X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=fc49bd325879f01d63c53efaaff25cc0e34e63e2;hp=1077c27b1f07a43014ea527de0f985037a0a2083;hb=306ddc9752eef70c3fbb111af63d197a3a6d447f;hpb=8fd15a9a8d64deae4c2f13b02e1df7e23ea3cafb

diff --git a/io_u.c b/io_u.c
index 1077c27b..fc49bd32 100644
--- a/io_u.c
+++ b/io_u.c
@@ -6,12 +6,7 @@
 #include <assert.h>
 
 #include "fio.h"
-#include "os.h"
-
-/*
- * Change this define to play with the timeout handling
- */
-#undef FIO_USE_TIMEOUT
+#include "hash.h"
 
 struct io_completion_data {
 	int nr;				/* input */
@@ -25,13 +20,14 @@ 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,
-			   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);
 
-	return (f->file_map[idx] & (1UL << bit)) == 0;
+	dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit);
+
+	return (f->file_map[idx] & (1 << bit)) == 0;
 }
 
 /*
@@ -42,50 +38,87 @@ 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 / (unsigned long long) min_bs;
-	blocks = 0;
+	block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs;
 	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 && !random_map_free(td, f, block))
+		if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) {
+			if (!blocks)
+				blocks = 1;
 			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;
+
+		if (this_blocks == BLOCKS_PER_MAP)
+			mask = -1U;
+		else
+			mask = ((1U << this_blocks) - 1) << bit;
+
+		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 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 = max_size / (unsigned long long) td->o.ba[ddir];
+	if (!max_blocks)
+		return 0;
+
+	return max_blocks;
+}
+
 /*
  * Return the next free block in the map.
  */
 static int get_next_free_block(struct thread_data *td, struct fio_file *f,
-			       unsigned long long *b)
+			       enum fio_ddir ddir, unsigned long long *b)
 {
+	unsigned long long min_bs = td->o.rw_min_bs;
 	int i;
 
 	i = f->last_free_lookup;
 	*b = (i * BLOCKS_PER_MAP);
-	while ((*b) * td->o.rw_min_bs < f->real_file_size) {
-		if (f->file_map[i] != -1UL) {
+	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;
 			return 0;
 		}
@@ -94,36 +127,58 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
 		i++;
 	}
 
+	dprint(FD_IO, "failed finding a free block\n");
 	return 1;
 }
 
 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
-				int ddir, unsigned long long *b)
+				enum fio_ddir ddir, unsigned long long *b)
 {
-	unsigned long long max_blocks = f->file_size / td->o.min_bs[ddir];
-	unsigned long long r, rb;
+	unsigned long long r;
 	int loops = 5;
 
 	do {
 		r = os_random_long(&td->random_state);
-		if (!max_blocks)
-			*b = 0;
-		else
-			*b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-		if (td->o.norandommap)
-			break;
-		rb = *b + (f->file_offset / td->o.min_bs[ddir]);
-		loops--;
-	} while (!random_map_free(td, f, rb) && loops);
+		dprint(FD_RANDOM, "off rand %llu\n", r);
+		*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 (!file_randommap(td, f))
+			return 0;
+
+		/*
+		 * calculate map offset and check if it's free
+		 */
+		if (random_map_free(f, *b))
+			return 0;
+
+		dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n",
+									*b);
+	} while (--loops);
 
 	/*
-	 * if we failed to retrieve a truly random offset within
-	 * the loops assigned, see if there are free ones left at all
+	 * we get here, if we didn't suceed in looking up a block. generate
+	 * a random start offset into the filemap, and find the first free
+	 * block from there.
 	 */
-	if (!loops && get_next_free_block(td, f, b))
-		return 1;
+	loops = 10;
+	do {
+		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;
 
-	return 0;
+		r = os_random_long(&td->random_state);
+	} while (--loops);
+
+	/*
+	 * that didn't work either, try exhaustive search from the start
+	 */
+	f->last_free_lookup = 0;
+	return get_next_free_block(td, f, ddir, b);
 }
 
 /*
@@ -134,51 +189,92 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
 static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f = io_u->file;
-	const int ddir = io_u->ddir;
 	unsigned long long b;
+	enum fio_ddir ddir = io_u->ddir;
 
 	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
-		b = f->last_pos / td->o.min_bs[ddir];
+		}
+	} else {
+		if (f->last_pos >= f->real_file_size) {
+			if (!td_random(td) ||
+			     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.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 = (b * td->o.min_bs[ddir]) + f->file_offset;
-	if (io_u->offset >= f->real_file_size)
+	io_u->offset += f->file_offset;
+	if (io_u->offset >= f->real_file_size) {
+		dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n",
+					io_u->offset, f->real_file_size);
 		return 1;
+	}
 
 	return 0;
 }
 
+static inline int is_power_of_2(unsigned int val)
+{
+	return (val != 0 && ((val & (val - 1)) == 0));
+}
+
 static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
-	struct fio_file *f = io_u->file;
 	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);
-		buflen = (unsigned int) (1 + (double) (td->o.max_bs[ddir] - 1) * r / (RAND_MAX + 1.0));
-		if (!td->o.bs_unaligned)
-			buflen = (buflen + td->o.min_bs[ddir] - 1) & ~(td->o.min_bs[ddir] - 1);
-	}
-
-	while (buflen + io_u->offset > f->real_file_size) {
-		if (buflen == td->o.min_bs[ddir]) {
-			if (!td->o.odirect) {
-				assert(io_u->offset <= f->real_file_size);
-				buflen = f->real_file_size - io_u->offset;
-				return buflen;
+		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[ddir]; i++) {
+				struct bssplit *bsp = &td->o.bssplit[ddir][i];
+
+				buflen = bsp->bs;
+				perc += bsp->perc;
+				if (r <= ((OS_RAND_MAX / 100L) * perc))
+					break;
 			}
-			return 0;
 		}
+		if (!td->o.bs_unaligned && is_power_of_2(minbs))
+			buflen = (buflen + minbs - 1) & ~(minbs - 1);
+	}
 
-		buflen = td->o.min_bs[ddir];
+	if (io_u->offset + buflen > io_u->file->real_file_size) {
+		dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
+						minbs, ddir);
+		buflen = minbs;
 	}
 
 	return buflen;
@@ -186,7 +282,6 @@ static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
 
 static void set_rwmix_bytes(struct thread_data *td)
 {
-	unsigned long long rbytes;
 	unsigned int diff;
 
 	/*
@@ -194,10 +289,8 @@ 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)
@@ -206,8 +299,8 @@ 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;
@@ -221,27 +314,12 @@ 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;
-
-		fio_gettime(&now, NULL);
-	 	elapsed = mtime_since_now(&td->rwmix_switch);
-
-		/*
-		 * if this is the first cycle, make it shorter
-		 */
-		cycle = td->o.rwmixcycle;
-		if (!td->rwmix_bytes)
-			cycle /= 10;
-
 		/*
 		 * Check if it's time to seed a new data direction.
 		 */
-		if (elapsed >= cycle ||
-		    td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
+		if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) {
 			unsigned long long max_bytes;
-			enum fio_ddir ddir;			
+			enum fio_ddir ddir;
 
 			/*
 			 * Put a top limit on how many bytes we do for
@@ -250,11 +328,13 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 			 */
 			ddir = get_rand_ddir(td);
 			max_bytes = td->this_io_bytes[ddir];
-			if (max_bytes >= (td->io_size * td->o.rwmix[ddir] / 100)) {
+			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));
+					fio_gettime(&td->rw_end[ddir], NULL);
 				}
+
 				ddir ^= 1;
 			}
 
@@ -262,7 +342,6 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 				set_rwmix_bytes(td);
 
 			td->rwmix_ddir = ddir;
-			memcpy(&td->rwmix_switch, &now, sizeof(now));
 		}
 		return td->rwmix_ddir;
 	} else if (td_read(td))
@@ -271,14 +350,25 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 		return DDIR_WRITE;
 }
 
+static 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;
 
+	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);
+	flist_del(&io_u->list);
+	flist_add(&io_u->list, &td->io_u_freelist);
 	td->cur_depth--;
 }
 
@@ -286,22 +376,24 @@ 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);
+
 	__io_u->flags |= IO_U_F_FREE;
+	if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC))
+		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);
+	flist_del(&__io_u->list);
+	flist_add_tail(&__io_u->list, &td->io_u_requeues);
 	td->cur_depth--;
 	*io_u = NULL;
 }
 
 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 {
-	/*
-	 * If using an iolog, grab next piece if any available.
-	 */
-	if (td->o.read_iolog)
-		return read_iolog_get(td, io_u);
+	if (td->io_ops->flags & FIO_NOIO)
+		goto out;
 
 	/*
 	 * see if it's time to sync
@@ -310,97 +402,220 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
 	     td->io_issues[DDIR_WRITE] && should_fsync(td)) {
 		io_u->ddir = DDIR_SYNC;
-		return 0;
+		goto out;
 	}
 
 	io_u->ddir = get_rw_ddir(td);
 
+	/*
+	 * See if it's time to switch to a new zone
+	 */
+	if (td->zone_bytes >= td->o.zone_size) {
+		td->zone_bytes = 0;
+		io_u->file->last_pos += td->o.zone_skip;
+		td->io_skip_bytes += td->o.zone_skip;
+	}
+
 	/*
 	 * No log, let the seq/rand engine retrieve the next buflen and
 	 * position.
 	 */
-	if (get_next_offset(td, io_u))
+	if (get_next_offset(td, io_u)) {
+		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
 		return 1;
+	}
 
 	io_u->buflen = get_next_buflen(td, io_u);
-	if (!io_u->buflen)
+	if (!io_u->buflen) {
+		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
 		return 1;
+	}
+
+	if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
+		dprint(FD_IO, "io_u %p, offset too large\n", io_u);
+		dprint(FD_IO, "  off=%llu/%lu > %llu\n", io_u->offset,
+				io_u->buflen, io_u->file->real_file_size);
+		return 1;
+	}
 
 	/*
 	 * mark entry before potentially trimming io_u
 	 */
-	if (td_random(td) && !td->o.norandommap)
+	if (td_random(td) && file_randommap(td, io_u->file))
 		mark_random_map(td, io_u);
 
 	/*
 	 * If using a write iolog, store this entry.
 	 */
-	if (td->o.write_iolog_file)
-		write_iolog_put(td, io_u);
-
+out:
+	dprint_io_u(io_u, "fill_io_u");
+	td->zone_bytes += io_u->buflen;
+	log_io_u(td, io_u);
 	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:
-		index++;
+		index = 6;
+		break;
 	case 32 ... 63:
-		index++;
+		index = 5;
+		break;
 	case 16 ... 31:
-		index++;
+		index = 4;
+		break;
 	case 8 ... 15:
-		index++;
+		index = 3;
+		break;
 	case 4 ... 7:
-		index++;
+		index = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		index = 1;
 	case 1:
 		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_latency(struct thread_data *td, unsigned long msec)
+static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
+{
+	int index = 0;
+
+	assert(usec < 1000);
+
+	switch (usec) {
+	case 750 ... 999:
+		index = 9;
+		break;
+	case 500 ... 749:
+		index = 8;
+		break;
+	case 250 ... 499:
+		index = 7;
+		break;
+	case 100 ... 249:
+		index = 6;
+		break;
+	case 50 ... 99:
+		index = 5;
+		break;
+	case 20 ... 49:
+		index = 4;
+		break;
+	case 10 ... 19:
+		index = 3;
+		break;
+	case 4 ... 9:
+		index = 2;
+		break;
+	case 2 ... 3:
+		index = 1;
+	case 0 ... 1:
+		break;
+	}
+
+	assert(index < FIO_IO_U_LAT_U_NR);
+	td->ts.io_u_lat_u[index]++;
+}
+
+static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
 {
 	int index = 0;
 
 	switch (msec) {
 	default:
-		index++;
+		index = 11;
+		break;
 	case 1000 ... 1999:
-		index++;
+		index = 10;
+		break;
 	case 750 ... 999:
-		index++;
+		index = 9;
+		break;
 	case 500 ... 749:
-		index++;
+		index = 8;
+		break;
 	case 250 ... 499:
-		index++;
+		index = 7;
+		break;
 	case 100 ... 249:
-		index++;
+		index = 6;
+		break;
 	case 50 ... 99:
-		index++;
+		index = 5;
+		break;
 	case 20 ... 49:
-		index++;
+		index = 4;
+		break;
 	case 10 ... 19:
-		index++;
+		index = 3;
+		break;
 	case 4 ... 9:
-		index++;
+		index = 2;
+		break;
 	case 2 ... 3:
-		index++;
+		index = 1;
 	case 0 ... 1:
 		break;
 	}
 
-	td->ts.io_u_lat[index]++;
+	assert(index < FIO_IO_U_LAT_M_NR);
+	td->ts.io_u_lat_m[index]++;
+}
+
+static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
+{
+	if (usec < 1000)
+		io_u_mark_lat_usec(td, usec);
+	else
+		io_u_mark_lat_msec(td, usec / 1000);
 }
 
 /*
@@ -414,12 +629,29 @@ static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf,
 
 	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];
+		fno = (unsigned int) ((double) td->o.nr_files
+			* (r / (OS_RAND_MAX + 1.0)));
+		f = td->files[fno];
+		if (f->flags & FIO_FILE_DONE)
+			continue;
 
-		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
+		if (!(f->flags & FIO_FILE_OPEN)) {
+			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);
 }
 
@@ -433,18 +665,43 @@ 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;
 
+		dprint(FD_FILE, "trying file %s %x\n", f->file_name, f->flags);
+		if (f->flags & FIO_FILE_DONE) {
+			f = NULL;
+			continue;
+		}
+
+		if (!(f->flags & FIO_FILE_OPEN)) {
+			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);
+				continue;
+			}
+			opened = 1;
+		}
+
+		dprint(FD_FILE, "goodf=%x, badf=%x, ff=%x\n", goodf, badf, f->flags);
 		if ((!goodf || (f->flags & goodf)) && !(f->flags & badf))
 			break;
 
+		if (opened)
+			td_io_close_file(td, f);
+
 		f = NULL;
 	} while (td->next_file != old_next_file);
 
+	dprint(FD_FILE, "get_next_file_rr: %p\n", f);
 	return f;
 }
 
@@ -454,43 +711,70 @@ static struct fio_file *get_next_file(struct thread_data *td)
 
 	assert(td->o.nr_files <= td->files_index);
 
-	if (!td->nr_open_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--)
-		return f;
+	if (f && (f->flags & FIO_FILE_OPEN) && !(f->flags & FIO_FILE_CLOSING)) {
+		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)
+	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);
 
 	td->file_service_file = f;
 	td->file_service_left = td->file_service_nr - 1;
+out:
+	dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
 	return f;
 }
 
-static struct fio_file *find_next_new_file(struct thread_data *td)
+static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 {
 	struct fio_file *f;
 
-	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);
+	do {
+		f = get_next_file(td);
+		if (!f)
+			return 1;
 
-	return f;
+		io_u->file = f;
+		get_file(f);
+
+		if (!fill_io_u(td, io_u))
+			break;
+
+		put_file_log(td, f);
+		td_io_close_file(td, f);
+		io_u->file = NULL;
+		f->flags |= FIO_FILE_DONE;
+		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;
 }
 
+
 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);
+	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;
@@ -503,8 +787,8 @@ struct io_u *__get_io_u(struct thread_data *td)
 		io_u->flags &= ~IO_U_F_FREE;
 
 		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++;
 	}
 
@@ -519,11 +803,12 @@ struct io_u *get_io_u(struct thread_data *td)
 {
 	struct fio_file *f;
 	struct io_u *io_u;
-	int ret;
 
 	io_u = __get_io_u(td);
-	if (!io_u)
+	if (!io_u) {
+		dprint(FD_IO, "__get_io_u failed\n");
 		return NULL;
+	}
 
 	/*
 	 * from a requeue, io_u already setup
@@ -531,81 +816,51 @@ struct io_u *get_io_u(struct thread_data *td)
 	if (io_u->file)
 		goto out;
 
-	do {
-		f = get_next_file(td);
-		if (!f) {
-			put_io_u(td, io_u);
-			return NULL;
-		}
-
-set_file:
-		io_u->file = f;
-
-		if (!fill_io_u(td, io_u))
-			break;
-
-		/*
-		 * No more to do for this file, close it
-		 */
-		io_u->file = NULL;
-		td_io_close_file(td, f);
-
-		/*
-		 * probably not the right place to do this, but see
-		 * if we need to open a new file
-		 */
-		if (td->nr_open_files < td->o.open_files &&
-		    td->o.open_files != td->o.nr_files) {
-			f = find_next_new_file(td);
-
-			if (!f || (ret = td_io_open_file(td, f))) {
-				put_io_u(td, io_u);
-				return NULL;
-			}
-			goto set_file;
-		}
-	} while (1);
-
-	if (td->zone_bytes >= td->o.zone_size) {
-		td->zone_bytes = 0;
-		f->last_pos += td->o.zone_skip;
+	/*
+	 * If using an iolog, grab next piece if any available.
+	 */
+	if (td->o.read_iolog_file) {
+		if (read_iolog_get(td, io_u))
+			goto err_put;
+	} else if (set_io_u_file(td, io_u)) {
+		dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
+		goto err_put;
 	}
 
-	if (io_u->buflen + io_u->offset > f->real_file_size) {
-		if (td->io_ops->flags & FIO_RAWIO) {
-			put_io_u(td, io_u);
-			return NULL;
-		}
-
-		io_u->buflen = f->real_file_size - io_u->offset;
-	}
+	f = io_u->file;
+	assert(f->flags & FIO_FILE_OPEN);
 
 	if (io_u->ddir != DDIR_SYNC) {
-		if (!io_u->buflen) {
-			put_io_u(td, io_u);
-			return NULL;
+		if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
+			dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
+			goto err_put;
 		}
 
 		f->last_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);
 	}
 
 	/*
 	 * Set io data pointers.
 	 */
-out:
+	io_u->endpos = io_u->offset + io_u->buflen;
 	io_u->xfer_buf = io_u->buf;
 	io_u->xfer_buflen = io_u->buflen;
 
-	if (td_io_prep(td, io_u)) {
-		put_io_u(td, io_u);
-		return NULL;
+out:
+	if (!td_io_prep(td, io_u)) {
+		if (!td->o.disable_slat)
+			fio_gettime(&io_u->start_time, NULL);
+		return io_u;
 	}
-
-	fio_gettime(&io_u->start_time, NULL);
-	return io_u;
+err_put:
+	dprint(FD_IO, "get_io_u failed\n");
+	put_io_u(td, io_u);
+	return NULL;
 }
 
 void io_u_log_error(struct thread_data *td, struct io_u *io_u)
@@ -619,7 +874,8 @@ 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");
@@ -628,13 +884,17 @@ 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 msec;
+	/*
+	 * Older gcc's are too dumb to realize that usec is always used
+	 * initialized, silence that warning.
+	 */
+	unsigned long uninitialized_var(usec);
+
+	dprint_io_u(io_u, "io complete");
 
 	assert(io_u->flags & IO_U_F_FLIGHT);
 	io_u->flags &= ~IO_U_F_FLIGHT;
 
-	put_file(td, io_u->file);
-
 	if (io_u->ddir == DDIR_SYNC) {
 		td->last_was_sync = 1;
 		return;
@@ -649,18 +909,23 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 
 		td->io_blocks[idx]++;
 		td->io_bytes[idx] += bytes;
-		td->zone_bytes += bytes;
 		td->this_io_bytes[idx] += bytes;
 
-		io_u->file->last_completed_pos = io_u->offset + io_u->buflen;
-
-		msec = mtime_since(&io_u->issue_time, &icd->time);
+		if (ramp_time_over(td)) {
+			if (!td->o.disable_clat || !td->o.disable_bw)
+				usec = utime_since(&io_u->issue_time,
+							&icd->time);
 
-		add_clat_sample(td, idx, msec);
-		add_bw_sample(td, idx, &icd->time);
-		io_u_mark_latency(td, msec);
+			if (!td->o.disable_clat) {
+				add_clat_sample(td, idx, usec, bytes);
+				io_u_mark_latency(td, usec);
+			}
+			if (!td->o.disable_bw)
+				add_bw_sample(td, idx, bytes, &icd->time);
+		}
 
-		if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
+		if (td_write(td) && idx == DDIR_WRITE &&
+		    td->o.do_verify &&
 		    td->o.verify != VERIFY_NONE)
 			log_io_piece(td, io_u);
 
@@ -677,9 +942,11 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	}
 }
 
-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;
 
@@ -708,7 +975,7 @@ long io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 {
 	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);
 
@@ -722,24 +989,26 @@ long 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.
  */
-long io_u_queued_complete(struct thread_data *td, int min_events)
+long io_u_queued_complete(struct thread_data *td, int min_evts)
 {
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
 	int ret;
 	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-	if (!min_events)
+	dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
+
+	if (!min_evts)
 		tvp = &ts;
 
-	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];
@@ -753,88 +1022,27 @@ long io_u_queued_complete(struct thread_data *td, int min_events)
  */
 void io_u_queued(struct thread_data *td, struct io_u *io_u)
 {
-	unsigned long slat_time;
+	if (!td->o.disable_slat) {
+		unsigned long slat_time;
 
-	slat_time = mtime_since(&io_u->start_time, &io_u->issue_time);
-	add_slat_sample(td, io_u->ddir, slat_time);
-}
-
-#ifdef FIO_USE_TIMEOUT
-void io_u_set_timeout(struct thread_data *td)
-{
-	assert(td->cur_depth);
-
-	td->timer.it_interval.tv_sec = 0;
-	td->timer.it_interval.tv_usec = 0;
-	td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC;
-	td->timer.it_value.tv_usec = 0;
-	setitimer(ITIMER_REAL, &td->timer, NULL);
-	fio_gettime(&td->timeout_end, NULL);
-}
-
-static void io_u_dump(struct io_u *io_u)
-{
-	unsigned long t_start = mtime_since_now(&io_u->start_time);
-	unsigned long t_issue = mtime_since_now(&io_u->issue_time);
-
-	log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue);
-	log_err("  buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, io_u->xfer_buf, io_u->buflen, io_u->xfer_buflen, io_u->offset);
-	log_err("  ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name);
-}
-#else
-void io_u_set_timeout(struct thread_data fio_unused *td)
-{
+		slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
+		add_slat_sample(td, io_u->ddir, io_u->xfer_buflen, slat_time);
+	}
 }
-#endif
 
-#ifdef FIO_USE_TIMEOUT
-static void io_u_timeout_handler(int fio_unused sig)
+/*
+ * "randomly" fill the buffer contents
+ */
+void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
+		      unsigned int max_bs)
 {
-	struct thread_data *td, *__td;
-	pid_t pid = getpid();
-	struct list_head *entry;
-	struct io_u *io_u;
-	int i;
-
-	log_err("fio: io_u timeout\n");
+	long *ptr = io_u->buf;
 
-	/*
-	 * TLS would be nice...
-	 */
-	td = NULL;
-	for_each_td(__td, i) {
-		if (__td->pid == pid) {
-			td = __td;
-			break;
+	if (!td->o.zero_buffers) {
+		while ((void *) ptr - io_u->buf < max_bs) {
+			*ptr = rand() * GOLDEN_RATIO_PRIME;
+			ptr++;
 		}
-	}
-
-	if (!td) {
-		log_err("fio: io_u timeout, can't find job\n");
-		exit(1);
-	}
-
-	if (!td->cur_depth) {
-		log_err("fio: timeout without pending work?\n");
-		return;
-	}
-
-	log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid);
-
-	list_for_each(entry, &td->io_u_busylist) {
-		io_u = list_entry(entry, struct io_u, list);
-
-		io_u_dump(io_u);
-	}
-
-	td_verror(td, ETIMEDOUT, "io_u timeout");
-	exit(1);
-}
-#endif
-
-void io_u_init_timeout(void)
-{
-#ifdef FIO_USE_TIMEOUT
-	signal(SIGALRM, io_u_timeout_handler);
-#endif
+	} else
+		memset(ptr, 0, max_bs);
 }