X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=16065128c324e76e3cf9326ff1824fec42baad5b;hp=f0b6170535d2a77a07861a4a0760d72a4130412e;hb=56d9fa4b8d4fa5166e3ec5dcdd37b5789b2cb01d;hpb=1f64bd7cbb8b16d64b30c65393adf9c028713e42

diff --git a/io_u.c b/io_u.c
index f0b61705..16065128 100644
--- a/io_u.c
+++ b/io_u.c
@@ -11,6 +11,7 @@
 #include "trim.h"
 #include "lib/rand.h"
 #include "lib/axmap.h"
+#include "err.h"
 
 struct io_completion_data {
 	int nr;				/* input */
@@ -67,6 +68,9 @@ static uint64_t last_block(struct thread_data *td, struct fio_file *f,
 	if (td->o.zone_range)
 		max_size = td->o.zone_range;
 
+	if (td->o.min_bs[ddir] > td->o.ba[ddir])
+		max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
+
 	max_blocks = max_size / (uint64_t) td->o.ba[ddir];
 	if (!max_blocks)
 		return 0;
@@ -82,32 +86,26 @@ struct rand_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;
+	uint64_t r;
 
 	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
-		uint64_t rmax;
+		uint64_t lastb;
 
-		rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
+		lastb = last_block(td, f, ddir);
+		if (!lastb)
+			return 1;
 
-		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);
-		}
+		r = __rand(&td->random_state);
 
 		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
 
-		*b = (lastb - 1) * (r / ((uint64_t) rmax + 1.0));
+		*b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
 	} else {
 		uint64_t off = 0;
 
-		if (lfsr_next(&f->lfsr, &off, lastb))
+		assert(fio_file_lfsr(f));
+
+		if (lfsr_next(&f->lfsr, &off))
 			return 1;
 
 		*b = off;
@@ -151,6 +149,15 @@ static int __get_next_rand_offset_pareto(struct thread_data *td,
 	return 0;
 }
 
+static int __get_next_rand_offset_gauss(struct thread_data *td,
+					struct fio_file *f, enum fio_ddir ddir,
+					uint64_t *b)
+{
+	*b = gauss_next(&f->gauss);
+	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);
@@ -168,6 +175,8 @@ static int get_off_from_method(struct thread_data *td, struct fio_file *f,
 		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);
+	else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
+		return __get_next_rand_offset_gauss(td, f, ddir, b);
 
 	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
 	return 1;
@@ -191,23 +200,18 @@ static inline int should_sort_io(struct thread_data *td)
 	return 1;
 }
 
-static int should_do_random(struct thread_data *td)
+static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
 {
 	unsigned int v;
 	unsigned long r;
 
-	if (td->o.perc_rand == 100)
+	if (td->o.perc_rand[ddir] == 100)
 		return 1;
 
-	if (td->o.use_os_rand) {
-		r = os_random_long(&td->seq_rand_state);
-		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
-	} else {
-		r = __rand(&td->__seq_rand_state);
-		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
-	}
+	r = __rand(&td->seq_rand_state[ddir]);
+	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
 
-	return v <= td->o.perc_rand;
+	return v <= td->o.perc_rand[ddir];
 }
 
 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
@@ -220,9 +224,8 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 		return get_off_from_method(td, f, ddir, b);
 
 	if (!flist_empty(&td->next_rand_list)) {
-		struct rand_off *r;
 fetch:
-		r = flist_entry(td->next_rand_list.next, struct rand_off, list);
+		r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
 		flist_del(&r->list);
 		*b = r->off;
 		free(r);
@@ -262,7 +265,7 @@ static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
 	}
 
 	dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
-			f->file_name, (unsigned long long) f->last_pos,
+			f->file_name, (unsigned long long) f->last_pos[ddir],
 			(unsigned long long) f->real_file_size);
 	return 1;
 }
@@ -270,20 +273,32 @@ static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
 			       enum fio_ddir ddir, uint64_t *offset)
 {
+	struct thread_options *o = &td->o;
+
 	assert(ddir_rw(ddir));
 
-	if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
-		f->last_pos = f->last_pos - f->io_size;
+	if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
+	    o->time_based)
+		f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
 
-	if (f->last_pos < f->real_file_size) {
+	if (f->last_pos[ddir] < 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;
+		if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
+			f->last_pos[ddir] = f->real_file_size;
 
-		pos = f->last_pos - f->file_offset;
-		if (pos)
-			pos += td->o.ddir_seq_add;
+		pos = f->last_pos[ddir] - f->file_offset;
+		if (pos && o->ddir_seq_add) {
+			pos += o->ddir_seq_add;
+
+			/*
+			 * If we reach beyond the end of the file
+			 * with holed IO, wrap around to the
+			 * beginning again.
+			 */
+			if (pos >= f->real_file_size)
+				pos = f->file_offset;
+		}
 
 		*offset = pos;
 		return 0;
@@ -293,7 +308,8 @@ static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
 }
 
 static int get_next_block(struct thread_data *td, struct io_u *io_u,
-			  enum fio_ddir ddir, int rw_seq)
+			  enum fio_ddir ddir, int rw_seq,
+			  unsigned int *is_random)
 {
 	struct fio_file *f = io_u->file;
 	uint64_t b, offset;
@@ -305,26 +321,33 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 
 	if (rw_seq) {
 		if (td_random(td)) {
-			if (should_do_random(td))
+			if (should_do_random(td, ddir)) {
 				ret = get_next_rand_block(td, f, ddir, &b);
-			else {
+				*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
+		} 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)
+			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;
+			if (f->last_start[ddir] != -1ULL)
+				offset = f->last_start[ddir] - f->file_offset;
 			else
 				offset = 0;
 			ret = 0;
@@ -353,7 +376,8 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
  * 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;
 	enum fio_ddir ddir = io_u->ddir;
@@ -366,7 +390,7 @@ static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 		td->ddir_seq_nr = td->o.ddir_seq_nr;
 	}
 
-	if (get_next_block(td, io_u, ddir, rw_seq_hit))
+	if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
 		return 1;
 
 	if (io_u->offset >= f->io_size) {
@@ -387,16 +411,17 @@ static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 	return 0;
 }
 
-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)
 {
 	if (td->flags & TD_F_PROFILE_OPS) {
 		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 ops->fill_io_u_off(td, io_u, is_random);
 	}
 
-	return __get_next_offset(td, io_u);
+	return __get_next_offset(td, io_u, is_random);
 }
 
 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
@@ -404,18 +429,22 @@ static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
 {
 	struct fio_file *f = io_u->file;
 
-	return io_u->offset + buflen <= f->io_size + get_start_offset(td);
+	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)
+static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
+				      unsigned int is_random)
 {
-	const int ddir = io_u->ddir;
+	int ddir = io_u->ddir;
 	unsigned int buflen = 0;
 	unsigned int minbs, maxbs;
-	unsigned long r, rand_max;
+	unsigned long r;
 
 	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];
 
@@ -428,20 +457,12 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 	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);
+		r = __rand(&td->bsrange_state);
 
 		if (!td->o.bssplit_nr[ddir]) {
 			buflen = 1 + (unsigned int) ((double) maxbs *
-					(r / (rand_max + 1.0)));
+					(r / (FRAND_MAX + 1.0)));
 			if (buflen < minbs)
 				buflen = minbs;
 		} else {
@@ -453,12 +474,16 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 
 				buflen = bsp->bs;
 				perc += bsp->perc;
-				if ((r <= ((rand_max / 100L) * perc)) &&
+				if ((r <= ((FRAND_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);
 
@@ -467,16 +492,17 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 	return buflen;
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+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);
+			return ops->fill_io_u_size(td, io_u, is_random);
 	}
 
-	return __get_next_buflen(td, io_u);
+	return __get_next_buflen(td, io_u, is_random);
 }
 
 static void set_rwmix_bytes(struct thread_data *td)
@@ -497,13 +523,8 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 	unsigned int v;
 	unsigned long r;
 
-	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)));
-	}
+	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;
@@ -522,6 +543,12 @@ void io_u_quiesce(struct thread_data *td)
 	 * io's that have been actually submitted to an async engine,
 	 * and cur_depth is meaningless for sync engines.
 	 */
+	if (td->io_u_queued || td->cur_depth) {
+		int fio_unused ret;
+
+		ret = td_io_commit(td);
+	}
+
 	while (td->io_u_in_flight) {
 		int fio_unused ret;
 
@@ -532,7 +559,6 @@ void io_u_quiesce(struct thread_data *td)
 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));
@@ -567,9 +593,7 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 
 	io_u_quiesce(td);
 
-	fio_gettime(&t, NULL);
-	usec_sleep(td, usec);
-	usec = utime_since_now(&t);
+	usec = usec_sleep(td, usec);
 
 	td->rate_pending_usleep[ddir] -= usec;
 
@@ -577,9 +601,6 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 	if (td_rw(td) && __should_check_rate(td, odir))
 		td->rate_pending_usleep[odir] -= usec;
 
-	if (ddir_trim(ddir))
-		return ddir;
-
 	return ddir;
 }
 
@@ -658,7 +679,7 @@ static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
 
 void put_file_log(struct thread_data *td, struct fio_file *f)
 {
-	int ret = put_file(td, f);
+	unsigned int ret = put_file(td, f);
 
 	if (ret)
 		td_verror(td, ret, "file close");
@@ -668,16 +689,15 @@ void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
 	td_io_u_lock(td);
 
-	if (io_u->file && !(io_u->flags & IO_U_F_FREE_DEF))
+	if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
 		put_file_log(td, io_u->file);
+
 	io_u->file = NULL;
-	io_u->flags &= ~IO_U_F_FREE_DEF;
 	io_u->flags |= IO_U_F_FREE;
 
 	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);
+	io_u_qpush(&td->io_u_freelist, io_u);
 	td_io_u_unlock(td);
 	td_io_u_free_notify(td);
 }
@@ -704,14 +724,16 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 	__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);
+
+	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;
 
@@ -727,9 +749,17 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	 * See if it's time to switch to a new zone
 	 */
 	if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
+		struct fio_file *f = io_u->file;
+
 		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;
+		f->file_offset += td->o.zone_range + td->o.zone_skip;
+
+		/*
+		 * Wrap from the beginning, if we exceed the file size
+		 */
+		if (f->file_offset >= f->real_file_size)
+			f->file_offset = f->real_file_size - f->file_offset;
+		f->last_pos[io_u->ddir] = f->file_offset;
 		td->io_skip_bytes += td->o.zone_skip;
 	}
 
@@ -737,12 +767,12 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	 * 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);
+	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;
@@ -946,15 +976,9 @@ static struct fio_file *get_next_file_rand(struct thread_data *td,
 		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 = __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))
@@ -963,6 +987,9 @@ static struct fio_file *get_next_file_rand(struct thread_data *td,
 		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;
@@ -1005,6 +1032,9 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 		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",
@@ -1058,16 +1088,22 @@ static struct fio_file *__get_next_file(struct thread_data *td)
 	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:
-	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
+	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)) {
+	if (td->flags & TD_F_PROFILE_OPS) {
 		struct prof_io_ops *ops = &td->prof_io_ops;
 
 		if (ops->get_next_file)
@@ -1077,14 +1113,14 @@ static struct fio_file *get_next_file(struct thread_data *td)
 	return __get_next_file(td);
 }
 
-static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
+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 (!f)
-			return 1;
+		if (IS_ERR_OR_NULL(f))
+			return PTR_ERR(f);
 
 		io_u->file = f;
 		get_file(f);
@@ -1104,37 +1140,194 @@ static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 	return 0;
 }
 
+static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
+		      unsigned long tusec, unsigned long max_usec)
+{
+	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;
+}
+
+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;
+}
+
+/*
+ * 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 int __lat_target_failed(struct thread_data *td)
+{
+	if (td->latency_qd == 1)
+		return 1;
+
+	td->latency_qd_high = td->latency_qd;
+
+	if (td->latency_qd == td->latency_qd_low)
+		td->latency_qd_low--;
+
+	td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
+
+	dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
+
+	/*
+	 * 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;
+}
+
+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;
+}
+
+void lat_target_init(struct thread_data *td)
+{
+	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;
+}
+
+void lat_target_reset(struct thread_data *td)
+{
+	if (!td->latency_end_run)
+		lat_target_init(td);
+}
+
+static void lat_target_success(struct thread_data *td)
+{
+	const unsigned int qd = td->latency_qd;
+	struct thread_options *o = &td->o;
+
+	td->latency_qd_low = td->latency_qd;
+
+	/*
+	 * 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
+		td->latency_qd *= 2;
+
+	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);
+}
+
+/*
+ * Check if we can bump the queue depth
+ */
+void lat_target_check(struct thread_data *td)
+{
+	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;
+
+	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
+		__lat_target_failed(td);
+}
+
+/*
+ * If latency target is enabled, we might be ramping up or down and not
+ * using the full queue depth available.
+ */
+int queue_full(const 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 (td->stop_io)
+		return NULL;
+
 	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);
+	if (!io_u_rempty(&td->io_u_requeues))
+		io_u = io_u_rpop(&td->io_u_requeues);
 	else if (!queue_full(td)) {
-		io_u = flist_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_F_FREE_DEF);
-		io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
-		io_u->flags &= ~IO_U_F_VER_LIST;
+		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;
 		io_u->acct_ddir = -1;
-		flist_del(&io_u->list);
-		flist_add_tail(&io_u->list, &td->io_u_busylist);
 		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
@@ -1249,6 +1442,7 @@ struct io_u *get_io_u(struct thread_data *td)
 	struct fio_file *f;
 	struct io_u *io_u;
 	int do_scramble = 0;
+	long ret = 0;
 
 	io_u = __get_io_u(td);
 	if (!io_u) {
@@ -1274,11 +1468,17 @@ struct io_u *get_io_u(struct thread_data *td)
 		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;
 	}
 
 	f = io_u->file;
+	if (!f) {
+		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
+		goto err_put;
+	}
+
 	assert(fio_file_open(f));
 
 	if (ddir_rw(io_u->ddir)) {
@@ -1287,14 +1487,16 @@ struct io_u *get_io_u(struct thread_data *td)
 			goto err_put;
 		}
 
-		f->last_start = io_u->offset;
-		f->last_pos = io_u->offset + io_u->buflen;
+		f->last_start[io_u->ddir] = io_u->offset;
+		f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
 
 		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->o.min_bs[DDIR_WRITE],
+					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);
@@ -1327,64 +1529,73 @@ out:
 err_put:
 	dprint(FD_IO, "get_io_u failed\n");
 	put_io_u(td, io_u);
-	return NULL;
+	return ERR_PTR(ret);
 }
 
 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 {
 	enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
-	const char *msg[] = { "read", "write", "sync", "datasync",
-				"sync_file_range", "wait", "trim" };
 
 	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
 		return;
 
-	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);
+	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 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 (!td->o.disable_clat || !td->o.disable_bw)
+	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);
+		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) {
-			if (!td->error)
-				log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
-			td_verror(td, ETIMEDOUT, "max latency exceeded");
-			icd->error = ETIMEDOUT;
+		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);
+		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);
 
-	add_iops_sample(td, idx, &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)
@@ -1398,21 +1609,38 @@ static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 	return remainder * 1000000 / bps + secs * 1000000;
 }
 
-static void io_completed(struct thread_data *td, struct io_u *io_u,
+static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 			 struct io_completion_data *icd)
 {
-	struct fio_file *f;
+	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_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();
+		}
+	}
+
 	td_io_u_unlock(td);
 
-	if (ddir_sync(io_u->ddir)) {
+	if (ddir_sync(ddir)) {
 		td->last_was_sync = 1;
-		f = io_u->file;
 		if (f) {
 			f->first_write = -1ULL;
 			f->last_write = -1ULL;
@@ -1421,23 +1649,21 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	}
 
 	td->last_was_sync = 0;
-	td->last_ddir = io_u->ddir;
+	td->last_ddir = ddir;
 
-	if (!io_u->error && ddir_rw(io_u->ddir)) {
+	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 odx = io_u->ddir ^ 1;
+		const enum fio_ddir oddir = ddir ^ 1;
 		int ret;
 
-		td->io_blocks[idx]++;
-		td->this_io_blocks[idx]++;
-		td->io_bytes[idx] += bytes;
+		td->io_blocks[ddir]++;
+		td->this_io_blocks[ddir]++;
+		td->io_bytes[ddir] += bytes;
 
 		if (!(io_u->flags & IO_U_F_VER_LIST))
-			td->this_io_bytes[idx] += bytes;
+			td->this_io_bytes[ddir] += bytes;
 
-		if (idx == DDIR_WRITE) {
-			f = io_u->file;
+		if (ddir == DDIR_WRITE) {
 			if (f) {
 				if (f->first_write == -1ULL ||
 				    io_u->offset < f->first_write)
@@ -1446,33 +1672,37 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 				    ((io_u->offset + bytes) > f->last_write))
 					f->last_write = io_u->offset + bytes;
 			}
+			if (td->last_write_comp) {
+				int idx = td->last_write_idx++;
+
+				td->last_write_comp[idx] = io_u->offset;
+				if (td->last_write_idx == td->o.iodepth)
+					td->last_write_idx = 0;
+			}
 		}
 
 		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
 					   td->runstate == TD_VERIFYING)) {
-			account_io_completion(td, io_u, icd, idx, bytes);
+			account_io_completion(td, io_u, icd, ddir, bytes);
 
-			if (__should_check_rate(td, idx)) {
-				td->rate_pending_usleep[idx] =
-					(usec_for_io(td, idx) -
+			if (__should_check_rate(td, ddir)) {
+				td->rate_pending_usleep[ddir] =
+					(usec_for_io(td, ddir) -
 					 utime_since_now(&td->start));
 			}
-			if (idx != DDIR_TRIM && __should_check_rate(td, odx))
-				td->rate_pending_usleep[odx] =
-					(usec_for_io(td, odx) -
+			if (ddir != DDIR_TRIM &&
+			    __should_check_rate(td, oddir)) {
+				td->rate_pending_usleep[oddir] =
+					(usec_for_io(td, oddir) -
 					 utime_since_now(&td->start));
+			}
 		}
 
-		if (td_write(td) && idx == DDIR_WRITE &&
-		    td->o.do_verify &&
-		    td->o.verify != VERIFY_NONE &&
-		    !td->o.experimental_verify)
-			log_io_piece(td, io_u);
-
-		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;
 		}
@@ -1481,9 +1711,11 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 		io_u_log_error(td, io_u);
 	}
 	if (icd->error) {
-		enum error_type_bit eb = td_error_type(io_u->ddir, 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.
@@ -1491,7 +1723,8 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 		update_error_count(td, icd->error);
 		td_clear_error(td);
 		icd->error = 0;
-		io_u->error = 0;
+		if (io_u)
+			io_u->error = 0;
 	}
 }
 
@@ -1499,7 +1732,8 @@ static void init_icd(struct thread_data *td, struct io_completion_data *icd,
 		     int nr)
 {
 	int ddir;
-	if (!td->o.disable_clat || !td->o.disable_bw)
+
+	if (!gtod_reduce(td))
 		fio_gettime(&icd->time, NULL);
 
 	icd->nr = nr;
@@ -1518,9 +1752,9 @@ 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);
+		io_completed(td, &io_u, icd);
 
-		if (!(io_u->flags & IO_U_F_FREE_DEF))
+		if (io_u)
 			put_io_u(td, io_u);
 	}
 }
@@ -1534,9 +1768,9 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
 	struct io_completion_data icd;
 
 	init_icd(td, &icd, 1);
-	io_completed(td, io_u, &icd);
+	io_completed(td, &io_u, &icd);
 
-	if (!(io_u->flags & IO_U_F_FREE_DEF))
+	if (io_u)
 		put_io_u(td, io_u);
 
 	if (icd.error) {
@@ -1569,6 +1803,8 @@ int io_u_queued_complete(struct thread_data *td, int 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_evts, td->o.iodepth_batch_complete, tvp);
 	if (ret < 0) {
@@ -1603,29 +1839,77 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 		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);
+		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
+				io_u->offset);
 	}
 }
 
-void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
-		    unsigned int max_bs)
+/*
+ * See if we should reuse the last seed, if dedupe is enabled
+ */
+static struct frand_state *get_buf_state(struct thread_data *td)
 {
-	if (!td->o.zero_buffers) {
-		unsigned int perc = td->o.compress_percentage;
+	unsigned int v;
+	unsigned long r;
 
-		if (perc) {
-			unsigned int seg = min_write;
+	if (!td->o.dedupe_percentage)
+		return &td->buf_state;
+	else if (td->o.dedupe_percentage == 100)
+		return &td->buf_state_prev;
 
-			seg = min(min_write, td->o.compress_chunk);
-			if (!seg)
-				seg = min_write;
+	r = __rand(&td->dedupe_state);
+	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
 
-			fill_random_buf_percentage(&td->buf_state, buf,
-						perc, seg, max_bs);
-		} else
-			fill_random_buf(&td->buf_state, buf, max_bs);
-	} else
+	if (v <= td->o.dedupe_percentage)
+		return &td->buf_state_prev;
+
+	return &td->buf_state;
+}
+
+static void save_buf_state(struct thread_data *td, struct frand_state *rs)
+{
+	if (rs == &td->buf_state)
+		frand_copy(&td->buf_state_prev, rs);
+}
+
+void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
+		    unsigned int max_bs)
+{
+	struct thread_options *o = &td->o;
+
+	if (o->compress_percentage || o->dedupe_percentage) {
+		unsigned int perc = td->o.compress_percentage;
+		struct frand_state *rs;
+		unsigned int left = max_bs;
+
+		do {
+			rs = get_buf_state(td);
+
+			min_write = min(min_write, left);
+
+			if (perc) {
+				unsigned int seg = min_write;
+
+				seg = min(min_write, td->o.compress_chunk);
+				if (!seg)
+					seg = min_write;
+
+				fill_random_buf_percentage(rs, buf, perc, seg,
+					min_write, o->buffer_pattern,
+						   o->buffer_pattern_bytes);
+			} else
+				fill_random_buf(rs, buf, min_write);
+
+			buf += min_write;
+			left -= min_write;
+			save_buf_state(td, rs);
+		} while (left);
+	} else if (o->buffer_pattern_bytes)
+		fill_buffer_pattern(td, buf, max_bs);
+	else if (o->zero_buffers)
 		memset(buf, 0, max_bs);
+	else
+		fill_random_buf(get_buf_state(td), buf, max_bs);
 }
 
 /*