X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=dc4473beab1737660f2ebeb2f1c91418b1a14da6;hp=0ffae29c9f035a2d387ad00b0ae7d882ac0515b3;hb=cbe8d7561cf6d81d741d87eb7940db2a111d2144;hpb=6977bcd0e4ee3faa7ffd8f208e4031bdf906ed88

diff --git a/io_u.c b/io_u.c
index 0ffae29c..dc4473be 100644
--- a/io_u.c
+++ b/io_u.c
@@ -6,11 +6,9 @@
 #include <assert.h>
 
 #include "fio.h"
-
-/*
- * Change this define to play with the timeout handling
- */
-#undef FIO_USE_TIMEOUT
+#include "hash.h"
+#include "verify.h"
+#include "lib/rand.h"
 
 struct io_completion_data {
 	int nr;				/* input */
@@ -24,15 +22,14 @@ struct io_completion_data {
  * The ->file_map[] contains a map of blocks we have or have not done io
  * to yet. Used to make sure we cover the entire range in a fair fashion.
  */
-static int random_map_free(struct thread_data *td, struct fio_file *f,
-			   const unsigned long long block)
+static int random_map_free(struct fio_file *f, const unsigned long long block)
 {
-	unsigned int idx = RAND_MAP_IDX(td, f, block);
-	unsigned int bit = RAND_MAP_BIT(td, f, block);
+	unsigned int idx = RAND_MAP_IDX(f, block);
+	unsigned int bit = RAND_MAP_BIT(f, block);
 
 	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
 
-	return (f->file_map[idx] & (1UL << bit)) == 0;
+	return (f->file_map[idx] & (1 << bit)) == 0;
 }
 
 /*
@@ -43,48 +40,75 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
 	unsigned int min_bs = td->o.rw_min_bs;
 	struct fio_file *f = io_u->file;
 	unsigned long long block;
-	unsigned int blocks;
-	unsigned int nr_blocks;
+	unsigned int blocks, nr_blocks;
 
 	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
-	blocks = 0;
 	nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
+	blocks = 0;
 
-	while (blocks < nr_blocks) {
+	while (nr_blocks) {
+		unsigned int this_blocks, mask;
 		unsigned int idx, bit;
 
 		/*
 		 * If we have a mixed random workload, we may
 		 * encounter blocks we already did IO to.
 		 */
-		if ((td->o.ddir_nr == 1) && !random_map_free(td, f, block))
+		if ((td->o.ddir_nr == 1) && !random_map_free(f, block))
 			break;
 
-		idx = RAND_MAP_IDX(td, f, block);
-		bit = RAND_MAP_BIT(td, f, block);
+		idx = RAND_MAP_IDX(f, block);
+		bit = RAND_MAP_BIT(f, block);
 
 		fio_assert(td, idx < f->num_maps);
 
-		f->file_map[idx] |= (1UL << bit);
-		block++;
-		blocks++;
+		this_blocks = nr_blocks;
+		if (this_blocks + bit > BLOCKS_PER_MAP)
+			this_blocks = BLOCKS_PER_MAP - bit;
+
+		do {
+			if (this_blocks == BLOCKS_PER_MAP)
+				mask = -1U;
+			else
+				mask = ((1U << this_blocks) - 1) << bit;
+	
+			if (!(f->file_map[idx] & mask))
+				break;
+
+			this_blocks--;
+		} while (this_blocks);
+
+		if (!this_blocks)
+			break;
+
+		f->file_map[idx] |= mask;
+		nr_blocks -= this_blocks;
+		blocks += this_blocks;
+		block += this_blocks;
 	}
 
 	if ((blocks * min_bs) < io_u->buflen)
 		io_u->buflen = blocks * min_bs;
 }
 
-static inline unsigned long long last_block(struct thread_data *td,
-					    struct fio_file *f,
-					    enum fio_ddir ddir)
+static unsigned long long last_block(struct thread_data *td, struct fio_file *f,
+				     enum fio_ddir ddir)
 {
 	unsigned long long max_blocks;
+	unsigned long long max_size;
+
+	/*
+	 * Hmm, should we make sure that ->io_size <= ->real_file_size?
+	 */
+	max_size = f->io_size;
+	if (max_size > f->real_file_size)
+		max_size = f->real_file_size;
 
-	max_blocks = f->io_size / (unsigned long long) td->o.min_bs[ddir];
+	max_blocks = max_size / (unsigned long long) td->o.ba[ddir];
 	if (!max_blocks)
 		return 0;
 
-	return max_blocks - 1;
+	return max_blocks;
 }
 
 /*
@@ -98,9 +122,10 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 
 	i = f->last_free_lookup;
 	*b = (i * BLOCKS_PER_MAP);
-	while ((*b) * min_bs < f->real_file_size) {
-		if (f->file_map[i] != -1UL) {
-			*b += fio_ffz(f->file_map[i]);
+	while ((*b) * min_bs < f->real_file_size &&
+		(*b) * min_bs < f->io_size) {
+		if (f->file_map[i] != (unsigned int) -1) {
+			*b += ffz(f->file_map[i]);
 			if (*b > last_block(td, f, ddir))
 				break;
 			f->last_free_lookup = i;
@@ -124,18 +149,19 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 	do {
 		r = os_random_long(&td->random_state);
 		dprint(FD_RANDOM, "off rand %llu\n", r);
-		*b = (last_block(td, f, ddir) - 1) * (r / ((unsigned long long) RAND_MAX + 1.0));
+		*b = (last_block(td, f, ddir) - 1)
+			* (r / ((unsigned long long) OS_RAND_MAX + 1.0));
 
 		/*
 		 * if we are not maintaining a random map, we are done.
 		 */
-		if (td->o.norandommap)
+		if (!file_randommap(td, f))
 			return 0;
 
 		/*
 		 * calculate map offset and check if it's free
 		 */
-		if (random_map_free(td, f, *b))
+		if (random_map_free(f, *b))
 			return 0;
 
 		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
@@ -149,7 +175,8 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 	 */
 	loops = 10;
 	do {
-		f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0));
+		f->last_free_lookup = (f->num_maps - 1) *
+					(r / (OS_RAND_MAX + 1.0));
 		if (!get_next_free_block(td, f, ddir, b))
 			return 0;
 
@@ -168,7 +195,7 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
  * until we find a free one. For sequential io, just return the end of
  * the last io issued.
  */
-static int get_next_offset(struct thread_data *td, struct io_u *io_u)
+static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f = io_u->file;
 	unsigned long long b;
@@ -177,18 +204,32 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 	if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
 		td->ddir_nr = td->o.ddir_nr;
 
-		if (get_next_rand_offset(td, f, ddir, &b))
+		if (get_next_rand_offset(td, f, ddir, &b)) {
+			dprint(FD_IO, "%s: getting rand offset failed\n",
+				f->file_name);
 			return 1;
+		}
 	} else {
 		if (f->last_pos >= f->real_file_size) {
 			if (!td_random(td) ||
-			     get_next_rand_offset(td, f, ddir, &b))
+			     get_next_rand_offset(td, f, ddir, &b)) {
+				dprint(FD_IO, "%s: pos %llu > size %llu\n",
+						f->file_name, f->last_pos,
+						f->real_file_size);
 				return 1;
+			}
 		} else
 			b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
 	}
 
-	io_u->offset = (b * td->o.min_bs[ddir]) + f->file_offset;
+	io_u->offset = b * td->o.ba[ddir];
+	if (io_u->offset >= f->io_size) {
+		dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n",
+					io_u->offset, f->io_size);
+		return 1;
+	}
+
+	io_u->offset += f->file_offset;
 	if (io_u->offset >= f->real_file_size) {
 		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
 					io_u->offset, f->real_file_size);
@@ -198,47 +239,73 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 	return 0;
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_off)
+		return ops->fill_io_u_off(td, io_u);
+
+	return __get_next_offset(td, io_u);
+}
+
+static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
 	const int ddir = io_u->ddir;
-	unsigned int buflen;
+	unsigned int uninitialized_var(buflen);
+	unsigned int minbs, maxbs;
 	long r;
 
-	if (td->o.min_bs[ddir] == td->o.max_bs[ddir])
-		buflen = td->o.min_bs[ddir];
+	minbs = td->o.min_bs[ddir];
+	maxbs = td->o.max_bs[ddir];
+
+	if (minbs == maxbs)
+		buflen = minbs;
 	else {
 		r = os_random_long(&td->bsrange_state);
-		if (!td->o.bssplit_nr)
-			buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
-		else {
+		if (!td->o.bssplit_nr[ddir]) {
+			buflen = 1 + (unsigned int) ((double) maxbs *
+					(r / (OS_RAND_MAX + 1.0)));
+			if (buflen < minbs)
+				buflen = minbs;
+		} else {
 			long perc = 0;
 			unsigned int i;
 
-			for (i = 0; i < td->o.bssplit_nr; i++) {
-				struct bssplit *bsp = &td->o.bssplit[i];
+			for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
+				struct bssplit *bsp = &td->o.bssplit[ddir][i];
 
 				buflen = bsp->bs;
 				perc += bsp->perc;
-				if (r <= ((LONG_MAX / 100L) * perc))
+				if (r <= ((OS_RAND_MAX / 100L) * perc))
 					break;
 			}
 		}
-		if (!td->o.bs_unaligned)
-			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
+		if (!td->o.bs_unaligned && is_power_of_2(minbs))
+			buflen = (buflen + minbs - 1) & ~(minbs - 1);
 	}
 
 	if (io_u->offset + buflen > io_u->file->real_file_size) {
 		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
-						td->o.min_bs[ddir], ddir);
-		buflen = td->o.min_bs[ddir];
+						minbs, ddir);
+		buflen = minbs;
 	}
 
 	return buflen;
 }
 
+static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+{
+	struct prof_io_ops *ops = &td->prof_io_ops;
+
+	if (ops->fill_io_u_size)
+		return ops->fill_io_u_size(td, io_u);
+
+	return __get_next_buflen(td, io_u);
+}
+
 static void set_rwmix_bytes(struct thread_data *td)
 {
-	unsigned long long rbytes;
 	unsigned int diff;
 
 	/*
@@ -246,10 +313,8 @@ static void set_rwmix_bytes(struct thread_data *td)
 	 * buffered writes may issue a lot quicker than they complete,
 	 * whereas reads do not.
 	 */
-	rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
 	diff = td->o.rwmix[td->rwmix_ddir ^ 1];
-
-	td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
+	td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100;
 }
 
 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
@@ -258,13 +323,60 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 	long r;
 
 	r = os_random_long(&td->rwmix_state);
-	v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
-	if (v < td->o.rwmix[DDIR_READ])
+	v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
+	if (v <= td->o.rwmix[DDIR_READ])
 		return DDIR_READ;
 
 	return DDIR_WRITE;
 }
 
+static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
+{
+	enum fio_ddir odir = ddir ^ 1;
+	struct timeval t;
+	long usec;
+
+	if (td->rate_pending_usleep[ddir] <= 0)
+		return ddir;
+
+	/*
+	 * We have too much pending sleep in this direction. See if we
+	 * should switch.
+	 */
+	if (td_rw(td)) {
+		/*
+		 * Other direction does not have too much pending, switch
+		 */
+		if (td->rate_pending_usleep[odir] < 100000)
+			return odir;
+
+		/*
+		 * Both directions have pending sleep. Sleep the minimum time
+		 * and deduct from both.
+		 */
+		if (td->rate_pending_usleep[ddir] <=
+			td->rate_pending_usleep[odir]) {
+			usec = td->rate_pending_usleep[ddir];
+		} else {
+			usec = td->rate_pending_usleep[odir];
+			ddir = odir;
+		}
+	} else
+		usec = td->rate_pending_usleep[ddir];
+
+	fio_gettime(&t, NULL);
+	usec_sleep(td, usec);
+	usec = utime_since_now(&t);
+
+	td->rate_pending_usleep[ddir] -= usec;
+
+	odir = ddir ^ 1;
+	if (td_rw(td) && __should_check_rate(td, odir))
+		td->rate_pending_usleep[odir] -= usec;
+
+	return ddir;
+}
+
 /*
  * Return the data direction for the next io_u. If the job is a
  * mixed read/write workload, check the rwmix cycle and switch if
@@ -272,88 +384,110 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
  */
 static enum fio_ddir get_rw_ddir(struct thread_data *td)
 {
-	if (td_rw(td)) {
-		struct timeval now;
-		unsigned long elapsed;
-		unsigned int cycle;
+	enum fio_ddir ddir;
 
-		fio_gettime(&now, NULL);
-	 	elapsed = mtime_since_now(&td->rwmix_switch);
+	/*
+	 * see if it's time to fsync
+	 */
+	if (td->o.fsync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC;
 
-		/*
-		 * if this is the first cycle, make it shorter
-		 */
-		cycle = td->o.rwmixcycle;
-		if (!td->rwmix_bytes)
-			cycle /= 10;
+	/*
+	 * see if it's time to fdatasync
+	 */
+	if (td->o.fdatasync_blocks &&
+	   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_DATASYNC;
+
+	/*
+	 * see if it's time to sync_file_range
+	 */
+	if (td->sync_file_range_nr &&
+	   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
+	     td->io_issues[DDIR_WRITE] && should_fsync(td))
+		return DDIR_SYNC_FILE_RANGE;
 
+	if (td_rw(td)) {
 		/*
 		 * Check if it's time to seed a new data direction.
 		 */
-		if (elapsed >= cycle ||
-		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
-			unsigned long long max_bytes;
-			enum fio_ddir ddir;			
-
+		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
 			/*
 			 * Put a top limit on how many bytes we do for
 			 * one data direction, to avoid overflowing the
 			 * ranges too much
 			 */
 			ddir = get_rand_ddir(td);
-			max_bytes = td->this_io_bytes[ddir];
-			if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
-				if (!td->rw_end_set[ddir]) {
-					td->rw_end_set[ddir] = 1;
-					memcpy(&td->rw_end[ddir], &now, sizeof(now));
-				}
-				ddir ^= 1;
-			}
 
 			if (ddir != td->rwmix_ddir)
 				set_rwmix_bytes(td);
 
 			td->rwmix_ddir = ddir;
-			memcpy(&td->rwmix_switch, &now, sizeof(now));
 		}
-		return td->rwmix_ddir;
+		ddir = td->rwmix_ddir;
 	} else if (td_read(td))
-		return DDIR_READ;
+		ddir = DDIR_READ;
 	else
-		return DDIR_WRITE;
+		ddir = DDIR_WRITE;
+
+	td->rwmix_ddir = rate_ddir(td, ddir);
+	return td->rwmix_ddir;
+}
+
+void put_file_log(struct thread_data *td, struct fio_file *f)
+{
+	int ret = put_file(td, f);
+
+	if (ret)
+		td_verror(td, ret, "file close");
 }
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
-	assert((io_u->flags & IO_U_F_FREE) == 0);
-	io_u->flags |= IO_U_F_FREE;
+	td_io_u_lock(td);
 
-	if (io_u->file) {
-		int ret = put_file(td, io_u->file);
+	io_u->flags |= IO_U_F_FREE;
+	io_u->flags &= ~IO_U_F_FREE_DEF;
 
-		if (ret)
-			td_verror(td, ret, "file close");
-	}
+	if (io_u->file)
+		put_file_log(td, io_u->file);
 
 	io_u->file = NULL;
-	list_del(&io_u->list);
-	list_add(&io_u->list, &td->io_u_freelist);
-	td->cur_depth--;
+	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
+	flist_del_init(&io_u->list);
+	flist_add(&io_u->list, &td->io_u_freelist);
+	td_io_u_unlock(td);
+	td_io_u_free_notify(td);
+}
+
+void clear_io_u(struct thread_data *td, struct io_u *io_u)
+{
+	io_u->flags &= ~IO_U_F_FLIGHT;
+	put_io_u(td, io_u);
 }
 
 void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 {
 	struct io_u *__io_u = *io_u;
 
+	dprint(FD_IO, "requeue %p\n", __io_u);
+
+	td_io_u_lock(td);
+
 	__io_u->flags |= IO_U_F_FREE;
-	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
+	if ((__io_u->flags & IO_U_F_FLIGHT) && !ddir_sync(__io_u->ddir))
 		td->io_issues[__io_u->ddir]--;
-		
-	__io_u->flags &= ~IO_U_F_FLIGHT;
 
-	list_del(&__io_u->list);
-	list_add_tail(&__io_u->list, &td->io_u_requeues);
-	td->cur_depth--;
+	__io_u->flags &= ~IO_U_F_FLIGHT;
+	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
+		td->cur_depth--;
+	flist_del(&__io_u->list);
+	flist_add_tail(&__io_u->list, &td->io_u_requeues);
+	td_io_u_unlock(td);
 	*io_u = NULL;
 }
 
@@ -362,17 +496,13 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	if (td->io_ops->flags & FIO_NOIO)
 		goto out;
 
+	io_u->ddir = get_rw_ddir(td);
+
 	/*
-	 * see if it's time to sync
+	 * fsync() or fdatasync(), we are done
 	 */
-	if (td->o.fsync_blocks &&
-	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
-	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
-		io_u->ddir = DDIR_SYNC;
+	if (ddir_sync(io_u->ddir))
 		goto out;
-	}
-
-	io_u->ddir = get_rw_ddir(td);
 
 	/*
 	 * See if it's time to switch to a new zone
@@ -408,7 +538,7 @@ 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) && !td->o.norandommap)
+	if (td_random(td) && file_randommap(td, io_u->file))
 		mark_random_map(td, io_u);
 
 	/*
@@ -421,12 +551,50 @@ out:
 	return 0;
 }
 
-void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
+static void __io_u_mark_map(unsigned int *map, unsigned int nr)
 {
 	int index = 0;
 
-	if (io_u->ddir == DDIR_SYNC)
-		return;
+	switch (nr) {
+	default:
+		index = 6;
+		break;
+	case 33 ... 64:
+		index = 5;
+		break;
+	case 17 ... 32:
+		index = 4;
+		break;
+	case 9 ... 16:
+		index = 3;
+		break;
+	case 5 ... 8:
+		index = 2;
+		break;
+	case 1 ... 4:
+		index = 1;
+	case 0:
+		break;
+	}
+
+	map[index]++;
+}
+
+void io_u_mark_submit(struct thread_data *td, unsigned int nr)
+{
+	__io_u_mark_map(td->ts.io_u_submit, nr);
+	td->ts.total_submit++;
+}
+
+void io_u_mark_complete(struct thread_data *td, unsigned int nr)
+{
+	__io_u_mark_map(td->ts.io_u_complete, nr);
+	td->ts.total_complete++;
+}
+
+void io_u_mark_depth(struct thread_data *td, unsigned int nr)
+{
+	int index = 0;
 
 	switch (td->cur_depth) {
 	default:
@@ -450,8 +618,7 @@ void io_u_mark_depth(struct thread_data *td, struct io_u *io_u)
 		break;
 	}
 
-	td->ts.io_u_map[index]++;
-	td->ts.total_io_u[io_u->ddir]++;
+	td->ts.io_u_map[index] += nr;
 }
 
 static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
@@ -551,24 +718,38 @@ static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
 /*
  * Get next file to service by choosing one at random
  */
-static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
-					   int badf)
+static struct fio_file *get_next_file_rand(struct thread_data *td,
+					   enum fio_file_flags goodf,
+					   enum fio_file_flags badf)
 {
 	struct fio_file *f;
 	int fno;
 
 	do {
 		long r = os_random_long(&td->next_file_state);
+		int opened = 0;
 
-		fno = (unsigned int) ((double) td->o.nr_files * (r / (RAND_MAX + 1.0)));
-		f = &td->files[fno];
-		if (f->flags & FIO_FILE_DONE)
+		fno = (unsigned int) ((double) td->o.nr_files
+			* (r / (OS_RAND_MAX + 1.0)));
+		f = td->files[fno];
+		if (fio_file_done(f))
 			continue;
 
+		if (!fio_file_open(f)) {
+			int err;
+
+			err = td_io_open_file(td, f);
+			if (err)
+				continue;
+			opened = 1;
+		}
+
 		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) {
 			dprint(FD_FILE, "get_next_file_rand: %p\n", f);
 			return f;
 		}
+		if (opened)
+			td_io_close_file(td, f);
 	} while (1);
 }
 
@@ -582,20 +763,41 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 	struct fio_file *f;
 
 	do {
-		f = &td->files[td->next_file];
+		int opened = 0;
+
+		f = td->files[td->next_file];
 
 		td->next_file++;
 		if (td->next_file >= td->o.nr_files)
 			td->next_file = 0;
 
-		if (f->flags & FIO_FILE_DONE) {
+		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
+		if (fio_file_done(f)) {
 			f = NULL;
 			continue;
 		}
 
+		if (!fio_file_open(f)) {
+			int err;
+
+			err = td_io_open_file(td, f);
+			if (err) {
+				dprint(FD_FILE, "error %d on open of %s\n",
+					err, f->file_name);
+				f = NULL;
+				continue;
+			}
+			opened = 1;
+		}
+
+		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf,
+								f->flags);
 		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			break;
 
+		if (opened)
+			td_io_close_file(td, f);
+
 		f = NULL;
 	} while (td->next_file != old_next_file);
 
@@ -603,46 +805,49 @@ static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf,
 	return f;
 }
 
-static struct fio_file *get_next_file(struct thread_data *td)
+static struct fio_file *__get_next_file(struct thread_data *td)
 {
 	struct fio_file *f;
 
 	assert(td->o.nr_files <= td->files_index);
 
-	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) {
-		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d, nr_files=%d\n", td->nr_open_files, td->nr_done_files, td->o.nr_files);
+	if (td->nr_done_files >= td->o.nr_files) {
+		dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d,"
+				" nr_files=%d\n", td->nr_open_files,
+						  td->nr_done_files,
+						  td->o.nr_files);
 		return NULL;
 	}
 
 	f = td->file_service_file;
-	if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--)
-		goto out;
+	if (f && fio_file_open(f) && !fio_file_closing(f)) {
+		if (td->o.file_service_type == FIO_FSERVICE_SEQ)
+			goto out;
+		if (td->file_service_left--)
+			goto out;
+	}
 
-	if (td->o.file_service_type == FIO_FSERVICE_RR)
-		f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+	if (td->o.file_service_type == FIO_FSERVICE_RR ||
+	    td->o.file_service_type == FIO_FSERVICE_SEQ)
+		f = get_next_file_rr(td, FIO_FILE_open, FIO_FILE_closing);
 	else
-		f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING);
+		f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
 
 	td->file_service_file = f;
 	td->file_service_left = td->file_service_nr - 1;
 out:
-	dprint(FD_FILE, "get_next_file: %p\n", f);
+	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
 	return f;
 }
 
-static struct fio_file *find_next_new_file(struct thread_data *td)
+static struct fio_file *get_next_file(struct thread_data *td)
 {
-	struct fio_file *f;
-
-	if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files)
-		return NULL;
+	struct prof_io_ops *ops = &td->prof_io_ops;
 
-	if (td->o.file_service_type == FIO_FSERVICE_RR)
-		f = get_next_file_rr(td, 0, FIO_FILE_OPEN);
-	else
-		f = get_next_file_rand(td, 0, FIO_FILE_OPEN);
+	if (ops->get_next_file)
+		return ops->get_next_file(td);
 
-	return f;
+	return __get_next_file(td);
 }
 
 static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
@@ -654,35 +859,19 @@ static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 		if (!f)
 			return 1;
 
-set_file:
 		io_u->file = f;
 		get_file(f);
 
 		if (!fill_io_u(td, io_u))
 			break;
 
-		/*
-		 * td_io_close() does a put_file() as well, so no need to
-		 * do that here.
-		 */
-		io_u->file = NULL;
+		put_file_log(td, f);
 		td_io_close_file(td, f);
-		f->flags |= FIO_FILE_DONE;
+		io_u->file = NULL;
+		fio_file_set_done(f);
 		td->nr_done_files++;
-
-		/*
-		 * probably not the right place to do this, but see
-		 * if we need to open a new file
-		 */
-		if (td->nr_open_files < td->o.open_files &&
-		    td->o.open_files != td->o.nr_files) {
-			f = find_next_new_file(td);
-
-			if (!f || td_io_open_file(td, f))
-				return 1;
-
-			goto set_file;
-		}
+		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;
@@ -693,10 +882,13 @@ struct io_u *__get_io_u(struct thread_data *td)
 {
 	struct io_u *io_u = NULL;
 
-	if (!list_empty(&td->io_u_requeues))
-		io_u = list_entry(td->io_u_requeues.next, struct io_u, list);
+	td_io_u_lock(td);
+
+again:
+	if (!flist_empty(&td->io_u_requeues))
+		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
 	else if (!queue_full(td)) {
-		io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
+		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
 
 		io_u->buflen = 0;
 		io_u->resid = 0;
@@ -706,14 +898,23 @@ struct io_u *__get_io_u(struct thread_data *td)
 
 	if (io_u) {
 		assert(io_u->flags & IO_U_F_FREE);
-		io_u->flags &= ~IO_U_F_FREE;
+		io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
 
 		io_u->error = 0;
-		list_del(&io_u->list);
-		list_add(&io_u->list, &td->io_u_busylist);
+		flist_del(&io_u->list);
+		flist_add(&io_u->list, &td->io_u_busylist);
 		td->cur_depth++;
+		io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+	} else if (td->o.verify_async) {
+		/*
+		 * We ran out, wait for async verify threads to finish and
+		 * return one
+		 */
+		pthread_cond_wait(&td->free_cond, &td->io_u_lock);
+		goto again;
 	}
 
+	td_io_u_unlock(td);
 	return io_u;
 }
 
@@ -732,6 +933,24 @@ struct io_u *get_io_u(struct thread_data *td)
 		return NULL;
 	}
 
+	if (td->o.verify_backlog && td->io_hist_len) {
+		int get_verify = 0;
+
+		if (td->verify_batch) {
+			td->verify_batch--;
+			get_verify = 1;
+		} else if (!(td->io_hist_len % td->o.verify_backlog) &&
+			 td->last_ddir != DDIR_READ) {
+			td->verify_batch = td->o.verify_batch;
+			if (!td->verify_batch)
+				td->verify_batch = td->o.verify_backlog;
+			get_verify = 1;
+		}
+
+		if (get_verify && !get_next_verify(td, io_u))
+			goto out;
+	}
+
 	/*
 	 * from a requeue, io_u already setup
 	 */
@@ -748,11 +967,11 @@ struct io_u *get_io_u(struct thread_data *td)
 		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
 		goto err_put;
 	}
-	
+
 	f = io_u->file;
-	assert(f->flags & FIO_FILE_OPEN);
+	assert(fio_file_open(f));
 
-	if (io_u->ddir != DDIR_SYNC) {
+	if (!ddir_sync(io_u->ddir)) {
 		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
 			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
 			goto err_put;
@@ -760,19 +979,30 @@ struct io_u *get_io_u(struct thread_data *td)
 
 		f->last_pos = io_u->offset + io_u->buflen;
 
-		if (td->o.verify != VERIFY_NONE)
+		if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_WRITE)
 			populate_verify_io_u(td, io_u);
+		else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE)
+			io_u_fill_buffer(td, io_u, io_u->xfer_buflen);
+		else if (io_u->ddir == DDIR_READ) {
+			/*
+			 * Reset the buf_filled parameters so next time if the
+			 * buffer is used for writes it is refilled.
+			 */
+			io_u->buf_filled = 0;
+			io_u->buf_filled_len = 0;
+		}
 	}
 
 	/*
 	 * Set io data pointers.
 	 */
-	io_u->endpos = io_u->offset + io_u->buflen;
 	io_u->xfer_buf = io_u->buf;
 	io_u->xfer_buflen = io_u->buflen;
+
 out:
 	if (!td_io_prep(td, io_u)) {
-		fio_gettime(&io_u->start_time, NULL);
+		if (!td->o.disable_slat)
+			fio_gettime(&io_u->start_time, NULL);
 		return io_u;
 	}
 err_put:
@@ -792,7 +1022,8 @@ void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 
 	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("     %s offset=%llu, buflen=%lu\n", msg[io_u->ddir],
+					io_u->offset, io_u->xfer_buflen);
 
 	if (!td->error)
 		td_verror(td, io_u->error, "io_u error");
@@ -801,36 +1032,86 @@ void io_u_log_error(struct thread_data *td, struct io_u *io_u)
 static void io_completed(struct thread_data *td, struct io_u *io_u,
 			 struct io_completion_data *icd)
 {
-	unsigned long usec;
+	/*
+	 * Older gcc's are too dumb to realize that usec is always used
+	 * initialized, silence that warning.
+	 */
+	unsigned long uninitialized_var(usec);
+	struct fio_file *f;
 
 	dprint_io_u(io_u, "io complete");
 
+	td_io_u_lock(td);
 	assert(io_u->flags & IO_U_F_FLIGHT);
 	io_u->flags &= ~IO_U_F_FLIGHT;
+	td_io_u_unlock(td);
 
-	if (io_u->ddir == DDIR_SYNC) {
+	if (ddir_sync(io_u->ddir)) {
 		td->last_was_sync = 1;
+		f = io_u->file;
+		if (f) {
+			f->first_write = -1ULL;
+			f->last_write = -1ULL;
+		}
 		return;
 	}
 
 	td->last_was_sync = 0;
+	td->last_ddir = io_u->ddir;
 
 	if (!io_u->error) {
 		unsigned int bytes = io_u->buflen - io_u->resid;
 		const enum fio_ddir idx = io_u->ddir;
+		const enum fio_ddir odx = io_u->ddir ^ 1;
 		int ret;
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
 		td->this_io_bytes[idx] += bytes;
 
-		io_u->file->last_completed_pos = io_u->endpos;
+		if (idx == DDIR_WRITE) {
+			f = io_u->file;
+			if (f) {
+				if (f->first_write == -1ULL ||
+				    io_u->offset < f->first_write)
+					f->first_write = io_u->offset;
+				if (f->last_write == -1ULL ||
+				    ((io_u->offset + bytes) > f->last_write))
+					f->last_write = io_u->offset + bytes;
+			}
+		}
+
+		if (ramp_time_over(td)) {
+			unsigned long uninitialized_var(lusec);
 
-		usec = utime_since(&io_u->issue_time, &icd->time);
+			if (!td->o.disable_clat || !td->o.disable_bw)
+				lusec = utime_since(&io_u->issue_time,
+							&icd->time);
+			if (!td->o.disable_lat) {
+				unsigned long tusec;
 
-		add_clat_sample(td, idx, usec);
-		add_bw_sample(td, idx, &icd->time);
-		io_u_mark_latency(td, usec);
+				tusec = utime_since(&io_u->start_time,
+							&icd->time);
+				add_lat_sample(td, idx, tusec, bytes);
+			}
+			if (!td->o.disable_clat) {
+				add_clat_sample(td, idx, lusec, bytes);
+				io_u_mark_latency(td, lusec);
+			}
+			if (!td->o.disable_bw)
+				add_bw_sample(td, idx, bytes, &icd->time);
+			if (__should_check_rate(td, idx)) {
+				td->rate_pending_usleep[idx] =
+					((td->this_io_bytes[idx] *
+					  td->rate_nsec_cycle[idx]) / 1000 -
+					 utime_since_now(&td->start));
+			}
+			if (__should_check_rate(td, idx ^ 1))
+				td->rate_pending_usleep[odx] =
+					((td->this_io_bytes[odx] *
+					  td->rate_nsec_cycle[odx]) / 1000 -
+					 utime_since_now(&td->start));
+		}
 
 		if (td_write(td) && idx == DDIR_WRITE &&
 		    td->o.do_verify &&
@@ -848,11 +1129,24 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 		icd->error = io_u->error;
 		io_u_log_error(td, io_u);
 	}
+	if (td->o.continue_on_error && icd->error &&
+	    td_non_fatal_error(icd->error)) {
+		/*
+		 * If there is a non_fatal error, then add to the error count
+		 * and clear all the errors.
+		 */
+		update_error_count(td, icd->error);
+		td_clear_error(td);
+		icd->error = 0;
+		io_u->error = 0;
+	}
 }
 
-static void init_icd(struct io_completion_data *icd, int nr)
+static void init_icd(struct thread_data *td, struct io_completion_data *icd,
+		     int nr)
 {
-	fio_gettime(&icd->time, NULL);
+	if (!td->o.disable_clat || !td->o.disable_bw)
+		fio_gettime(&icd->time, NULL);
 
 	icd->nr = nr;
 
@@ -870,57 +1164,75 @@ static void ios_completed(struct thread_data *td,
 		io_u = td->io_ops->event(td, i);
 
 		io_completed(td, io_u, icd);
-		put_io_u(td, io_u);
+
+		if (!(io_u->flags & IO_U_F_FREE_DEF))
+			put_io_u(td, io_u);
 	}
 }
 
 /*
  * Complete a single io_u for the sync engines.
  */
-long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
+int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
+		       unsigned long *bytes)
 {
 	struct io_completion_data icd;
 
-	init_icd(&icd, 1);
+	init_icd(td, &icd, 1);
 	io_completed(td, io_u, &icd);
-	put_io_u(td, io_u);
 
-	if (!icd.error)
-		return icd.bytes_done[0] + icd.bytes_done[1];
+	if (!(io_u->flags & IO_U_F_FREE_DEF))
+		put_io_u(td, io_u);
+
+	if (icd.error) {
+		td_verror(td, icd.error, "io_u_sync_complete");
+		return -1;
+	}
+
+	if (bytes) {
+		bytes[0] += icd.bytes_done[0];
+		bytes[1] += icd.bytes_done[1];
+	}
 
-	td_verror(td, icd.error, "io_u_sync_complete");
-	return -1;
+	return 0;
 }
 
 /*
  * Called to complete min_events number of io for the async engines.
  */
-long io_u_queued_complete(struct thread_data *td, int min_events)
+int io_u_queued_complete(struct thread_data *td, int min_evts,
+			 unsigned long *bytes)
 {
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
 	int ret;
 	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events);
+	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
 
-	if (!min_events)
+	if (!min_evts)
 		tvp = &ts;
 
-	ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
+	ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
 	if (ret < 0) {
 		td_verror(td, -ret, "td_io_getevents");
 		return ret;
 	} else if (!ret)
 		return ret;
 
-	init_icd(&icd, ret);
+	init_icd(td, &icd, ret);
 	ios_completed(td, &icd);
-	if (!icd.error)
-		return icd.bytes_done[0] + icd.bytes_done[1];
+	if (icd.error) {
+		td_verror(td, icd.error, "io_u_queued_complete");
+		return -1;
+	}
+
+	if (bytes) {
+		bytes[0] += icd.bytes_done[0];
+		bytes[1] += icd.bytes_done[1];
+	}
 
-	td_verror(td, icd.error, "io_u_queued_complete");
-	return -1;
+	return 0;
 }
 
 /*
@@ -928,88 +1240,22 @@ long io_u_queued_complete(struct thread_data *td, int min_events)
  */
 void io_u_queued(struct thread_data *td, struct io_u *io_u)
 {
-	unsigned long slat_time;
-
-	slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
-	add_slat_sample(td, io_u->ddir, slat_time);
-}
-
-#ifdef FIO_USE_TIMEOUT
-void io_u_set_timeout(struct thread_data *td)
-{
-	assert(td->cur_depth);
-
-	td->timer.it_interval.tv_sec = 0;
-	td->timer.it_interval.tv_usec = 0;
-	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
-	td->timer.it_value.tv_usec = 0;
-	setitimer(ITIMER_REAL, &td->timer, NULL);
-	fio_gettime(&td->timeout_end, NULL);
-}
-
-static void io_u_dump(struct io_u *io_u)
-{
-	unsigned long t_start = mtime_since_now(&io_u->start_time);
-	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
-
-	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
-	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
-	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
-}
-#else
-void io_u_set_timeout(struct thread_data fio_unused *td)
-{
-}
-#endif
-
-#ifdef FIO_USE_TIMEOUT
-static void io_u_timeout_handler(int fio_unused sig)
-{
-	struct thread_data *td, *__td;
-	pid_t pid = getpid();
-	struct list_head *entry;
-	struct io_u *io_u;
-	int i;
-
-	log_err("fio: io_u timeout\n");
-
-	/*
-	 * TLS would be nice...
-	 */
-	td = NULL;
-	for_each_td(__td, i) {
-		if (__td->pid == pid) {
-			td = __td;
-			break;
-		}
-	}
-
-	if (!td) {
-		log_err("fio: io_u timeout, can't find job\n");
-		exit(1);
-	}
-
-	if (!td->cur_depth) {
-		log_err("fio: timeout without pending work?\n");
-		return;
-	}
-
-	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
+	if (!td->o.disable_slat) {
+		unsigned long slat_time;
 
-	list_for_each(entry, &td->io_u_busylist) {
-		io_u = list_entry(entry, struct io_u, list);
-
-		io_u_dump(io_u);
+		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
+		add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen);
 	}
-
-	td_verror(td, ETIMEDOUT, "io_u timeout");
-	exit(1);
 }
-#endif
 
-void io_u_init_timeout(void)
+/*
+ * "randomly" fill the buffer contents
+ */
+void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
+		      unsigned int max_bs)
 {
-#ifdef FIO_USE_TIMEOUT
-	signal(SIGALRM, io_u_timeout_handler);
-#endif
+	if (!td->o.zero_buffers)
+		fill_random_buf(io_u->buf, max_bs);
+	else
+		memset(io_u->buf, 0, max_bs);
 }