X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=7ac31a0d297872fe82168e74cff63cfce1880dcb;hp=b0e91e7a1bcc2b52d479df079e671629b8cf3ac8;hb=0d29de831183dfd049c97a03008d425ce21e2fa4;hpb=97601024a53586d77a368763f08be9ee483fdc9c

diff --git a/io_u.c b/io_u.c
index b0e91e7a..ea0d46c5 100644
--- a/io_u.c
+++ b/io_u.c
@@ -3,13 +3,16 @@
 #include <string.h>
 #include <signal.h>
 #include <time.h>
+#include <assert.h>
 
 #include "fio.h"
-#include "os.h"
+#include "hash.h"
+#include "verify.h"
+#include "trim.h"
+#include "lib/rand.h"
 
 struct io_completion_data {
 	int nr;				/* input */
-	endio_handler *handler;		/* input */
 
 	int error;			/* output */
 	unsigned long bytes_done[2];	/* output */
@@ -20,63 +23,120 @@ struct io_completion_data {
  * 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)
+static int random_map_free(struct fio_file *f, const unsigned long long block)
 {
-	unsigned int idx = RAND_MAP_IDX(td, f, block);
-	unsigned int bit = RAND_MAP_BIT(td, f, block);
+	unsigned int idx = RAND_MAP_IDX(f, block);
+	unsigned int bit = RAND_MAP_BIT(f, block);
 
-	return (f->file_map[idx] & (1UL << bit)) == 0;
+	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
+
+	return (f->file_map[idx] & (1 << bit)) == 0;
 }
 
 /*
  * 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)
+static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 {
-	unsigned int min_bs = td->rw_min_bs;
+	unsigned int min_bs = td->o.rw_min_bs;
+	struct fio_file *f = io_u->file;
 	unsigned long long block;
-	unsigned int blocks;
-	unsigned int nr_blocks;
+	unsigned int blocks, nr_blocks;
+	int busy_check;
 
-	block = io_u->offset / (unsigned long long) min_bs;
-	blocks = 0;
+	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
 	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
+	blocks = 0;
+	busy_check = !(io_u->flags & IO_U_F_BUSY_OK);
 
-	while (blocks < nr_blocks) {
+	while (nr_blocks) {
+		unsigned int this_blocks, mask;
 		unsigned int idx, bit;
 
-		if (!random_map_free(td, f, block))
+		/*
+		 * If we have a mixed random workload, we may
+		 * encounter blocks we already did IO to.
+		 */
+		if (!busy_check) {
+			blocks = nr_blocks;
+			break;
+		}
+		if ((td->o.ddir_seq_nr == 1) && !random_map_free(f, block))
 			break;
 
-		idx = RAND_MAP_IDX(td, f, block);
-		bit = RAND_MAP_BIT(td, f, block);
+		idx = RAND_MAP_IDX(f, block);
+		bit = RAND_MAP_BIT(f, block);
 
 		fio_assert(td, idx < f->num_maps);
 
-		f->file_map[idx] |= (1UL << bit);
-		block++;
-		blocks++;
+		this_blocks = nr_blocks;
+		if (this_blocks + bit > BLOCKS_PER_MAP)
+			this_blocks = BLOCKS_PER_MAP - bit;
+
+		do {
+			if (this_blocks == BLOCKS_PER_MAP)
+				mask = -1U;
+			else
+				mask = ((1U << this_blocks) - 1) << bit;
+	
+			if (!(f->file_map[idx] & mask))
+				break;
+
+			this_blocks--;
+		} while (this_blocks);
+
+		if (!this_blocks)
+			break;
+
+		f->file_map[idx] |= mask;
+		nr_blocks -= this_blocks;
+		blocks += this_blocks;
+		block += this_blocks;
 	}
 
 	if ((blocks * min_bs) < io_u->buflen)
 		io_u->buflen = blocks * min_bs;
 }
 
+static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
+				     enum fio_ddir ddir)
+{
+	unsigned long long max_blocks;
+	unsigned long long max_size;
+
+	assert(ddir_rw(ddir));
+
+	/*
+	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
+	 */
+	max_size = f->io_size;
+	if (max_size > f->real_file_size)
+		max_size = f->real_file_size;
+
+	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
+	if (!max_blocks)
+		return 0;
+
+	return max_blocks;
+}
+
 /*
  * 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)
+			       enum fio_ddir ddir, unsigned long long *b)
 {
+	unsigned long long min_bs = td->o.rw_min_bs;
 	int i;
 
 	i = f->last_free_lookup;
 	*b = (i * BLOCKS_PER_MAP);
-	while ((*b) * td->rw_min_bs < f->real_file_size) {
-		if (f->file_map[i] != -1UL) {
+	while ((*b) * min_bs < f->real_file_size &&
+		(*b) * min_bs < f->io_size) {
+		if (f->file_map[i] != (unsigned int) -1) {
 			*b += ffz(f->file_map[i]);
+			if (*b > last_block(td, f, ddir))
+				break;
 			f->last_free_lookup = i;
 			return 0;
 		}
@@ -85,76 +145,303 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 		i++;
 	}
 
+	dprint(FD_IO, "failed finding a free block\n");
+	return 1;
+}
+
+static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
+				enum fio_ddir ddir, unsigned long long *b)
+{
+	unsigned long long r;
+	int loops = 5;
+
+	do {
+		r = os_random_long(&td->random_state);
+		dprint(FD_RANDOM, "off rand %llu\n", r);
+		*b = (last_block(td, f, ddir) - 1)
+			* (r / ((unsigned long long) OS_RAND_MAX + 1.0));
+
+		/*
+		 * if we are not maintaining a random map, we are done.
+		 */
+		if (!file_randommap(td, f))
+			return 0;
+
+		/*
+		 * calculate map offset and check if it's free
+		 */
+		if (random_map_free(f, *b))
+			return 0;
+
+		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
+									*b);
+	} while (--loops);
+
+	/*
+	 * we get here, if we didn't suceed in looking up a block. generate
+	 * a random start offset into the filemap, and find the first free
+	 * block from there.
+	 */
+	loops = 10;
+	do {
+		f->last_free_lookup = (f->num_maps - 1) *
+					(r / (OS_RAND_MAX + 1.0));
+		if (!get_next_free_block(td, f, ddir, b))
+			return 0;
+
+		r = os_random_long(&td->random_state);
+	} while (--loops);
+
+	/*
+	 * that didn't work either, try exhaustive search from the start
+	 */
+	f->last_free_lookup = 0;
+	return get_next_free_block(td, f, ddir, b);
+}
+
+static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
+			       enum fio_ddir ddir, unsigned long long *b)
+{
+	if (get_next_rand_offset(td, f, ddir, b)) {
+		dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
+				f->file_name, f->last_pos, f->real_file_size);
+		return 1;
+	}
+
+	return 0;
+}
+
+static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
+			      enum fio_ddir ddir, unsigned long long *b)
+{
+	assert(ddir_rw(ddir));
+
+	if (f->last_pos < f->real_file_size) {
+		*b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
+		return 0;
+	}
+
 	return 1;
 }
 
+static int get_next_block(struct thread_data *td, struct io_u *io_u,
+			  enum fio_ddir ddir, int rw_seq, unsigned long long *b)
+{
+	struct fio_file *f = io_u->file;
+	int ret;
+
+	assert(ddir_rw(ddir));
+
+	if (rw_seq) {
+		if (td_random(td))
+			ret = get_next_rand_block(td, f, ddir, b);
+		else
+			ret = get_next_seq_block(td, f, ddir, b);
+	} else {
+		io_u->flags |= IO_U_F_BUSY_OK;
+
+		if (td->o.rw_seq == RW_SEQ_SEQ) {
+			ret = get_next_seq_block(td, f, ddir, b);
+			if (ret)
+				ret = get_next_rand_block(td, f, ddir, b);
+		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
+			if (f->last_start != -1ULL)
+				*b = (f->last_start - f->file_offset) / td->o.min_bs[ddir];
+			else
+				*b = 0;
+			ret = 0;
+		} else {
+			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
+			ret = 1;
+		}
+	}
+	
+	return ret;
+}
+
 /*
  * 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,
-			   struct io_u *io_u)
+static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
-	const int ddir = io_u->ddir;
-	unsigned long long b, rb;
-	long r;
+	struct fio_file *f = io_u->file;
+	unsigned long long b;
+	enum fio_ddir ddir = io_u->ddir;
+	int rw_seq_hit = 0;
 
-	if (!td->sequential) {
-		unsigned long long max_blocks = f->file_size / td->min_bs[ddir];
-		int loops = 5;
+	assert(ddir_rw(ddir));
 
-		do {
-			r = os_random_long(&td->random_state);
-			b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-			if (td->norandommap)
-				break;
-			rb = b + (f->file_offset / td->min_bs[ddir]);
-			loops--;
-		} while (!random_map_free(td, f, rb) && loops);
+	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
+		rw_seq_hit = 1;
+		td->ddir_seq_nr = td->o.ddir_seq_nr;
+	}
 
-		/*
-		 * if we failed to retrieve a truly random offset within
-		 * the loops assigned, see if there are free ones left at all
-		 */
-		if (!loops && get_next_free_block(td, f, &b))
-			return 1;
-	} else
-		b = f->last_pos / td->min_bs[ddir];
+	if (get_next_block(td, io_u, ddir, rw_seq_hit, &b)) {
+		printf("fail\n");
+		return 1;
+	}
 
-	io_u->offset = (b * td->min_bs[ddir]) + f->file_offset;
-	if (io_u->offset >= f->real_file_size)
+	io_u->offset = b * td->o.ba[ddir];
+	if (io_u->offset >= f->io_size) {
+		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
+					io_u->offset, f->io_size);
 		return 1;
+	}
+
+	io_u->offset += f->file_offset;
+	if (io_u->offset >= f->real_file_size) {
+		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
+					io_u->offset, f->real_file_size);
+		return 1;
+	}
 
 	return 0;
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f,
-				    struct io_u *io_u)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_off)
+		return ops->fill_io_u_off(td, io_u);
+
+	return __get_next_offset(td, io_u);
+}
+
+static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
 	const int ddir = io_u->ddir;
-	unsigned int buflen;
+	unsigned int uninitialized_var(buflen);
+	unsigned int minbs, maxbs;
 	long r;
 
-	if (td->min_bs[ddir] == td->max_bs[ddir])
-		buflen = td->min_bs[ddir];
+	assert(ddir_rw(ddir));
+
+	minbs = td->o.min_bs[ddir];
+	maxbs = td->o.max_bs[ddir];
+
+	if (minbs == maxbs)
+		buflen = minbs;
 	else {
 		r = os_random_long(&td->bsrange_state);
-		buflen = (unsigned int) (1 + (double) (td->max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
-		if (!td->bs_unaligned)
-			buflen = (buflen + td->min_bs[ddir] - 1) & ~(td->min_bs[ddir] - 1);
+		if (!td->o.bssplit_nr[ddir]) {
+			buflen = 1 + (unsigned int) ((double) maxbs *
+					(r / (OS_RAND_MAX + 1.0)));
+			if (buflen < minbs)
+				buflen = minbs;
+		} else {
+			long perc = 0;
+			unsigned int i;
+
+			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
+				struct bssplit *bsp = &td->o.bssplit[ddir][i];
+
+				buflen = bsp->bs;
+				perc += bsp->perc;
+				if (r <= ((OS_RAND_MAX / 100L) * perc))
+					break;
+			}
+		}
+		if (!td->o.bs_unaligned && is_power_of_2(minbs))
+			buflen = (buflen + minbs - 1) & ~(minbs - 1);
 	}
 
-	while (buflen + io_u->offset > f->real_file_size) {
-		if (buflen == td->min_bs[ddir])
-			return 0;
-
-		buflen = td->min_bs[ddir];
+	if (io_u->offset + buflen > io_u->file->real_file_size) {
+		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
+						minbs, ddir);
+		buflen = minbs;
 	}
 
 	return buflen;
 }
 
+static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_size)
+		return ops->fill_io_u_size(td, io_u);
+
+	return __get_next_buflen(td, io_u);
+}
+
+static void set_rwmix_bytes(struct thread_data *td)
+{
+	unsigned int diff;
+
+	/*
+	 * we do time or byte based switch. this is needed because
+	 * buffered writes may issue a lot quicker than they complete,
+	 * whereas reads do not.
+	 */
+	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
+	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
+}
+
+static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
+{
+	unsigned int v;
+	long r;
+
+	r = os_random_long(&td->rwmix_state);
+	v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
+	if (v <= td->o.rwmix[DDIR_READ])
+		return DDIR_READ;
+
+	return DDIR_WRITE;
+}
+
+static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
+{
+	enum fio_ddir odir = ddir ^ 1;
+	struct timeval t;
+	long usec;
+
+	assert(ddir_rw(ddir));
+
+	if (td->rate_pending_usleep[ddir] <= 0)
+		return ddir;
+
+	/*
+	 * We have too much pending sleep in this direction. See if we
+	 * should switch.
+	 */
+	if (td_rw(td)) {
+		/*
+		 * Other direction does not have too much pending, switch
+		 */
+		if (td->rate_pending_usleep[odir] < 100000)
+			return odir;
+
+		/*
+		 * Both directions have pending sleep. Sleep the minimum time
+		 * and deduct from both.
+		 */
+		if (td->rate_pending_usleep[ddir] <=
+			td->rate_pending_usleep[odir]) {
+			usec = td->rate_pending_usleep[ddir];
+		} else {
+			usec = td->rate_pending_usleep[odir];
+			ddir = odir;
+		}
+	} else
+		usec = td->rate_pending_usleep[ddir];
+
+	fio_gettime(&t, NULL);
+	usec_sleep(td, usec);
+	usec = utime_since_now(&t);
+
+	td->rate_pending_usleep[ddir] -= usec;
+
+	odir = ddir ^ 1;
+	if (td_rw(td) && __should_check_rate(td, odir))
+		td->rate_pending_usleep[odir] -= usec;
+
+	return ddir;
+}
+
 /*
  * 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
@@ -162,210 +449,646 @@ static unsigned int get_next_buflen(struct thread_data *td, struct fio_file *f,
  */
 static enum fio_ddir get_rw_ddir(struct thread_data *td)
 {
-	if (td_rw(td)) {
-		struct timeval now;
-		unsigned long elapsed;
+	enum fio_ddir ddir;
 
-		fio_gettime(&now, NULL);
-	 	elapsed = mtime_since_now(&td->rwmix_switch);
+	/*
+	 * see if it's time to fsync
+	 */
+	if (td->o.fsync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC;
 
+	/*
+	 * see if it's time to fdatasync
+	 */
+	if (td->o.fdatasync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_DATASYNC;
+
+	/*
+	 * see if it's time to sync_file_range
+	 */
+	if (td->sync_file_range_nr &&
+	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC_FILE_RANGE;
+
+	if (td_rw(td)) {
 		/*
 		 * Check if it's time to seed a new data direction.
 		 */
-		if (elapsed >= td->rwmixcycle) {
-			unsigned int v;
-			long r;
-
-			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
-				td->rwmix_ddir = DDIR_WRITE;
-			memcpy(&td->rwmix_switch, &now, sizeof(now));
+		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
+			/*
+			 * Put a top limit on how many bytes we do for
+			 * one data direction, to avoid overflowing the
+			 * ranges too much
+			 */
+			ddir = get_rand_ddir(td);
+
+			if (ddir != td->rwmix_ddir)
+				set_rwmix_bytes(td);
+
+			td->rwmix_ddir = ddir;
 		}
-		return td->rwmix_ddir;
+		ddir = td->rwmix_ddir;
 	} else if (td_read(td))
-		return DDIR_READ;
+		ddir = DDIR_READ;
 	else
-		return DDIR_WRITE;
+		ddir = DDIR_WRITE;
+
+	td->rwmix_ddir = rate_ddir(td, ddir);
+	return td->rwmix_ddir;
+}
+
+void put_file_log(struct thread_data *td, struct fio_file *f)
+{
+	int ret = put_file(td, f);
+
+	if (ret)
+		td_verror(td, ret, "file close");
 }
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
+	td_io_u_lock(td);
+
+	io_u->flags |= IO_U_F_FREE;
+	io_u->flags &= ~IO_U_F_FREE_DEF;
+
+	if (io_u->file)
+		put_file_log(td, io_u->file);
+
 	io_u->file = NULL;
-	list_del(&io_u->list);
-	list_add(&io_u->list, &td->io_u_freelist);
-	td->cur_depth--;
+	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
+	flist_del_init(&io_u->list);
+	flist_add(&io_u->list, &td->io_u_freelist);
+	td_io_u_unlock(td);
+	td_io_u_free_notify(td);
+}
+
+void clear_io_u(struct thread_data *td, struct io_u *io_u)
+{
+	io_u->flags &= ~IO_U_F_FLIGHT;
+	put_io_u(td, io_u);
 }
 
-static int fill_io_u(struct thread_data *td, struct fio_file *f,
-		     struct io_u *io_u)
+void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 {
+	struct io_u *__io_u = *io_u;
+
+	dprint(FD_IO, "requeue %p\n", __io_u);
+
+	td_io_u_lock(td);
+
+	__io_u->flags |= IO_U_F_FREE;
+	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(__io_u->ddir))
+		td->io_issues[__io_u->ddir]--;
+
+	__io_u->flags &= ~IO_U_F_FLIGHT;
+	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
+	flist_del(&__io_u->list);
+	flist_add_tail(&__io_u->list, &td->io_u_requeues);
+	td_io_u_unlock(td);
+	*io_u = NULL;
+}
+
+static int fill_io_u(struct thread_data *td, struct io_u *io_u)
+{
+	if (td->io_ops->flags & FIO_NOIO)
+		goto out;
+
+	io_u->ddir = get_rw_ddir(td);
+
 	/*
-	 * If using an iolog, grab next piece if any available.
+	 * fsync() or fdatasync() or trim etc, we are done
 	 */
-	if (td->read_iolog)
-		return read_iolog_get(td, io_u);
+	if (!ddir_rw(io_u->ddir))
+		goto out;
 
 	/*
-	 * see if it's time to sync
+	 * See if it's time to switch to a new zone
 	 */
-	if (td->fsync_blocks && !(td->io_blocks[DDIR_WRITE] % td->fsync_blocks)
-	    && should_fsync(td)) {
-		io_u->ddir = DDIR_SYNC;
-		io_u->file = f;
-		return 0;
+	if (td->zone_bytes >= td->o.zone_size) {
+		td->zone_bytes = 0;
+		io_u->file->last_pos += td->o.zone_skip;
+		td->io_skip_bytes += td->o.zone_skip;
 	}
 
-	io_u->ddir = get_rw_ddir(td);
-
 	/*
 	 * No log, let the seq/rand engine retrieve the next buflen and
 	 * position.
 	 */
-	if (get_next_offset(td, f, io_u))
+	if (get_next_offset(td, io_u)) {
+		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
+		return 1;
+	}
+
+	io_u->buflen = get_next_buflen(td, io_u);
+	if (!io_u->buflen) {
+		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
 		return 1;
+	}
 
-	io_u->buflen = get_next_buflen(td, f, io_u);
-	if (!io_u->buflen)
+	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
+		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
+		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
+				io_u->buflen, io_u->file->real_file_size);
 		return 1;
+	}
 
 	/*
 	 * mark entry before potentially trimming io_u
 	 */
-	if (!td->read_iolog && !td->sequential && !td->norandommap)
-		mark_random_map(td, f, io_u);
+	if (td_random(td) && file_randommap(td, io_u->file))
+		mark_random_map(td, io_u);
 
 	/*
 	 * If using a write iolog, store this entry.
 	 */
-	if (td->write_iolog_file)
-		write_iolog_put(td, io_u);
-
-	io_u->file = f;
+out:
+	dprint_io_u(io_u, "fill_io_u");
+	td->zone_bytes += io_u->buflen;
+	log_io_u(td, io_u);
 	return 0;
 }
 
-static void io_u_mark_depth(struct thread_data *td)
+static void __io_u_mark_map(unsigned int *map, unsigned int nr)
+{
+	int index = 0;
+
+	switch (nr) {
+	default:
+		index = 6;
+		break;
+	case 33 ... 64:
+		index = 5;
+		break;
+	case 17 ... 32:
+		index = 4;
+		break;
+	case 9 ... 16:
+		index = 3;
+		break;
+	case 5 ... 8:
+		index = 2;
+		break;
+	case 1 ... 4:
+		index = 1;
+	case 0:
+		break;
+	}
+
+	map[index]++;
+}
+
+void io_u_mark_submit(struct thread_data *td, unsigned int nr)
+{
+	__io_u_mark_map(td->ts.io_u_submit, nr);
+	td->ts.total_submit++;
+}
+
+void io_u_mark_complete(struct thread_data *td, unsigned int nr)
+{
+	__io_u_mark_map(td->ts.io_u_complete, nr);
+	td->ts.total_complete++;
+}
+
+void io_u_mark_depth(struct thread_data *td, unsigned int nr)
 {
 	int index = 0;
 
 	switch (td->cur_depth) {
 	default:
-		index++;
+		index = 6;
+		break;
 	case 32 ... 63:
-		index++;
+		index = 5;
+		break;
 	case 16 ... 31:
-		index++;
+		index = 4;
+		break;
 	case 8 ... 15:
-		index++;
+		index = 3;
+		break;
 	case 4 ... 7:
-		index++;
+		index = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		index = 1;
 	case 1:
 		break;
 	}
 
-	td->io_u_map[index]++;
-	td->total_io_u++;
+	td->ts.io_u_map[index] += nr;
 }
 
-static void io_u_mark_latency(struct thread_data *td, unsigned long msec)
+static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
+{
+	int index = 0;
+
+	assert(usec < 1000);
+
+	switch (usec) {
+	case 750 ... 999:
+		index = 9;
+		break;
+	case 500 ... 749:
+		index = 8;
+		break;
+	case 250 ... 499:
+		index = 7;
+		break;
+	case 100 ... 249:
+		index = 6;
+		break;
+	case 50 ... 99:
+		index = 5;
+		break;
+	case 20 ... 49:
+		index = 4;
+		break;
+	case 10 ... 19:
+		index = 3;
+		break;
+	case 4 ... 9:
+		index = 2;
+		break;
+	case 2 ... 3:
+		index = 1;
+	case 0 ... 1:
+		break;
+	}
+
+	assert(index < FIO_IO_U_LAT_U_NR);
+	td->ts.io_u_lat_u[index]++;
+}
+
+static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
 {
 	int index = 0;
 
 	switch (msec) {
 	default:
-		index++;
-	case 1024 ... 2047:
-		index++;
-	case 512 ... 1023:
-		index++;
-	case 256 ... 511:
-		index++;
-	case 128 ... 255:
-		index++;
-	case 64 ... 127:
-		index++;
-	case 32 ... 63:
-		index++;
-	case 16 ... 31:
-		index++;
-	case 8 ... 15:
-		index++;
-	case 4 ... 7:
-		index++;
+		index = 11;
+		break;
+	case 1000 ... 1999:
+		index = 10;
+		break;
+	case 750 ... 999:
+		index = 9;
+		break;
+	case 500 ... 749:
+		index = 8;
+		break;
+	case 250 ... 499:
+		index = 7;
+		break;
+	case 100 ... 249:
+		index = 6;
+		break;
+	case 50 ... 99:
+		index = 5;
+		break;
+	case 20 ... 49:
+		index = 4;
+		break;
+	case 10 ... 19:
+		index = 3;
+		break;
+	case 4 ... 9:
+		index = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		index = 1;
 	case 0 ... 1:
 		break;
 	}
 
-	td->io_u_lat[index]++;
+	assert(index < FIO_IO_U_LAT_M_NR);
+	td->ts.io_u_lat_m[index]++;
+}
+
+static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
+{
+	if (usec < 1000)
+		io_u_mark_lat_usec(td, usec);
+	else
+		io_u_mark_lat_msec(td, usec / 1000);
+}
+
+/*
+ * Get next file to service by choosing one at random
+ */
+static struct fio_file *get_next_file_rand(struct thread_data *td,
+					   enum fio_file_flags goodf,
+					   enum fio_file_flags badf)
+{
+	struct fio_file *f;
+	int fno;
+
+	do {
+		long r = os_random_long(&td->next_file_state);
+		int opened = 0;
+
+		fno = (unsigned int) ((double) td->o.nr_files
+			* (r / (OS_RAND_MAX + 1.0)));
+		f = td->files[fno];
+		if (fio_file_done(f))
+			continue;
+
+		if (!fio_file_open(f)) {
+			int err;
+
+			err = td_io_open_file(td, f);
+			if (err)
+				continue;
+			opened = 1;
+		}
+
+		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
+			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
+			return f;
+		}
+		if (opened)
+			td_io_close_file(td, f);
+	} while (1);
+}
+
+/*
+ * Get next file to service by doing round robin between all available ones
+ */
+static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
+					 int badf)
+{
+	unsigned int old_next_file = td->next_file;
+	struct fio_file *f;
+
+	do {
+		int opened = 0;
+
+		f = td->files[td->next_file];
+
+		td->next_file++;
+		if (td->next_file >= td->o.nr_files)
+			td->next_file = 0;
+
+		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
+		if (fio_file_done(f)) {
+			f = NULL;
+			continue;
+		}
+
+		if (!fio_file_open(f)) {
+			int err;
+
+			err = td_io_open_file(td, f);
+			if (err) {
+				dprint(FD_FILE, "error %d on open of %s\n",
+					err, f->file_name);
+				f = NULL;
+				continue;
+			}
+			opened = 1;
+		}
+
+		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
+								f->flags);
+		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
+			break;
+
+		if (opened)
+			td_io_close_file(td, f);
+
+		f = NULL;
+	} while (td->next_file != old_next_file);
+
+	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
+	return f;
+}
+
+static struct fio_file *__get_next_file(struct thread_data *td)
+{
+	struct fio_file *f;
+
+	assert(td->o.nr_files <= td->files_index);
+
+	if (td->nr_done_files >= td->o.nr_files) {
+		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
+				" nr_files=%d\n", td->nr_open_files,
+						  td->nr_done_files,
+						  td->o.nr_files);
+		return NULL;
+	}
+
+	f = td->file_service_file;
+	if (f && fio_file_open(f) && !fio_file_closing(f)) {
+		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
+			goto out;
+		if (td->file_service_left--)
+			goto out;
+	}
+
+	if (td->o.file_service_type == FIO_FSERVICE_RR ||
+	    td->o.file_service_type == FIO_FSERVICE_SEQ)
+		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
+	else
+		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
+
+	td->file_service_file = f;
+	td->file_service_left = td->file_service_nr - 1;
+out:
+	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
+	return f;
 }
 
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->get_next_file)
+		return ops->get_next_file(td);
+
+	return __get_next_file(td);
+}
+
+static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
+{
+	struct fio_file *f;
+
+	do {
+		f = get_next_file(td);
+		if (!f)
+			return 1;
+
+		io_u->file = f;
+		get_file(f);
+
+		if (!fill_io_u(td, io_u))
+			break;
+
+		put_file_log(td, f);
+		td_io_close_file(td, f);
+		io_u->file = NULL;
+		fio_file_set_done(f);
+		td->nr_done_files++;
+		dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
+					td->nr_done_files, td->o.nr_files);
+	} while (1);
+
+	return 0;
+}
+
+
 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);
+	td_io_u_lock(td);
+
+again:
+	if (!flist_empty(&td->io_u_requeues))
+		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
+	else if (!queue_full(td)) {
+		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
 
 		io_u->buflen = 0;
-		io_u->error = 0;
 		io_u->resid = 0;
-		list_del(&io_u->list);
-		list_add(&io_u->list, &td->io_u_busylist);
+		io_u->file = NULL;
+		io_u->end_io = NULL;
+	}
+
+	if (io_u) {
+		assert(io_u->flags & IO_U_F_FREE);
+		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
+
+		io_u->error = 0;
+		flist_del(&io_u->list);
+		flist_add(&io_u->list, &td->io_u_busylist);
 		td->cur_depth++;
-		io_u_mark_depth(td);
+		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+	} else if (td->o.verify_async) {
+		/*
+		 * We ran out, wait for async verify threads to finish and
+		 * return one
+		 */
+		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
+		goto again;
 	}
 
+	td_io_u_unlock(td);
 	return io_u;
 }
 
+static int check_get_trim(struct thread_data *td, struct io_u *io_u)
+{
+	if (td->o.trim_backlog && td->trim_entries) {
+		int get_trim = 0;
+
+		if (td->trim_batch) {
+			td->trim_batch--;
+			get_trim = 1;
+		} else if (!(td->io_hist_len % td->o.trim_backlog) &&
+			 td->last_ddir != DDIR_READ) {
+			td->trim_batch = td->o.trim_batch;
+			if (!td->trim_batch)
+				td->trim_batch = td->o.trim_backlog;
+			get_trim = 1;
+		}
+
+		if (get_trim && !get_next_trim(td, io_u))
+			return 1;
+	}
+
+	return 0;
+}
+
+static int check_get_verify(struct thread_data *td, struct io_u *io_u)
+{
+	if (td->o.verify_backlog && td->io_hist_len) {
+		int get_verify = 0;
+
+		if (td->verify_batch) {
+			td->verify_batch--;
+			get_verify = 1;
+		} else if (!(td->io_hist_len % td->o.verify_backlog) &&
+			 td->last_ddir != DDIR_READ) {
+			td->verify_batch = td->o.verify_batch;
+			if (!td->verify_batch)
+				td->verify_batch = td->o.verify_backlog;
+			get_verify = 1;
+		}
+
+		if (get_verify && !get_next_verify(td, io_u))
+			return 1;
+	}
+
+	return 0;
+}
+
 /*
  * 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.
  */
-struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
+struct io_u *get_io_u(struct thread_data *td)
 {
+	struct fio_file *f;
 	struct io_u *io_u;
 
 	io_u = __get_io_u(td);
-	if (!io_u)
+	if (!io_u) {
+		dprint(FD_IO, "__get_io_u failed\n");
 		return NULL;
-
-	if (td->zone_bytes >= td->zone_size) {
-		td->zone_bytes = 0;
-		f->last_pos += td->zone_skip;
 	}
 
-	if (fill_io_u(td, f, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
-	}
+	if (check_get_verify(td, io_u))
+		goto out;
+	if (check_get_trim(td, io_u))
+		goto out;
 
-	if (io_u->buflen + io_u->offset > f->real_file_size) {
-		if (td->io_ops->flags & FIO_RAWIO) {
-			put_io_u(td, io_u);
-			return NULL;
-		}
+	/*
+	 * from a requeue, io_u already setup
+	 */
+	if (io_u->file)
+		goto out;
 
-		io_u->buflen = f->real_file_size - io_u->offset;
+	/*
+	 * If using an iolog, grab next piece if any available.
+	 */
+	if (td->o.read_iolog_file) {
+		if (read_iolog_get(td, io_u))
+			goto err_put;
+	} else if (set_io_u_file(td, io_u)) {
+		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
+		goto err_put;
 	}
 
-	if (io_u->ddir != DDIR_SYNC) {
-		if (!io_u->buflen) {
-			put_io_u(td, io_u);
-			return NULL;
+	f = io_u->file;
+	assert(fio_file_open(f));
+
+	if (ddir_rw(io_u->ddir)) {
+		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
+			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
+			goto err_put;
 		}
 
+		f->last_start = io_u->offset;
 		f->last_pos = io_u->offset + io_u->buflen;
 
-		if (td->verify != VERIFY_NONE)
+		if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
 			populate_verify_io_u(td, io_u);
+		else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
+			io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
+		else if (io_u->ddir == DDIR_READ) {
+			/*
+			 * Reset the buf_filled parameters so next time if the
+			 * buffer is used for writes it is refilled.
+			 */
+			io_u->buf_filled_len = 0;
+		}
 	}
 
 	/*
@@ -374,65 +1097,159 @@ struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
 	io_u->xfer_buf = io_u->buf;
 	io_u->xfer_buflen = io_u->buflen;
 
-	if (td_io_prep(td, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
+out:
+	assert(io_u->file);
+	if (!td_io_prep(td, io_u)) {
+		if (!td->o.disable_slat)
+			fio_gettime(&io_u->start_time, NULL);
+		return io_u;
 	}
+err_put:
+	dprint(FD_IO, "get_io_u failed\n");
+	put_io_u(td, io_u);
+	return NULL;
+}
 
-	fio_gettime(&io_u->start_time, NULL);
-	return io_u;
+void io_u_log_error(struct thread_data *td, struct io_u *io_u)
+{
+	const char *msg[] = { "read", "write", "sync", "datasync",
+				"sync_file_range", "wait", "trim" };
+
+
+
+	log_err("fio: io_u error");
+
+	if (io_u->file)
+		log_err(" on file %s", io_u->file->file_name);
+
+	log_err(": %s\n", strerror(io_u->error));
+
+	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
+					io_u->offset, io_u->xfer_buflen);
+
+	if (!td->error)
+		td_verror(td, io_u->error, "io_u error");
 }
 
 static void io_completed(struct thread_data *td, struct io_u *io_u,
 			 struct io_completion_data *icd)
 {
-	unsigned long msec;
+	/*
+	 * Older gcc's are too dumb to realize that usec is always used
+	 * initialized, silence that warning.
+	 */
+	unsigned long uninitialized_var(usec);
+	struct fio_file *f;
+
+	dprint_io_u(io_u, "io complete");
+
+	td_io_u_lock(td);
+	assert(io_u->flags & IO_U_F_FLIGHT);
+	io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
+	td_io_u_unlock(td);
 
-	if (io_u->ddir == DDIR_SYNC) {
+	if (ddir_sync(io_u->ddir)) {
 		td->last_was_sync = 1;
+		f = io_u->file;
+		if (f) {
+			f->first_write = -1ULL;
+			f->last_write = -1ULL;
+		}
 		return;
 	}
 
 	td->last_was_sync = 0;
+	td->last_ddir = io_u->ddir;
 
-	if (!io_u->error) {
+	if (!io_u->error && ddir_rw(io_u->ddir)) {
 		unsigned int bytes = io_u->buflen - io_u->resid;
 		const enum fio_ddir idx = io_u->ddir;
+		const enum fio_ddir odx = io_u->ddir ^ 1;
 		int ret;
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
-		td->zone_bytes += bytes;
 		td->this_io_bytes[idx] += bytes;
 
-		io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
-
-		msec = mtime_since(&io_u->issue_time, &icd->time);
+		if (idx == DDIR_WRITE) {
+			f = io_u->file;
+			if (f) {
+				if (f->first_write == -1ULL ||
+				    io_u->offset < f->first_write)
+					f->first_write = io_u->offset;
+				if (f->last_write == -1ULL ||
+				    ((io_u->offset + bytes) > f->last_write))
+					f->last_write = io_u->offset + bytes;
+			}
+		}
 
-		add_clat_sample(td, idx, msec);
-		add_bw_sample(td, idx, &icd->time);
-		io_u_mark_latency(td, msec);
+		if (ramp_time_over(td)) {
+			unsigned long uninitialized_var(lusec);
+
+			if (!td->o.disable_clat || !td->o.disable_bw)
+				lusec = utime_since(&io_u->issue_time,
+							&icd->time);
+			if (!td->o.disable_lat) {
+				unsigned long tusec;
+
+				tusec = utime_since(&io_u->start_time,
+							&icd->time);
+				add_lat_sample(td, idx, tusec, bytes);
+			}
+			if (!td->o.disable_clat) {
+				add_clat_sample(td, idx, lusec, bytes);
+				io_u_mark_latency(td, lusec);
+			}
+			if (!td->o.disable_bw)
+				add_bw_sample(td, idx, bytes, &icd->time);
+			if (__should_check_rate(td, idx)) {
+				td->rate_pending_usleep[idx] =
+					((td->this_io_bytes[idx] *
+					  td->rate_nsec_cycle[idx]) / 1000 -
+					 utime_since_now(&td->start));
+			}
+			if (__should_check_rate(td, idx ^ 1))
+				td->rate_pending_usleep[odx] =
+					((td->this_io_bytes[odx] *
+					  td->rate_nsec_cycle[odx]) / 1000 -
+					 utime_since_now(&td->start));
+		}
 
-		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
+		if (td_write(td) && idx == DDIR_WRITE &&
+		    td->o.do_verify &&
+		    td->o.verify != VERIFY_NONE)
 			log_io_piece(td, io_u);
 
 		icd->bytes_done[idx] += bytes;
 
-		if (icd->handler) {
-			ret = icd->handler(io_u);
+		if (io_u->end_io) {
+			ret = io_u->end_io(td, io_u);
 			if (ret && !icd->error)
 				icd->error = ret;
 		}
-	} else
+	} else if (io_u->error) {
 		icd->error = io_u->error;
+		io_u_log_error(td, io_u);
+	}
+	if (td->o.continue_on_error && icd->error &&
+	    td_non_fatal_error(icd->error)) {
+		/*
+		 * If there is a non_fatal error, then add to the error count
+		 * and clear all the errors.
+		 */
+		update_error_count(td, icd->error);
+		td_clear_error(td);
+		icd->error = 0;
+		io_u->error = 0;
+	}
 }
 
-static void init_icd(struct io_completion_data *icd, endio_handler *handler,
+static void init_icd(struct thread_data *td, struct io_completion_data *icd,
 		     int nr)
 {
-	fio_gettime(&icd->time, NULL);
+	if (!td->o.disable_clat || !td->o.disable_bw)
+		fio_gettime(&icd->time, NULL);
 
-	icd->handler = handler;
 	icd->nr = nr;
 
 	icd->error = 0;
@@ -449,50 +1266,98 @@ static void ios_completed(struct thread_data *td,
 		io_u = td->io_ops->event(td, i);
 
 		io_completed(td, io_u, icd);
-		put_io_u(td, io_u);
+
+		if (!(io_u->flags & IO_U_F_FREE_DEF))
+			put_io_u(td, io_u);
 	}
 }
 
-long io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
-			endio_handler *handler)
+/*
+ * Complete a single io_u for the sync engines.
+ */
+int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
+		       unsigned long *bytes)
 {
 	struct io_completion_data icd;
 
-	init_icd(&icd, handler, 1);
+	init_icd(td, &icd, 1);
 	io_completed(td, io_u, &icd);
-	put_io_u(td, io_u);
 
-	if (!icd.error)
-		return icd.bytes_done[0] + icd.bytes_done[1];
+	if (!(io_u->flags & IO_U_F_FREE_DEF))
+		put_io_u(td, io_u);
 
-	td_verror(td, icd.error);
-	return -1;
-}
+	if (icd.error) {
+		td_verror(td, icd.error, "io_u_sync_complete");
+		return -1;
+	}
 
-long io_u_queued_complete(struct thread_data *td, int min_events,
-			  endio_handler *handler)
+	if (bytes) {
+		bytes[0] += icd.bytes_done[0];
+		bytes[1] += icd.bytes_done[1];
+	}
 
+	return 0;
+}
+
+/*
+ * Called to complete min_events number of io for the async engines.
+ */
+int io_u_queued_complete(struct thread_data *td, int min_evts,
+			 unsigned long *bytes)
 {
-	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
-	struct timespec *tsp = NULL;
 	struct io_completion_data icd;
+	struct timespec *tvp = NULL;
 	int ret;
+	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-	if (min_events > 0)
-		tsp = &ts;
+	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
 
-	ret = td_io_getevents(td, min_events, td->cur_depth, tsp);
+	if (!min_evts)
+		tvp = &ts;
+
+	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
 	if (ret < 0) {
-		td_verror(td, -ret);
+		td_verror(td, -ret, "td_io_getevents");
 		return ret;
 	} else if (!ret)
 		return ret;
 
-	init_icd(&icd, handler, ret);
+	init_icd(td, &icd, ret);
 	ios_completed(td, &icd);
-	if (!icd.error)
-		return icd.bytes_done[0] + icd.bytes_done[1];
+	if (icd.error) {
+		td_verror(td, icd.error, "io_u_queued_complete");
+		return -1;
+	}
+
+	if (bytes) {
+		bytes[0] += icd.bytes_done[0];
+		bytes[1] += icd.bytes_done[1];
+	}
+
+	return 0;
+}
+
+/*
+ * Call when io_u is really queued, to update the submission latency.
+ */
+void io_u_queued(struct thread_data *td, struct io_u *io_u)
+{
+	if (!td->o.disable_slat) {
+		unsigned long slat_time;
+
+		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
+		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
+	}
+}
 
-	td_verror(td, icd.error);
-	return -1;
+/*
+ * "randomly" fill the buffer contents
+ */
+void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
+		      unsigned int max_bs)
+{
+	if (!td->o.zero_buffers)
+		fill_random_buf(io_u->buf, max_bs);
+	else
+		memset(io_u->buf, 0, max_bs);
 }