X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=405346b6f310788728b4223d96ba9b5d7d9aa046;hp=241e64e6cee1c0f3e018357933f2447d3daf2443;hb=b2a151925a91f38aeb298d693687a47269ad4e94;hpb=22f78b320a8d2ffa32b5736fe754c108a8d21525

diff --git a/fio.c b/fio.c
index 241e64e6..405346b6 100644
--- a/fio.c
+++ b/fio.c
@@ -2,6 +2,7 @@
  * fio - the flexible io tester
  *
  * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
+ * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
  *
  *  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
@@ -43,13 +44,18 @@ int thread_number = 0;
 static char run_str[MAX_JOBS + 1];
 int shm_id = 0;
 static struct timeval genesis;
+int temp_stall_ts;
+char *fio_inst_prefix = _INST_PREFIX;
 
 static void print_thread_status(void);
 
 extern unsigned long long mlock_size;
 
 /*
- * thread life cycle
+ * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
+ * will never back again. It may cycle between running/verififying/fsyncing.
+ * Once the thread reaches TD_EXITED, it is just waiting for the core to
+ * reap it.
  */
 enum {
 	TD_NOT_CREATED = 0,
@@ -57,13 +63,14 @@ enum {
 	TD_INITIALIZED,
 	TD_RUNNING,
 	TD_VERIFYING,
+	TD_FSYNCING,
 	TD_EXITED,
 	TD_REAPED,
 };
 
 #define should_fsync(td)	((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
 
-static sem_t startup_sem;
+static volatile int startup_sem;
 
 #define TERMINATE_ALL		(-1)
 #define JOB_START_TIMEOUT	(5 * 1000)
@@ -98,23 +105,32 @@ static void sig_handler(int sig)
 	}
 }
 
-static int random_map_free(struct thread_data *td, unsigned long long block)
+/*
+ * The ->file_map[] contains a map of blocks we have or have not done io
+ * to yet. Used to make sure we cover the entire range in a fair fashion.
+ */
+static int random_map_free(struct thread_data *td, struct fio_file *f,
+			   unsigned long long block)
 {
-	unsigned int idx = RAND_MAP_IDX(td, block);
-	unsigned int bit = RAND_MAP_BIT(td, block);
+	unsigned int idx = RAND_MAP_IDX(td, f, block);
+	unsigned int bit = RAND_MAP_BIT(td, f, block);
 
-	return (td->file_map[idx] & (1UL << bit)) == 0;
+	return (f->file_map[idx] & (1UL << bit)) == 0;
 }
 
-static int get_next_free_block(struct thread_data *td, unsigned long long *b)
+/*
+ * Return the next free block in the map.
+ */
+static int get_next_free_block(struct thread_data *td, struct fio_file *f,
+			       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]);
+	while ((*b) * td->min_bs < f->file_size) {
+		if (f->file_map[i] != -1UL) {
+			*b += ffz(f->file_map[i]);
 			return 0;
 		}
 
@@ -125,7 +141,11 @@ static int get_next_free_block(struct thread_data *td, unsigned long long *b)
 	return 1;
 }
 
-static void mark_random_map(struct thread_data *td, struct io_u *io_u)
+/*
+ * Mark a given offset as used in the map.
+ */
+static void mark_random_map(struct thread_data *td, struct fio_file *f,
+			    struct io_u *io_u)
 {
 	unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
 	unsigned int blocks = 0;
@@ -133,15 +153,15 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 	while (blocks < (io_u->buflen / td->min_bs)) {
 		unsigned int idx, bit;
 
-		if (!random_map_free(td, block))
+		if (!random_map_free(td, f, block))
 			break;
 
-		idx = RAND_MAP_IDX(td, block);
-		bit = RAND_MAP_BIT(td, block);
+		idx = RAND_MAP_IDX(td, f, block);
+		bit = RAND_MAP_BIT(td, f, block);
 
-		assert(idx < td->num_maps);
+		assert(idx < f->num_maps);
 
-		td->file_map[idx] |= (1UL << bit);
+		f->file_map[idx] |= (1UL << bit);
 		block++;
 		blocks++;
 	}
@@ -150,7 +170,13 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 		io_u->buflen = blocks * td->min_bs;
 }
 
-static int get_next_offset(struct thread_data *td, unsigned long long *offset)
+/*
+ * For random io, generate a random new block and see if it's used. Repeat
+ * until we find a free one. For sequential io, just return the end of
+ * the last io issued.
+ */
+static int get_next_offset(struct thread_data *td, struct fio_file *f,
+			   unsigned long long *offset)
 {
 	unsigned long long b, rb;
 	long r;
@@ -160,21 +186,21 @@ static int get_next_offset(struct thread_data *td, unsigned long long *offset)
 		int loops = 50;
 
 		do {
-			lrand48_r(&td->random_state, &r);
+			r = os_random_long(&td->random_state);
 			b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-			rb = b + (td->file_offset / td->min_bs);
+			rb = b + (f->file_offset / td->min_bs);
 			loops--;
-		} while (!random_map_free(td, rb) && loops);
+		} while (!random_map_free(td, f, rb) && loops);
 
 		if (!loops) {
-			if (get_next_free_block(td, &b))
+			if (get_next_free_block(td, f, &b))
 				return 1;
 		}
 	} else
-		b = td->last_pos / td->min_bs;
+		b = f->last_pos / td->min_bs;
 
-	*offset = (b * td->min_bs) + td->file_offset;
-	if (*offset > td->real_file_size)
+	*offset = (b * td->min_bs) + f->file_offset;
+	if (*offset > f->file_size)
 		return 1;
 
 	return 0;
@@ -188,17 +214,28 @@ static unsigned int get_next_buflen(struct thread_data *td)
 	if (td->min_bs == td->max_bs)
 		buflen = td->min_bs;
 	else {
-		lrand48_r(&td->bsrange_state, &r);
+		r = os_random_long(&td->bsrange_state);
 		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])
+	if (buflen > td->io_size - td->this_io_bytes[td->ddir]) {
+		/*
+		 * if using direct/raw io, we may not be able to
+		 * shrink the size. so just fail it.
+		 */
+		if (td->io_ops->flags & FIO_RAWIO)
+			return 0;
+
 		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;
@@ -221,7 +258,7 @@ 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);
+			fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
 			if (rate_quit)
 				terminate_threads(td->groupid);
 			return 1;
@@ -250,7 +287,7 @@ static void fill_random_bytes(struct thread_data *td,
 	double r;
 
 	while (len) {
-		drand48_r(&td->verify_state, &r);
+		r = os_random_double(&td->verify_state);
 
 		/*
 		 * lrand48_r seems to be broken and only fill the bottom
@@ -273,8 +310,8 @@ static void hexdump(void *buffer, int len)
 	int i;
 
 	for (i = 0; i < len; i++)
-		printf("%02x", p[i]);
-	printf("\n");
+		fprintf(f_out, "%02x", p[i]);
+	fprintf(f_out, "\n");
 }
 
 static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
@@ -286,8 +323,8 @@ static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
 	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);
+		log_err("crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
+		log_err("crc32: wanted %lx, got %lx\n", hdr->crc32, c);
 		return 1;
 	}
 
@@ -304,7 +341,7 @@ static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
 	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);
+		log_err("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;
@@ -326,7 +363,7 @@ static int verify_io_u(struct io_u *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);
+		log_err("Bad verify type %d\n", hdr->verify_type);
 		ret = 1;
 	}
 
@@ -347,6 +384,11 @@ static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
 	memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
 }
 
+/*
+ * Return the data direction for the next io_u. If the job is a
+ * mixed read/write workload, check the rwmix cycle and switch if
+ * necessary.
+ */
 static int get_rw_ddir(struct thread_data *td)
 {
 	if (td_rw(td)) {
@@ -360,11 +402,11 @@ static int get_rw_ddir(struct thread_data *td)
 		 * Check if it's time to seed a new data direction.
 		 */
 		if (elapsed >= td->rwmixcycle) {
-			unsigned long v;
+			int v;
 			long r;
 
-			lrand48_r(&td->random_state, &r);
-			v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
+			r = os_random_long(&td->rwmix_state);
+			v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
 			if (v < td->rwmixread)
 				td->rwmix_ddir = DDIR_READ;
 			else
@@ -405,7 +447,7 @@ static void populate_io_u(struct thread_data *td, struct io_u *io_u)
 
 static int td_io_prep(struct thread_data *td, struct io_u *io_u)
 {
-	if (td->io_prep && td->io_prep(td, io_u))
+	if (td->io_ops->prep && td->io_ops->prep(td, io_u))
 		return 1;
 
 	return 0;
@@ -413,12 +455,14 @@ static int td_io_prep(struct thread_data *td, struct io_u *io_u)
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
+	io_u->file = NULL;
 	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)
+static int fill_io_u(struct thread_data *td, struct fio_file *f,
+		     struct io_u *io_u)
 {
 	/*
 	 * If using an iolog, grab next piece if any available.
@@ -429,7 +473,7 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	/*
 	 * No log, let the seq/rand engine retrieve the next position.
 	 */
-	if (!get_next_offset(td, &io_u->offset)) {
+	if (!get_next_offset(td, f, &io_u->offset)) {
 		io_u->buflen = get_next_buflen(td);
 
 		if (io_u->buflen) {
@@ -441,6 +485,7 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 			if (td->write_iolog)
 				write_iolog_put(td, io_u);
 
+			io_u->file = f;
 			return 0;
 		}
 	}
@@ -467,7 +512,11 @@ struct io_u *__get_io_u(struct thread_data *td)
 	return io_u;
 }
 
-static struct io_u *get_io_u(struct thread_data *td)
+/*
+ * Return an io_u to be processed. Gets a buflen and offset, sets direction,
+ * etc. The returned io_u is fully ready to be prepped and submitted.
+ */
+static struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
 {
 	struct io_u *io_u;
 
@@ -477,16 +526,22 @@ static struct io_u *get_io_u(struct thread_data *td)
 
 	if (td->zone_bytes >= td->zone_size) {
 		td->zone_bytes = 0;
-		td->last_pos += td->zone_skip;
+		f->last_pos += td->zone_skip;
 	}
 
-	if (fill_io_u(td, io_u)) {
+	if (fill_io_u(td, f, 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 + io_u->offset > f->file_size) {
+		if (td->io_ops->flags & FIO_RAWIO) {
+			put_io_u(td, io_u);
+			return NULL;
+		}
+
+		io_u->buflen = f->file_size - io_u->offset;
+	}
 
 	if (!io_u->buflen) {
 		put_io_u(td, io_u);
@@ -494,9 +549,9 @@ static struct io_u *get_io_u(struct thread_data *td)
 	}
 
 	if (!td->read_iolog && !td->sequential)
-		mark_random_map(td, io_u);
+		mark_random_map(td, f, io_u);
 
-	td->last_pos += io_u->buflen;
+	f->last_pos += io_u->buflen;
 
 	if (td->verify != VERIFY_NONE)
 		populate_io_u(td, io_u);
@@ -512,7 +567,6 @@ static struct io_u *get_io_u(struct thread_data *td)
 
 static inline void td_set_runstate(struct thread_data *td, int runstate)
 {
-	td->old_runstate = td->runstate;
 	td->runstate = runstate;
 }
 
@@ -535,10 +589,31 @@ static int get_next_verify(struct thread_data *td, struct io_u *io_u)
 	return 1;
 }
 
-static int sync_td(struct thread_data *td)
+static struct fio_file *get_next_file(struct thread_data *td)
 {
-	if (td->io_sync)
-		return td->io_sync(td);
+	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;
+}
+
+static int td_io_sync(struct thread_data *td, struct fio_file *f)
+{
+	if (td->io_ops->sync)
+		return td->io_ops->sync(td, f);
 
 	return 0;
 }
@@ -546,14 +621,14 @@ static int sync_td(struct thread_data *td)
 static int io_u_getevents(struct thread_data *td, int min, int max,
 			  struct timespec *t)
 {
-	return td->io_getevents(td, min, max, t);
+	return td->io_ops->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);
+	return td->io_ops->queue(td, io_u);
 }
 
 #define iocb_time(iocb)	((unsigned long) (iocb)->data)
@@ -597,13 +672,17 @@ static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
 	icd->bytes_done[0] = icd->bytes_done[1] = 0;
 
 	for (i = 0; i < icd->nr; i++) {
-		io_u = td->io_event(td, i);
+		io_u = td->io_ops->event(td, i);
 
 		io_completed(td, io_u, icd);
 		put_io_u(td, io_u);
 	}
 }
 
+/*
+ * 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};
@@ -624,11 +703,11 @@ 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);
 		}
@@ -657,11 +736,16 @@ static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
 	return ret;
 }
 
+/*
+ * The main verify engine. Runs over the writes we previusly submitted,
+ * reads the blocks back in, and checks the crc/md5 of the data.
+ */
 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;
 
 	td_set_runstate(td, TD_VERIFYING);
@@ -683,6 +767,12 @@ 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;
@@ -709,7 +799,7 @@ static void do_verify(struct thread_data *td)
 			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);
@@ -737,11 +827,44 @@ static void do_verify(struct thread_data *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;
+
+	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};
@@ -752,7 +875,11 @@ static void do_io(struct thread_data *td)
 		if (td->terminate)
 			break;
 
-		io_u = get_io_u(td);
+		f = get_next_file(td);
+		if (!f)
+			break;
+
+		io_u = get_io_u(td, f);
 		if (!io_u)
 			break;
 
@@ -812,40 +939,25 @@ static void do_io(struct thread_data *td)
 
 		if (should_fsync(td) && td->fsync_blocks &&
 		    (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
-			sync_td(td);
+			td_io_sync(td, f);
 	}
 
 	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);
+	if (should_fsync(td) && td->end_fsync) {
+		td_set_runstate(td, TD_FSYNCING);
+		for_each_file(td, f, i)
+			td_io_sync(td, f);
+	}
 }
 
 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;
-	}
+	if (td->io_ops->init)
+		return td->io_ops->init(td);
+
+	return 0;
 }
 
 static void cleanup_io_u(struct thread_data *td)
@@ -870,7 +982,7 @@ static void cleanup_io_u(struct thread_data *td)
 	} 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);
+		log_err("Bad memory type %d\n", td->mem_type);
 
 	td->orig_buffer = NULL;
 }
@@ -881,7 +993,10 @@ static int init_io_u(struct thread_data *td)
 	int i, max_units;
 	char *p;
 
-	if (td->io_engine & FIO_SYNCIO)
+	if (td->io_ops->flags & FIO_CPUIO)
+		return 0;
+
+	if (td->io_ops->flags & FIO_SYNCIO)
 		max_units = 1;
 	else
 		max_units = td->iodepth;
@@ -929,300 +1044,6 @@ static int init_io_u(struct thread_data *td)
 	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 (fstat(td->fd, &st) == -1) {
-		td_verror(td, errno);
-		return 1;
-	}
-
-	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)
-{
-	unsigned long long bytes;
-	int r;
-
-	r = blockdev_size(td->fd, &bytes);
-	if (r) {
-		td_verror(td, r);
-		return 1;
-	}
-
-	td->real_file_size = bytes;
-
-	/*
-	 * 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;
-
-	td->file_size -= td->file_offset;
-	return 0;
-}
-
-static int get_file_size(struct thread_data *td)
-{
-	int ret = 0;
-
-	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;
-
-	if (ret)
-		return ret;
-
-	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;
-	}
-
-	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->zone_size)
-		td->zone_size = td->io_size;
-
-	td->total_io_size = td->io_size * td->loops;
-	return 0;
-}
-
-static int setup_file_mmap(struct thread_data *td)
-{
-	int flags;
-
-	if (td_rw(td))
-		flags = PROT_READ | PROT_WRITE;
-	else if (td_write(td)) {
-		flags = PROT_WRITE;
-
-		if (td->verify != VERIFY_NONE)
-			flags |= PROT_READ;
-	} else
-		flags = PROT_READ;
-
-	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->invalidate_cache) {
-		if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
-			td_verror(td, errno);
-			return 1;
-		}
-	}
-
-	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;
-		}
-	}
-
-	return 0;
-}
-
-static int setup_file_plain(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;
-		}
-	}
-
-	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;
-		}
-	}
-
-	return 0;
-}
-
-static int setup_file(struct thread_data *td)
-{
-	struct stat st;
-	int flags = 0;
-
-	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;
-		}
-	}
-
-	if (td->odirect)
-		flags |= O_DIRECT;
-
-	if (td_write(td) || td_rw(td)) {
-		if (td->filetype == FIO_TYPE_FILE) {
-			if (!td->overwrite)
-				flags |= O_TRUNC;
-
-			flags |= O_CREAT;
-		}
-		if (td->sync_io)
-			flags |= O_SYNC;
-
-		flags |= O_RDWR;
-
-		td->fd = open(td->file_name, flags, 0600);
-	} else {
-		if (td->filetype == FIO_TYPE_CHAR)
-			flags |= O_RDWR;
-		else
-			flags |= O_RDONLY;
-
-		td->fd = open(td->file_name, flags);
-	}
-
-	if (td->fd == -1) {
-		td_verror(td, errno);
-		return 1;
-	}
-
-	if (get_file_size(td))
-		return 1;
-
-	if (td->io_engine != FIO_MMAPIO)
-		return setup_file_plain(td);
-	else
-		return setup_file_mmap(td);
-}
-
 static int switch_ioscheduler(struct thread_data *td)
 {
 	char tmp[256], tmp2[128];
@@ -1261,7 +1082,7 @@ 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 +1094,27 @@ 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;
@@ -1320,7 +1150,7 @@ static void *thread_main(void *data)
 		}
 	}
 
-	if (nice(td->nice) < 0) {
+	if (nice(td->nice) == -1) {
 		td_verror(td, errno);
 		goto err;
 	}
@@ -1332,10 +1162,10 @@ 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;
 
 	gettimeofday(&td->epoch, NULL);
@@ -1354,7 +1184,10 @@ 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])
@@ -1394,20 +1227,18 @@ 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;
@@ -1425,13 +1256,13 @@ static void *fork_main(int shmid, int offset)
 	return NULL;
 }
 
+/*
+ * Sets the status of the 'td' in the printed status map.
+ */
 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 = '_';
@@ -1460,6 +1291,9 @@ static void check_str_update(struct thread_data *td)
 		case TD_VERIFYING:
 			c = 'V';
 			break;
+		case TD_FSYNCING:
+			c = 'F';
+			break;
 		case TD_CREATED:
 			c = 'C';
 			break;
@@ -1470,13 +1304,15 @@ static void check_str_update(struct thread_data *td)
 			c = 'P';
 			break;
 		default:
-			printf("state %d\n", td->runstate);
+			log_err("state %d\n", td->runstate);
 	}
 
 	run_str[td->thread_number - 1] = c;
-	td->old_runstate = td->runstate;
 }
 
+/*
+ * Convert seconds to a printable string.
+ */
 static void eta_to_str(char *str, int eta_sec)
 {
 	unsigned int d, h, m, s;
@@ -1505,6 +1341,9 @@ static void eta_to_str(char *str, int eta_sec)
 	str += sprintf(str, "%02ds", s);
 }
 
+/*
+ * Best effort calculation of the estimated pending runtime of a job.
+ */
 static int thread_eta(struct thread_data *td, unsigned long elapsed)
 {
 	unsigned long long bytes_total, bytes_done;
@@ -1559,10 +1398,10 @@ static int thread_eta(struct thread_data *td, unsigned long elapsed)
 		else if (t_eta)
 			eta_sec = t_eta;
 		else
-			eta_sec = INT_MAX;
+			eta_sec = 0;
 	} else {
 		/*
-		 * thread is already done
+		 * thread is already done or waiting for fsync
 		 */
 		eta_sec = 0;
 	}
@@ -1570,25 +1409,34 @@ static int thread_eta(struct thread_data *td, unsigned long elapsed)
 	return eta_sec;
 }
 
+/*
+ * Print status of the jobs we know about. This includes rate estimates,
+ * ETA, thread state, etc.
+ */
 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;
+	int i, nr_running, nr_pending, t_rate, m_rate, *eta_secs, eta_sec;
 	char eta_str[32];
 	double perc = 0.0;
 
+	if (temp_stall_ts || terse_output)
+		return;
+
 	eta_secs = malloc(thread_number * sizeof(int));
 	memset(eta_secs, 0, thread_number * sizeof(int));
 
-	nr_running = t_rate = m_rate = 0;
+	nr_pending = 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){
+		if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING||
+		    td->runstate == TD_FSYNCING) {
 			nr_running++;
 			t_rate += td->rate;
 			m_rate += td->ratemin;
-		}
+		} else if (td->runstate < TD_RUNNING)
+			nr_pending++;
 
 		if (elapsed >= 3)
 			eta_secs[i] = thread_eta(td, elapsed);
@@ -1618,10 +1466,13 @@ static void print_thread_status(void)
 		eta_to_str(eta_str, eta_sec);
 	}
 
-	printf("Threads now running (%d)", nr_running);
+	if (!nr_running && !nr_pending)
+		return;
+
+	printf("Threads running: %d", nr_running);
 	if (m_rate || t_rate)
 		printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
-	if (eta_sec != INT_MAX) {
+	if (eta_sec != INT_MAX && nr_running) {
 		perc *= 100.0;
 		printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
 	}
@@ -1630,16 +1481,22 @@ static void print_thread_status(void)
 	free(eta_secs);
 }
 
+/*
+ * 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)
 {
-	int i;
+	int i, cputhreads;
 
 	/*
 	 * reap exited threads (TD_EXITED -> TD_REAPED)
 	 */
-	for (i = 0; i < thread_number; i++) {
+	for (i = 0, cputhreads = 0; i < thread_number; i++) {
 		struct thread_data *td = &threads[i];
 
+		if (td->io_ops->flags & FIO_CPUIO)
+			cputhreads++;
+
 		if (td->runstate != TD_EXITED)
 			continue;
 
@@ -1657,6 +1514,9 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 		(*m_rate) -= td->ratemin;
 		(*t_rate) -= td->rate;
 	}
+
+	if (*nr_running == cputhreads)
+		terminate_threads(TERMINATE_ALL);
 }
 
 static void fio_unpin_memory(void *pinned)
@@ -1670,7 +1530,7 @@ static void fio_unpin_memory(void *pinned)
 
 static void *fio_pin_memory(void)
 {
-	long pagesize, pages;
+	unsigned long long phys_mem;
 	void *ptr;
 
 	if (!mlock_size)
@@ -1679,15 +1539,11 @@ static void *fio_pin_memory(void)
 	/*
 	 * 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);
+	phys_mem = os_phys_mem();
+	if (phys_mem) {
+		if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
+			mlock_size = phys_mem - 128 * 1024 * 1024;
+			fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
 		}
 	}
 
@@ -1705,6 +1561,9 @@ static void *fio_pin_memory(void)
 	return ptr;
 }
 
+/*
+ * Main function for kicking off and reaping jobs, as needed.
+ */
 static void run_threads(void)
 {
 	struct thread_data *td;
@@ -1714,8 +1573,10 @@ static void run_threads(void)
 
 	mlocked_mem = fio_pin_memory();
 
-	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);
@@ -1740,7 +1601,7 @@ 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--;
 		}
@@ -1787,7 +1648,7 @@ 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 +1658,7 @@ 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 +1670,6 @@ 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 +1695,7 @@ 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)
@@ -1848,7 +1708,6 @@ 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];
 
@@ -1861,7 +1720,7 @@ 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);
@@ -1885,7 +1744,7 @@ int main(int argc, char *argv[])
 		return 1;
 
 	if (!thread_number) {
-		printf("Nothing to do\n");
+		log_err("Nothing to do\n");
 		return 1;
 	}