X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=ad2282d7516fe7fc0033b3bc721a12b16d370fe1;hp=d8c143332823bf2c488ef0799951c8ce70c71fe3;hb=4b341fca0dab0caa41d000e8db36f1530c26616f;hpb=7a93ab190e5fea78edb21affeb0189a9d61a2cd8

diff --git a/fio.c b/fio.c
index d8c14333..ad2282d7 100644
--- a/fio.c
+++ b/fio.c
@@ -36,6 +36,7 @@
 
 #include "fio.h"
 #include "hash.h"
+#include "smalloc.h"
 
 unsigned long page_mask;
 unsigned long page_size;
@@ -48,18 +49,26 @@ int nr_process = 0;
 int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
+unsigned long done_secs = 0;
 
-static struct fio_sem *startup_sem;
+static struct fio_mutex *startup_mutex;
 static volatile int fio_abort;
 static int exit_value;
+static struct itimerval itimer;
+static pthread_t gtod_thread;
 
 struct io_log *agg_io_log[2];
 
 #define TERMINATE_ALL		(-1)
 #define JOB_START_TIMEOUT	(5 * 1000)
 
-static inline void td_set_runstate(struct thread_data *td, int runstate)
+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;
 }
 
@@ -68,35 +77,111 @@ 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->o.start_delay = 0;
+
 			/*
 			 * if the thread is running, just let it exit
 			 */
 			if (td->runstate < TD_RUNNING)
 				kill(td->pid, SIGQUIT);
-			td->terminate = 1;
-			td->o.start_delay = 0;
+			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 %d\n", sig);
-			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)
+{
+	struct thread_options *o = &td->o;
+
+	/*
+	 * If some rate setting was given, we need to check it
+	 */
+	if (o->rate || o->ratemin || o->rate_iops || o->rate_iops_min)
+		return 1;
+
+	return 0;
+}
+
 /*
  * Check if we are above the minimum rate given.
  */
@@ -107,12 +192,6 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 	unsigned long spent;
 	unsigned long rate;
 
-	/*
-	 * No minimum rate set, always ok
-	 */
-	if (!td->o.ratemin && !td->o.rate_iops_min)
-		return 0;
-
 	/*
 	 * allow a 2 second settle period in the beginning
 	 */
@@ -141,12 +220,16 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 			 * check bandwidth specified rate
 			 */
 			if (bytes < td->rate_bytes) {
-				log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
+				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);
+				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;
 				}
 			}
@@ -155,12 +238,17 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 			 * 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);
+				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);
+				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);
 				}
 			}
 		}
@@ -188,7 +276,7 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-	struct list_head *entry, *n;
+	struct flist_head *entry, *n;
 	struct io_u *io_u;
 	int r;
 
@@ -203,8 +291,8 @@ static void cleanup_pending_aio(struct thread_data *td)
 	 * 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);
 
 			/*
 			 * if the io_u isn't in flight, then that generally
@@ -272,6 +360,12 @@ 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 previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
@@ -303,13 +397,15 @@ static void do_verify(struct thread_data *td)
 
 	io_u = NULL;
 	while (!td->terminate) {
-		int ret2;
+		int ret2, full;
 
 		io_u = __get_io_u(td);
 		if (!io_u)
 			break;
 
-		if (runtime_exceeded(td, &io_u->start_time)) {
+		update_tv_cache(td);
+
+		if (runtime_exceeded(td, &td->tv_cache)) {
 			put_io_u(td, io_u);
 			td->terminate = 1;
 			break;
@@ -340,7 +436,7 @@ static void do_verify(struct thread_data *td)
 				 * zero read, fail
 				 */
 				if (!bytes) {
-					td_verror(td, ENODATA, "full resid");
+					td_verror(td, EIO, "full resid");
 					put_io_u(td, io_u);
 					break;
 				}
@@ -348,7 +444,6 @@ static void do_verify(struct thread_data *td)
 				io_u->xfer_buflen = io_u->resid;
 				io_u->xfer_buf += bytes;
 				io_u->offset += bytes;
-				f->last_completed_pos = io_u->offset;
 
 				td->ts.short_io_u[io_u->ddir]++;
 
@@ -384,19 +479,25 @@ sync_done:
 		 * if we can queue more, do so. but check if there are
 		 * completed io_u's first.
 		 */
-		min_events = 0;
-		if (queue_full(td) || ret == FIO_Q_BUSY) {
-			min_events = 1;
+		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;
 
-			if (td->cur_depth > td->o.iodepth_low)
-				min_events = td->cur_depth - td->o.iodepth_low;
+			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));
 		}
-
-		/*
-		 * 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)
+		if (ret < 0)
 			break;
 	}
 
@@ -417,19 +518,21 @@ sync_done:
  */
 static void do_io(struct thread_data *td)
 {
-	struct timeval s;
 	unsigned long usec;
 	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[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;
+		int ret2, full;
 
 		if (td->terminate)
 			break;
@@ -438,14 +541,26 @@ static void do_io(struct thread_data *td)
 		if (!io_u)
 			break;
 
-		memcpy(&s, &io_u->start_time, sizeof(s));
+		update_tv_cache(td);
 
-		if (runtime_exceeded(td, &s)) {
+		if (runtime_exceeded(td, &td->tv_cache)) {
 			put_io_u(td, io_u);
 			td->terminate = 1;
 			break;
 		}
 
+		/*
+		 * 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);
 		switch (ret) {
 		case FIO_Q_COMPLETED:
@@ -459,7 +574,7 @@ static void do_io(struct thread_data *td)
 				 * zero read, fail
 				 */
 				if (!bytes) {
-					td_verror(td, ENODATA, "full resid");
+					td_verror(td, EIO, "full resid");
 					put_io_u(td, io_u);
 					break;
 				}
@@ -467,7 +582,6 @@ static void do_io(struct thread_data *td)
 				io_u->xfer_buflen = io_u->resid;
 				io_u->xfer_buf += bytes;
 				io_u->offset += bytes;
-				f->last_completed_pos = io_u->offset;
 
 				td->ts.short_io_u[io_u->ddir]++;
 
@@ -477,7 +591,9 @@ static void do_io(struct thread_data *td)
 				requeue_io_u(td, &io_u);
 			} else {
 sync_done:
-				fio_gettime(&comp_time, NULL);
+				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;
@@ -510,21 +626,26 @@ sync_done:
 		/*
 		 * See if we need to complete some commands
 		 */
-		if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
-			min_evts = 0;
-			if (queue_full(td) || ret == FIO_Q_BUSY) {
+		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 (td->cur_depth > td->o.iodepth_low)
-					min_evts = td->cur_depth - td->o.iodepth_low;
-			}
+			if (should_check_rate(td))
+				fio_gettime(&comp_time, NULL);
 
-			fio_gettime(&comp_time, NULL);
-			bytes_done = io_u_queued_complete(td, min_evts);
-			if (bytes_done < 0)
-				break;
+			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));
 		}
 
+		if (ret < 0)
+			break;
 		if (!bytes_done)
 			continue;
 
@@ -533,15 +654,17 @@ sync_done:
 		 * of completions except the very first one which may look
 		 * a little bursty
 		 */
-		usec = utime_since(&s, &comp_time);
+		if (!in_ramp_time(td) && should_check_rate(td)) {
+			usec = utime_since(&td->tv_cache, &comp_time);
 
-		rate_throttle(td, usec, bytes_done);
+			rate_throttle(td, usec, bytes_done);
 
-		if (check_min_rate(td, &comp_time)) {
-			if (exitall_on_terminate)
-				terminate_threads(td->groupid);
-			td_verror(td, ENODATA, "check_min_rate");
-			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 (td->o.thinktime) {
@@ -552,7 +675,7 @@ sync_done:
 				int left;
 
 				if (td->o.thinktime_spin)
-					__usec_sleep(td->o.thinktime_spin);
+					usec_spin(td->o.thinktime_spin);
 
 				left = td->o.thinktime - td->o.thinktime_spin;
 				if (left)
@@ -561,6 +684,10 @@ sync_done:
 		}
 	}
 
+	if (td->o.fill_device && td->error == ENOSPC) {
+		td->error = 0;
+		td->terminate = 1;
+	}
 	if (!td->error) {
 		struct fio_file *f;
 
@@ -579,88 +706,80 @@ sync_done:
 		}
 	} 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);
 	}
 
 	free_io_mem(td);
 }
 
-/*
- * "randomly" fill the buffer contents
- */
-static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
-{
-	long *ptr = io_u->buf;
-
-	if (!td->o.zero_buffers) {
-		while ((void *) ptr - io_u->buf < max_bs) {
-			*ptr = rand() * GOLDEN_RATIO_PRIME;
-			ptr++;
-		}
-	} else
-		memset(ptr, 0, max_bs);
-}
-
 static int init_io_u(struct thread_data *td)
 {
-	unsigned long long buf_size;
 	struct io_u *io_u;
 	unsigned int max_bs;
 	int i, max_units;
 	char *p;
 
-	if (td->io_ops->flags & FIO_SYNCIO)
-		max_units = 1;
-	else
-		max_units = td->o.iodepth;
-
+	max_units = td->o.iodepth;
 	max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
-	buf_size = (unsigned long long) max_bs * (unsigned long long) max_units;
-	buf_size += page_mask;
-	if (buf_size != (size_t) buf_size) {
+	td->orig_buffer_size = (unsigned long long) max_bs
+					* (unsigned long long) max_units;
+
+	if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
+		unsigned long bs;
+
+		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;
 	}
 
-	td->orig_buffer_size = buf_size;
-
-	if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
-		td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(td->o.hugepage_size - 1);
-	else if (td->orig_buffer_size & page_mask)
-		td->orig_buffer_size = (td->orig_buffer_size + page_mask) & ~page_mask;
-
 	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);
 
-		io_u->buf = p + max_bs * i;
+		if (!(td->io_ops->flags & FIO_NOIO)) {
+			io_u->buf = p + max_bs * i;
 
-		if (td_write(td))
-			fill_io_buf(td, io_u, max_bs);
+			if (td_write(td) && !td->o.refill_buffers)
+				io_u_fill_buffer(td, io_u, max_bs);
+		}
 
 		io_u->index = i;
 		io_u->flags = IO_U_F_FREE;
-		list_add(&io_u->list, &td->io_u_freelist);
+		flist_add(&io_u->list, &td->io_u_freelist);
 	}
 
-	io_u_init_timeout();
-
 	return 0;
 }
 
@@ -678,7 +797,8 @@ static int switch_ioscheduler(struct thread_data *td)
 	f = fopen(tmp, "r+");
 	if (!f) {
 		if (errno == ENOENT) {
-			log_err("fio: os or kernel doesn't support IO scheduler switching\n");
+			log_err("fio: os or kernel doesn't support IO scheduler"
+				" switching\n");
 			return 0;
 		}
 		td_verror(td, errno, "fopen iosched");
@@ -719,12 +839,29 @@ static int switch_ioscheduler(struct thread_data *td)
 	return 0;
 }
 
-static int clear_io_state(struct thread_data *td)
+static int keep_running(struct thread_data *td)
 {
-	struct fio_file *f;
-	unsigned int i;
-	int ret;
+	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;
 
+	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;
@@ -734,17 +871,47 @@ static int clear_io_state(struct thread_data *td)
 
 	td->last_was_sync = 0;
 
-	for_each_file(td, f, i)
-		td_io_close_file(td, f);
+	/*
+	 * reset file done count if we are to start over
+	 */
+	if (td->o.time_based || td->o.loops)
+		td->nr_done_files = 0;
 
-	ret = 0;
-	for_each_file(td, f, i) {
-		ret = td_io_open_file(td, f);
-		if (ret)
-			break;
+	/*
+	 * Set the same seed to get repeatable runs
+	 */
+	td_fill_rand_seeds(td);
+}
+
+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 void clear_io_state(struct thread_data *td)
+{
+	struct fio_file *f;
+	unsigned int i;
+
+	reset_io_counters(td);
 
-	return ret;
+	close_files(td);
+	for_each_file(td, f, i)
+		f->flags &= ~FIO_FILE_DONE;
 }
 
 /*
@@ -753,9 +920,8 @@ static int clear_io_state(struct thread_data *td)
  */
 static void *thread_main(void *data)
 {
-	unsigned long long runtime[2];
+	unsigned long long runtime[2], elapsed;
 	struct thread_data *td = data;
-	unsigned long elapsed;
 	int clear_state;
 
 	if (!td->o.use_thread)
@@ -763,48 +929,69 @@ static void *thread_main(void *data)
 
 	td->pid = getpid();
 
-	INIT_LIST_HEAD(&td->io_u_freelist);
-	INIT_LIST_HEAD(&td->io_u_busylist);
-	INIT_LIST_HEAD(&td->io_u_requeues);
-	INIT_LIST_HEAD(&td->io_log_list);
-	INIT_LIST_HEAD(&td->io_hist_list);
+	dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
+
+	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);
+	dprint(FD_MUTEX, "up startup_mutex\n");
+	fio_mutex_up(startup_mutex);
+	dprint(FD_MUTEX, "wait on td->mutex\n");
+	fio_mutex_down(td->mutex);
+	dprint(FD_MUTEX, "done waiting on td->mutex\n");
+
+	/*
+	 * 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 (init_io_u(td))
-		goto err_sem;
+		goto err;
 
-	if (fio_setaffinity(td) == -1) {
+	if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
 		td_verror(td, errno, "cpu_set_affinity");
-		goto err_sem;
+		goto err;
 	}
 
-	if (init_iolog(td))
-		goto err_sem;
+	/*
+	 * If we have a gettimeofday() thread, make sure we exclude that
+	 * thread from this job
+	 */
+	if (td->o.gtod_cpu) {
+		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, "ioprio_set");
-			goto err_sem;
+			goto err;
 		}
 	}
 
 	if (nice(td->o.nice) == -1) {
 		td_verror(td, errno, "nice");
-		goto err_sem;
+		goto err;
 	}
 
 	if (td->o.ioscheduler && switch_ioscheduler(td))
-		goto err_sem;
-
-	td_set_runstate(td, TD_INITIALIZED);
-	fio_sem_up(startup_sem);
-	fio_sem_down(td->mutex);
-
-	/*
-	 * the ->mutex semaphore is now no longer used, close it to avoid
-	 * eating a file descriptor
-	 */
-	fio_sem_remove(td->mutex);
+		goto err;
 
 	if (!td->o.create_serialize && setup_files(td))
 		goto err;
@@ -812,9 +999,6 @@ static void *thread_main(void *data)
 	if (td_io_init(td))
 		goto err;
 
-	if (open_files(td))
-		goto err;
-
 	if (init_random_map(td))
 		goto err;
 
@@ -824,20 +1008,21 @@ static void *thread_main(void *data)
 	}
 
 	fio_gettime(&td->epoch, NULL);
-	memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
 	getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
 	runtime[0] = runtime[1] = 0;
 	clear_state = 0;
-	while (td->o.time_based || td->o.loops--) {
+	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->o.ratemin)
-			memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
+			memcpy(&td->lastrate, &td->ts.stat_sample_time,
+							sizeof(td->lastrate));
 
-		if (clear_state && clear_io_state(td))
-			break;
+		if (clear_state)
+			clear_io_state(td);
 
 		prune_io_piece_log(td);
 
@@ -847,7 +1032,8 @@ static void *thread_main(void *data)
 
 		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]);
+				elapsed = utime_since(&td->start,
+						      &td->rw_end[DDIR_READ]);
 			else
 				elapsed = utime_since_now(&td->start);
 
@@ -855,21 +1041,23 @@ static void *thread_main(void *data)
 		}
 		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]);
+				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->o.verify == VERIFY_NONE)
+		if (!td->o.do_verify ||
+		    td->o.verify == VERIFY_NONE ||
+		    (td->io_ops->flags & FIO_UNIDIR))
 			continue;
 
-		if (clear_io_state(td))
-			break;
+		clear_io_state(td);
 
 		fio_gettime(&td->start, NULL);
 
@@ -882,20 +1070,33 @@ static void *thread_main(void *data)
 	}
 
 	update_rusage_stat(td);
-	td->ts.runtime[0] = runtime[0] / 1000;
-	td->ts.runtime[1] = runtime[1] / 1000;
+	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)
-		finish_log(td, td->ts.bw_log, "bw");
-	if (td->ts.slat_log)
-		finish_log(td, td->ts.slat_log, "slat");
-	if (td->ts.clat_log)
-		finish_log(td, td->ts.clat_log, "clat");
-	if (td->o.write_iolog_file)
-		write_iolog_close(td);
+	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);
@@ -906,16 +1107,27 @@ static void *thread_main(void *data)
 
 err:
 	if (td->error)
-		printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
-	close_files(td);
+		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);
+
+	if (td->o.cpumask_set) {
+		int ret = fio_cpuset_exit(&td->o.cpumask);
+
+		td_verror(td, ret, "fio_cpuset_exit");
+	}
+
+	/*
+	 * 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;
-err_sem:
-	fio_sem_up(startup_sem);
-	goto err;
 }
 
 /*
@@ -947,12 +1159,12 @@ static int 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, pending, status, ret;
+	int i, cputhreads, realthreads, pending, status, ret;
 
 	/*
 	 * reap exited threads (TD_EXITED -> TD_REAPED)
 	 */
-	pending = cputhreads = 0;
+	realthreads = pending = cputhreads = 0;
 	for_each_td(td, i) {
 		int flags = 0;
 
@@ -962,8 +1174,14 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		 */
 		if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
 			cputhreads++;
+		else
+			realthreads++;
 
-		if (!td->pid || td->runstate == TD_REAPED)
+		if (!td->pid) {
+			pending++;
+			continue;
+		}
+		if (td->runstate == TD_REAPED)
 			continue;
 		if (td->o.use_thread) {
 			if (td->runstate == TD_EXITED) {
@@ -983,7 +1201,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		ret = waitpid(td->pid, &status, flags);
 		if (ret < 0) {
 			if (errno == ECHILD) {
-				log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
+				log_err("fio: pid=%d disappeared %d\n",
+						(int) td->pid, td->runstate);
 				td_set_runstate(td, TD_REAPED);
 				goto reaped;
 			}
@@ -993,7 +1212,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 				int sig = WTERMSIG(status);
 
 				if (sig != SIGQUIT)
-					log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+					log_err("fio: pid=%d, got signal=%d\n",
+							(int) td->pid, sig);
 				td_set_runstate(td, TD_REAPED);
 				goto reaped;
 			}
@@ -1012,26 +1232,62 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		pending++;
 		continue;
 reaped:
-		if (td->o.use_thread) {
-			long ret;
-
-			if (pthread_join(td->thread, (void *) &ret))
-				perror("pthread_join");
-		}
-
 		(*nr_running)--;
 		(*m_rate) -= td->o.ratemin;
 		(*t_rate) -= td->o.rate;
-		pending--;
+		if (!td->pid)
+			pending--;
 
 		if (td->error)
 			exit_value++;
+
+		done_secs += mtime_since_now(&td->epoch) / 1000;
 	}
 
-	if (*nr_running == cputhreads && !pending)
+	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)
+{
+	int ret;
+
+	ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
+	if (ret) {
+		log_err("Can't create gtod thread: %s\n", strerror(ret));
+		return 1;
+	}
+
+	ret = pthread_detach(gtod_thread);
+	if (ret) {
+		log_err("Can't detatch gtod thread: %s\n", strerror(ret));
+		return 1;
+	}
+
+	dprint(FD_MUTEX, "wait on startup_mutex\n");
+	fio_mutex_down(startup_mutex);
+	dprint(FD_MUTEX, "done waiting on startup_mutex\n");
+	return 0;
+}
+
 /*
  * Main function for kicking off and reaping jobs, as needed.
  */
@@ -1044,21 +1300,25 @@ static void run_threads(void)
 	if (fio_pin_memory())
 		return;
 
+	if (fio_gtod_offload && fio_start_gtod_thread())
+		return;
+
 	if (!terse_output) {
 		printf("Starting ");
 		if (nr_thread)
-			printf("%d thread%s", nr_thread, nr_thread > 1 ? "s" : "");
+			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("%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;
@@ -1081,9 +1341,21 @@ static void run_threads(void)
 		if (setup_files(td)) {
 			exit_value++;
 			if (td->error)
-				log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
+				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);
@@ -1119,8 +1391,11 @@ static void run_threads(void)
 					continue;
 			}
 
-			if (td->o.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
@@ -1131,19 +1406,35 @@ static void run_threads(void)
 			nr_started++;
 
 			if (td->o.use_thread) {
-				if (pthread_create(&td->thread, NULL, thread_main, td)) {
-					perror("thread_create");
+				int ret;
+
+				dprint(FD_PROCESS, "will pthread_create\n");
+				ret = pthread_create(&td->thread, NULL,
+							thread_main, td);
+				if (ret) {
+					log_err("pthread_create: %s\n",
+							strerror(ret));
 					nr_started--;
 					break;
 				}
+				ret = pthread_detach(td->thread);
+				if (ret)
+					log_err("pthread_detach: %s",
+							strerror(ret));
 			} else {
-				if (!fork()) {
+				pid_t pid;
+				dprint(FD_PROCESS, "will fork\n");
+				pid = fork();
+				if (!pid) {
 					int ret = fork_main(shm_id, i);
 
-					exit(ret);
-				}
+					_exit(ret);
+				} else if (i == fio_debug_jobno)
+					*fio_debug_jobp = pid;
 			}
-			fio_sem_down(startup_sem);
+			dprint(FD_MUTEX, "wait on startup_mutex\n");
+			fio_mutex_down(startup_mutex);
+			dprint(FD_MUTEX, "done waiting on startup_mutex\n");
 		}
 
 		/*
@@ -1192,13 +1483,16 @@ 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->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);
@@ -1220,6 +1514,8 @@ 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.
@@ -1247,22 +1543,23 @@ int main(int argc, char *argv[])
 		setup_log(&agg_io_log[DDIR_WRITE]);
 	}
 
-	startup_sem = fio_sem_init(0);
+	startup_mutex = fio_mutex_init(0);
 
 	set_genesis_time();
 
-	disk_util_timer_arm();
+	status_timer_arm();
 
 	run_threads();
 
 	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");
+			__finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
+			__finish_log(agg_io_log[DDIR_WRITE],
+					"agg-write_bw.log");
 		}
 	}
 
-	fio_sem_remove(startup_sem);
+	fio_mutex_remove(startup_mutex);
 	return exit_value;
 }