2 * Copyright (C) 2018 Western Digital Corporation or its affiliates.
4 * This file is released under the GPL.
12 #include <sys/ioctl.h>
15 #include <linux/blkzoned.h>
25 * zbd_zone_idx - convert an offset into a zone number
27 * @offset: offset in bytes. If this offset is in the first zone_size bytes
28 * past the disk size then the index of the sentinel is returned.
30 static uint32_t zbd_zone_idx(const struct fio_file *f, uint64_t offset)
34 if (f->zbd_info->zone_size_log2 > 0)
35 zone_idx = offset >> f->zbd_info->zone_size_log2;
37 zone_idx = offset / f->zbd_info->zone_size;
39 return min(zone_idx, f->zbd_info->nr_zones);
43 * zbd_zone_full - verify whether a minimum number of bytes remain in a zone
45 * @z: zone info pointer.
46 * @required: minimum number of bytes that must remain in a zone.
48 * The caller must hold z->mutex.
50 static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
53 assert((required & 511) == 0);
55 return z->type == BLK_ZONE_TYPE_SEQWRITE_REQ &&
56 z->wp + required > z->start + f->zbd_info->zone_size;
59 static bool is_valid_offset(const struct fio_file *f, uint64_t offset)
61 return (uint64_t)(offset - f->file_offset) < f->io_size;
64 /* Verify whether direct I/O is used for all host-managed zoned drives. */
65 static bool zbd_using_direct_io(void)
67 struct thread_data *td;
72 if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE))
74 for_each_file(td, f, j) {
76 f->zbd_info->model == ZBD_DM_HOST_MANAGED)
84 /* Whether or not the I/O range for f includes one or more sequential zones */
85 static bool zbd_is_seq_job(struct fio_file *f)
87 uint32_t zone_idx, zone_idx_b, zone_idx_e;
92 zone_idx_b = zbd_zone_idx(f, f->file_offset);
93 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size - 1);
94 for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++)
95 if (f->zbd_info->zone_info[zone_idx].type ==
96 BLK_ZONE_TYPE_SEQWRITE_REQ)
103 * Verify whether offset and size parameters are aligned with zone boundaries.
105 static bool zbd_verify_sizes(void)
107 const struct fio_zone_info *z;
108 struct thread_data *td;
110 uint64_t new_offset, new_end;
115 for_each_file(td, f, j) {
118 if (f->file_offset >= f->real_file_size)
120 if (!zbd_is_seq_job(f))
122 zone_idx = zbd_zone_idx(f, f->file_offset);
123 z = &f->zbd_info->zone_info[zone_idx];
124 if (f->file_offset != z->start) {
125 new_offset = (z+1)->start;
126 if (new_offset >= f->file_offset + f->io_size) {
127 log_info("%s: io_size must be at least one zone\n",
131 log_info("%s: rounded up offset from %llu to %llu\n",
132 f->file_name, (unsigned long long) f->file_offset,
133 (unsigned long long) new_offset);
134 f->io_size -= (new_offset - f->file_offset);
135 f->file_offset = new_offset;
137 zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size);
138 z = &f->zbd_info->zone_info[zone_idx];
140 if (f->file_offset + f->io_size != new_end) {
141 if (new_end <= f->file_offset) {
142 log_info("%s: io_size must be at least one zone\n",
146 log_info("%s: rounded down io_size from %llu to %llu\n",
147 f->file_name, (unsigned long long) f->io_size,
148 (unsigned long long) new_end - f->file_offset);
149 f->io_size = new_end - f->file_offset;
157 static bool zbd_verify_bs(void)
159 struct thread_data *td;
165 for_each_file(td, f, j) {
168 zone_size = f->zbd_info->zone_size;
169 for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
170 if (td->o.verify != VERIFY_NONE &&
171 zone_size % td->o.bs[k] != 0) {
172 log_info("%s: block size %llu is not a divisor of the zone size %d\n",
173 f->file_name, td->o.bs[k],
184 * Read zone information into @buf starting from sector @start_sector.
185 * @fd is a file descriptor that refers to a block device and @bufsz is the
188 * Returns 0 upon success and a negative error code upon failure.
190 static int read_zone_info(int fd, uint64_t start_sector,
191 void *buf, unsigned int bufsz)
193 struct blk_zone_report *hdr = buf;
195 if (bufsz < sizeof(*hdr))
198 memset(hdr, 0, sizeof(*hdr));
200 hdr->nr_zones = (bufsz - sizeof(*hdr)) / sizeof(struct blk_zone);
201 hdr->sector = start_sector;
202 return ioctl(fd, BLKREPORTZONE, hdr) >= 0 ? 0 : -errno;
206 * Read up to 255 characters from the first line of a file. Strip the trailing
209 static char *read_file(const char *path)
211 char line[256], *p = line;
214 f = fopen(path, "rb");
217 if (!fgets(line, sizeof(line), f))
225 static enum blk_zoned_model get_zbd_model(const char *file_name)
227 enum blk_zoned_model model = ZBD_DM_NONE;
228 char *zoned_attr_path = NULL;
229 char *model_str = NULL;
231 char *sys_devno_path = NULL;
232 char *part_attr_path = NULL;
233 char *part_str = NULL;
234 char sys_path[PATH_MAX];
238 if (stat(file_name, &statbuf) < 0)
241 if (asprintf(&sys_devno_path, "/sys/dev/block/%d:%d",
242 major(statbuf.st_rdev), minor(statbuf.st_rdev)) < 0)
245 sz = readlink(sys_devno_path, sys_path, sizeof(sys_path) - 1);
251 * If the device is a partition device, cut the device name in the
252 * canonical sysfs path to obtain the sysfs path of the holder device.
253 * e.g.: /sys/devices/.../sda/sda1 -> /sys/devices/.../sda
255 if (asprintf(&part_attr_path, "/sys/dev/block/%s/partition",
258 part_str = read_file(part_attr_path);
259 if (part_str && *part_str == '1') {
260 delim = strrchr(sys_path, '/');
266 if (asprintf(&zoned_attr_path,
267 "/sys/dev/block/%s/queue/zoned", sys_path) < 0)
270 model_str = read_file(zoned_attr_path);
273 dprint(FD_ZBD, "%s: zbd model string: %s\n", file_name, model_str);
274 if (strcmp(model_str, "host-aware") == 0)
275 model = ZBD_DM_HOST_AWARE;
276 else if (strcmp(model_str, "host-managed") == 0)
277 model = ZBD_DM_HOST_MANAGED;
281 free(zoned_attr_path);
283 free(part_attr_path);
284 free(sys_devno_path);
288 static int ilog2(uint64_t i)
300 * Initialize f->zbd_info for devices that are not zoned block devices. This
301 * allows to execute a ZBD workload against a non-ZBD device.
303 static int init_zone_info(struct thread_data *td, struct fio_file *f)
306 struct fio_zone_info *p;
308 struct zoned_block_device_info *zbd_info = NULL;
309 pthread_mutexattr_t attr;
312 zone_size = td->o.zone_size;
314 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
315 zbd_info = scalloc(1, sizeof(*zbd_info) +
316 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
320 pthread_mutexattr_init(&attr);
321 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
322 pthread_mutexattr_setpshared(&attr, true);
323 pthread_mutex_init(&zbd_info->mutex, &attr);
324 zbd_info->refcount = 1;
325 p = &zbd_info->zone_info[0];
326 for (i = 0; i < nr_zones; i++, p++) {
327 pthread_mutex_init(&p->mutex, &attr);
328 p->start = i * zone_size;
329 p->wp = p->start + zone_size;
330 p->type = BLK_ZONE_TYPE_SEQWRITE_REQ;
331 p->cond = BLK_ZONE_COND_EMPTY;
334 p->start = nr_zones * zone_size;
336 f->zbd_info = zbd_info;
337 f->zbd_info->zone_size = zone_size;
338 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
339 ilog2(zone_size) : -1;
340 f->zbd_info->nr_zones = nr_zones;
341 pthread_mutexattr_destroy(&attr);
346 * Parse the BLKREPORTZONE output and store it in f->zbd_info. Must be called
347 * only for devices that support this ioctl, namely zoned block devices.
349 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
351 const unsigned int bufsz = sizeof(struct blk_zone_report) +
352 4096 * sizeof(struct blk_zone);
354 struct blk_zone_report *hdr;
355 const struct blk_zone *z;
356 struct fio_zone_info *p;
357 uint64_t zone_size, start_sector;
358 struct zoned_block_device_info *zbd_info = NULL;
359 pthread_mutexattr_t attr;
361 int fd, i, j, ret = 0;
363 pthread_mutexattr_init(&attr);
364 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
365 pthread_mutexattr_setpshared(&attr, true);
371 fd = open(f->file_name, O_RDONLY | O_LARGEFILE);
377 ret = read_zone_info(fd, 0, buf, bufsz);
379 log_info("fio: BLKREPORTZONE(%lu) failed for %s (%d).\n",
380 0UL, f->file_name, -ret);
384 if (hdr->nr_zones < 1) {
385 log_info("fio: %s has invalid zone information.\n",
389 z = (void *)(hdr + 1);
390 zone_size = z->len << 9;
391 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
393 if (td->o.zone_size == 0) {
394 td->o.zone_size = zone_size;
395 } else if (td->o.zone_size != zone_size) {
396 log_info("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n",
397 f->file_name, (unsigned long long) td->o.zone_size,
398 (unsigned long long) zone_size);
403 dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name,
404 nr_zones, (unsigned long long) zone_size / 1024);
406 zbd_info = scalloc(1, sizeof(*zbd_info) +
407 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
411 pthread_mutex_init(&zbd_info->mutex, &attr);
412 zbd_info->refcount = 1;
413 p = &zbd_info->zone_info[0];
414 for (start_sector = 0, j = 0; j < nr_zones;) {
415 z = (void *)(hdr + 1);
416 for (i = 0; i < hdr->nr_zones; i++, j++, z++, p++) {
417 pthread_mutex_init(&p->mutex, &attr);
418 p->start = z->start << 9;
420 case BLK_ZONE_COND_NOT_WP:
423 case BLK_ZONE_COND_FULL:
424 p->wp = p->start + zone_size;
427 assert(z->start <= z->wp);
428 assert(z->wp <= z->start + (zone_size >> 9));
434 if (j > 0 && p->start != p[-1].start + zone_size) {
435 log_info("%s: invalid zone data\n",
442 start_sector = z->start + z->len;
445 ret = read_zone_info(fd, start_sector, buf, bufsz);
447 log_info("fio: BLKREPORTZONE(%llu) failed for %s (%d).\n",
448 (unsigned long long) start_sector, f->file_name, -ret);
453 zbd_info->zone_info[nr_zones].start = start_sector << 9;
455 f->zbd_info = zbd_info;
456 f->zbd_info->zone_size = zone_size;
457 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
458 ilog2(zone_size) : -1;
459 f->zbd_info->nr_zones = nr_zones;
469 pthread_mutexattr_destroy(&attr);
474 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
476 * Returns 0 upon success and a negative error code upon failure.
478 int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
480 enum blk_zoned_model zbd_model;
483 assert(td->o.zone_mode == ZONE_MODE_ZBD);
485 zbd_model = get_zbd_model(f->file_name);
487 case ZBD_DM_HOST_AWARE:
488 case ZBD_DM_HOST_MANAGED:
489 ret = parse_zone_info(td, f);
492 ret = init_zone_info(td, f);
496 f->zbd_info->model = zbd_model;
500 void zbd_free_zone_info(struct fio_file *f)
507 pthread_mutex_lock(&f->zbd_info->mutex);
508 refcount = --f->zbd_info->refcount;
509 pthread_mutex_unlock(&f->zbd_info->mutex);
511 assert((int32_t)refcount >= 0);
518 * Initialize f->zbd_info.
520 * Returns 0 upon success and a negative error code upon failure.
522 * Note: this function can only work correctly if it is called before the first
525 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
527 struct thread_data *td2;
531 for_each_td(td2, i) {
532 for_each_file(td2, f2, j) {
533 if (td2 == td && f2 == file)
536 strcmp(f2->file_name, file->file_name) != 0)
538 file->zbd_info = f2->zbd_info;
539 file->zbd_info->refcount++;
544 ret = zbd_create_zone_info(td, file);
546 td_verror(td, -ret, "BLKREPORTZONE failed");
550 int zbd_init(struct thread_data *td)
555 for_each_file(td, f, i) {
556 if (f->filetype != FIO_TYPE_BLOCK)
558 if (td->o.zone_size && td->o.zone_size < 512) {
559 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
563 if (td->o.zone_size == 0 &&
564 get_zbd_model(f->file_name) == ZBD_DM_NONE) {
565 log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n",
569 zbd_init_zone_info(td, f);
572 if (!zbd_using_direct_io()) {
573 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
577 if (!zbd_verify_sizes())
580 if (!zbd_verify_bs())
587 * zbd_reset_range - reset zones for a range of sectors
588 * @td: FIO thread data.
589 * @f: Fio file for which to reset zones
590 * @sector: Starting sector in units of 512 bytes
591 * @nr_sectors: Number of sectors in units of 512 bytes
593 * Returns 0 upon success and a negative error code upon failure.
595 static int zbd_reset_range(struct thread_data *td, const struct fio_file *f,
596 uint64_t offset, uint64_t length)
598 struct blk_zone_range zr = {
599 .sector = offset >> 9,
600 .nr_sectors = length >> 9,
602 uint32_t zone_idx_b, zone_idx_e;
603 struct fio_zone_info *zb, *ze, *z;
607 assert(is_valid_offset(f, offset + length - 1));
608 switch (f->zbd_info->model) {
609 case ZBD_DM_HOST_AWARE:
610 case ZBD_DM_HOST_MANAGED:
611 ret = ioctl(f->fd, BLKRESETZONE, &zr);
613 td_verror(td, errno, "resetting wp failed");
614 log_err("%s: resetting wp for %llu sectors at sector %llu failed (%d).\n",
615 f->file_name, zr.nr_sectors, zr.sector, errno);
623 zone_idx_b = zbd_zone_idx(f, offset);
624 zb = &f->zbd_info->zone_info[zone_idx_b];
625 zone_idx_e = zbd_zone_idx(f, offset + length);
626 ze = &f->zbd_info->zone_info[zone_idx_e];
627 for (z = zb; z < ze; z++) {
628 pthread_mutex_lock(&z->mutex);
629 pthread_mutex_lock(&f->zbd_info->mutex);
630 f->zbd_info->sectors_with_data -= z->wp - z->start;
631 pthread_mutex_unlock(&f->zbd_info->mutex);
634 pthread_mutex_unlock(&z->mutex);
637 td->ts.nr_zone_resets += ze - zb;
642 static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info,
643 struct fio_zone_info *zone)
645 return zone - zbd_info->zone_info;
649 * zbd_reset_zone - reset the write pointer of a single zone
650 * @td: FIO thread data.
651 * @f: FIO file associated with the disk for which to reset a write pointer.
654 * Returns 0 upon success and a negative error code upon failure.
656 static int zbd_reset_zone(struct thread_data *td, const struct fio_file *f,
657 struct fio_zone_info *z)
659 dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
660 zbd_zone_nr(f->zbd_info, z));
662 return zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
666 * Reset a range of zones. Returns 0 upon success and 1 upon failure.
667 * @td: fio thread data.
668 * @f: fio file for which to reset zones
669 * @zb: first zone to reset.
670 * @ze: first zone not to reset.
671 * @all_zones: whether to reset all zones or only those zones for which the
672 * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE].
674 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
675 struct fio_zone_info *const zb,
676 struct fio_zone_info *const ze, bool all_zones)
678 struct fio_zone_info *z, *start_z = ze;
679 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
683 dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name,
684 zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze));
686 for (z = zb; z < ze; z++) {
687 pthread_mutex_lock(&z->mutex);
689 case BLK_ZONE_TYPE_SEQWRITE_REQ:
690 reset_wp = all_zones ? z->wp != z->start :
691 (td->o.td_ddir & TD_DDIR_WRITE) &&
693 if (start_z == ze && reset_wp) {
695 } else if (start_z < ze && !reset_wp) {
697 "%s: resetting zones %u .. %u\n",
699 zbd_zone_nr(f->zbd_info, start_z),
700 zbd_zone_nr(f->zbd_info, z));
701 if (zbd_reset_range(td, f, start_z->start,
702 z->start - start_z->start) < 0)
710 dprint(FD_ZBD, "%s: resetting zones %u .. %u\n",
711 f->file_name, zbd_zone_nr(f->zbd_info, start_z),
712 zbd_zone_nr(f->zbd_info, z));
713 if (zbd_reset_range(td, f, start_z->start,
714 z->start - start_z->start) < 0)
721 dprint(FD_ZBD, "%s: resetting zones %u .. %u\n", f->file_name,
722 zbd_zone_nr(f->zbd_info, start_z),
723 zbd_zone_nr(f->zbd_info, z));
724 if (zbd_reset_range(td, f, start_z->start,
725 z->start - start_z->start) < 0)
728 for (z = zb; z < ze; z++)
729 pthread_mutex_unlock(&z->mutex);
735 * Reset zbd_info.write_cnt, the counter that counts down towards the next
738 static void zbd_reset_write_cnt(const struct thread_data *td,
739 const struct fio_file *f)
741 assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
743 pthread_mutex_lock(&f->zbd_info->mutex);
744 f->zbd_info->write_cnt = td->o.zrf.u.f ?
745 min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
746 pthread_mutex_unlock(&f->zbd_info->mutex);
749 static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
750 const struct fio_file *f)
752 uint32_t write_cnt = 0;
754 pthread_mutex_lock(&f->zbd_info->mutex);
755 assert(f->zbd_info->write_cnt);
756 if (f->zbd_info->write_cnt)
757 write_cnt = --f->zbd_info->write_cnt;
759 zbd_reset_write_cnt(td, f);
760 pthread_mutex_unlock(&f->zbd_info->mutex);
762 return write_cnt == 0;
770 /* Calculate the number of sectors with data (swd) and perform action 'a' */
771 static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a)
773 struct fio_zone_info *zb, *ze, *z;
776 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
777 ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset +
779 for (z = zb; z < ze; z++) {
780 pthread_mutex_lock(&z->mutex);
781 swd += z->wp - z->start;
783 pthread_mutex_lock(&f->zbd_info->mutex);
786 assert(f->zbd_info->sectors_with_data == swd);
789 f->zbd_info->sectors_with_data = swd;
792 pthread_mutex_unlock(&f->zbd_info->mutex);
793 for (z = zb; z < ze; z++)
794 pthread_mutex_unlock(&z->mutex);
800 * The swd check is useful for debugging but takes too much time to leave
801 * it enabled all the time. Hence it is disabled by default.
803 static const bool enable_check_swd = false;
805 /* Check whether the value of zbd_info.sectors_with_data is correct. */
806 static void zbd_check_swd(const struct fio_file *f)
808 if (!enable_check_swd)
811 zbd_process_swd(f, CHECK_SWD);
814 static void zbd_init_swd(struct fio_file *f)
818 swd = zbd_process_swd(f, SET_SWD);
819 dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name,
823 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
825 struct fio_zone_info *zb, *ze;
831 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
832 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size);
833 ze = &f->zbd_info->zone_info[zone_idx_e];
836 * If data verification is enabled reset the affected zones before
837 * writing any data to avoid that a zone reset has to be issued while
838 * writing data, which causes data loss.
840 zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE &&
841 (td->o.td_ddir & TD_DDIR_WRITE) &&
842 td->runstate != TD_VERIFYING);
843 zbd_reset_write_cnt(td, f);
846 /* The caller must hold f->zbd_info->mutex. */
847 static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
848 unsigned int zone_idx)
850 struct zoned_block_device_info *zbdi = f->zbd_info;
853 assert(td->o.max_open_zones <= ARRAY_SIZE(zbdi->open_zones));
854 assert(zbdi->num_open_zones <= td->o.max_open_zones);
856 for (i = 0; i < zbdi->num_open_zones; i++)
857 if (zbdi->open_zones[i] == zone_idx)
864 * Open a ZBD zone if it was not yet open. Returns true if either the zone was
865 * already open or if opening a new zone is allowed. Returns false if the zone
866 * was not yet open and opening a new zone would cause the zone limit to be
869 static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
872 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
873 const struct fio_file *f = io_u->file;
874 struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
877 if (z->cond == BLK_ZONE_COND_OFFLINE)
881 * Skip full zones with data verification enabled because resetting a
882 * zone causes data loss and hence causes verification to fail.
884 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
887 /* Zero means no limit */
888 if (!td->o.max_open_zones)
891 pthread_mutex_lock(&f->zbd_info->mutex);
892 if (is_zone_open(td, f, zone_idx))
895 if (f->zbd_info->num_open_zones >= td->o.max_open_zones)
897 dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx);
898 f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx;
903 pthread_mutex_unlock(&f->zbd_info->mutex);
907 /* The caller must hold f->zbd_info->mutex */
908 static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
909 unsigned int open_zone_idx)
913 assert(open_zone_idx < f->zbd_info->num_open_zones);
914 zone_idx = f->zbd_info->open_zones[open_zone_idx];
915 memmove(f->zbd_info->open_zones + open_zone_idx,
916 f->zbd_info->open_zones + open_zone_idx + 1,
917 (FIO_MAX_OPEN_ZBD_ZONES - (open_zone_idx + 1)) *
918 sizeof(f->zbd_info->open_zones[0]));
919 f->zbd_info->num_open_zones--;
920 f->zbd_info->zone_info[zone_idx].open = 0;
924 * Modify the offset of an I/O unit that does not refer to an open zone such
925 * that it refers to an open zone. Close an open zone and open a new zone if
926 * necessary. This algorithm can only work correctly if all write pointers are
927 * a multiple of the fio block size. The caller must neither hold z->mutex
928 * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
930 struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
933 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
934 const struct fio_file *f = io_u->file;
935 struct fio_zone_info *z;
936 unsigned int open_zone_idx = -1;
937 uint32_t zone_idx, new_zone_idx;
940 assert(is_valid_offset(f, io_u->offset));
942 if (td->o.max_open_zones) {
944 * This statement accesses f->zbd_info->open_zones[] on purpose
947 zone_idx = f->zbd_info->open_zones[(io_u->offset -
949 f->zbd_info->num_open_zones / f->io_size];
951 zone_idx = zbd_zone_idx(f, io_u->offset);
953 dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
954 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
957 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
958 * lock it can happen that the state of the zone with index zone_idx
959 * has changed after 'z' has been assigned and before f->zbd_info->mutex
960 * has been obtained. Hence the loop.
963 z = &f->zbd_info->zone_info[zone_idx];
965 pthread_mutex_lock(&z->mutex);
966 pthread_mutex_lock(&f->zbd_info->mutex);
967 if (td->o.max_open_zones == 0)
969 if (f->zbd_info->num_open_zones == 0) {
970 pthread_mutex_unlock(&f->zbd_info->mutex);
971 pthread_mutex_unlock(&z->mutex);
972 dprint(FD_ZBD, "%s(%s): no zones are open\n",
973 __func__, f->file_name);
976 open_zone_idx = (io_u->offset - f->file_offset) *
977 f->zbd_info->num_open_zones / f->io_size;
978 assert(open_zone_idx < f->zbd_info->num_open_zones);
979 new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
980 if (new_zone_idx == zone_idx)
982 zone_idx = new_zone_idx;
983 pthread_mutex_unlock(&f->zbd_info->mutex);
984 pthread_mutex_unlock(&z->mutex);
987 /* Both z->mutex and f->zbd_info->mutex are held. */
990 if (z->wp + min_bs <= (z+1)->start) {
991 pthread_mutex_unlock(&f->zbd_info->mutex);
994 dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name,
996 if (td->o.max_open_zones)
997 zbd_close_zone(td, f, open_zone_idx);
998 pthread_mutex_unlock(&f->zbd_info->mutex);
1000 /* Only z->mutex is held. */
1002 /* Zone 'z' is full, so try to open a new zone. */
1003 for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
1005 pthread_mutex_unlock(&z->mutex);
1007 if (!is_valid_offset(f, z->start)) {
1009 zone_idx = zbd_zone_idx(f, f->file_offset);
1010 z = &f->zbd_info->zone_info[zone_idx];
1012 assert(is_valid_offset(f, z->start));
1013 pthread_mutex_lock(&z->mutex);
1016 if (zbd_open_zone(td, io_u, zone_idx))
1020 /* Only z->mutex is held. */
1022 /* Check whether the write fits in any of the already opened zones. */
1023 pthread_mutex_lock(&f->zbd_info->mutex);
1024 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1025 zone_idx = f->zbd_info->open_zones[i];
1026 pthread_mutex_unlock(&f->zbd_info->mutex);
1027 pthread_mutex_unlock(&z->mutex);
1029 z = &f->zbd_info->zone_info[zone_idx];
1031 pthread_mutex_lock(&z->mutex);
1032 if (z->wp + min_bs <= (z+1)->start)
1034 pthread_mutex_lock(&f->zbd_info->mutex);
1036 pthread_mutex_unlock(&f->zbd_info->mutex);
1037 pthread_mutex_unlock(&z->mutex);
1038 dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
1043 dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
1045 io_u->offset = z->start;
1049 /* The caller must hold z->mutex. */
1050 static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
1052 struct fio_zone_info *z)
1054 const struct fio_file *f = io_u->file;
1055 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
1057 if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
1058 pthread_mutex_unlock(&z->mutex);
1059 z = zbd_convert_to_open_zone(td, io_u);
1063 if (z->verify_block * min_bs >= f->zbd_info->zone_size)
1064 log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
1065 min_bs, (unsigned long long) f->zbd_info->zone_size);
1066 io_u->offset = z->start + z->verify_block++ * min_bs;
1071 * Find another zone for which @io_u fits below the write pointer. Start
1072 * searching in zones @zb + 1 .. @zl and continue searching in zones
1075 * Either returns NULL or returns a zone pointer and holds the mutex for that
1078 static struct fio_zone_info *
1079 zbd_find_zone(struct thread_data *td, struct io_u *io_u,
1080 struct fio_zone_info *zb, struct fio_zone_info *zl)
1082 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1083 const struct fio_file *f = io_u->file;
1084 struct fio_zone_info *z1, *z2;
1085 const struct fio_zone_info *const zf =
1086 &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
1089 * Skip to the next non-empty zone in case of sequential I/O and to
1090 * the nearest non-empty zone in case of random I/O.
1092 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1093 if (z1 < zl && z1->cond != BLK_ZONE_COND_OFFLINE) {
1094 pthread_mutex_lock(&z1->mutex);
1095 if (z1->start + min_bs <= z1->wp)
1097 pthread_mutex_unlock(&z1->mutex);
1098 } else if (!td_random(td)) {
1101 if (td_random(td) && z2 >= zf &&
1102 z2->cond != BLK_ZONE_COND_OFFLINE) {
1103 pthread_mutex_lock(&z2->mutex);
1104 if (z2->start + min_bs <= z2->wp)
1106 pthread_mutex_unlock(&z2->mutex);
1109 dprint(FD_ZBD, "%s: adjusting random read offset failed\n",
1115 * zbd_queue_io - update the write pointer of a sequential zone
1117 * @success: Whether or not the I/O unit has been queued successfully
1118 * @q: queueing status (busy, completed or queued).
1120 * For write and trim operations, update the write pointer of the I/O unit
1123 static void zbd_queue_io(struct io_u *io_u, int q, bool success)
1125 const struct fio_file *f = io_u->file;
1126 struct zoned_block_device_info *zbd_info = f->zbd_info;
1127 struct fio_zone_info *z;
1134 zone_idx = zbd_zone_idx(f, io_u->offset);
1135 assert(zone_idx < zbd_info->nr_zones);
1136 z = &zbd_info->zone_info[zone_idx];
1138 if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ)
1145 "%s: queued I/O (%lld, %llu) for zone %u\n",
1146 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1148 switch (io_u->ddir) {
1150 zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
1152 pthread_mutex_lock(&zbd_info->mutex);
1154 * z->wp > zone_end means that one or more I/O errors
1157 if (z->wp <= zone_end)
1158 zbd_info->sectors_with_data += zone_end - z->wp;
1159 pthread_mutex_unlock(&zbd_info->mutex);
1163 assert(z->wp == z->start);
1170 if (!success || q != FIO_Q_QUEUED) {
1171 /* BUSY or COMPLETED: unlock the zone */
1172 pthread_mutex_unlock(&z->mutex);
1173 io_u->zbd_put_io = NULL;
1178 * zbd_put_io - Unlock an I/O unit target zone lock
1181 static void zbd_put_io(const struct io_u *io_u)
1183 const struct fio_file *f = io_u->file;
1184 struct zoned_block_device_info *zbd_info = f->zbd_info;
1185 struct fio_zone_info *z;
1191 zone_idx = zbd_zone_idx(f, io_u->offset);
1192 assert(zone_idx < zbd_info->nr_zones);
1193 z = &zbd_info->zone_info[zone_idx];
1195 if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ)
1199 "%s: terminate I/O (%lld, %llu) for zone %u\n",
1200 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1202 assert(pthread_mutex_unlock(&z->mutex) == 0);
1206 bool zbd_unaligned_write(int error_code)
1208 switch (error_code) {
1217 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1218 * @td: FIO thread data.
1219 * @io_u: FIO I/O unit.
1221 * Locking strategy: returns with z->mutex locked if and only if z refers
1222 * to a sequential zone and if io_u_accept is returned. z is the zone that
1223 * corresponds to io_u->offset at the end of this function.
1225 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1227 const struct fio_file *f = io_u->file;
1228 uint32_t zone_idx_b;
1229 struct fio_zone_info *zb, *zl, *orig_zb;
1230 uint32_t orig_len = io_u->buflen;
1231 uint32_t min_bs = td->o.min_bs[io_u->ddir];
1238 assert(is_valid_offset(f, io_u->offset));
1239 assert(io_u->buflen);
1240 zone_idx_b = zbd_zone_idx(f, io_u->offset);
1241 zb = &f->zbd_info->zone_info[zone_idx_b];
1244 /* Accept the I/O offset for conventional zones. */
1245 if (zb->type == BLK_ZONE_TYPE_CONVENTIONAL)
1249 * Accept the I/O offset for reads if reading beyond the write pointer
1252 if (zb->cond != BLK_ZONE_COND_OFFLINE &&
1253 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1259 * Lock the io_u target zone. The zone will be unlocked if io_u offset
1260 * is changed or when io_u completes and zbd_put_io() executed.
1261 * To avoid multiple jobs doing asynchronous I/Os from deadlocking each
1262 * other waiting for zone locks when building an io_u batch, first
1263 * only trylock the zone. If the zone is already locked by another job,
1264 * process the currently queued I/Os so that I/O progress is made and
1267 if (pthread_mutex_trylock(&zb->mutex) != 0) {
1268 if (!td_ioengine_flagged(td, FIO_SYNCIO))
1270 pthread_mutex_lock(&zb->mutex);
1273 switch (io_u->ddir) {
1275 if (td->runstate == TD_VERIFYING) {
1276 zb = zbd_replay_write_order(td, io_u, zb);
1280 * Check that there is enough written data in the zone to do an
1281 * I/O of at least min_bs B. If there isn't, find a new zone for
1284 range = zb->cond != BLK_ZONE_COND_OFFLINE ?
1285 zb->wp - zb->start : 0;
1286 if (range < min_bs ||
1287 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1288 pthread_mutex_unlock(&zb->mutex);
1289 zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
1290 f->file_offset + f->io_size)];
1291 zb = zbd_find_zone(td, io_u, zb, zl);
1294 "%s: zbd_find_zone(%lld, %llu) failed\n",
1295 f->file_name, io_u->offset,
1300 * zbd_find_zone() returned a zone with a range of at
1303 range = zb->wp - zb->start;
1304 assert(range >= min_bs);
1307 io_u->offset = zb->start;
1310 * Make sure the I/O is within the zone valid data range while
1311 * maximizing the I/O size and preserving randomness.
1313 if (range <= io_u->buflen)
1314 io_u->offset = zb->start;
1315 else if (td_random(td))
1316 io_u->offset = zb->start +
1317 ((io_u->offset - orig_zb->start) %
1318 (range - io_u->buflen)) / min_bs * min_bs;
1320 * Make sure the I/O does not cross over the zone wp position.
1322 new_len = min((unsigned long long)io_u->buflen,
1323 (unsigned long long)(zb->wp - io_u->offset));
1324 new_len = new_len / min_bs * min_bs;
1325 if (new_len < io_u->buflen) {
1326 io_u->buflen = new_len;
1327 dprint(FD_IO, "Changed length from %u into %llu\n",
1328 orig_len, io_u->buflen);
1330 assert(zb->start <= io_u->offset);
1331 assert(io_u->offset + io_u->buflen <= zb->wp);
1334 if (io_u->buflen > f->zbd_info->zone_size)
1336 if (!zbd_open_zone(td, io_u, zone_idx_b)) {
1337 pthread_mutex_unlock(&zb->mutex);
1338 zb = zbd_convert_to_open_zone(td, io_u);
1341 zone_idx_b = zb - f->zbd_info->zone_info;
1343 /* Check whether the zone reset threshold has been exceeded */
1344 if (td->o.zrf.u.f) {
1345 if (f->zbd_info->sectors_with_data >=
1346 f->io_size * td->o.zrt.u.f &&
1347 zbd_dec_and_reset_write_cnt(td, f)) {
1351 /* Reset the zone pointer if necessary */
1352 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
1353 assert(td->o.verify == VERIFY_NONE);
1355 * Since previous write requests may have been submitted
1356 * asynchronously and since we will submit the zone
1357 * reset synchronously, wait until previously submitted
1358 * write requests have completed before issuing a
1363 if (zbd_reset_zone(td, f, zb) < 0)
1366 /* Make writes occur at the write pointer */
1367 assert(!zbd_zone_full(f, zb, min_bs));
1368 io_u->offset = zb->wp;
1369 if (!is_valid_offset(f, io_u->offset)) {
1370 dprint(FD_ZBD, "Dropped request with offset %llu\n",
1375 * Make sure that the buflen is a multiple of the minimal
1376 * block size. Give up if shrinking would make the request too
1379 new_len = min((unsigned long long)io_u->buflen,
1380 (zb + 1)->start - io_u->offset);
1381 new_len = new_len / min_bs * min_bs;
1382 if (new_len == io_u->buflen)
1384 if (new_len >= min_bs) {
1385 io_u->buflen = new_len;
1386 dprint(FD_IO, "Changed length from %u into %llu\n",
1387 orig_len, io_u->buflen);
1390 log_err("Zone remainder %lld smaller than minimum block size %d\n",
1391 ((zb + 1)->start - io_u->offset),
1398 case DDIR_SYNC_FILE_RANGE:
1409 assert(zb->cond != BLK_ZONE_COND_OFFLINE);
1410 assert(!io_u->zbd_queue_io);
1411 assert(!io_u->zbd_put_io);
1412 io_u->zbd_queue_io = zbd_queue_io;
1413 io_u->zbd_put_io = zbd_put_io;
1418 pthread_mutex_unlock(&zb->mutex);
1422 /* Return a string with ZBD statistics */
1423 char *zbd_write_status(const struct thread_stat *ts)
1427 if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0)