_PREFIX_STR_SZ (sizeof(CMD_LINE_PREFIX_STR) - 1)
-
-    int ifd[2], ofd[2], len, pid, i;
-    char **volatile new_environ; /* volatile since a vfork() then cannot
-				    cause 'new_environ' to be clobbered
-				    in the parent process. */
-    int saved_errno;
-    long res;
-    char *cmd_line;
-#ifndef QNX
-    int unbind;
-#endif
-#if !DISABLE_VFORK
-    int no_vfork;
-    size_t no_vfork_sz = sizeof(no_vfork);
-
-    no_vfork = (erts_sys_getenv_raw("ERL_NO_VFORK",
-				    (char *) &no_vfork,
-				    &no_vfork_sz) >= 0);
-#endif
-
-    switch (opts->read_write) {
-    case DO_READ:
-	if (pipe(ifd) < 0)
-	    return ERL_DRV_ERROR_ERRNO;
-	if (ifd[0] >= max_files) {
-	    close_pipes(ifd, ofd, opts->read_write);
-	    errno = EMFILE;
-	    return ERL_DRV_ERROR_ERRNO;
-	}
-	ofd[1] = -1;		/* keep purify happy */
-	break;
-    case DO_WRITE:
-	if (pipe(ofd) < 0) return ERL_DRV_ERROR_ERRNO;
-	if (ofd[1] >= max_files) {
-	    close_pipes(ifd, ofd, opts->read_write);
-	    errno = EMFILE;
-	    return ERL_DRV_ERROR_ERRNO;
-	}
-	ifd[0] = -1;		/* keep purify happy */
-	break;
-    case DO_READ|DO_WRITE:
-	if (pipe(ifd) < 0) return ERL_DRV_ERROR_ERRNO;
-	errno = EMFILE;		/* default for next two conditions */
-	if (ifd[0] >= max_files || pipe(ofd) < 0) {
-	    close_pipes(ifd, ofd, DO_READ);
-	    return ERL_DRV_ERROR_ERRNO;
-	}
-	if (ofd[1] >= max_files) {
-	    close_pipes(ifd, ofd, opts->read_write);
-	    errno = EMFILE;
-	    return ERL_DRV_ERROR_ERRNO;
-	}
-	break;
-    default:
-	ASSERT(0);
-	return ERL_DRV_ERROR_GENERAL;
-    }
-
-    if (opts->spawn_type == ERTS_SPAWN_EXECUTABLE) {
-	/* started with spawn_executable, not with spawn */
-	len = strlen(name);
-	cmd_line = (char *) erts_alloc_fnf(ERTS_ALC_T_TMP, len + 1);
-	if (!cmd_line) {
-	    close_pipes(ifd, ofd, opts->read_write);
-	    errno = ENOMEM;
-	    return ERL_DRV_ERROR_ERRNO;
-	}
-	memcpy((void *) cmd_line,(void *) name, len);
-	cmd_line[len] = '\0';
-	if (access(cmd_line,X_OK) != 0) {
-	    int save_errno = errno;
-	    erts_free(ERTS_ALC_T_TMP, cmd_line);
-	    errno = save_errno;
-	    return ERL_DRV_ERROR_ERRNO;
-	}
-    } else {
-	/* make the string suitable for giving to "sh" */
-	len = strlen(name);
-	cmd_line = (char *) erts_alloc_fnf(ERTS_ALC_T_TMP,
-					   CMD_LINE_PREFIX_STR_SZ + len + 1);
-	if (!cmd_line) {
-	    close_pipes(ifd, ofd, opts->read_write);
-	    errno = ENOMEM;
-	    return ERL_DRV_ERROR_ERRNO;
-	}
-	memcpy((void *) cmd_line,
-	       (void *) CMD_LINE_PREFIX_STR,
-	       CMD_LINE_PREFIX_STR_SZ);
-	memcpy((void *) (cmd_line + CMD_LINE_PREFIX_STR_SZ), (void *) name, len);
-	cmd_line[CMD_LINE_PREFIX_STR_SZ + len] = '\0';
-    }
-
-    erts_smp_rwmtx_rlock(&environ_rwmtx);
-
-    if (opts->envir == NULL) {
-	new_environ = environ;
-    } else if ((new_environ = build_unix_environment(opts->envir)) == NULL) {
-	erts_smp_rwmtx_runlock(&environ_rwmtx);
-	erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
-	errno = ENOMEM;
-	return ERL_DRV_ERROR_ERRNO;
-    } 
-
-#ifndef QNX
-    /* Block child from SIGINT and SIGUSR1. Must be before fork()
-       to be safe. */
-    block_signals();
-
-    CHLD_STAT_LOCK;
-
-    unbind = erts_sched_bind_atfork_prepare();
-
-#if !DISABLE_VFORK
-    /* See fork/vfork discussion before this function. */
-    if (no_vfork) {
-#endif
-
-	DEBUGF(("Using fork\n"));
-	pid = fork();
-
-	if (pid == 0) {
-	    /* The child! Setup child... */
-
-	    if (erts_sched_bind_atfork_child(unbind) != 0)
-		goto child_error;
-
-	    /* OBSERVE!
-	     * Keep child setup after vfork() (implemented below and in
-	     * erl_child_setup.c) up to date if changes are made here.
-	     */
-
-	    if (opts->use_stdio) {
-		if (opts->read_write & DO_READ) {
-		    /* stdout for process */
-		    if (dup2(ifd[1], 1) < 0)
-			goto child_error;
-		    if(opts->redir_stderr)
-			/* stderr for process */
-			if (dup2(ifd[1], 2) < 0)
-			    goto child_error;
-		}
-		if (opts->read_write & DO_WRITE)
-		    /* stdin for process */
-		    if (dup2(ofd[0], 0) < 0)
-			goto child_error;
-	    }
-	    else {	/* XXX will fail if ofd[0] == 4 (unlikely..) */
-		if (opts->read_write & DO_READ)
-		    if (dup2(ifd[1], 4) < 0)
-			goto child_error;
-		if (opts->read_write & DO_WRITE)
-		    if (dup2(ofd[0], 3) < 0)
-			goto child_error;
-	    }
-
-#if defined(HAVE_CLOSEFROM)
-	    closefrom(opts->use_stdio ? 3 : 5);
-#else
-	    for (i = opts->use_stdio ? 3 : 5; i < max_files; i++)
-		(void) close(i);
-#endif
-
-	    if (opts->wd && chdir(opts->wd) < 0)
-		goto child_error;
-
-#if defined(USE_SETPGRP_NOARGS)		/* SysV */
-	    (void) setpgrp();
-#elif defined(USE_SETPGRP)		/* BSD */
-	    (void) setpgrp(0, getpid());
-#else					/* POSIX */
-	    (void) setsid();
-#endif
-	    
-	    unblock_signals();
-
-	    if (opts->spawn_type == ERTS_SPAWN_EXECUTABLE) {
-		if (opts->argv == NULL) {
-		    execle(cmd_line,cmd_line,(char *) NULL, new_environ);
-		} else {
-		    if (opts->argv[0] == erts_default_arg0) {
-			opts->argv[0] = cmd_line;
-		    }
-		    execve(cmd_line, opts->argv, new_environ);
-		    if (opts->argv[0] == cmd_line) {
-			opts->argv[0] = erts_default_arg0;
-		    }
-		}
-	    } else {
-		execle(SHELL, "sh", "-c", cmd_line, (char *) NULL, new_environ);
-	    }
-	child_error:
-	    _exit(1);
-	}
-#if !DISABLE_VFORK
-    }
-#define ENOUGH_BYTES (44)
-    else { /* Use vfork() */
-	char **cs_argv= erts_alloc(ERTS_ALC_T_TMP,(CS_ARGV_NO_OF_ARGS + 1)*
-				   sizeof(char *));
-	char fd_close_range[ENOUGH_BYTES];                  /* 44 bytes are enough to  */
-	char dup2_op[CS_ARGV_NO_OF_DUP2_OPS][ENOUGH_BYTES]; /* hold any "%d:%d" string */
-                                                            /* on a 64-bit machine.    */
-
-	/* Setup argv[] for the child setup program (implemented in
-	   erl_child_setup.c) */
-	i = 0;
-	if (opts->use_stdio) {
-	    if (opts->read_write & DO_READ){
-		/* stdout for process */
-		erts_snprintf(&dup2_op[i++][0], ENOUGH_BYTES, "%d:%d", ifd[1], 1);
-		if(opts->redir_stderr)
-		    /* stderr for process */
-		    erts_snprintf(&dup2_op[i++][0], ENOUGH_BYTES, "%d:%d", ifd[1], 2);
-	    }
-	    if (opts->read_write & DO_WRITE)
-		/* stdin for process */
-		erts_snprintf(&dup2_op[i++][0], ENOUGH_BYTES, "%d:%d", ofd[0], 0);
-	} else {	/* XXX will fail if ofd[0] == 4 (unlikely..) */
-	    if (opts->read_write & DO_READ)
-		erts_snprintf(&dup2_op[i++][0], ENOUGH_BYTES, "%d:%d", ifd[1], 4);
-	    if (opts->read_write & DO_WRITE)
-		erts_snprintf(&dup2_op[i++][0], ENOUGH_BYTES, "%d:%d", ofd[0], 3);
-	}
-	for (; i < CS_ARGV_NO_OF_DUP2_OPS; i++)
-	    strcpy(&dup2_op[i][0], "-");
-	erts_snprintf(fd_close_range, ENOUGH_BYTES, "%d:%d", opts->use_stdio ? 3 : 5, max_files-1);
-
-	cs_argv[CS_ARGV_PROGNAME_IX] = child_setup_prog;
-	cs_argv[CS_ARGV_WD_IX] = opts->wd ? opts->wd : ".";
-	cs_argv[CS_ARGV_UNBIND_IX] = erts_sched_bind_atvfork_child(unbind);
-	cs_argv[CS_ARGV_FD_CR_IX] = fd_close_range;
-	for (i = 0; i < CS_ARGV_NO_OF_DUP2_OPS; i++)
-	    cs_argv[CS_ARGV_DUP2_OP_IX(i)] = &dup2_op[i][0];
-
-	if (opts->spawn_type == ERTS_SPAWN_EXECUTABLE) {
-	    int num = 0;
-	    int j = 0;
-	    if (opts->argv != NULL) {
-		for(; opts->argv[num] != NULL; ++num)
-		    ;
-	    }
-	    cs_argv = erts_realloc(ERTS_ALC_T_TMP,cs_argv, (CS_ARGV_NO_OF_ARGS + 1 + num + 1) * sizeof(char *));
-	    cs_argv[CS_ARGV_CMD_IX] = "-";
-	    cs_argv[CS_ARGV_NO_OF_ARGS] = cmd_line;
-	    if (opts->argv != NULL) {
-		for (;opts->argv[j] != NULL; ++j) {
-		    if (opts->argv[j] == erts_default_arg0) {
-			cs_argv[CS_ARGV_NO_OF_ARGS + 1 + j] = cmd_line;
-		    } else {
-			cs_argv[CS_ARGV_NO_OF_ARGS + 1 + j] = opts->argv[j];
-		    }
-		}
-	    }
-	    cs_argv[CS_ARGV_NO_OF_ARGS + 1 + j] = NULL;
-	} else {
-	    cs_argv[CS_ARGV_CMD_IX] = cmd_line; /* Command */
-	    cs_argv[CS_ARGV_NO_OF_ARGS] = NULL;
-	}
-	DEBUGF(("Using vfork\n"));
-	pid = vfork();
-
-	if (pid == 0) {
-	    /* The child! */
-
-	    /* Observe!
-	     * OTP-4389: The child setup program (implemented in
-	     * erl_child_setup.c) will perform the necessary setup of the
-	     * child before it execs to the user program. This because
-	     * vfork() only allow an *immediate* execve() or _exit() in the
-	     * child.
-	     */
-	    execve(child_setup_prog, cs_argv, new_environ);
-	    _exit(1);
-	}
-	erts_free(ERTS_ALC_T_TMP,cs_argv);
-    }
-#undef ENOUGH_BYTES
-#endif
-
-    erts_sched_bind_atfork_parent(unbind);
-
-    if (pid == -1) {
-        saved_errno = errno;
-	CHLD_STAT_UNLOCK;
-	erts_smp_rwmtx_runlock(&environ_rwmtx);
-	erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
-        unblock_signals();
-        close_pipes(ifd, ofd, opts->read_write);
-	errno = saved_errno;
-	return ERL_DRV_ERROR_ERRNO;
-    }
-#else /* QNX */
-    if (opts->use_stdio) {
-	if (opts->read_write & DO_READ)
-	    qnx_spawn_options.iov[1] = ifd[1];  /* stdout for process */
-	if (opts->read_write & DO_WRITE)
-	    qnx_spawn_options.iov[0] = ofd[0];  /* stdin for process */
-	} 
-    else {
-	if (opts->read_write & DO_READ)
-	    qnx_spawn_options.iov[4] = ifd[1];
-	if (opts->read_write & DO_WRITE)
-	    qnx_spawn_options.iov[3] = ofd[0];
-    }
-    /* Close fds on exec */
-    for (i = 3; i < max_files; i++)
-	fcntl(i, F_SETFD, 1);
-
-    qnx_spawn_options.flags = _SPAWN_SETSID;
-    if ((pid = spawnl(P_NOWAIT, SHELL, SHELL, "-c", cmd_line, 
-                      (char *) 0)) < 0) {
-	erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
-        reset_qnx_spawn();
-	erts_smp_rwmtx_runlock(&environ_rwmtx);
-	close_pipes(ifd, ofd, opts->read_write);
-	return ERL_DRV_ERROR_GENERAL;
-    }
-    reset_qnx_spawn();
-#endif /* QNX */
-
-    erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
-
-    if (new_environ != environ)
-	erts_free(ERTS_ALC_T_ENVIRONMENT, (void *) new_environ);
-
-    if (opts->read_write & DO_READ) 
-	(void) close(ifd[1]);
-    if (opts->read_write & DO_WRITE)
-	(void) close(ofd[0]);
-	
-    if (opts->read_write & DO_READ) {
-	SET_NONBLOCKING(ifd[0]);
-	init_fd_data(ifd[0], port_num);
-    }
-    if (opts->read_write & DO_WRITE) {
-	SET_NONBLOCKING(ofd[1]);
-        init_fd_data(ofd[1], port_num);
-    }
-
-    res = set_driver_data(port_num, ifd[0], ofd[1], opts->packet_bytes,
-			  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();
-
-#if CHLDWTHR
-    ASSERT(children_alive >= 0);
-
-    if (!(children_alive++))
-	CHLD_STAT_SIGNAL; /* Wake up child waiter thread if no children
-			     was alive before we fork()ed ... */
-#endif
-    /* Don't unlock chld_stat_mtx until now of the same reason as above */
-    CHLD_STAT_UNLOCK;
-
-    erts_smp_rwmtx_runlock(&environ_rwmtx);
-
-    return (ErlDrvData)res;
-#undef CMD_LINE_PREFIX_STR
-#undef CMD_LINE_PREFIX_STR_SZ
-}
-
-#ifdef QNX
-static reset_qnx_spawn()
-{
-    int i;
-
-    /* Reset qnx_spawn_options */
-    qnx_spawn_options.flags = 0; 
-    qnx_spawn_options.iov[0] = 0xff;
-    qnx_spawn_options.iov[1] = 0xff;
-    qnx_spawn_options.iov[2] = 0xff;
-    qnx_spawn_options.iov[3] = 0xff;
-}
-#endif
-
-#define FD_DEF_HEIGHT 24
-#define FD_DEF_WIDTH 80
-/* Control op */
-#define FD_CTRL_OP_GET_WINSIZE 100
-
-static int fd_get_window_size(int fd, Uint32 *width, Uint32 *height)
-{
-#ifdef TIOCGWINSZ 
-    struct winsize ws;
-    if (ioctl(fd,TIOCGWINSZ,&ws) == 0) {
-	*width = (Uint32) ws.ws_col;
-	*height = (Uint32) ws.ws_row;
-	return 0;
-    }
-#endif
-    return -1;
-}
-
-static ErlDrvSSizeT fd_control(ErlDrvData drv_data,
-			       unsigned int command,
-			       char *buf, ErlDrvSizeT len,
-			       char **rbuf, ErlDrvSizeT rlen)
-{
-    int fd = (int)(long)drv_data;
-    char resbuff[2*sizeof(Uint32)];
-    switch (command) {
-    case FD_CTRL_OP_GET_WINSIZE:
-	{
-	    Uint32 w,h;
-	    if (fd_get_window_size(fd,&w,&h)) 
-		return 0;
-	    memcpy(resbuff,&w,sizeof(Uint32));
-	    memcpy(resbuff+sizeof(Uint32),&h,sizeof(Uint32));
-	}
-	break;
-    default:
-	return 0;
-    }
-    if (rlen < 2*sizeof(Uint32)) {
-	*rbuf = driver_alloc(2*sizeof(Uint32));
-    }
-    memcpy(*rbuf,resbuff,2*sizeof(Uint32));
-    return 2*sizeof(Uint32);
-}
-
-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))
-	return ERL_DRV_ERROR_GENERAL;
-
-    /*
-     * Historical:
-     *
-     * "Note about nonblocking I/O.
-     *
-     * At least on Solaris, setting the write end of a TTY to nonblocking,
-     * will set the input end to nonblocking as well (and vice-versa).
-     * If erl is run in a pipeline like this:  cat | erl
-     * the input end of the TTY will be the standard input of cat.
-     * And cat is not prepared to handle nonblocking I/O."
-     *
-     * Actually, the reason for this is not that the tty itself gets set
-     * in non-blocking mode, but that the "input end" (cat's stdin) and
-     * the "output end" (erlang's stdout) are typically the "same" file
-     * descriptor, dup()'ed from a single fd by one of this process'
-     * ancestors.
-     *
-     * The workaround for this problem used to be a rather bad kludge,
-     * interposing an extra process ("internal cat") between erlang's
-     * stdout and the original stdout, allowing erlang to set its stdout
-     * in non-blocking mode without affecting the stdin of the preceding
-     * process in the pipeline - and being a kludge, it caused all kinds
-     * of weird problems.
-     *
-     * So, this is the current logic:
-     *
-     * The only reason to set non-blocking mode on the output fd at all is
-     * if it's something that can cause a write() to block, of course,
-     * i.e. primarily if it points to a tty, socket, pipe, or fifo. 
-     *
-     * If we don't set non-blocking mode when we "should" have, and output
-     * becomes blocked, the entire runtime system will be suspended - this
-     * is normally bad of course, and can happen fairly "easily" - e.g. user
-     * hits ^S on tty - but doesn't necessarily happen.
-     * 
-     * If we do set non-blocking mode when we "shouldn't" have, the runtime
-     * system will end up seeing EOF on the input fd (due to the preceding
-     * process dying), which typically will cause the entire runtime system
-     * to terminate immediately (due to whatever erlang process is seeing
-     * the EOF taking it as a signal to halt the system). This is *very* bad.
-     * 
-     * I.e. we should take a conservative approach, and only set non-
-     * blocking mode when we a) need to, and b) are reasonably certain
-     * that it won't be a problem. And as in the example above, the problem
-     * occurs when input fd and output fd point to different "things".
-     *
-     * However, determining that they are not just the same "type" of
-     * "thing", but actually the same instance of that type of thing, is
-     * unreasonably complex in many/most cases.
-     *
-     * Also, with pipes, sockets, and fifos it's far from obvious that the
-     * user *wants* non-blocking output: If you're running erlang inside
-     * some complex pipeline, you're probably not running a real-time system
-     * that must never stop, but rather *want* it to suspend if the output
-     * channel is "full".
-     *
-     * So, the bottom line: We will only set the output fd non-blocking if
-     * it points to a tty, and either a) the input fd also points to a tty,
-     * or b) we can make sure that setting the output fd non-blocking
-     * doesn't interfere with someone else's input, via a somewhat milder
-     * kludge than the above.
-     *
-     * Also keep in mind that while this code is almost exclusively run as
-     * a result of an erlang open_port({fd,0,1}, ...), that isn't the only
-     * 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) {
-	init_fd_data(opts->ifd, port_num);
-    }
-    if (opts->read_write & DO_WRITE) {
-	init_fd_data(opts->ofd, port_num);
-
-	/* If we don't have a read end, all bets are off - no non-blocking. */
-	if (opts->read_write & DO_READ) {
-
-	    if (isatty(opts->ofd)) { /* output fd is a tty:-) */
-
-		if (isatty(opts->ifd)) { /* input fd is also a tty */
-
-		    /* To really do this "right", we should also check that
-		       input and output fd point to the *same* tty - but
-		       this seems like overkill; ttyname() isn't for free,
-		       and this is a very common case - and it's hard to
-		       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 */
-
-		    /* This is a "problem case", but also common (see the
-		       example above) - i.e. it makes sense to try a bit
-		       harder before giving up on non-blocking mode: Try to
-		       re-open the tty that the output fd points to, and if
-		       successful replace the original one with the "new" fd
-		       obtained this way, and set *that* one in non-blocking
-		       mode. (Yes, this is a kludge.)
-
-		       However, re-opening the tty may fail in a couple of
-		       (unusual) cases:
-
-		       1) The name of the tty (or an equivalent one, i.e.
-			  same major/minor number) can't be found, because
-			  it actually lives somewhere other than /dev (or
-			  wherever ttyname() looks for it), and isn't
-			  equivalent to any of those that do live in the
-			  "standard" place - this should be *very* unusual.
-
-		       2) Permissions on the tty don't allow us to open it -
-			  it's perfectly possible to have an fd open to an
-			  object whose permissions wouldn't allow us to open
-			  it. This is not as unusual as it sounds, one case
-			  is if the user has su'ed to someone else (not
-			  root) - we have a read/write fd open to the tty
-			  (because it has been inherited all the way down
-			  here), but we have neither read nor write
-			  permission for the tty.
-
-		       In these cases, we finally give up, and don't set the
-		       output fd in non-blocking mode. */
-
-		    char *tty;
-		    int nfd;
-
-		    if ((tty = ttyname(opts->ofd)) != NULL &&
-			(nfd = open(tty, O_WRONLY)) != -1) {
-			dup2(nfd, opts->ofd);
-			close(nfd);
-                        non_blocking = 1;
-			SET_NONBLOCKING(opts->ofd);
-		    }
-		}
-	    }
-	}
-    }
-    CHLD_STAT_LOCK;
-    res = (ErlDrvData)(long)set_driver_data(port_num, opts->ifd, opts->ofd,
-                                            opts->packet_bytes,
-                                            opts->read_write, 0, -1,
-                                            !non_blocking);
-    CHLD_STAT_UNLOCK;
-    return res;
-}
-
-static void clear_fd_data(int fd) 
-{
-    if (fd_data[fd].sz > 0) {
-	erts_free(ERTS_ALC_T_FD_ENTRY_BUF, (void *) fd_data[fd].buf);
-	ASSERT(erts_smp_atomic_read_nob(&sys_misc_mem_sz) >= fd_data[fd].sz);
-	erts_smp_atomic_add_nob(&sys_misc_mem_sz, -1*fd_data[fd].sz);
-    }
-    fd_data[fd].buf = NULL;
-    fd_data[fd].sz = 0;
-    fd_data[fd].remain = 0;
-    fd_data[fd].cpos = NULL;
-    fd_data[fd].psz = 0;
-}
-
-static void nbio_stop_fd(ErlDrvPort prt, int fd)
-{
-    driver_select(prt,fd,DO_READ|DO_WRITE,0);
-    clear_fd_data(fd);
-    SET_BLOCKING(fd);
-}
-
-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;
-    
-#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,
-				SysDriverOpts* opts)
-{
-    int flags, fd;
-    ErlDrvData res;
-
-    flags = (opts->read_write == DO_READ ? O_RDONLY :
-	     opts->read_write == DO_WRITE ? O_WRONLY|O_CREAT|O_TRUNC :
-	     O_RDWR|O_CREAT);
-    if ((fd = open(name, flags, 0666)) < 0)
-	return ERL_DRV_ERROR_GENERAL;
-    if (fd >= max_files) {
-	close(fd);
-	return ERL_DRV_ERROR_GENERAL;
-    }
-    SET_NONBLOCKING(fd);
-    init_fd_data(fd, port_num);
-
-    CHLD_STAT_LOCK;
-    res = (ErlDrvData)(long)set_driver_data(port_num, fd, fd,
-                                            opts->packet_bytes,
-                                            opts->read_write, 0, -1, 0);
-    CHLD_STAT_UNLOCK;
-    return res;
-}
-
-/* Note that driver_data[fd].ifd == fd if the port was opened for reading, */
-/* otherwise (i.e. write only) driver_data[fd].ofd = fd.  */
-
-static void stop(ErlDrvData fd)
-{
-    ErlDrvPort prt;
-    int ofd;
-
-    prt = driver_data[(int)(long)fd].port_num;
-    nbio_stop_fd(prt, (int)(long)fd);
-
-    ofd = driver_data[(int)(long)fd].ofd;
-    if (ofd != (int)(long)fd && (int)(long)ofd != -1)
-	nbio_stop_fd(prt, ofd);
-    else
-	ofd = -1;
-
-    CHLD_STAT_LOCK;
-
-    /* Mark as unused. */
-    driver_data[(int)(long)fd].pid = -1;
-
-    CHLD_STAT_UNLOCK;
-
-    /* SMP note: Close has to be last thing done (open file descriptors work
-       as locks on driver_data[] entries) */
-    driver_select(prt, (int)(long)fd, ERL_DRV_USE, 0);  /* close(fd); */
-    if (ofd >= 0) {
-	driver_select(prt, (int)(long)ofd, ERL_DRV_USE, 0);  /* close(ofd); */
-    }
-}
-
-/* used by fd_driver */
-static void outputv(ErlDrvData e, ErlIOVec* ev)
-{
-    int fd = (int)(long)e;
-    ErlDrvPort ix = driver_data[fd].port_num;
-    int pb = driver_data[fd].packet_bytes;
-    int ofd = driver_data[fd].ofd;
-    ssize_t n;
-    ErlDrvSizeT sz;
-    char lb[4];
-    char* lbp;
-    ErlDrvSizeT len = ev->size;
-
-    /* (len > ((unsigned long)-1 >> (4-pb)*8)) */
-    /*    if (pb >= 0 && (len & (((ErlDrvSizeT)1 << (pb*8))) - 1) != len) {*/
-    if (((pb == 2) && (len > 0xffff)) || (pb == 1 && len > 0xff)) {
-	driver_failure_posix(ix, EINVAL);
-	return; /* -1; */
-    }
-    /* Handles 0 <= pb <= 4 only */
-    put_int32((Uint32) len, lb);
-    lbp = lb + (4-pb);
-
-    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 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);
-	if (n == ev->size)
-	    return; /* 0;*/
-	if (n < 0) {
-	    if ((errno != EINTR) && (errno != ERRNO_BLOCK)) {
-		driver_failure_posix(ix, errno);
-		return; /* -1;*/
-	    }
-	    n = 0;
-	}
-	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;
-    ErlDrvPort ix = driver_data[fd].port_num;
-    int pb = driver_data[fd].packet_bytes;
-    int ofd = driver_data[fd].ofd;
-    ssize_t n;
-    ErlDrvSizeT sz;
-    char lb[4];
-    char* lbp;
-    struct iovec iv[2];
-
-    /* (len > ((unsigned long)-1 >> (4-pb)*8)) */
-    if (((pb == 2) && (len > 0xffff)) || (pb == 1 && len > 0xff)) {
-	driver_failure_posix(ix, EINVAL);
-	return; /* -1; */
-    }
-    put_int32(len, lb);
-    lbp = lb + (4-pb);
-
-    if ((sz = driver_sizeq(ix)) > 0) {
-	driver_enq(ix, lbp, pb);
-	driver_enq(ix, buf, len);
-	if (sz + len + pb >= (1 << 13))
-	    set_busy_port(ix, 1);
-    }
-    else {
-	iv[0].iov_base = lbp;
-	iv[0].iov_len = pb;  /* should work for pb=0 */
-	iv[1].iov_base = buf;
-	iv[1].iov_len = len;
-	n = writev(ofd, iv, 2);
-	if (n == pb+len)
-	    return; /* 0; */
-	if (n < 0) {
-	    if ((errno != EINTR) && (errno != ERRNO_BLOCK)) {
-		driver_failure_posix(ix, errno);
-		return; /* -1; */
-	    }
-	    n = 0;
-	}
-	if (n < pb) {
-	    driver_enq(ix, lbp+n, pb-n);
-	    driver_enq(ix, buf, len);
-	}
-	else {
-	    n -= pb;
-	    driver_enq(ix, buf+n, len-n);
-	}
-	driver_select(ix, ofd, ERL_DRV_WRITE|ERL_DRV_USE, 1);
-    }
-    return; /* 0; */
-}
-
-static int port_inp_failure(ErlDrvPort port_num, int ready_fd, int res)
-				/* Result: 0 (eof) or -1 (error) */
-{
-    int err = errno;
-
-    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;
-
-	    if (driver_data[ready_fd].alive) {
-		/*
-		 * We have eof and want to report exit status, but the process
-		 * hasn't exited yet. When it does report_exit_status() will
-		 * driver_select() this fd which will make sure that we get
-		 * back here with driver_data[ready_fd].alive == 0 and
-		 * driver_data[ready_fd].status set.
-		 */
-		CHLD_STAT_UNLOCK;
-		return 0;
-	    }
-	    else {
-		int status = driver_data[ready_fd].status;
-		CHLD_STAT_UNLOCK;
-
-		/* We need not be prepared for stopped/continued processes. */
-		if (WIFSIGNALED(status))
-		    status = 128 + WTERMSIG(status);
-		else
-		    status = WEXITSTATUS(status);
-
-		driver_report_exit(driver_data[ready_fd].port_num, status);
-	    }
-       }
-       driver_failure_eof(port_num);
-    } else {
-	driver_failure_posix(port_num, err);
-    }
-    return 0;
-}
-
-/* fd is the drv_data that is returned from the */
-/* initial start routine                        */
-/* ready_fd is the descriptor that is ready to read */
-
-static void ready_input(ErlDrvData e, ErlDrvEvent ready_fd)
-{
-    int fd = (int)(long)e;
-    ErlDrvPort port_num;
-    int packet_bytes;
-    int res;
-    Uint h;
-
-    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);
-	res = read(ready_fd, read_buf, ERTS_SYS_READ_BUF_SZ);
-	if (res < 0) {
-	    if ((errno != EINTR) && (errno != ERRNO_BLOCK))
-		port_inp_failure(port_num, ready_fd, res);
-	}
-	else if (res == 0)
-	    port_inp_failure(port_num, ready_fd, res);
-	else 
-	    driver_output(port_num, (char*) read_buf, res);
-	erts_free(ERTS_ALC_T_SYS_READ_BUF, (void *) read_buf);
-    }
-    else if (fd_data[ready_fd].remain > 0) { /* We try to read the remainder */
-	/* space is allocated in buf */
-	res = read(ready_fd, fd_data[ready_fd].cpos, 
-		   fd_data[ready_fd].remain);
-	if (res < 0) {
-	    if ((errno != EINTR) && (errno != ERRNO_BLOCK))
-		port_inp_failure(port_num, ready_fd, res);
-	}
-	else if (res == 0) {
-	    port_inp_failure(port_num, ready_fd, res);
-	}
-	else if (res == fd_data[ready_fd].remain) { /* we're done  */
-	    driver_output(port_num, fd_data[ready_fd].buf, 
-			  fd_data[ready_fd].sz);
-	    clear_fd_data(ready_fd);
-	}
-	else { /*  if (res < fd_data[ready_fd].remain) */
-	    fd_data[ready_fd].cpos += res;
-	    fd_data[ready_fd].remain -= res;
-	}
-    }
-    else if (fd_data[ready_fd].remain == 0) { /* clean fd */
-	byte *read_buf = (byte *) erts_alloc(ERTS_ALC_T_SYS_READ_BUF,
-					     ERTS_SYS_READ_BUF_SZ);
-	/* We make one read attempt and see what happens */
-	res = read(ready_fd, read_buf, ERTS_SYS_READ_BUF_SZ);
-	if (res < 0) {  
-	    if ((errno != EINTR) && (errno != ERRNO_BLOCK))
-		port_inp_failure(port_num, ready_fd, res);
-	}
-	else if (res == 0) {     	/* eof */
-	    port_inp_failure(port_num, ready_fd, res);
-	} 
-	else if (res < packet_bytes - fd_data[ready_fd].psz) { 
-	    memcpy(fd_data[ready_fd].pbuf+fd_data[ready_fd].psz,
-		   read_buf, res);
-	    fd_data[ready_fd].psz += res;
-	}
-	else  { /* if (res >= packet_bytes) */
-	    unsigned char* cpos = read_buf;
-	    int bytes_left = res;
-
-	    while (1) {
-		int psz = fd_data[ready_fd].psz;
-		char* pbp = fd_data[ready_fd].pbuf + psz;
-
-		while(bytes_left && (psz < packet_bytes)) {
-		    *pbp++ = *cpos++;
-		    bytes_left--;
-		    psz++;
-		}
-		
-		if (psz < packet_bytes) {
-		    fd_data[ready_fd].psz = psz;
-		    break;
-		}
-		fd_data[ready_fd].psz = 0;
-
-		switch (packet_bytes) {
-		case 1: h = get_int8(fd_data[ready_fd].pbuf);  break;
-		case 2: h = get_int16(fd_data[ready_fd].pbuf); break;
-		case 4: h = get_int32(fd_data[ready_fd].pbuf); break;
-		default: ASSERT(0); return; /* -1; */
-		}
-
-		if (h <= (bytes_left)) {
-		    driver_output(port_num, (char*) cpos, h);
-		    cpos += h;
-		    bytes_left -= h;
-		    continue;
-		}
-		else {		/* The last message we got was split */
-		        char *buf = erts_alloc_fnf(ERTS_ALC_T_FD_ENTRY_BUF, h);
-		    if (!buf) {
-			errno = ENOMEM;
-			port_inp_failure(port_num, ready_fd, -1);
-		    }
-		    else {
-			erts_smp_atomic_add_nob(&sys_misc_mem_sz, h);
-			sys_memcpy(buf, cpos, bytes_left);
-			fd_data[ready_fd].buf = buf;
-			fd_data[ready_fd].sz = h;
-			fd_data[ready_fd].remain = h - bytes_left;
-			fd_data[ready_fd].cpos = buf + bytes_left;
-		    }
-		    break;
-		}
-	    }
-	}
-	erts_free(ERTS_ALC_T_SYS_READ_BUF, (void *) read_buf);
-    }
-}
-
-
-/* fd is the drv_data that is returned from the */
-/* initial start routine                        */
-/* ready_fd is the descriptor that is ready to read */
-
-static void ready_output(ErlDrvData e, ErlDrvEvent ready_fd)
-{
-    int fd = (int)(long)e;
-    ErlDrvPort ix = driver_data[fd].port_num;
-    int n;
-    struct iovec* iv;
-    int vsize;
-    
-
-    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;
-    if ((n = writev(ready_fd, iv, vsize)) > 0) {
-	if (driver_deq(ix, n) == 0)
-	    set_busy_port(ix, 0);
-    }
-    else if (n < 0) {
-	if (errno == ERRNO_BLOCK || errno == EINTR)
-	    return; /* 0; */
-	else {
-	    int res = errno;
-	    driver_select(ix, ready_fd, ERL_DRV_WRITE, 0);
-	    driver_failure_posix(ix, res);
-	    return; /* -1; */
-	}
-    }
-    return; /* 0; */
-}
-
-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            err = 0;
-    /* much of this code is stolen from efile_drv:invoke_writev */
-    driver_pdl_lock(dd->blocking->pdl);
-    iov0 = driver_peekq(dd->port_num, &iovlen);
-    iovlen = iovlen < MAXIOV ? iovlen : MAXIOV;
-    iov = erts_alloc_fnf(ERTS_ALC_T_SYS_WRITE_BUF,
-                         sizeof(SysIOVec)*iovlen);
-    if (!iov) {
-        res = -1;
-        err = ENOMEM;
-        driver_pdl_unlock(dd->blocking->pdl);
-    } else {
-        memcpy(iov,iov0,iovlen*sizeof(SysIOVec));
-        driver_pdl_unlock(dd->blocking->pdl);
-
-        do {
-            res = writev(dd->ofd, iov, iovlen);
-        } while (res < 0 && errno == EINTR);
-        if (res < 0)
-            err = errno;
-
-        erts_free(ERTS_ALC_T_SYS_WRITE_BUF, iov);
-    }
-    dd->blocking->res = res;
-    dd->blocking->err = err;
-}
-
-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) {
-        if (dd->blocking->err == ERRNO_BLOCK) {
-            set_busy_port(port_num, 1);
-            /* still data left to write in queue */
-            driver_async(port_num, &dd->blocking->pkey, fd_async, dd, NULL);
-        } else
-            driver_failure_posix(port_num, dd->blocking->err);
-        return; /* -1; */
-    }
-    return; /* 0; */
-}
-
-#endif
-
 void erts_do_break_handling(void)
 {
     struct termios temp_mode;
@@ -2738,10 +1011,6 @@ erts_sys_unsetenv(char *key)
 void
 sys_init_io(void)
 {
-    fd_data = (struct fd_data *)
-	erts_alloc(ERTS_ALC_T_FD_TAB, max_files * sizeof(struct fd_data));
-    erts_smp_atomic_add_nob(&sys_misc_mem_sz,
-			    max_files * sizeof(struct fd_data));
 }
 
 #if (0) /* unused? */
@@ -2935,179 +1204,6 @@ erl_debug(char* fmt, ...)
 
 #endif /* DEBUG */
 
-static ERTS_INLINE void
-report_exit_status(ErtsSysReportExit *rep, int status)
-{
-    Port *pp;
-#ifdef ERTS_SMP
-    CHLD_STAT_UNLOCK;
-    pp = erts_thr_id2port_sflgs(rep->port,
-				ERTS_PORT_SFLGS_INVALID_DRIVER_LOOKUP);
-    CHLD_STAT_LOCK;
-#else
-    pp = erts_id2port_sflgs(rep->port,
-			    NULL,
-			    0,
-			    ERTS_PORT_SFLGS_INVALID_DRIVER_LOOKUP);
-#endif
-    if (pp) {
-	if (rep->ifd >= 0) {
-	    driver_data[rep->ifd].alive = 0;
-	    driver_data[rep->ifd].status = status;
-	    (void) driver_select(ERTS_Port2ErlDrvPort(pp),
-				 rep->ifd,
-				 (ERL_DRV_READ|ERL_DRV_USE),
-				 1);
-	}
-	if (rep->ofd >= 0) {
-	    driver_data[rep->ofd].alive = 0;
-	    driver_data[rep->ofd].status = status;
-	    (void) driver_select(ERTS_Port2ErlDrvPort(pp),
-				 rep->ofd,
-				 (ERL_DRV_WRITE|ERL_DRV_USE),
-				 1);
-	}
-#ifdef ERTS_SMP
-	erts_thr_port_release(pp);
-#else
-	erts_port_release(pp);
-#endif
-    }
-    erts_free(ERTS_ALC_T_PRT_REP_EXIT, rep);
-}
-
-#if !CHLDWTHR  /* ---------------------------------------------------------- */
-
-#define ERTS_REPORT_EXIT_STATUS report_exit_status
-
-static int check_children(void)
-{
-    int res = 0;
-    int pid;
-    int status;
-
-#ifndef ERTS_SMP
-    if (children_died)
-#endif
-    {
-	sys_sigblock(SIGCHLD);
-	CHLD_STAT_LOCK;
-	while ((pid = waitpid(-1, &status, WNOHANG)) > 0)
-	    note_child_death(pid, status);
-#ifndef ERTS_SMP
-	children_died = 0;
-#endif
-	CHLD_STAT_UNLOCK;
-	sys_sigrelease(SIGCHLD);
-	res = 1;
-    }
-    return res;
-}
-
-#ifdef ERTS_SMP
-
-void
-erts_check_children(void)
-{
-    (void) check_children();
-}
-
-#endif
-
-#elif CHLDWTHR && defined(ERTS_SMP) /* ------------------------------------- */
-
-#define ERTS_REPORT_EXIT_STATUS report_exit_status
-
-#define check_children() (0)
-
-
-#else /* CHLDWTHR && !defined(ERTS_SMP) ------------------------------------ */
-
-#define ERTS_REPORT_EXIT_STATUS initiate_report_exit_status
-
-static ERTS_INLINE void
-initiate_report_exit_status(ErtsSysReportExit *rep, int status)
-{
-    rep->next = report_exit_transit_list;
-    rep->status = status;
-    report_exit_transit_list = rep;
-    erts_sys_schedule_interrupt(1);
-}
-
-static int check_children(void)
-{
-    int res;
-    ErtsSysReportExit *rep;
-    CHLD_STAT_LOCK;
-    rep = report_exit_transit_list;
-    res = rep != NULL;
-    while (rep) {
-	ErtsSysReportExit *curr_rep = rep;
-	rep = rep->next;
-	report_exit_status(curr_rep, curr_rep->status);
-    }
-    report_exit_transit_list = NULL;
-    CHLD_STAT_UNLOCK;
-    return res;
-}
-
-#endif /* ------------------------------------------------------------------ */
-
-static void note_child_death(int pid, int status)
-{
-    ErtsSysReportExit **repp = &report_exit_list;
-    ErtsSysReportExit *rep = report_exit_list;
-
-    while (rep) {
-	if (pid == rep->pid) {
-	    *repp = rep->next;
-	    ERTS_REPORT_EXIT_STATUS(rep, status);
-	    break;
-	}
-	repp = &rep->next;
-	rep = rep->next;
-    }
-}
-
-#if CHLDWTHR
-
-static void *
-child_waiter(void *unused)
-{
-  int pid;
-  int status;
-
-#ifdef ERTS_ENABLE_LOCK_CHECK
-  erts_lc_set_thread_name("child waiter");
-#endif
-
-  while(1) {
-#ifdef DEBUG
-      int waitpid_errno;
-#endif
-      pid = waitpid(-1, &status, 0);
-#ifdef DEBUG
-      waitpid_errno = errno;
-#endif
-      CHLD_STAT_LOCK;
-      if (pid < 0) {
-	  ASSERT(waitpid_errno == ECHILD);
-      }
-      else {
-	  children_alive--;
-	  ASSERT(children_alive >= 0);
-	  note_child_death(pid, status);
-      }
-      while (!children_alive)
-	  CHLD_STAT_WAIT; /* Wait for children to wait on... :) */
-      CHLD_STAT_UNLOCK;
-  }
-
-  return NULL;
-}
-
-#endif
-
 /*
  * Called from schedule() when it runs out of runnable processes,
  * or when Erlang code has performed INPUT_REDUCTIONS reduction
@@ -3116,13 +1212,8 @@ child_waiter(void *unused)
 void
 erl_sys_schedule(int runnable)
 {
-#ifdef ERTS_SMP
     ERTS_CHK_IO(!runnable);
-#else
-    ERTS_CHK_IO(runnable ? 0 : !check_children());
-#endif
     ERTS_SMP_LC_ASSERT(!erts_thr_progress_is_blocking());
-    (void) check_children();
 }
 
 
@@ -3150,10 +1241,6 @@ smp_sig_notify(char c)
 static void *
 signal_dispatcher_thread_func(void *unused)
 {
-#if !CHLDWTHR
-    int initialized = 0;
-    int notify_check_children = 0;
-#endif
 #ifdef ERTS_ENABLE_LOCK_CHECK
     erts_lc_set_thread_name("signal_dispatcher");
 #endif
@@ -3191,19 +1278,7 @@ signal_dispatcher_thread_func(void *unused)
 	     */
 	    switch (buf[i]) {
 	    case 0: /* Emulator initialized */
-#if !CHLDWTHR
-		initialized = 1;
-		if (!notify_check_children)
-#endif
-		    break;
-#if !CHLDWTHR
-	    case 'C': /* SIGCHLD */
-		if (initialized)
-		    erts_smp_notify_check_children_needed();
-		else
-		    notify_check_children = 1;
-		break;
-#endif
+                break;
 	    case 'I': /* SIGINT */
 		break_requested();
 		break;
diff --git a/erts/emulator/sys/unix/sys_drivers.c b/erts/emulator/sys/unix/sys_drivers.c
new file mode 100644
index 0000000000..82fea1b330
--- /dev/null
+++ b/erts/emulator/sys/unix/sys_drivers.c
@@ -0,0 +1,1862 @@
+/*
+ * %CopyrightBegin%
+ *
+ * Copyright Ericsson AB 1996-2014. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ * %CopyrightEnd%
+ */
+
+#ifdef HAVE_CONFIG_H
+#  include "config.h"
+#endif
+
+#ifdef ISC32
+#define _POSIX_SOURCE
+#define _XOPEN_SOURCE
+#endif
+
+#include <sys/times.h>		/* ! */
+#include <time.h>
+#include <signal.h>
+#include <sys/wait.h>
+#include <sys/uio.h>
+#include <termios.h>
+#include <ctype.h>
+#include <sys/utsname.h>
+#include <sys/select.h>
+#include <arpa/inet.h>
+
+#ifdef ISC32
+#include <sys/bsdtypes.h>
+#endif
+
+#include <termios.h>
+#ifdef HAVE_FCNTL_H
+#include <fcntl.h>
+#endif
+#ifdef HAVE_SYS_IOCTL_H
+#include <sys/ioctl.h>
+#endif
+
+#define WANT_NONBLOCKING    /* must define this to pull in defs from sys.h */
+#include "sys.h"
+
+#ifdef USE_THREADS
+#include "erl_threads.h"
+#endif
+
+extern char **environ;
+extern erts_smp_rwmtx_t environ_rwmtx;
+
+extern erts_smp_atomic_t sys_misc_mem_sz;
+
+static Eterm forker_port;
+
+#define MAX_VSIZE 16		/* Max number of entries allowed in an I/O
+				 * vector sock_sendv().
+				 */
+/*
+ * Don't need global.h, but erl_cpu_topology.h won't compile otherwise
+ */
+#include "global.h"
+#include "erl_cpu_topology.h"
+
+#include "erl_sys_driver.h"
+#include "sys_uds.h"
+
+#include "erl_child_setup.h"
+
+#if defined IOV_MAX
+#define MAXIOV IOV_MAX
+#elif defined UIO_MAXIOV
+#define MAXIOV UIO_MAXIOV
+#else
+#define MAXIOV 16
+#endif
+
+#ifdef USE_THREADS
+#  define FDBLOCK 1
+#else
+#  define FDBLOCK 0
+#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;
+
+typedef struct fd_data {
+    int   fd;
+    char  pbuf[4];   /* hold partial packet bytes */
+    int   psz;       /* size of pbuf */
+    char  *buf;
+    char  *cpos;
+    int   sz;
+    int   remain;  /* for input on fd */
+} ErtsSysFdData;
+
+typedef struct driver_data {
+    ErlDrvPort port_num;
+    ErtsSysFdData *ofd;
+    ErtsSysFdData *ifd;
+    int packet_bytes;
+    int pid;
+    int alive;
+    int status;
+    int terminating;
+    ErtsSysBlocking *blocking;
+} ErtsSysDriverData;
+
+#define DIR_SEPARATOR_CHAR    '/'
+
+#if defined(__ANDROID__)
+#define SHELL "/system/bin/sh"
+#else
+#define SHELL "/bin/sh"
+#endif /* __ANDROID__ */
+
+#if defined(DEBUG)
+#define ERL_BUILD_TYPE_MARKER ".debug"
+#elif defined(PURIFY)
+#define ERL_BUILD_TYPE_MARKER ".purify"
+#elif defined(QUANTIFY)
+#define ERL_BUILD_TYPE_MARKER ".quantify"
+#elif defined(PURECOV)
+#define ERL_BUILD_TYPE_MARKER ".purecov"
+#elif defined(VALGRIND)
+#define ERL_BUILD_TYPE_MARKER ".valgrind"
+#else /* opt */
+#define ERL_BUILD_TYPE_MARKER
+#endif
+
+#ifdef DEBUG
+#define close(fd) do { int res = close(fd); ASSERT(res > -1); } while(0)
+#endif
+
+#define CHILD_SETUP_PROG_NAME	"erl_child_setup" ERL_BUILD_TYPE_MARKER
+
+// #define HARD_DEBUG
+#ifdef HARD_DEBUG
+#define driver_select(port_num, fd, flags, onoff)                       \
+    do {                                                                \
+        if (((flags) & ERL_DRV_READ) && onoff)                          \
+            fprintf(stderr,"%010d %p: read select %d\r\n", __LINE__, port_num, (int)fd); \
+        if (((flags) & ERL_DRV_WRITE) && onoff)                         \
+            fprintf(stderr,"%010d %p: writ select %d\r\n", __LINE__, port_num, (int)fd); \
+        if (((flags) & ERL_DRV_READ) && !onoff)                          \
+            fprintf(stderr,"%010d %p: read unsele %d\r\n", __LINE__, port_num, (int)fd); \
+        if (((flags) & ERL_DRV_WRITE) && !onoff)                         \
+            fprintf(stderr,"%010d %p: writ unsele %d\r\n", __LINE__, port_num, (int)fd); \
+        driver_select_nkp(port_num, fd, flags, onoff);                  \
+    } while(0)
+#endif
+
+/*
+ * Decreasing the size of it below 16384 is not allowed.
+ */
+
+#define ERTS_SYS_READ_BUF_SZ (64*1024)
+
+/* I. Initialization */
+
+void
+erl_sys_late_init(void)
+{
+    SysDriverOpts opts;
+#ifdef ERTS_SMP
+    Port *port;
+#endif
+
+    sys_signal(SIGPIPE, SIG_IGN); /* Ignore - we'll handle the write failure */
+
+    opts.packet_bytes = 0;
+    opts.use_stdio = 1;
+    opts.redir_stderr = 0;
+    opts.read_write = 0;
+    opts.hide_window = 0;
+    opts.wd = NULL;
+    opts.envir = NULL;
+    opts.exit_status = 0;
+    opts.overlapped_io = 0;
+    opts.spawn_type = ERTS_SPAWN_ANY;
+    opts.argv = NULL;
+    opts.parallelism = erts_port_parallelism;
+
+#ifdef ERTS_SMP
+    port =
+#endif
+        erts_open_driver(&forker_driver, make_internal_pid(0), "forker", &opts, NULL, NULL);
+#ifdef ERTS_SMP
+    erts_mtx_unlock(port->lock);
+#endif
+}
+
+/* II. Prototypes */
+
+/* II.I Spawn prototypes */
+static ErlDrvData spawn_start(ErlDrvPort, char*, SysDriverOpts*);
+static ErlDrvSSizeT spawn_control(ErlDrvData, unsigned int, char *,
+                                  ErlDrvSizeT, char **, ErlDrvSizeT);
+
+/* II.II Vanilla prototypes */
+static ErlDrvData vanilla_start(ErlDrvPort, char*, SysDriverOpts*);
+
+
+/* II.III FD prototypes */
+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 void fd_stop(ErlDrvData);
+static void fd_flush(ErlDrvData);
+
+/* II.IV Common prototypes */
+static void stop(ErlDrvData);
+static void ready_input(ErlDrvData, ErlDrvEvent);
+static void ready_output(ErlDrvData, ErlDrvEvent);
+static void output(ErlDrvData, char*, ErlDrvSizeT);
+static void outputv(ErlDrvData, ErlIOVec*);
+static void stop_select(ErlDrvEvent, void*);
+
+/* II.V Forker prototypes */
+static ErlDrvData forker_start(ErlDrvPort, char*, SysDriverOpts*);
+static void forker_stop(ErlDrvData);
+static void forker_ready_input(ErlDrvData, ErlDrvEvent);
+static void forker_ready_output(ErlDrvData, ErlDrvEvent);
+static ErlDrvSSizeT forker_control(ErlDrvData, unsigned int, char *,
+                                   ErlDrvSizeT, char **, ErlDrvSizeT);
+
+/* III Driver entries */
+
+/* III.I The spawn driver */
+struct erl_drv_entry spawn_driver_entry = {
+    NULL,
+    spawn_start,
+    stop,
+    output,
+    ready_input,
+    ready_output,
+    "spawn",
+    NULL,
+    NULL,
+    spawn_control,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    ERL_DRV_EXTENDED_MARKER,
+    ERL_DRV_EXTENDED_MAJOR_VERSION,
+    ERL_DRV_EXTENDED_MINOR_VERSION,
+    ERL_DRV_FLAG_USE_PORT_LOCKING | ERL_DRV_FLAG_USE_INIT_ACK,
+    NULL, NULL,
+    stop_select
+};
+
+/* III.II The fd driver */
+struct erl_drv_entry fd_driver_entry = {
+    NULL,
+    fd_start,
+    fd_stop,
+    output,
+    ready_input,
+    ready_output,
+    "fd",
+    NULL,
+    NULL,
+    fd_control,
+    NULL,
+    outputv,
+#if FDBLOCK
+    fd_ready_async, /* ready_async */
+#else
+    NULL,
+#endif
+    fd_flush, /* flush */
+    NULL, /* call */
+    NULL, /* event */
+    ERL_DRV_EXTENDED_MARKER,
+    ERL_DRV_EXTENDED_MAJOR_VERSION,
+    ERL_DRV_EXTENDED_MINOR_VERSION,
+    0, /* ERL_DRV_FLAGs */
+    NULL, /* handle2 */
+    NULL, /* process_exit */
+    stop_select
+};
+
+/* III.III The vanilla driver */
+struct erl_drv_entry vanilla_driver_entry = {
+    NULL,
+    vanilla_start,
+    stop,
+    output,
+    ready_input,
+    ready_output,
+    "vanilla",
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL, /* flush */
+    NULL, /* call */
+    NULL, /* event */
+    ERL_DRV_EXTENDED_MARKER,
+    ERL_DRV_EXTENDED_MAJOR_VERSION,
+    ERL_DRV_EXTENDED_MINOR_VERSION,
+    0, /* ERL_DRV_FLAGs */
+    NULL, /* handle2 */
+    NULL, /* process_exit */
+    stop_select
+};
+
+/* III.III The forker driver */
+struct erl_drv_entry forker_driver_entry = {
+    NULL,
+    forker_start,
+    forker_stop,
+    NULL,
+    forker_ready_input,
+    forker_ready_output,
+    "spawn_forker",
+    NULL,
+    NULL,
+    forker_control,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    ERL_DRV_EXTENDED_MARKER,
+    ERL_DRV_EXTENDED_MAJOR_VERSION,
+    ERL_DRV_EXTENDED_MINOR_VERSION,
+    0,
+    NULL, NULL,
+    stop_select
+};
+
+/* Untility functions */
+
+static int set_blocking_data(ErtsSysDriverData *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 void init_fd_data(ErtsSysFdData *fd_data, int fd)
+{
+    fd_data->fd = fd;
+    fd_data->buf = NULL;
+    fd_data->cpos = NULL;
+    fd_data->remain = 0;
+    fd_data->sz = 0;
+    fd_data->psz = 0;
+}
+
+static ErtsSysDriverData *
+create_driver_data(ErlDrvPort port_num,
+                   int ifd,
+                   int ofd,
+                   int packet_bytes,
+                   int read_write,
+                   int exit_status,
+                   int pid,
+                   int is_blocking)
+{
+    Port *prt;
+    ErtsSysDriverData *driver_data;
+    char *data;
+    int size = sizeof(ErtsSysDriverData);
+
+    if (read_write & DO_READ)
+        size += sizeof(ErtsSysFdData);
+
+    if ((read_write & DO_WRITE) &&
+        ((ifd != ofd || ofd == -1) || !(read_write & DO_READ)))
+        size += sizeof(ErtsSysFdData);
+
+    data = erts_alloc(ERTS_ALC_T_DRV_TAB,size);
+    erts_smp_atomic_add_nob(&sys_misc_mem_sz, size);
+
+    driver_data = (ErtsSysDriverData*)data;
+    data += sizeof(*driver_data);
+
+    prt = erts_drvport2port(port_num);
+    if (prt != ERTS_INVALID_ERL_DRV_PORT)
+	prt->os_pid = pid;
+
+    driver_data->packet_bytes = packet_bytes;
+    driver_data->port_num = port_num;
+    driver_data->pid = pid;
+    driver_data->alive = exit_status ? 1 : 0;
+    driver_data->status = 0;
+    driver_data->terminating = 0;
+    driver_data->blocking = NULL;
+
+    if (read_write & DO_READ) {
+        driver_data->ifd = (ErtsSysFdData*)data;
+        data += sizeof(*driver_data->ifd);
+        init_fd_data(driver_data->ifd, ifd);
+        driver_select(port_num, ifd, (ERL_DRV_READ|ERL_DRV_USE), 1);
+    } else {
+        driver_data->ifd = NULL;
+    }
+
+    if (read_write & DO_WRITE) {
+        if (ofd != -1 && ifd == ofd && read_write & DO_READ) {
+            /* This is for when ifd and ofd are the same fd */
+            driver_data->ofd = driver_data->ifd;
+        } else {
+            driver_data->ofd = (ErtsSysFdData*)data;
+            data += sizeof(*driver_data->ofd);
+            init_fd_data(driver_data->ofd, ofd);
+        }
+        if (is_blocking && FDBLOCK)
+            if (!set_blocking_data(driver_data)) {
+                erts_free(ERTS_ALC_T_DRV_TAB, driver_data);
+                return NULL;
+            }
+    } else {
+        driver_data->ofd = NULL;
+    }
+
+    return driver_data;
+}
+
+/* Spawn driver */
+
+static void close_pipes(int ifd[2], int ofd[2])
+{
+    close(ifd[0]);
+    close(ifd[1]);
+    close(ofd[0]);
+    close(ofd[1]);
+}
+
+static char **build_unix_environment(char *block)
+{
+    int i;
+    int j;
+    int len;
+    char *cp;
+    char **cpp;
+    char** old_env;
+    
+    ERTS_SMP_LC_ASSERT(erts_smp_lc_rwmtx_is_rlocked(&environ_rwmtx));
+
+    cp = block;
+    len = 0;
+    while (*cp != '\0') {
+	cp += strlen(cp) + 1;
+	len++;
+    }
+    old_env = environ;
+    while (*old_env++ != NULL) {
+	len++;
+    }
+    
+    cpp = (char **) erts_alloc_fnf(ERTS_ALC_T_ENVIRONMENT,
+				   sizeof(char *) * (len+1));
+    if (cpp == NULL) {
+	return NULL;
+    }
+
+    cp = block;
+    len = 0;
+    while (*cp != '\0') {
+	cpp[len] = cp;
+	cp += strlen(cp) + 1;
+	len++;
+    }
+    
+    i = len;
+    for (old_env = environ; *old_env; old_env++) {
+	char* old = *old_env;
+
+	for (j = 0; j < len; j++) {
+	    char *s, *t;
+
+	    /* check if cpp[j] equals old
+	       before the = sign,
+	       i.e.
+	       "TMPDIR=/tmp/" */
+	    s = cpp[j];
+	    t = old;
+	    while (*s == *t && *s != '=') {
+		s++, t++;
+	    }
+	    if (*s == '=' && *t == '=') {
+		break;
+	    }
+	}
+
+	if (j == len) {		/* New version not found */
+	    cpp[len++] = old;
+	}
+    }
+
+    for (j = 0; j < i; ) {
+        size_t last = strlen(cpp[j])-1;
+	if (cpp[j][last] == '=' && strchr(cpp[j], '=') == cpp[j]+last) {
+	    cpp[j] = cpp[--len];
+	    if (len < i) {
+		i--;
+	    } else {
+		j++;
+	    }
+	}
+	else {
+	    j++;
+	}
+    }
+
+    cpp[len] = NULL;
+    return cpp;
+}
+
+static ErlDrvData spawn_start(ErlDrvPort port_num, char* name,
+                              SysDriverOpts* opts)
+{
+#define CMD_LINE_PREFIX_STR "exec "
+#define CMD_LINE_PREFIX_STR_SZ (sizeof(CMD_LINE_PREFIX_STR) - 1)
+
+    int len;
+    char **new_environ;
+    ErtsSysDriverData *dd;
+    char *cmd_line;
+    char wd_buff[MAXPATHLEN+1];
+    char *wd;
+    int ifd[2], ofd[2], stderrfd;
+
+    if (pipe(ifd) < 0) return ERL_DRV_ERROR_ERRNO;
+    errno = EMFILE;		/* default for next three conditions */
+    if (ifd[0] >= sys_max_files() || pipe(ofd) < 0) {
+        close(ifd[0]);
+        close(ifd[1]);
+        return ERL_DRV_ERROR_ERRNO;
+    }
+    if (ofd[1] >= sys_max_files()) {
+        close_pipes(ifd, ofd);
+        errno = EMFILE;
+        return ERL_DRV_ERROR_ERRNO;
+    }
+
+    SET_NONBLOCKING(ifd[0]);
+    SET_NONBLOCKING(ofd[1]);
+
+    stderrfd = opts->redir_stderr ? ifd[1] : dup(2);
+
+    if (stderrfd >= sys_max_files() || stderrfd < 0) {
+        close_pipes(ifd, ofd);
+        if (stderrfd > -1)
+            close(stderrfd);
+        return ERL_DRV_ERROR_ERRNO;
+    }
+
+    if (opts->spawn_type == ERTS_SPAWN_EXECUTABLE) {
+	/* started with spawn_executable, not with spawn */
+	len = strlen(name);
+	cmd_line = (char *) erts_alloc_fnf(ERTS_ALC_T_TMP, len + 1);
+	if (!cmd_line) {
+            close_pipes(ifd, ofd);
+	    errno = ENOMEM;
+	    return ERL_DRV_ERROR_ERRNO;
+	}
+	memcpy((void *) cmd_line,(void *) name, len);
+	cmd_line[len] = '\0';
+	len = len + 1;
+	if (access(cmd_line,X_OK) != 0) {
+	    int save_errno = errno;
+	    erts_free(ERTS_ALC_T_TMP, cmd_line);
+            close_pipes(ifd, ofd);
+	    errno = save_errno;
+	    return ERL_DRV_ERROR_ERRNO;
+	}
+    } else {
+	/* make the string suitable for giving to "sh" */
+	len = strlen(name);
+	cmd_line = (char *) erts_alloc_fnf(ERTS_ALC_T_TMP,
+					   CMD_LINE_PREFIX_STR_SZ + len + 1);
+	if (!cmd_line) {
+            close_pipes(ifd, ofd);
+	    errno = ENOMEM;
+	    return ERL_DRV_ERROR_ERRNO;
+	}
+	memcpy((void *) cmd_line,
+	       (void *) CMD_LINE_PREFIX_STR,
+	       CMD_LINE_PREFIX_STR_SZ);
+	memcpy((void *) (cmd_line + CMD_LINE_PREFIX_STR_SZ), (void *) name, len);
+	cmd_line[CMD_LINE_PREFIX_STR_SZ + len] = '\0';
+	len = CMD_LINE_PREFIX_STR_SZ + len + 1;
+    }
+
+    erts_smp_rwmtx_rlock(&environ_rwmtx);
+
+    if (opts->envir == NULL) {
+	new_environ = environ;
+    } else if ((new_environ = build_unix_environment(opts->envir)) == NULL) {
+	erts_smp_rwmtx_runlock(&environ_rwmtx);
+        close_pipes(ifd, ofd);
+	erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
+	errno = ENOMEM;
+	return ERL_DRV_ERROR_ERRNO;
+    }
+
+    if (opts->wd == NULL) {
+        if ((wd = getcwd(wd_buff, MAXPATHLEN+1)) == NULL) {
+            /* on some OSs this call opens a fd in the
+               background which means that this can
+               return EMFILE */
+            int err = errno;
+            close_pipes(ifd, ofd);
+            erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
+            if (new_environ != environ)
+                erts_free(ERTS_ALC_T_ENVIRONMENT, (void *) new_environ);
+            erts_smp_rwmtx_runlock(&environ_rwmtx);
+            errno = err;
+            return ERL_DRV_ERROR_ERRNO;
+        }
+    } else {
+        wd = opts->wd;
+    }
+
+    {
+        struct iovec *io_vector;
+        int iov_len = 5;
+        char nullbuff[] = "\0";
+        int j, i = 0, res;
+        Sint32 buffsz = 0, env_len = 0, argv_len = 0,
+            flags = (opts->use_stdio ? FORKER_FLAG_USE_STDIO : 0)
+            | (opts->exit_status ? FORKER_FLAG_EXIT_STATUS : 0)
+            | (opts->read_write & DO_READ ? FORKER_FLAG_DO_READ : 0)
+            | (opts->read_write & DO_WRITE ? FORKER_FLAG_DO_WRITE : 0);
+
+        /* count number of elements in environment */
+        while(new_environ[env_len] != NULL)
+            env_len++;
+        iov_len += 1 + env_len; /* num envs including size int */
+
+        /* count number of element in argument list */
+        if (opts->spawn_type == ERTS_SPAWN_EXECUTABLE) {
+            if (opts->argv != NULL) {
+                while(opts->argv[argv_len] != NULL)
+                    argv_len++;
+            } else {
+                argv_len++;
+            }
+            iov_len += 1 + argv_len; /* num argvs including size int */
+        }
+
+        io_vector = erts_alloc_fnf(ERTS_ALC_T_TMP, sizeof(struct iovec) * iov_len);
+
+        if (!io_vector) {
+            close_pipes(ifd, ofd);
+            erts_smp_rwmtx_runlock(&environ_rwmtx);
+            erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
+            if (new_environ != environ)
+                erts_free(ERTS_ALC_T_ENVIRONMENT, (void *) new_environ);
+            errno = ENOMEM;
+            return ERL_DRV_ERROR_ERRNO;
+        }
+
+        io_vector[i].iov_base = (void*)&buffsz;
+        io_vector[i++].iov_len = sizeof(buffsz);
+
+        io_vector[i].iov_base = (void*)&flags;
+        flags = htonl(flags);
+        io_vector[i++].iov_len = sizeof(flags);
+        buffsz += sizeof(flags);
+
+        io_vector[i].iov_base = cmd_line;
+        io_vector[i++].iov_len = len;
+        buffsz += len;
+
+        io_vector[i].iov_base = wd;
+        io_vector[i].iov_len = strlen(io_vector[i].iov_base) + 1;
+        buffsz += io_vector[i++].iov_len;
+
+        io_vector[i].iov_base = nullbuff;
+        io_vector[i++].iov_len = 1;
+        buffsz += io_vector[i-1].iov_len;
+
+        io_vector[i].iov_base = (void*)&env_len;
+        env_len = htonl(env_len);
+        io_vector[i++].iov_len = sizeof(env_len);
+        buffsz += io_vector[i-1].iov_len;
+
+        for (j = 0; new_environ[j] != NULL; j++) {
+            io_vector[i].iov_base = new_environ[j];
+            io_vector[i++].iov_len = strlen(new_environ[j]) + 1;
+            buffsz += io_vector[i-1].iov_len;
+        }
+
+        /* only append arguments if this was a spawn_executable */
+        if (opts->spawn_type == ERTS_SPAWN_EXECUTABLE) {
+
+            io_vector[i].iov_base = (void*)&argv_len;
+            argv_len = htonl(argv_len);
+            io_vector[i++].iov_len = sizeof(argv_len);
+            buffsz += io_vector[i-1].iov_len;
+
+            if (opts->argv) {
+                /* If there are arguments we copy in the references to
+                   them into the iov */
+                for (j = 0; opts->argv[j]; j++) {
+                    if (opts->argv[j] == erts_default_arg0)
+                        io_vector[i].iov_base = cmd_line;
+                    else
+                        io_vector[i].iov_base = opts->argv[j];
+                    io_vector[i].iov_len = strlen(io_vector[i].iov_base) + 1;
+                    buffsz += io_vector[i++].iov_len;
+                }
+            } else {
+                io_vector[i].iov_base = cmd_line;
+                io_vector[i].iov_len = strlen(io_vector[i].iov_base) + 1;
+                buffsz += io_vector[i++].iov_len;
+            }
+        }
+
+        /* we send the request to do the fork */
+        if ((res = writev(ofd[1], io_vector, iov_len > MAXIOV ? MAXIOV : iov_len)) < 0) {
+            if (errno == ERRNO_BLOCK) {
+                res = 0;
+            } else {
+                int err = errno;
+                close_pipes(ifd, ofd);
+                erts_free(ERTS_ALC_T_TMP, io_vector);
+                if (new_environ != environ)
+                    erts_free(ERTS_ALC_T_ENVIRONMENT, (void *) new_environ);
+                erts_smp_rwmtx_runlock(&environ_rwmtx);
+                erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
+                errno = err;
+                return ERL_DRV_ERROR_ERRNO;
+            }
+        }
+
+        if (res < buffsz) {
+            /* we only wrote part of the command payload. Enqueue the rest. */
+            for (i = 0; i < iov_len; i++) {
+                driver_enq(port_num, io_vector[i].iov_base, io_vector[i].iov_len);
+            }
+            driver_deq(port_num, res);
+            driver_select(port_num, ofd[1], ERL_DRV_WRITE|ERL_DRV_USE, 1);
+        }
+
+        erts_free(ERTS_ALC_T_TMP, io_vector);
+    }
+
+    erts_free(ERTS_ALC_T_TMP, (void *) cmd_line);
+
+    if (new_environ != environ)
+	erts_free(ERTS_ALC_T_ENVIRONMENT, (void *) new_environ);
+
+    erts_smp_rwmtx_runlock(&environ_rwmtx);
+
+    dd = create_driver_data(port_num, ifd[0], ofd[1], opts->packet_bytes,
+                             DO_WRITE | DO_READ, opts->exit_status,
+                             0, 0);
+
+    {
+        /* send ofd[0] + ifd[1] + stderrfd to forker port */
+        ErtsSysForkerProto *proto =
+            erts_alloc(ERTS_ALC_T_DRV_CTRL_DATA,
+                       sizeof(ErtsSysForkerProto));
+        memset(proto, 0, sizeof(ErtsSysForkerProto));
+        proto->action = ErtsSysForkerProtoAction_Start;
+        proto->u.start.fds[0] = ofd[0];
+        proto->u.start.fds[1] = ifd[1];
+        proto->u.start.fds[2] = stderrfd;
+        proto->u.start.port_id = opts->exit_status ? erts_drvport2id(port_num) : THE_NON_VALUE;
+        if (erl_drv_port_control(forker_port, 'S', (char*)proto, sizeof(*proto))) {
+            /* The forker port has been killed, we close both fd's which will
+               make open_port throw an epipe error */
+            close(ofd[0]);
+            close(ifd[1]);
+        }
+    }
+
+    /* we set these fds to negative to mark if
+       they should be closed after the handshake */
+    if (!(opts->read_write & DO_READ))
+        dd->ifd->fd *= -1;
+
+    if (!(opts->read_write & DO_WRITE))
+        dd->ofd->fd *= -1;
+
+    return (ErlDrvData)dd;
+#undef CMD_LINE_PREFIX_STR
+#undef CMD_LINE_PREFIX_STR_SZ
+}
+
+static ErlDrvSSizeT spawn_control(ErlDrvData e, unsigned int cmd, char *buf,
+                                  ErlDrvSizeT len, char **rbuf, ErlDrvSizeT rlen)
+{
+    ErtsSysDriverData *dd = (ErtsSysDriverData*)e;
+    ErtsSysForkerProto *proto = (ErtsSysForkerProto *)buf;
+
+    ASSERT(len == sizeof(*proto));
+    ASSERT(proto->action == ErtsSysForkerProtoAction_SigChld);
+
+    dd->status = proto->u.sigchld.error_number;
+    dd->alive = -1;
+
+    if (dd->ifd)
+        driver_select(dd->port_num, abs(dd->ifd->fd), ERL_DRV_READ | ERL_DRV_USE, 1);
+
+    if (dd->ofd)
+        driver_select(dd->port_num, abs(dd->ofd->fd), ERL_DRV_WRITE | ERL_DRV_USE, 1);
+
+    return 0;
+}
+
+#define FD_DEF_HEIGHT 24
+#define FD_DEF_WIDTH 80
+/* Control op */
+#define FD_CTRL_OP_GET_WINSIZE 100
+
+static int fd_get_window_size(int fd, Uint32 *width, Uint32 *height)
+{
+#ifdef TIOCGWINSZ 
+    struct winsize ws;
+    if (ioctl(fd,TIOCGWINSZ,&ws) == 0) {
+	*width = (Uint32) ws.ws_col;
+	*height = (Uint32) ws.ws_row;
+	return 0;
+    }
+#endif
+    return -1;
+}
+
+static ErlDrvSSizeT fd_control(ErlDrvData drv_data,
+			       unsigned int command,
+			       char *buf, ErlDrvSizeT len,
+			       char **rbuf, ErlDrvSizeT rlen)
+{
+    int fd = (int)(long)drv_data;
+    char resbuff[2*sizeof(Uint32)];
+    switch (command) {
+    case FD_CTRL_OP_GET_WINSIZE:
+	{
+	    Uint32 w,h;
+	    if (fd_get_window_size(fd,&w,&h)) 
+		return 0;
+	    memcpy(resbuff,&w,sizeof(Uint32));
+	    memcpy(resbuff+sizeof(Uint32),&h,sizeof(Uint32));
+	}
+	break;
+    default:
+	return 0;
+    }
+    if (rlen < 2*sizeof(Uint32)) {
+	*rbuf = driver_alloc(2*sizeof(Uint32));
+    }
+    memcpy(*rbuf,resbuff,2*sizeof(Uint32));
+    return 2*sizeof(Uint32);
+}
+
+static ErlDrvData fd_start(ErlDrvPort port_num, char* name,
+			   SysDriverOpts* opts)
+{
+    int non_blocking = 0;
+
+    if (((opts->read_write & DO_READ) && opts->ifd >= sys_max_files()) ||
+	((opts->read_write & DO_WRITE) && opts->ofd >= sys_max_files()))
+	return ERL_DRV_ERROR_GENERAL;
+
+    /*
+     * Historical:
+     *
+     * "Note about nonblocking I/O.
+     *
+     * At least on Solaris, setting the write end of a TTY to nonblocking,
+     * will set the input end to nonblocking as well (and vice-versa).
+     * If erl is run in a pipeline like this:  cat | erl
+     * the input end of the TTY will be the standard input of cat.
+     * And cat is not prepared to handle nonblocking I/O."
+     *
+     * Actually, the reason for this is not that the tty itself gets set
+     * in non-blocking mode, but that the "input end" (cat's stdin) and
+     * the "output end" (erlang's stdout) are typically the "same" file
+     * descriptor, dup()'ed from a single fd by one of this process'
+     * ancestors.
+     *
+     * The workaround for this problem used to be a rather bad kludge,
+     * interposing an extra process ("internal cat") between erlang's
+     * stdout and the original stdout, allowing erlang to set its stdout
+     * in non-blocking mode without affecting the stdin of the preceding
+     * process in the pipeline - and being a kludge, it caused all kinds
+     * of weird problems.
+     *
+     * So, this is the current logic:
+     *
+     * The only reason to set non-blocking mode on the output fd at all is
+     * if it's something that can cause a write() to block, of course,
+     * i.e. primarily if it points to a tty, socket, pipe, or fifo. 
+     *
+     * If we don't set non-blocking mode when we "should" have, and output
+     * becomes blocked, the entire runtime system will be suspended - this
+     * is normally bad of course, and can happen fairly "easily" - e.g. user
+     * hits ^S on tty - but doesn't necessarily happen.
+     * 
+     * If we do set non-blocking mode when we "shouldn't" have, the runtime
+     * system will end up seeing EOF on the input fd (due to the preceding
+     * process dying), which typically will cause the entire runtime system
+     * to terminate immediately (due to whatever erlang process is seeing
+     * the EOF taking it as a signal to halt the system). This is *very* bad.
+     * 
+     * I.e. we should take a conservative approach, and only set non-
+     * blocking mode when we a) need to, and b) are reasonably certain
+     * that it won't be a problem. And as in the example above, the problem
+     * occurs when input fd and output fd point to different "things".
+     *
+     * However, determining that they are not just the same "type" of
+     * "thing", but actually the same instance of that type of thing, is
+     * unreasonably complex in many/most cases.
+     *
+     * Also, with pipes, sockets, and fifos it's far from obvious that the
+     * user *wants* non-blocking output: If you're running erlang inside
+     * some complex pipeline, you're probably not running a real-time system
+     * that must never stop, but rather *want* it to suspend if the output
+     * channel is "full".
+     *
+     * So, the bottom line: We will only set the output fd non-blocking if
+     * it points to a tty, and either a) the input fd also points to a tty,
+     * or b) we can make sure that setting the output fd non-blocking
+     * doesn't interfere with someone else's input, via a somewhat milder
+     * kludge than the above.
+     *
+     * Also keep in mind that while this code is almost exclusively run as
+     * a result of an erlang open_port({fd,0,1}, ...), that isn't the only
+     * 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.
+     */
+
+    /* Try to figure out if we can use non-blocking writes */
+    if (opts->read_write & DO_WRITE) {
+
+	/* If we don't have a read end, all bets are off - no non-blocking. */
+	if (opts->read_write & DO_READ) {
+
+	    if (isatty(opts->ofd)) { /* output fd is a tty:-) */
+
+		if (isatty(opts->ifd)) { /* input fd is also a tty */
+
+		    /* To really do this "right", we should also check that
+		       input and output fd point to the *same* tty - but
+		       this seems like overkill; ttyname() isn't for free,
+		       and this is a very common case - and it's hard to
+		       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 */
+
+		    /* This is a "problem case", but also common (see the
+		       example above) - i.e. it makes sense to try a bit
+		       harder before giving up on non-blocking mode: Try to
+		       re-open the tty that the output fd points to, and if
+		       successful replace the original one with the "new" fd
+		       obtained this way, and set *that* one in non-blocking
+		       mode. (Yes, this is a kludge.)
+
+		       However, re-opening the tty may fail in a couple of
+		       (unusual) cases:
+
+		       1) The name of the tty (or an equivalent one, i.e.
+			  same major/minor number) can't be found, because
+			  it actually lives somewhere other than /dev (or
+			  wherever ttyname() looks for it), and isn't
+			  equivalent to any of those that do live in the
+			  "standard" place - this should be *very* unusual.
+
+		       2) Permissions on the tty don't allow us to open it -
+			  it's perfectly possible to have an fd open to an
+			  object whose permissions wouldn't allow us to open
+			  it. This is not as unusual as it sounds, one case
+			  is if the user has su'ed to someone else (not
+			  root) - we have a read/write fd open to the tty
+			  (because it has been inherited all the way down
+			  here), but we have neither read nor write
+			  permission for the tty.
+
+		       In these cases, we finally give up, and don't set the
+		       output fd in non-blocking mode. */
+
+		    char *tty;
+		    int nfd;
+
+		    if ((tty = ttyname(opts->ofd)) != NULL &&
+			(nfd = open(tty, O_WRONLY)) != -1) {
+			dup2(nfd, opts->ofd);
+			close(nfd);
+                        non_blocking = 1;
+			SET_NONBLOCKING(opts->ofd);
+		    }
+		}
+	    }
+	}
+    }
+    return (ErlDrvData)create_driver_data(port_num, opts->ifd, opts->ofd,
+                                          opts->packet_bytes,
+                                          opts->read_write, 0, -1,
+                                          !non_blocking);
+}
+
+static void clear_fd_data(ErtsSysFdData *fdd)
+{
+    if (fdd->sz > 0) {
+	erts_free(ERTS_ALC_T_FD_ENTRY_BUF, (void *) fdd->buf);
+	ASSERT(erts_smp_atomic_read_nob(&sys_misc_mem_sz) >= fdd->sz);
+	erts_smp_atomic_add_nob(&sys_misc_mem_sz, -1*fdd->sz);
+    }
+    fdd->buf = NULL;
+    fdd->sz = 0;
+    fdd->remain = 0;
+    fdd->cpos = NULL;
+    fdd->psz = 0;
+}
+
+static void nbio_stop_fd(ErlDrvPort prt, ErtsSysFdData *fdd)
+{
+    driver_select(prt, abs(fdd->fd), DO_READ|DO_WRITE, 0);
+    clear_fd_data(fdd);
+    SET_BLOCKING(abs(fdd->fd));
+
+}
+
+static void fd_stop(ErlDrvData ev)  /* Does not close the fds */
+{
+    ErtsSysDriverData* dd = (ErtsSysDriverData*)ev;
+    ErlDrvPort prt = dd->port_num;
+    int sz = sizeof(ErtsSysDriverData);
+
+#if FDBLOCK
+    if (dd->blocking) {
+        erts_free(ERTS_ALC_T_SYS_BLOCKING, dd->blocking);
+        dd->blocking = NULL;
+        sz += sizeof(ErtsSysBlocking);
+    }
+#endif
+
+    if (dd->ifd) {
+        sz += sizeof(ErtsSysFdData);
+        nbio_stop_fd(prt, dd->ifd);
+    }
+    if (dd->ofd && dd->ofd != dd->ifd) {
+        sz += sizeof(ErtsSysFdData);
+        nbio_stop_fd(prt, dd->ofd);
+    }
+
+     erts_free(ERTS_ALC_T_DRV_TAB, dd);
+     erts_smp_atomic_add_nob(&sys_misc_mem_sz, -sz);
+}
+
+static void fd_flush(ErlDrvData ev)
+{
+    ErtsSysDriverData* dd = (ErtsSysDriverData*)ev;
+    if (!dd->terminating)
+        dd->terminating = 1;
+}
+
+static ErlDrvData vanilla_start(ErlDrvPort port_num, char* name,
+				SysDriverOpts* opts)
+{
+    int flags, fd;
+    ErlDrvData res;
+
+    flags = (opts->read_write == DO_READ ? O_RDONLY :
+	     opts->read_write == DO_WRITE ? O_WRONLY|O_CREAT|O_TRUNC :
+	     O_RDWR|O_CREAT);
+    if ((fd = open(name, flags, 0666)) < 0)
+	return ERL_DRV_ERROR_GENERAL;
+    if (fd >= sys_max_files()) {
+	close(fd);
+	return ERL_DRV_ERROR_GENERAL;
+    }
+    SET_NONBLOCKING(fd);
+
+    res = (ErlDrvData)(long)create_driver_data(port_num, fd, fd,
+                                               opts->packet_bytes,
+                                               opts->read_write, 0, -1, 0);
+    return res;
+}
+
+/* Note that driver_data[fd].ifd == fd if the port was opened for reading, */
+/* otherwise (i.e. write only) driver_data[fd].ofd = fd.  */
+
+static void stop(ErlDrvData ev)
+{
+    ErtsSysDriverData* dd = (ErtsSysDriverData*)ev;
+    ErlDrvPort prt = dd->port_num;
+
+    if (dd->ifd) {
+        nbio_stop_fd(prt, dd->ifd);
+        driver_select(prt, abs(dd->ifd->fd), ERL_DRV_USE, 0);  /* close(ifd); */
+    }
+
+    if (dd->ofd && dd->ofd != dd->ifd) {
+	nbio_stop_fd(prt, dd->ofd);
+	driver_select(prt, abs(dd->ofd->fd), ERL_DRV_USE, 0);  /* close(ofd); */
+    }
+
+    erts_free(ERTS_ALC_T_DRV_TAB, dd);
+}
+
+/* used by fd_driver */
+static void outputv(ErlDrvData e, ErlIOVec* ev)
+{
+    ErtsSysDriverData *dd = (ErtsSysDriverData*)e;
+    ErlDrvPort ix = dd->port_num;
+    int pb = dd->packet_bytes;
+    int ofd = dd->ofd ? dd->ofd->fd : -1;
+    ssize_t n;
+    ErlDrvSizeT sz;
+    char lb[4];
+    char* lbp;
+    ErlDrvSizeT len = ev->size;
+
+    /* (len > ((unsigned long)-1 >> (4-pb)*8)) */
+    /*    if (pb >= 0 && (len & (((ErlDrvSizeT)1 << (pb*8))) - 1) != len) {*/
+    if (((pb == 2) && (len > 0xffff)) || (pb == 1 && len > 0xff)) {
+	driver_failure_posix(ix, EINVAL);
+	return; /* -1; */
+    }
+    /* Handles 0 <= pb <= 4 only */
+    put_int32((Uint32) len, lb);
+    lbp = lb + (4-pb);
+
+    ev->iov[0].iov_base = lbp;
+    ev->iov[0].iov_len = pb;
+    ev->size += pb;
+
+    if (dd->blocking && FDBLOCK)
+        driver_pdl_lock(dd->blocking->pdl);
+
+    if ((sz = driver_sizeq(ix)) > 0) {
+	driver_enqv(ix, ev, 0);
+
+        if (dd->blocking && FDBLOCK)
+            driver_pdl_unlock(dd->blocking->pdl);
+
+	if (sz + ev->size >= (1 << 13))
+	    set_busy_port(ix, 1);
+    }
+    else if (!dd->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);
+	if (n == ev->size)
+	    return; /* 0;*/
+	if (n < 0) {
+	    if ((errno != EINTR) && (errno != ERRNO_BLOCK)) {
+		driver_failure_posix(ix, errno);
+		return; /* -1;*/
+	    }
+	    n = 0;
+	}
+	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(dd->blocking->pdl);
+            driver_async(ix, &dd->blocking->pkey,
+                         fd_async, dd, NULL);
+        } else {
+            driver_pdl_unlock(dd->blocking->pdl);
+        }
+    }
+#endif
+    /* return 0;*/
+}
+
+/* Used by spawn_driver and vanilla driver */
+static void output(ErlDrvData e, char* buf, ErlDrvSizeT len)
+{
+    ErtsSysDriverData *dd = (ErtsSysDriverData*)e;
+    ErlDrvPort ix = dd->port_num;
+    int pb = dd->packet_bytes;
+    int ofd = dd->ofd ? dd->ofd->fd : -1;
+    ssize_t n;
+    ErlDrvSizeT sz;
+    char lb[4];
+    char* lbp;
+    struct iovec iv[2];
+
+    /* (len > ((unsigned long)-1 >> (4-pb)*8)) */
+    if (((pb == 2) && (len > 0xffff))
+        || (pb == 1 && len > 0xff)
+        || dd->pid == 0 /* Attempt at output before port is ready */) {
+	driver_failure_posix(ix, EINVAL);
+	return; /* -1; */
+    }
+    put_int32(len, lb);
+    lbp = lb + (4-pb);
+
+    if ((sz = driver_sizeq(ix)) > 0) {
+	driver_enq(ix, lbp, pb);
+	driver_enq(ix, buf, len);
+	if (sz + len + pb >= (1 << 13))
+	    set_busy_port(ix, 1);
+    }
+    else {
+	iv[0].iov_base = lbp;
+	iv[0].iov_len = pb;  /* should work for pb=0 */
+	iv[1].iov_base = buf;
+	iv[1].iov_len = len;
+	n = writev(ofd, iv, 2);
+	if (n == pb+len)
+	    return; /* 0; */
+	if (n < 0) {
+	    if ((errno != EINTR) && (errno != ERRNO_BLOCK)) {
+		driver_failure_posix(ix, errno);
+		return; /* -1; */
+	    }
+	    n = 0;
+	}
+	if (n < pb) {
+	    driver_enq(ix, lbp+n, pb-n);
+	    driver_enq(ix, buf, len);
+	}
+	else {
+	    n -= pb;
+	    driver_enq(ix, buf+n, len-n);
+	}
+	driver_select(ix, ofd, ERL_DRV_WRITE|ERL_DRV_USE, 1);
+    }
+    return; /* 0; */
+}
+
+static int port_inp_failure(ErtsSysDriverData *dd, int res)
+				/* Result: 0 (eof) or -1 (error) */
+{
+    int err = errno;
+
+    ASSERT(res <= 0);
+    if (dd->ifd) {
+        driver_select(dd->port_num, dd->ifd->fd, ERL_DRV_READ|ERL_DRV_WRITE, 0);
+        clear_fd_data(dd->ifd);
+    }
+
+    if (dd->blocking && FDBLOCK) {
+        driver_pdl_lock(dd->blocking->pdl);
+        if (driver_sizeq(dd->port_num) > 0) {
+            driver_pdl_unlock(dd->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)
+                dd->terminating = 2;
+            else
+                dd->terminating = -err;
+            return 0;
+        }
+        driver_pdl_unlock(dd->blocking->pdl);
+    }
+
+    if (res == 0) {
+        if (dd->alive == 1) {
+            /*
+             * We have eof and want to report exit status, but the process
+             * hasn't exited yet. When it does ready_input will
+             * driver_select() this fd which will make sure that we get
+             * back here with dd->alive == -1 and dd->status set.
+             */
+            return 0;
+        }
+        else if (dd->alive == -1) {
+            int status = dd->status;
+
+            /* We need not be prepared for stopped/continued processes. */
+            if (WIFSIGNALED(status))
+                status = 128 + WTERMSIG(status);
+            else
+                status = WEXITSTATUS(status);
+            driver_report_exit(dd->port_num, status);
+        }
+       driver_failure_eof(dd->port_num);
+    } else if (dd->ifd) {
+        erl_drv_init_ack(dd->port_num, ERL_DRV_ERROR_ERRNO);
+    } else {
+	driver_failure_posix(dd->port_num, err);
+    }
+    return 0;
+}
+
+/* fd is the drv_data that is returned from the */
+/* initial start routine                        */
+/* ready_fd is the descriptor that is ready to read */
+
+static void ready_input(ErlDrvData e, ErlDrvEvent ready_fd)
+{
+    ErtsSysDriverData *dd = (ErtsSysDriverData*)e;
+    ErlDrvPort port_num;
+    int packet_bytes;
+    int res;
+    Uint h;
+
+    port_num = dd->port_num;
+    packet_bytes = dd->packet_bytes;
+
+    ASSERT(abs(dd->ifd->fd) == ready_fd);
+
+    if (dd->pid == 0) {
+        /* the pid is sent from erl_child_setup. spawn driver only. */
+        ErtsSysForkerProto proto;
+        int res;
+
+        if((res = read(ready_fd, &proto, sizeof(proto))) <= 0) {
+            /* hmm, child setup seems to have closed the pipe too early...
+               we close the port as there is not much else we can do */
+            if (res < 0 && errno == ERRNO_BLOCK)
+                return;
+            driver_select(port_num, ready_fd, ERL_DRV_READ, 0);
+            if (res == 0)
+                errno = EPIPE;
+            port_inp_failure(dd, -1);
+            return;
+        }
+
+        ASSERT(proto.action == ErtsSysForkerProtoAction_Go);
+        dd->pid = proto.u.go.os_pid;
+
+        if (dd->pid == -1) {
+            /* Setup failed! The only reason why this should happen is if
+               the fork fails. */
+            errno = proto.u.go.error_number;
+            port_inp_failure(dd, -1);
+            return;
+        }
+
+        proto.action = ErtsSysForkerProtoAction_Ack;
+
+        if (driver_sizeq(port_num) > 0) {
+            driver_enq(port_num, (char*)&proto, sizeof(proto));
+            } else {
+                if (write(abs(dd->ofd->fd), &proto, sizeof(proto)) < 0)
+                    if (errno == ERRNO_BLOCK || errno == EINTR)
+                        driver_enq(port_num, (char*)&proto, sizeof(proto));
+                /* do nothing on failure here. If the ofd is broken, then
+                   the ifd will probably also be broken and trigger
+                   a port_inp_failure */
+            }
+
+            if (dd->ifd->fd < 0) {
+                driver_select(port_num, abs(dd->ifd->fd), ERL_DRV_READ|ERL_DRV_USE, 0);
+                erts_smp_atomic_add_nob(&sys_misc_mem_sz, -sizeof(ErtsSysFdData));
+                dd->ifd = NULL;
+            }
+
+            if (dd->ofd->fd < 0  || driver_sizeq(port_num) > 0)
+                /* we select in order to close fd or write to queue,
+                   child setup will close this fd if fd < 0 */
+                driver_select(port_num, abs(dd->ofd->fd), ERL_DRV_WRITE|ERL_DRV_USE, 1);
+
+            erl_drv_set_os_pid(port_num, dd->pid);
+            erl_drv_init_ack(port_num, e);
+            return;
+    }
+
+    if (packet_bytes == 0) {
+	byte *read_buf = (byte *) erts_alloc(ERTS_ALC_T_SYS_READ_BUF,
+					     ERTS_SYS_READ_BUF_SZ);
+	res = read(ready_fd, read_buf, ERTS_SYS_READ_BUF_SZ);
+	if (res < 0) {
+	    if ((errno != EINTR) && (errno != ERRNO_BLOCK))
+		port_inp_failure(dd, res);
+	}
+	else if (res == 0)
+	    port_inp_failure(dd, res);
+	else
+	    driver_output(port_num, (char*) read_buf, res);
+	erts_free(ERTS_ALC_T_SYS_READ_BUF, (void *) read_buf);
+    }
+    else if (dd->ifd->remain > 0) { /* We try to read the remainder */
+	/* space is allocated in buf */
+	res = read(ready_fd, dd->ifd->cpos,
+		   dd->ifd->remain);
+	if (res < 0) {
+	    if ((errno != EINTR) && (errno != ERRNO_BLOCK))
+		port_inp_failure(dd, res);
+	}
+	else if (res == 0) {
+	    port_inp_failure(dd, res);
+	}
+	else if (res == dd->ifd->remain) { /* we're done  */
+	    driver_output(port_num, dd->ifd->buf,
+			  dd->ifd->sz);
+	    clear_fd_data(dd->ifd);
+	}
+	else { /*  if (res < dd->ifd->remain) */
+	    dd->ifd->cpos += res;
+	    dd->ifd->remain -= res;
+	}
+    }
+    else if (dd->ifd->remain == 0) { /* clean fd */
+	byte *read_buf = (byte *) erts_alloc(ERTS_ALC_T_SYS_READ_BUF,
+					     ERTS_SYS_READ_BUF_SZ);
+	/* We make one read attempt and see what happens */
+	res = read(ready_fd, read_buf, ERTS_SYS_READ_BUF_SZ);
+	if (res < 0) {
+	    if ((errno != EINTR) && (errno != ERRNO_BLOCK))
+		port_inp_failure(dd, res);
+	}
+	else if (res == 0) {     	/* eof */
+	    port_inp_failure(dd, res);
+	}
+	else if (res < packet_bytes - dd->ifd->psz) {
+	    memcpy(dd->ifd->pbuf+dd->ifd->psz,
+		   read_buf, res);
+	    dd->ifd->psz += res;
+	}
+	else  { /* if (res >= packet_bytes) */
+	    unsigned char* cpos = read_buf;
+	    int bytes_left = res;
+
+	    while (1) {
+		int psz = dd->ifd->psz;
+		char* pbp = dd->ifd->pbuf + psz;
+
+		while(bytes_left && (psz < packet_bytes)) {
+		    *pbp++ = *cpos++;
+		    bytes_left--;
+		    psz++;
+		}
+
+		if (psz < packet_bytes) {
+		    dd->ifd->psz = psz;
+		    break;
+		}
+		dd->ifd->psz = 0;
+
+		switch (packet_bytes) {
+		case 1: h = get_int8(dd->ifd->pbuf);  break;
+		case 2: h = get_int16(dd->ifd->pbuf); break;
+		case 4: h = get_int32(dd->ifd->pbuf); break;
+		default: ASSERT(0); return; /* -1; */
+		}
+
+		if (h <= (bytes_left)) {
+		    driver_output(port_num, (char*) cpos, h);
+		    cpos += h;
+		    bytes_left -= h;
+		    continue;
+		}
+		else {		/* The last message we got was split */
+		        char *buf = erts_alloc_fnf(ERTS_ALC_T_FD_ENTRY_BUF, h);
+		    if (!buf) {
+			errno = ENOMEM;
+			port_inp_failure(dd, -1);
+		    }
+		    else {
+			erts_smp_atomic_add_nob(&sys_misc_mem_sz, h);
+			sys_memcpy(buf, cpos, bytes_left);
+			dd->ifd->buf = buf;
+			dd->ifd->sz = h;
+			dd->ifd->remain = h - bytes_left;
+			dd->ifd->cpos = buf + bytes_left;
+		    }
+		    break;
+		}
+	    }
+	}
+	erts_free(ERTS_ALC_T_SYS_READ_BUF, (void *) read_buf);
+    }
+}
+
+
+/* fd is the drv_data that is returned from the */
+/* initial start routine                        */
+/* ready_fd is the descriptor that is ready to read */
+
+static void ready_output(ErlDrvData e, ErlDrvEvent ready_fd)
+{
+    ErtsSysDriverData *dd = (ErtsSysDriverData*)e;
+    ErlDrvPort ix = dd->port_num;
+    int n;
+    struct iovec* iv;
+    int vsize;
+
+    if ((iv = (struct iovec*) driver_peekq(ix, &vsize)) == NULL) {
+	driver_select(ix, ready_fd, ERL_DRV_WRITE, 0);
+        if (dd->pid > 0 && dd->ofd->fd < 0) {
+            /* The port was opened with 'in' option, which means we
+               should close the output fd as soon as the command has
+               been sent. */
+            driver_select(ix, ready_fd, ERL_DRV_WRITE|ERL_DRV_USE, 0);
+            erts_smp_atomic_add_nob(&sys_misc_mem_sz, -sizeof(ErtsSysFdData));
+            dd->ofd = NULL;
+        }
+        if (dd->terminating)
+            driver_failure_atom(dd->port_num,"normal");
+	return; /* 0; */
+    }
+    vsize = vsize > MAX_VSIZE ? MAX_VSIZE : vsize;
+    if ((n = writev(ready_fd, iv, vsize)) > 0) {
+	if (driver_deq(ix, n) == 0)
+	    set_busy_port(ix, 0);
+    }
+    else if (n < 0) {
+	if (errno == ERRNO_BLOCK || errno == EINTR)
+	    return; /* 0; */
+	else {
+	    int res = errno;
+	    driver_select(ix, ready_fd, ERL_DRV_WRITE, 0);
+	    driver_failure_posix(ix, res);
+	    return; /* -1; */
+	}
+    }
+    return; /* 0; */
+}
+
+static void stop_select(ErlDrvEvent fd, void* _)
+{
+    close((int)fd);
+}
+
+#if FDBLOCK
+
+static void
+fd_async(void *async_data)
+{
+    int res;
+    ErtsSysDriverData *dd = (ErtsSysDriverData *)async_data;
+    SysIOVec      *iov0;
+    SysIOVec      *iov;
+    int            iovlen;
+    int            err = 0;
+    /* much of this code is stolen from efile_drv:invoke_writev */
+    driver_pdl_lock(dd->blocking->pdl);
+    iov0 = driver_peekq(dd->port_num, &iovlen);
+    iovlen = iovlen < MAXIOV ? iovlen : MAXIOV;
+    iov = erts_alloc_fnf(ERTS_ALC_T_SYS_WRITE_BUF,
+                         sizeof(SysIOVec)*iovlen);
+    if (!iov) {
+        res = -1;
+        err = ENOMEM;
+        driver_pdl_unlock(dd->blocking->pdl);
+    } else {
+        memcpy(iov,iov0,iovlen*sizeof(SysIOVec));
+        driver_pdl_unlock(dd->blocking->pdl);
+
+        do {
+            res = writev(dd->ofd->fd, iov, iovlen);
+        } while (res < 0 && errno == EINTR);
+        if (res < 0)
+            err = errno;
+        err = errno;
+
+        erts_free(ERTS_ALC_T_SYS_WRITE_BUF, iov);
+    }
+    dd->blocking->res = res;
+    dd->blocking->err = err;
+}
+
+void fd_ready_async(ErlDrvData drv_data,
+                    ErlDrvThreadData thread_data) {
+    ErtsSysDriverData *dd = (ErtsSysDriverData *)thread_data;
+    ErlDrvPort port_num = dd->port_num;
+
+    ASSERT(dd->blocking);
+
+    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) {
+        if (dd->blocking->err == ERRNO_BLOCK) {
+            set_busy_port(port_num, 1);
+            /* still data left to write in queue */
+            driver_async(port_num, &dd->blocking->pkey, fd_async, dd, NULL);
+        } else
+            driver_failure_posix(port_num, dd->blocking->err);
+        return; /* -1; */
+    }
+    return; /* 0; */
+}
+
+#endif
+
+/* Forker driver */
+
+static int forker_fd;
+
+static ErlDrvData forker_start(ErlDrvPort port_num, char* name,
+                               SysDriverOpts* opts)
+{
+
+    int i;
+    int fds[2];
+    int res, unbind;
+    char bindir[MAXPATHLEN];
+    size_t bindirsz = sizeof(bindir);
+    Uint csp_path_sz;
+    char *child_setup_prog;
+
+    forker_port = erts_drvport2id(port_num);
+
+    res = erts_sys_getenv_raw("BINDIR", bindir, &bindirsz);
+    if (res != 0) {
+        if (res < 0)
+            erts_exit(1,
+                     "Environment variable BINDIR is not set\n");
+        if (res > 0)
+            erts_exit(1,
+                     "Value of environment variable BINDIR is too large\n");
+    }
+    if (bindir[0] != DIR_SEPARATOR_CHAR)
+        erts_exit(1,
+                 "Environment variable BINDIR does not contain an"
+                 " absolute path\n");
+    csp_path_sz = (strlen(bindir)
+                   + 1 /* DIR_SEPARATOR_CHAR */
+                   + sizeof(CHILD_SETUP_PROG_NAME)
+                   + 1);
+    child_setup_prog = erts_alloc(ERTS_ALC_T_CS_PROG_PATH, csp_path_sz);
+    erts_snprintf(child_setup_prog, csp_path_sz,
+                  "%s%c%s",
+                  bindir,
+                  DIR_SEPARATOR_CHAR,
+                  CHILD_SETUP_PROG_NAME);
+    if (socketpair(AF_UNIX, SOCK_STREAM, 0, fds) < 0) {
+        erts_exit(ERTS_ABORT_EXIT,
+                 "Could not open unix domain socket in spawn_init: %d\n",
+                 errno);
+    }
+
+    forker_fd = fds[0];
+
+    unbind = erts_sched_bind_atfork_prepare();
+
+    i = fork();
+
+    if (i == 0) {
+        /* The child */
+        char *cs_argv[FORKER_ARGV_NO_OF_ARGS] =
+            {CHILD_SETUP_PROG_NAME, NULL, NULL};
+        char buff[128];
+
+        erts_sched_bind_atfork_child(unbind);
+
+        snprintf(buff, 128, "%d", sys_max_files());
+        cs_argv[FORKER_ARGV_MAX_FILES] = buff;
+
+        /* We preallocate fd 3 for the uds fd */
+        if (fds[1] != 3) {
+            dup2(fds[1], 3);
+        }
+
+#if defined(USE_SETPGRP_NOARGS)		/* SysV */
+    (void) setpgrp();
+#elif defined(USE_SETPGRP)		/* BSD */
+    (void) setpgrp(0, getpid());
+#else					/* POSIX */
+    (void) setsid();
+#endif
+
+        execv(child_setup_prog, cs_argv);
+        _exit(1);
+    }
+
+    erts_sched_bind_atfork_parent(unbind);
+
+    erts_free(ERTS_ALC_T_CS_PROG_PATH, child_setup_prog);
+
+    close(fds[1]);
+
+    SET_NONBLOCKING(forker_fd);
+
+    driver_select(port_num, forker_fd, ERL_DRV_READ|ERL_DRV_USE, 1);
+
+    return (ErlDrvData)port_num;
+}
+
+static void forker_stop(ErlDrvData e)
+{
+    /* we probably should do something here,
+       the port has been closed by the user. */
+}
+
+static void forker_ready_input(ErlDrvData e, ErlDrvEvent fd)
+{
+    int res;
+    ErtsSysForkerProto *proto;
+
+    proto = erts_alloc(ERTS_ALC_T_DRV_CTRL_DATA, sizeof(*proto));
+
+    if ((res = read(fd, proto, sizeof(*proto))) < 0) {
+        if (errno == ERRNO_BLOCK)
+            return;
+        erts_exit(ERTS_DUMP_EXIT, "Failed to read from erl_child_setup: %d\n", errno);
+    }
+
+    if (res == 0)
+        erts_exit(ERTS_DUMP_EXIT, "erl_child_setup closed\n");
+
+    ASSERT(res == sizeof(*proto));
+
+#ifdef FORKER_PROTO_START_ACK
+    if (proto->action == ErtsSysForkerProtoAction_StartAck) {
+        /* Ideally we would like to not have to ack each Start
+           command being sent over the uds, but it would seem
+           that some operating systems (only observed on FreeBSD)
+           throw away data on the uds when the socket becomes full,
+           so we have to.
+        */
+        ErlDrvPort port_num = (ErlDrvPort)e;
+        int vlen;
+        SysIOVec *iov = driver_peekq(port_num, &vlen);
+        ErtsSysForkerProto *proto = (ErtsSysForkerProto *)iov[0].iov_base;
+
+        close(proto->u.start.fds[0]);
+        close(proto->u.start.fds[1]);
+        if (proto->u.start.fds[1] != proto->u.start.fds[2])
+            close(proto->u.start.fds[2]);
+
+        driver_deq(port_num, sizeof(*proto));
+
+        if (driver_sizeq(port_num) > 0)
+            driver_select(port_num, forker_fd, ERL_DRV_WRITE|ERL_DRV_USE, 1);
+    } else
+#endif
+    {
+        ASSERT(proto->action == ErtsSysForkerProtoAction_SigChld);
+
+        /* ideally this would be a port_command call, but as command is
+           already used by the spawn_driver, we use control instead.
+           Note that when using erl_drv_port_control it is an asynchronous
+           control. */
+        erl_drv_port_control(proto->u.sigchld.port_id, 'S',
+                             (char*)proto, sizeof(*proto));
+    }
+
+}
+
+static void forker_ready_output(ErlDrvData e, ErlDrvEvent fd)
+{
+    ErlDrvPort port_num = (ErlDrvPort)e;
+
+#ifndef FORKER_PROTO_START_ACK
+    while (driver_sizeq(port_num) > 0) {
+#endif
+        int vlen;
+        SysIOVec *iov = driver_peekq(port_num, &vlen);
+        ErtsSysForkerProto *proto = (ErtsSysForkerProto *)iov[0].iov_base;
+        ASSERT(iov[0].iov_len >= (sizeof(*proto)));
+        if (sys_uds_write(forker_fd, (char*)proto, sizeof(*proto),
+                          proto->u.start.fds, 3, 0) < 0) {
+            if (errno == ERRNO_BLOCK)
+                return;
+            erts_exit(ERTS_DUMP_EXIT, "Failed to write to erl_child_setup: %d\n", errno);
+        }
+#ifndef FORKER_PROTO_START_ACK
+        close(proto->u.start.fds[0]);
+        close(proto->u.start.fds[1]);
+        if (proto->u.start.fds[1] != proto->u.start.fds[2])
+            close(proto->u.start.fds[2]);
+        driver_deq(port_num, sizeof(*proto));
+    }
+#endif
+
+    driver_select(port_num, forker_fd, ERL_DRV_WRITE, 0);
+}
+
+static ErlDrvSSizeT forker_control(ErlDrvData e, unsigned int cmd, char *buf,
+                                   ErlDrvSizeT len, char **rbuf, ErlDrvSizeT rlen)
+{
+    ErtsSysForkerProto *proto = (ErtsSysForkerProto *)buf;
+    ErlDrvPort port_num = (ErlDrvPort)e;
+    int res;
+
+    driver_enq(port_num, buf, len);
+    if (driver_sizeq(port_num) > sizeof(*proto)) {
+        return 0;
+    }
+
+    if ((res = sys_uds_write(forker_fd, (char*)proto, sizeof(*proto),
+                             proto->u.start.fds, 3, 0)) < 0) {
+        if (errno == ERRNO_BLOCK) {
+            driver_select(port_num, forker_fd, ERL_DRV_WRITE|ERL_DRV_USE, 1);
+            return 0;
+        }
+        erts_exit(ERTS_DUMP_EXIT, "Failed to write to erl_child_setup: %d\n", errno);
+    }
+
+#ifndef FORKER_PROTO_START_ACK
+    ASSERT(res == sizeof(*proto));
+    close(proto->u.start.fds[0]);
+    close(proto->u.start.fds[1]);
+    if (proto->u.start.fds[1] != proto->u.start.fds[2])
+        close(proto->u.start.fds[2]);
+    driver_deq(port_num, sizeof(*proto));
+#endif
+
+    return 0;
+}
diff --git a/erts/emulator/sys/unix/sys_time.c b/erts/emulator/sys/unix/sys_time.c
index 95b3daa4ed..8ce02506ab 100644
--- a/erts/emulator/sys/unix/sys_time.c
+++ b/erts/emulator/sys/unix/sys_time.c
@@ -65,6 +65,8 @@
 #  include <fcntl.h>
 #endif
 
+static void init_perf_counter(void);
+
 /******************* Routines for time measurement *********************/
 
 #undef ERTS_SYS_TIME_INTERNAL_STATE_WRITE_FREQ__
@@ -404,6 +406,8 @@ sys_init_time(ErtsSysInitTimeResult *init_resp)
 #  error Missing erts_os_system_time() implementation
 #endif
 
+    init_perf_counter();
+
 }
 
 void
@@ -908,10 +912,120 @@ erts_os_times(ErtsMonotonicTime *mtimep, ErtsSystemTime *stimep)
 
 #endif
 
+/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *\
+ * Performance counter functions                                             *
+\*                                                                           */
+
+
+/* What resolution to spin to in micro seconds */
+#define RESOLUTION 100
+/* How many iterations to spin */
+#define ITERATIONS 1
+/* How many significant figures to round to */
+#define SIGFIGS 3
+
+static ErtsSysPerfCounter calculate_perf_counter_unit(void) {
+    int i;
+    ErtsSysPerfCounter pre, post;
+    double value = 0;
+    double round_factor;
+#if defined(HAVE_GETHRTIME) && defined(GETHRTIME_WITH_CLOCK_GETTIME)
+    struct timespec basetime,comparetime;
+#define __GETTIME(arg) clock_gettime(CLOCK_MONOTONIC,arg)
+#define __GETUSEC(arg) (arg.tv_nsec / 1000)
+#else
+    SysTimeval basetime,comparetime;
+#define __GETTIME(arg) sys_gettimeofday(arg)
+#define __GETUSEC(arg) arg.tv_usec
+#endif
+
+    for (i = 0; i < ITERATIONS; i++) {
+        /* Make sure usec just flipped over at current resolution */
+        __GETTIME(&basetime);
+        do {
+            __GETTIME(&comparetime);
+        } while ((__GETUSEC(basetime) / RESOLUTION) == (__GETUSEC(comparetime) / RESOLUTION));
+
+        pre = erts_sys_perf_counter();
+
+        __GETTIME(&basetime);
+        do {
+            __GETTIME(&comparetime);
+        } while ((__GETUSEC(basetime) / RESOLUTION) == (__GETUSEC(comparetime) / RESOLUTION));
+
+        post = erts_sys_perf_counter();
+
+        value += post - pre;
+    }
+    /* After this value is ticks per us */
+    value /= (RESOLUTION*ITERATIONS);
+
+    /* We round to 3 significant figures */
+    round_factor = pow(10.0, SIGFIGS - ceil(log10(value)));
+    value = ((ErtsSysPerfCounter)(value * round_factor + 0.5)) / round_factor;
+
+    /* convert to ticks per second */
+    return 1000000 * value;
+}
+
+static int have_rdtscp(void)
+{
+#if defined(ETHR_X86_RUNTIME_CONF__)
+    /* On early x86 cpu's the tsc varies with the current speed of the cpu,
+       which means that the time per tick vary depending on the current
+       load of the cpu. We do not want this as it would give very scewed
+       numbers when the cpu is mostly idle.
+       The linux kernel seems to think that checking for constant and
+       reliable is enough to trust the counter so we do the same.
+
+       If this test is not good enough, I don't know what we'll do.
+       Maybe fallback on erts_sys_hrtime always, but that would be a shame as
+       rdtsc is about 3 times faster than hrtime... */
+    return ETHR_X86_RUNTIME_CONF_HAVE_CONSTANT_TSC__ &&
+        ETHR_X86_RUNTIME_CONF_HAVE_TSC_RELIABLE__;
+#else
+    return 0;
+#endif
+}
+
+static ErtsSysPerfCounter rdtsc(void)
+{
+ /*  It may have been a good idea to put the cpuid instruction before
+     the rdtsc, but I decided against it because it is not really
+     needed for msacc, and it slows it down by quite a bit (5-7 times slower).
+     As a result though, this timestamp becomes much less
+     accurate as it might be re-ordered to be executed way before or after this
+     function is called.
+ */
+    ErtsSysPerfCounter ts;
+#if defined(__x86_64__)
+    __asm__ __volatile__ ("rdtsc\n\t"
+                          "shl $32, %%rdx\n\t"
+                          "or %%rdx, %0" : "=a" (ts) : : "rdx");
+#elif defined(__i386__)
+    __asm__ __volatile__ ("rdtsc\n\t"
+                           : "=A" (ts) );
+#endif
+    return ts;
+}
+
+static void init_perf_counter(void)
+{
+    if (have_rdtscp()) {
+        erts_sys_time_data__.r.o.perf_counter = rdtsc;
+        erts_sys_time_data__.r.o.perf_counter_unit = calculate_perf_counter_unit();
+    } else {
+        erts_sys_time_data__.r.o.perf_counter = erts_sys_hrtime;
+        erts_sys_time_data__.r.o.perf_counter_unit = ERTS_HRTIME_UNIT;
+    }
+}
+
 /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
 
 #ifdef HAVE_GETHRVTIME_PROCFS_IOCTL
 
+/* The code below only has effect on solaris < 10,
+   needed in order for gehhrvtime to work properly */
 int sys_start_hrvtime(void)
 {
     long msacct = PR_MSACCT;
diff --git a/erts/emulator/sys/unix/sys_uds.c b/erts/emulator/sys/unix/sys_uds.c
new file mode 100644
index 0000000000..015d0346a1
--- /dev/null
+++ b/erts/emulator/sys/unix/sys_uds.c
@@ -0,0 +1,155 @@
+/*
+ * %CopyrightBegin%
+ * 
+ * Copyright Ericsson AB 2002-2009. All Rights Reserved.
+ * 
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * 
+ * %CopyrightEnd%
+ */
+
+#include "sys_uds.h"
+
+int
+sys_uds_readv(int fd, struct iovec *iov, size_t iov_len,
+             int *fds, int fd_count, int flags) {
+    struct msghdr msg;
+    struct cmsghdr *cmsg = NULL;
+    char ancillary_buff[256] = {0};
+    int res, i = 0;
+
+    /* setup a place to fill in message contents */
+    memset(&msg, 0, sizeof(struct msghdr));
+    msg.msg_iov = iov;
+    msg.msg_iovlen = iov_len;
+
+    /* provide space for the ancillary data */
+    msg.msg_control = ancillary_buff;
+    msg.msg_controllen = sizeof(ancillary_buff);
+
+    if((res = recvmsg(fd, &msg, flags)) < 0) {
+#if defined(__APPLE__) && defined(__MACH__) && !defined(__DARWIN__)
+        /* When some OS X versions run out of fd's
+           they give EMSGSIZE instead of EMFILE.
+           We remap this as we want the correct
+           error to appear for the user */
+        if (errno == EMSGSIZE)
+            errno = EMFILE;
+#endif
+        return res;
+    }
+
+    if((msg.msg_flags & MSG_CTRUNC) == MSG_CTRUNC)
+    {
+        /* We assume that we have given enough space for any header
+           that are sent to us. So the only remaining reason to get
+           this flag set is if the caller has run out of file descriptors.
+        */
+        errno = EMFILE;
+        return -1;
+    }
+
+    for (cmsg = CMSG_FIRSTHDR(&msg); cmsg; cmsg = CMSG_NXTHDR(&msg, cmsg) ) {
+        if ((cmsg->cmsg_level == SOL_SOCKET) &&
+            (cmsg->cmsg_type == SCM_RIGHTS)) {
+                int *cmsg_data = (int *)CMSG_DATA(cmsg);
+                while ((char*)cmsg_data < (char*)cmsg + cmsg->cmsg_len) {
+                    if (i < fd_count) {
+                        fds[i++] = *cmsg_data++;
+                    } else {
+                        /* for some strange reason, we have received more FD's
+                           than we wanted... close them if we are not running
+                           debug. */
+                        if(i >= fd_count) abort();
+                        close(*cmsg_data++);
+                    }
+                }
+            }
+    }
+
+    return res;
+}
+
+int
+sys_uds_read(int fd, char *buff, size_t len,
+             int *fds, int fd_count, int flags) {
+    struct iovec iov;
+    iov.iov_base = buff;
+    iov.iov_len = len;
+    return sys_uds_readv(fd, &iov, 1, fds, fd_count, flags);
+}
+
+
+int
+sys_uds_writev(int fd, struct iovec *iov, size_t iov_len,
+               int *fds, int fd_count, int flags) {
+
+    struct msghdr msg;
+    struct cmsghdr *cmsg = NULL;
+    int res, i;
+
+    /* initialize socket message */
+    memset(&msg, 0, sizeof(struct msghdr));
+
+    /* We flatten the iov if it is too long */
+    if (iov_len > MAXIOV) {
+        int size = 0;
+        char *buff;
+        for (i = 0; i < iov_len; i++)
+            size += iov[i].iov_len;
+        buff = malloc(size);
+
+        for (i = 0; i < iov_len; i++) {
+            memcpy(buff, iov[i].iov_base, iov[i].iov_len);
+            buff += iov[i].iov_len;
+        }
+
+        iov[0].iov_base = buff - size;
+        iov[0].iov_len = size;
+        msg.msg_iov = iov;
+        msg.msg_iovlen = 1;
+    } else {
+        msg.msg_iov = iov;
+        msg.msg_iovlen = iov_len;
+    }
+
+    /* initialize the ancillary data */
+    msg.msg_control = calloc(1, CMSG_SPACE(sizeof(int) * fd_count));
+    msg.msg_controllen = CMSG_SPACE(sizeof(int) * fd_count);
+
+    /* copy the fd array into the ancillary data */
+    cmsg = CMSG_FIRSTHDR(&msg);
+    if(!cmsg) abort();
+    cmsg->cmsg_level = SOL_SOCKET;
+    cmsg->cmsg_type = SCM_RIGHTS;
+    cmsg->cmsg_len = CMSG_LEN(sizeof(int) * fd_count);
+    memcpy(CMSG_DATA(cmsg), fds, sizeof(int) * fd_count);
+
+    res = sendmsg(fd, &msg, flags);
+
+    if (iov_len > MAXIOV)
+        free(iov[0].iov_base);
+
+    free(msg.msg_control);
+
+    return res;
+}
+
+int
+sys_uds_write(int fd, char *buff, size_t len,
+              int *fds, int fd_count, int flags) {
+    struct iovec iov;
+    iov.iov_base = buff;
+    iov.iov_len = len;
+    return sys_uds_writev(fd, &iov, 1, fds, fd_count, flags);
+}
diff --git a/erts/emulator/sys/unix/sys_uds.h b/erts/emulator/sys/unix/sys_uds.h
new file mode 100644
index 0000000000..844a2804d8
--- /dev/null
+++ b/erts/emulator/sys/unix/sys_uds.h
@@ -0,0 +1,57 @@
+/*
+ * %CopyrightBegin%
+ *
+ * Copyright Ericsson AB 2002-2009. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ * %CopyrightEnd%
+ */
+
+#ifndef _ERL_UNIX_UDS_H
+#define _ERL_UNIX_UDS_H
+
+#ifdef HAVE_CONFIG_H
+#  include "config.h"
+#endif
+
+#if defined(__sun__) && !defined(_XOPEN_SOURCE)
+#define _XOPEN_SOURCE 500
+#endif
+
+#include <limits.h>
+
+#include <sys/types.h>
+#include <sys/socket.h>
+#include <sys/uio.h>
+
+#if defined IOV_MAX
+#define MAXIOV IOV_MAX
+#elif defined UIO_MAXIOV
+#define MAXIOV UIO_MAXIOV
+#else
+#define MAXIOV 16
+#endif
+
+#include "sys.h"
+
+int sys_uds_readv(int fd, struct iovec *iov, size_t iov_len,
+                  int *fds, int fd_count, int flags);
+int sys_uds_read(int fd, char *buff, size_t len,
+                 int *fds, int fd_count, int flags);
+int sys_uds_writev(int fd, struct iovec *iov, size_t iov_len,
+                   int *fds, int fd_count, int flags);
+int sys_uds_write(int fd, char *buff, size_t len,
+                  int *fds, int fd_count, int flags);
+
+#endif /* #ifndef _ERL_UNIX_UDS_H */