X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=46d97319fc174512fc4ca353486144c57097db31;hp=7cbdb915a19629fcbd3b500b9255c918bfbd6dd1;hb=79591fa9555024aa9564656bbbb9d4c828a3e189;hpb=709c8313cf171e63c92740c6b2f38041cb11c722

diff --git a/io_u.c b/io_u.c
index 7cbdb915..46d97319 100644
--- a/io_u.c
+++ b/io_u.c
@@ -12,6 +12,8 @@
 #include "lib/rand.h"
 #include "lib/axmap.h"
 #include "err.h"
+#include "lib/pow2.h"
+#include "minmax.h"
 
 struct io_completion_data {
 	int nr;				/* input */
@@ -25,7 +27,7 @@ struct io_completion_data {
  * 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);
 }
@@ -60,6 +62,7 @@ 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)
@@ -68,6 +71,9 @@ 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;
@@ -81,34 +87,25 @@ struct rand_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, lastb;
+	uint64_t r;
 
-	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) {
 
-	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);
-		}
+		r = __rand(&td->random_state);
 
 		dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
 
-		*b = lastb * (r / ((uint64_t) rmax + 1.0));
+		*b = lastb * (r / (rand_max(&td->random_state) + 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;
@@ -152,6 +149,79 @@ 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 __get_next_rand_offset_zoned(struct thread_data *td,
+					struct fio_file *f, enum fio_ddir ddir,
+					uint64_t *b)
+{
+	unsigned int v, send, stotal;
+	uint64_t offset, lastb;
+	static int warned;
+	struct zone_split_index *zsi;
+
+	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 = rand32_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 (!warned) {
+			log_err("fio: bug in zoned generation\n");
+			warned = 1;
+		}
+		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 flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
 {
 	struct rand_off *r1 = flist_entry(a, struct rand_off, list);
@@ -163,12 +233,22 @@ static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
 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)
+	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);
 
 	log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
 	return 1;
@@ -178,35 +258,29 @@ static int get_off_from_method(struct thread_data *td, struct fio_file *f,
  * Sort the reads for a verify phase in batches of verifysort_nr, if
  * specified.
  */
-static inline int should_sort_io(struct thread_data *td)
+static inline bool should_sort_io(struct thread_data *td)
 {
 	if (!td->o.verifysort_nr || !td->o.do_verify)
-		return 0;
+		return false;
 	if (!td_random(td))
-		return 0;
+		return false;
 	if (td->runstate != TD_VERIFYING)
-		return 0;
-	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
-		return 0;
+		return false;
+	if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
+	    td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
+		return false;
 
-	return 1;
+	return true;
 }
 
-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;
 
-	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)));
-	}
+	v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
 
 	return v <= td->o.perc_rand[ddir];
 }
@@ -221,7 +295,6 @@ 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_first_entry(&td->next_rand_list, struct rand_off, list);
 		flist_del(&r->list);
@@ -256,14 +329,15 @@ 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);
 		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,
+			f->file_name, (unsigned long long) f->last_pos[ddir],
 			(unsigned long long) f->real_file_size);
 	return 1;
 }
@@ -271,20 +345,54 @@ 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, f) && 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) {
+		struct thread_options *o = &td->o;
+		uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
+
+		if (io_size > f->last_pos[ddir])
+			f->last_pos[ddir] = 0;
+		else
+			f->last_pos[ddir] = f->last_pos[ddir] - 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) {
+			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 - 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 we're doing backwards IO,
+			 * wrap to the end.
+			 */
+			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;
 		return 0;
@@ -312,7 +420,7 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 				*is_random = 1;
 			} else {
 				*is_random = 0;
-				io_u->flags |= IO_U_F_BUSY_OK;
+				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);
@@ -322,7 +430,7 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 			ret = get_next_seq_offset(td, f, ddir, &offset);
 		}
 	} else {
-		io_u->flags |= IO_U_F_BUSY_OK;
+		io_u_set(td, io_u, IO_U_F_BUSY_OK);
 		*is_random = 0;
 
 		if (td->o.rw_seq == RW_SEQ_SEQ) {
@@ -332,8 +440,8 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 				*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;
@@ -410,8 +518,8 @@ static int get_next_offset(struct thread_data *td, struct io_u *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,
-			    unsigned int buflen)
+static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
+			     unsigned int buflen)
 {
 	struct fio_file *f = io_u->file;
 
@@ -424,7 +532,7 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
 	int ddir = io_u->ddir;
 	unsigned int buflen = 0;
 	unsigned int minbs, maxbs;
-	unsigned long r, rand_max;
+	uint64_t frand_max, r;
 
 	assert(ddir_rw(ddir));
 
@@ -443,24 +551,17 @@ 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;
-
+	frand_max = rand_max(&td->bsrange_state);
 	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 {
-			long perc = 0;
+			long long perc = 0;
 			unsigned int i;
 
 			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
@@ -468,18 +569,16 @@ 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 (!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);
+			buflen &= ~(minbs - 1);
 
 	} while (!io_u_fits(td, io_u, buflen));
 
@@ -515,15 +614,8 @@ 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;
 
-	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)));
-	}
+	v = rand32_between(&td->rwmix_state, 1, 100);
 
 	if (v <= td->o.rwmix[DDIR_READ])
 		return DDIR_READ;
@@ -531,8 +623,10 @@ 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.
@@ -542,63 +636,70 @@ 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;
 
-		ret = io_u_queued_complete(td, 1, NULL);
+		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;
-	struct timeval t;
 	long 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 necissary
 		 */
-		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];
-
-	io_u_quiesce(td);
-
-	fio_gettime(&t, NULL);
-	usec_sleep(td, usec);
-	usec = utime_since_now(&t);
+		usec = td->rate_next_io_time[ddir] - now;
 
-	td->rate_pending_usleep[ddir] -= usec;
+	if (td->o.io_submit_mode == IO_MODE_INLINE)
+		io_u_quiesce(td);
 
-	odir = ddir ^ 1;
-	if (td_rw(td) && __should_check_rate(td, odir))
-		td->rate_pending_usleep[odir] -= usec;
-
-	if (ddir_trim(ddir))
-		return ddir;
+	usec = usec_sleep(td, usec);
 
 	return ddir;
 }
@@ -658,8 +759,10 @@ 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;
@@ -667,13 +770,23 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 
 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);
+	enum fio_ddir ddir = get_rw_ddir(td);
+
+	if (td_trimwrite(td)) {
+		struct fio_file *f = io_u->file;
+		if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
+			ddir = DDIR_TRIM;
+		else
+			ddir = DDIR_WRITE;
+	}
+
+	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->flags |= 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)
@@ -686,16 +799,21 @@ 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)
 {
+	if (td->parent)
+		td = td->parent;
+
 	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->flags |= IO_U_F_FREE;
+	io_u_set(td, io_u, IO_U_F_FREE);
 
-	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
+	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);
@@ -703,7 +821,7 @@ void put_io_u(struct thread_data *td, struct io_u *io_u)
 
 void clear_io_u(struct thread_data *td, struct io_u *io_u)
 {
-	io_u->flags &= ~IO_U_F_FLIGHT;
+	io_u_clear(td, io_u, IO_U_F_FLIGHT);
 	put_io_u(td, io_u);
 }
 
@@ -714,18 +832,24 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 
 	dprint(FD_IO, "requeue %p\n", __io_u);
 
+	if (td->parent)
+		td = td->parent;
+
 	td_io_u_lock(td);
 
-	__io_u->flags |= 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->flags &= ~IO_U_F_FLIGHT;
-	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
+	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);
 	*io_u = NULL;
 }
 
@@ -733,7 +857,7 @@ 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)
+	if (td_ioengine_flagged(td, FIO_NOIO))
 		goto out;
 
 	set_rw_ddir(td, io_u);
@@ -748,9 +872,17 @@ 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;
 	}
 
@@ -953,6 +1085,34 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
 		io_u_mark_lat_msec(td, usec / 1000);
 }
 
+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;
+}
+
 /*
  * Get next file to service by choosing one at random
  */
@@ -965,17 +1125,8 @@ static struct fio_file *get_next_file_rand(struct thread_data *td,
 
 	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)));
-		}
+		fno = __get_next_fileno_rand(td);
 
 		f = td->files[fno];
 		if (fio_file_done(f))
@@ -1128,10 +1279,14 @@ static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
 		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;
@@ -1157,10 +1312,10 @@ 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;
 
@@ -1177,16 +1332,16 @@ 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)
@@ -1281,14 +1436,14 @@ 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(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;
 }
@@ -1297,6 +1452,9 @@ 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:
@@ -1313,21 +1471,23 @@ again:
 
 	if (io_u) {
 		assert(io_u->flags & IO_U_F_FREE);
-		io_u->flags &= ~(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);
 
 		io_u->error = 0;
 		io_u->acct_ddir = -1;
 		td->cur_depth++;
-		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+		assert(!(td->flags & TD_F_CHILD));
+		io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
 		io_u->ipo = NULL;
-	} else if (td->o.verify_async) {
+	} else if (td_async_processing(td)) {
 		/*
 		 * We ran out, wait for async verify threads to finish and
 		 * return one
 		 */
-		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
+		assert(!(td->flags & TD_F_CHILD));
+		assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
 		goto again;
 	}
 
@@ -1335,10 +1495,10 @@ again:
 	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;
+		return false;
 
 	if (td->trim_entries) {
 		int get_trim = 0;
@@ -1354,17 +1514,17 @@ static int check_get_trim(struct thread_data *td, struct io_u *io_u)
 			get_trim = 1;
 		}
 
-		if (get_trim && !get_next_trim(td, io_u))
-			return 1;
+		if (get_trim && 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;
@@ -1381,11 +1541,11 @@ 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;
 }
 
 /*
@@ -1476,18 +1636,19 @@ 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;
 		}
 
-		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);
+					td->o.min_bs[DDIR_WRITE],
+					io_u->buflen);
 			} else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
 				   !(td->flags & TD_F_COMPRESS))
 				do_scramble = 1;
@@ -1513,7 +1674,7 @@ 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);
@@ -1525,7 +1686,7 @@ err_put:
 	return ERR_PTR(ret);
 }
 
-void io_u_log_error(struct thread_data *td, struct io_u *io_u)
+static 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);
 
@@ -1539,23 +1700,41 @@ void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 		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");
 }
 
-static inline int gtod_reduce(struct thread_data *td)
+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->parent, io_u);
+}
+
+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,
 				  struct io_completion_data *icd,
 				  const enum fio_ddir idx, unsigned int bytes)
 {
+	const int no_reduce = !gtod_reduce(td);
 	unsigned long lusec = 0;
 
-	if (!gtod_reduce(td))
+	if (td->parent)
+		td = td->parent;
+
+	if (no_reduce)
 		lusec = utime_since(&io_u->issue_time, &icd->time);
 
 	if (!td->o.disable_lat) {
@@ -1579,27 +1758,53 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 		}
 	}
 
-	if (!td->o.disable_clat) {
-		add_clat_sample(td, idx, lusec, bytes, io_u->offset);
-		io_u_mark_latency(td, lusec);
-	}
+	if (ddir_rw(idx)) {
+		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 (!td->o.disable_bw && per_unit_log(td->bw_log))
+			add_bw_sample(td, io_u, bytes, lusec);
 
-	if (!gtod_reduce(td))
-		add_iops_sample(td, idx, bytes, &icd->time);
+		if (no_reduce && per_unit_log(td->iops_log))
+			add_iops_sample(td, io_u, bytes);
+	}
+
+	if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
+		uint32_t *info = io_u_block_info(td, io_u);
+		if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
+			if (io_u->ddir == DDIR_TRIM) {
+				*info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
+						BLOCK_INFO_TRIMS(*info) + 1);
+			} else if (io_u->ddir == DDIR_WRITE) {
+				*info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
+								*info);
+			}
+		}
+	}
 }
 
-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;
+
+	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;
 
-	bytes = td->this_io_bytes[ddir];
-	bps = td->rate_bps[ddir];
-	secs = bytes / bps;
-	remainder = bytes % bps;
-	return remainder * 1000000 / bps + secs * 1000000;
+	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 void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
@@ -1611,9 +1816,8 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 
 	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);
+	io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
 
 	/*
 	 * Mark IO ok to verify
@@ -1630,8 +1834,6 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 		}
 	}
 
-	td_io_u_unlock(td);
-
 	if (ddir_sync(ddir)) {
 		td->last_was_sync = 1;
 		if (f) {
@@ -1646,7 +1848,6 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 
 	if (!io_u->error && ddir_rw(ddir)) {
 		unsigned int bytes = io_u->buflen - io_u->resid;
-		const enum fio_ddir oddir = ddir ^ 1;
 		int ret;
 
 		td->io_blocks[ddir]++;
@@ -1656,32 +1857,13 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 		if (!(io_u->flags & IO_U_F_VER_LIST))
 			td->this_io_bytes[ddir] += bytes;
 
-		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;
-		}
+		if (ddir == DDIR_WRITE)
+			file_log_write_comp(td, f, io_u->offset, bytes);
 
 		if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
-					   td->runstate == TD_VERIFYING)) {
+					   td->runstate == TD_VERIFYING))
 			account_io_completion(td, io_u, icd, ddir, bytes);
 
-			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) {
@@ -1723,7 +1905,7 @@ 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;
 }
 
@@ -1746,10 +1928,10 @@ static void ios_completed(struct thread_data *td,
 /*
  * Complete a single io_u for the sync engines.
  */
-int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
-		       uint64_t *bytes)
+int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 {
 	struct io_completion_data icd;
+	int ddir;
 
 	init_icd(td, &icd, 1);
 	io_completed(td, &io_u, &icd);
@@ -1762,12 +1944,8 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
 		return -1;
 	}
 
-	if (bytes) {
-		int ddir;
-
-		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
-			bytes[ddir] += icd.bytes_done[ddir];
-	}
+	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
+		td->bytes_done[ddir] += icd.bytes_done[ddir];
 
 	return 0;
 }
@@ -1775,12 +1953,11 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
 /*
  * Called to complete min_events number of io for the async engines.
  */
-int io_u_queued_complete(struct thread_data *td, int min_evts,
-			 uint64_t *bytes)
+int io_u_queued_complete(struct thread_data *td, int min_evts)
 {
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
-	int ret;
+	int ret, ddir;
 	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
 	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
@@ -1790,7 +1967,9 @@ int io_u_queued_complete(struct thread_data *td, int min_evts,
 	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;
@@ -1804,14 +1983,10 @@ int io_u_queued_complete(struct thread_data *td, int min_evts,
 		return -1;
 	}
 
-	if (bytes) {
-		int ddir;
+	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
+		td->bytes_done[ddir] += icd.bytes_done[ddir];
 
-		for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
-			bytes[ddir] += icd.bytes_done[ddir];
-	}
-
-	return 0;
+	return ret;
 }
 
 /*
@@ -1823,32 +1998,84 @@ 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);
+
+		if (td->parent)
+			td = td->parent;
+
 		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
 				io_u->offset);
 	}
 }
 
+/*
+ * See if we should reuse the last seed, if dedupe is enabled
+ */
+static struct frand_state *get_buf_state(struct thread_data *td)
+{
+	unsigned int v;
+
+	if (!td->o.dedupe_percentage)
+		return &td->buf_state;
+	else if (td->o.dedupe_percentage == 100) {
+		frand_copy(&td->buf_state_prev, &td->buf_state);
+		return &td->buf_state;
+	}
+
+	v = rand32_between(&td->dedupe_state, 1, 100);
+
+	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 (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)
 {
-	if (td->o.buffer_pattern_bytes)
-		fill_buffer_pattern(td, buf, max_bs);
-	else if (!td->o.zero_buffers) {
+	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;
+		unsigned int this_write;
 
-		if (perc) {
-			unsigned int seg = min_write;
+		do {
+			rs = get_buf_state(td);
 
-			seg = min(min_write, td->o.compress_chunk);
-			if (!seg)
-				seg = min_write;
+			min_write = min(min_write, left);
 
-			fill_random_buf_percentage(&td->buf_state, buf,
-						perc, seg, max_bs);
-		} else
-			fill_random_buf(&td->buf_state, buf, max_bs);
-	} else
+			if (perc) {
+				this_write = min_not_zero(min_write,
+							td->o.compress_chunk);
+
+				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 += this_write;
+			left -= this_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);
 }
 
 /*