X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=zbd.c;h=d7e91e37e010c0194da07636b29ce058ebd2db13;hp=42487780ca77e5cf7f67665545f47c856bbd2919;hb=04ba61dfa67784d4dfcc22a2b3de7ede28e22e40;hpb=bfbdd35b3e8f7de1bf1f48e7087c04a6b37e9c61

diff --git a/zbd.c b/zbd.c
index 42487780..d7e91e37 100644
--- a/zbd.c
+++ b/zbd.c
@@ -31,10 +31,10 @@ static uint32_t zbd_zone_idx(const struct fio_file *f, uint64_t offset)
 {
 	uint32_t zone_idx;
 
-	if (f->zbd_info->zone_size_log2)
+	if (f->zbd_info->zone_size_log2 > 0)
 		zone_idx = offset >> f->zbd_info->zone_size_log2;
 	else
-		zone_idx = (offset >> 9) / f->zbd_info->zone_size;
+		zone_idx = offset / f->zbd_info->zone_size;
 
 	return min(zone_idx, f->zbd_info->nr_zones);
 }
@@ -53,7 +53,7 @@ static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
 	assert((required & 511) == 0);
 
 	return z->type == BLK_ZONE_TYPE_SEQWRITE_REQ &&
-		z->wp + (required >> 9) > z->start + f->zbd_info->zone_size;
+		z->wp + required > z->start + f->zbd_info->zone_size;
 }
 
 static bool is_valid_offset(const struct fio_file *f, uint64_t offset)
@@ -121,31 +121,31 @@ static bool zbd_verify_sizes(void)
 				continue;
 			zone_idx = zbd_zone_idx(f, f->file_offset);
 			z = &f->zbd_info->zone_info[zone_idx];
-			if (f->file_offset != (z->start << 9)) {
-				new_offset = (z+1)->start << 9;
+			if (f->file_offset != z->start) {
+				new_offset = (z+1)->start;
 				if (new_offset >= f->file_offset + f->io_size) {
 					log_info("%s: io_size must be at least one zone\n",
 						 f->file_name);
 					return false;
 				}
-				log_info("%s: rounded up offset from %lu to %lu\n",
-					 f->file_name, f->file_offset,
-					 new_offset);
+				log_info("%s: rounded up offset from %llu to %llu\n",
+					 f->file_name, (unsigned long long) f->file_offset,
+					 (unsigned long long) new_offset);
 				f->io_size -= (new_offset - f->file_offset);
 				f->file_offset = new_offset;
 			}
 			zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size);
 			z = &f->zbd_info->zone_info[zone_idx];
-			new_end = z->start << 9;
+			new_end = z->start;
 			if (f->file_offset + f->io_size != new_end) {
 				if (new_end <= f->file_offset) {
 					log_info("%s: io_size must be at least one zone\n",
 						 f->file_name);
 					return false;
 				}
-				log_info("%s: rounded down io_size from %lu to %lu\n",
-					 f->file_name, f->io_size,
-					 new_end - f->file_offset);
+				log_info("%s: rounded down io_size from %llu to %llu\n",
+					 f->file_name, (unsigned long long) f->io_size,
+					 (unsigned long long) new_end - f->file_offset);
 				f->io_size = new_end - f->file_offset;
 			}
 		}
@@ -168,10 +168,10 @@ static bool zbd_verify_bs(void)
 			zone_size = f->zbd_info->zone_size;
 			for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
 				if (td->o.verify != VERIFY_NONE &&
-				    (zone_size << 9) % td->o.bs[k] != 0) {
+				    zone_size % td->o.bs[k] != 0) {
 					log_info("%s: block size %llu is not a divisor of the zone size %d\n",
 						 f->file_name, td->o.bs[k],
-						 zone_size << 9);
+						 zone_size);
 					return false;
 				}
 			}
@@ -186,11 +186,14 @@ static bool zbd_verify_bs(void)
  * size of @buf.
  *
  * Returns 0 upon success and a negative error code upon failure.
+ * If the zone report is empty, always assume an error (device problem) and
+ * return -EIO.
  */
 static int read_zone_info(int fd, uint64_t start_sector,
 			  void *buf, unsigned int bufsz)
 {
 	struct blk_zone_report *hdr = buf;
+	int ret;
 
 	if (bufsz < sizeof(*hdr))
 		return -EINVAL;
@@ -199,7 +202,12 @@ static int read_zone_info(int fd, uint64_t start_sector,
 
 	hdr->nr_zones = (bufsz - sizeof(*hdr)) / sizeof(struct blk_zone);
 	hdr->sector = start_sector;
-	return ioctl(fd, BLKREPORTZONE, hdr) >= 0 ? 0 : -errno;
+	ret = ioctl(fd, BLKREPORTZONE, hdr);
+	if (ret)
+		return -errno;
+	if (!hdr->nr_zones)
+		return -EIO;
+	return 0;
 }
 
 /*
@@ -228,12 +236,45 @@ static enum blk_zoned_model get_zbd_model(const char *file_name)
 	char *zoned_attr_path = NULL;
 	char *model_str = NULL;
 	struct stat statbuf;
+	char *sys_devno_path = NULL;
+	char *part_attr_path = NULL;
+	char *part_str = NULL;
+	char sys_path[PATH_MAX];
+	ssize_t sz;
+	char *delim = NULL;
 
 	if (stat(file_name, &statbuf) < 0)
 		goto out;
-	if (asprintf(&zoned_attr_path, "/sys/dev/block/%d:%d/queue/zoned",
+
+	if (asprintf(&sys_devno_path, "/sys/dev/block/%d:%d",
 		     major(statbuf.st_rdev), minor(statbuf.st_rdev)) < 0)
 		goto out;
+
+	sz = readlink(sys_devno_path, sys_path, sizeof(sys_path) - 1);
+	if (sz < 0)
+		goto out;
+	sys_path[sz] = '\0';
+
+	/*
+	 * If the device is a partition device, cut the device name in the
+	 * canonical sysfs path to obtain the sysfs path of the holder device.
+	 *   e.g.:  /sys/devices/.../sda/sda1 -> /sys/devices/.../sda
+	 */
+	if (asprintf(&part_attr_path, "/sys/dev/block/%s/partition",
+		     sys_path) < 0)
+		goto out;
+	part_str = read_file(part_attr_path);
+	if (part_str && *part_str == '1') {
+		delim = strrchr(sys_path, '/');
+		if (!delim)
+			goto out;
+		*delim = '\0';
+	}
+
+	if (asprintf(&zoned_attr_path,
+		     "/sys/dev/block/%s/queue/zoned", sys_path) < 0)
+		goto out;
+
 	model_str = read_file(zoned_attr_path);
 	if (!model_str)
 		goto out;
@@ -246,6 +287,9 @@ static enum blk_zoned_model get_zbd_model(const char *file_name)
 out:
 	free(model_str);
 	free(zoned_attr_path);
+	free(part_str);
+	free(part_attr_path);
+	free(sys_devno_path);
 	return model;
 }
 
@@ -273,9 +317,9 @@ static int init_zone_info(struct thread_data *td, struct fio_file *f)
 	pthread_mutexattr_t attr;
 	int i;
 
-	zone_size = td->o.zone_size >> 9;
+	zone_size = td->o.zone_size;
 	assert(zone_size);
-	nr_zones = ((f->real_file_size >> 9) + zone_size - 1) / zone_size;
+	nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
 	zbd_info = scalloc(1, sizeof(*zbd_info) +
 			   (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
 	if (!zbd_info)
@@ -300,7 +344,7 @@ static int init_zone_info(struct thread_data *td, struct fio_file *f)
 	f->zbd_info = zbd_info;
 	f->zbd_info->zone_size = zone_size;
 	f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
-		ilog2(zone_size) + 9 : -1;
+		ilog2(zone_size) : -1;
 	f->zbd_info->nr_zones = nr_zones;
 	pthread_mutexattr_destroy(&attr);
 	return 0;
@@ -351,20 +395,21 @@ static int parse_zone_info(struct thread_data *td, struct fio_file *f)
 		goto close;
 	}
 	z = (void *)(hdr + 1);
-	zone_size = z->len;
-	nr_zones = ((f->real_file_size >> 9) + zone_size - 1) / zone_size;
+	zone_size = z->len << 9;
+	nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
 
 	if (td->o.zone_size == 0) {
-		td->o.zone_size = zone_size << 9;
-	} else if (td->o.zone_size != zone_size << 9) {
-		log_info("fio: %s job parameter zonesize %lld does not match disk zone size %ld.\n",
-			 f->file_name, td->o.zone_size, zone_size << 9);
+		td->o.zone_size = zone_size;
+	} else if (td->o.zone_size != zone_size) {
+		log_info("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n",
+			 f->file_name, (unsigned long long) td->o.zone_size,
+			(unsigned long long) zone_size);
 		ret = -EINVAL;
 		goto close;
 	}
 
-	dprint(FD_ZBD, "Device %s has %d zones of size %lu KB\n", f->file_name,
-	       nr_zones, zone_size / 2);
+	dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name,
+	       nr_zones, (unsigned long long) zone_size / 1024);
 
 	zbd_info = scalloc(1, sizeof(*zbd_info) +
 			   (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
@@ -378,18 +423,16 @@ static int parse_zone_info(struct thread_data *td, struct fio_file *f)
 		z = (void *)(hdr + 1);
 		for (i = 0; i < hdr->nr_zones; i++, j++, z++, p++) {
 			pthread_mutex_init(&p->mutex, &attr);
-			p->start = z->start;
+			p->start = z->start << 9;
 			switch (z->cond) {
 			case BLK_ZONE_COND_NOT_WP:
-				p->wp = z->start;
-				break;
 			case BLK_ZONE_COND_FULL:
-				p->wp = z->start + zone_size;
+				p->wp = p->start + zone_size;
 				break;
 			default:
 				assert(z->start <= z->wp);
-				assert(z->wp <= z->start + zone_size);
-				p->wp = z->wp;
+				assert(z->wp <= z->start + (zone_size >> 9));
+				p->wp = z->wp << 9;
 				break;
 			}
 			p->type = z->type;
@@ -407,18 +450,18 @@ static int parse_zone_info(struct thread_data *td, struct fio_file *f)
 			break;
 		ret = read_zone_info(fd, start_sector, buf, bufsz);
 		if (ret < 0) {
-			log_info("fio: BLKREPORTZONE(%lu) failed for %s (%d).\n",
-				 start_sector, f->file_name, -ret);
+			log_info("fio: BLKREPORTZONE(%llu) failed for %s (%d).\n",
+				 (unsigned long long) start_sector, f->file_name, -ret);
 			goto close;
 		}
 	}
 	/* a sentinel */
-	zbd_info->zone_info[nr_zones].start = start_sector;
+	zbd_info->zone_info[nr_zones].start = start_sector << 9;
 
 	f->zbd_info = zbd_info;
 	f->zbd_info->zone_size = zone_size;
 	f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
-		ilog2(zone_size) + 9 : -1;
+		ilog2(zone_size) : -1;
 	f->zbd_info->nr_zones = nr_zones;
 	zbd_info = NULL;
 	ret = 0;
@@ -556,18 +599,18 @@ int zbd_init(struct thread_data *td)
  * Returns 0 upon success and a negative error code upon failure.
  */
 static int zbd_reset_range(struct thread_data *td, const struct fio_file *f,
-			   uint64_t sector, uint64_t nr_sectors)
+			   uint64_t offset, uint64_t length)
 {
 	struct blk_zone_range zr = {
-		.sector         = sector,
-		.nr_sectors     = nr_sectors,
+		.sector         = offset >> 9,
+		.nr_sectors     = length >> 9,
 	};
 	uint32_t zone_idx_b, zone_idx_e;
 	struct fio_zone_info *zb, *ze, *z;
 	int ret = 0;
 
 	assert(f->fd != -1);
-	assert(is_valid_offset(f, ((sector + nr_sectors) << 9) - 1));
+	assert(is_valid_offset(f, offset + length - 1));
 	switch (f->zbd_info->model) {
 	case ZBD_DM_HOST_AWARE:
 	case ZBD_DM_HOST_MANAGED:
@@ -583,20 +626,31 @@ static int zbd_reset_range(struct thread_data *td, const struct fio_file *f,
 		break;
 	}
 
-	zone_idx_b = zbd_zone_idx(f, sector << 9);
+	zone_idx_b = zbd_zone_idx(f, offset);
 	zb = &f->zbd_info->zone_info[zone_idx_b];
-	zone_idx_e = zbd_zone_idx(f, (sector + nr_sectors) << 9);
+	zone_idx_e = zbd_zone_idx(f, offset + length);
 	ze = &f->zbd_info->zone_info[zone_idx_e];
 	for (z = zb; z < ze; z++) {
 		pthread_mutex_lock(&z->mutex);
+		pthread_mutex_lock(&f->zbd_info->mutex);
+		f->zbd_info->sectors_with_data -= z->wp - z->start;
+		pthread_mutex_unlock(&f->zbd_info->mutex);
 		z->wp = z->start;
 		z->verify_block = 0;
 		pthread_mutex_unlock(&z->mutex);
 	}
 
+	td->ts.nr_zone_resets += ze - zb;
+
 	return ret;
 }
 
+static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info,
+				struct fio_zone_info *zone)
+{
+	return zone - zbd_info->zone_info;
+}
+
 /**
  * zbd_reset_zone - reset the write pointer of a single zone
  * @td: FIO thread data.
@@ -608,12 +662,10 @@ static int zbd_reset_range(struct thread_data *td, const struct fio_file *f,
 static int zbd_reset_zone(struct thread_data *td, const struct fio_file *f,
 			  struct fio_zone_info *z)
 {
-	int ret;
+	dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
+		zbd_zone_nr(f->zbd_info, z));
 
-	dprint(FD_ZBD, "%s: resetting wp of zone %lu.\n", f->file_name,
-	       z - f->zbd_info->zone_info);
-	ret = zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
-	return ret;
+	return zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
 }
 
 /*
@@ -630,12 +682,12 @@ static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
 			   struct fio_zone_info *const ze, bool all_zones)
 {
 	struct fio_zone_info *z, *start_z = ze;
-	const uint32_t min_bs = td->o.min_bs[DDIR_WRITE] >> 9;
+	const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
 	bool reset_wp;
 	int res = 0;
 
-	dprint(FD_ZBD, "%s: examining zones %lu .. %lu\n", f->file_name,
-	       zb - f->zbd_info->zone_info, ze - f->zbd_info->zone_info);
+	dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name,
+		zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze));
 	assert(f->fd != -1);
 	for (z = zb; z < ze; z++) {
 		pthread_mutex_lock(&z->mutex);
@@ -648,10 +700,10 @@ static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
 				start_z = z;
 			} else if (start_z < ze && !reset_wp) {
 				dprint(FD_ZBD,
-				       "%s: resetting zones %lu .. %lu\n",
+				       "%s: resetting zones %u .. %u\n",
 				       f->file_name,
-				       start_z - f->zbd_info->zone_info,
-				       z - f->zbd_info->zone_info);
+					zbd_zone_nr(f->zbd_info, start_z),
+					zbd_zone_nr(f->zbd_info, z));
 				if (zbd_reset_range(td, f, start_z->start,
 						z->start - start_z->start) < 0)
 					res = 1;
@@ -661,9 +713,9 @@ static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
 		default:
 			if (start_z == ze)
 				break;
-			dprint(FD_ZBD, "%s: resetting zones %lu .. %lu\n",
-			       f->file_name, start_z - f->zbd_info->zone_info,
-			       z - f->zbd_info->zone_info);
+			dprint(FD_ZBD, "%s: resetting zones %u .. %u\n",
+			       f->file_name, zbd_zone_nr(f->zbd_info, start_z),
+			       zbd_zone_nr(f->zbd_info, z));
 			if (zbd_reset_range(td, f, start_z->start,
 					    z->start - start_z->start) < 0)
 				res = 1;
@@ -672,9 +724,9 @@ static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
 		}
 	}
 	if (start_z < ze) {
-		dprint(FD_ZBD, "%s: resetting zones %lu .. %lu\n", f->file_name,
-		       start_z - f->zbd_info->zone_info,
-		       z - f->zbd_info->zone_info);
+		dprint(FD_ZBD, "%s: resetting zones %u .. %u\n", f->file_name,
+			zbd_zone_nr(f->zbd_info, start_z),
+			zbd_zone_nr(f->zbd_info, z));
 		if (zbd_reset_range(td, f, start_z->start,
 				    z->start - start_z->start) < 0)
 			res = 1;
@@ -685,6 +737,95 @@ static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
 	return res;
 }
 
+/*
+ * Reset zbd_info.write_cnt, the counter that counts down towards the next
+ * zone reset.
+ */
+static void zbd_reset_write_cnt(const struct thread_data *td,
+				const struct fio_file *f)
+{
+	assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
+
+	pthread_mutex_lock(&f->zbd_info->mutex);
+	f->zbd_info->write_cnt = td->o.zrf.u.f ?
+		min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
+	pthread_mutex_unlock(&f->zbd_info->mutex);
+}
+
+static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
+					const struct fio_file *f)
+{
+	uint32_t write_cnt = 0;
+
+	pthread_mutex_lock(&f->zbd_info->mutex);
+	assert(f->zbd_info->write_cnt);
+	if (f->zbd_info->write_cnt)
+		write_cnt = --f->zbd_info->write_cnt;
+	if (write_cnt == 0)
+		zbd_reset_write_cnt(td, f);
+	pthread_mutex_unlock(&f->zbd_info->mutex);
+
+	return write_cnt == 0;
+}
+
+enum swd_action {
+	CHECK_SWD,
+	SET_SWD,
+};
+
+/* Calculate the number of sectors with data (swd) and perform action 'a' */
+static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a)
+{
+	struct fio_zone_info *zb, *ze, *z;
+	uint64_t swd = 0;
+
+	zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
+	ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset +
+						  f->io_size)];
+	for (z = zb; z < ze; z++) {
+		pthread_mutex_lock(&z->mutex);
+		swd += z->wp - z->start;
+	}
+	pthread_mutex_lock(&f->zbd_info->mutex);
+	switch (a) {
+	case CHECK_SWD:
+		assert(f->zbd_info->sectors_with_data == swd);
+		break;
+	case SET_SWD:
+		f->zbd_info->sectors_with_data = swd;
+		break;
+	}
+	pthread_mutex_unlock(&f->zbd_info->mutex);
+	for (z = zb; z < ze; z++)
+		pthread_mutex_unlock(&z->mutex);
+
+	return swd;
+}
+
+/*
+ * The swd check is useful for debugging but takes too much time to leave
+ * it enabled all the time. Hence it is disabled by default.
+ */
+static const bool enable_check_swd = false;
+
+/* Check whether the value of zbd_info.sectors_with_data is correct. */
+static void zbd_check_swd(const struct fio_file *f)
+{
+	if (!enable_check_swd)
+		return;
+
+	zbd_process_swd(f, CHECK_SWD);
+}
+
+static void zbd_init_swd(struct fio_file *f)
+{
+	uint64_t swd;
+
+	swd = zbd_process_swd(f, SET_SWD);
+	dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name,
+	       swd);
+}
+
 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
 {
 	struct fio_zone_info *zb, *ze;
@@ -696,6 +837,7 @@ void zbd_file_reset(struct thread_data *td, struct fio_file *f)
 	zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
 	zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size);
 	ze = &f->zbd_info->zone_info[zone_idx_e];
+	zbd_init_swd(f);
 	/*
 	 * If data verification is enabled reset the affected zones before
 	 * writing any data to avoid that a zone reset has to be issued while
@@ -704,19 +846,231 @@ void zbd_file_reset(struct thread_data *td, struct fio_file *f)
 	zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE &&
 			(td->o.td_ddir & TD_DDIR_WRITE) &&
 			td->runstate != TD_VERIFYING);
+	zbd_reset_write_cnt(td, f);
+}
+
+/* The caller must hold f->zbd_info->mutex. */
+static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
+			 unsigned int zone_idx)
+{
+	struct zoned_block_device_info *zbdi = f->zbd_info;
+	int i;
+
+	assert(td->o.max_open_zones <= ARRAY_SIZE(zbdi->open_zones));
+	assert(zbdi->num_open_zones <= td->o.max_open_zones);
+
+	for (i = 0; i < zbdi->num_open_zones; i++)
+		if (zbdi->open_zones[i] == zone_idx)
+			return true;
+
+	return false;
+}
+
+/*
+ * Open a ZBD zone if it was not yet open. Returns true if either the zone was
+ * already open or if opening a new zone is allowed. Returns false if the zone
+ * was not yet open and opening a new zone would cause the zone limit to be
+ * exceeded.
+ */
+static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
+			  uint32_t zone_idx)
+{
+	const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
+	const struct fio_file *f = io_u->file;
+	struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
+	bool res = true;
+
+	if (z->cond == BLK_ZONE_COND_OFFLINE)
+		return false;
+
+	/*
+	 * Skip full zones with data verification enabled because resetting a
+	 * zone causes data loss and hence causes verification to fail.
+	 */
+	if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
+		return false;
+
+	/* Zero means no limit */
+	if (!td->o.max_open_zones)
+		return true;
+
+	pthread_mutex_lock(&f->zbd_info->mutex);
+	if (is_zone_open(td, f, zone_idx))
+		goto out;
+	res = false;
+	if (f->zbd_info->num_open_zones >= td->o.max_open_zones)
+		goto out;
+	dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx);
+	f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx;
+	z->open = 1;
+	res = true;
+
+out:
+	pthread_mutex_unlock(&f->zbd_info->mutex);
+	return res;
+}
+
+/* The caller must hold f->zbd_info->mutex */
+static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
+			   unsigned int open_zone_idx)
+{
+	uint32_t zone_idx;
+
+	assert(open_zone_idx < f->zbd_info->num_open_zones);
+	zone_idx = f->zbd_info->open_zones[open_zone_idx];
+	memmove(f->zbd_info->open_zones + open_zone_idx,
+		f->zbd_info->open_zones + open_zone_idx + 1,
+		(FIO_MAX_OPEN_ZBD_ZONES - (open_zone_idx + 1)) *
+		sizeof(f->zbd_info->open_zones[0]));
+	f->zbd_info->num_open_zones--;
+	f->zbd_info->zone_info[zone_idx].open = 0;
+}
+
+/*
+ * Modify the offset of an I/O unit that does not refer to an open zone such
+ * that it refers to an open zone. Close an open zone and open a new zone if
+ * necessary. This algorithm can only work correctly if all write pointers are
+ * a multiple of the fio block size. The caller must neither hold z->mutex
+ * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
+ */
+struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
+					       struct io_u *io_u)
+{
+	const uint32_t min_bs = td->o.min_bs[io_u->ddir];
+	const struct fio_file *f = io_u->file;
+	struct fio_zone_info *z;
+	unsigned int open_zone_idx = -1;
+	uint32_t zone_idx, new_zone_idx;
+	int i;
+
+	assert(is_valid_offset(f, io_u->offset));
+
+	if (td->o.max_open_zones) {
+		/*
+		 * This statement accesses f->zbd_info->open_zones[] on purpose
+		 * without locking.
+		 */
+		zone_idx = f->zbd_info->open_zones[(io_u->offset -
+						    f->file_offset) *
+				f->zbd_info->num_open_zones / f->io_size];
+	} else {
+		zone_idx = zbd_zone_idx(f, io_u->offset);
+	}
+	dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
+	       __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
+
+	/*
+	 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
+	 * lock it can happen that the state of the zone with index zone_idx
+	 * has changed after 'z' has been assigned and before f->zbd_info->mutex
+	 * has been obtained. Hence the loop.
+	 */
+	for (;;) {
+		z = &f->zbd_info->zone_info[zone_idx];
+
+		pthread_mutex_lock(&z->mutex);
+		pthread_mutex_lock(&f->zbd_info->mutex);
+		if (td->o.max_open_zones == 0)
+			goto examine_zone;
+		if (f->zbd_info->num_open_zones == 0) {
+			pthread_mutex_unlock(&f->zbd_info->mutex);
+			pthread_mutex_unlock(&z->mutex);
+			dprint(FD_ZBD, "%s(%s): no zones are open\n",
+			       __func__, f->file_name);
+			return NULL;
+		}
+		open_zone_idx = (io_u->offset - f->file_offset) *
+			f->zbd_info->num_open_zones / f->io_size;
+		assert(open_zone_idx < f->zbd_info->num_open_zones);
+		new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
+		if (new_zone_idx == zone_idx)
+			break;
+		zone_idx = new_zone_idx;
+		pthread_mutex_unlock(&f->zbd_info->mutex);
+		pthread_mutex_unlock(&z->mutex);
+	}
+
+	/* Both z->mutex and f->zbd_info->mutex are held. */
+
+examine_zone:
+	if (z->wp + min_bs <= (z+1)->start) {
+		pthread_mutex_unlock(&f->zbd_info->mutex);
+		goto out;
+	}
+	dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name,
+	       zone_idx);
+	if (td->o.max_open_zones)
+		zbd_close_zone(td, f, open_zone_idx);
+	pthread_mutex_unlock(&f->zbd_info->mutex);
+
+	/* Only z->mutex is held. */
+
+	/* Zone 'z' is full, so try to open a new zone. */
+	for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
+		zone_idx++;
+		pthread_mutex_unlock(&z->mutex);
+		z++;
+		if (!is_valid_offset(f, z->start)) {
+			/* Wrap-around. */
+			zone_idx = zbd_zone_idx(f, f->file_offset);
+			z = &f->zbd_info->zone_info[zone_idx];
+		}
+		assert(is_valid_offset(f, z->start));
+		pthread_mutex_lock(&z->mutex);
+		if (z->open)
+			continue;
+		if (zbd_open_zone(td, io_u, zone_idx))
+			goto out;
+	}
+
+	/* Only z->mutex is held. */
+
+	/* Check whether the write fits in any of the already opened zones. */
+	pthread_mutex_lock(&f->zbd_info->mutex);
+	for (i = 0; i < f->zbd_info->num_open_zones; i++) {
+		zone_idx = f->zbd_info->open_zones[i];
+		pthread_mutex_unlock(&f->zbd_info->mutex);
+		pthread_mutex_unlock(&z->mutex);
+
+		z = &f->zbd_info->zone_info[zone_idx];
+
+		pthread_mutex_lock(&z->mutex);
+		if (z->wp + min_bs <= (z+1)->start)
+			goto out;
+		pthread_mutex_lock(&f->zbd_info->mutex);
+	}
+	pthread_mutex_unlock(&f->zbd_info->mutex);
+	pthread_mutex_unlock(&z->mutex);
+	dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
+	       f->file_name);
+	return NULL;
+
+out:
+	dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
+	       zone_idx);
+	io_u->offset = z->start;
+	return z;
 }
 
 /* The caller must hold z->mutex. */
-static void zbd_replay_write_order(struct thread_data *td, struct io_u *io_u,
-				   struct fio_zone_info *z)
+static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
+						    struct io_u *io_u,
+						    struct fio_zone_info *z)
 {
 	const struct fio_file *f = io_u->file;
 	const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
 
+	if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
+		pthread_mutex_unlock(&z->mutex);
+		z = zbd_convert_to_open_zone(td, io_u);
+		assert(z);
+	}
+
 	if (z->verify_block * min_bs >= f->zbd_info->zone_size)
-		log_err("%s: %d * %d >= %ld\n", f->file_name, z->verify_block,
-			min_bs, f->zbd_info->zone_size);
-	io_u->offset = (z->start << 9) + z->verify_block++ * min_bs;
+		log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
+			min_bs, (unsigned long long) f->zbd_info->zone_size);
+	io_u->offset = z->start + z->verify_block++ * min_bs;
+	return z;
 }
 
 /*
@@ -744,7 +1098,7 @@ zbd_find_zone(struct thread_data *td, struct io_u *io_u,
 	for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
 		if (z1 < zl && z1->cond != BLK_ZONE_COND_OFFLINE) {
 			pthread_mutex_lock(&z1->mutex);
-			if (z1->start + (min_bs >> 9) <= z1->wp)
+			if (z1->start + min_bs <= z1->wp)
 				return z1;
 			pthread_mutex_unlock(&z1->mutex);
 		} else if (!td_random(td)) {
@@ -753,7 +1107,7 @@ zbd_find_zone(struct thread_data *td, struct io_u *io_u,
 		if (td_random(td) && z2 >= zf &&
 		    z2->cond != BLK_ZONE_COND_OFFLINE) {
 			pthread_mutex_lock(&z2->mutex);
-			if (z2->start + (min_bs >> 9) <= z2->wp)
+			if (z2->start + min_bs <= z2->wp)
 				return z2;
 			pthread_mutex_unlock(&z2->mutex);
 		}
@@ -763,37 +1117,52 @@ zbd_find_zone(struct thread_data *td, struct io_u *io_u,
 	return NULL;
 }
 
-
 /**
- * zbd_post_submit - update the write pointer and unlock the zone lock
+ * zbd_queue_io - update the write pointer of a sequential zone
  * @io_u: I/O unit
- * @success: Whether or not the I/O unit has been executed successfully
+ * @success: Whether or not the I/O unit has been queued successfully
+ * @q: queueing status (busy, completed or queued).
  *
- * For write and trim operations, update the write pointer of all affected
- * zones.
+ * For write and trim operations, update the write pointer of the I/O unit
+ * target zone.
  */
-static void zbd_post_submit(const struct io_u *io_u, bool success)
+static void zbd_queue_io(struct io_u *io_u, int q, bool success)
 {
-	struct zoned_block_device_info *zbd_info;
+	const struct fio_file *f = io_u->file;
+	struct zoned_block_device_info *zbd_info = f->zbd_info;
 	struct fio_zone_info *z;
 	uint32_t zone_idx;
-	uint64_t end, zone_end;
+	uint64_t zone_end;
 
-	zbd_info = io_u->file->zbd_info;
 	if (!zbd_info)
 		return;
 
-	zone_idx = zbd_zone_idx(io_u->file, io_u->offset);
-	end = (io_u->offset + io_u->buflen) >> 9;
-	z = &zbd_info->zone_info[zone_idx];
+	zone_idx = zbd_zone_idx(f, io_u->offset);
 	assert(zone_idx < zbd_info->nr_zones);
+	z = &zbd_info->zone_info[zone_idx];
+
 	if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ)
 		return;
+
 	if (!success)
 		goto unlock;
+
+	dprint(FD_ZBD,
+	       "%s: queued I/O (%lld, %llu) for zone %u\n",
+	       f->file_name, io_u->offset, io_u->buflen, zone_idx);
+
 	switch (io_u->ddir) {
 	case DDIR_WRITE:
-		zone_end = min(end, (z + 1)->start);
+		zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
+			       (z + 1)->start);
+		pthread_mutex_lock(&zbd_info->mutex);
+		/*
+		 * z->wp > zone_end means that one or more I/O errors
+		 * have occurred.
+		 */
+		if (z->wp <= zone_end)
+			zbd_info->sectors_with_data += zone_end - z->wp;
+		pthread_mutex_unlock(&zbd_info->mutex);
 		z->wp = zone_end;
 		break;
 	case DDIR_TRIM:
@@ -802,8 +1171,42 @@ static void zbd_post_submit(const struct io_u *io_u, bool success)
 	default:
 		break;
 	}
+
 unlock:
-	pthread_mutex_unlock(&z->mutex);
+	if (!success || q != FIO_Q_QUEUED) {
+		/* BUSY or COMPLETED: unlock the zone */
+		pthread_mutex_unlock(&z->mutex);
+		io_u->zbd_put_io = NULL;
+	}
+}
+
+/**
+ * zbd_put_io - Unlock an I/O unit target zone lock
+ * @io_u: I/O unit
+ */
+static void zbd_put_io(const struct io_u *io_u)
+{
+	const struct fio_file *f = io_u->file;
+	struct zoned_block_device_info *zbd_info = f->zbd_info;
+	struct fio_zone_info *z;
+	uint32_t zone_idx;
+
+	if (!zbd_info)
+		return;
+
+	zone_idx = zbd_zone_idx(f, io_u->offset);
+	assert(zone_idx < zbd_info->nr_zones);
+	z = &zbd_info->zone_info[zone_idx];
+
+	if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ)
+		return;
+
+	dprint(FD_ZBD,
+	       "%s: terminate I/O (%lld, %llu) for zone %u\n",
+	       f->file_name, io_u->offset, io_u->buflen, zone_idx);
+
+	assert(pthread_mutex_unlock(&z->mutex) == 0);
+	zbd_check_swd(f);
 }
 
 bool zbd_unaligned_write(int error_code)
@@ -829,7 +1232,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 {
 	const struct fio_file *f = io_u->file;
 	uint32_t zone_idx_b;
-	struct fio_zone_info *zb, *zl;
+	struct fio_zone_info *zb, *zl, *orig_zb;
 	uint32_t orig_len = io_u->buflen;
 	uint32_t min_bs = td->o.min_bs[io_u->ddir];
 	uint64_t new_len;
@@ -842,6 +1245,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 	assert(io_u->buflen);
 	zone_idx_b = zbd_zone_idx(f, io_u->offset);
 	zb = &f->zbd_info->zone_info[zone_idx_b];
+	orig_zb = zb;
 
 	/* Accept the I/O offset for conventional zones. */
 	if (zb->type == BLK_ZONE_TYPE_CONVENTIONAL)
@@ -855,29 +1259,38 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 	    io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
 		return io_u_accept;
 
-	pthread_mutex_lock(&zb->mutex);
+	zbd_check_swd(f);
+
+	/*
+	 * Lock the io_u target zone. The zone will be unlocked if io_u offset
+	 * is changed or when io_u completes and zbd_put_io() executed.
+	 * To avoid multiple jobs doing asynchronous I/Os from deadlocking each
+	 * other waiting for zone locks when building an io_u batch, first
+	 * only trylock the zone. If the zone is already locked by another job,
+	 * process the currently queued I/Os so that I/O progress is made and
+	 * zones unlocked.
+	 */
+	if (pthread_mutex_trylock(&zb->mutex) != 0) {
+		if (!td_ioengine_flagged(td, FIO_SYNCIO))
+			io_u_quiesce(td);
+		pthread_mutex_lock(&zb->mutex);
+	}
+
 	switch (io_u->ddir) {
 	case DDIR_READ:
 		if (td->runstate == TD_VERIFYING) {
-			zbd_replay_write_order(td, io_u, zb);
+			zb = zbd_replay_write_order(td, io_u, zb);
 			goto accept;
 		}
 		/*
-		 * Avoid reads past the write pointer because such reads do not
-		 * hit the medium.
+		 * Check that there is enough written data in the zone to do an
+		 * I/O of at least min_bs B. If there isn't, find a new zone for
+		 * the I/O.
 		 */
 		range = zb->cond != BLK_ZONE_COND_OFFLINE ?
-			((zb->wp - zb->start) << 9) - io_u->buflen : 0;
-		if (td_random(td) && range >= 0) {
-			io_u->offset = (zb->start << 9) +
-				((io_u->offset - (zb->start << 9)) %
-				 (range + 1)) / min_bs * min_bs;
-			assert(zb->start << 9 <= io_u->offset);
-			assert(io_u->offset + io_u->buflen <= zb->wp << 9);
-			goto accept;
-		}
-		if (zb->cond == BLK_ZONE_COND_OFFLINE ||
-		    (io_u->offset + io_u->buflen) >> 9 > zb->wp) {
+			zb->wp - zb->start : 0;
+		if (range < min_bs ||
+		    ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
 			pthread_mutex_unlock(&zb->mutex);
 			zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
 						f->file_offset + f->io_size)];
@@ -889,18 +1302,58 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 				       io_u->buflen);
 				goto eof;
 			}
-			io_u->offset = zb->start << 9;
+			/*
+			 * zbd_find_zone() returned a zone with a range of at
+			 * least min_bs.
+			 */
+			range = zb->wp - zb->start;
+			assert(range >= min_bs);
+
+			if (!td_random(td))
+				io_u->offset = zb->start;
 		}
-		if ((io_u->offset + io_u->buflen) >> 9 > zb->wp) {
-			dprint(FD_ZBD, "%s: %lld + %lld > %" PRIu64 "\n",
-			       f->file_name, io_u->offset, io_u->buflen,
-			       zb->wp);
-			goto eof;
+		/*
+		 * Make sure the I/O is within the zone valid data range while
+		 * maximizing the I/O size and preserving randomness.
+		 */
+		if (range <= io_u->buflen)
+			io_u->offset = zb->start;
+		else if (td_random(td))
+			io_u->offset = zb->start +
+				((io_u->offset - orig_zb->start) %
+				 (range - io_u->buflen)) / min_bs * min_bs;
+		/*
+		 * Make sure the I/O does not cross over the zone wp position.
+		 */
+		new_len = min((unsigned long long)io_u->buflen,
+			      (unsigned long long)(zb->wp - io_u->offset));
+		new_len = new_len / min_bs * min_bs;
+		if (new_len < io_u->buflen) {
+			io_u->buflen = new_len;
+			dprint(FD_IO, "Changed length from %u into %llu\n",
+			       orig_len, io_u->buflen);
 		}
+		assert(zb->start <= io_u->offset);
+		assert(io_u->offset + io_u->buflen <= zb->wp);
 		goto accept;
 	case DDIR_WRITE:
-		if (io_u->buflen > (f->zbd_info->zone_size << 9))
+		if (io_u->buflen > f->zbd_info->zone_size)
 			goto eof;
+		if (!zbd_open_zone(td, io_u, zone_idx_b)) {
+			pthread_mutex_unlock(&zb->mutex);
+			zb = zbd_convert_to_open_zone(td, io_u);
+			if (!zb)
+				goto eof;
+			zone_idx_b = zb - f->zbd_info->zone_info;
+		}
+		/* Check whether the zone reset threshold has been exceeded */
+		if (td->o.zrf.u.f) {
+			if (f->zbd_info->sectors_with_data >=
+			    f->io_size * td->o.zrt.u.f &&
+			    zbd_dec_and_reset_write_cnt(td, f)) {
+				zb->reset_zone = 1;
+			}
+		}
 		/* Reset the zone pointer if necessary */
 		if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
 			assert(td->o.verify == VERIFY_NONE);
@@ -918,7 +1371,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 		}
 		/* Make writes occur at the write pointer */
 		assert(!zbd_zone_full(f, zb, min_bs));
-		io_u->offset = zb->wp << 9;
+		io_u->offset = zb->wp;
 		if (!is_valid_offset(f, io_u->offset)) {
 			dprint(FD_ZBD, "Dropped request with offset %llu\n",
 			       io_u->offset);
@@ -930,7 +1383,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 		 * small.
 		 */
 		new_len = min((unsigned long long)io_u->buflen,
-			      ((zb + 1)->start << 9) - io_u->offset);
+			      (zb + 1)->start - io_u->offset);
 		new_len = new_len / min_bs * min_bs;
 		if (new_len == io_u->buflen)
 			goto accept;
@@ -941,7 +1394,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 			goto accept;
 		}
 		log_err("Zone remainder %lld smaller than minimum block size %d\n",
-			(((zb + 1)->start << 9) - io_u->offset),
+			((zb + 1)->start - io_u->offset),
 			min_bs);
 		goto eof;
 	case DDIR_TRIM:
@@ -960,8 +1413,10 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
 accept:
 	assert(zb);
 	assert(zb->cond != BLK_ZONE_COND_OFFLINE);
-	assert(!io_u->post_submit);
-	io_u->post_submit = zbd_post_submit;
+	assert(!io_u->zbd_queue_io);
+	assert(!io_u->zbd_put_io);
+	io_u->zbd_queue_io = zbd_queue_io;
+	io_u->zbd_put_io = zbd_put_io;
 	return io_u_accept;
 
 eof:
@@ -969,3 +1424,13 @@ eof:
 		pthread_mutex_unlock(&zb->mutex);
 	return io_u_eof;
 }
+
+/* Return a string with ZBD statistics */
+char *zbd_write_status(const struct thread_stat *ts)
+{
+	char *res;
+
+	if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0)
+		return NULL;
+	return res;
+}