X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=6ce46ba5a31e9be12d4c2cc6ef524ec5297837c4;hp=0dbce58835ad83e76e6855f7d4f100c9509decaa;hb=be4ecfdf6c8daa75c4df8ac875c9a87da80d81a0;hpb=858a3d476060cfb1ea5dfbdffd94ed3108f194eb

diff --git a/fio.c b/fio.c
index 0dbce588..a58effcf 100644
--- a/fio.c
+++ b/fio.c
@@ -4,10 +4,12 @@
  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
  * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
  *
+ * The license below covers all files distributed with fio unless otherwise
+ * noted in the file itself.
+ *
  *  This program is free software; you can redistribute it and/or modify
- *  it under the terms of the GNU General Public License as published by
- *  the Free Software Foundation; either version 2 of the License, or
- *  (at your option) any later version.
+ *  it under the terms of the GNU General Public License version 2 as
+ *  published by the Free Software Foundation.
  *
  *  This program is distributed in the hope that it will be useful,
  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
@@ -24,69 +26,169 @@
 #include <string.h>
 #include <signal.h>
 #include <time.h>
+#include <locale.h>
 #include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
 #include <sys/shm.h>
-#include <sys/ioctl.h>
 #include <sys/mman.h>
 
 #include "fio.h"
-#include "os.h"
-
-#define MASK	(4095)
+#include "hash.h"
+#include "smalloc.h"
 
-#define ALIGN(buf)	(char *) (((unsigned long) (buf) + MASK) & ~(MASK))
+unsigned long page_mask;
+unsigned long page_size;
+#define ALIGN(buf)	\
+	(char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
 int thread_number = 0;
+int nr_process = 0;
+int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
-char *fio_inst_prefix = _INST_PREFIX;
+unsigned long done_secs = 0;
+
+static struct fio_mutex *startup_mutex;
+static volatile int fio_abort;
+static int exit_value;
+static struct itimerval itimer;
+static pthread_t gtod_thread;
 
-static volatile int startup_sem;
+struct io_log *agg_io_log[2];
 
 #define TERMINATE_ALL		(-1)
 #define JOB_START_TIMEOUT	(5 * 1000)
 
+void td_set_runstate(struct thread_data *td, int runstate)
+{
+	if (td->runstate == runstate)
+		return;
+
+	dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
+						td->runstate, runstate);
+	td->runstate = runstate;
+}
+
 static void terminate_threads(int group_id)
 {
 	struct thread_data *td;
 	int i;
 
+	dprint(FD_PROCESS, "terminate group_id=%d\n", group_id);
+
 	for_each_td(td, i) {
 		if (group_id == TERMINATE_ALL || groupid == td->groupid) {
+			dprint(FD_PROCESS, "setting terminate on %s/%d\n",
+						td->o.name, (int) td->pid);
 			td->terminate = 1;
-			td->start_delay = 0;
+			td->o.start_delay = 0;
+
+			/*
+			 * if the thread is running, just let it exit
+			 */
+			if (td->runstate < TD_RUNNING)
+				kill(td->pid, SIGQUIT);
+			else {
+				struct ioengine_ops *ops = td->io_ops;
+
+				if (ops && (ops->flags & FIO_SIGQUIT))
+					kill(td->pid, SIGQUIT);
+			}
 		}
 	}
 }
 
-static void sig_handler(int sig)
+static void status_timer_arm(void)
 {
-	switch (sig) {
-		case SIGALRM:
-			update_io_ticks();
-			disk_util_timer_arm();
-			print_thread_status();
-			break;
-		default:
-			printf("\nfio: terminating on signal\n");
-			fflush(stdout);
-			terminate_threads(TERMINATE_ALL);
-			break;
+	itimer.it_value.tv_sec = 0;
+	itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
+	setitimer(ITIMER_REAL, &itimer, NULL);
+}
+
+static void sig_alrm(int fio_unused sig)
+{
+	if (threads) {
+		update_io_ticks();
+		print_thread_status();
+		status_timer_arm();
+	}
+}
+
+/*
+ * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
+ */
+static void sig_quit(int sig)
+{
+}
+
+static void sig_int(int sig)
+{
+	if (threads) {
+		printf("\nfio: terminating on signal %d\n", sig);
+		fflush(stdout);
+		terminate_threads(TERMINATE_ALL);
 	}
 }
 
+static void sig_ill(int fio_unused sig)
+{
+	if (!threads)
+		return;
+
+	log_err("fio: illegal instruction. your cpu does not support "
+		"the sse4.2 instruction for crc32c\n");
+	terminate_threads(TERMINATE_ALL);
+	exit(4);
+}
+
+static void set_sig_handlers(void)
+{
+	struct sigaction act;
+
+	memset(&act, 0, sizeof(act));
+	act.sa_handler = sig_alrm;
+	act.sa_flags = SA_RESTART;
+	sigaction(SIGALRM, &act, NULL);
+
+	memset(&act, 0, sizeof(act));
+	act.sa_handler = sig_int;
+	act.sa_flags = SA_RESTART;
+	sigaction(SIGINT, &act, NULL);
+
+	memset(&act, 0, sizeof(act));
+	act.sa_handler = sig_ill;
+	act.sa_flags = SA_RESTART;
+	sigaction(SIGILL, &act, NULL);
+
+	memset(&act, 0, sizeof(act));
+	act.sa_handler = sig_quit;
+	act.sa_flags = SA_RESTART;
+	sigaction(SIGQUIT, &act, NULL);
+}
+
+static inline int should_check_rate(struct thread_data *td)
+{
+	/*
+	 * No minimum rate set, always ok
+	 */
+	if (!td->o.ratemin && !td->o.rate_iops_min)
+		return 0;
+
+	return 1;
+}
+
 /*
  * Check if we are above the minimum rate given.
  */
 static int check_min_rate(struct thread_data *td, struct timeval *now)
 {
+	unsigned long long bytes = 0;
+	unsigned long iops = 0;
 	unsigned long spent;
 	unsigned long rate;
-	int ddir = td->ddir;
 
 	/*
 	 * allow a 2 second settle period in the beginning
@@ -94,103 +196,120 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 	if (mtime_since(&td->start, now) < 2000)
 		return 0;
 
+	if (td_read(td)) {
+		iops += td->io_blocks[DDIR_READ];
+		bytes += td->this_io_bytes[DDIR_READ];
+	}
+	if (td_write(td)) {
+		iops += td->io_blocks[DDIR_WRITE];
+		bytes += td->this_io_bytes[DDIR_WRITE];
+	}
+
 	/*
 	 * if rate blocks is set, sample is running
 	 */
-	if (td->rate_bytes) {
+	if (td->rate_bytes || td->rate_blocks) {
 		spent = mtime_since(&td->lastrate, now);
-		if (spent < td->ratecycle)
+		if (spent < td->o.ratecycle)
 			return 0;
 
-		rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
-		if (rate < td->ratemin) {
-			fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
-			return 1;
+		if (td->o.rate) {
+			/*
+			 * check bandwidth specified rate
+			 */
+			if (bytes < td->rate_bytes) {
+				log_err("%s: min rate %u not met\n", td->o.name,
+								td->o.ratemin);
+				return 1;
+			} else {
+				rate = (bytes - td->rate_bytes) / spent;
+				if (rate < td->o.ratemin ||
+				    bytes < td->rate_bytes) {
+					log_err("%s: min rate %u not met, got"
+						" %luKiB/sec\n", td->o.name,
+							td->o.ratemin, rate);
+					return 1;
+				}
+			}
+		} else {
+			/*
+			 * checks iops specified rate
+			 */
+			if (iops < td->o.rate_iops) {
+				log_err("%s: min iops rate %u not met\n",
+						td->o.name, td->o.rate_iops);
+				return 1;
+			} else {
+				rate = (iops - td->rate_blocks) / spent;
+				if (rate < td->o.rate_iops_min ||
+				    iops < td->rate_blocks) {
+					log_err("%s: min iops rate %u not met,"
+						" got %lu\n", td->o.name,
+							td->o.rate_iops_min,
+							rate);
+				}
+			}
 		}
 	}
 
-	td->rate_bytes = td->this_io_bytes[ddir];
+	td->rate_bytes = bytes;
+	td->rate_blocks = iops;
 	memcpy(&td->lastrate, now, sizeof(*now));
 	return 0;
 }
 
 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 {
-	if (!td->timeout)
+	if (!td->o.timeout)
 		return 0;
-	if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
+	if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
 		return 1;
 
 	return 0;
 }
 
-static inline void td_set_runstate(struct thread_data *td, int runstate)
-{
-	td->runstate = runstate;
-}
-
-static struct fio_file *get_next_file(struct thread_data *td)
-{
-	unsigned int old_next_file = td->next_file;
-	struct fio_file *f;
-
-	do {
-		f = &td->files[td->next_file];
-
-		td->next_file++;
-		if (td->next_file >= td->nr_files)
-			td->next_file = 0;
-
-		if (f->fd != -1)
-			break;
-
-		f = NULL;
-	} while (td->next_file != old_next_file);
-
-	return f;
-}
-
 /*
  * When job exits, we can cancel the in-flight IO if we are using async
  * io. Attempt to do so.
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
-	struct list_head *entry, *n;
-	struct io_completion_data icd;
+	struct flist_head *entry, *n;
 	struct io_u *io_u;
 	int r;
 
 	/*
 	 * get immediately available events, if any
 	 */
-	r = td_io_getevents(td, 0, td->cur_depth, &ts);
-	if (r > 0) {
-		icd.nr = r;
-		ios_completed(td, &icd);
-	}
+	r = io_u_queued_complete(td, 0);
+	if (r < 0)
+		return;
 
 	/*
 	 * now cancel remaining active events
 	 */
 	if (td->io_ops->cancel) {
-		list_for_each_safe(entry, n, &td->io_u_busylist) {
-			io_u = list_entry(entry, struct io_u, list);
+		flist_for_each_safe(entry, n, &td->io_u_busylist) {
+			io_u = flist_entry(entry, struct io_u, list);
 
-			r = td->io_ops->cancel(td, io_u);
-			if (!r)
+			/*
+			 * if the io_u isn't in flight, then that generally
+			 * means someone leaked an io_u. complain but fix
+			 * it up, so we don't stall here.
+			 */
+			if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
+				log_err("fio: non-busy IO on busy list\n");
 				put_io_u(td, io_u);
+			} else {
+				r = td->io_ops->cancel(td, io_u);
+				if (!r)
+					put_io_u(td, io_u);
+			}
 		}
 	}
 
-	if (td->cur_depth) {
-		r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
-		if (r > 0) {
-			icd.nr = r;
-			ios_completed(td, &icd);
-		}
-	}
+	if (td->cur_depth)
+		r = io_u_queued_complete(td, td->cur_depth);
 }
 
 /*
@@ -200,7 +319,6 @@ static void cleanup_pending_aio(struct thread_data *td)
 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
 	struct io_u *io_u = __get_io_u(td);
-	struct io_completion_data icd;
 	int ret;
 
 	if (!io_u)
@@ -214,263 +332,395 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 		return 1;
 	}
 
+requeue:
 	ret = td_io_queue(td, io_u);
-	if (ret) {
-		put_io_u(td, io_u);
-		td_verror(td, ret);
-		return 1;
-	}
-
-	ret = td_io_getevents(td, 1, td->cur_depth, NULL);
 	if (ret < 0) {
-		td_verror(td, -ret);
+		td_verror(td, io_u->error, "td_io_queue");
+		put_io_u(td, io_u);
 		return 1;
-	}
+	} else if (ret == FIO_Q_QUEUED) {
+		if (io_u_queued_complete(td, 1) < 0)
+			return 1;
+	} else if (ret == FIO_Q_COMPLETED) {
+		if (io_u->error) {
+			td_verror(td, io_u->error, "td_io_queue");
+			return 1;
+		}
 
-	icd.nr = ret;
-	ios_completed(td, &icd);
-	if (icd.error) {
-		td_verror(td, icd.error);
-		return 1;
+		if (io_u_sync_complete(td, io_u) < 0)
+			return 1;
+	} else if (ret == FIO_Q_BUSY) {
+		if (td_io_commit(td))
+			return 1;
+		goto requeue;
 	}
 
 	return 0;
 }
 
+static inline void update_tv_cache(struct thread_data *td)
+{
+	if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
+		fio_gettime(&td->tv_cache, NULL);
+}
+
 /*
- * The main verify engine. Runs over the writes we previusly submitted,
+ * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
  */
-void do_verify(struct thread_data *td)
+static void do_verify(struct thread_data *td)
 {
-	struct timeval t;
-	struct io_u *io_u, *v_io_u = NULL;
-	struct io_completion_data icd;
 	struct fio_file *f;
-	int ret, i;
+	struct io_u *io_u;
+	int ret, min_events;
+	unsigned int i;
 
 	/*
 	 * sync io first and invalidate cache, to make sure we really
 	 * read from disk.
 	 */
 	for_each_file(td, f, i) {
-		fio_io_sync(td, f);
-		file_invalidate_cache(td, f);
+		if (!(f->flags & FIO_FILE_OPEN))
+			continue;
+		if (fio_io_sync(td, f))
+			break;
+		if (file_invalidate_cache(td, f))
+			break;
 	}
 
-	td_set_runstate(td, TD_VERIFYING);
+	if (td->error)
+		return;
 
-	do {
-		if (td->terminate)
-			break;
+	td_set_runstate(td, TD_VERIFYING);
 
-		gettimeofday(&t, NULL);
-		if (runtime_exceeded(td, &t))
-			break;
+	io_u = NULL;
+	while (!td->terminate) {
+		int ret2, full;
 
 		io_u = __get_io_u(td);
 		if (!io_u)
 			break;
 
-		if (get_next_verify(td, io_u)) {
+		update_tv_cache(td);
+
+		if (runtime_exceeded(td, &td->tv_cache)) {
 			put_io_u(td, io_u);
+			td->terminate = 1;
 			break;
 		}
 
-		f = get_next_file(td);
-		if (!f)
+		if (get_next_verify(td, io_u)) {
+			put_io_u(td, io_u);
 			break;
-
-		io_u->file = f;
+		}
 
 		if (td_io_prep(td, io_u)) {
 			put_io_u(td, io_u);
 			break;
 		}
 
+		io_u->end_io = verify_io_u;
+
 		ret = td_io_queue(td, io_u);
-		if (ret) {
-			put_io_u(td, io_u);
-			td_verror(td, ret);
-			break;
-		}
+		switch (ret) {
+		case FIO_Q_COMPLETED:
+			if (io_u->error)
+				ret = -io_u->error;
+			else if (io_u->resid) {
+				int bytes = io_u->xfer_buflen - io_u->resid;
+				struct fio_file *f = io_u->file;
+
+				/*
+				 * zero read, fail
+				 */
+				if (!bytes) {
+					td_verror(td, EIO, "full resid");
+					put_io_u(td, io_u);
+					break;
+				}
 
-		/*
-		 * we have one pending to verify, do that while
-		 * we are doing io on the next one
-		 */
-		if (do_io_u_verify(td, &v_io_u))
-			break;
+				io_u->xfer_buflen = io_u->resid;
+				io_u->xfer_buf += bytes;
+				io_u->offset += bytes;
 
-		ret = td_io_getevents(td, 1, 1, NULL);
-		if (ret != 1) {
-			if (ret < 0)
-				td_verror(td, ret);
-			break;
-		}
+				td->ts.short_io_u[io_u->ddir]++;
 
-		v_io_u = td->io_ops->event(td, 0);
-		icd.nr = 1;
-		icd.error = 0;
-		io_completed(td, v_io_u, &icd);
+				if (io_u->offset == f->real_file_size)
+					goto sync_done;
 
-		if (icd.error) {
-			td_verror(td, icd.error);
-			put_io_u(td, v_io_u);
-			v_io_u = NULL;
+				requeue_io_u(td, &io_u);
+			} else {
+sync_done:
+				ret = io_u_sync_complete(td, io_u);
+				if (ret < 0)
+					break;
+			}
+			continue;
+		case FIO_Q_QUEUED:
+			break;
+		case FIO_Q_BUSY:
+			requeue_io_u(td, &io_u);
+			ret2 = td_io_commit(td);
+			if (ret2 < 0)
+				ret = ret2;
+			break;
+		default:
+			assert(ret < 0);
+			td_verror(td, -ret, "td_io_queue");
 			break;
 		}
 
+		if (ret < 0 || td->error)
+			break;
+
 		/*
-		 * if we can't submit more io, we need to verify now
+		 * if we can queue more, do so. but check if there are
+		 * completed io_u's first.
 		 */
-		if (queue_full(td) && do_io_u_verify(td, &v_io_u))
+		full = queue_full(td) || ret == FIO_Q_BUSY;
+		if (full || !td->o.iodepth_batch_complete) {
+			min_events = td->o.iodepth_batch_complete;
+			if (full && !min_events)
+				min_events = 1;
+
+			do {
+				/*
+				 * Reap required number of io units, if any,
+				 * and do the verification on them through
+				 * the callback handler
+				 */
+				if (io_u_queued_complete(td, min_events) < 0) {
+					ret = -1;
+					break;
+				}
+			} while (full && (td->cur_depth > td->o.iodepth_low));
+		}
+		if (ret < 0)
 			break;
+	}
 
-	} while (1);
-
-	do_io_u_verify(td, &v_io_u);
+	if (!td->error) {
+		min_events = td->cur_depth;
 
-	if (td->cur_depth)
+		if (min_events)
+			ret = io_u_queued_complete(td, min_events);
+	} else
 		cleanup_pending_aio(td);
 
 	td_set_runstate(td, TD_RUNNING);
 }
 
-/*
- * Not really an io thread, all it does is burn CPU cycles in the specified
- * manner.
- */
-static void do_cpuio(struct thread_data *td)
-{
-	struct timeval e;
-	int split = 100 / td->cpuload;
-	int i = 0;
-
-	while (!td->terminate) {
-		gettimeofday(&e, NULL);
-
-		if (runtime_exceeded(td, &e))
-			break;
-
-		if (!(i % split))
-			__usec_sleep(10000);
-		else
-			usec_sleep(td, 10000);
-
-		i++;
-	}
-}
-
 /*
  * Main IO worker function. It retrieves io_u's to process and queues
  * and reaps them, checking for rate and errors along the way.
  */
 static void do_io(struct thread_data *td)
 {
-	struct io_completion_data icd;
-	struct timeval s, e;
 	unsigned long usec;
-	struct fio_file *f;
-	int i, ret = 0;
+	unsigned int i;
+	int ret = 0;
 
-	td_set_runstate(td, TD_RUNNING);
+	if (in_ramp_time(td))
+		td_set_runstate(td, TD_RAMP);
+	else
+		td_set_runstate(td, TD_RUNNING);
 
-	while (td->this_io_bytes[td->ddir] < td->io_size) {
-		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
-		struct timespec *timeout;
+	while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
+		struct timeval comp_time;
+		long bytes_done = 0;
 		int min_evts = 0;
 		struct io_u *io_u;
+		int ret2, full;
 
 		if (td->terminate)
 			break;
 
-		f = get_next_file(td);
-		if (!f)
+		io_u = get_io_u(td);
+		if (!io_u)
 			break;
 
-		io_u = get_io_u(td, f);
-		if (!io_u)
+		update_tv_cache(td);
+
+		if (runtime_exceeded(td, &td->tv_cache)) {
+			put_io_u(td, io_u);
+			td->terminate = 1;
 			break;
+		}
 
-		memcpy(&s, &io_u->start_time, sizeof(s));
+		/*
+		 * Add verification end_io handler, if asked to verify
+		 * a previously written file.
+		 */
+		if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
+			io_u->end_io = verify_io_u;
+			td_set_runstate(td, TD_VERIFYING);
+		} else if (in_ramp_time(td))
+			td_set_runstate(td, TD_RAMP);
+		else
+			td_set_runstate(td, TD_RUNNING);
 
 		ret = td_io_queue(td, io_u);
-		if (ret) {
+		switch (ret) {
+		case FIO_Q_COMPLETED:
+			if (io_u->error)
+				ret = -io_u->error;
+			else if (io_u->resid) {
+				int bytes = io_u->xfer_buflen - io_u->resid;
+				struct fio_file *f = io_u->file;
+
+				/*
+				 * zero read, fail
+				 */
+				if (!bytes) {
+					td_verror(td, EIO, "full resid");
+					put_io_u(td, io_u);
+					break;
+				}
+
+				io_u->xfer_buflen = io_u->resid;
+				io_u->xfer_buf += bytes;
+				io_u->offset += bytes;
+
+				td->ts.short_io_u[io_u->ddir]++;
+
+				if (io_u->offset == f->real_file_size)
+					goto sync_done;
+
+				requeue_io_u(td, &io_u);
+			} else {
+sync_done:
+				if (should_check_rate(td))
+					fio_gettime(&comp_time, NULL);
+
+				bytes_done = io_u_sync_complete(td, io_u);
+				if (bytes_done < 0)
+					ret = bytes_done;
+			}
+			break;
+		case FIO_Q_QUEUED:
+			/*
+			 * if the engine doesn't have a commit hook,
+			 * the io_u is really queued. if it does have such
+			 * a hook, it has to call io_u_queued() itself.
+			 */
+			if (td->io_ops->commit == NULL)
+				io_u_queued(td, io_u);
+			break;
+		case FIO_Q_BUSY:
+			requeue_io_u(td, &io_u);
+			ret2 = td_io_commit(td);
+			if (ret2 < 0)
+				ret = ret2;
+			break;
+		default:
+			assert(ret < 0);
 			put_io_u(td, io_u);
-			td_verror(td, ret);
 			break;
 		}
 
-		add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
+		if (ret < 0 || td->error)
+			break;
 
-		if (td->cur_depth < td->iodepth) {
-			timeout = &ts;
-			min_evts = 0;
-		} else {
-			timeout = NULL;
-			min_evts = 1;
+		/*
+		 * See if we need to complete some commands
+		 */
+		full = queue_full(td) || ret == FIO_Q_BUSY;
+		if (full || !td->o.iodepth_batch_complete) {
+			min_evts = td->o.iodepth_batch_complete;
+			if (full && !min_evts)
+				min_evts = 1;
+
+			if (should_check_rate(td))
+				fio_gettime(&comp_time, NULL);
+
+			do {
+				ret = io_u_queued_complete(td, min_evts);
+				if (ret <= 0)
+					break;
+
+				bytes_done += ret;
+			} while (full && (td->cur_depth > td->o.iodepth_low));
 		}
 
-		ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
-		if (ret < 0) {
-			td_verror(td, ret);
+		if (ret < 0)
 			break;
-		} else if (!ret)
+		if (!bytes_done)
 			continue;
 
-		icd.nr = ret;
-		ios_completed(td, &icd);
-		if (icd.error) {
-			td_verror(td, icd.error);
-			break;
-		}
-
 		/*
 		 * the rate is batched for now, it should work for batches
 		 * of completions except the very first one which may look
 		 * a little bursty
 		 */
-		gettimeofday(&e, NULL);
-		usec = utime_since(&s, &e);
+		if (!in_ramp_time(td) && should_check_rate(td)) {
+			usec = utime_since(&td->tv_cache, &comp_time);
 
-		rate_throttle(td, usec, icd.bytes_done[td->ddir]);
+			rate_throttle(td, usec, bytes_done);
 
-		if (check_min_rate(td, &e)) {
-			if (rate_quit)
-				terminate_threads(td->groupid);
-			td_verror(td, ENOMEM);
-			break;
+			if (check_min_rate(td, &comp_time)) {
+				if (exitall_on_terminate)
+					terminate_threads(td->groupid);
+				td_verror(td, EIO, "check_min_rate");
+				break;
+			}
 		}
 
-		if (runtime_exceeded(td, &e))
-			break;
+		if (td->o.thinktime) {
+			unsigned long long b;
+
+			b = td->io_blocks[0] + td->io_blocks[1];
+			if (!(b % td->o.thinktime_blocks)) {
+				int left;
+
+				if (td->o.thinktime_spin)
+					__usec_sleep(td->o.thinktime_spin);
+
+				left = td->o.thinktime - td->o.thinktime_spin;
+				if (left)
+					usec_sleep(td, left);
+			}
+		}
+	}
 
-		if (td->thinktime)
-			usec_sleep(td, td->thinktime);
+	if (td->o.fill_device && td->error == ENOSPC) {
+		td->error = 0;
+		td->terminate = 1;
 	}
+	if (!td->error) {
+		struct fio_file *f;
 
-	if (!ret) {
-		if (td->cur_depth)
-			cleanup_pending_aio(td);
+		i = td->cur_depth;
+		if (i)
+			ret = io_u_queued_complete(td, i);
 
-		if (should_fsync(td) && td->end_fsync) {
+		if (should_fsync(td) && td->o.end_fsync) {
 			td_set_runstate(td, TD_FSYNCING);
-			for_each_file(td, f, i)
+
+			for_each_file(td, f, i) {
+				if (!(f->flags & FIO_FILE_OPEN))
+					continue;
 				fio_io_sync(td, f);
+			}
 		}
-	}
+	} else
+		cleanup_pending_aio(td);
+
+	/*
+	 * stop job if we failed doing any IO
+	 */
+	if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
+		td->done = 1;
 }
 
 static void cleanup_io_u(struct thread_data *td)
 {
-	struct list_head *entry, *n;
+	struct flist_head *entry, *n;
 	struct io_u *io_u;
 
-	list_for_each_safe(entry, n, &td->io_u_freelist) {
-		io_u = list_entry(entry, struct io_u, list);
+	flist_for_each_safe(entry, n, &td->io_u_freelist) {
+		io_u = flist_entry(entry, struct io_u, list);
 
-		list_del(&io_u->list);
+		flist_del(&io_u->list);
 		free(io_u);
 	}
 
@@ -480,31 +730,52 @@ static void cleanup_io_u(struct thread_data *td)
 static int init_io_u(struct thread_data *td)
 {
 	struct io_u *io_u;
+	unsigned int max_bs;
 	int i, max_units;
 	char *p;
 
-	if (td->io_ops->flags & FIO_CPUIO)
-		return 0;
+	max_units = td->o.iodepth;
+	max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
+	td->orig_buffer_size = (unsigned long long) max_bs
+					* (unsigned long long) max_units;
 
-	if (td->io_ops->flags & FIO_SYNCIO)
-		max_units = 1;
-	else
-		max_units = td->iodepth;
+	if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
+		unsigned long bs;
 
-	td->orig_buffer_size = td->max_bs * max_units + MASK;
+		bs = td->orig_buffer_size + td->o.hugepage_size - 1;
+		td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
+	}
+
+	if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
+		log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
+		return 1;
+	}
 
 	if (allocate_io_mem(td))
 		return 1;
 
-	p = ALIGN(td->orig_buffer);
+	if (td->o.odirect)
+		p = ALIGN(td->orig_buffer);
+	else
+		p = td->orig_buffer;
+
 	for (i = 0; i < max_units; i++) {
+		if (td->terminate)
+			return 1;
 		io_u = malloc(sizeof(*io_u));
 		memset(io_u, 0, sizeof(*io_u));
-		INIT_LIST_HEAD(&io_u->list);
+		INIT_FLIST_HEAD(&io_u->list);
+
+		if (!(td->io_ops->flags & FIO_NOIO)) {
+			io_u->buf = p + max_bs * i;
+
+			if (td_write(td) && !td->o.refill_buffers)
+				io_u_fill_buffer(td, io_u, max_bs);
+		}
 
-		io_u->buf = p + td->max_bs * i;
 		io_u->index = i;
-		list_add(&io_u->list, &td->io_u_freelist);
+		io_u->flags = IO_U_F_FREE;
+		flist_add(&io_u->list, &td->io_u_freelist);
 	}
 
 	return 0;
@@ -516,20 +787,28 @@ static int switch_ioscheduler(struct thread_data *td)
 	FILE *f;
 	int ret;
 
+	if (td->io_ops->flags & FIO_DISKLESSIO)
+		return 0;
+
 	sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 
 	f = fopen(tmp, "r+");
 	if (!f) {
-		td_verror(td, errno);
+		if (errno == ENOENT) {
+			log_err("fio: os or kernel doesn't support IO scheduler"
+				" switching\n");
+			return 0;
+		}
+		td_verror(td, errno, "fopen iosched");
 		return 1;
 	}
 
 	/*
 	 * Set io scheduler.
 	 */
-	ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
+	ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
 	if (ferror(f) || ret != 1) {
-		td_verror(td, errno);
+		td_verror(td, errno, "fwrite");
 		fclose(f);
 		return 1;
 	}
@@ -541,15 +820,15 @@ static int switch_ioscheduler(struct thread_data *td)
 	 */
 	ret = fread(tmp, 1, sizeof(tmp), f);
 	if (ferror(f) || ret < 0) {
-		td_verror(td, errno);
+		td_verror(td, errno, "fread");
 		fclose(f);
 		return 1;
 	}
 
-	sprintf(tmp2, "[%s]", td->ioscheduler);
+	sprintf(tmp2, "[%s]", td->o.ioscheduler);
 	if (!strstr(tmp, tmp2)) {
-		log_err("fio: io scheduler %s not found\n", td->ioscheduler);
-		td_verror(td, EINVAL);
+		log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
+		td_verror(td, EINVAL, "iosched_switch");
 		fclose(f);
 		return 1;
 	}
@@ -558,23 +837,83 @@ static int switch_ioscheduler(struct thread_data *td)
 	return 0;
 }
 
-static void clear_io_state(struct thread_data *td)
+static int keep_running(struct thread_data *td)
 {
-	struct fio_file *f;
-	int i;
+	unsigned long long io_done;
+
+	if (td->done)
+		return 0;
+	if (td->o.time_based)
+		return 1;
+	if (td->o.loops) {
+		td->o.loops--;
+		return 1;
+	}
+
+	io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
+			+ td->io_skip_bytes;
+	if (io_done < td->o.size)
+		return 1;
 
-	td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
+	return 0;
+}
+
+static void reset_io_counters(struct thread_data *td)
+{
+	td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
 	td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 	td->zone_bytes = 0;
+	td->rate_bytes = 0;
+	td->rate_blocks = 0;
+	td->rw_end_set[0] = td->rw_end_set[1] = 0;
 
-	for_each_file(td, f, i) {
-		f->last_pos = 0;
-		if (td->io_ops->flags & FIO_SYNCIO)
-			lseek(f->fd, SEEK_SET, 0);
+	td->last_was_sync = 0;
+
+	/*
+	 * reset file done count if we are to start over
+	 */
+	if (td->o.time_based || td->o.loops)
+		td->nr_done_files = 0;
+}
+
+void reset_all_stats(struct thread_data *td)
+{
+	struct timeval tv;
+	int i;
+
+	reset_io_counters(td);
+
+	for (i = 0; i < 2; i++) {
+		td->io_bytes[i] = 0;
+		td->io_blocks[i] = 0;
+		td->io_issues[i] = 0;
+		td->ts.total_io_u[i] = 0;
+	}
+	
+	fio_gettime(&tv, NULL);
+	memcpy(&td->epoch, &tv, sizeof(tv));
+	memcpy(&td->start, &tv, sizeof(tv));
+}
+
+static int clear_io_state(struct thread_data *td)
+{
+	struct fio_file *f;
+	unsigned int i;
+	int ret;
+
+	reset_io_counters(td);
 
-		if (f->file_map)
-			memset(f->file_map, 0, f->num_maps * sizeof(long));
+	close_files(td);
+
+	ret = 0;
+	for_each_file(td, f, i) {
+		f->flags &= ~FIO_FILE_DONE;
+		ret = td_io_open_file(td, f);
+		if (ret)
+			break;
 	}
+
+	return ret;
 }
 
 /*
@@ -583,143 +922,229 @@ static void clear_io_state(struct thread_data *td)
  */
 static void *thread_main(void *data)
 {
+	unsigned long long runtime[2], elapsed;
 	struct thread_data *td = data;
+	int clear_state;
 
-	if (!td->use_thread)
+	if (!td->o.use_thread)
 		setsid();
 
 	td->pid = getpid();
 
-	INIT_LIST_HEAD(&td->io_u_freelist);
-	INIT_LIST_HEAD(&td->io_u_busylist);
-	INIT_LIST_HEAD(&td->io_hist_list);
-	INIT_LIST_HEAD(&td->io_log_list);
+	dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
-	if (init_io_u(td))
-		goto err;
+	INIT_FLIST_HEAD(&td->io_u_freelist);
+	INIT_FLIST_HEAD(&td->io_u_busylist);
+	INIT_FLIST_HEAD(&td->io_u_requeues);
+	INIT_FLIST_HEAD(&td->io_log_list);
+	INIT_FLIST_HEAD(&td->io_hist_list);
+	td->io_hist_tree = RB_ROOT;
+
+	td_set_runstate(td, TD_INITIALIZED);
+	fio_mutex_up(startup_mutex);
+	fio_mutex_down(td->mutex);
 
-	if (fio_setaffinity(td) == -1) {
-		td_verror(td, errno);
+	/*
+	 * the ->mutex mutex is now no longer used, close it to avoid
+	 * eating a file descriptor
+	 */
+	fio_mutex_remove(td->mutex);
+
+	/*
+	 * May alter parameters that init_io_u() will use, so we need to
+	 * do this first.
+	 */
+	if (init_iolog(td))
 		goto err;
-	}
 
-	if (td_io_init(td))
+	if (init_io_u(td))
 		goto err;
 
-	if (init_iolog(td))
+	if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
+		td_verror(td, errno, "cpu_set_affinity");
 		goto err;
+	}
+
+	if (td->o.gtod_cpu) {
+		if (fio_getaffinity(td->pid, &td->o.cpumask) == -1) {
+			td_verror(td, errno, "cpu_get_affinity");
+			goto err;
+		}
+		fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
+		if (fio_setaffinity(td) == -1) {
+			td_verror(td, errno, "cpu_set_affinity");
+			goto err;
+		}
+	}
 
-	if (td->ioprio) {
+	if (td->ioprio_set) {
 		if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
-			td_verror(td, errno);
+			td_verror(td, errno, "ioprio_set");
 			goto err;
 		}
 	}
 
-	if (nice(td->nice) == -1) {
-		td_verror(td, errno);
+	if (nice(td->o.nice) == -1) {
+		td_verror(td, errno, "nice");
 		goto err;
 	}
 
-	if (init_random_state(td))
+	if (td->o.ioscheduler && switch_ioscheduler(td))
 		goto err;
 
-	if (td->ioscheduler && switch_ioscheduler(td))
+	if (!td->o.create_serialize && setup_files(td))
 		goto err;
 
-	td_set_runstate(td, TD_INITIALIZED);
-	fio_sem_up(&startup_sem);
-	fio_sem_down(&td->mutex);
+	if (td_io_init(td))
+		goto err;
+
+	if (open_files(td))
+		goto err;
 
-	if (!td->create_serialize && setup_files(td))
+	if (init_random_map(td))
 		goto err;
 
-	gettimeofday(&td->epoch, NULL);
+	if (td->o.exec_prerun) {
+		if (system(td->o.exec_prerun) < 0)
+			goto err;
+	}
 
-	if (td->exec_prerun)
-		system(td->exec_prerun);
+	fio_gettime(&td->epoch, NULL);
+	getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
-	while (td->loops--) {
-		getrusage(RUSAGE_SELF, &td->ru_start);
-		gettimeofday(&td->start, NULL);
-		memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+	runtime[0] = runtime[1] = 0;
+	clear_state = 0;
+	while (keep_running(td)) {
+		fio_gettime(&td->start, NULL);
+		memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
+		memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 
-		if (td->ratemin)
-			memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+		if (td->o.ratemin)
+			memcpy(&td->lastrate, &td->ts.stat_sample_time,
+							sizeof(td->lastrate));
+
+		if (clear_state && clear_io_state(td))
+			break;
 
-		clear_io_state(td);
 		prune_io_piece_log(td);
 
-		if (td->io_ops->flags & FIO_CPUIO)
-			do_cpuio(td);
-		else
-			do_io(td);
+		do_io(td);
 
-		td->runtime[td->ddir] += mtime_since_now(&td->start);
-		if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
-			td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
+		clear_state = 1;
 
-		update_rusage_stat(td);
+		if (td_read(td) && td->io_bytes[DDIR_READ]) {
+			if (td->rw_end_set[DDIR_READ])
+				elapsed = utime_since(&td->start,
+						      &td->rw_end[DDIR_READ]);
+			else
+				elapsed = utime_since_now(&td->start);
+
+			runtime[DDIR_READ] += elapsed;
+		}
+		if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
+			if (td->rw_end_set[DDIR_WRITE])
+				elapsed = utime_since(&td->start,
+						      &td->rw_end[DDIR_WRITE]);
+			else
+				elapsed = utime_since_now(&td->start);
+
+			runtime[DDIR_WRITE] += elapsed;
+		}
 
 		if (td->error || td->terminate)
 			break;
 
-		if (td->verify == VERIFY_NONE)
+		if (!td->o.do_verify ||
+		    td->o.verify == VERIFY_NONE ||
+		    (td->io_ops->flags & FIO_UNIDIR))
 			continue;
 
-		clear_io_state(td);
-		gettimeofday(&td->start, NULL);
+		if (clear_io_state(td))
+			break;
+
+		fio_gettime(&td->start, NULL);
 
 		do_verify(td);
 
-		td->runtime[DDIR_READ] += mtime_since_now(&td->start);
+		runtime[DDIR_READ] += utime_since_now(&td->start);
 
 		if (td->error || td->terminate)
 			break;
 	}
 
-	if (td->bw_log)
-		finish_log(td, td->bw_log, "bw");
-	if (td->slat_log)
-		finish_log(td, td->slat_log, "slat");
-	if (td->clat_log)
-		finish_log(td, td->clat_log, "clat");
-	if (td->write_iolog)
-		write_iolog_close(td);
-	if (td->exec_postrun)
-		system(td->exec_postrun);
+	update_rusage_stat(td);
+	td->ts.runtime[0] = (runtime[0] + 999) / 1000;
+	td->ts.runtime[1] = (runtime[1] + 999) / 1000;
+	td->ts.total_run_time = mtime_since_now(&td->epoch);
+	td->ts.io_bytes[0] = td->io_bytes[0];
+	td->ts.io_bytes[1] = td->io_bytes[1];
+
+	if (td->ts.bw_log) {
+		if (td->o.bw_log_file)
+			finish_log_named(td, td->ts.bw_log, td->o.bw_log_file, "bw");
+		else
+			finish_log(td, td->ts.bw_log, "bw");
+	}
+	if (td->ts.slat_log) {
+		if (td->o.lat_log_file)
+			finish_log_named(td, td->ts.slat_log, td->o.lat_log_file, "clat");
+		else
+			finish_log(td, td->ts.slat_log, "slat");
+	}
+	if (td->ts.clat_log) {
+		if (td->o.lat_log_file)
+			finish_log_named(td, td->ts.clat_log, td->o.lat_log_file, "clat");
+		else
+			finish_log(td, td->ts.clat_log, "clat");
+	}
+	if (td->o.exec_postrun) {
+		if (system(td->o.exec_postrun) < 0)
+			log_err("fio: postrun %s failed\n", td->o.exec_postrun);
+	}
 
 	if (exitall_on_terminate)
 		terminate_threads(td->groupid);
 
 err:
-	close_files(td);
+	if (td->error)
+		printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
+							td->verror);
+	close_and_free_files(td);
 	close_ioengine(td);
 	cleanup_io_u(td);
-	td_set_runstate(td, TD_EXITED);
-	return NULL;
 
+	/*
+	 * do this very late, it will log file closing as well
+	 */
+	if (td->o.write_iolog_file)
+		write_iolog_close(td);
+
+	options_mem_free(td);
+	td_set_runstate(td, TD_EXITED);
+	return (void *) (unsigned long) td->error;
 }
 
 /*
  * We cannot pass the td data into a forked process, so attach the td and
  * pass it to the thread worker.
  */
-static void *fork_main(int shmid, int offset)
+static int fork_main(int shmid, int offset)
 {
 	struct thread_data *td;
-	void *data;
+	void *data, *ret;
 
 	data = shmat(shmid, NULL, 0);
 	if (data == (void *) -1) {
+		int __err = errno;
+
 		perror("shmat");
-		return NULL;
+		return __err;
 	}
 
 	td = data + offset * sizeof(struct thread_data);
-	thread_main(td);
+	ret = thread_main(td);
 	shmdt(data);
-	return NULL;
+	return (int) (unsigned long) ret;
 }
 
 /*
@@ -728,42 +1153,128 @@ static void *fork_main(int shmid, int offset)
 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 {
 	struct thread_data *td;
-	int i, cputhreads;
+	int i, cputhreads, realthreads, pending, status, ret;
 
 	/*
 	 * reap exited threads (TD_EXITED -> TD_REAPED)
 	 */
-	cputhreads = 0;
+	realthreads = pending = cputhreads = 0;
 	for_each_td(td, i) {
+		int flags = 0;
+
 		/*
 		 * ->io_ops is NULL for a thread that has closed its
 		 * io engine
 		 */
-		if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
+		if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 			cputhreads++;
+		else
+			realthreads++;
 
-		if (td->runstate != TD_EXITED)
+		if (!td->pid) {
+			pending++;
+			continue;
+		}
+		if (td->runstate == TD_REAPED)
+			continue;
+		if (td->o.use_thread) {
+			if (td->runstate == TD_EXITED) {
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
 			continue;
+		}
 
-		td_set_runstate(td, TD_REAPED);
+		flags = WNOHANG;
+		if (td->runstate == TD_EXITED)
+			flags = 0;
 
-		if (td->use_thread) {
-			long ret;
+		/*
+		 * check if someone quit or got killed in an unusual way
+		 */
+		ret = waitpid(td->pid, &status, flags);
+		if (ret < 0) {
+			if (errno == ECHILD) {
+				log_err("fio: pid=%d disappeared %d\n",
+						(int) td->pid, td->runstate);
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
+			perror("waitpid");
+		} else if (ret == td->pid) {
+			if (WIFSIGNALED(status)) {
+				int sig = WTERMSIG(status);
+
+				if (sig != SIGQUIT)
+					log_err("fio: pid=%d, got signal=%d\n",
+							(int) td->pid, sig);
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
+			if (WIFEXITED(status)) {
+				if (WEXITSTATUS(status) && !td->error)
+					td->error = WEXITSTATUS(status);
 
-			if (pthread_join(td->thread, (void *) &ret))
-				perror("thread_join");
-		} else
-			waitpid(td->pid, NULL, 0);
+				td_set_runstate(td, TD_REAPED);
+				goto reaped;
+			}
+		}
 
+		/*
+		 * thread is not dead, continue
+		 */
+		pending++;
+		continue;
+reaped:
 		(*nr_running)--;
-		(*m_rate) -= td->ratemin;
-		(*t_rate) -= td->rate;
+		(*m_rate) -= td->o.ratemin;
+		(*t_rate) -= td->o.rate;
+		if (!td->pid)
+			pending--;
+
+		if (td->error)
+			exit_value++;
+
+		done_secs += mtime_since_now(&td->epoch) / 1000;
 	}
 
-	if (*nr_running == cputhreads)
+	if (*nr_running == cputhreads && !pending && realthreads)
 		terminate_threads(TERMINATE_ALL);
 }
 
+static void *gtod_thread_main(void *data)
+{
+	fio_mutex_up(startup_mutex);
+
+	/*
+	 * As long as we have jobs around, update the clock. It would be nice
+	 * to have some way of NOT hammering that CPU with gettimeofday(),
+	 * but I'm not sure what to use outside of a simple CPU nop to relax
+	 * it - we don't want to lose precision.
+	 */
+	while (threads) {
+		fio_gtod_update();
+		nop;
+	}
+
+	return NULL;
+}
+
+static int fio_start_gtod_thread(void)
+{
+	if (pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL)) {
+		perror("Can't create gtod thread");
+		return 1;
+	}
+	if (pthread_detach(gtod_thread) < 0) {
+		perror("Can't detatch gtod thread");
+		return 1;
+	}
+
+	fio_mutex_down(startup_mutex);
+	return 0;
+}
+
 /*
  * Main function for kicking off and reaping jobs, as needed.
  */
@@ -776,13 +1287,25 @@ static void run_threads(void)
 	if (fio_pin_memory())
 		return;
 
+	if (fio_gtod_offload && fio_start_gtod_thread())
+		return;
+
 	if (!terse_output) {
-		printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
+		printf("Starting ");
+		if (nr_thread)
+			printf("%d thread%s", nr_thread,
+						nr_thread > 1 ? "s" : "");
+		if (nr_process) {
+			if (nr_thread)
+				printf(" and ");
+			printf("%d process%s", nr_process,
+						nr_process > 1 ? "es" : "");
+		}
+		printf("\n");
 		fflush(stdout);
 	}
 
-	signal(SIGINT, sig_handler);
-	signal(SIGALRM, sig_handler);
+	set_sig_handlers();
 
 	todo = thread_number;
 	nr_running = 0;
@@ -792,10 +1315,10 @@ static void run_threads(void)
 	for_each_td(td, i) {
 		print_status_init(td->thread_number - 1);
 
-		init_disk_util(td);
-
-		if (!td->create_serialize)
+		if (!td->o.create_serialize) {
+			init_disk_util(td);
 			continue;
+		}
 
 		/*
 		 * do file setup here so it happens sequentially,
@@ -803,12 +1326,29 @@ static void run_threads(void)
 		 * client data interspersed on disk
 		 */
 		if (setup_files(td)) {
+			exit_value++;
+			if (td->error)
+				log_err("fio: pid=%d, err=%d/%s\n",
+					(int) td->pid, td->error, td->verror);
 			td_set_runstate(td, TD_REAPED);
 			todo--;
+		} else {
+			struct fio_file *f;
+			unsigned int i;
+
+			/*
+			 * for sharing to work, each job must always open
+			 * its own files. so close them, if we opened them
+			 * for creation
+			 */
+			for_each_file(td, f, i)
+				td_io_close_file(td, f);
 		}
+
+		init_disk_util(td);
 	}
 
-	time_init();
+	set_genesis_time();
 
 	while (todo) {
 		struct thread_data *map[MAX_JOBS];
@@ -831,15 +1371,18 @@ static void run_threads(void)
 				continue;
 			}
 
-			if (td->start_delay) {
+			if (td->o.start_delay) {
 				spent = mtime_since_genesis();
 
-				if (td->start_delay * 1000 > spent)
+				if (td->o.start_delay * 1000 > spent)
 					continue;
 			}
 
-			if (td->stonewall && (nr_started || nr_running))
+			if (td->o.stonewall && (nr_started || nr_running)) {
+				dprint(FD_PROCESS, "%s: stonewall wait\n",
+							td->o.name);
 				break;
+			}
 
 			/*
 			 * Set state to created. Thread will transition
@@ -847,31 +1390,39 @@ static void run_threads(void)
 			 */
 			td_set_runstate(td, TD_CREATED);
 			map[this_jobs++] = td;
-			fio_sem_init(&startup_sem, 1);
 			nr_started++;
 
-			if (td->use_thread) {
-				if (pthread_create(&td->thread, NULL, thread_main, td)) {
-					perror("thread_create");
+			if (td->o.use_thread) {
+				dprint(FD_PROCESS, "will pthread_create\n");
+				if (pthread_create(&td->thread, NULL,
+						   thread_main, td)) {
+					perror("pthread_create");
 					nr_started--;
+					break;
 				}
+				if (pthread_detach(td->thread) < 0)
+					perror("pthread_detach");
 			} else {
-				if (fork())
-					fio_sem_down(&startup_sem);
-				else {
-					fork_main(shm_id, i);
-					exit(0);
-				}
+				pid_t pid;
+				dprint(FD_PROCESS, "will fork\n");
+				pid = fork();
+				if (!pid) {
+					int ret = fork_main(shm_id, i);
+
+					_exit(ret);
+				} else if (i == fio_debug_jobno)
+					*fio_debug_jobp = pid;
 			}
+			fio_mutex_down(startup_mutex);
 		}
 
 		/*
 		 * Wait for the started threads to transition to
 		 * TD_INITIALIZED.
 		 */
-		gettimeofday(&this_start, NULL);
+		fio_gettime(&this_start, NULL);
 		left = this_jobs;
-		while (left) {
+		while (left && !fio_abort) {
 			if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
 				break;
 
@@ -911,13 +1462,16 @@ static void run_threads(void)
 			if (td->runstate != TD_INITIALIZED)
 				continue;
 
-			td_set_runstate(td, TD_RUNNING);
+			if (in_ramp_time(td))
+				td_set_runstate(td, TD_RAMP);
+			else
+				td_set_runstate(td, TD_RUNNING);
 			nr_running++;
 			nr_started--;
-			m_rate += td->ratemin;
-			t_rate += td->rate;
+			m_rate += td->o.ratemin;
+			t_rate += td->o.rate;
 			todo--;
-			fio_sem_up(&td->mutex);
+			fio_mutex_up(td->mutex);
 		}
 
 		reap_threads(&nr_running, &t_rate, &m_rate);
@@ -937,18 +1491,54 @@ static void run_threads(void)
 
 int main(int argc, char *argv[])
 {
+	long ps;
+
+	sinit();
+
+	/*
+	 * We need locale for number printing, if it isn't set then just
+	 * go with the US format.
+	 */
+	if (!getenv("LC_NUMERIC"))
+		setlocale(LC_NUMERIC, "en_US");
+
 	if (parse_options(argc, argv))
 		return 1;
 
-	if (!thread_number) {
-		log_err("Nothing to do\n");
+	if (!thread_number)
+		return 0;
+
+	ps = sysconf(_SC_PAGESIZE);
+	if (ps < 0) {
+		log_err("Failed to get page size\n");
 		return 1;
 	}
 
-	disk_util_timer_arm();
+	page_size = ps;
+	page_mask = ps - 1;
+
+	if (write_bw_log) {
+		setup_log(&agg_io_log[DDIR_READ]);
+		setup_log(&agg_io_log[DDIR_WRITE]);
+	}
+
+	startup_mutex = fio_mutex_init(0);
+
+	set_genesis_time();
+
+	status_timer_arm();
 
 	run_threads();
-	show_run_stats();
 
-	return 0;
+	if (!fio_abort) {
+		show_run_stats();
+		if (write_bw_log) {
+			__finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
+			__finish_log(agg_io_log[DDIR_WRITE],
+					"agg-write_bw.log");
+		}
+	}
+
+	fio_mutex_remove(startup_mutex);
+	return exit_value;
 }