X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=e4671e42ac741557006ce7056dd863508b60280c;hp=17e188ad022ee05f3bb6dceedb153a24027b2fb3;hb=709c8313cf171e63c92740c6b2f38041cb11c722;hpb=8400d9b26a4f24f3b56cc9dad02be65791b58a83

diff --git a/io_u.c b/io_u.c
index 17e188ad..7cbdb915 100644
--- a/io_u.c
+++ b/io_u.c
@@ -6,32 +6,28 @@
 #include <assert.h>
 
 #include "fio.h"
-#include "os.h"
-
-/*
- * Change this define to play with the timeout handling
- */
-#undef FIO_USE_TIMEOUT
+#include "hash.h"
+#include "verify.h"
+#include "trim.h"
+#include "lib/rand.h"
+#include "lib/axmap.h"
+#include "err.h"
 
 struct io_completion_data {
 	int nr;				/* input */
 
 	int error;			/* output */
-	unsigned long bytes_done[2];	/* output */
+	uint64_t bytes_done[DDIR_RWDIR_CNT];	/* output */
 	struct timeval time;		/* output */
 };
 
 /*
- * The ->file_map[] contains a map of blocks we have or have not done io
+ * The ->io_axmap 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 uint64_t block)
 {
-	unsigned int idx = RAND_MAP_IDX(td, f, block);
-	unsigned int bit = RAND_MAP_BIT(td, f, block);
-
-	return (f->file_map[idx] & (1UL << bit)) == 0;
+	return !axmap_isset(f->io_axmap, block);
 }
 
 /*
@@ -41,89 +37,324 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 {
 	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;
+	uint64_t block;
 
-	block = io_u->offset / (unsigned long long) min_bs;
-	blocks = 0;
+	block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
 	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
 
-	while (blocks < nr_blocks) {
-		unsigned int idx, bit;
+	if (!(io_u->flags & IO_U_F_BUSY_OK))
+		nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
 
-		/*
-		 * If we have a mixed random workload, we may
-		 * encounter blocks we already did IO to.
-		 */
-		if (!td->o.ddir_nr && !random_map_free(td, f, block))
-			break;
+	if ((nr_blocks * min_bs) < io_u->buflen)
+		io_u->buflen = nr_blocks * min_bs;
+}
+
+static uint64_t last_block(struct thread_data *td, struct fio_file *f,
+			   enum fio_ddir ddir)
+{
+	uint64_t max_blocks;
+	uint64_t 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;
+
+	if (td->o.zone_range)
+		max_size = td->o.zone_range;
 
-		idx = RAND_MAP_IDX(td, f, block);
-		bit = RAND_MAP_BIT(td, f, block);
+	max_blocks = max_size / (uint64_t) td->o.ba[ddir];
+	if (!max_blocks)
+		return 0;
+
+	return max_blocks;
+}
 
-		fio_assert(td, idx < f->num_maps);
+struct rand_off {
+	struct flist_head list;
+	uint64_t off;
+};
+
+static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
+				  enum fio_ddir ddir, uint64_t *b)
+{
+	uint64_t r, lastb;
+
+	lastb = last_block(td, f, ddir);
+	if (!lastb)
+		return 1;
+
+	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
+		uint64_t rmax;
+
+		rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
+
+		if (td->o.use_os_rand) {
+			rmax = OS_RAND_MAX;
+			r = os_random_long(&td->random_state);
+		} else {
+			rmax = FRAND_MAX;
+			r = __rand(&td->__random_state);
+		}
 
-		f->file_map[idx] |= (1UL << bit);
-		block++;
-		blocks++;
+		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
+
+		*b = lastb * (r / ((uint64_t) rmax + 1.0));
+	} else {
+		uint64_t off = 0;
+
+		if (lfsr_next(&f->lfsr, &off, lastb))
+			return 1;
+
+		*b = off;
 	}
 
-	if ((blocks * min_bs) < io_u->buflen)
-		io_u->buflen = blocks * min_bs;
+	/*
+	 * if we are not maintaining a random map, we are done.
+	 */
+	if (!file_randommap(td, f))
+		goto ret;
+
+	/*
+	 * calculate map offset and check if it's free
+	 */
+	if (random_map_free(f, *b))
+		goto ret;
+
+	dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
+						(unsigned long long) *b);
+
+	*b = axmap_next_free(f->io_axmap, *b);
+	if (*b == (uint64_t) -1ULL)
+		return 1;
+ret:
+	return 0;
+}
+
+static int __get_next_rand_offset_zipf(struct thread_data *td,
+				       struct fio_file *f, enum fio_ddir ddir,
+				       uint64_t *b)
+{
+	*b = zipf_next(&f->zipf);
+	return 0;
+}
+
+static int __get_next_rand_offset_pareto(struct thread_data *td,
+					 struct fio_file *f, enum fio_ddir ddir,
+					 uint64_t *b)
+{
+	*b = pareto_next(&f->zipf);
+	return 0;
+}
+
+static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
+{
+	struct rand_off *r1 = flist_entry(a, struct rand_off, list);
+	struct rand_off *r2 = flist_entry(b, struct rand_off, list);
+
+	return r1->off - r2->off;
+}
+
+static int get_off_from_method(struct thread_data *td, struct fio_file *f,
+			       enum fio_ddir ddir, uint64_t *b)
+{
+	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
+		return __get_next_rand_offset(td, f, ddir, b);
+	else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
+		return __get_next_rand_offset_zipf(td, f, ddir, b);
+	else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
+		return __get_next_rand_offset_pareto(td, f, ddir, b);
+
+	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
+	return 1;
 }
 
 /*
- * Return the next free block in the map.
+ * Sort the reads for a verify phase in batches of verifysort_nr, if
+ * specified.
  */
-static int get_next_free_block(struct thread_data *td, struct fio_file *f,
-			       unsigned long long *b)
+static inline int should_sort_io(struct thread_data *td)
 {
-	int i;
+	if (!td->o.verifysort_nr || !td->o.do_verify)
+		return 0;
+	if (!td_random(td))
+		return 0;
+	if (td->runstate != TD_VERIFYING)
+		return 0;
+	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
+		return 0;
 
-	i = f->last_free_lookup;
-	*b = (i * BLOCKS_PER_MAP);
-	while ((*b) * td->o.rw_min_bs < f->real_file_size) {
-		if (f->file_map[i] != -1UL) {
-			*b += ffz(f->file_map[i]);
-			f->last_free_lookup = i;
-			return 0;
-		}
+	return 1;
+}
+
+static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
+{
+	unsigned int v;
+	unsigned long r;
 
-		*b += BLOCKS_PER_MAP;
-		i++;
+	if (td->o.perc_rand[ddir] == 100)
+		return 1;
+
+	if (td->o.use_os_rand) {
+		r = os_random_long(&td->seq_rand_state[ddir]);
+		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
+	} else {
+		r = __rand(&td->__seq_rand_state[ddir]);
+		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
 	}
 
-	return 1;
+	return v <= td->o.perc_rand[ddir];
 }
 
 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
-				int ddir, unsigned long long *b)
+				enum fio_ddir ddir, uint64_t *b)
 {
-	unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
-	unsigned long long r, rb;
-	int loops = 5;
+	struct rand_off *r;
+	int i, ret = 1;
+
+	if (!should_sort_io(td))
+		return get_off_from_method(td, f, ddir, b);
+
+	if (!flist_empty(&td->next_rand_list)) {
+		struct rand_off *r;
+fetch:
+		r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
+		flist_del(&r->list);
+		*b = r->off;
+		free(r);
+		return 0;
+	}
 
-	do {
-		r = os_random_long(&td->random_state);
-		if (!max_blocks)
-			*b = 0;
-		else
-			*b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-		if (td->o.norandommap)
+	for (i = 0; i < td->o.verifysort_nr; i++) {
+		r = malloc(sizeof(*r));
+
+		ret = get_off_from_method(td, f, ddir, &r->off);
+		if (ret) {
+			free(r);
 			break;
-		rb = *b + (f->file_offset / td->o.min_bs[ddir]);
-		loops--;
-	} while (!random_map_free(td, f, rb) && loops);
+		}
 
-	/*
-	 * 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;
+		flist_add(&r->list, &td->next_rand_list);
+	}
 
-	return 0;
+	if (ret && !i)
+		return ret;
+
+	assert(!flist_empty(&td->next_rand_list));
+	flist_sort(NULL, &td->next_rand_list, flist_cmp);
+	goto fetch;
+}
+
+static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
+			       enum fio_ddir ddir, uint64_t *b)
+{
+	if (!get_next_rand_offset(td, f, ddir, b))
+		return 0;
+
+	if (td->o.time_based) {
+		fio_file_reset(td, f);
+		if (!get_next_rand_offset(td, f, ddir, b))
+			return 0;
+	}
+
+	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
+			f->file_name, (unsigned long long) f->last_pos,
+			(unsigned long long) f->real_file_size);
+	return 1;
+}
+
+static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
+			       enum fio_ddir ddir, uint64_t *offset)
+{
+	assert(ddir_rw(ddir));
+
+	if (f->last_pos >= f->io_size + get_start_offset(td, f) && td->o.time_based)
+		f->last_pos = f->last_pos - f->io_size;
+
+	if (f->last_pos < f->real_file_size) {
+		uint64_t pos;
+
+		if (f->last_pos == f->file_offset && td->o.ddir_seq_add < 0)
+			f->last_pos = f->real_file_size;
+
+		pos = f->last_pos - f->file_offset;
+		if (pos)
+			pos += td->o.ddir_seq_add;
+
+		*offset = pos;
+		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 int *is_random)
+{
+	struct fio_file *f = io_u->file;
+	uint64_t b, offset;
+	int ret;
+
+	assert(ddir_rw(ddir));
+
+	b = offset = -1ULL;
+
+	if (rw_seq) {
+		if (td_random(td)) {
+			if (should_do_random(td, ddir)) {
+				ret = get_next_rand_block(td, f, ddir, &b);
+				*is_random = 1;
+			} else {
+				*is_random = 0;
+				io_u->flags |= IO_U_F_BUSY_OK;
+				ret = get_next_seq_offset(td, f, ddir, &offset);
+				if (ret)
+					ret = get_next_rand_block(td, f, ddir, &b);
+			}
+		} else {
+			*is_random = 0;
+			ret = get_next_seq_offset(td, f, ddir, &offset);
+		}
+	} else {
+		io_u->flags |= IO_U_F_BUSY_OK;
+		*is_random = 0;
+
+		if (td->o.rw_seq == RW_SEQ_SEQ) {
+			ret = get_next_seq_offset(td, f, ddir, &offset);
+			if (ret) {
+				ret = get_next_rand_block(td, f, ddir, &b);
+				*is_random = 0;
+			}
+		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
+			if (f->last_start != -1ULL)
+				offset = f->last_start - f->file_offset;
+			else
+				offset = 0;
+			ret = 0;
+		} else {
+			log_err("fio: unknown rw_seq=%d\n", td->o.rw_seq);
+			ret = 1;
+		}
+	}
+
+	if (!ret) {
+		if (offset != -1ULL)
+			io_u->offset = offset;
+		else if (b != -1ULL)
+			io_u->offset = b * td->o.ba[ddir];
+		else {
+			log_err("fio: bug in offset generation: offset=%llu, b=%llu\n", (unsigned long long) offset, (unsigned long long) b);
+			ret = 1;
+		}
+	}
+
+	return ret;
 }
 
 /*
@@ -131,52 +362,145 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
  * 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 io_u *io_u)
+static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
+			     unsigned int *is_random)
 {
 	struct fio_file *f = io_u->file;
-	const int ddir = io_u->ddir;
-	unsigned long long b;
+	enum fio_ddir ddir = io_u->ddir;
+	int rw_seq_hit = 0;
 
-	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
-		td->ddir_nr = td->o.ddir_nr;
+	assert(ddir_rw(ddir));
 
-		if (get_next_rand_offset(td, f, ddir, &b))
-			return 1;
-	} else {
-		if (f->last_pos >= f->real_file_size)
-			return 1;
+	if (td->o.ddir_seq_nr && !--td->ddir_seq_nr) {
+		rw_seq_hit = 1;
+		td->ddir_seq_nr = td->o.ddir_seq_nr;
+	}
 
-		b = f->last_pos / td->o.min_bs[ddir];
+	if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
+		return 1;
+
+	if (io_u->offset >= f->io_size) {
+		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
+					(unsigned long long) io_u->offset,
+					(unsigned long long) f->io_size);
+		return 1;
 	}
 
-	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
-	if (io_u->offset >= f->real_file_size)
+	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",
+					(unsigned long long) io_u->offset,
+					(unsigned long long) f->real_file_size);
 		return 1;
+	}
 
 	return 0;
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u,
+			   unsigned int *is_random)
 {
-	const int ddir = io_u->ddir;
-	unsigned int buflen;
-	long r;
+	if (td->flags & TD_F_PROFILE_OPS) {
+		struct prof_io_ops *ops = &td->prof_io_ops;
 
-	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
-		buflen = td->o.min_bs[ddir];
-	else {
-		r = os_random_long(&td->bsrange_state);
-		buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
-		if (!td->o.bs_unaligned)
-			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
+		if (ops->fill_io_u_off)
+			return ops->fill_io_u_off(td, io_u, is_random);
 	}
 
+	return __get_next_offset(td, io_u, is_random);
+}
+
+static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
+			    unsigned int buflen)
+{
+	struct fio_file *f = io_u->file;
+
+	return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
+}
+
+static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
+				      unsigned int is_random)
+{
+	int ddir = io_u->ddir;
+	unsigned int buflen = 0;
+	unsigned int minbs, maxbs;
+	unsigned long r, rand_max;
+
+	assert(ddir_rw(ddir));
+
+	if (td->o.bs_is_seq_rand)
+		ddir = is_random ? DDIR_WRITE: DDIR_READ;
+
+	minbs = td->o.min_bs[ddir];
+	maxbs = td->o.max_bs[ddir];
+
+	if (minbs == maxbs)
+		return minbs;
+
+	/*
+	 * If we can't satisfy the min block size from here, then fail
+	 */
+	if (!io_u_fits(td, io_u, minbs))
+		return 0;
+
+	if (td->o.use_os_rand)
+		rand_max = OS_RAND_MAX;
+	else
+		rand_max = FRAND_MAX;
+
+	do {
+		if (td->o.use_os_rand)
+			r = os_random_long(&td->bsrange_state);
+		else
+			r = __rand(&td->__bsrange_state);
+
+		if (!td->o.bssplit_nr[ddir]) {
+			buflen = 1 + (unsigned int) ((double) maxbs *
+					(r / (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 <= ((rand_max / 100L) * perc)) &&
+				    io_u_fits(td, io_u, buflen))
+					break;
+			}
+		}
+
+		if (td->o.do_verify && td->o.verify != VERIFY_NONE)
+			buflen = (buflen + td->o.verify_interval - 1) &
+				~(td->o.verify_interval - 1);
+
+		if (!td->o.bs_unaligned && is_power_of_2(minbs))
+			buflen = (buflen + minbs - 1) & ~(minbs - 1);
+
+	} while (!io_u_fits(td, io_u, buflen));
+
 	return buflen;
 }
 
+static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
+				    unsigned int is_random)
+{
+	if (td->flags & TD_F_PROFILE_OPS) {
+		struct prof_io_ops *ops = &td->prof_io_ops;
+
+		if (ops->fill_io_u_size)
+			return ops->fill_io_u_size(td, io_u, is_random);
+	}
+
+	return __get_next_buflen(td, io_u, is_random);
+}
+
 static void set_rwmix_bytes(struct thread_data *td)
 {
-	unsigned long long rbytes;
 	unsigned int diff;
 
 	/*
@@ -184,25 +508,101 @@ static void set_rwmix_bytes(struct thread_data *td)
 	 * buffered writes may issue a lot quicker than they complete,
 	 * whereas reads do not.
 	 */
-	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
 	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
-
-	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
+	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;
+	unsigned long r;
 
-	r = os_random_long(&td->rwmix_state);
-	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
-	if (v < td->o.rwmix[DDIR_READ])
+	if (td->o.use_os_rand) {
+		r = os_random_long(&td->rwmix_state);
+		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
+	} else {
+		r = __rand(&td->__rwmix_state);
+		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+	}
+
+	if (v <= td->o.rwmix[DDIR_READ])
 		return DDIR_READ;
 
 	return DDIR_WRITE;
 }
 
+void io_u_quiesce(struct thread_data *td)
+{
+	/*
+	 * We are going to sleep, ensure that we flush anything pending as
+	 * not to skew our latency numbers.
+	 *
+	 * Changed to only monitor 'in flight' requests here instead of the
+	 * td->cur_depth, b/c td->cur_depth does not accurately represent
+	 * io's that have been actually submitted to an async engine,
+	 * and cur_depth is meaningless for sync engines.
+	 */
+	while (td->io_u_in_flight) {
+		int fio_unused ret;
+
+		ret = io_u_queued_complete(td, 1, NULL);
+	}
+}
+
+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) && td->o.rwmix[odir]) {
+		/*
+		 * 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];
+
+	io_u_quiesce(td);
+
+	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;
+
+	if (ddir_trim(ddir))
+		return ddir;
+
+	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
@@ -210,305 +610,823 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
  */
 static enum fio_ddir get_rw_ddir(struct thread_data *td)
 {
-	if (td_rw(td)) {
-		struct timeval now;
-		unsigned long elapsed;
-		unsigned int cycle;
+	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;
 
-		/*
-		 * if this is the first cycle, make it shorter
-		 */
-		cycle = td->o.rwmixcycle;
-		if (!td->rwmix_bytes)
-			cycle /= 10;
+	/*
+	 * 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 >= cycle ||
-		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
-			unsigned long long max_bytes;
-			enum fio_ddir ddir;			
-
+		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);
-			max_bytes = td->this_io_bytes[ddir];
-			if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
-				if (!td->rw_end_set[ddir]) {
-					td->rw_end_set[ddir] = 1;
-					memcpy(&td->rw_end[ddir], &now, sizeof(now));
-				}
-				ddir ^= 1;
-			}
 
 			if (ddir != td->rwmix_ddir)
 				set_rwmix_bytes(td);
 
 			td->rwmix_ddir = ddir;
-			memcpy(&td->rwmix_switch, &now, sizeof(now));
 		}
-		return td->rwmix_ddir;
+		ddir = td->rwmix_ddir;
 	} else if (td_read(td))
-		return DDIR_READ;
+		ddir = DDIR_READ;
+	else if (td_write(td))
+		ddir = DDIR_WRITE;
 	else
-		return DDIR_WRITE;
+		ddir = DDIR_TRIM;
+
+	td->rwmix_ddir = rate_ddir(td, ddir);
+	return td->rwmix_ddir;
+}
+
+static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
+{
+	io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
+
+	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
+	    td->o.barrier_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
+	     td->io_issues[DDIR_WRITE])
+		io_u->flags |= IO_U_F_BARRIER;
+}
+
+void put_file_log(struct thread_data *td, struct fio_file *f)
+{
+	unsigned 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)
 {
-	assert((io_u->flags & IO_U_F_FREE) == 0);
-	io_u->flags |= IO_U_F_FREE;
+	td_io_u_lock(td);
+
+	if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
+		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--;
+	io_u->flags |= IO_U_F_FREE;
+
+	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
+	io_u_qpush(&td->io_u_freelist, io_u);
+	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);
 }
 
 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 {
 	struct io_u *__io_u = *io_u;
+	enum fio_ddir ddir = acct_ddir(__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(ddir))
+		td->io_issues[ddir]--;
+
 	__io_u->flags &= ~IO_U_F_FLIGHT;
+	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
 
-	list_del(&__io_u->list);
-	list_add_tail(&__io_u->list, &td->io_u_requeues);
-	td->cur_depth--;
+	io_u_rpush(&td->io_u_requeues, __io_u);
+	td_io_u_unlock(td);
 	*io_u = NULL;
 }
 
 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 {
+	unsigned int is_random;
+
+	if (td->io_ops->flags & FIO_NOIO)
+		goto out;
+
+	set_rw_ddir(td, io_u);
+
 	/*
-	 * If using an iolog, grab next piece if any available.
+	 * fsync() or fdatasync() or trim etc, we are done
 	 */
-	if (td->o.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->o.fsync_blocks &&
-	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
-	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
-		io_u->ddir = DDIR_SYNC;
-		goto out;
+	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
+		td->zone_bytes = 0;
+		io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
+		io_u->file->last_pos = io_u->file->file_offset;
+		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, io_u))
+	if (get_next_offset(td, io_u, &is_random)) {
+		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)
+	io_u->buflen = get_next_buflen(td, io_u, is_random);
+	if (!io_u->buflen) {
+		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
 		return 1;
+	}
+
+	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",
+			(unsigned long long) io_u->offset, io_u->buflen,
+			(unsigned long long) io_u->file->real_file_size);
+		return 1;
+	}
 
 	/*
 	 * mark entry before potentially trimming io_u
 	 */
-	if (td_random(td) && !td->o.norandommap)
+	if (td_random(td) && file_randommap(td, io_u->file))
 		mark_random_map(td, io_u);
 
-	/*
-	 * If using a write iolog, store this entry.
-	 */
 out:
-	if (td->o.write_iolog_file)
-		write_iolog_put(td, io_u);
-
+	dprint_io_u(io_u, "fill_io_u");
+	td->zone_bytes += io_u->buflen;
 	return 0;
 }
 
-void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
+static void __io_u_mark_map(unsigned int *map, unsigned int nr)
 {
-	int index = 0;
+	int idx = 0;
 
-	if (io_u->ddir == DDIR_SYNC)
-		return;
+	switch (nr) {
+	default:
+		idx = 6;
+		break;
+	case 33 ... 64:
+		idx = 5;
+		break;
+	case 17 ... 32:
+		idx = 4;
+		break;
+	case 9 ... 16:
+		idx = 3;
+		break;
+	case 5 ... 8:
+		idx = 2;
+		break;
+	case 1 ... 4:
+		idx = 1;
+	case 0:
+		break;
+	}
+
+	map[idx]++;
+}
+
+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 idx = 0;
 
 	switch (td->cur_depth) {
 	default:
-		index++;
+		idx = 6;
+		break;
 	case 32 ... 63:
-		index++;
+		idx = 5;
+		break;
 	case 16 ... 31:
-		index++;
+		idx = 4;
+		break;
 	case 8 ... 15:
-		index++;
+		idx = 3;
+		break;
 	case 4 ... 7:
-		index++;
+		idx = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		idx = 1;
 	case 1:
 		break;
 	}
 
-	td->ts.io_u_map[index]++;
-	td->ts.total_io_u[io_u->ddir]++;
+	td->ts.io_u_map[idx] += 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;
+	int idx = 0;
+
+	assert(usec < 1000);
+
+	switch (usec) {
+	case 750 ... 999:
+		idx = 9;
+		break;
+	case 500 ... 749:
+		idx = 8;
+		break;
+	case 250 ... 499:
+		idx = 7;
+		break;
+	case 100 ... 249:
+		idx = 6;
+		break;
+	case 50 ... 99:
+		idx = 5;
+		break;
+	case 20 ... 49:
+		idx = 4;
+		break;
+	case 10 ... 19:
+		idx = 3;
+		break;
+	case 4 ... 9:
+		idx = 2;
+		break;
+	case 2 ... 3:
+		idx = 1;
+	case 0 ... 1:
+		break;
+	}
+
+	assert(idx < FIO_IO_U_LAT_U_NR);
+	td->ts.io_u_lat_u[idx]++;
+}
+
+static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
+{
+	int idx = 0;
 
 	switch (msec) {
 	default:
-		index++;
+		idx = 11;
+		break;
 	case 1000 ... 1999:
-		index++;
+		idx = 10;
+		break;
 	case 750 ... 999:
-		index++;
+		idx = 9;
+		break;
 	case 500 ... 749:
-		index++;
+		idx = 8;
+		break;
 	case 250 ... 499:
-		index++;
+		idx = 7;
+		break;
 	case 100 ... 249:
-		index++;
+		idx = 6;
+		break;
 	case 50 ... 99:
-		index++;
+		idx = 5;
+		break;
 	case 20 ... 49:
-		index++;
+		idx = 4;
+		break;
 	case 10 ... 19:
-		index++;
+		idx = 3;
+		break;
 	case 4 ... 9:
-		index++;
+		idx = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		idx = 1;
 	case 0 ... 1:
 		break;
 	}
 
-	td->ts.io_u_lat[index]++;
+	assert(idx < FIO_IO_U_LAT_M_NR);
+	td->ts.io_u_lat_m[idx]++;
+}
+
+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 {
+		int opened = 0;
+		unsigned long r;
+
+		if (td->o.use_os_rand) {
+			r = os_random_long(&td->next_file_state);
+			fno = (unsigned int) ((double) td->o.nr_files
+				* (r / (OS_RAND_MAX + 1.0)));
+		} else {
+			r = __rand(&td->__next_file_state);
+			fno = (unsigned int) ((double) td->o.nr_files
+				* (r / (FRAND_MAX + 1.0)));
+		}
+
+		f = td->files[fno];
+		if (fio_file_done(f))
+			continue;
+
+		if (!fio_file_open(f)) {
+			int err;
+
+			if (td->nr_open_files >= td->o.open_files)
+				return ERR_PTR(-EBUSY);
+
+			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;
+
+			if (td->nr_open_files >= td->o.open_files)
+				return ERR_PTR(-EBUSY);
+
+			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);
+
+	if (IS_ERR(f))
+		return f;
+
+	td->file_service_file = f;
+	td->file_service_left = td->file_service_nr - 1;
+out:
+	if (f)
+		dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
+	else
+		dprint(FD_FILE, "get_next_file: NULL\n");
+	return f;
+}
+
+static struct fio_file *get_next_file(struct thread_data *td)
+{
+	if (td->flags & TD_F_PROFILE_OPS) {
+		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 long set_io_u_file(struct thread_data *td, struct io_u *io_u)
+{
+	struct fio_file *f;
+
+	do {
+		f = get_next_file(td);
+		if (IS_ERR_OR_NULL(f))
+			return PTR_ERR(f);
+
+		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;
 }
 
-/*
- * Get next file to service by choosing one at random
- */
-static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
-					   int badf)
+static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
+		      unsigned long tusec, unsigned long max_usec)
 {
-	struct fio_file *f;
-	int fno;
-
-	do {
-		long r = os_random_long(&td->next_file_state);
-
-		fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
-		f = &td->files[fno];
-		if (f->flags & FIO_FILE_DONE)
-			continue;
+	if (!td->error)
+		log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
+	td_verror(td, ETIMEDOUT, "max latency exceeded");
+	icd->error = ETIMEDOUT;
+}
 
-		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
-			return f;
-	} while (1);
+static void lat_new_cycle(struct thread_data *td)
+{
+	fio_gettime(&td->latency_ts, NULL);
+	td->latency_ios = ddir_rw_sum(td->io_blocks);
+	td->latency_failed = 0;
 }
 
 /*
- * Get next file to service by doing round robin between all available ones
+ * We had an IO outside the latency target. Reduce the queue depth. If we
+ * are at QD=1, then it's time to give up.
  */
-static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
-					 int badf)
+static int __lat_target_failed(struct thread_data *td)
 {
-	unsigned int old_next_file = td->next_file;
-	struct fio_file *f;
+	if (td->latency_qd == 1)
+		return 1;
 
-	do {
-		f = &td->files[td->next_file];
+	td->latency_qd_high = td->latency_qd;
 
-		td->next_file++;
-		if (td->next_file >= td->o.nr_files)
-			td->next_file = 0;
+	if (td->latency_qd == td->latency_qd_low)
+		td->latency_qd_low--;
 
-		if (f->flags & FIO_FILE_DONE) {
-			f = NULL;
-			continue;
-		}
+	td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
 
-		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
-			break;
+	dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
 
-		f = NULL;
-	} while (td->next_file != old_next_file);
+	/*
+	 * When we ramp QD down, quiesce existing IO to prevent
+	 * a storm of ramp downs due to pending higher depth.
+	 */
+	io_u_quiesce(td);
+	lat_new_cycle(td);
+	return 0;
+}
 
-	return f;
+static int lat_target_failed(struct thread_data *td)
+{
+	if (td->o.latency_percentile.u.f == 100.0)
+		return __lat_target_failed(td);
+
+	td->latency_failed++;
+	return 0;
 }
 
-static struct fio_file *get_next_file(struct thread_data *td)
+void lat_target_init(struct thread_data *td)
 {
-	struct fio_file *f;
+	td->latency_end_run = 0;
+
+	if (td->o.latency_target) {
+		dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
+		fio_gettime(&td->latency_ts, NULL);
+		td->latency_qd = 1;
+		td->latency_qd_high = td->o.iodepth;
+		td->latency_qd_low = 1;
+		td->latency_ios = ddir_rw_sum(td->io_blocks);
+	} else
+		td->latency_qd = td->o.iodepth;
+}
 
-	assert(td->o.nr_files <= td->files_index);
+void lat_target_reset(struct thread_data *td)
+{
+	if (!td->latency_end_run)
+		lat_target_init(td);
+}
 
-	if (!td->nr_open_files)
-		return NULL;
+static void lat_target_success(struct thread_data *td)
+{
+	const unsigned int qd = td->latency_qd;
+	struct thread_options *o = &td->o;
 
-	f = td->file_service_file;
-	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
-		return f;
+	td->latency_qd_low = td->latency_qd;
 
-	if (td->o.file_service_type == FIO_FSERVICE_RR)
-		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+	/*
+	 * If we haven't failed yet, we double up to a failing value instead
+	 * of bisecting from highest possible queue depth. If we have set
+	 * a limit other than td->o.iodepth, bisect between that.
+	 */
+	if (td->latency_qd_high != o->iodepth)
+		td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
 	else
-		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+		td->latency_qd *= 2;
 
-	td->file_service_file = f;
-	td->file_service_left = td->file_service_nr - 1;
-	return f;
+	if (td->latency_qd > o->iodepth)
+		td->latency_qd = o->iodepth;
+
+	dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
+
+	/*
+	 * Same as last one, we are done. Let it run a latency cycle, so
+	 * we get only the results from the targeted depth.
+	 */
+	if (td->latency_qd == qd) {
+		if (td->latency_end_run) {
+			dprint(FD_RATE, "We are done\n");
+			td->done = 1;
+		} else {
+			dprint(FD_RATE, "Quiesce and final run\n");
+			io_u_quiesce(td);
+			td->latency_end_run = 1;
+			reset_all_stats(td);
+			reset_io_stats(td);
+		}
+	}
+
+	lat_new_cycle(td);
 }
 
-static struct fio_file *find_next_new_file(struct thread_data *td)
+/*
+ * Check if we can bump the queue depth
+ */
+void lat_target_check(struct thread_data *td)
 {
-	struct fio_file *f;
+	uint64_t usec_window;
+	uint64_t ios;
+	double success_ios;
+
+	usec_window = utime_since_now(&td->latency_ts);
+	if (usec_window < td->o.latency_window)
+		return;
 
-	if (td->o.file_service_type == FIO_FSERVICE_RR)
-		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
+	ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
+	success_ios = (double) (ios - td->latency_failed) / (double) ios;
+	success_ios *= 100.0;
+
+	dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
+
+	if (success_ios >= td->o.latency_percentile.u.f)
+		lat_target_success(td);
 	else
-		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
+		__lat_target_failed(td);
+}
 
-	return f;
+/*
+ * If latency target is enabled, we might be ramping up or down and not
+ * using the full queue depth available.
+ */
+int queue_full(struct thread_data *td)
+{
+	const int qempty = io_u_qempty(&td->io_u_freelist);
+
+	if (qempty)
+		return 1;
+	if (!td->o.latency_target)
+		return 0;
+
+	return td->cur_depth >= td->latency_qd;
 }
 
 struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
 
-	if (!list_empty(&td->io_u_requeues))
-		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
+	td_io_u_lock(td);
+
+again:
+	if (!io_u_rempty(&td->io_u_requeues))
+		io_u = io_u_rpop(&td->io_u_requeues);
 	else if (!queue_full(td)) {
-		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
+		io_u = io_u_qpop(&td->io_u_freelist);
 
+		io_u->file = NULL;
 		io_u->buflen = 0;
 		io_u->resid = 0;
-		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->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
+				 IO_U_F_TRIMMED | IO_U_F_BARRIER |
+				 IO_U_F_VER_LIST);
 
 		io_u->error = 0;
-		list_del(&io_u->list);
-		list_add(&io_u->list, &td->io_u_busylist);
+		io_u->acct_ddir = -1;
 		td->cur_depth++;
+		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+		io_u->ipo = NULL;
+	} 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->flags & TD_F_TRIM_BACKLOG))
+		return 0;
+
+	if (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->flags & TD_F_VER_BACKLOG))
+		return 0;
+
+	if (td->io_hist_len) {
+		int get_verify = 0;
+
+		if (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)) {
+			td->verify_batch--;
+			return 1;
+		}
+	}
+
+	return 0;
+}
+
+/*
+ * Fill offset and start time into the buffer content, to prevent too
+ * easy compressible data for simple de-dupe attempts. Do this for every
+ * 512b block in the range, since that should be the smallest block size
+ * we can expect from a device.
+ */
+static void small_content_scramble(struct io_u *io_u)
+{
+	unsigned int i, nr_blocks = io_u->buflen / 512;
+	uint64_t boffset;
+	unsigned int offset;
+	void *p, *end;
+
+	if (!nr_blocks)
+		return;
+
+	p = io_u->xfer_buf;
+	boffset = io_u->offset;
+	io_u->buf_filled_len = 0;
+
+	for (i = 0; i < nr_blocks; i++) {
+		/*
+		 * Fill the byte offset into a "random" start offset of
+		 * the buffer, given by the product of the usec time
+		 * and the actual offset.
+		 */
+		offset = (io_u->start_time.tv_usec ^ boffset) & 511;
+		offset &= ~(sizeof(uint64_t) - 1);
+		if (offset >= 512 - sizeof(uint64_t))
+			offset -= sizeof(uint64_t);
+		memcpy(p + offset, &boffset, sizeof(boffset));
+
+		end = p + 512 - sizeof(io_u->start_time);
+		memcpy(end, &io_u->start_time, sizeof(io_u->start_time));
+		p += 512;
+		boffset += 512;
+	}
+}
+
 /*
  * 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.
@@ -517,11 +1435,19 @@ struct io_u *get_io_u(struct thread_data *td)
 {
 	struct fio_file *f;
 	struct io_u *io_u;
-	int ret;
+	int do_scramble = 0;
+	long ret = 0;
 
 	io_u = __get_io_u(td);
-	if (!io_u)
+	if (!io_u) {
+		dprint(FD_IO, "__get_io_u failed\n");
 		return NULL;
+	}
+
+	if (check_get_verify(td, io_u))
+		goto out;
+	if (check_get_trim(td, io_u))
+		goto out;
 
 	/*
 	 * from a requeue, io_u already setup
@@ -529,152 +1455,276 @@ struct io_u *get_io_u(struct thread_data *td)
 	if (io_u->file)
 		goto out;
 
-	do {
-		f = get_next_file(td);
-		if (!f) {
-			put_io_u(td, io_u);
-			return NULL;
-		}
-
-set_file:
-		io_u->file = f;
-
-		if (!fill_io_u(td, io_u))
-			break;
-
-		/*
-		 * No more to do for this file, close it
-		 */
-		io_u->file = NULL;
-		td_io_close_file(td, f);
-		f->flags |= FIO_FILE_DONE;
-
-		/*
-		 * probably not the right place to do this, but see
-		 * if we need to open a new file
-		 */
-		if (td->nr_open_files < td->o.open_files &&
-		    td->o.open_files != td->o.nr_files) {
-			f = find_next_new_file(td);
-
-			if (!f || (ret = td_io_open_file(td, f))) {
-				put_io_u(td, io_u);
-				return NULL;
-			}
-			goto set_file;
-		}
-	} while (1);
+	/*
+	 * If using an iolog, grab next piece if any available.
+	 */
+	if (td->flags & TD_F_READ_IOLOG) {
+		if (read_iolog_get(td, io_u))
+			goto err_put;
+	} else if (set_io_u_file(td, io_u)) {
+		ret = -EBUSY;
+		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
+		goto err_put;
+	}
 
-	if (td->zone_bytes >= td->o.zone_size) {
-		td->zone_bytes = 0;
-		f->last_pos += td->o.zone_skip;
+	f = io_u->file;
+	if (!f) {
+		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;
+	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->o.verify != VERIFY_NONE)
-			populate_verify_io_u(td, io_u);
+		if (io_u->ddir == DDIR_WRITE) {
+			if (td->flags & TD_F_REFILL_BUFFERS) {
+				io_u_fill_buffer(td, io_u,
+					io_u->xfer_buflen, io_u->xfer_buflen);
+			} else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
+				   !(td->flags & TD_F_COMPRESS))
+				do_scramble = 1;
+			if (td->flags & TD_F_VER_NONE) {
+				populate_verify_io_u(td, io_u);
+				do_scramble = 0;
+			}
+		} 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;
+		}
 	}
 
 	/*
 	 * Set io data pointers.
 	 */
-out:
 	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);
+		if (do_scramble)
+			small_content_scramble(io_u);
+		return io_u;
 	}
-
-	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 ERR_PTR(ret);
 }
 
 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 {
-	const char *msg[] = { "read", "write", "sync" };
-
-	log_err("fio: io_u error");
-
-	if (io_u->file)
-		log_err(" on file %s", io_u->file->file_name);
+	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
 
-	log_err(": %s\n", strerror(io_u->error));
+	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
+		return;
 
-	log_err("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], io_u->offset, io_u->xfer_buflen);
+	log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
+		io_u->file ? " on file " : "",
+		io_u->file ? io_u->file->file_name : "",
+		strerror(io_u->error),
+		io_ddir_name(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,
+static inline int gtod_reduce(struct thread_data *td)
+{
+	return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
+		&& td->o.disable_bw;
+}
+
+static void account_io_completion(struct thread_data *td, struct io_u *io_u,
+				  struct io_completion_data *icd,
+				  const enum fio_ddir idx, unsigned int bytes)
+{
+	unsigned long lusec = 0;
+
+	if (!gtod_reduce(td))
+		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, io_u->offset);
+
+		if (td->flags & TD_F_PROFILE_OPS) {
+			struct prof_io_ops *ops = &td->prof_io_ops;
+
+			if (ops->io_u_lat)
+				icd->error = ops->io_u_lat(td, tusec);
+		}
+
+		if (td->o.max_latency && tusec > td->o.max_latency)
+			lat_fatal(td, icd, tusec, td->o.max_latency);
+		if (td->o.latency_target && tusec > td->o.latency_target) {
+			if (lat_target_failed(td))
+				lat_fatal(td, icd, tusec, td->o.latency_target);
+		}
+	}
+
+	if (!td->o.disable_clat) {
+		add_clat_sample(td, idx, lusec, bytes, io_u->offset);
+		io_u_mark_latency(td, lusec);
+	}
+
+	if (!td->o.disable_bw)
+		add_bw_sample(td, idx, bytes, &icd->time);
+
+	if (!gtod_reduce(td))
+		add_iops_sample(td, idx, bytes, &icd->time);
+}
+
+static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
+{
+	uint64_t secs, remainder, bps, bytes;
+
+	bytes = td->this_io_bytes[ddir];
+	bps = td->rate_bps[ddir];
+	secs = bytes / bps;
+	remainder = bytes % bps;
+	return remainder * 1000000 / bps + secs * 1000000;
+}
+
+static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 			 struct io_completion_data *icd)
 {
-	unsigned long msec;
+	struct io_u *io_u = *io_u_ptr;
+	enum fio_ddir ddir = io_u->ddir;
+	struct fio_file *f = io_u->file;
 
+	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->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
+
+	/*
+	 * Mark IO ok to verify
+	 */
+	if (io_u->ipo) {
+		/*
+		 * Remove errored entry from the verification list
+		 */
+		if (io_u->error)
+			unlog_io_piece(td, io_u);
+		else {
+			io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
+			write_barrier();
+		}
+	}
 
-	put_file(td, io_u->file);
+	td_io_u_unlock(td);
 
-	if (io_u->ddir == DDIR_SYNC) {
+	if (ddir_sync(ddir)) {
 		td->last_was_sync = 1;
+		if (f) {
+			f->first_write = -1ULL;
+			f->last_write = -1ULL;
+		}
 		return;
 	}
 
 	td->last_was_sync = 0;
+	td->last_ddir = ddir;
 
-	if (!io_u->error) {
+	if (!io_u->error && ddir_rw(ddir)) {
 		unsigned int bytes = io_u->buflen - io_u->resid;
-		const enum fio_ddir idx = io_u->ddir;
+		const enum fio_ddir oddir = ddir ^ 1;
 		int ret;
 
-		td->io_blocks[idx]++;
-		td->io_bytes[idx] += bytes;
-		td->zone_bytes += bytes;
-		td->this_io_bytes[idx] += bytes;
+		td->io_blocks[ddir]++;
+		td->this_io_blocks[ddir]++;
+		td->io_bytes[ddir] += bytes;
 
-		io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
+		if (!(io_u->flags & IO_U_F_VER_LIST))
+			td->this_io_bytes[ddir] += bytes;
 
-		msec = mtime_since(&io_u->issue_time, &icd->time);
+		if (ddir == DDIR_WRITE && 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) && (td->runstate == TD_RUNNING ||
+					   td->runstate == TD_VERIFYING)) {
+			account_io_completion(td, io_u, icd, ddir, bytes);
 
-		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
-		    td->o.verify != VERIFY_NONE)
-			log_io_piece(td, io_u);
+			if (__should_check_rate(td, ddir)) {
+				td->rate_pending_usleep[ddir] =
+					(usec_for_io(td, ddir) -
+					 utime_since_now(&td->start));
+			}
+			if (ddir != DDIR_TRIM &&
+			    __should_check_rate(td, oddir)) {
+				td->rate_pending_usleep[oddir] =
+					(usec_for_io(td, oddir) -
+					 utime_since_now(&td->start));
+			}
+		}
 
-		icd->bytes_done[idx] += bytes;
+		icd->bytes_done[ddir] += bytes;
 
 		if (io_u->end_io) {
-			ret = io_u->end_io(td, io_u);
+			ret = io_u->end_io(td, io_u_ptr);
+			io_u = *io_u_ptr;
 			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 (icd->error) {
+		enum error_type_bit eb = td_error_type(ddir, icd->error);
+
+		if (!td_non_fatal_error(td, eb, icd->error))
+			return;
+
+		/*
+		 * 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;
+		if (io_u)
+			io_u->error = 0;
+	}
 }
 
-static void init_icd(struct io_completion_data *icd, int nr)
+static void init_icd(struct thread_data *td, struct io_completion_data *icd,
+		     int nr)
 {
-	fio_gettime(&icd->time, NULL);
+	int ddir;
+
+	if (!gtod_reduce(td))
+		fio_gettime(&icd->time, NULL);
 
 	icd->nr = nr;
 
 	icd->error = 0;
-	icd->bytes_done[0] = icd->bytes_done[1] = 0;
+	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+		icd->bytes_done[ddir] = 0;
 }
 
 static void ios_completed(struct thread_data *td,
@@ -686,56 +1736,82 @@ static void ios_completed(struct thread_data *td,
 	for (i = 0; i < icd->nr; i++) {
 		io_u = td->io_ops->event(td, i);
 
-		io_completed(td, io_u, icd);
-		put_io_u(td, io_u);
+		io_completed(td, &io_u, icd);
+
+		if (io_u)
+			put_io_u(td, io_u);
 	}
 }
 
 /*
  * Complete a single io_u for the sync engines.
  */
-long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
+int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
+		       uint64_t *bytes)
 {
 	struct io_completion_data icd;
 
-	init_icd(&icd, 1);
-	io_completed(td, io_u, &icd);
-	put_io_u(td, io_u);
+	init_icd(td, &icd, 1);
+	io_completed(td, &io_u, &icd);
+
+	if (io_u)
+		put_io_u(td, io_u);
+
+	if (icd.error) {
+		td_verror(td, icd.error, "io_u_sync_complete");
+		return -1;
+	}
+
+	if (bytes) {
+		int ddir;
 
-	if (!icd.error)
-		return icd.bytes_done[0] + icd.bytes_done[1];
+		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+			bytes[ddir] += icd.bytes_done[ddir];
+	}
 
-	td_verror(td, icd.error, "io_u_sync_complete");
-	return -1;
+	return 0;
 }
 
 /*
  * Called to complete min_events number of io for the async engines.
  */
-long io_u_queued_complete(struct thread_data *td, int min_events)
+int io_u_queued_complete(struct thread_data *td, int min_evts,
+			 uint64_t *bytes)
 {
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
 	int ret;
 	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-	if (!min_events)
+	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
+
+	if (!min_evts)
 		tvp = &ts;
+	else if (min_evts > td->cur_depth)
+		min_evts = td->cur_depth;
 
-	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
+	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
 	if (ret < 0) {
 		td_verror(td, -ret, "td_io_getevents");
 		return ret;
 	} else if (!ret)
 		return ret;
 
-	init_icd(&icd, 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;
+	}
 
-	td_verror(td, icd.error, "io_u_queued_complete");
-	return -1;
+	if (bytes) {
+		int ddir;
+
+		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+			bytes[ddir] += icd.bytes_done[ddir];
+	}
+
+	return 0;
 }
 
 /*
@@ -743,88 +1819,44 @@ long io_u_queued_complete(struct thread_data *td, int min_events)
  */
 void io_u_queued(struct thread_data *td, struct io_u *io_u)
 {
-	unsigned long slat_time;
+	if (!td->o.disable_slat) {
+		unsigned long slat_time;
 
-	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
-	add_slat_sample(td, io_u->ddir, slat_time);
-}
-
-#ifdef FIO_USE_TIMEOUT
-void io_u_set_timeout(struct thread_data *td)
-{
-	assert(td->cur_depth);
-
-	td->timer.it_interval.tv_sec = 0;
-	td->timer.it_interval.tv_usec = 0;
-	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
-	td->timer.it_value.tv_usec = 0;
-	setitimer(ITIMER_REAL, &td->timer, NULL);
-	fio_gettime(&td->timeout_end, NULL);
-}
-
-static void io_u_dump(struct io_u *io_u)
-{
-	unsigned long t_start = mtime_since_now(&io_u->start_time);
-	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
-
-	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
-	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
-	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
-}
-#else
-void io_u_set_timeout(struct thread_data fio_unused *td)
-{
+		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,
+				io_u->offset);
+	}
 }
-#endif
 
-#ifdef FIO_USE_TIMEOUT
-static void io_u_timeout_handler(int fio_unused sig)
+void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
+		    unsigned int max_bs)
 {
-	struct thread_data *td, *__td;
-	pid_t pid = getpid();
-	struct list_head *entry;
-	struct io_u *io_u;
-	int i;
-
-	log_err("fio: io_u timeout\n");
-
-	/*
-	 * TLS would be nice...
-	 */
-	td = NULL;
-	for_each_td(__td, i) {
-		if (__td->pid == pid) {
-			td = __td;
-			break;
-		}
-	}
-
-	if (!td) {
-		log_err("fio: io_u timeout, can't find job\n");
-		exit(1);
-	}
-
-	if (!td->cur_depth) {
-		log_err("fio: timeout without pending work?\n");
-		return;
-	}
-
-	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
-
-	list_for_each(entry, &td->io_u_busylist) {
-		io_u = list_entry(entry, struct io_u, list);
-
-		io_u_dump(io_u);
-	}
-
-	td_verror(td, ETIMEDOUT, "io_u timeout");
-	exit(1);
+	if (td->o.buffer_pattern_bytes)
+		fill_buffer_pattern(td, buf, max_bs);
+	else if (!td->o.zero_buffers) {
+		unsigned int perc = td->o.compress_percentage;
+
+		if (perc) {
+			unsigned int seg = min_write;
+
+			seg = min(min_write, td->o.compress_chunk);
+			if (!seg)
+				seg = min_write;
+
+			fill_random_buf_percentage(&td->buf_state, buf,
+						perc, seg, max_bs);
+		} else
+			fill_random_buf(&td->buf_state, buf, max_bs);
+	} else
+		memset(buf, 0, max_bs);
 }
-#endif
 
-void io_u_init_timeout(void)
+/*
+ * "randomly" fill the buffer contents
+ */
+void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
+		      unsigned int min_write, unsigned int max_bs)
 {
-#ifdef FIO_USE_TIMEOUT
-	signal(SIGALRM, io_u_timeout_handler);
-#endif
+	io_u->buf_filled_len = 0;
+	fill_io_buffer(td, io_u->buf, min_write, max_bs);
 }