2 * Copyright (C) 2018 Western Digital Corporation or its affiliates.
4 * This file is released under the GPL.
19 #include "oslib/asprintf.h"
26 * zbd_get_zoned_model - Get a device zoned model
27 * @td: FIO thread data
28 * @f: FIO file for which to get model information
30 int zbd_get_zoned_model(struct thread_data *td, struct fio_file *f,
31 enum zbd_zoned_model *model)
35 if (td->io_ops && td->io_ops->get_zoned_model)
36 ret = td->io_ops->get_zoned_model(td, f, model);
38 ret = blkzoned_get_zoned_model(td, f, model);
40 td_verror(td, errno, "get zoned model failed");
41 log_err("%s: get zoned model failed (%d).\n",
49 * zbd_report_zones - Get zone information
50 * @td: FIO thread data.
51 * @f: FIO file for which to get zone information
52 * @offset: offset from which to report zones
53 * @zones: Array of struct zbd_zone
54 * @nr_zones: Size of @zones array
56 * Get zone information into @zones starting from the zone at offset @offset
57 * for the device specified by @f.
59 * Returns the number of zones reported upon success and a negative error code
60 * upon failure. If the zone report is empty, always assume an error (device
61 * problem) and return -EIO.
63 int zbd_report_zones(struct thread_data *td, struct fio_file *f,
64 uint64_t offset, struct zbd_zone *zones,
65 unsigned int nr_zones)
69 if (td->io_ops && td->io_ops->report_zones)
70 ret = td->io_ops->report_zones(td, f, offset, zones, nr_zones);
72 ret = blkzoned_report_zones(td, f, offset, zones, nr_zones);
74 td_verror(td, errno, "report zones failed");
75 log_err("%s: report zones from sector %llu failed (%d).\n",
76 f->file_name, (unsigned long long)offset >> 9, errno);
77 } else if (ret == 0) {
78 td_verror(td, errno, "Empty zone report");
79 log_err("%s: report zones from sector %llu is empty.\n",
80 f->file_name, (unsigned long long)offset >> 9);
88 * zbd_reset_wp - reset the write pointer of a range of zones
89 * @td: FIO thread data.
90 * @f: FIO file for which to reset zones
91 * @offset: Starting offset of the first zone to reset
92 * @length: Length of the range of zones to reset
94 * Reset the write pointer of all zones in the range @offset...@offset+@length.
95 * Returns 0 upon success and a negative error code upon failure.
97 int zbd_reset_wp(struct thread_data *td, struct fio_file *f,
98 uint64_t offset, uint64_t length)
102 if (td->io_ops && td->io_ops->reset_wp)
103 ret = td->io_ops->reset_wp(td, f, offset, length);
105 ret = blkzoned_reset_wp(td, f, offset, length);
107 td_verror(td, errno, "resetting wp failed");
108 log_err("%s: resetting wp for %llu sectors at sector %llu failed (%d).\n",
109 f->file_name, (unsigned long long)length >> 9,
110 (unsigned long long)offset >> 9, errno);
117 * zbd_zone_idx - convert an offset into a zone number
119 * @offset: offset in bytes. If this offset is in the first zone_size bytes
120 * past the disk size then the index of the sentinel is returned.
122 static uint32_t zbd_zone_idx(const struct fio_file *f, uint64_t offset)
126 if (f->zbd_info->zone_size_log2 > 0)
127 zone_idx = offset >> f->zbd_info->zone_size_log2;
129 zone_idx = offset / f->zbd_info->zone_size;
131 return min(zone_idx, f->zbd_info->nr_zones);
135 * zbd_zone_swr - Test whether a zone requires sequential writes
136 * @z: zone info pointer.
138 static inline bool zbd_zone_swr(struct fio_zone_info *z)
140 return z->type == ZBD_ZONE_TYPE_SWR;
144 * zbd_zone_full - verify whether a minimum number of bytes remain in a zone
146 * @z: zone info pointer.
147 * @required: minimum number of bytes that must remain in a zone.
149 * The caller must hold z->mutex.
151 static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
154 assert((required & 511) == 0);
156 return zbd_zone_swr(z) &&
157 z->wp + required > z->start + f->zbd_info->zone_size;
160 static void zone_lock(struct thread_data *td, struct fio_file *f, struct fio_zone_info *z)
162 struct zoned_block_device_info *zbd = f->zbd_info;
163 uint32_t nz = z - zbd->zone_info;
165 /* A thread should never lock zones outside its working area. */
166 assert(f->min_zone <= nz && nz < f->max_zone);
169 * Lock the io_u target zone. The zone will be unlocked if io_u offset
170 * is changed or when io_u completes and zbd_put_io() executed.
171 * To avoid multiple jobs doing asynchronous I/Os from deadlocking each
172 * other waiting for zone locks when building an io_u batch, first
173 * only trylock the zone. If the zone is already locked by another job,
174 * process the currently queued I/Os so that I/O progress is made and
177 if (pthread_mutex_trylock(&z->mutex) != 0) {
178 if (!td_ioengine_flagged(td, FIO_SYNCIO))
180 pthread_mutex_lock(&z->mutex);
184 static bool is_valid_offset(const struct fio_file *f, uint64_t offset)
186 return (uint64_t)(offset - f->file_offset) < f->io_size;
189 /* Verify whether direct I/O is used for all host-managed zoned drives. */
190 static bool zbd_using_direct_io(void)
192 struct thread_data *td;
197 if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE))
199 for_each_file(td, f, j) {
201 f->zbd_info->model == ZBD_HOST_MANAGED)
209 /* Whether or not the I/O range for f includes one or more sequential zones */
210 static bool zbd_is_seq_job(struct fio_file *f)
212 uint32_t zone_idx, zone_idx_b, zone_idx_e;
217 zone_idx_b = zbd_zone_idx(f, f->file_offset);
218 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size - 1);
219 for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++)
220 if (zbd_zone_swr(&f->zbd_info->zone_info[zone_idx]))
227 * Verify whether offset and size parameters are aligned with zone boundaries.
229 static bool zbd_verify_sizes(void)
231 const struct fio_zone_info *z;
232 struct thread_data *td;
234 uint64_t new_offset, new_end;
239 for_each_file(td, f, j) {
242 if (f->file_offset >= f->real_file_size)
244 if (!zbd_is_seq_job(f))
247 if (!td->o.zone_size) {
248 td->o.zone_size = f->zbd_info->zone_size;
249 if (!td->o.zone_size) {
250 log_err("%s: invalid 0 zone size\n",
254 } else if (td->o.zone_size != f->zbd_info->zone_size) {
255 log_err("%s: job parameter zonesize %llu does not match disk zone size %llu.\n",
256 f->file_name, (unsigned long long) td->o.zone_size,
257 (unsigned long long) f->zbd_info->zone_size);
261 if (td->o.zone_skip &&
262 (td->o.zone_skip < td->o.zone_size ||
263 td->o.zone_skip % td->o.zone_size)) {
264 log_err("%s: zoneskip %llu is not a multiple of the device zone size %llu.\n",
265 f->file_name, (unsigned long long) td->o.zone_skip,
266 (unsigned long long) td->o.zone_size);
270 zone_idx = zbd_zone_idx(f, f->file_offset);
271 z = &f->zbd_info->zone_info[zone_idx];
272 if ((f->file_offset != z->start) &&
273 (td->o.td_ddir != TD_DDIR_READ)) {
274 new_offset = (z+1)->start;
275 if (new_offset >= f->file_offset + f->io_size) {
276 log_info("%s: io_size must be at least one zone\n",
280 log_info("%s: rounded up offset from %llu to %llu\n",
281 f->file_name, (unsigned long long) f->file_offset,
282 (unsigned long long) new_offset);
283 f->io_size -= (new_offset - f->file_offset);
284 f->file_offset = new_offset;
286 zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size);
287 z = &f->zbd_info->zone_info[zone_idx];
289 if ((td->o.td_ddir != TD_DDIR_READ) &&
290 (f->file_offset + f->io_size != new_end)) {
291 if (new_end <= f->file_offset) {
292 log_info("%s: io_size must be at least one zone\n",
296 log_info("%s: rounded down io_size from %llu to %llu\n",
297 f->file_name, (unsigned long long) f->io_size,
298 (unsigned long long) new_end - f->file_offset);
299 f->io_size = new_end - f->file_offset;
302 f->min_zone = zbd_zone_idx(f, f->file_offset);
303 f->max_zone = zbd_zone_idx(f, f->file_offset + f->io_size);
310 static bool zbd_verify_bs(void)
312 struct thread_data *td;
318 for_each_file(td, f, j) {
321 zone_size = f->zbd_info->zone_size;
322 for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
323 if (td->o.verify != VERIFY_NONE &&
324 zone_size % td->o.bs[k] != 0) {
325 log_info("%s: block size %llu is not a divisor of the zone size %d\n",
326 f->file_name, td->o.bs[k],
336 static int ilog2(uint64_t i)
348 * Initialize f->zbd_info for devices that are not zoned block devices. This
349 * allows to execute a ZBD workload against a non-ZBD device.
351 static int init_zone_info(struct thread_data *td, struct fio_file *f)
354 struct fio_zone_info *p;
355 uint64_t zone_size = td->o.zone_size;
356 struct zoned_block_device_info *zbd_info = NULL;
359 if (zone_size == 0) {
360 log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n",
365 if (zone_size < 512) {
366 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
371 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
372 zbd_info = scalloc(1, sizeof(*zbd_info) +
373 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
377 mutex_init_pshared(&zbd_info->mutex);
378 zbd_info->refcount = 1;
379 p = &zbd_info->zone_info[0];
380 for (i = 0; i < nr_zones; i++, p++) {
381 mutex_init_pshared_with_type(&p->mutex,
382 PTHREAD_MUTEX_RECURSIVE);
383 p->start = i * zone_size;
384 p->wp = p->start + zone_size;
385 p->type = ZBD_ZONE_TYPE_SWR;
386 p->cond = ZBD_ZONE_COND_EMPTY;
389 p->start = nr_zones * zone_size;
391 f->zbd_info = zbd_info;
392 f->zbd_info->zone_size = zone_size;
393 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
394 ilog2(zone_size) : 0;
395 f->zbd_info->nr_zones = nr_zones;
400 * Maximum number of zones to report in one operation.
402 #define ZBD_REPORT_MAX_ZONES 8192U
405 * Parse the device zone report and store it in f->zbd_info. Must be called
406 * only for devices that are zoned, namely those with a model != ZBD_NONE.
408 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
411 struct zbd_zone *zones, *z;
412 struct fio_zone_info *p;
413 uint64_t zone_size, offset;
414 struct zoned_block_device_info *zbd_info = NULL;
417 zones = calloc(ZBD_REPORT_MAX_ZONES, sizeof(struct zbd_zone));
421 nrz = zbd_report_zones(td, f, 0, zones, ZBD_REPORT_MAX_ZONES);
424 log_info("fio: report zones (offset 0) failed for %s (%d).\n",
429 zone_size = zones[0].len;
430 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
432 if (td->o.zone_size == 0) {
433 td->o.zone_size = zone_size;
434 } else if (td->o.zone_size != zone_size) {
435 log_err("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n",
436 f->file_name, (unsigned long long) td->o.zone_size,
437 (unsigned long long) zone_size);
442 dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name,
443 nr_zones, (unsigned long long) zone_size / 1024);
445 zbd_info = scalloc(1, sizeof(*zbd_info) +
446 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
450 mutex_init_pshared(&zbd_info->mutex);
451 zbd_info->refcount = 1;
452 p = &zbd_info->zone_info[0];
453 for (offset = 0, j = 0; j < nr_zones;) {
455 for (i = 0; i < nrz; i++, j++, z++, p++) {
456 mutex_init_pshared_with_type(&p->mutex,
457 PTHREAD_MUTEX_RECURSIVE);
460 case ZBD_ZONE_COND_NOT_WP:
461 case ZBD_ZONE_COND_FULL:
462 p->wp = p->start + zone_size;
465 assert(z->start <= z->wp);
466 assert(z->wp <= z->start + zone_size);
472 if (j > 0 && p->start != p[-1].start + zone_size) {
473 log_info("%s: invalid zone data\n",
480 offset = z->start + z->len;
483 nrz = zbd_report_zones(td, f, offset,
484 zones, ZBD_REPORT_MAX_ZONES);
487 log_info("fio: report zones (offset %llu) failed for %s (%d).\n",
488 (unsigned long long)offset,
495 zbd_info->zone_info[nr_zones].start = offset;
497 f->zbd_info = zbd_info;
498 f->zbd_info->zone_size = zone_size;
499 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
500 ilog2(zone_size) : 0;
501 f->zbd_info->nr_zones = nr_zones;
512 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
514 * Returns 0 upon success and a negative error code upon failure.
516 static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
518 enum zbd_zoned_model zbd_model;
521 assert(td->o.zone_mode == ZONE_MODE_ZBD);
523 ret = zbd_get_zoned_model(td, f, &zbd_model);
531 case ZBD_HOST_MANAGED:
532 ret = parse_zone_info(td, f);
535 ret = init_zone_info(td, f);
538 td_verror(td, EINVAL, "Unsupported zoned model");
539 log_err("Unsupported zoned model\n");
544 f->zbd_info->model = zbd_model;
545 f->zbd_info->max_open_zones = td->o.max_open_zones;
550 void zbd_free_zone_info(struct fio_file *f)
556 pthread_mutex_lock(&f->zbd_info->mutex);
557 refcount = --f->zbd_info->refcount;
558 pthread_mutex_unlock(&f->zbd_info->mutex);
560 assert((int32_t)refcount >= 0);
567 * Initialize f->zbd_info.
569 * Returns 0 upon success and a negative error code upon failure.
571 * Note: this function can only work correctly if it is called before the first
574 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
576 struct thread_data *td2;
580 for_each_td(td2, i) {
581 for_each_file(td2, f2, j) {
582 if (td2 == td && f2 == file)
585 strcmp(f2->file_name, file->file_name) != 0)
587 file->zbd_info = f2->zbd_info;
588 file->zbd_info->refcount++;
593 ret = zbd_create_zone_info(td, file);
595 td_verror(td, -ret, "zbd_create_zone_info() failed");
599 int zbd_setup_files(struct thread_data *td)
604 for_each_file(td, f, i) {
605 if (zbd_init_zone_info(td, f))
609 if (!zbd_using_direct_io()) {
610 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
614 if (!zbd_verify_sizes())
617 if (!zbd_verify_bs())
620 for_each_file(td, f, i) {
621 struct zoned_block_device_info *zbd = f->zbd_info;
626 zbd->max_open_zones = zbd->max_open_zones ?: ZBD_MAX_OPEN_ZONES;
628 if (td->o.max_open_zones > 0 &&
629 zbd->max_open_zones != td->o.max_open_zones) {
630 log_err("Different 'max_open_zones' values\n");
633 if (zbd->max_open_zones > ZBD_MAX_OPEN_ZONES) {
634 log_err("'max_open_zones' value is limited by %u\n", ZBD_MAX_OPEN_ZONES);
643 * zbd_reset_range - reset zones for a range of sectors
644 * @td: FIO thread data.
645 * @f: Fio file for which to reset zones
646 * @sector: Starting sector in units of 512 bytes
647 * @nr_sectors: Number of sectors in units of 512 bytes
649 * Returns 0 upon success and a negative error code upon failure.
651 static int zbd_reset_range(struct thread_data *td, struct fio_file *f,
652 uint64_t offset, uint64_t length)
654 uint32_t zone_idx_b, zone_idx_e;
655 struct fio_zone_info *zb, *ze, *z;
658 assert(is_valid_offset(f, offset + length - 1));
660 switch (f->zbd_info->model) {
662 case ZBD_HOST_MANAGED:
663 ret = zbd_reset_wp(td, f, offset, length);
671 zone_idx_b = zbd_zone_idx(f, offset);
672 zb = &f->zbd_info->zone_info[zone_idx_b];
673 zone_idx_e = zbd_zone_idx(f, offset + length);
674 ze = &f->zbd_info->zone_info[zone_idx_e];
675 for (z = zb; z < ze; z++) {
676 pthread_mutex_lock(&z->mutex);
677 pthread_mutex_lock(&f->zbd_info->mutex);
678 f->zbd_info->sectors_with_data -= z->wp - z->start;
679 pthread_mutex_unlock(&f->zbd_info->mutex);
682 pthread_mutex_unlock(&z->mutex);
685 td->ts.nr_zone_resets += ze - zb;
690 static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info,
691 struct fio_zone_info *zone)
693 return zone - zbd_info->zone_info;
697 * zbd_reset_zone - reset the write pointer of a single zone
698 * @td: FIO thread data.
699 * @f: FIO file associated with the disk for which to reset a write pointer.
702 * Returns 0 upon success and a negative error code upon failure.
704 static int zbd_reset_zone(struct thread_data *td, struct fio_file *f,
705 struct fio_zone_info *z)
707 dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
708 zbd_zone_nr(f->zbd_info, z));
710 return zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
713 /* The caller must hold f->zbd_info->mutex */
714 static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
715 unsigned int open_zone_idx)
719 assert(open_zone_idx < f->zbd_info->num_open_zones);
720 zone_idx = f->zbd_info->open_zones[open_zone_idx];
721 memmove(f->zbd_info->open_zones + open_zone_idx,
722 f->zbd_info->open_zones + open_zone_idx + 1,
723 (ZBD_MAX_OPEN_ZONES - (open_zone_idx + 1)) *
724 sizeof(f->zbd_info->open_zones[0]));
725 f->zbd_info->num_open_zones--;
726 td->num_open_zones--;
727 f->zbd_info->zone_info[zone_idx].open = 0;
731 * Reset a range of zones. Returns 0 upon success and 1 upon failure.
732 * @td: fio thread data.
733 * @f: fio file for which to reset zones
734 * @zb: first zone to reset.
735 * @ze: first zone not to reset.
736 * @all_zones: whether to reset all zones or only those zones for which the
737 * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE].
739 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
740 struct fio_zone_info *const zb,
741 struct fio_zone_info *const ze, bool all_zones)
743 struct fio_zone_info *z;
744 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
750 dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name,
751 zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze));
752 for (z = zb; z < ze; z++) {
753 uint32_t nz = z - f->zbd_info->zone_info;
755 if (!zbd_zone_swr(z))
761 pthread_mutex_lock(&f->zbd_info->mutex);
762 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
763 if (f->zbd_info->open_zones[i] == nz)
764 zbd_close_zone(td, f, i);
766 pthread_mutex_unlock(&f->zbd_info->mutex);
768 reset_wp = z->wp != z->start;
770 reset_wp = z->wp % min_bs != 0;
773 dprint(FD_ZBD, "%s: resetting zone %u\n",
775 zbd_zone_nr(f->zbd_info, z));
776 if (zbd_reset_zone(td, f, z) < 0)
779 pthread_mutex_unlock(&z->mutex);
786 * Reset zbd_info.write_cnt, the counter that counts down towards the next
789 static void _zbd_reset_write_cnt(const struct thread_data *td,
790 const struct fio_file *f)
792 assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
794 f->zbd_info->write_cnt = td->o.zrf.u.f ?
795 min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
798 static void zbd_reset_write_cnt(const struct thread_data *td,
799 const struct fio_file *f)
801 pthread_mutex_lock(&f->zbd_info->mutex);
802 _zbd_reset_write_cnt(td, f);
803 pthread_mutex_unlock(&f->zbd_info->mutex);
806 static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
807 const struct fio_file *f)
809 uint32_t write_cnt = 0;
811 pthread_mutex_lock(&f->zbd_info->mutex);
812 assert(f->zbd_info->write_cnt);
813 if (f->zbd_info->write_cnt)
814 write_cnt = --f->zbd_info->write_cnt;
816 _zbd_reset_write_cnt(td, f);
817 pthread_mutex_unlock(&f->zbd_info->mutex);
819 return write_cnt == 0;
827 /* Calculate the number of sectors with data (swd) and perform action 'a' */
828 static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a)
830 struct fio_zone_info *zb, *ze, *z;
833 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
834 ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset +
836 for (z = zb; z < ze; z++) {
837 pthread_mutex_lock(&z->mutex);
838 swd += z->wp - z->start;
840 pthread_mutex_lock(&f->zbd_info->mutex);
843 assert(f->zbd_info->sectors_with_data == swd);
846 f->zbd_info->sectors_with_data = swd;
849 pthread_mutex_unlock(&f->zbd_info->mutex);
850 for (z = zb; z < ze; z++)
851 pthread_mutex_unlock(&z->mutex);
857 * The swd check is useful for debugging but takes too much time to leave
858 * it enabled all the time. Hence it is disabled by default.
860 static const bool enable_check_swd = false;
862 /* Check whether the value of zbd_info.sectors_with_data is correct. */
863 static void zbd_check_swd(const struct fio_file *f)
865 if (!enable_check_swd)
868 zbd_process_swd(f, CHECK_SWD);
871 static void zbd_init_swd(struct fio_file *f)
875 if (!enable_check_swd)
878 swd = zbd_process_swd(f, SET_SWD);
879 dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name,
883 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
885 struct fio_zone_info *zb, *ze;
887 if (!f->zbd_info || !td_write(td))
890 zb = &f->zbd_info->zone_info[f->min_zone];
891 ze = &f->zbd_info->zone_info[f->max_zone];
894 * If data verification is enabled reset the affected zones before
895 * writing any data to avoid that a zone reset has to be issued while
896 * writing data, which causes data loss.
898 zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE &&
899 td->runstate != TD_VERIFYING);
900 zbd_reset_write_cnt(td, f);
903 /* The caller must hold f->zbd_info->mutex. */
904 static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
905 unsigned int zone_idx)
907 struct zoned_block_device_info *zbdi = f->zbd_info;
910 assert(td->o.job_max_open_zones == 0 || td->num_open_zones <= td->o.job_max_open_zones);
911 assert(td->o.job_max_open_zones <= zbdi->max_open_zones);
912 assert(zbdi->num_open_zones <= zbdi->max_open_zones);
914 for (i = 0; i < zbdi->num_open_zones; i++)
915 if (zbdi->open_zones[i] == zone_idx)
922 * Open a ZBD zone if it was not yet open. Returns true if either the zone was
923 * already open or if opening a new zone is allowed. Returns false if the zone
924 * was not yet open and opening a new zone would cause the zone limit to be
927 static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
930 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
931 const struct fio_file *f = io_u->file;
932 struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
935 if (z->cond == ZBD_ZONE_COND_OFFLINE)
939 * Skip full zones with data verification enabled because resetting a
940 * zone causes data loss and hence causes verification to fail.
942 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
945 pthread_mutex_lock(&f->zbd_info->mutex);
946 if (is_zone_open(td, f, zone_idx))
949 /* Zero means no limit */
950 if (td->o.job_max_open_zones > 0 &&
951 td->num_open_zones >= td->o.job_max_open_zones)
953 if (f->zbd_info->num_open_zones >= f->zbd_info->max_open_zones)
955 dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx);
956 f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx;
957 td->num_open_zones++;
962 pthread_mutex_unlock(&f->zbd_info->mutex);
966 /* Anything goes as long as it is not a constant. */
967 static uint32_t pick_random_zone_idx(const struct fio_file *f,
968 const struct io_u *io_u)
970 return io_u->offset * f->zbd_info->num_open_zones / f->real_file_size;
974 * Modify the offset of an I/O unit that does not refer to an open zone such
975 * that it refers to an open zone. Close an open zone and open a new zone if
976 * necessary. This algorithm can only work correctly if all write pointers are
977 * a multiple of the fio block size. The caller must neither hold z->mutex
978 * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
980 static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
983 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
984 struct fio_file *f = io_u->file;
985 struct fio_zone_info *z;
986 unsigned int open_zone_idx = -1;
987 uint32_t zone_idx, new_zone_idx;
990 assert(is_valid_offset(f, io_u->offset));
992 if (td->o.max_open_zones || td->o.job_max_open_zones) {
994 * This statement accesses f->zbd_info->open_zones[] on purpose
997 zone_idx = f->zbd_info->open_zones[pick_random_zone_idx(f, io_u)];
999 zone_idx = zbd_zone_idx(f, io_u->offset);
1001 if (zone_idx < f->min_zone)
1002 zone_idx = f->min_zone;
1003 else if (zone_idx >= f->max_zone)
1004 zone_idx = f->max_zone - 1;
1005 dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
1006 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
1009 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
1010 * lock it can happen that the state of the zone with index zone_idx
1011 * has changed after 'z' has been assigned and before f->zbd_info->mutex
1012 * has been obtained. Hence the loop.
1017 z = &f->zbd_info->zone_info[zone_idx];
1019 zone_lock(td, f, z);
1020 pthread_mutex_lock(&f->zbd_info->mutex);
1021 if (td->o.max_open_zones == 0 && td->o.job_max_open_zones == 0)
1023 if (f->zbd_info->num_open_zones == 0) {
1024 pthread_mutex_unlock(&f->zbd_info->mutex);
1025 pthread_mutex_unlock(&z->mutex);
1026 dprint(FD_ZBD, "%s(%s): no zones are open\n",
1027 __func__, f->file_name);
1032 * List of opened zones is per-device, shared across all threads.
1033 * Start with quasi-random candidate zone.
1034 * Ignore zones which don't belong to thread's offset/size area.
1036 open_zone_idx = pick_random_zone_idx(f, io_u);
1037 assert(open_zone_idx < f->zbd_info->num_open_zones);
1038 tmp_idx = open_zone_idx;
1039 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1042 if (tmp_idx >= f->zbd_info->num_open_zones)
1044 tmpz = f->zbd_info->open_zones[tmp_idx];
1045 if (f->min_zone <= tmpz && tmpz < f->max_zone) {
1046 open_zone_idx = tmp_idx;
1047 goto found_candidate_zone;
1053 dprint(FD_ZBD, "%s(%s): no candidate zone\n",
1054 __func__, f->file_name);
1055 pthread_mutex_unlock(&f->zbd_info->mutex);
1056 pthread_mutex_unlock(&z->mutex);
1059 found_candidate_zone:
1060 new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
1061 if (new_zone_idx == zone_idx)
1063 zone_idx = new_zone_idx;
1064 pthread_mutex_unlock(&f->zbd_info->mutex);
1065 pthread_mutex_unlock(&z->mutex);
1068 /* Both z->mutex and f->zbd_info->mutex are held. */
1071 if (z->wp + min_bs <= (z+1)->start) {
1072 pthread_mutex_unlock(&f->zbd_info->mutex);
1075 dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name,
1077 if (td->o.max_open_zones || td->o.job_max_open_zones)
1078 zbd_close_zone(td, f, open_zone_idx);
1079 pthread_mutex_unlock(&f->zbd_info->mutex);
1081 /* Only z->mutex is held. */
1083 /* Zone 'z' is full, so try to open a new zone. */
1084 for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
1086 pthread_mutex_unlock(&z->mutex);
1088 if (!is_valid_offset(f, z->start)) {
1090 zone_idx = f->min_zone;
1091 z = &f->zbd_info->zone_info[zone_idx];
1093 assert(is_valid_offset(f, z->start));
1094 zone_lock(td, f, z);
1097 if (zbd_open_zone(td, io_u, zone_idx))
1101 /* Only z->mutex is held. */
1103 /* Check whether the write fits in any of the already opened zones. */
1104 pthread_mutex_lock(&f->zbd_info->mutex);
1105 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1106 zone_idx = f->zbd_info->open_zones[i];
1107 if (zone_idx < f->min_zone || zone_idx >= f->max_zone)
1109 pthread_mutex_unlock(&f->zbd_info->mutex);
1110 pthread_mutex_unlock(&z->mutex);
1112 z = &f->zbd_info->zone_info[zone_idx];
1114 zone_lock(td, f, z);
1115 if (z->wp + min_bs <= (z+1)->start)
1117 pthread_mutex_lock(&f->zbd_info->mutex);
1119 pthread_mutex_unlock(&f->zbd_info->mutex);
1120 pthread_mutex_unlock(&z->mutex);
1121 dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
1126 dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
1128 io_u->offset = z->start;
1132 /* The caller must hold z->mutex. */
1133 static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
1135 struct fio_zone_info *z)
1137 const struct fio_file *f = io_u->file;
1138 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
1140 if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
1141 pthread_mutex_unlock(&z->mutex);
1142 z = zbd_convert_to_open_zone(td, io_u);
1146 if (z->verify_block * min_bs >= f->zbd_info->zone_size)
1147 log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
1148 min_bs, (unsigned long long) f->zbd_info->zone_size);
1149 io_u->offset = z->start + z->verify_block++ * min_bs;
1154 * Find another zone for which @io_u fits below the write pointer. Start
1155 * searching in zones @zb + 1 .. @zl and continue searching in zones
1158 * Either returns NULL or returns a zone pointer and holds the mutex for that
1161 static struct fio_zone_info *
1162 zbd_find_zone(struct thread_data *td, struct io_u *io_u,
1163 struct fio_zone_info *zb, struct fio_zone_info *zl)
1165 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1166 struct fio_file *f = io_u->file;
1167 struct fio_zone_info *z1, *z2;
1168 const struct fio_zone_info *const zf =
1169 &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
1172 * Skip to the next non-empty zone in case of sequential I/O and to
1173 * the nearest non-empty zone in case of random I/O.
1175 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1176 if (z1 < zl && z1->cond != ZBD_ZONE_COND_OFFLINE) {
1177 zone_lock(td, f, z1);
1178 if (z1->start + min_bs <= z1->wp)
1180 pthread_mutex_unlock(&z1->mutex);
1181 } else if (!td_random(td)) {
1184 if (td_random(td) && z2 >= zf &&
1185 z2->cond != ZBD_ZONE_COND_OFFLINE) {
1186 zone_lock(td, f, z2);
1187 if (z2->start + min_bs <= z2->wp)
1189 pthread_mutex_unlock(&z2->mutex);
1192 dprint(FD_ZBD, "%s: adjusting random read offset failed\n",
1198 * zbd_queue_io - update the write pointer of a sequential zone
1200 * @success: Whether or not the I/O unit has been queued successfully
1201 * @q: queueing status (busy, completed or queued).
1203 * For write and trim operations, update the write pointer of the I/O unit
1206 static void zbd_queue_io(struct io_u *io_u, int q, bool success)
1208 const struct fio_file *f = io_u->file;
1209 struct zoned_block_device_info *zbd_info = f->zbd_info;
1210 struct fio_zone_info *z;
1217 zone_idx = zbd_zone_idx(f, io_u->offset);
1218 assert(zone_idx < zbd_info->nr_zones);
1219 z = &zbd_info->zone_info[zone_idx];
1221 if (!zbd_zone_swr(z))
1228 "%s: queued I/O (%lld, %llu) for zone %u\n",
1229 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1231 switch (io_u->ddir) {
1233 zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
1235 pthread_mutex_lock(&zbd_info->mutex);
1237 * z->wp > zone_end means that one or more I/O errors
1240 if (z->wp <= zone_end)
1241 zbd_info->sectors_with_data += zone_end - z->wp;
1242 pthread_mutex_unlock(&zbd_info->mutex);
1246 assert(z->wp == z->start);
1253 if (!success || q != FIO_Q_QUEUED) {
1254 /* BUSY or COMPLETED: unlock the zone */
1255 pthread_mutex_unlock(&z->mutex);
1256 io_u->zbd_put_io = NULL;
1261 * zbd_put_io - Unlock an I/O unit target zone lock
1264 static void zbd_put_io(const struct io_u *io_u)
1266 const struct fio_file *f = io_u->file;
1267 struct zoned_block_device_info *zbd_info = f->zbd_info;
1268 struct fio_zone_info *z;
1275 zone_idx = zbd_zone_idx(f, io_u->offset);
1276 assert(zone_idx < zbd_info->nr_zones);
1277 z = &zbd_info->zone_info[zone_idx];
1279 if (!zbd_zone_swr(z))
1283 "%s: terminate I/O (%lld, %llu) for zone %u\n",
1284 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1286 ret = pthread_mutex_unlock(&z->mutex);
1292 * Windows and MacOS do not define this.
1295 #define EREMOTEIO 121 /* POSIX value */
1298 bool zbd_unaligned_write(int error_code)
1300 switch (error_code) {
1309 * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives
1310 * @td: FIO thread data.
1311 * @io_u: FIO I/O unit.
1313 * For sequential workloads, change the file offset to skip zoneskip bytes when
1314 * no more IO can be performed in the current zone.
1315 * - For read workloads, zoneskip is applied when the io has reached the end of
1316 * the zone or the zone write position (when td->o.read_beyond_wp is false).
1317 * - For write workloads, zoneskip is applied when the zone is full.
1318 * This applies only to read and write operations.
1320 void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
1322 struct fio_file *f = io_u->file;
1323 enum fio_ddir ddir = io_u->ddir;
1324 struct fio_zone_info *z;
1327 assert(td->o.zone_mode == ZONE_MODE_ZBD);
1328 assert(td->o.zone_size);
1331 * zone_skip is valid only for sequential workloads.
1333 if (td_random(td) || !td->o.zone_skip)
1337 * It is time to switch to a new zone if:
1338 * - zone_bytes == zone_size bytes have already been accessed
1339 * - The last position reached the end of the current zone.
1340 * - For reads with td->o.read_beyond_wp == false, the last position
1341 * reached the zone write pointer.
1343 zone_idx = zbd_zone_idx(f, f->last_pos[ddir]);
1344 z = &f->zbd_info->zone_info[zone_idx];
1346 if (td->zone_bytes >= td->o.zone_size ||
1347 f->last_pos[ddir] >= (z+1)->start ||
1348 (ddir == DDIR_READ &&
1349 (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) {
1354 f->file_offset += td->o.zone_size + td->o.zone_skip;
1357 * Wrap from the beginning, if we exceed the file size
1359 if (f->file_offset >= f->real_file_size)
1360 f->file_offset = get_start_offset(td, f);
1362 f->last_pos[ddir] = f->file_offset;
1363 td->io_skip_bytes += td->o.zone_skip;
1368 * zbd_adjust_ddir - Adjust an I/O direction for zonemode=zbd.
1370 * @td: FIO thread data.
1371 * @io_u: FIO I/O unit.
1372 * @ddir: I/O direction before adjustment.
1374 * Return adjusted I/O direction.
1376 enum fio_ddir zbd_adjust_ddir(struct thread_data *td, struct io_u *io_u,
1380 * In case read direction is chosen for the first random I/O, fio with
1381 * zonemode=zbd stops because no data can be read from zoned block
1382 * devices with all empty zones. Overwrite the first I/O direction as
1383 * write to make sure data to read exists.
1385 if (ddir != DDIR_READ || !td_rw(td))
1388 if (io_u->file->zbd_info->sectors_with_data ||
1389 td->o.read_beyond_wp)
1396 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1397 * @td: FIO thread data.
1398 * @io_u: FIO I/O unit.
1400 * Locking strategy: returns with z->mutex locked if and only if z refers
1401 * to a sequential zone and if io_u_accept is returned. z is the zone that
1402 * corresponds to io_u->offset at the end of this function.
1404 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1406 struct fio_file *f = io_u->file;
1407 uint32_t zone_idx_b;
1408 struct fio_zone_info *zb, *zl, *orig_zb;
1409 uint32_t orig_len = io_u->buflen;
1410 uint32_t min_bs = td->o.min_bs[io_u->ddir];
1418 assert(is_valid_offset(f, io_u->offset));
1419 assert(io_u->buflen);
1420 zone_idx_b = zbd_zone_idx(f, io_u->offset);
1421 zb = &f->zbd_info->zone_info[zone_idx_b];
1424 /* Accept the I/O offset for conventional zones. */
1425 if (!zbd_zone_swr(zb))
1429 * Accept the I/O offset for reads if reading beyond the write pointer
1432 if (zb->cond != ZBD_ZONE_COND_OFFLINE &&
1433 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1438 zone_lock(td, f, zb);
1440 switch (io_u->ddir) {
1442 if (td->runstate == TD_VERIFYING) {
1444 zb = zbd_replay_write_order(td, io_u, zb);
1448 * Check that there is enough written data in the zone to do an
1449 * I/O of at least min_bs B. If there isn't, find a new zone for
1452 range = zb->cond != ZBD_ZONE_COND_OFFLINE ?
1453 zb->wp - zb->start : 0;
1454 if (range < min_bs ||
1455 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1456 pthread_mutex_unlock(&zb->mutex);
1457 zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
1458 f->file_offset + f->io_size)];
1459 zb = zbd_find_zone(td, io_u, zb, zl);
1462 "%s: zbd_find_zone(%lld, %llu) failed\n",
1463 f->file_name, io_u->offset,
1468 * zbd_find_zone() returned a zone with a range of at
1471 range = zb->wp - zb->start;
1472 assert(range >= min_bs);
1475 io_u->offset = zb->start;
1478 * Make sure the I/O is within the zone valid data range while
1479 * maximizing the I/O size and preserving randomness.
1481 if (range <= io_u->buflen)
1482 io_u->offset = zb->start;
1483 else if (td_random(td))
1484 io_u->offset = zb->start +
1485 ((io_u->offset - orig_zb->start) %
1486 (range - io_u->buflen)) / min_bs * min_bs;
1488 * Make sure the I/O does not cross over the zone wp position.
1490 new_len = min((unsigned long long)io_u->buflen,
1491 (unsigned long long)(zb->wp - io_u->offset));
1492 new_len = new_len / min_bs * min_bs;
1493 if (new_len < io_u->buflen) {
1494 io_u->buflen = new_len;
1495 dprint(FD_IO, "Changed length from %u into %llu\n",
1496 orig_len, io_u->buflen);
1498 assert(zb->start <= io_u->offset);
1499 assert(io_u->offset + io_u->buflen <= zb->wp);
1502 if (io_u->buflen > f->zbd_info->zone_size)
1504 if (!zbd_open_zone(td, io_u, zone_idx_b)) {
1505 pthread_mutex_unlock(&zb->mutex);
1506 zb = zbd_convert_to_open_zone(td, io_u);
1509 zone_idx_b = zb - f->zbd_info->zone_info;
1511 /* Check whether the zone reset threshold has been exceeded */
1512 if (td->o.zrf.u.f) {
1513 if (f->zbd_info->sectors_with_data >=
1514 f->io_size * td->o.zrt.u.f &&
1515 zbd_dec_and_reset_write_cnt(td, f)) {
1519 /* Reset the zone pointer if necessary */
1520 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
1521 assert(td->o.verify == VERIFY_NONE);
1523 * Since previous write requests may have been submitted
1524 * asynchronously and since we will submit the zone
1525 * reset synchronously, wait until previously submitted
1526 * write requests have completed before issuing a
1531 if (zbd_reset_zone(td, f, zb) < 0)
1534 /* Make writes occur at the write pointer */
1535 assert(!zbd_zone_full(f, zb, min_bs));
1536 io_u->offset = zb->wp;
1537 if (!is_valid_offset(f, io_u->offset)) {
1538 dprint(FD_ZBD, "Dropped request with offset %llu\n",
1543 * Make sure that the buflen is a multiple of the minimal
1544 * block size. Give up if shrinking would make the request too
1547 new_len = min((unsigned long long)io_u->buflen,
1548 (zb + 1)->start - io_u->offset);
1549 new_len = new_len / min_bs * min_bs;
1550 if (new_len == io_u->buflen)
1552 if (new_len >= min_bs) {
1553 io_u->buflen = new_len;
1554 dprint(FD_IO, "Changed length from %u into %llu\n",
1555 orig_len, io_u->buflen);
1558 log_err("Zone remainder %lld smaller than minimum block size %d\n",
1559 ((zb + 1)->start - io_u->offset),
1566 case DDIR_SYNC_FILE_RANGE:
1577 assert(zb->cond != ZBD_ZONE_COND_OFFLINE);
1578 assert(!io_u->zbd_queue_io);
1579 assert(!io_u->zbd_put_io);
1580 io_u->zbd_queue_io = zbd_queue_io;
1581 io_u->zbd_put_io = zbd_put_io;
1586 pthread_mutex_unlock(&zb->mutex);
1590 /* Return a string with ZBD statistics */
1591 char *zbd_write_status(const struct thread_stat *ts)
1595 if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0)