2 * Copyright (C) 2018 Western Digital Corporation or its affiliates.
4 * This file is released under the GPL.
19 #include "oslib/asprintf.h"
25 * zbd_get_zoned_model - Get a device zoned model
26 * @td: FIO thread data
27 * @f: FIO file for which to get model information
29 int zbd_get_zoned_model(struct thread_data *td, struct fio_file *f,
30 enum zbd_zoned_model *model)
34 if (td->io_ops && td->io_ops->get_zoned_model)
35 ret = td->io_ops->get_zoned_model(td, f, model);
37 ret = blkzoned_get_zoned_model(td, f, model);
39 td_verror(td, errno, "get zoned model failed");
40 log_err("%s: get zoned model failed (%d).\n",
48 * zbd_report_zones - Get zone information
49 * @td: FIO thread data.
50 * @f: FIO file for which to get zone information
51 * @offset: offset from which to report zones
52 * @zones: Array of struct zbd_zone
53 * @nr_zones: Size of @zones array
55 * Get zone information into @zones starting from the zone at offset @offset
56 * for the device specified by @f.
58 * Returns the number of zones reported upon success and a negative error code
59 * upon failure. If the zone report is empty, always assume an error (device
60 * problem) and return -EIO.
62 int zbd_report_zones(struct thread_data *td, struct fio_file *f,
63 uint64_t offset, struct zbd_zone *zones,
64 unsigned int nr_zones)
68 if (td->io_ops && td->io_ops->report_zones)
69 ret = td->io_ops->report_zones(td, f, offset, zones, nr_zones);
71 ret = blkzoned_report_zones(td, f, offset, zones, nr_zones);
73 td_verror(td, errno, "report zones failed");
74 log_err("%s: report zones from sector %llu failed (%d).\n",
75 f->file_name, (unsigned long long)offset >> 9, errno);
76 } else if (ret == 0) {
77 td_verror(td, errno, "Empty zone report");
78 log_err("%s: report zones from sector %llu is empty.\n",
79 f->file_name, (unsigned long long)offset >> 9);
87 * zbd_reset_wp - reset the write pointer of a range of zones
88 * @td: FIO thread data.
89 * @f: FIO file for which to reset zones
90 * @offset: Starting offset of the first zone to reset
91 * @length: Length of the range of zones to reset
93 * Reset the write pointer of all zones in the range @offset...@offset+@length.
94 * Returns 0 upon success and a negative error code upon failure.
96 int zbd_reset_wp(struct thread_data *td, struct fio_file *f,
97 uint64_t offset, uint64_t length)
101 if (td->io_ops && td->io_ops->reset_wp)
102 ret = td->io_ops->reset_wp(td, f, offset, length);
104 ret = blkzoned_reset_wp(td, f, offset, length);
106 td_verror(td, errno, "resetting wp failed");
107 log_err("%s: resetting wp for %llu sectors at sector %llu failed (%d).\n",
108 f->file_name, (unsigned long long)length >> 9,
109 (unsigned long long)offset >> 9, errno);
116 * zbd_zone_idx - convert an offset into a zone number
118 * @offset: offset in bytes. If this offset is in the first zone_size bytes
119 * past the disk size then the index of the sentinel is returned.
121 static uint32_t zbd_zone_idx(const struct fio_file *f, uint64_t offset)
125 if (f->zbd_info->zone_size_log2 > 0)
126 zone_idx = offset >> f->zbd_info->zone_size_log2;
128 zone_idx = offset / f->zbd_info->zone_size;
130 return min(zone_idx, f->zbd_info->nr_zones);
134 * zbd_zone_swr - Test whether a zone requires sequential writes
135 * @z: zone info pointer.
137 static inline bool zbd_zone_swr(struct fio_zone_info *z)
139 return z->type == ZBD_ZONE_TYPE_SWR;
143 * zbd_zone_full - verify whether a minimum number of bytes remain in a zone
145 * @z: zone info pointer.
146 * @required: minimum number of bytes that must remain in a zone.
148 * The caller must hold z->mutex.
150 static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
153 assert((required & 511) == 0);
155 return zbd_zone_swr(z) &&
156 z->wp + required > z->start + f->zbd_info->zone_size;
159 static void zone_lock(struct thread_data *td, struct fio_zone_info *z)
162 * Lock the io_u target zone. The zone will be unlocked if io_u offset
163 * is changed or when io_u completes and zbd_put_io() executed.
164 * To avoid multiple jobs doing asynchronous I/Os from deadlocking each
165 * other waiting for zone locks when building an io_u batch, first
166 * only trylock the zone. If the zone is already locked by another job,
167 * process the currently queued I/Os so that I/O progress is made and
170 if (pthread_mutex_trylock(&z->mutex) != 0) {
171 if (!td_ioengine_flagged(td, FIO_SYNCIO))
173 pthread_mutex_lock(&z->mutex);
177 static bool is_valid_offset(const struct fio_file *f, uint64_t offset)
179 return (uint64_t)(offset - f->file_offset) < f->io_size;
182 /* Verify whether direct I/O is used for all host-managed zoned drives. */
183 static bool zbd_using_direct_io(void)
185 struct thread_data *td;
190 if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE))
192 for_each_file(td, f, j) {
194 f->zbd_info->model == ZBD_HOST_MANAGED)
202 /* Whether or not the I/O range for f includes one or more sequential zones */
203 static bool zbd_is_seq_job(struct fio_file *f)
205 uint32_t zone_idx, zone_idx_b, zone_idx_e;
210 zone_idx_b = zbd_zone_idx(f, f->file_offset);
211 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size - 1);
212 for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++)
213 if (zbd_zone_swr(&f->zbd_info->zone_info[zone_idx]))
220 * Verify whether offset and size parameters are aligned with zone boundaries.
222 static bool zbd_verify_sizes(void)
224 const struct fio_zone_info *z;
225 struct thread_data *td;
227 uint64_t new_offset, new_end;
232 for_each_file(td, f, j) {
235 if (f->file_offset >= f->real_file_size)
237 if (!zbd_is_seq_job(f))
240 if (!td->o.zone_size) {
241 td->o.zone_size = f->zbd_info->zone_size;
242 if (!td->o.zone_size) {
243 log_err("%s: invalid 0 zone size\n",
247 } else if (td->o.zone_size != f->zbd_info->zone_size) {
248 log_err("%s: job parameter zonesize %llu does not match disk zone size %llu.\n",
249 f->file_name, (unsigned long long) td->o.zone_size,
250 (unsigned long long) f->zbd_info->zone_size);
254 if (td->o.zone_skip &&
255 (td->o.zone_skip < td->o.zone_size ||
256 td->o.zone_skip % td->o.zone_size)) {
257 log_err("%s: zoneskip %llu is not a multiple of the device zone size %llu.\n",
258 f->file_name, (unsigned long long) td->o.zone_skip,
259 (unsigned long long) td->o.zone_size);
263 zone_idx = zbd_zone_idx(f, f->file_offset);
264 z = &f->zbd_info->zone_info[zone_idx];
265 if ((f->file_offset != z->start) &&
266 (td->o.td_ddir != TD_DDIR_READ)) {
267 new_offset = (z+1)->start;
268 if (new_offset >= f->file_offset + f->io_size) {
269 log_info("%s: io_size must be at least one zone\n",
273 log_info("%s: rounded up offset from %llu to %llu\n",
274 f->file_name, (unsigned long long) f->file_offset,
275 (unsigned long long) new_offset);
276 f->io_size -= (new_offset - f->file_offset);
277 f->file_offset = new_offset;
279 zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size);
280 z = &f->zbd_info->zone_info[zone_idx];
282 if ((td->o.td_ddir != TD_DDIR_READ) &&
283 (f->file_offset + f->io_size != new_end)) {
284 if (new_end <= f->file_offset) {
285 log_info("%s: io_size must be at least one zone\n",
289 log_info("%s: rounded down io_size from %llu to %llu\n",
290 f->file_name, (unsigned long long) f->io_size,
291 (unsigned long long) new_end - f->file_offset);
292 f->io_size = new_end - f->file_offset;
300 static bool zbd_verify_bs(void)
302 struct thread_data *td;
308 for_each_file(td, f, j) {
311 zone_size = f->zbd_info->zone_size;
312 for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
313 if (td->o.verify != VERIFY_NONE &&
314 zone_size % td->o.bs[k] != 0) {
315 log_info("%s: block size %llu is not a divisor of the zone size %d\n",
316 f->file_name, td->o.bs[k],
326 static int ilog2(uint64_t i)
338 * Initialize f->zbd_info for devices that are not zoned block devices. This
339 * allows to execute a ZBD workload against a non-ZBD device.
341 static int init_zone_info(struct thread_data *td, struct fio_file *f)
344 struct fio_zone_info *p;
345 uint64_t zone_size = td->o.zone_size;
346 struct zoned_block_device_info *zbd_info = NULL;
347 pthread_mutexattr_t attr;
350 if (zone_size == 0) {
351 log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n",
356 if (zone_size < 512) {
357 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
362 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
363 zbd_info = scalloc(1, sizeof(*zbd_info) +
364 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
368 pthread_mutexattr_init(&attr);
369 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
370 pthread_mutexattr_setpshared(&attr, true);
371 pthread_mutex_init(&zbd_info->mutex, &attr);
372 zbd_info->refcount = 1;
373 p = &zbd_info->zone_info[0];
374 for (i = 0; i < nr_zones; i++, p++) {
375 pthread_mutex_init(&p->mutex, &attr);
376 p->start = i * zone_size;
377 p->wp = p->start + zone_size;
378 p->type = ZBD_ZONE_TYPE_SWR;
379 p->cond = ZBD_ZONE_COND_EMPTY;
382 p->start = nr_zones * zone_size;
384 f->zbd_info = zbd_info;
385 f->zbd_info->zone_size = zone_size;
386 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
387 ilog2(zone_size) : 0;
388 f->zbd_info->nr_zones = nr_zones;
389 pthread_mutexattr_destroy(&attr);
394 * Maximum number of zones to report in one operation.
396 #define ZBD_REPORT_MAX_ZONES 8192U
399 * Parse the device zone report and store it in f->zbd_info. Must be called
400 * only for devices that are zoned, namely those with a model != ZBD_NONE.
402 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
405 struct zbd_zone *zones, *z;
406 struct fio_zone_info *p;
407 uint64_t zone_size, offset;
408 struct zoned_block_device_info *zbd_info = NULL;
409 pthread_mutexattr_t attr;
412 pthread_mutexattr_init(&attr);
413 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
414 pthread_mutexattr_setpshared(&attr, true);
416 zones = calloc(ZBD_REPORT_MAX_ZONES, sizeof(struct zbd_zone));
420 nrz = zbd_report_zones(td, f, 0, zones, ZBD_REPORT_MAX_ZONES);
423 log_info("fio: report zones (offset 0) failed for %s (%d).\n",
428 zone_size = zones[0].len;
429 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
431 if (td->o.zone_size == 0) {
432 td->o.zone_size = zone_size;
433 } else if (td->o.zone_size != zone_size) {
434 log_err("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n",
435 f->file_name, (unsigned long long) td->o.zone_size,
436 (unsigned long long) zone_size);
441 dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name,
442 nr_zones, (unsigned long long) zone_size / 1024);
444 zbd_info = scalloc(1, sizeof(*zbd_info) +
445 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
449 pthread_mutex_init(&zbd_info->mutex, &attr);
450 zbd_info->refcount = 1;
451 p = &zbd_info->zone_info[0];
452 for (offset = 0, j = 0; j < nr_zones;) {
454 for (i = 0; i < nrz; i++, j++, z++, p++) {
455 pthread_mutex_init(&p->mutex, &attr);
458 case ZBD_ZONE_COND_NOT_WP:
459 case ZBD_ZONE_COND_FULL:
460 p->wp = p->start + zone_size;
463 assert(z->start <= z->wp);
464 assert(z->wp <= z->start + zone_size);
470 if (j > 0 && p->start != p[-1].start + zone_size) {
471 log_info("%s: invalid zone data\n",
478 offset = z->start + z->len;
481 nrz = zbd_report_zones(td, f, offset,
482 zones, ZBD_REPORT_MAX_ZONES);
485 log_info("fio: report zones (offset %llu) failed for %s (%d).\n",
486 (unsigned long long)offset,
493 zbd_info->zone_info[nr_zones].start = offset;
495 f->zbd_info = zbd_info;
496 f->zbd_info->zone_size = zone_size;
497 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
498 ilog2(zone_size) : 0;
499 f->zbd_info->nr_zones = nr_zones;
506 pthread_mutexattr_destroy(&attr);
511 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
513 * Returns 0 upon success and a negative error code upon failure.
515 static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
517 enum zbd_zoned_model zbd_model;
520 assert(td->o.zone_mode == ZONE_MODE_ZBD);
522 ret = zbd_get_zoned_model(td, f, &zbd_model);
530 case ZBD_HOST_MANAGED:
531 ret = parse_zone_info(td, f);
534 ret = init_zone_info(td, f);
537 td_verror(td, EINVAL, "Unsupported zoned model");
538 log_err("Unsupported zoned model\n");
543 f->zbd_info->model = zbd_model;
547 void zbd_free_zone_info(struct fio_file *f)
553 pthread_mutex_lock(&f->zbd_info->mutex);
554 refcount = --f->zbd_info->refcount;
555 pthread_mutex_unlock(&f->zbd_info->mutex);
557 assert((int32_t)refcount >= 0);
564 * Initialize f->zbd_info.
566 * Returns 0 upon success and a negative error code upon failure.
568 * Note: this function can only work correctly if it is called before the first
571 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
573 struct thread_data *td2;
577 for_each_td(td2, i) {
578 for_each_file(td2, f2, j) {
579 if (td2 == td && f2 == file)
582 strcmp(f2->file_name, file->file_name) != 0)
584 file->zbd_info = f2->zbd_info;
585 file->zbd_info->refcount++;
590 ret = zbd_create_zone_info(td, file);
592 td_verror(td, -ret, "zbd_create_zone_info() failed");
596 int zbd_setup_files(struct thread_data *td)
601 for_each_file(td, f, i) {
602 if (zbd_init_zone_info(td, f))
606 if (!zbd_using_direct_io()) {
607 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
611 if (!zbd_verify_sizes())
614 if (!zbd_verify_bs())
621 * zbd_reset_range - reset zones for a range of sectors
622 * @td: FIO thread data.
623 * @f: Fio file for which to reset zones
624 * @sector: Starting sector in units of 512 bytes
625 * @nr_sectors: Number of sectors in units of 512 bytes
627 * Returns 0 upon success and a negative error code upon failure.
629 static int zbd_reset_range(struct thread_data *td, struct fio_file *f,
630 uint64_t offset, uint64_t length)
632 uint32_t zone_idx_b, zone_idx_e;
633 struct fio_zone_info *zb, *ze, *z;
636 assert(is_valid_offset(f, offset + length - 1));
638 switch (f->zbd_info->model) {
640 case ZBD_HOST_MANAGED:
641 ret = zbd_reset_wp(td, f, offset, length);
649 zone_idx_b = zbd_zone_idx(f, offset);
650 zb = &f->zbd_info->zone_info[zone_idx_b];
651 zone_idx_e = zbd_zone_idx(f, offset + length);
652 ze = &f->zbd_info->zone_info[zone_idx_e];
653 for (z = zb; z < ze; z++) {
654 pthread_mutex_lock(&z->mutex);
655 pthread_mutex_lock(&f->zbd_info->mutex);
656 f->zbd_info->sectors_with_data -= z->wp - z->start;
657 pthread_mutex_unlock(&f->zbd_info->mutex);
660 pthread_mutex_unlock(&z->mutex);
663 td->ts.nr_zone_resets += ze - zb;
668 static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info,
669 struct fio_zone_info *zone)
671 return zone - zbd_info->zone_info;
675 * zbd_reset_zone - reset the write pointer of a single zone
676 * @td: FIO thread data.
677 * @f: FIO file associated with the disk for which to reset a write pointer.
680 * Returns 0 upon success and a negative error code upon failure.
682 static int zbd_reset_zone(struct thread_data *td, struct fio_file *f,
683 struct fio_zone_info *z)
685 dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
686 zbd_zone_nr(f->zbd_info, z));
688 return zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
691 /* The caller must hold f->zbd_info->mutex */
692 static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
693 unsigned int open_zone_idx)
697 assert(open_zone_idx < f->zbd_info->num_open_zones);
698 zone_idx = f->zbd_info->open_zones[open_zone_idx];
699 memmove(f->zbd_info->open_zones + open_zone_idx,
700 f->zbd_info->open_zones + open_zone_idx + 1,
701 (ZBD_MAX_OPEN_ZONES - (open_zone_idx + 1)) *
702 sizeof(f->zbd_info->open_zones[0]));
703 f->zbd_info->num_open_zones--;
704 f->zbd_info->zone_info[zone_idx].open = 0;
708 * Reset a range of zones. Returns 0 upon success and 1 upon failure.
709 * @td: fio thread data.
710 * @f: fio file for which to reset zones
711 * @zb: first zone to reset.
712 * @ze: first zone not to reset.
713 * @all_zones: whether to reset all zones or only those zones for which the
714 * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE].
716 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
717 struct fio_zone_info *const zb,
718 struct fio_zone_info *const ze, bool all_zones)
720 struct fio_zone_info *z;
721 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
727 dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name,
728 zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze));
729 for (z = zb; z < ze; z++) {
730 uint32_t nz = z - f->zbd_info->zone_info;
732 if (!zbd_zone_swr(z))
738 pthread_mutex_lock(&f->zbd_info->mutex);
739 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
740 if (f->zbd_info->open_zones[i] == nz)
741 zbd_close_zone(td, f, i);
743 pthread_mutex_unlock(&f->zbd_info->mutex);
745 reset_wp = z->wp != z->start;
747 reset_wp = z->wp % min_bs != 0;
750 dprint(FD_ZBD, "%s: resetting zone %u\n",
752 zbd_zone_nr(f->zbd_info, z));
753 if (zbd_reset_zone(td, f, z) < 0)
756 pthread_mutex_unlock(&z->mutex);
763 * Reset zbd_info.write_cnt, the counter that counts down towards the next
766 static void zbd_reset_write_cnt(const struct thread_data *td,
767 const struct fio_file *f)
769 assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
771 pthread_mutex_lock(&f->zbd_info->mutex);
772 f->zbd_info->write_cnt = td->o.zrf.u.f ?
773 min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
774 pthread_mutex_unlock(&f->zbd_info->mutex);
777 static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
778 const struct fio_file *f)
780 uint32_t write_cnt = 0;
782 pthread_mutex_lock(&f->zbd_info->mutex);
783 assert(f->zbd_info->write_cnt);
784 if (f->zbd_info->write_cnt)
785 write_cnt = --f->zbd_info->write_cnt;
787 zbd_reset_write_cnt(td, f);
788 pthread_mutex_unlock(&f->zbd_info->mutex);
790 return write_cnt == 0;
798 /* Calculate the number of sectors with data (swd) and perform action 'a' */
799 static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a)
801 struct fio_zone_info *zb, *ze, *z;
804 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
805 ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset +
807 for (z = zb; z < ze; z++) {
808 pthread_mutex_lock(&z->mutex);
809 swd += z->wp - z->start;
811 pthread_mutex_lock(&f->zbd_info->mutex);
814 assert(f->zbd_info->sectors_with_data == swd);
817 f->zbd_info->sectors_with_data = swd;
820 pthread_mutex_unlock(&f->zbd_info->mutex);
821 for (z = zb; z < ze; z++)
822 pthread_mutex_unlock(&z->mutex);
828 * The swd check is useful for debugging but takes too much time to leave
829 * it enabled all the time. Hence it is disabled by default.
831 static const bool enable_check_swd = false;
833 /* Check whether the value of zbd_info.sectors_with_data is correct. */
834 static void zbd_check_swd(const struct fio_file *f)
836 if (!enable_check_swd)
839 zbd_process_swd(f, CHECK_SWD);
842 static void zbd_init_swd(struct fio_file *f)
846 if (!enable_check_swd)
849 swd = zbd_process_swd(f, SET_SWD);
850 dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name,
854 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
856 struct fio_zone_info *zb, *ze;
859 if (!f->zbd_info || !td_write(td))
862 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
863 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size);
864 ze = &f->zbd_info->zone_info[zone_idx_e];
867 * If data verification is enabled reset the affected zones before
868 * writing any data to avoid that a zone reset has to be issued while
869 * writing data, which causes data loss.
871 zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE &&
872 td->runstate != TD_VERIFYING);
873 zbd_reset_write_cnt(td, f);
876 /* The caller must hold f->zbd_info->mutex. */
877 static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
878 unsigned int zone_idx)
880 struct zoned_block_device_info *zbdi = f->zbd_info;
883 assert(td->o.max_open_zones <= ARRAY_SIZE(zbdi->open_zones));
884 assert(zbdi->num_open_zones <= td->o.max_open_zones);
886 for (i = 0; i < zbdi->num_open_zones; i++)
887 if (zbdi->open_zones[i] == zone_idx)
894 * Open a ZBD zone if it was not yet open. Returns true if either the zone was
895 * already open or if opening a new zone is allowed. Returns false if the zone
896 * was not yet open and opening a new zone would cause the zone limit to be
899 static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
902 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
903 const struct fio_file *f = io_u->file;
904 struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
907 if (z->cond == ZBD_ZONE_COND_OFFLINE)
911 * Skip full zones with data verification enabled because resetting a
912 * zone causes data loss and hence causes verification to fail.
914 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
917 /* Zero means no limit */
918 if (!td->o.max_open_zones)
921 pthread_mutex_lock(&f->zbd_info->mutex);
922 if (is_zone_open(td, f, zone_idx))
925 if (f->zbd_info->num_open_zones >= td->o.max_open_zones)
927 dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx);
928 f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx;
933 pthread_mutex_unlock(&f->zbd_info->mutex);
937 /* Anything goes as long as it is not a constant. */
938 static uint32_t pick_random_zone_idx(const struct fio_file *f,
939 const struct io_u *io_u)
941 return io_u->offset * f->zbd_info->num_open_zones / f->real_file_size;
945 * Modify the offset of an I/O unit that does not refer to an open zone such
946 * that it refers to an open zone. Close an open zone and open a new zone if
947 * necessary. This algorithm can only work correctly if all write pointers are
948 * a multiple of the fio block size. The caller must neither hold z->mutex
949 * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
951 static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
954 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
955 const struct fio_file *f = io_u->file;
956 struct fio_zone_info *z;
957 unsigned int open_zone_idx = -1;
958 uint32_t zone_idx, new_zone_idx;
961 assert(is_valid_offset(f, io_u->offset));
963 if (td->o.max_open_zones) {
965 * This statement accesses f->zbd_info->open_zones[] on purpose
968 zone_idx = f->zbd_info->open_zones[pick_random_zone_idx(f, io_u)];
970 zone_idx = zbd_zone_idx(f, io_u->offset);
972 dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
973 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
976 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
977 * lock it can happen that the state of the zone with index zone_idx
978 * has changed after 'z' has been assigned and before f->zbd_info->mutex
979 * has been obtained. Hence the loop.
984 z = &f->zbd_info->zone_info[zone_idx];
987 pthread_mutex_lock(&f->zbd_info->mutex);
988 if (td->o.max_open_zones == 0)
990 if (f->zbd_info->num_open_zones == 0) {
991 pthread_mutex_unlock(&f->zbd_info->mutex);
992 pthread_mutex_unlock(&z->mutex);
993 dprint(FD_ZBD, "%s(%s): no zones are open\n",
994 __func__, f->file_name);
999 * List of opened zones is per-device, shared across all threads.
1000 * Start with quasi-random candidate zone.
1001 * Ignore zones which don't belong to thread's offset/size area.
1003 open_zone_idx = pick_random_zone_idx(f, io_u);
1004 assert(open_zone_idx < f->zbd_info->num_open_zones);
1005 tmp_idx = open_zone_idx;
1006 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1009 if (tmp_idx >= f->zbd_info->num_open_zones)
1011 tmpz = f->zbd_info->open_zones[tmp_idx];
1013 if (is_valid_offset(f, f->zbd_info->zone_info[tmpz].start)) {
1014 open_zone_idx = tmp_idx;
1015 goto found_candidate_zone;
1021 dprint(FD_ZBD, "%s(%s): no candidate zone\n",
1022 __func__, f->file_name);
1023 pthread_mutex_unlock(&f->zbd_info->mutex);
1024 pthread_mutex_unlock(&z->mutex);
1027 found_candidate_zone:
1028 new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
1029 if (new_zone_idx == zone_idx)
1031 zone_idx = new_zone_idx;
1032 pthread_mutex_unlock(&f->zbd_info->mutex);
1033 pthread_mutex_unlock(&z->mutex);
1036 /* Both z->mutex and f->zbd_info->mutex are held. */
1039 if (z->wp + min_bs <= (z+1)->start) {
1040 pthread_mutex_unlock(&f->zbd_info->mutex);
1043 dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name,
1045 if (td->o.max_open_zones)
1046 zbd_close_zone(td, f, open_zone_idx);
1047 pthread_mutex_unlock(&f->zbd_info->mutex);
1049 /* Only z->mutex is held. */
1051 /* Zone 'z' is full, so try to open a new zone. */
1052 for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
1054 pthread_mutex_unlock(&z->mutex);
1056 if (!is_valid_offset(f, z->start)) {
1058 zone_idx = zbd_zone_idx(f, f->file_offset);
1059 z = &f->zbd_info->zone_info[zone_idx];
1061 assert(is_valid_offset(f, z->start));
1065 if (zbd_open_zone(td, io_u, zone_idx))
1069 /* Only z->mutex is held. */
1071 /* Check whether the write fits in any of the already opened zones. */
1072 pthread_mutex_lock(&f->zbd_info->mutex);
1073 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1074 zone_idx = f->zbd_info->open_zones[i];
1075 pthread_mutex_unlock(&f->zbd_info->mutex);
1076 pthread_mutex_unlock(&z->mutex);
1078 z = &f->zbd_info->zone_info[zone_idx];
1081 if (z->wp + min_bs <= (z+1)->start)
1083 pthread_mutex_lock(&f->zbd_info->mutex);
1085 pthread_mutex_unlock(&f->zbd_info->mutex);
1086 pthread_mutex_unlock(&z->mutex);
1087 dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
1092 dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
1094 io_u->offset = z->start;
1098 /* The caller must hold z->mutex. */
1099 static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
1101 struct fio_zone_info *z)
1103 const struct fio_file *f = io_u->file;
1104 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
1106 if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
1107 pthread_mutex_unlock(&z->mutex);
1108 z = zbd_convert_to_open_zone(td, io_u);
1112 if (z->verify_block * min_bs >= f->zbd_info->zone_size)
1113 log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
1114 min_bs, (unsigned long long) f->zbd_info->zone_size);
1115 io_u->offset = z->start + z->verify_block++ * min_bs;
1120 * Find another zone for which @io_u fits below the write pointer. Start
1121 * searching in zones @zb + 1 .. @zl and continue searching in zones
1124 * Either returns NULL or returns a zone pointer and holds the mutex for that
1127 static struct fio_zone_info *
1128 zbd_find_zone(struct thread_data *td, struct io_u *io_u,
1129 struct fio_zone_info *zb, struct fio_zone_info *zl)
1131 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1132 const struct fio_file *f = io_u->file;
1133 struct fio_zone_info *z1, *z2;
1134 const struct fio_zone_info *const zf =
1135 &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
1138 * Skip to the next non-empty zone in case of sequential I/O and to
1139 * the nearest non-empty zone in case of random I/O.
1141 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1142 if (z1 < zl && z1->cond != ZBD_ZONE_COND_OFFLINE) {
1144 if (z1->start + min_bs <= z1->wp)
1146 pthread_mutex_unlock(&z1->mutex);
1147 } else if (!td_random(td)) {
1150 if (td_random(td) && z2 >= zf &&
1151 z2->cond != ZBD_ZONE_COND_OFFLINE) {
1153 if (z2->start + min_bs <= z2->wp)
1155 pthread_mutex_unlock(&z2->mutex);
1158 dprint(FD_ZBD, "%s: adjusting random read offset failed\n",
1164 * zbd_queue_io - update the write pointer of a sequential zone
1166 * @success: Whether or not the I/O unit has been queued successfully
1167 * @q: queueing status (busy, completed or queued).
1169 * For write and trim operations, update the write pointer of the I/O unit
1172 static void zbd_queue_io(struct io_u *io_u, int q, bool success)
1174 const struct fio_file *f = io_u->file;
1175 struct zoned_block_device_info *zbd_info = f->zbd_info;
1176 struct fio_zone_info *z;
1183 zone_idx = zbd_zone_idx(f, io_u->offset);
1184 assert(zone_idx < zbd_info->nr_zones);
1185 z = &zbd_info->zone_info[zone_idx];
1187 if (!zbd_zone_swr(z))
1194 "%s: queued I/O (%lld, %llu) for zone %u\n",
1195 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1197 switch (io_u->ddir) {
1199 zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
1201 pthread_mutex_lock(&zbd_info->mutex);
1203 * z->wp > zone_end means that one or more I/O errors
1206 if (z->wp <= zone_end)
1207 zbd_info->sectors_with_data += zone_end - z->wp;
1208 pthread_mutex_unlock(&zbd_info->mutex);
1212 assert(z->wp == z->start);
1219 if (!success || q != FIO_Q_QUEUED) {
1220 /* BUSY or COMPLETED: unlock the zone */
1221 pthread_mutex_unlock(&z->mutex);
1222 io_u->zbd_put_io = NULL;
1227 * zbd_put_io - Unlock an I/O unit target zone lock
1230 static void zbd_put_io(const struct io_u *io_u)
1232 const struct fio_file *f = io_u->file;
1233 struct zoned_block_device_info *zbd_info = f->zbd_info;
1234 struct fio_zone_info *z;
1241 zone_idx = zbd_zone_idx(f, io_u->offset);
1242 assert(zone_idx < zbd_info->nr_zones);
1243 z = &zbd_info->zone_info[zone_idx];
1245 if (!zbd_zone_swr(z))
1249 "%s: terminate I/O (%lld, %llu) for zone %u\n",
1250 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1252 ret = pthread_mutex_unlock(&z->mutex);
1258 * Windows and MacOS do not define this.
1261 #define EREMOTEIO 121 /* POSIX value */
1264 bool zbd_unaligned_write(int error_code)
1266 switch (error_code) {
1275 * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives
1276 * @td: FIO thread data.
1277 * @io_u: FIO I/O unit.
1279 * For sequential workloads, change the file offset to skip zoneskip bytes when
1280 * no more IO can be performed in the current zone.
1281 * - For read workloads, zoneskip is applied when the io has reached the end of
1282 * the zone or the zone write position (when td->o.read_beyond_wp is false).
1283 * - For write workloads, zoneskip is applied when the zone is full.
1284 * This applies only to read and write operations.
1286 void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
1288 struct fio_file *f = io_u->file;
1289 enum fio_ddir ddir = io_u->ddir;
1290 struct fio_zone_info *z;
1293 assert(td->o.zone_mode == ZONE_MODE_ZBD);
1294 assert(td->o.zone_size);
1297 * zone_skip is valid only for sequential workloads.
1299 if (td_random(td) || !td->o.zone_skip)
1303 * It is time to switch to a new zone if:
1304 * - zone_bytes == zone_size bytes have already been accessed
1305 * - The last position reached the end of the current zone.
1306 * - For reads with td->o.read_beyond_wp == false, the last position
1307 * reached the zone write pointer.
1309 zone_idx = zbd_zone_idx(f, f->last_pos[ddir]);
1310 z = &f->zbd_info->zone_info[zone_idx];
1312 if (td->zone_bytes >= td->o.zone_size ||
1313 f->last_pos[ddir] >= (z+1)->start ||
1314 (ddir == DDIR_READ &&
1315 (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) {
1320 f->file_offset += td->o.zone_size + td->o.zone_skip;
1323 * Wrap from the beginning, if we exceed the file size
1325 if (f->file_offset >= f->real_file_size)
1326 f->file_offset = get_start_offset(td, f);
1328 f->last_pos[ddir] = f->file_offset;
1329 td->io_skip_bytes += td->o.zone_skip;
1334 * zbd_adjust_ddir - Adjust an I/O direction for zonedmode=zbd.
1336 * @td: FIO thread data.
1337 * @io_u: FIO I/O unit.
1338 * @ddir: I/O direction before adjustment.
1340 * Return adjusted I/O direction.
1342 enum fio_ddir zbd_adjust_ddir(struct thread_data *td, struct io_u *io_u,
1346 * In case read direction is chosen for the first random I/O, fio with
1347 * zonemode=zbd stops because no data can be read from zoned block
1348 * devices with all empty zones. Overwrite the first I/O direction as
1349 * write to make sure data to read exists.
1351 if (ddir != DDIR_READ || !td_rw(td))
1354 if (io_u->file->zbd_info->sectors_with_data ||
1355 td->o.read_beyond_wp)
1362 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1363 * @td: FIO thread data.
1364 * @io_u: FIO I/O unit.
1366 * Locking strategy: returns with z->mutex locked if and only if z refers
1367 * to a sequential zone and if io_u_accept is returned. z is the zone that
1368 * corresponds to io_u->offset at the end of this function.
1370 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1372 struct fio_file *f = io_u->file;
1373 uint32_t zone_idx_b;
1374 struct fio_zone_info *zb, *zl, *orig_zb;
1375 uint32_t orig_len = io_u->buflen;
1376 uint32_t min_bs = td->o.min_bs[io_u->ddir];
1384 assert(is_valid_offset(f, io_u->offset));
1385 assert(io_u->buflen);
1386 zone_idx_b = zbd_zone_idx(f, io_u->offset);
1387 zb = &f->zbd_info->zone_info[zone_idx_b];
1390 /* Accept the I/O offset for conventional zones. */
1391 if (!zbd_zone_swr(zb))
1395 * Accept the I/O offset for reads if reading beyond the write pointer
1398 if (zb->cond != ZBD_ZONE_COND_OFFLINE &&
1399 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1406 switch (io_u->ddir) {
1408 if (td->runstate == TD_VERIFYING) {
1410 zb = zbd_replay_write_order(td, io_u, zb);
1414 * Check that there is enough written data in the zone to do an
1415 * I/O of at least min_bs B. If there isn't, find a new zone for
1418 range = zb->cond != ZBD_ZONE_COND_OFFLINE ?
1419 zb->wp - zb->start : 0;
1420 if (range < min_bs ||
1421 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1422 pthread_mutex_unlock(&zb->mutex);
1423 zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
1424 f->file_offset + f->io_size)];
1425 zb = zbd_find_zone(td, io_u, zb, zl);
1428 "%s: zbd_find_zone(%lld, %llu) failed\n",
1429 f->file_name, io_u->offset,
1434 * zbd_find_zone() returned a zone with a range of at
1437 range = zb->wp - zb->start;
1438 assert(range >= min_bs);
1441 io_u->offset = zb->start;
1444 * Make sure the I/O is within the zone valid data range while
1445 * maximizing the I/O size and preserving randomness.
1447 if (range <= io_u->buflen)
1448 io_u->offset = zb->start;
1449 else if (td_random(td))
1450 io_u->offset = zb->start +
1451 ((io_u->offset - orig_zb->start) %
1452 (range - io_u->buflen)) / min_bs * min_bs;
1454 * Make sure the I/O does not cross over the zone wp position.
1456 new_len = min((unsigned long long)io_u->buflen,
1457 (unsigned long long)(zb->wp - io_u->offset));
1458 new_len = new_len / min_bs * min_bs;
1459 if (new_len < io_u->buflen) {
1460 io_u->buflen = new_len;
1461 dprint(FD_IO, "Changed length from %u into %llu\n",
1462 orig_len, io_u->buflen);
1464 assert(zb->start <= io_u->offset);
1465 assert(io_u->offset + io_u->buflen <= zb->wp);
1468 if (io_u->buflen > f->zbd_info->zone_size)
1470 if (!zbd_open_zone(td, io_u, zone_idx_b)) {
1471 pthread_mutex_unlock(&zb->mutex);
1472 zb = zbd_convert_to_open_zone(td, io_u);
1475 zone_idx_b = zb - f->zbd_info->zone_info;
1477 /* Check whether the zone reset threshold has been exceeded */
1478 if (td->o.zrf.u.f) {
1479 if (f->zbd_info->sectors_with_data >=
1480 f->io_size * td->o.zrt.u.f &&
1481 zbd_dec_and_reset_write_cnt(td, f)) {
1485 /* Reset the zone pointer if necessary */
1486 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
1487 assert(td->o.verify == VERIFY_NONE);
1489 * Since previous write requests may have been submitted
1490 * asynchronously and since we will submit the zone
1491 * reset synchronously, wait until previously submitted
1492 * write requests have completed before issuing a
1497 if (zbd_reset_zone(td, f, zb) < 0)
1500 /* Make writes occur at the write pointer */
1501 assert(!zbd_zone_full(f, zb, min_bs));
1502 io_u->offset = zb->wp;
1503 if (!is_valid_offset(f, io_u->offset)) {
1504 dprint(FD_ZBD, "Dropped request with offset %llu\n",
1509 * Make sure that the buflen is a multiple of the minimal
1510 * block size. Give up if shrinking would make the request too
1513 new_len = min((unsigned long long)io_u->buflen,
1514 (zb + 1)->start - io_u->offset);
1515 new_len = new_len / min_bs * min_bs;
1516 if (new_len == io_u->buflen)
1518 if (new_len >= min_bs) {
1519 io_u->buflen = new_len;
1520 dprint(FD_IO, "Changed length from %u into %llu\n",
1521 orig_len, io_u->buflen);
1524 log_err("Zone remainder %lld smaller than minimum block size %d\n",
1525 ((zb + 1)->start - io_u->offset),
1532 case DDIR_SYNC_FILE_RANGE:
1543 assert(zb->cond != ZBD_ZONE_COND_OFFLINE);
1544 assert(!io_u->zbd_queue_io);
1545 assert(!io_u->zbd_put_io);
1546 io_u->zbd_queue_io = zbd_queue_io;
1547 io_u->zbd_put_io = zbd_put_io;
1552 pthread_mutex_unlock(&zb->mutex);
1556 /* Return a string with ZBD statistics */
1557 char *zbd_write_status(const struct thread_stat *ts)
1561 if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0)