X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=ed14c41f0c645bc5c9378b84d183e89c81955d34;hp=241e64e6cee1c0f3e018357933f2447d3daf2443;hb=21972cdef7db6b557c2d56138f8434658a9d1e49;hpb=22f78b320a8d2ffa32b5736fe754c108a8d21525

diff --git a/fio.c b/fio.c
index 241e64e6..ed14c41f 100644
--- a/fio.c
+++ b/fio.c
@@ -2,11 +2,14 @@
  * fio - the flexible io tester
  *
  * 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
@@ -40,41 +43,21 @@
 
 int groupid = 0;
 int thread_number = 0;
-static char run_str[MAX_JOBS + 1];
 int shm_id = 0;
-static struct timeval genesis;
-
-static void print_thread_status(void);
-
-extern unsigned long long mlock_size;
-
-/*
- * thread life cycle
- */
-enum {
-	TD_NOT_CREATED = 0,
-	TD_CREATED,
-	TD_INITIALIZED,
-	TD_RUNNING,
-	TD_VERIFYING,
-	TD_EXITED,
-	TD_REAPED,
-};
-
-#define should_fsync(td)	((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
+int temp_stall_ts;
+const char *fio_inst_prefix = _INST_PREFIX;
 
-static sem_t startup_sem;
+static volatile int startup_sem;
 
 #define TERMINATE_ALL		(-1)
 #define JOB_START_TIMEOUT	(5 * 1000)
 
 static void terminate_threads(int group_id)
 {
+	struct thread_data *td;
 	int i;
 
-	for (i = 0; i < thread_number; i++) {
-		struct thread_data *td = &threads[i];
-
+	for_each_td(td, i) {
 		if (group_id == TERMINATE_ALL || groupid == td->groupid) {
 			td->terminate = 1;
 			td->start_delay = 0;
@@ -98,107 +81,9 @@ static void sig_handler(int sig)
 	}
 }
 
-static int random_map_free(struct thread_data *td, unsigned long long block)
-{
-	unsigned int idx = RAND_MAP_IDX(td, block);
-	unsigned int bit = RAND_MAP_BIT(td, block);
-
-	return (td->file_map[idx] & (1UL << bit)) == 0;
-}
-
-static int get_next_free_block(struct thread_data *td, unsigned long long *b)
-{
-	int i;
-
-	*b = 0;
-	i = 0;
-	while ((*b) * td->min_bs < td->io_size) {
-		if (td->file_map[i] != -1UL) {
-			*b += ffz(td->file_map[i]);
-			return 0;
-		}
-
-		*b += BLOCKS_PER_MAP;
-		i++;
-	}
-
-	return 1;
-}
-
-static void mark_random_map(struct thread_data *td, struct io_u *io_u)
-{
-	unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
-	unsigned int blocks = 0;
-
-	while (blocks < (io_u->buflen / td->min_bs)) {
-		unsigned int idx, bit;
-
-		if (!random_map_free(td, block))
-			break;
-
-		idx = RAND_MAP_IDX(td, block);
-		bit = RAND_MAP_BIT(td, block);
-
-		assert(idx < td->num_maps);
-
-		td->file_map[idx] |= (1UL << bit);
-		block++;
-		blocks++;
-	}
-
-	if ((blocks * td->min_bs) < io_u->buflen)
-		io_u->buflen = blocks * td->min_bs;
-}
-
-static int get_next_offset(struct thread_data *td, unsigned long long *offset)
-{
-	unsigned long long b, rb;
-	long r;
-
-	if (!td->sequential) {
-		unsigned long long max_blocks = td->io_size / td->min_bs;
-		int loops = 50;
-
-		do {
-			lrand48_r(&td->random_state, &r);
-			b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-			rb = b + (td->file_offset / td->min_bs);
-			loops--;
-		} while (!random_map_free(td, rb) && loops);
-
-		if (!loops) {
-			if (get_next_free_block(td, &b))
-				return 1;
-		}
-	} else
-		b = td->last_pos / td->min_bs;
-
-	*offset = (b * td->min_bs) + td->file_offset;
-	if (*offset > td->real_file_size)
-		return 1;
-
-	return 0;
-}
-
-static unsigned int get_next_buflen(struct thread_data *td)
-{
-	unsigned int buflen;
-	long r;
-
-	if (td->min_bs == td->max_bs)
-		buflen = td->min_bs;
-	else {
-		lrand48_r(&td->bsrange_state, &r);
-		buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
-		buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
-	}
-
-	if (buflen > td->io_size - td->this_io_bytes[td->ddir])
-		buflen = td->io_size - td->this_io_bytes[td->ddir];
-
-	return buflen;
-}
-
+/*
+ * Check if we are above the minimum rate given.
+ */
 static int check_min_rate(struct thread_data *td, struct timeval *now)
 {
 	unsigned long spent;
@@ -221,9 +106,7 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 
 		rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
 		if (rate < td->ratemin) {
-			printf("%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
-			if (rate_quit)
-				terminate_threads(td->groupid);
+			fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
 			return 1;
 		}
 	}
@@ -243,367 +126,36 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 	return 0;
 }
 
-static void fill_random_bytes(struct thread_data *td,
-			      unsigned char *p, unsigned int len)
-{
-	unsigned int todo;
-	double r;
-
-	while (len) {
-		drand48_r(&td->verify_state, &r);
-
-		/*
-		 * lrand48_r seems to be broken and only fill the bottom
-		 * 32-bits, even on 64-bit archs with 64-bit longs
-		 */
-		todo = sizeof(r);
-		if (todo > len)
-			todo = len;
-
-		memcpy(p, &r, todo);
-
-		len -= todo;
-		p += todo;
-	}
-}
-
-static void hexdump(void *buffer, int len)
-{
-	unsigned char *p = buffer;
-	int i;
-
-	for (i = 0; i < len; i++)
-		printf("%02x", p[i]);
-	printf("\n");
-}
-
-static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
-{
-	unsigned char *p = (unsigned char *) io_u->buf;
-	unsigned long c;
-
-	p += sizeof(*hdr);
-	c = crc32(p, hdr->len - sizeof(*hdr));
-
-	if (c != hdr->crc32) {
-		fprintf(stderr, "crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
-		fprintf(stderr, "crc32: wanted %lx, got %lx\n", hdr->crc32, c);
-		return 1;
-	}
-
-	return 0;
-}
-
-static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
-{
-	unsigned char *p = (unsigned char *) io_u->buf;
-	struct md5_ctx md5_ctx;
-
-	memset(&md5_ctx, 0, sizeof(md5_ctx));
-	p += sizeof(*hdr);
-	md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
-
-	if (memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash))) {
-		fprintf(stderr, "md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
-		hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
-		hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
-		return 1;
-	}
-
-	return 0;
-}
-
-static int verify_io_u(struct io_u *io_u)
-{
-	struct verify_header *hdr = (struct verify_header *) io_u->buf;
-	int ret;
-
-	if (hdr->fio_magic != FIO_HDR_MAGIC)
-		return 1;
-
-	if (hdr->verify_type == VERIFY_MD5)
-		ret = verify_io_u_md5(hdr, io_u);
-	else if (hdr->verify_type == VERIFY_CRC32)
-		ret = verify_io_u_crc32(hdr, io_u);
-	else {
-		fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
-		ret = 1;
-	}
-
-	return ret;
-}
-
-static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
-{
-	hdr->crc32 = crc32(p, len);
-}
-
-static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
-{
-	struct md5_ctx md5_ctx;
-
-	memset(&md5_ctx, 0, sizeof(md5_ctx));
-	md5_update(&md5_ctx, p, len);
-	memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
-}
-
-static int get_rw_ddir(struct thread_data *td)
-{
-	if (td_rw(td)) {
-		struct timeval now;
-		unsigned long elapsed;
-
-		gettimeofday(&now, NULL);
-	 	elapsed = mtime_since_now(&td->rwmix_switch);
-
-		/*
-		 * Check if it's time to seed a new data direction.
-		 */
-		if (elapsed >= td->rwmixcycle) {
-			unsigned long v;
-			long r;
-
-			lrand48_r(&td->random_state, &r);
-			v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
-			if (v < td->rwmixread)
-				td->rwmix_ddir = DDIR_READ;
-			else
-				td->rwmix_ddir = DDIR_WRITE;
-			memcpy(&td->rwmix_switch, &now, sizeof(now));
-		}
-		return td->rwmix_ddir;
-	} else if (td_read(td))
-		return DDIR_READ;
-	else
-		return DDIR_WRITE;
-}
-
-/*
- * fill body of io_u->buf with random data and add a header with the
- * crc32 or md5 sum of that data.
- */
-static void populate_io_u(struct thread_data *td, struct io_u *io_u)
-{
-	unsigned char *p = (unsigned char *) io_u->buf;
-	struct verify_header hdr;
-
-	hdr.fio_magic = FIO_HDR_MAGIC;
-	hdr.len = io_u->buflen;
-	p += sizeof(hdr);
-	fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
-
-	if (td->verify == VERIFY_MD5) {
-		fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
-		hdr.verify_type = VERIFY_MD5;
-	} else {
-		fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
-		hdr.verify_type = VERIFY_CRC32;
-	}
-
-	memcpy(io_u->buf, &hdr, sizeof(hdr));
-}
-
-static int td_io_prep(struct thread_data *td, struct io_u *io_u)
-{
-	if (td->io_prep && td->io_prep(td, io_u))
-		return 1;
-
-	return 0;
-}
-
-void put_io_u(struct thread_data *td, struct io_u *io_u)
-{
-	list_del(&io_u->list);
-	list_add(&io_u->list, &td->io_u_freelist);
-	td->cur_depth--;
-}
-
-static int fill_io_u(struct thread_data *td, struct io_u *io_u)
-{
-	/*
-	 * If using an iolog, grab next piece if any available.
-	 */
-	if (td->read_iolog)
-		return read_iolog_get(td, io_u);
-
-	/*
-	 * No log, let the seq/rand engine retrieve the next position.
-	 */
-	if (!get_next_offset(td, &io_u->offset)) {
-		io_u->buflen = get_next_buflen(td);
-
-		if (io_u->buflen) {
-			io_u->ddir = get_rw_ddir(td);
-
-			/*
-			 * If using a write iolog, store this entry.
-			 */
-			if (td->write_iolog)
-				write_iolog_put(td, io_u);
-
-			return 0;
-		}
-	}
-
-	return 1;
-}
-
-#define queue_full(td)	list_empty(&(td)->io_u_freelist)
-
-struct io_u *__get_io_u(struct thread_data *td)
-{
-	struct io_u *io_u = NULL;
-
-	if (!queue_full(td)) {
-		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
-
-		io_u->error = 0;
-		io_u->resid = 0;
-		list_del(&io_u->list);
-		list_add(&io_u->list, &td->io_u_busylist);
-		td->cur_depth++;
-	}
-
-	return io_u;
-}
-
-static struct io_u *get_io_u(struct thread_data *td)
-{
-	struct io_u *io_u;
-
-	io_u = __get_io_u(td);
-	if (!io_u)
-		return NULL;
-
-	if (td->zone_bytes >= td->zone_size) {
-		td->zone_bytes = 0;
-		td->last_pos += td->zone_skip;
-	}
-
-	if (fill_io_u(td, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
-
-	if (io_u->buflen + io_u->offset > td->real_file_size)
-		io_u->buflen = td->real_file_size - io_u->offset;
-
-	if (!io_u->buflen) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
-
-	if (!td->read_iolog && !td->sequential)
-		mark_random_map(td, io_u);
-
-	td->last_pos += io_u->buflen;
-
-	if (td->verify != VERIFY_NONE)
-		populate_io_u(td, io_u);
-
-	if (td_io_prep(td, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
-
-	gettimeofday(&io_u->start_time, NULL);
-	return io_u;
-}
-
 static inline void td_set_runstate(struct thread_data *td, int runstate)
 {
-	td->old_runstate = td->runstate;
 	td->runstate = runstate;
 }
 
-static int get_next_verify(struct thread_data *td, struct io_u *io_u)
-{
-	struct io_piece *ipo;
-
-	if (!list_empty(&td->io_hist_list)) {
-		ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
-
-		list_del(&ipo->list);
-
-		io_u->offset = ipo->offset;
-		io_u->buflen = ipo->len;
-		io_u->ddir = DDIR_READ;
-		free(ipo);
-		return 0;
-	}
-
-	return 1;
-}
-
-static int sync_td(struct thread_data *td)
-{
-	if (td->io_sync)
-		return td->io_sync(td);
-
-	return 0;
-}
-
-static int io_u_getevents(struct thread_data *td, int min, int max,
-			  struct timespec *t)
-{
-	return td->io_getevents(td, min, max, t);
-}
-
-static int io_u_queue(struct thread_data *td, struct io_u *io_u)
-{
-	gettimeofday(&io_u->issue_time, NULL);
-
-	return td->io_queue(td, io_u);
-}
-
-#define iocb_time(iocb)	((unsigned long) (iocb)->data)
-
-static void io_completed(struct thread_data *td, struct io_u *io_u,
-			 struct io_completion_data *icd)
+static struct fio_file *get_next_file(struct thread_data *td)
 {
-	struct timeval e;
-	unsigned long msec;
-
-	gettimeofday(&e, NULL);
-
-	if (!io_u->error) {
-		unsigned int bytes = io_u->buflen - io_u->resid;
-		const int idx = io_u->ddir;
+	unsigned int old_next_file = td->next_file;
+	struct fio_file *f;
 
-		td->io_blocks[idx]++;
-		td->io_bytes[idx] += bytes;
-		td->zone_bytes += bytes;
-		td->this_io_bytes[idx] += bytes;
-
-		msec = mtime_since(&io_u->issue_time, &e);
-
-		add_clat_sample(td, idx, msec);
-		add_bw_sample(td, idx);
+	do {
+		f = &td->files[td->next_file];
 
-		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
-			log_io_piece(td, io_u);
+		td->next_file++;
+		if (td->next_file >= td->nr_files)
+			td->next_file = 0;
 
-		icd->bytes_done[idx] += bytes;
-	} else
-		icd->error = io_u->error;
-}
-
-static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
-{
-	struct io_u *io_u;
-	int i;
-
-	icd->error = 0;
-	icd->bytes_done[0] = icd->bytes_done[1] = 0;
+		if (f->fd != -1)
+			break;
 
-	for (i = 0; i < icd->nr; i++) {
-		io_u = td->io_event(td, i);
+		f = NULL;
+	} while (td->next_file != old_next_file);
 
-		io_completed(td, io_u, icd);
-		put_io_u(td, io_u);
-	}
+	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};
@@ -615,7 +167,7 @@ static void cleanup_pending_aio(struct thread_data *td)
 	/*
 	 * get immediately available events, if any
 	 */
-	r = io_u_getevents(td, 0, td->cur_depth, &ts);
+	r = td_io_getevents(td, 0, td->cur_depth, &ts);
 	if (r > 0) {
 		icd.nr = r;
 		ios_completed(td, &icd);
@@ -624,18 +176,18 @@ static void cleanup_pending_aio(struct thread_data *td)
 	/*
 	 * now cancel remaining active events
 	 */
-	if (td->io_cancel) {
+	if (td->io_ops->cancel) {
 		list_for_each_safe(entry, n, &td->io_u_busylist) {
 			io_u = list_entry(entry, struct io_u, list);
 
-			r = td->io_cancel(td, io_u);
+			r = td->io_ops->cancel(td, io_u);
 			if (!r)
 				put_io_u(td, io_u);
 		}
 	}
 
 	if (td->cur_depth) {
-		r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
+		r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
 		if (r > 0) {
 			icd.nr = r;
 			ios_completed(td, &icd);
@@ -643,26 +195,70 @@ static void cleanup_pending_aio(struct thread_data *td)
 	}
 }
 
-static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
+/*
+ * Helper to handle the final sync of a file. Works just like the normal
+ * io path, just does everything sync.
+ */
+static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
-	struct io_u *v_io_u = *io_u;
-	int ret = 0;
+	struct io_u *io_u = __get_io_u(td);
+	struct io_completion_data icd;
+	int ret;
+
+	if (!io_u)
+		return 1;
+
+	io_u->ddir = DDIR_SYNC;
+	io_u->file = f;
+
+	if (td_io_prep(td, io_u)) {
+		put_io_u(td, io_u);
+		return 1;
+	}
 
-	if (v_io_u) {
-		ret = verify_io_u(v_io_u);
-		put_io_u(td, v_io_u);
-		*io_u = NULL;
+	ret = td_io_queue(td, io_u);
+	if (ret) {
+		td_verror(td, io_u->error);
+		put_io_u(td, io_u);
+		return 1;
+	}
+
+	ret = td_io_getevents(td, 1, td->cur_depth, NULL);
+	if (ret < 0) {
+		td_verror(td, ret);
+		return 1;
+	}
+
+	icd.nr = ret;
+	ios_completed(td, &icd);
+	if (icd.error) {
+		td_verror(td, icd.error);
+		return 1;
 	}
 
-	return ret;
+	return 0;
 }
 
-static void do_verify(struct thread_data *td)
+/*
+ * The main verify engine. Runs over the writes we previusly submitted,
+ * reads the blocks back in, and checks the crc/md5 of the data.
+ */
+void do_verify(struct thread_data *td)
 {
 	struct timeval t;
 	struct io_u *io_u, *v_io_u = NULL;
 	struct io_completion_data icd;
-	int ret;
+	struct fio_file *f;
+	int ret, 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);
+	}
 
 	td_set_runstate(td, TD_VERIFYING);
 
@@ -683,15 +279,21 @@ static void do_verify(struct thread_data *td)
 			break;
 		}
 
+		f = get_next_file(td);
+		if (!f)
+			break;
+
+		io_u->file = f;
+
 		if (td_io_prep(td, io_u)) {
 			put_io_u(td, io_u);
 			break;
 		}
 
-		ret = io_u_queue(td, io_u);
+		ret = td_io_queue(td, io_u);
 		if (ret) {
+			td_verror(td, io_u->error);
 			put_io_u(td, io_u);
-			td_verror(td, ret);
 			break;
 		}
 
@@ -702,14 +304,14 @@ static void do_verify(struct thread_data *td)
 		if (do_io_u_verify(td, &v_io_u))
 			break;
 
-		ret = io_u_getevents(td, 1, 1, NULL);
+		ret = td_io_getevents(td, 1, 1, NULL);
 		if (ret != 1) {
 			if (ret < 0)
 				td_verror(td, ret);
 			break;
 		}
 
-		v_io_u = td->io_event(td, 0);
+		v_io_u = td->io_ops->event(td, 0);
 		icd.nr = 1;
 		icd.error = 0;
 		io_completed(td, v_io_u, &icd);
@@ -737,490 +339,195 @@ static void do_verify(struct thread_data *td)
 	td_set_runstate(td, TD_RUNNING);
 }
 
-static void do_io(struct thread_data *td)
+/*
+ * 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 io_completion_data icd;
-	struct timeval s, e;
-	unsigned long usec;
-
-	while (td->this_io_bytes[td->ddir] < td->io_size) {
-		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
-		struct timespec *timeout;
-		int ret, min_evts = 0;
-		struct io_u *io_u;
-
-		if (td->terminate)
-			break;
-
-		io_u = get_io_u(td);
-		if (!io_u)
-			break;
-
-		memcpy(&s, &io_u->start_time, sizeof(s));
-
-		ret = io_u_queue(td, io_u);
-		if (ret) {
-			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 (td->cur_depth < td->iodepth) {
-			timeout = &ts;
-			min_evts = 0;
-		} else {
-			timeout = NULL;
-			min_evts = 1;
-		}
-
-		ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
-		if (ret < 0) {
-			td_verror(td, ret);
-			break;
-		} else if (!ret)
-			continue;
-
-		icd.nr = ret;
-		ios_completed(td, &icd);
-		if (icd.error) {
-			td_verror(td, icd.error);
-			break;
-		}
+	struct timeval e;
+	int split = 100 / td->cpuload;
+	int i = 0;
 
-		/*
-		 * the rate is batched for now, it should work for batches
-		 * of completions except the very first one which may look
-		 * a little bursty
-		 */
+	while (!td->terminate) {
 		gettimeofday(&e, NULL);
-		usec = utime_since(&s, &e);
-
-		rate_throttle(td, usec, icd.bytes_done[td->ddir]);
-
-		if (check_min_rate(td, &e)) {
-			td_verror(td, ENOMEM);
-			break;
-		}
 
 		if (runtime_exceeded(td, &e))
 			break;
 
-		if (td->thinktime)
-			usec_sleep(td, td->thinktime);
-
-		if (should_fsync(td) && td->fsync_blocks &&
-		    (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
-			sync_td(td);
-	}
-
-	if (td->cur_depth)
-		cleanup_pending_aio(td);
-
-	if (should_fsync(td) && td->end_fsync)
-		sync_td(td);
-}
-
-static void cleanup_io(struct thread_data *td)
-{
-	if (td->io_cleanup)
-		td->io_cleanup(td);
-}
-
-static int init_io(struct thread_data *td)
-{
-	if (td->io_engine == FIO_SYNCIO)
-		return fio_syncio_init(td);
-	else if (td->io_engine == FIO_MMAPIO)
-		return fio_mmapio_init(td);
-	else if (td->io_engine == FIO_LIBAIO)
-		return fio_libaio_init(td);
-	else if (td->io_engine == FIO_POSIXAIO)
-		return fio_posixaio_init(td);
-	else if (td->io_engine == FIO_SGIO)
-		return fio_sgio_init(td);
-	else if (td->io_engine == FIO_SPLICEIO)
-		return fio_spliceio_init(td);
-	else {
-		fprintf(stderr, "bad io_engine %d\n", td->io_engine);
-		return 1;
-	}
-}
-
-static void cleanup_io_u(struct thread_data *td)
-{
-	struct list_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);
-
-		list_del(&io_u->list);
-		free(io_u);
-	}
-
-	if (td->mem_type == MEM_MALLOC)
-		free(td->orig_buffer);
-	else if (td->mem_type == MEM_SHM) {
-		struct shmid_ds sbuf;
-
-		shmdt(td->orig_buffer);
-		shmctl(td->shm_id, IPC_RMID, &sbuf);
-	} else if (td->mem_type == MEM_MMAP)
-		munmap(td->orig_buffer, td->orig_buffer_size);
-	else
-		fprintf(stderr, "Bad memory type %d\n", td->mem_type);
-
-	td->orig_buffer = NULL;
-}
-
-static int init_io_u(struct thread_data *td)
-{
-	struct io_u *io_u;
-	int i, max_units;
-	char *p;
-
-	if (td->io_engine & FIO_SYNCIO)
-		max_units = 1;
-	else
-		max_units = td->iodepth;
-
-	td->orig_buffer_size = td->max_bs * max_units + MASK;
-
-	if (td->mem_type == MEM_MALLOC)
-		td->orig_buffer = malloc(td->orig_buffer_size);
-	else if (td->mem_type == MEM_SHM) {
-		td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
-		if (td->shm_id < 0) {
-			td_verror(td, errno);
-			perror("shmget");
-			return 1;
-		}
-
-		td->orig_buffer = shmat(td->shm_id, NULL, 0);
-		if (td->orig_buffer == (void *) -1) {
-			td_verror(td, errno);
-			perror("shmat");
-			td->orig_buffer = NULL;
-			return 1;
-		}
-	} else if (td->mem_type == MEM_MMAP) {
-		td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
-		if (td->orig_buffer == MAP_FAILED) {
-			td_verror(td, errno);
-			perror("mmap");
-			td->orig_buffer = NULL;
-			return 1;
-		}
-	}
-
-	p = ALIGN(td->orig_buffer);
-	for (i = 0; i < max_units; i++) {
-		io_u = malloc(sizeof(*io_u));
-		memset(io_u, 0, sizeof(*io_u));
-		INIT_LIST_HEAD(&io_u->list);
-
-		io_u->buf = p + td->max_bs * i;
-		io_u->index = i;
-		list_add(&io_u->list, &td->io_u_freelist);
-	}
-
-	return 0;
-}
-
-static void cleanup_allocs(struct thread_data *td)
-{
-	if (td->directory)
-		free(td->directory);
-	if (td->iolog_file)
-		free(td->iolog_file);
-	if (td->exec_prerun)
-		free(td->exec_prerun);
-	if (td->exec_postrun)
-		free(td->exec_postrun);
-	if (td->ioscheduler)
-		free(td->ioscheduler);
-	if (td->sysfs_root)
-		free(td->sysfs_root);
-}
-
-static int create_file(struct thread_data *td, unsigned long long size,
-		       int extend)
-{
-	unsigned long long left;
-	unsigned int bs;
-	int r, oflags;
-	char *b;
-
-	/*
-	 * unless specifically asked for overwrite, let normal io extend it
-	 */
-	if (td_write(td) && !td->overwrite)
-		return 0;
-
-	if (!size) {
-		fprintf(stderr, "Need size for create\n");
-		td_verror(td, EINVAL);
-		return 1;
-	}
-
-	if (!extend) {
-		oflags = O_CREAT | O_TRUNC;
-		printf("%s: Laying out IO file (%LuMiB)\n", td->name, size >> 20);
-	} else {
-		oflags = O_APPEND;
-		printf("%s: Extending IO file (%Lu -> %LuMiB)\n", td->name, (td->file_size - size) >> 20, td->file_size >> 20);
-	}
-
-	td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
-	if (td->fd < 0) {
-		td_verror(td, errno);
-		return 1;
-	}
-
-	if (!extend && ftruncate(td->fd, td->file_size) == -1) {
-		td_verror(td, errno);
-		return 1;
-	}
-
-	td->io_size = td->file_size;
-	b = malloc(td->max_bs);
-	memset(b, 0, td->max_bs);
-
-	left = size;
-	while (left && !td->terminate) {
-		bs = td->max_bs;
-		if (bs > left)
-			bs = left;
-
-		r = write(td->fd, b, bs);
-
-		if (r == (int) bs) {
-			left -= bs;
-			continue;
-		} else {
-			if (r < 0)
-				td_verror(td, errno);
-			else
-				td_verror(td, EIO);
-
-			break;
-		}
-	}
-
-	if (td->terminate)
-		unlink(td->file_name);
-	else if (td->create_fsync)
-		fsync(td->fd);
-
-	close(td->fd);
-	td->fd = -1;
-	free(b);
-	return 0;
-}
-
-static int file_size(struct thread_data *td)
-{
-	struct stat st;
+		if (!(i % split))
+			__usec_sleep(10000);
+		else
+			usec_sleep(td, 10000);
 
-	if (fstat(td->fd, &st) == -1) {
-		td_verror(td, errno);
-		return 1;
+		i++;
 	}
-
-	td->real_file_size = st.st_size;
-
-	if (!td->file_size || td->file_size > td->real_file_size)
-		td->file_size = td->real_file_size;
-
-	td->file_size -= td->file_offset;
-	return 0;
 }
 
-static int bdev_size(struct thread_data *td)
+/*
+ * 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)
 {
-	unsigned long long bytes;
-	int r;
+	struct io_completion_data icd;
+	struct timeval s, e;
+	unsigned long usec;
+	struct fio_file *f;
+	int i, ret = 0;
 
-	r = blockdev_size(td->fd, &bytes);
-	if (r) {
-		td_verror(td, r);
-		return 1;
-	}
+	td_set_runstate(td, TD_RUNNING);
 
-	td->real_file_size = bytes;
+	while (td->this_io_bytes[td->ddir] < td->io_size) {
+		struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+		struct timespec *timeout;
+		int min_evts = 0;
+		struct io_u *io_u;
 
-	/*
-	 * no extend possibilities, so limit size to device size if too large
-	 */
-	if (!td->file_size || td->file_size > td->real_file_size)
-		td->file_size = td->real_file_size;
+		if (td->terminate)
+			break;
 
-	td->file_size -= td->file_offset;
-	return 0;
-}
+		f = get_next_file(td);
+		if (!f)
+			break;
 
-static int get_file_size(struct thread_data *td)
-{
-	int ret = 0;
+		io_u = get_io_u(td, f);
+		if (!io_u)
+			break;
 
-	if (td->filetype == FIO_TYPE_FILE)
-		ret = file_size(td);
-	else if (td->filetype == FIO_TYPE_BD)
-		ret = bdev_size(td);
-	else
-		td->real_file_size = -1;
+		memcpy(&s, &io_u->start_time, sizeof(s));
 
-	if (ret)
-		return ret;
+		ret = td_io_queue(td, io_u);
+		if (ret) {
+			td_verror(td, io_u->error);
+			put_io_u(td, io_u);
+			break;
+		}
 
-	if (td->file_offset > td->real_file_size) {
-		fprintf(stderr, "%s: offset extends end (%Lu > %Lu)\n", td->name, td->file_offset, td->real_file_size);
-		return 1;
-	}
+		add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 
-	td->io_size = td->file_size;
-	if (td->io_size == 0) {
-		fprintf(stderr, "%s: no io blocks\n", td->name);
-		td_verror(td, EINVAL);
-		return 1;
-	}
+		if (td->cur_depth < td->iodepth) {
+			timeout = &ts;
+			min_evts = 0;
+		} else {
+			timeout = NULL;
+			min_evts = 1;
+		}
 
-	if (!td->zone_size)
-		td->zone_size = td->io_size;
+		ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
+		if (ret < 0) {
+			td_verror(td, ret);
+			break;
+		} else if (!ret)
+			continue;
 
-	td->total_io_size = td->io_size * td->loops;
-	return 0;
-}
+		icd.nr = ret;
+		ios_completed(td, &icd);
+		if (icd.error) {
+			td_verror(td, icd.error);
+			break;
+		}
 
-static int setup_file_mmap(struct thread_data *td)
-{
-	int flags;
+		/*
+		 * 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);
+
+		rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 
-	if (td_rw(td))
-		flags = PROT_READ | PROT_WRITE;
-	else if (td_write(td)) {
-		flags = PROT_WRITE;
+		if (check_min_rate(td, &e)) {
+			if (exitall_on_terminate)
+				terminate_threads(td->groupid);
+			td_verror(td, ENOMEM);
+			break;
+		}
 
-		if (td->verify != VERIFY_NONE)
-			flags |= PROT_READ;
-	} else
-		flags = PROT_READ;
+		if (runtime_exceeded(td, &e))
+			break;
 
-	td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
-	if (td->mmap == MAP_FAILED) {
-		td->mmap = NULL;
-		td_verror(td, errno);
-		return 1;
+		if (td->thinktime)
+			usec_sleep(td, td->thinktime);
 	}
 
-	if (td->invalidate_cache) {
-		if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
-			td_verror(td, errno);
-			return 1;
-		}
-	}
+	if (!td->error) {
+		if (td->cur_depth)
+			cleanup_pending_aio(td);
 
-	if (td->sequential) {
-		if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
-			td_verror(td, errno);
-			return 1;
-		}
-	} else {
-		if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
-			td_verror(td, errno);
-			return 1;
+		if (should_fsync(td) && td->end_fsync) {
+			td_set_runstate(td, TD_FSYNCING);
+			for_each_file(td, f, i)
+				fio_io_sync(td, f);
 		}
 	}
-
-	return 0;
 }
 
-static int setup_file_plain(struct thread_data *td)
+static void cleanup_io_u(struct thread_data *td)
 {
-	if (td->invalidate_cache) {
-		if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
-			td_verror(td, errno);
-			return 1;
-		}
-	}
+	struct list_head *entry, *n;
+	struct io_u *io_u;
 
-	if (td->sequential) {
-		if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
-			td_verror(td, errno);
-			return 1;
-		}
-	} else {
-		if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
-			td_verror(td, errno);
-			return 1;
-		}
+	list_for_each_safe(entry, n, &td->io_u_freelist) {
+		io_u = list_entry(entry, struct io_u, list);
+
+		list_del(&io_u->list);
+		free(io_u);
 	}
 
-	return 0;
+	free_io_mem(td);
 }
 
-static int setup_file(struct thread_data *td)
+/*
+ * "randomly" fill the buffer contents
+ */
+static void fill_rand_buf(struct io_u *io_u, int max_bs)
 {
-	struct stat st;
-	int flags = 0;
+	int *ptr = io_u->buf;
 
-	if (stat(td->file_name, &st) == -1) {
-		if (errno != ENOENT) {
-			td_verror(td, errno);
-			return 1;
-		}
-		if (!td->create_file) {
-			td_verror(td, ENOENT);
-			return 1;
-		}
-		if (create_file(td, td->file_size, 0))
-			return 1;
-	} else if (td->filetype == FIO_TYPE_FILE) {
-		if (st.st_size < (off_t) td->file_size) {
-			if (create_file(td, td->file_size - st.st_size, 1))
-				return 1;
-		}
+	while ((void *) ptr - io_u->buf < max_bs) {
+		*ptr = rand() * 0x9e370001;
+		ptr++;
 	}
+}
 
-	if (td->odirect)
-		flags |= O_DIRECT;
+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_write(td) || td_rw(td)) {
-		if (td->filetype == FIO_TYPE_FILE) {
-			if (!td->overwrite)
-				flags |= O_TRUNC;
+	if (td->io_ops->flags & FIO_CPUIO)
+		return 0;
 
-			flags |= O_CREAT;
-		}
-		if (td->sync_io)
-			flags |= O_SYNC;
+	if (td->io_ops->flags & FIO_SYNCIO)
+		max_units = 1;
+	else
+		max_units = td->iodepth;
 
-		flags |= O_RDWR;
+	max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
+	td->orig_buffer_size = max_bs * max_units + MASK;
 
-		td->fd = open(td->file_name, flags, 0600);
-	} else {
-		if (td->filetype == FIO_TYPE_CHAR)
-			flags |= O_RDWR;
-		else
-			flags |= O_RDONLY;
+	if (allocate_io_mem(td))
+		return 1;
 
-		td->fd = open(td->file_name, flags);
-	}
+	p = ALIGN(td->orig_buffer);
+	for (i = 0; i < max_units; i++) {
+		io_u = malloc(sizeof(*io_u));
+		memset(io_u, 0, sizeof(*io_u));
+		INIT_LIST_HEAD(&io_u->list);
 
-	if (td->fd == -1) {
-		td_verror(td, errno);
-		return 1;
-	}
+		io_u->buf = p + max_bs * i;
+		if (td_write(td) || td_rw(td))
+			fill_rand_buf(io_u, max_bs);
 
-	if (get_file_size(td))
-		return 1;
+		io_u->index = i;
+		list_add(&io_u->list, &td->io_u_freelist);
+	}
 
-	if (td->io_engine != FIO_MMAPIO)
-		return setup_file_plain(td);
-	else
-		return setup_file_mmap(td);
+	return 0;
 }
 
 static int switch_ioscheduler(struct thread_data *td)
@@ -1229,6 +536,9 @@ static int switch_ioscheduler(struct thread_data *td)
 	FILE *f;
 	int ret;
 
+	if (td->io_ops->flags & FIO_CPUIO)
+		return 0;
+
 	sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 
 	f = fopen(tmp, "r+");
@@ -1261,7 +571,7 @@ static int switch_ioscheduler(struct thread_data *td)
 
 	sprintf(tmp2, "[%s]", td->ioscheduler);
 	if (!strstr(tmp, tmp2)) {
-		fprintf(stderr, "fio: io scheduler %s not found\n", td->ioscheduler);
+		log_err("fio: io scheduler %s not found\n", td->ioscheduler);
 		td_verror(td, EINVAL);
 		fclose(f);
 		return 1;
@@ -1273,18 +583,27 @@ static int switch_ioscheduler(struct thread_data *td)
 
 static void clear_io_state(struct thread_data *td)
 {
-	if (td->io_engine == FIO_SYNCIO)
-		lseek(td->fd, SEEK_SET, 0);
+	struct fio_file *f;
+	int i;
 
-	td->last_pos = 0;
 	td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
 	td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
 	td->zone_bytes = 0;
 
-	if (td->file_map)
-		memset(td->file_map, 0, td->num_maps * sizeof(long));
+	for_each_file(td, f, i) {
+		f->last_pos = 0;
+		if (td->io_ops->flags & FIO_SYNCIO)
+			lseek(f->fd, SEEK_SET, 0);
+
+		if (f->file_map)
+			memset(f->file_map, 0, f->num_maps * sizeof(long));
+	}
 }
 
+/*
+ * Entry point for the thread based jobs. The process based jobs end up
+ * here as well, after a little setup.
+ */
 static void *thread_main(void *data)
 {
 	struct thread_data *td = data;
@@ -1307,7 +626,7 @@ static void *thread_main(void *data)
 		goto err;
 	}
 
-	if (init_io(td))
+	if (td_io_init(td))
 		goto err;
 
 	if (init_iolog(td))
@@ -1320,7 +639,7 @@ static void *thread_main(void *data)
 		}
 	}
 
-	if (nice(td->nice) < 0) {
+	if (nice(td->nice) == -1) {
 		td_verror(td, errno);
 		goto err;
 	}
@@ -1332,10 +651,12 @@ static void *thread_main(void *data)
 		goto err;
 
 	td_set_runstate(td, TD_INITIALIZED);
-	sem_post(&startup_sem);
-	sem_wait(&td->mutex);
+	fio_sem_up(&startup_sem);
+	fio_sem_down(&td->mutex);
 
-	if (!td->create_serialize && setup_file(td))
+	if (!td->create_serialize && setup_files(td))
+		goto err;
+	if (open_files(td))
 		goto err;
 
 	gettimeofday(&td->epoch, NULL);
@@ -1354,7 +675,10 @@ static void *thread_main(void *data)
 		clear_io_state(td);
 		prune_io_piece_log(td);
 
-		do_io(td);
+		if (td->io_ops->flags & FIO_CPUIO)
+			do_cpuio(td);
+		else
+			do_io(td);
 
 		td->runtime[td->ddir] += mtime_since_now(&td->start);
 		if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
@@ -1385,7 +709,7 @@ static void *thread_main(void *data)
 		finish_log(td, td->slat_log, "slat");
 	if (td->clat_log)
 		finish_log(td, td->clat_log, "clat");
-	if (td->write_iolog)
+	if (td->write_iolog_file)
 		write_iolog_close(td);
 	if (td->exec_postrun)
 		system(td->exec_postrun);
@@ -1394,20 +718,18 @@ static void *thread_main(void *data)
 		terminate_threads(td->groupid);
 
 err:
-	if (td->fd != -1) {
-		close(td->fd);
-		td->fd = -1;
-	}
-	if (td->mmap)
-		munmap(td->mmap, td->file_size);
-	cleanup_allocs(td);
-	cleanup_io(td);
+	close_files(td);
+	close_ioengine(td);
 	cleanup_io_u(td);
 	td_set_runstate(td, TD_EXITED);
 	return NULL;
 
 }
 
+/*
+ * 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)
 {
 	struct thread_data *td;
@@ -1425,223 +747,32 @@ static void *fork_main(int shmid, int offset)
 	return NULL;
 }
 
-static void check_str_update(struct thread_data *td)
-{
-	char c = run_str[td->thread_number - 1];
-
-	if (td->runstate == td->old_runstate)
-		return;
-
-	switch (td->runstate) {
-		case TD_REAPED:
-			c = '_';
-			break;
-		case TD_EXITED:
-			c = 'E';
-			break;
-		case TD_RUNNING:
-			if (td_rw(td)) {
-				if (td->sequential)
-					c = 'M';
-				else
-					c = 'm';
-			} else if (td_read(td)) {
-				if (td->sequential)
-					c = 'R';
-				else
-					c = 'r';
-			} else {
-				if (td->sequential)
-					c = 'W';
-				else
-					c = 'w';
-			}
-			break;
-		case TD_VERIFYING:
-			c = 'V';
-			break;
-		case TD_CREATED:
-			c = 'C';
-			break;
-		case TD_INITIALIZED:
-			c = 'I';
-			break;
-		case TD_NOT_CREATED:
-			c = 'P';
-			break;
-		default:
-			printf("state %d\n", td->runstate);
-	}
-
-	run_str[td->thread_number - 1] = c;
-	td->old_runstate = td->runstate;
-}
-
-static void eta_to_str(char *str, int eta_sec)
-{
-	unsigned int d, h, m, s;
-	static int always_d, always_h;
-
-	d = h = m = s = 0;
-
-	s = eta_sec % 60;
-	eta_sec /= 60;
-	m = eta_sec % 60;
-	eta_sec /= 60;
-	h = eta_sec % 24;
-	eta_sec /= 24;
-	d = eta_sec;
-
-	if (d || always_d) {
-		always_d = 1;
-		str += sprintf(str, "%02dd:", d);
-	}
-	if (h || always_h) {
-		always_h = 1;
-		str += sprintf(str, "%02dh:", h);
-	}
-
-	str += sprintf(str, "%02dm:", m);
-	str += sprintf(str, "%02ds", s);
-}
-
-static int thread_eta(struct thread_data *td, unsigned long elapsed)
+/*
+ * Run over the job map and reap the threads that have exited, if any.
+ */
+static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 {
-	unsigned long long bytes_total, bytes_done;
-	unsigned int eta_sec = 0;
-
-	bytes_total = td->total_io_size;
+	struct thread_data *td;
+	int i, cputhreads, pending;
 
 	/*
-	 * if writing, bytes_total will be twice the size. If mixing,
-	 * assume a 50/50 split and thus bytes_total will be 50% larger.
+	 * reap exited threads (TD_EXITED -> TD_REAPED)
 	 */
-	if (td->verify) {
-		if (td_rw(td))
-			bytes_total = bytes_total * 3 / 2;
-		else
-			bytes_total <<= 1;
-	}
-	if (td->zone_size && td->zone_skip)
-		bytes_total /= (td->zone_skip / td->zone_size);
-
-	if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
-		double perc;
-
-		bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
-		perc = (double) bytes_done / (double) bytes_total;
-		if (perc > 1.0)
-			perc = 1.0;
-
-		eta_sec = (elapsed * (1.0 / perc)) - elapsed;
-
-		if (td->timeout && eta_sec > (td->timeout - elapsed))
-			eta_sec = td->timeout - elapsed;
-	} else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
-			|| td->runstate == TD_INITIALIZED) {
-		int t_eta = 0, r_eta = 0;
-
-		/*
-		 * We can only guess - assume it'll run the full timeout
-		 * if given, otherwise assume it'll run at the specified rate.
-		 */
-		if (td->timeout)
-			t_eta = td->timeout + td->start_delay - elapsed;
-		if (td->rate) {
-			r_eta = (bytes_total / 1024) / td->rate;
-			r_eta += td->start_delay - elapsed;
-		}
-
-		if (r_eta && t_eta)
-			eta_sec = min(r_eta, t_eta);
-		else if (r_eta)
-			eta_sec = r_eta;
-		else if (t_eta)
-			eta_sec = t_eta;
-		else
-			eta_sec = INT_MAX;
-	} else {
+	pending = cputhreads = 0;
+	for_each_td(td, i) {
 		/*
-		 * thread is already done
+		 * ->io_ops is NULL for a thread that has closed its
+		 * io engine
 		 */
-		eta_sec = 0;
-	}
-
-	return eta_sec;
-}
-
-static void print_thread_status(void)
-{
-	unsigned long elapsed = time_since_now(&genesis);
-	int i, nr_running, t_rate, m_rate, *eta_secs, eta_sec;
-	char eta_str[32];
-	double perc = 0.0;
-
-	eta_secs = malloc(thread_number * sizeof(int));
-	memset(eta_secs, 0, thread_number * sizeof(int));
-
-	nr_running = t_rate = m_rate = 0;
-	for (i = 0; i < thread_number; i++) {
-		struct thread_data *td = &threads[i];
-
-		if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING){
-			nr_running++;
-			t_rate += td->rate;
-			m_rate += td->ratemin;
-		}
-
-		if (elapsed >= 3)
-			eta_secs[i] = thread_eta(td, elapsed);
-		else
-			eta_secs[i] = INT_MAX;
-
-		check_str_update(td);
-	}
-
-	if (exitall_on_terminate)
-		eta_sec = INT_MAX;
-	else
-		eta_sec = 0;
-
-	for (i = 0; i < thread_number; i++) {
-		if (exitall_on_terminate) {
-			if (eta_secs[i] < eta_sec)
-				eta_sec = eta_secs[i];
-		} else {
-			if (eta_secs[i] > eta_sec)
-				eta_sec = eta_secs[i];
-		}
-	}
-
-	if (eta_sec != INT_MAX && elapsed) {
-		perc = (double) elapsed / (double) (elapsed + eta_sec);
-		eta_to_str(eta_str, eta_sec);
-	}
-
-	printf("Threads now running (%d)", nr_running);
-	if (m_rate || t_rate)
-		printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
-	if (eta_sec != INT_MAX) {
-		perc *= 100.0;
-		printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
-	}
-	printf("\r");
-	fflush(stdout);
-	free(eta_secs);
-}
-
-static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
-{
-	int i;
+		if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
+			cputhreads++;
 
-	/*
-	 * reap exited threads (TD_EXITED -> TD_REAPED)
-	 */
-	for (i = 0; i < thread_number; i++) {
-		struct thread_data *td = &threads[i];
+		if (td->runstate != TD_EXITED) {
+			if (td->runstate < TD_RUNNING)
+				pending++;
 
-		if (td->runstate != TD_EXITED)
 			continue;
+		}
 
 		td_set_runstate(td, TD_REAPED);
 
@@ -1657,65 +788,27 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		(*m_rate) -= td->ratemin;
 		(*t_rate) -= td->rate;
 	}
-}
-
-static void fio_unpin_memory(void *pinned)
-{
-	if (pinned) {
-		if (munlock(pinned, mlock_size) < 0)
-			perror("munlock");
-		munmap(pinned, mlock_size);
-	}
-}
-
-static void *fio_pin_memory(void)
-{
-	long pagesize, pages;
-	void *ptr;
-
-	if (!mlock_size)
-		return NULL;
-
-	/*
-	 * Don't allow mlock of more than real_mem-128MB
-	 */
-	pagesize = sysconf(_SC_PAGESIZE);
-	pages = sysconf(_SC_PHYS_PAGES);
-	if (pages != -1 && pagesize != -1) {
-		unsigned long long real_mem = pages * pagesize;
-
-		if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
-			mlock_size = real_mem - 128 * 1024 * 1024;
-			printf("fio: limiting mlocked memory to %lluMiB\n",
-							mlock_size >> 20);
-		}
-	}
-
-	ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
-	if (!ptr) {
-		perror("malloc locked mem");
-		return NULL;
-	}
-	if (mlock(ptr, mlock_size) < 0) {
-		munmap(ptr, mlock_size);
-		perror("mlock");
-		return NULL;
-	}
 
-	return ptr;
+	if (*nr_running == cputhreads && !pending)
+		terminate_threads(TERMINATE_ALL);
 }
 
+/*
+ * Main function for kicking off and reaping jobs, as needed.
+ */
 static void run_threads(void)
 {
 	struct thread_data *td;
 	unsigned long spent;
 	int i, todo, nr_running, m_rate, t_rate, nr_started;
-	void *mlocked_mem;
 
-	mlocked_mem = fio_pin_memory();
+	if (fio_pin_memory())
+		return;
 
-	printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
-	fflush(stdout);
+	if (!terse_output) {
+		printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
+		fflush(stdout);
+	}
 
 	signal(SIGINT, sig_handler);
 	signal(SIGALRM, sig_handler);
@@ -1725,10 +818,8 @@ static void run_threads(void)
 	nr_started = 0;
 	m_rate = t_rate = 0;
 
-	for (i = 0; i < thread_number; i++) {
-		td = &threads[i];
-
-		run_str[td->thread_number - 1] = 'P';
+	for_each_td(td, i) {
+		print_status_init(td->thread_number - 1);
 
 		init_disk_util(td);
 
@@ -1740,13 +831,13 @@ static void run_threads(void)
 		 * we don't want X number of threads getting their
 		 * client data interspersed on disk
 		 */
-		if (setup_file(td)) {
+		if (setup_files(td)) {
 			td_set_runstate(td, TD_REAPED);
 			todo--;
 		}
 	}
 
-	gettimeofday(&genesis, NULL);
+	time_init();
 
 	while (todo) {
 		struct thread_data *map[MAX_JOBS];
@@ -1756,9 +847,7 @@ static void run_threads(void)
 		/*
 		 * create threads (TD_NOT_CREATED -> TD_CREATED)
 		 */
-		for (i = 0; i < thread_number; i++) {
-			td = &threads[i];
-
+		for_each_td(td, i) {
 			if (td->runstate != TD_NOT_CREATED)
 				continue;
 
@@ -1772,7 +861,7 @@ static void run_threads(void)
 			}
 
 			if (td->start_delay) {
-				spent = mtime_since_now(&genesis);
+				spent = mtime_since_genesis();
 
 				if (td->start_delay * 1000 > spent)
 					continue;
@@ -1787,7 +876,7 @@ static void run_threads(void)
 			 */
 			td_set_runstate(td, TD_CREATED);
 			map[this_jobs++] = td;
-			sem_init(&startup_sem, 0, 1);
+			fio_sem_init(&startup_sem, 1);
 			nr_started++;
 
 			if (td->use_thread) {
@@ -1797,7 +886,7 @@ static void run_threads(void)
 				}
 			} else {
 				if (fork())
-					sem_wait(&startup_sem);
+					fio_sem_down(&startup_sem);
 				else {
 					fork_main(shm_id, i);
 					exit(0);
@@ -1809,7 +898,6 @@ static void run_threads(void)
 		 * Wait for the started threads to transition to
 		 * TD_INITIALIZED.
 		 */
-		printf("fio: Waiting for threads to initialize...\n");
 		gettimeofday(&this_start, NULL);
 		left = this_jobs;
 		while (left) {
@@ -1835,7 +923,7 @@ static void run_threads(void)
 		}
 
 		if (left) {
-			fprintf(stderr, "fio: %d jobs failed to start\n", left);
+			log_err("fio: %d jobs failed to start\n", left);
 			for (i = 0; i < this_jobs; i++) {
 				td = map[i];
 				if (!td)
@@ -1848,10 +936,7 @@ static void run_threads(void)
 		/*
 		 * start created threads (TD_INITIALIZED -> TD_RUNNING).
 		 */
-		printf("fio: Go for launch\n");
-		for (i = 0; i < thread_number; i++) {
-			td = &threads[i];
-
+		for_each_td(td, i) {
 			if (td->runstate != TD_INITIALIZED)
 				continue;
 
@@ -1861,7 +946,7 @@ static void run_threads(void)
 			m_rate += td->ratemin;
 			t_rate += td->rate;
 			todo--;
-			sem_post(&td->mutex);
+			fio_sem_up(&td->mutex);
 		}
 
 		reap_threads(&nr_running, &t_rate, &m_rate);
@@ -1876,7 +961,7 @@ static void run_threads(void)
 	}
 
 	update_io_ticks();
-	fio_unpin_memory(mlocked_mem);
+	fio_unpin_memory();
 }
 
 int main(int argc, char *argv[])
@@ -1885,7 +970,7 @@ int main(int argc, char *argv[])
 		return 1;
 
 	if (!thread_number) {
-		printf("Nothing to do\n");
+		log_err("Nothing to do\n");
 		return 1;
 	}