X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=a3540d14714ef45f59f7fcf22744bae507f6b491;hp=ba3f7ca00fcd1db4f657f5dcf7321fc8c9f7ca0e;hb=fb1a0dcdc0869f0c92c49f13e640c15f4ab3b7e2;hpb=36870c47801ca85d1a98c7ec47950b0aa360f865

diff --git a/io_u.c b/io_u.c
index ba3f7ca0..a3540d14 100644
--- a/io_u.c
+++ b/io_u.c
@@ -1,31 +1,30 @@
 #include <unistd.h>
-#include <fcntl.h>
 #include <string.h>
-#include <signal.h>
-#include <time.h>
 #include <assert.h>
 
 #include "fio.h"
-#include "hash.h"
 #include "verify.h"
 #include "trim.h"
 #include "lib/rand.h"
 #include "lib/axmap.h"
 #include "err.h"
+#include "lib/pow2.h"
+#include "minmax.h"
+#include "zbd.h"
 
 struct io_completion_data {
 	int nr;				/* input */
 
 	int error;			/* output */
 	uint64_t bytes_done[DDIR_RWDIR_CNT];	/* output */
-	struct timeval time;		/* output */
+	struct timespec time;		/* output */
 };
 
 /*
  * 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 fio_file *f, const uint64_t block)
+static bool random_map_free(struct fio_file *f, const uint64_t block)
 {
 	return !axmap_isset(f->io_axmap, block);
 }
@@ -33,21 +32,27 @@ static int random_map_free(struct fio_file *f, const uint64_t block)
 /*
  * Mark a given offset as used in the map.
  */
-static void mark_random_map(struct thread_data *td, struct io_u *io_u)
+static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u,
+				uint64_t offset, uint64_t buflen)
 {
-	unsigned int min_bs = td->o.rw_min_bs;
+	unsigned long long min_bs = td->o.min_bs[io_u->ddir];
 	struct fio_file *f = io_u->file;
-	unsigned int nr_blocks;
+	unsigned long long nr_blocks;
 	uint64_t block;
 
-	block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
-	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
+	block = (offset - f->file_offset) / (uint64_t) min_bs;
+	nr_blocks = (buflen + min_bs - 1) / min_bs;
+	assert(nr_blocks > 0);
 
-	if (!(io_u->flags & IO_U_F_BUSY_OK))
+	if (!(io_u->flags & IO_U_F_BUSY_OK)) {
 		nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
+		assert(nr_blocks > 0);
+	}
+
+	if ((nr_blocks * min_bs) < buflen)
+		buflen = nr_blocks * min_bs;
 
-	if ((nr_blocks * min_bs) < io_u->buflen)
-		io_u->buflen = nr_blocks * min_bs;
+	return buflen;
 }
 
 static uint64_t last_block(struct thread_data *td, struct fio_file *f,
@@ -60,12 +65,13 @@ static uint64_t last_block(struct thread_data *td, struct fio_file *f,
 
 	/*
 	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
+	 * -> not for now since there is code assuming it could go either.
 	 */
 	max_size = f->io_size;
 	if (max_size > f->real_file_size)
 		max_size = f->real_file_size;
 
-	if (td->o.zone_range)
+	if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
 		max_size = td->o.zone_range;
 
 	if (td->o.min_bs[ddir] > td->o.ba[ddir])
@@ -78,28 +84,20 @@ static uint64_t last_block(struct thread_data *td, struct fio_file *f,
 	return max_blocks;
 }
 
-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)
+				  enum fio_ddir ddir, uint64_t *b,
+				  uint64_t lastb)
 {
 	uint64_t r;
 
-	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
-		uint64_t lastb;
-
-		lastb = last_block(td, f, ddir);
-		if (!lastb)
-			return 1;
+	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
+	    td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
 
 		r = __rand(&td->random_state);
 
 		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
 
-		*b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
+		*b = lastb * (r / (rand_max(&td->random_state) + 1.0));
 	} else {
 		uint64_t off = 0;
 
@@ -157,99 +155,172 @@ static int __get_next_rand_offset_gauss(struct thread_data *td,
 	return 0;
 }
 
+static int __get_next_rand_offset_zoned_abs(struct thread_data *td,
+					    struct fio_file *f,
+					    enum fio_ddir ddir, uint64_t *b)
+{
+	struct zone_split_index *zsi;
+	uint64_t lastb, send, stotal;
+	unsigned int v;
+
+	lastb = last_block(td, f, ddir);
+	if (!lastb)
+		return 1;
+
+	if (!td->o.zone_split_nr[ddir]) {
+bail:
+		return __get_next_rand_offset(td, f, ddir, b, lastb);
+	}
+
+	/*
+	 * Generate a value, v, between 1 and 100, both inclusive
+	 */
+	v = rand_between(&td->zone_state, 1, 100);
+
+	/*
+	 * Find our generated table. 'send' is the end block of this zone,
+	 * 'stotal' is our start offset.
+	 */
+	zsi = &td->zone_state_index[ddir][v - 1];
+	stotal = zsi->size_prev / td->o.ba[ddir];
+	send = zsi->size / td->o.ba[ddir];
+
+	/*
+	 * Should never happen
+	 */
+	if (send == -1U) {
+		if (!fio_did_warn(FIO_WARN_ZONED_BUG))
+			log_err("fio: bug in zoned generation\n");
+		goto bail;
+	} else if (send > lastb) {
+		/*
+		 * This happens if the user specifies ranges that exceed
+		 * the file/device size. We can't handle that gracefully,
+		 * so error and exit.
+		 */
+		log_err("fio: zoned_abs sizes exceed file size\n");
+		return 1;
+	}
+
+	/*
+	 * Generate index from 0..send-stotal
+	 */
+	if (__get_next_rand_offset(td, f, ddir, b, send - stotal) == 1)
+		return 1;
+
+	*b += stotal;
+	return 0;
+}
 
-static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
+static int __get_next_rand_offset_zoned(struct thread_data *td,
+					struct fio_file *f, enum fio_ddir ddir,
+					uint64_t *b)
 {
-	struct rand_off *r1 = flist_entry(a, struct rand_off, list);
-	struct rand_off *r2 = flist_entry(b, struct rand_off, list);
+	unsigned int v, send, stotal;
+	uint64_t offset, lastb;
+	struct zone_split_index *zsi;
 
-	return r1->off - r2->off;
+	lastb = last_block(td, f, ddir);
+	if (!lastb)
+		return 1;
+
+	if (!td->o.zone_split_nr[ddir]) {
+bail:
+		return __get_next_rand_offset(td, f, ddir, b, lastb);
+	}
+
+	/*
+	 * Generate a value, v, between 1 and 100, both inclusive
+	 */
+	v = rand_between(&td->zone_state, 1, 100);
+
+	zsi = &td->zone_state_index[ddir][v - 1];
+	stotal = zsi->size_perc_prev;
+	send = zsi->size_perc;
+
+	/*
+	 * Should never happen
+	 */
+	if (send == -1U) {
+		if (!fio_did_warn(FIO_WARN_ZONED_BUG))
+			log_err("fio: bug in zoned generation\n");
+		goto bail;
+	}
+
+	/*
+	 * 'send' is some percentage below or equal to 100 that
+	 * marks the end of the current IO range. 'stotal' marks
+	 * the start, in percent.
+	 */
+	if (stotal)
+		offset = stotal * lastb / 100ULL;
+	else
+		offset = 0;
+
+	lastb = lastb * (send - stotal) / 100ULL;
+
+	/*
+	 * Generate index from 0..send-of-lastb
+	 */
+	if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
+		return 1;
+
+	/*
+	 * Add our start offset, if any
+	 */
+	if (offset)
+		*b += offset;
+
+	return 0;
 }
 
-static int get_off_from_method(struct thread_data *td, struct fio_file *f,
-			       enum fio_ddir ddir, uint64_t *b)
+static int get_next_rand_offset(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)
+	if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
+		uint64_t lastb;
+
+		lastb = last_block(td, f, ddir);
+		if (!lastb)
+			return 1;
+
+		return __get_next_rand_offset(td, f, ddir, b, lastb);
+	} 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);
 	else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
 		return __get_next_rand_offset_gauss(td, f, ddir, b);
+	else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
+		return __get_next_rand_offset_zoned(td, f, ddir, b);
+	else if (td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
+		return __get_next_rand_offset_zoned_abs(td, f, ddir, b);
 
 	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
 	return 1;
 }
 
-/*
- * Sort the reads for a verify phase in batches of verifysort_nr, if
- * specified.
- */
-static inline int should_sort_io(struct thread_data *td)
-{
-	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;
-
-	return 1;
-}
-
-static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
+static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
 {
 	unsigned int v;
-	unsigned long r;
 
 	if (td->o.perc_rand[ddir] == 100)
-		return 1;
+		return true;
 
-	r = __rand(&td->seq_rand_state[ddir]);
-	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+	v = rand_between(&td->seq_rand_state[ddir], 1, 100);
 
 	return v <= td->o.perc_rand[ddir];
 }
 
-static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
-				enum fio_ddir ddir, uint64_t *b)
+static void loop_cache_invalidate(struct thread_data *td, struct fio_file *f)
 {
-	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)) {
-fetch:
-		r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
-		flist_del(&r->list);
-		*b = r->off;
-		free(r);
-		return 0;
-	}
-
-	for (i = 0; i < td->o.verifysort_nr; i++) {
-		r = malloc(sizeof(*r));
+	struct thread_options *o = &td->o;
 
-		ret = get_off_from_method(td, f, ddir, &r->off);
-		if (ret) {
-			free(r);
-			break;
-		}
+	if (o->invalidate_cache && !o->odirect) {
+		int fio_unused ret;
 
-		flist_add(&r->list, &td->next_rand_list);
+		ret = file_invalidate_cache(td, f);
 	}
-
-	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,
@@ -258,8 +329,10 @@ static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
 	if (!get_next_rand_offset(td, f, ddir, b))
 		return 0;
 
-	if (td->o.time_based) {
+	if (td->o.time_based ||
+	    (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
 		fio_file_reset(td, f);
+		loop_cache_invalidate(td, f);
 		if (!get_next_rand_offset(td, f, ddir, b))
 			return 0;
 	}
@@ -277,15 +350,29 @@ static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
 
 	assert(ddir_rw(ddir));
 
+	/*
+	 * If we reach the end for a time based run, reset us back to 0
+	 * and invalidate the cache, if we need to.
+	 */
 	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;
+	    o->time_based) {
+		f->last_pos[ddir] = f->file_offset;
+		loop_cache_invalidate(td, f);
+	}
 
 	if (f->last_pos[ddir] < f->real_file_size) {
 		uint64_t pos;
 
-		if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0)
-			f->last_pos[ddir] = f->real_file_size;
+		/*
+		 * Only rewind if we already hit the end
+		 */
+		if (f->last_pos[ddir] == f->file_offset &&
+		    f->file_offset && o->ddir_seq_add < 0) {
+			if (f->real_file_size > f->io_size)
+				f->last_pos[ddir] = f->io_size;
+			else
+				f->last_pos[ddir] = f->real_file_size;
+		}
 
 		pos = f->last_pos[ddir] - f->file_offset;
 		if (pos && o->ddir_seq_add) {
@@ -294,10 +381,21 @@ static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
 			/*
 			 * If we reach beyond the end of the file
 			 * with holed IO, wrap around to the
-			 * beginning again.
+			 * beginning again. If we're doing backwards IO,
+			 * wrap to the end.
 			 */
-			if (pos >= f->real_file_size)
-				pos = f->file_offset;
+			if (pos >= f->real_file_size) {
+				if (o->ddir_seq_add > 0)
+					pos = f->file_offset;
+				else {
+					if (f->real_file_size > f->io_size)
+						pos = f->io_size;
+					else
+						pos = f->real_file_size;
+
+					pos += o->ddir_seq_add;
+				}
+			}
 		}
 
 		*offset = pos;
@@ -309,7 +407,7 @@ 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,
-			  unsigned int *is_random)
+			  bool *is_random)
 {
 	struct fio_file *f = io_u->file;
 	uint64_t b, offset;
@@ -323,27 +421,27 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 		if (td_random(td)) {
 			if (should_do_random(td, ddir)) {
 				ret = get_next_rand_block(td, f, ddir, &b);
-				*is_random = 1;
+				*is_random = true;
 			} else {
-				*is_random = 0;
-				io_u_set(io_u, IO_U_F_BUSY_OK);
+				*is_random = false;
+				io_u_set(td, io_u, 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;
+			*is_random = false;
 			ret = get_next_seq_offset(td, f, ddir, &offset);
 		}
 	} else {
-		io_u_set(io_u, IO_U_F_BUSY_OK);
-		*is_random = 0;
+		io_u_set(td, io_u, IO_U_F_BUSY_OK);
+		*is_random = false;
 
 		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;
+				*is_random = false;
 			}
 		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
 			if (f->last_start[ddir] != -1ULL)
@@ -376,8 +474,8 @@ 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,
-			     unsigned int *is_random)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u,
+			   bool *is_random)
 {
 	struct fio_file *f = io_u->file;
 	enum fio_ddir ddir = io_u->ddir;
@@ -411,39 +509,27 @@ 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,
-			   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, 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)
+static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
+			     unsigned long long 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)
+static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *io_u,
+				    bool is_random)
 {
 	int ddir = io_u->ddir;
-	unsigned int buflen = 0;
-	unsigned int minbs, maxbs;
-	unsigned long r;
+	unsigned long long buflen = 0;
+	unsigned long long minbs, maxbs;
+	uint64_t frand_max, r;
+	bool power_2;
 
 	assert(ddir_rw(ddir));
 
 	if (td->o.bs_is_seq_rand)
-		ddir = is_random ? DDIR_WRITE: DDIR_READ;
+		ddir = is_random ? DDIR_WRITE : DDIR_READ;
 
 	minbs = td->o.min_bs[ddir];
 	maxbs = td->o.max_bs[ddir];
@@ -457,16 +543,15 @@ 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;
 
+	frand_max = rand_max(&td->bsrange_state[ddir]);
 	do {
-		r = __rand(&td->bsrange_state);
+		r = __rand(&td->bsrange_state[ddir]);
 
 		if (!td->o.bssplit_nr[ddir]) {
-			buflen = 1 + (unsigned int) ((double) maxbs *
-					(r / (FRAND_MAX + 1.0)));
-			if (buflen < minbs)
-				buflen = minbs;
+			buflen = minbs + (unsigned long long) ((double) maxbs *
+					(r / (frand_max + 1.0)));
 		} else {
-			long perc = 0;
+			long long perc = 0;
 			unsigned int i;
 
 			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
@@ -474,37 +559,24 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
 
 				buflen = bsp->bs;
 				perc += bsp->perc;
-				if ((r <= ((FRAND_MAX / 100L) * perc)) &&
+				if (!perc)
+					break;
+				if ((r / perc <= frand_max / 100ULL) &&
 				    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);
-
+		power_2 = is_power_of_2(minbs);
+		if (!td->o.bs_unaligned && power_2)
+			buflen &= ~(minbs - 1);
+		else if (!td->o.bs_unaligned && !power_2) 
+			buflen -= buflen % minbs; 
 	} 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 int diff;
@@ -521,10 +593,8 @@ static void set_rwmix_bytes(struct thread_data *td)
 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 {
 	unsigned int v;
-	unsigned long r;
 
-	r = __rand(&td->rwmix_state);
-	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+	v = rand_between(&td->rwmix_state, 1, 100);
 
 	if (v <= td->o.rwmix[DDIR_READ])
 		return DDIR_READ;
@@ -532,8 +602,10 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 	return DDIR_WRITE;
 }
 
-void io_u_quiesce(struct thread_data *td)
+int io_u_quiesce(struct thread_data *td)
 {
+	int completed = 0;
+
 	/*
 	 * We are going to sleep, ensure that we flush anything pending as
 	 * not to skew our latency numbers.
@@ -543,65 +615,67 @@ 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);
-	}
+	if (td->io_u_queued || td->cur_depth)
+		td_io_commit(td);
 
 	while (td->io_u_in_flight) {
-		int fio_unused ret;
+		int ret;
 
 		ret = io_u_queued_complete(td, 1);
+		if (ret > 0)
+			completed += ret;
 	}
+
+	if (td->flags & TD_F_REGROW_LOGS)
+		regrow_logs(td);
+
+	return completed;
 }
 
 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 {
 	enum fio_ddir odir = ddir ^ 1;
-	long usec;
+	uint64_t usec;
+	uint64_t now;
 
 	assert(ddir_rw(ddir));
+	now = utime_since_now(&td->start);
 
-	if (td->rate_pending_usleep[ddir] <= 0)
+	/*
+	 * if rate_next_io_time is in the past, need to catch up to rate
+	 */
+	if (td->rate_next_io_time[ddir] <= now)
 		return ddir;
 
 	/*
-	 * We have too much pending sleep in this direction. See if we
+	 * We are ahead of rate 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
+		 * Other direction is behind rate, switch
 		 */
-		if (td->rate_pending_usleep[odir] < 100000)
+		if (td->rate_next_io_time[odir] <= now)
 			return odir;
 
 		/*
-		 * Both directions have pending sleep. Sleep the minimum time
-		 * and deduct from both.
+		 * Both directions are ahead of rate. sleep the min,
+		 * switch if necessary
 		 */
-		if (td->rate_pending_usleep[ddir] <=
-			td->rate_pending_usleep[odir]) {
-			usec = td->rate_pending_usleep[ddir];
+		if (td->rate_next_io_time[ddir] <=
+		    td->rate_next_io_time[odir]) {
+			usec = td->rate_next_io_time[ddir] - now;
 		} else {
-			usec = td->rate_pending_usleep[odir];
+			usec = td->rate_next_io_time[odir] - now;
 			ddir = odir;
 		}
 	} else
-		usec = td->rate_pending_usleep[ddir];
+		usec = td->rate_next_io_time[ddir] - now;
 
 	if (td->o.io_submit_mode == IO_MODE_INLINE)
 		io_u_quiesce(td);
 
-	usec = usec_sleep(td, usec);
-
-	td->rate_pending_usleep[ddir] -= usec;
-
-	odir = ddir ^ 1;
-	if (td_rw(td) && __should_check_rate(td, odir))
-		td->rate_pending_usleep[odir] -= usec;
-
+	usec_sleep(td, usec);
 	return ddir;
 }
 
@@ -615,28 +689,22 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 	enum fio_ddir ddir;
 
 	/*
-	 * 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
+	 * See if it's time to fsync/fdatasync/sync_file_range first,
+	 * and if not then move on to check regular I/Os.
 	 */
-	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 (should_fsync(td)) {
+		if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
+		    !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
+			return DDIR_SYNC;
+
+		if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
+		    !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
+			return DDIR_DATASYNC;
+
+		if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
+		    !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
+			return DDIR_SYNC_FILE_RANGE;
+	}
 
 	if (td_rw(td)) {
 		/*
@@ -660,8 +728,10 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 		ddir = DDIR_READ;
 	else if (td_write(td))
 		ddir = DDIR_WRITE;
-	else
+	else if (td_trim(td))
 		ddir = DDIR_TRIM;
+	else
+		ddir = DDIR_INVAL;
 
 	td->rwmix_ddir = rate_ddir(td, ddir);
 	return td->rwmix_ddir;
@@ -681,11 +751,11 @@ static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
 
 	io_u->ddir = io_u->acct_ddir = ddir;
 
-	if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
+	if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
 	    td->o.barrier_blocks &&
 	   !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
 	     td->io_issues[DDIR_WRITE])
-		io_u_set(io_u, IO_U_F_BARRIER);
+		io_u_set(td, io_u, IO_U_F_BARRIER);
 }
 
 void put_file_log(struct thread_data *td, struct fio_file *f)
@@ -698,34 +768,45 @@ void put_file_log(struct thread_data *td, struct fio_file *f)
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
+	const bool needs_lock = td_async_processing(td);
+
+	if (io_u->post_submit) {
+		io_u->post_submit(io_u, io_u->error == 0);
+		io_u->post_submit = NULL;
+	}
+
 	if (td->parent)
 		td = td->parent;
 
-	td_io_u_lock(td);
+	if (needs_lock)
+		__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;
-	io_u_set(io_u, IO_U_F_FREE);
+	io_u_set(td, io_u, IO_U_F_FREE);
 
 	if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
 		td->cur_depth--;
 		assert(!(td->flags & TD_F_CHILD));
 	}
 	io_u_qpush(&td->io_u_freelist, io_u);
-	td_io_u_unlock(td);
 	td_io_u_free_notify(td);
+
+	if (needs_lock)
+		__td_io_u_unlock(td);
 }
 
 void clear_io_u(struct thread_data *td, struct io_u *io_u)
 {
-	io_u_clear(io_u, IO_U_F_FLIGHT);
+	io_u_clear(td, io_u, IO_U_F_FLIGHT);
 	put_io_u(td, io_u);
 }
 
 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 {
+	const bool needs_lock = td_async_processing(td);
 	struct io_u *__io_u = *io_u;
 	enum fio_ddir ddir = acct_ddir(__io_u);
 
@@ -734,45 +815,40 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 	if (td->parent)
 		td = td->parent;
 
-	td_io_u_lock(td);
+	if (needs_lock)
+		__td_io_u_lock(td);
 
-	io_u_set(__io_u, IO_U_F_FREE);
+	io_u_set(td, __io_u, IO_U_F_FREE);
 	if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
 		td->io_issues[ddir]--;
 
-	io_u_clear(__io_u, IO_U_F_FLIGHT);
+	io_u_clear(td, __io_u, IO_U_F_FLIGHT);
 	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
 		td->cur_depth--;
 		assert(!(td->flags & TD_F_CHILD));
 	}
 
 	io_u_rpush(&td->io_u_requeues, __io_u);
-	td_io_u_unlock(td);
 	td_io_u_free_notify(td);
+
+	if (needs_lock)
+		__td_io_u_unlock(td);
+
 	*io_u = NULL;
 }
 
-static int fill_io_u(struct thread_data *td, struct io_u *io_u)
+static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
 {
-	unsigned int is_random;
-
-	if (td->io_ops->flags & FIO_NOIO)
-		goto out;
+	struct fio_file *f = io_u->file;
 
-	set_rw_ddir(td, io_u);
-
-	/*
-	 * fsync() or fdatasync() or trim etc, we are done
-	 */
-	if (!ddir_rw(io_u->ddir))
-		goto out;
+	assert(td->o.zone_mode == ZONE_MODE_STRIDED);
+	assert(td->o.zone_size);
+	assert(td->o.zone_range);
 
 	/*
 	 * 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;
 		f->file_offset += td->o.zone_range + td->o.zone_skip;
 
@@ -780,11 +856,50 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 		 * 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->file_offset = get_start_offset(td, f);
+
 		f->last_pos[io_u->ddir] = f->file_offset;
 		td->io_skip_bytes += td->o.zone_skip;
 	}
 
+	/*
+	 * If zone_size > zone_range, then maintain the same zone until
+	 * zone_bytes >= zone_size.
+	 */
+	if (f->last_pos[io_u->ddir] >= (f->file_offset + td->o.zone_range)) {
+		dprint(FD_IO, "io_u maintain zone offset=%" PRIu64 "/last_pos=%" PRIu64 "\n",
+				f->file_offset, f->last_pos[io_u->ddir]);
+		f->last_pos[io_u->ddir] = f->file_offset;
+	}
+
+	/*
+	 * For random: if 'norandommap' is not set and zone_size > zone_range,
+	 * map needs to be reset as it's done with zone_range everytime.
+	 */
+	if ((td->zone_bytes % td->o.zone_range) == 0)
+		fio_file_reset(td, f);
+}
+
+static int fill_io_u(struct thread_data *td, struct io_u *io_u)
+{
+	bool is_random;
+	uint64_t offset;
+	enum io_u_action ret;
+
+	if (td_ioengine_flagged(td, FIO_NOIO))
+		goto out;
+
+	set_rw_ddir(td, io_u);
+
+	/*
+	 * fsync() or fdatasync() or trim etc, we are done
+	 */
+	if (!ddir_rw(io_u->ddir))
+		goto out;
+
+	if (td->o.zone_mode == ZONE_MODE_STRIDED)
+		setup_strided_zone_mode(td, io_u);
+
 	/*
 	 * No log, let the seq/rand engine retrieve the next buflen and
 	 * position.
@@ -800,9 +915,16 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 		return 1;
 	}
 
+	offset = io_u->offset;
+	if (td->o.zone_mode == ZONE_MODE_ZBD) {
+		ret = zbd_adjust_block(td, io_u);
+		if (ret == io_u_eof)
+			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",
+		dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
+			io_u,
 			(unsigned long long) io_u->offset, io_u->buflen,
 			(unsigned long long) io_u->file->real_file_size);
 		return 1;
@@ -812,15 +934,15 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	 * mark entry before potentially trimming io_u
 	 */
 	if (td_random(td) && file_randommap(td, io_u->file))
-		mark_random_map(td, io_u);
+		io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
 
 out:
-	dprint_io_u(io_u, "fill_io_u");
+	dprint_io_u(io_u, "fill");
 	td->zone_bytes += io_u->buflen;
 	return 0;
 }
 
-static void __io_u_mark_map(unsigned int *map, unsigned int nr)
+static void __io_u_mark_map(uint64_t *map, unsigned int nr)
 {
 	int idx = 0;
 
@@ -890,11 +1012,52 @@ void io_u_mark_depth(struct thread_data *td, unsigned int nr)
 	td->ts.io_u_map[idx] += nr;
 }
 
-static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
+static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
 {
 	int idx = 0;
 
-	assert(usec < 1000);
+	assert(nsec < 1000);
+
+	switch (nsec) {
+	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_N_NR);
+	td->ts.io_u_lat_n[idx]++;
+}
+
+static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
+{
+	int idx = 0;
+
+	assert(usec < 1000 && usec >= 1);
 
 	switch (usec) {
 	case 750 ... 999:
@@ -931,10 +1094,12 @@ static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
 	td->ts.io_u_lat_u[idx]++;
 }
 
-static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
+static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
 {
 	int idx = 0;
 
+	assert(msec >= 1);
+
 	switch (msec) {
 	default:
 		idx = 11;
@@ -976,12 +1141,42 @@ static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
 	td->ts.io_u_lat_m[idx]++;
 }
 
-static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
+static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
 {
-	if (usec < 1000)
-		io_u_mark_lat_usec(td, usec);
+	if (nsec < 1000)
+		io_u_mark_lat_nsec(td, nsec);
+	else if (nsec < 1000000)
+		io_u_mark_lat_usec(td, nsec / 1000);
 	else
-		io_u_mark_lat_msec(td, usec / 1000);
+		io_u_mark_lat_msec(td, nsec / 1000000);
+}
+
+static unsigned int __get_next_fileno_rand(struct thread_data *td)
+{
+	unsigned long fileno;
+
+	if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
+		uint64_t frand_max = rand_max(&td->next_file_state);
+		unsigned long r;
+
+		r = __rand(&td->next_file_state);
+		return (unsigned int) ((double) td->o.nr_files
+				* (r / (frand_max + 1.0)));
+	}
+
+	if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
+		fileno = zipf_next(&td->next_file_zipf);
+	else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
+		fileno = pareto_next(&td->next_file_zipf);
+	else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
+		fileno = gauss_next(&td->next_file_gauss);
+	else {
+		log_err("fio: bad file service type: %d\n", td->o.file_service_type);
+		assert(0);
+		return 0;
+	}
+
+	return fileno >> FIO_FSERVICE_SHIFT;
 }
 
 /*
@@ -996,11 +1191,8 @@ static struct fio_file *get_next_file_rand(struct thread_data *td,
 
 	do {
 		int opened = 0;
-		unsigned long r;
 
-		r = __rand(&td->next_file_state);
-		fno = (unsigned int) ((double) td->o.nr_files
-				* (r / (FRAND_MAX + 1.0)));
+		fno = __get_next_fileno_rand(td);
 
 		f = td->files[fno];
 		if (fio_file_done(f))
@@ -1125,13 +1317,6 @@ out:
 
 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);
 }
 
@@ -1150,23 +1335,32 @@ static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
 		if (!fill_io_u(td, io_u))
 			break;
 
+		if (io_u->post_submit) {
+			io_u->post_submit(io_u, false);
+			io_u->post_submit = NULL;
+		}
+
 		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,
+		if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
+			fio_file_reset(td, f);
+		else {
+			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;
 }
 
 static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
-		      unsigned long tusec, unsigned long max_usec)
+		      unsigned long long tnsec, unsigned long long max_nsec)
 {
 	if (!td->error)
-		log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
+		log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
 	td_verror(td, ETIMEDOUT, "max latency exceeded");
 	icd->error = ETIMEDOUT;
 }
@@ -1182,10 +1376,10 @@ static void lat_new_cycle(struct thread_data *td)
  * 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)
+static bool __lat_target_failed(struct thread_data *td)
 {
 	if (td->latency_qd == 1)
-		return 1;
+		return true;
 
 	td->latency_qd_high = td->latency_qd;
 
@@ -1202,16 +1396,16 @@ static int __lat_target_failed(struct thread_data *td)
 	 */
 	io_u_quiesce(td);
 	lat_new_cycle(td);
-	return 0;
+	return false;
 }
 
-static int lat_target_failed(struct thread_data *td)
+static bool 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;
+	return false;
 }
 
 void lat_target_init(struct thread_data *td)
@@ -1306,26 +1500,29 @@ void lat_target_check(struct thread_data *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)
+bool queue_full(const struct thread_data *td)
 {
 	const int qempty = io_u_qempty(&td->io_u_freelist);
 
 	if (qempty)
-		return 1;
+		return true;
 	if (!td->o.latency_target)
-		return 0;
+		return false;
 
 	return td->cur_depth >= td->latency_qd;
 }
 
 struct io_u *__get_io_u(struct thread_data *td)
 {
+	const bool needs_lock = td_async_processing(td);
 	struct io_u *io_u = NULL;
+	int ret;
 
 	if (td->stop_io)
 		return NULL;
 
-	td_io_u_lock(td);
+	if (needs_lock)
+		__td_io_u_lock(td);
 
 again:
 	if (!io_u_rempty(&td->io_u_requeues))
@@ -1341,7 +1538,7 @@ again:
 
 	if (io_u) {
 		assert(io_u->flags & IO_U_F_FREE);
-		io_u_clear(io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
+		io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
 				 IO_U_F_TRIMMED | IO_U_F_BARRIER |
 				 IO_U_F_VER_LIST);
 
@@ -1349,7 +1546,7 @@ again:
 		io_u->acct_ddir = -1;
 		td->cur_depth++;
 		assert(!(td->flags & TD_F_CHILD));
-		io_u_set(io_u, IO_U_F_IN_CUR_DEPTH);
+		io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
 		io_u->ipo = NULL;
 	} else if (td_async_processing(td)) {
 		/*
@@ -1357,44 +1554,44 @@ again:
 		 * return one
 		 */
 		assert(!(td->flags & TD_F_CHILD));
-		assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
+		ret = pthread_cond_wait(&td->free_cond, &td->io_u_lock);
+		assert(ret == 0);
 		goto again;
 	}
 
-	td_io_u_unlock(td);
+	if (needs_lock)
+		__td_io_u_unlock(td);
+
 	return io_u;
 }
 
-static int check_get_trim(struct thread_data *td, struct io_u *io_u)
+static bool 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 false;
+	if (!td->trim_entries)
+		return false;
+
+	if (td->trim_batch) {
+		td->trim_batch--;
+		if (get_next_trim(td, io_u))
+			return true;
+	} 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;
+		if (get_next_trim(td, io_u))
+			return true;
 	}
 
-	return 0;
+	return false;
 }
 
-static int check_get_verify(struct thread_data *td, struct io_u *io_u)
+static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
 {
 	if (!(td->flags & TD_F_VER_BACKLOG))
-		return 0;
+		return false;
 
 	if (td->io_hist_len) {
 		int get_verify = 0;
@@ -1411,11 +1608,11 @@ static int check_get_verify(struct thread_data *td, struct io_u *io_u)
 
 		if (get_verify && !get_next_verify(td, io_u)) {
 			td->verify_batch--;
-			return 1;
+			return true;
 		}
 	}
 
-	return 0;
+	return false;
 }
 
 /*
@@ -1426,32 +1623,40 @@ static int check_get_verify(struct thread_data *td, struct io_u *io_u)
  */
 static void small_content_scramble(struct io_u *io_u)
 {
-	unsigned int i, nr_blocks = io_u->buflen / 512;
-	uint64_t boffset;
+	unsigned long long i, nr_blocks = io_u->buflen >> 9;
 	unsigned int offset;
-	void *p, *end;
+	uint64_t boffset, *iptr;
+	char *p;
 
 	if (!nr_blocks)
 		return;
 
 	p = io_u->xfer_buf;
 	boffset = io_u->offset;
-	io_u->buf_filled_len = 0;
+
+	if (io_u->buf_filled_len)
+		io_u->buf_filled_len = 0;
+
+	/*
+	 * Generate random index between 0..7. We do chunks of 512b, if
+	 * we assume a cacheline is 64 bytes, then we have 8 of those.
+	 * Scramble content within the blocks in the same cacheline to
+	 * speed things up.
+	 */
+	offset = (io_u->start_time.tv_nsec ^ boffset) & 7;
 
 	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.
+		 * Fill offset into start of cacheline, time into end
+		 * of cacheline
 		 */
-		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));
+		iptr = (void *) p + (offset << 6);
+		*iptr = boffset;
+
+		iptr = (void *) p + 64 - 2 * sizeof(uint64_t);
+		iptr[0] = io_u->start_time.tv_sec;
+		iptr[1] = io_u->start_time.tv_nsec;
+
 		p += 512;
 		boffset += 512;
 	}
@@ -1459,7 +1664,7 @@ static void small_content_scramble(struct io_u *io_u)
 
 /*
  * 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.
+ * etc. The returned io_u is fully ready to be prepped, populated and submitted.
  */
 struct io_u *get_io_u(struct thread_data *td)
 {
@@ -1506,7 +1711,7 @@ struct io_u *get_io_u(struct thread_data *td)
 	assert(fio_file_open(f));
 
 	if (ddir_rw(io_u->ddir)) {
-		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
+		if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
 			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
 			goto err_put;
 		}
@@ -1518,14 +1723,11 @@ struct io_u *get_io_u(struct thread_data *td)
 			if (td->flags & TD_F_REFILL_BUFFERS) {
 				io_u_fill_buffer(td, io_u,
 					td->o.min_bs[DDIR_WRITE],
-					io_u->xfer_buflen);
+					io_u->buflen);
 			} else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
-				   !(td->flags & TD_F_COMPRESS))
+				   !(td->flags & TD_F_COMPRESS) &&
+				   !(td->flags & TD_F_DO_VERIFY))
 				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
@@ -1544,10 +1746,12 @@ struct io_u *get_io_u(struct thread_data *td)
 out:
 	assert(io_u->file);
 	if (!td_io_prep(td, io_u)) {
-		if (!td->o.disable_slat)
+		if (!td->o.disable_lat)
 			fio_gettime(&io_u->start_time, NULL);
+
 		if (do_scramble)
 			small_content_scramble(io_u);
+
 		return io_u;
 	}
 err_put:
@@ -1563,13 +1767,20 @@ static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
 	if (td_non_fatal_error(td, eb, io_u->error) && !td->o.error_dump)
 		return;
 
-	log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%lu\n",
+	log_err("fio: io_u error%s%s: %s: %s offset=%llu, buflen=%llu\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->io_ops->errdetails) {
+		char *err = td->io_ops->errdetails(io_u);
+
+		log_err("fio: %s\n", err);
+		free(err);
+	}
+
 	if (!td->error)
 		td_verror(td, io_u->error, "io_u error");
 }
@@ -1578,13 +1789,13 @@ void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 {
 	__io_u_log_error(td, io_u);
 	if (td->parent)
-		__io_u_log_error(td, io_u);
+		__io_u_log_error(td->parent, io_u);
 }
 
-static inline int gtod_reduce(struct thread_data *td)
+static inline bool gtod_reduce(struct thread_data *td)
 {
-	return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
-		&& td->o.disable_bw;
+	return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
+			|| td->o.gtod_reduce;
 }
 
 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
@@ -1592,45 +1803,51 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 				  const enum fio_ddir idx, unsigned int bytes)
 {
 	const int no_reduce = !gtod_reduce(td);
-	unsigned long lusec = 0;
+	unsigned long long llnsec = 0;
+
+	if (td->parent)
+		td = td->parent;
+
+	if (!td->o.stats || td_ioengine_flagged(td, FIO_NOSTATS))
+		return;
 
 	if (no_reduce)
-		lusec = utime_since(&io_u->issue_time, &icd->time);
+		llnsec = ntime_since(&io_u->issue_time, &icd->time);
 
 	if (!td->o.disable_lat) {
-		unsigned long tusec;
+		unsigned long long tnsec;
 
-		tusec = utime_since(&io_u->start_time, &icd->time);
-		add_lat_sample(td, idx, tusec, bytes, io_u->offset);
+		tnsec = ntime_since(&io_u->start_time, &icd->time);
+		add_lat_sample(td, idx, tnsec, 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);
+				icd->error = ops->io_u_lat(td, tnsec);
 		}
 
-		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 (td->o.max_latency && tnsec > td->o.max_latency)
+			lat_fatal(td, icd, tnsec, td->o.max_latency);
+		if (td->o.latency_target && tnsec > td->o.latency_target) {
 			if (lat_target_failed(td))
-				lat_fatal(td, icd, tusec, td->o.latency_target);
+				lat_fatal(td, icd, tnsec, 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->parent)
-		td = td->parent;
+	if (ddir_rw(idx)) {
+		if (!td->o.disable_clat) {
+			add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
+			io_u_mark_latency(td, llnsec);
+		}
 
-	if (!td->o.disable_bw)
-		add_bw_sample(td, idx, bytes, &icd->time);
+		if (!td->o.disable_bw && per_unit_log(td->bw_log))
+			add_bw_sample(td, io_u, bytes, llnsec);
 
-	if (no_reduce)
-		add_iops_sample(td, idx, bytes, &icd->time);
+		if (no_reduce && per_unit_log(td->iops_log))
+			add_iops_sample(td, io_u, bytes);
+	} else if (ddir_sync(idx) && !td->o.disable_clat)
+		add_sync_clat_sample(&td->ts, llnsec);
 
 	if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
 		uint32_t *info = io_u_block_info(td, io_u);
@@ -1646,16 +1863,32 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 	}
 }
 
-static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
+static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
+				uint64_t offset, unsigned int bytes)
 {
-	uint64_t secs, remainder, bps, bytes;
+	int idx;
 
-	assert(!(td->flags & TD_F_CHILD));
-	bytes = td->this_io_bytes[ddir];
-	bps = td->rate_bps[ddir];
-	secs = bytes / bps;
-	remainder = bytes % bps;
-	return remainder * 1000000 / bps + secs * 1000000;
+	if (!f)
+		return;
+
+	if (f->first_write == -1ULL || offset < f->first_write)
+		f->first_write = offset;
+	if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
+		f->last_write = offset + bytes;
+
+	if (!f->last_write_comp)
+		return;
+
+	idx = f->last_write_idx++;
+	f->last_write_comp[idx] = offset;
+	if (f->last_write_idx == td->o.iodepth)
+		f->last_write_idx = 0;
+}
+
+static bool should_account(struct thread_data *td)
+{
+	return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
+					   td->runstate == TD_VERIFYING);
 }
 
 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
@@ -1665,10 +1898,10 @@ static void io_completed(struct thread_data *td, struct 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");
+	dprint_io_u(io_u, "complete");
 
 	assert(io_u->flags & IO_U_F_FLIGHT);
-	io_u_clear(io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
+	io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
 
 	/*
 	 * Mark IO ok to verify
@@ -1686,69 +1919,37 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 	}
 
 	if (ddir_sync(ddir)) {
-		td->last_was_sync = 1;
+		td->last_was_sync = true;
 		if (f) {
 			f->first_write = -1ULL;
 			f->last_write = -1ULL;
 		}
+		if (should_account(td))
+			account_io_completion(td, io_u, icd, ddir, io_u->buflen);
 		return;
 	}
 
-	td->last_was_sync = 0;
+	td->last_was_sync = false;
 	td->last_ddir = ddir;
 
 	if (!io_u->error && ddir_rw(ddir)) {
-		unsigned int bytes = io_u->buflen - io_u->resid;
-		const enum fio_ddir oddir = ddir ^ 1;
+		unsigned long long bytes = io_u->buflen - io_u->resid;
 		int ret;
 
 		td->io_blocks[ddir]++;
-		td->this_io_blocks[ddir]++;
 		td->io_bytes[ddir] += bytes;
 
-		if (!(io_u->flags & IO_U_F_VER_LIST))
+		if (!(io_u->flags & IO_U_F_VER_LIST)) {
+			td->this_io_blocks[ddir]++;
 			td->this_io_bytes[ddir] += bytes;
-
-		if (ddir == DDIR_WRITE) {
-			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;
-			}
-			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)) {
-			struct thread_data *__td = td;
+		if (ddir == DDIR_WRITE)
+			file_log_write_comp(td, f, io_u->offset, bytes);
 
+		if (should_account(td))
 			account_io_completion(td, io_u, icd, ddir, bytes);
 
-			if (td->parent)
-				__td = td->parent;
-
-			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[ddir] += bytes;
 
 		if (io_u->end_io) {
@@ -1790,7 +1991,7 @@ static void init_icd(struct thread_data *td, struct io_completion_data *icd,
 	icd->nr = nr;
 
 	icd->error = 0;
-	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
 		icd->bytes_done[ddir] = 0;
 }
 
@@ -1829,7 +2030,7 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 		return -1;
 	}
 
-	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
 		td->bytes_done[ddir] += icd.bytes_done[ddir];
 
 	return 0;
@@ -1845,14 +2046,16 @@ int io_u_queued_complete(struct thread_data *td, int min_evts)
 	int ret, ddir;
 	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
+	dprint(FD_IO, "io_u_queued_complete: 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_evts, td->o.iodepth_batch_complete, tvp);
+	/* No worries, td_io_getevents fixes min and max if they are
+	 * set incorrectly */
+	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
 	if (ret < 0) {
 		td_verror(td, -ret, "td_io_getevents");
 		return ret;
@@ -1866,10 +2069,10 @@ int io_u_queued_complete(struct thread_data *td, int min_evts)
 		return -1;
 	}
 
-	for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
 		td->bytes_done[ddir] += icd.bytes_done[ddir];
 
-	return 0;
+	return ret;
 }
 
 /*
@@ -1877,10 +2080,14 @@ int io_u_queued_complete(struct thread_data *td, int min_evts)
  */
 void io_u_queued(struct thread_data *td, struct io_u *io_u)
 {
-	if (!td->o.disable_slat) {
+	if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
 		unsigned long slat_time;
 
-		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
+		slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
+
+		if (td->parent)
+			td = td->parent;
+
 		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
 				io_u->offset);
 	}
@@ -1892,15 +2099,15 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 static struct frand_state *get_buf_state(struct thread_data *td)
 {
 	unsigned int v;
-	unsigned long r;
 
 	if (!td->o.dedupe_percentage)
 		return &td->buf_state;
-	else if (td->o.dedupe_percentage == 100)
-		return &td->buf_state_prev;
+	else if (td->o.dedupe_percentage == 100) {
+		frand_copy(&td->buf_state_prev, &td->buf_state);
+		return &td->buf_state;
+	}
 
-	r = __rand(&td->dedupe_state);
-	v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+	v = rand_between(&td->dedupe_state, 1, 100);
 
 	if (v <= td->o.dedupe_percentage)
 		return &td->buf_state_prev;
@@ -1910,19 +2117,25 @@ static struct frand_state *get_buf_state(struct thread_data *td)
 
 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
 {
-	if (rs == &td->buf_state)
+	if (td->o.dedupe_percentage == 100)
+		frand_copy(rs, &td->buf_state_prev);
+	else 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)
+void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_write,
+		    unsigned long long max_bs)
 {
 	struct thread_options *o = &td->o;
 
+	if (o->mem_type == MEM_CUDA_MALLOC)
+		return;
+
 	if (o->compress_percentage || o->dedupe_percentage) {
 		unsigned int perc = td->o.compress_percentage;
 		struct frand_state *rs;
-		unsigned int left = max_bs;
+		unsigned long long left = max_bs;
+		unsigned long long this_write;
 
 		do {
 			rs = get_buf_state(td);
@@ -1930,20 +2143,20 @@ void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
 			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;
+				this_write = min_not_zero(min_write,
+							(unsigned long long) td->o.compress_chunk);
 
-				fill_random_buf_percentage(rs, buf, perc, seg,
-					min_write, o->buffer_pattern,
-						   o->buffer_pattern_bytes);
-			} else
+				fill_random_buf_percentage(rs, buf, perc,
+					this_write, this_write,
+					o->buffer_pattern,
+					o->buffer_pattern_bytes);
+			} else {
 				fill_random_buf(rs, buf, min_write);
+				this_write = min_write;
+			}
 
-			buf += min_write;
-			left -= min_write;
+			buf += this_write;
+			left -= this_write;
 			save_buf_state(td, rs);
 		} while (left);
 	} else if (o->buffer_pattern_bytes)
@@ -1958,8 +2171,66 @@ void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
  * "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)
+		      unsigned long long min_write, unsigned long long max_bs)
 {
 	io_u->buf_filled_len = 0;
 	fill_io_buffer(td, io_u->buf, min_write, max_bs);
 }
+
+static int do_sync_file_range(const struct thread_data *td,
+			      struct fio_file *f)
+{
+	off64_t offset, nbytes;
+
+	offset = f->first_write;
+	nbytes = f->last_write - f->first_write;
+
+	if (!nbytes)
+		return 0;
+
+	return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
+}
+
+int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
+{
+	int ret;
+
+	if (io_u->ddir == DDIR_SYNC) {
+		ret = fsync(io_u->file->fd);
+	} else if (io_u->ddir == DDIR_DATASYNC) {
+#ifdef CONFIG_FDATASYNC
+		ret = fdatasync(io_u->file->fd);
+#else
+		ret = io_u->xfer_buflen;
+		io_u->error = EINVAL;
+#endif
+	} else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
+		ret = do_sync_file_range(td, io_u->file);
+	else {
+		ret = io_u->xfer_buflen;
+		io_u->error = EINVAL;
+	}
+
+	if (ret < 0)
+		io_u->error = errno;
+
+	return ret;
+}
+
+int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
+{
+#ifndef FIO_HAVE_TRIM
+	io_u->error = EINVAL;
+	return 0;
+#else
+	struct fio_file *f = io_u->file;
+	int ret;
+
+	ret = os_trim(f, io_u->offset, io_u->xfer_buflen);
+	if (!ret)
+		return io_u->xfer_buflen;
+
+	io_u->error = ret;
+	return 0;
+#endif
+}