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_end - Return zone end location
145 * @z: zone info pointer.
147 static inline uint64_t zbd_zone_end(const struct fio_zone_info *z)
153 * zbd_zone_capacity_end - Return zone capacity limit end location
154 * @z: zone info pointer.
156 static inline uint64_t zbd_zone_capacity_end(const struct fio_zone_info *z)
158 return z->start + z->capacity;
162 * zbd_zone_full - verify whether a minimum number of bytes remain in a zone
164 * @z: zone info pointer.
165 * @required: minimum number of bytes that must remain in a zone.
167 * The caller must hold z->mutex.
169 static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
172 assert((required & 511) == 0);
174 return zbd_zone_swr(z) &&
175 z->wp + required > zbd_zone_capacity_end(z);
178 static void zone_lock(struct thread_data *td, struct fio_file *f, struct fio_zone_info *z)
180 struct zoned_block_device_info *zbd = f->zbd_info;
181 uint32_t nz = z - zbd->zone_info;
183 /* A thread should never lock zones outside its working area. */
184 assert(f->min_zone <= nz && nz < f->max_zone);
187 * Lock the io_u target zone. The zone will be unlocked if io_u offset
188 * is changed or when io_u completes and zbd_put_io() executed.
189 * To avoid multiple jobs doing asynchronous I/Os from deadlocking each
190 * other waiting for zone locks when building an io_u batch, first
191 * only trylock the zone. If the zone is already locked by another job,
192 * process the currently queued I/Os so that I/O progress is made and
195 if (pthread_mutex_trylock(&z->mutex) != 0) {
196 if (!td_ioengine_flagged(td, FIO_SYNCIO))
198 pthread_mutex_lock(&z->mutex);
202 static bool is_valid_offset(const struct fio_file *f, uint64_t offset)
204 return (uint64_t)(offset - f->file_offset) < f->io_size;
207 /* Verify whether direct I/O is used for all host-managed zoned drives. */
208 static bool zbd_using_direct_io(void)
210 struct thread_data *td;
215 if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE))
217 for_each_file(td, f, j) {
219 f->zbd_info->model == ZBD_HOST_MANAGED)
227 /* Whether or not the I/O range for f includes one or more sequential zones */
228 static bool zbd_is_seq_job(struct fio_file *f)
230 uint32_t zone_idx, zone_idx_b, zone_idx_e;
235 zone_idx_b = zbd_zone_idx(f, f->file_offset);
236 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size - 1);
237 for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++)
238 if (zbd_zone_swr(&f->zbd_info->zone_info[zone_idx]))
245 * Verify whether offset and size parameters are aligned with zone boundaries.
247 static bool zbd_verify_sizes(void)
249 const struct fio_zone_info *z;
250 struct thread_data *td;
252 uint64_t new_offset, new_end;
257 for_each_file(td, f, j) {
260 if (f->file_offset >= f->real_file_size)
262 if (!zbd_is_seq_job(f))
265 if (!td->o.zone_size) {
266 td->o.zone_size = f->zbd_info->zone_size;
267 if (!td->o.zone_size) {
268 log_err("%s: invalid 0 zone size\n",
272 } else if (td->o.zone_size != f->zbd_info->zone_size) {
273 log_err("%s: job parameter zonesize %llu does not match disk zone size %llu.\n",
274 f->file_name, (unsigned long long) td->o.zone_size,
275 (unsigned long long) f->zbd_info->zone_size);
279 if (td->o.zone_skip &&
280 (td->o.zone_skip < td->o.zone_size ||
281 td->o.zone_skip % td->o.zone_size)) {
282 log_err("%s: zoneskip %llu is not a multiple of the device zone size %llu.\n",
283 f->file_name, (unsigned long long) td->o.zone_skip,
284 (unsigned long long) td->o.zone_size);
288 zone_idx = zbd_zone_idx(f, f->file_offset);
289 z = &f->zbd_info->zone_info[zone_idx];
290 if ((f->file_offset != z->start) &&
291 (td->o.td_ddir != TD_DDIR_READ)) {
292 new_offset = zbd_zone_end(z);
293 if (new_offset >= f->file_offset + f->io_size) {
294 log_info("%s: io_size must be at least one zone\n",
298 log_info("%s: rounded up offset from %llu to %llu\n",
299 f->file_name, (unsigned long long) f->file_offset,
300 (unsigned long long) new_offset);
301 f->io_size -= (new_offset - f->file_offset);
302 f->file_offset = new_offset;
304 zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size);
305 z = &f->zbd_info->zone_info[zone_idx];
307 if ((td->o.td_ddir != TD_DDIR_READ) &&
308 (f->file_offset + f->io_size != new_end)) {
309 if (new_end <= f->file_offset) {
310 log_info("%s: io_size must be at least one zone\n",
314 log_info("%s: rounded down io_size from %llu to %llu\n",
315 f->file_name, (unsigned long long) f->io_size,
316 (unsigned long long) new_end - f->file_offset);
317 f->io_size = new_end - f->file_offset;
320 f->min_zone = zbd_zone_idx(f, f->file_offset);
321 f->max_zone = zbd_zone_idx(f, f->file_offset + f->io_size);
322 assert(f->min_zone < f->max_zone);
329 static bool zbd_verify_bs(void)
331 struct thread_data *td;
337 for_each_file(td, f, j) {
340 zone_size = f->zbd_info->zone_size;
341 for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
342 if (td->o.verify != VERIFY_NONE &&
343 zone_size % td->o.bs[k] != 0) {
344 log_info("%s: block size %llu is not a divisor of the zone size %d\n",
345 f->file_name, td->o.bs[k],
355 static int ilog2(uint64_t i)
367 * Initialize f->zbd_info for devices that are not zoned block devices. This
368 * allows to execute a ZBD workload against a non-ZBD device.
370 static int init_zone_info(struct thread_data *td, struct fio_file *f)
373 struct fio_zone_info *p;
374 uint64_t zone_size = td->o.zone_size;
375 uint64_t zone_capacity = td->o.zone_capacity;
376 struct zoned_block_device_info *zbd_info = NULL;
379 if (zone_size == 0) {
380 log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n",
385 if (zone_size < 512) {
386 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
391 if (zone_capacity == 0)
392 zone_capacity = zone_size;
394 if (zone_capacity > zone_size) {
395 log_err("%s: job parameter zonecapacity %llu is larger than zone size %llu\n",
396 f->file_name, (unsigned long long) td->o.zone_capacity,
397 (unsigned long long) td->o.zone_size);
401 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
402 zbd_info = scalloc(1, sizeof(*zbd_info) +
403 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
407 mutex_init_pshared(&zbd_info->mutex);
408 zbd_info->refcount = 1;
409 p = &zbd_info->zone_info[0];
410 for (i = 0; i < nr_zones; i++, p++) {
411 mutex_init_pshared_with_type(&p->mutex,
412 PTHREAD_MUTEX_RECURSIVE);
413 p->start = i * zone_size;
415 p->type = ZBD_ZONE_TYPE_SWR;
416 p->cond = ZBD_ZONE_COND_EMPTY;
417 p->capacity = zone_capacity;
420 p->start = nr_zones * zone_size;
422 f->zbd_info = zbd_info;
423 f->zbd_info->zone_size = zone_size;
424 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
425 ilog2(zone_size) : 0;
426 f->zbd_info->nr_zones = nr_zones;
431 * Maximum number of zones to report in one operation.
433 #define ZBD_REPORT_MAX_ZONES 8192U
436 * Parse the device zone report and store it in f->zbd_info. Must be called
437 * only for devices that are zoned, namely those with a model != ZBD_NONE.
439 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
442 struct zbd_zone *zones, *z;
443 struct fio_zone_info *p;
444 uint64_t zone_size, offset;
445 struct zoned_block_device_info *zbd_info = NULL;
448 zones = calloc(ZBD_REPORT_MAX_ZONES, sizeof(struct zbd_zone));
452 nrz = zbd_report_zones(td, f, 0, zones, ZBD_REPORT_MAX_ZONES);
455 log_info("fio: report zones (offset 0) failed for %s (%d).\n",
460 zone_size = zones[0].len;
461 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
463 if (td->o.zone_size == 0) {
464 td->o.zone_size = zone_size;
465 } else if (td->o.zone_size != zone_size) {
466 log_err("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n",
467 f->file_name, (unsigned long long) td->o.zone_size,
468 (unsigned long long) zone_size);
473 dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name,
474 nr_zones, (unsigned long long) zone_size / 1024);
476 zbd_info = scalloc(1, sizeof(*zbd_info) +
477 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
481 mutex_init_pshared(&zbd_info->mutex);
482 zbd_info->refcount = 1;
483 p = &zbd_info->zone_info[0];
484 for (offset = 0, j = 0; j < nr_zones;) {
486 for (i = 0; i < nrz; i++, j++, z++, p++) {
487 mutex_init_pshared_with_type(&p->mutex,
488 PTHREAD_MUTEX_RECURSIVE);
490 p->capacity = z->capacity;
492 case ZBD_ZONE_COND_NOT_WP:
493 case ZBD_ZONE_COND_FULL:
494 p->wp = p->start + p->capacity;
497 assert(z->start <= z->wp);
498 assert(z->wp <= z->start + zone_size);
504 if (j > 0 && p->start != p[-1].start + zone_size) {
505 log_info("%s: invalid zone data\n",
512 offset = z->start + z->len;
515 nrz = zbd_report_zones(td, f, offset,
516 zones, ZBD_REPORT_MAX_ZONES);
519 log_info("fio: report zones (offset %llu) failed for %s (%d).\n",
520 (unsigned long long)offset,
527 zbd_info->zone_info[nr_zones].start = offset;
529 f->zbd_info = zbd_info;
530 f->zbd_info->zone_size = zone_size;
531 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
532 ilog2(zone_size) : 0;
533 f->zbd_info->nr_zones = nr_zones;
544 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
546 * Returns 0 upon success and a negative error code upon failure.
548 static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
550 enum zbd_zoned_model zbd_model;
553 assert(td->o.zone_mode == ZONE_MODE_ZBD);
555 ret = zbd_get_zoned_model(td, f, &zbd_model);
563 case ZBD_HOST_MANAGED:
564 ret = parse_zone_info(td, f);
567 ret = init_zone_info(td, f);
570 td_verror(td, EINVAL, "Unsupported zoned model");
571 log_err("Unsupported zoned model\n");
576 f->zbd_info->model = zbd_model;
577 f->zbd_info->max_open_zones = td->o.max_open_zones;
582 void zbd_free_zone_info(struct fio_file *f)
588 pthread_mutex_lock(&f->zbd_info->mutex);
589 refcount = --f->zbd_info->refcount;
590 pthread_mutex_unlock(&f->zbd_info->mutex);
592 assert((int32_t)refcount >= 0);
599 * Initialize f->zbd_info.
601 * Returns 0 upon success and a negative error code upon failure.
603 * Note: this function can only work correctly if it is called before the first
606 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
608 struct thread_data *td2;
612 for_each_td(td2, i) {
613 for_each_file(td2, f2, j) {
614 if (td2 == td && f2 == file)
617 strcmp(f2->file_name, file->file_name) != 0)
619 file->zbd_info = f2->zbd_info;
620 file->zbd_info->refcount++;
625 ret = zbd_create_zone_info(td, file);
627 td_verror(td, -ret, "zbd_create_zone_info() failed");
631 int zbd_setup_files(struct thread_data *td)
636 for_each_file(td, f, i) {
637 if (zbd_init_zone_info(td, f))
641 if (!zbd_using_direct_io()) {
642 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
646 if (!zbd_verify_sizes())
649 if (!zbd_verify_bs())
652 for_each_file(td, f, i) {
653 struct zoned_block_device_info *zbd = f->zbd_info;
658 zbd->max_open_zones = zbd->max_open_zones ?: ZBD_MAX_OPEN_ZONES;
660 if (td->o.max_open_zones > 0 &&
661 zbd->max_open_zones != td->o.max_open_zones) {
662 log_err("Different 'max_open_zones' values\n");
665 if (zbd->max_open_zones > ZBD_MAX_OPEN_ZONES) {
666 log_err("'max_open_zones' value is limited by %u\n", ZBD_MAX_OPEN_ZONES);
674 static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info,
675 struct fio_zone_info *zone)
677 return zone - zbd_info->zone_info;
681 * zbd_reset_zone - reset the write pointer of a single zone
682 * @td: FIO thread data.
683 * @f: FIO file associated with the disk for which to reset a write pointer.
686 * Returns 0 upon success and a negative error code upon failure.
688 * The caller must hold z->mutex.
690 static int zbd_reset_zone(struct thread_data *td, struct fio_file *f,
691 struct fio_zone_info *z)
693 uint64_t offset = z->start;
694 uint64_t length = (z+1)->start - offset;
697 assert(is_valid_offset(f, offset + length - 1));
699 dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
700 zbd_zone_nr(f->zbd_info, z));
701 switch (f->zbd_info->model) {
703 case ZBD_HOST_MANAGED:
704 ret = zbd_reset_wp(td, f, offset, length);
712 pthread_mutex_lock(&f->zbd_info->mutex);
713 f->zbd_info->sectors_with_data -= z->wp - z->start;
714 pthread_mutex_unlock(&f->zbd_info->mutex);
718 td->ts.nr_zone_resets++;
723 /* The caller must hold f->zbd_info->mutex */
724 static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
725 unsigned int zone_idx)
727 uint32_t open_zone_idx = 0;
729 for (; open_zone_idx < f->zbd_info->num_open_zones; open_zone_idx++) {
730 if (f->zbd_info->open_zones[open_zone_idx] == zone_idx)
733 if (open_zone_idx == f->zbd_info->num_open_zones) {
734 dprint(FD_ZBD, "%s: zone %d is not open\n",
735 f->file_name, zone_idx);
739 dprint(FD_ZBD, "%s: closing zone %d\n", f->file_name, zone_idx);
740 memmove(f->zbd_info->open_zones + open_zone_idx,
741 f->zbd_info->open_zones + open_zone_idx + 1,
742 (ZBD_MAX_OPEN_ZONES - (open_zone_idx + 1)) *
743 sizeof(f->zbd_info->open_zones[0]));
744 f->zbd_info->num_open_zones--;
745 td->num_open_zones--;
746 f->zbd_info->zone_info[zone_idx].open = 0;
750 * Reset a range of zones. Returns 0 upon success and 1 upon failure.
751 * @td: fio thread data.
752 * @f: fio file for which to reset zones
753 * @zb: first zone to reset.
754 * @ze: first zone not to reset.
755 * @all_zones: whether to reset all zones or only those zones for which the
756 * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE].
758 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
759 struct fio_zone_info *const zb,
760 struct fio_zone_info *const ze, bool all_zones)
762 struct fio_zone_info *z;
763 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
769 dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name,
770 zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze));
771 for (z = zb; z < ze; z++) {
772 uint32_t nz = z - f->zbd_info->zone_info;
774 if (!zbd_zone_swr(z))
778 pthread_mutex_lock(&f->zbd_info->mutex);
779 zbd_close_zone(td, f, nz);
780 pthread_mutex_unlock(&f->zbd_info->mutex);
782 reset_wp = z->wp != z->start;
784 reset_wp = z->wp % min_bs != 0;
787 dprint(FD_ZBD, "%s: resetting zone %u\n",
789 zbd_zone_nr(f->zbd_info, z));
790 if (zbd_reset_zone(td, f, z) < 0)
793 pthread_mutex_unlock(&z->mutex);
800 * Reset zbd_info.write_cnt, the counter that counts down towards the next
803 static void _zbd_reset_write_cnt(const struct thread_data *td,
804 const struct fio_file *f)
806 assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
808 f->zbd_info->write_cnt = td->o.zrf.u.f ?
809 min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
812 static void zbd_reset_write_cnt(const struct thread_data *td,
813 const struct fio_file *f)
815 pthread_mutex_lock(&f->zbd_info->mutex);
816 _zbd_reset_write_cnt(td, f);
817 pthread_mutex_unlock(&f->zbd_info->mutex);
820 static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
821 const struct fio_file *f)
823 uint32_t write_cnt = 0;
825 pthread_mutex_lock(&f->zbd_info->mutex);
826 assert(f->zbd_info->write_cnt);
827 if (f->zbd_info->write_cnt)
828 write_cnt = --f->zbd_info->write_cnt;
830 _zbd_reset_write_cnt(td, f);
831 pthread_mutex_unlock(&f->zbd_info->mutex);
833 return write_cnt == 0;
841 /* Calculate the number of sectors with data (swd) and perform action 'a' */
842 static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a)
844 struct fio_zone_info *zb, *ze, *z;
847 zb = &f->zbd_info->zone_info[f->min_zone];
848 ze = &f->zbd_info->zone_info[f->max_zone];
849 for (z = zb; z < ze; z++) {
850 pthread_mutex_lock(&z->mutex);
851 swd += z->wp - z->start;
853 pthread_mutex_lock(&f->zbd_info->mutex);
856 assert(f->zbd_info->sectors_with_data == swd);
859 f->zbd_info->sectors_with_data = swd;
862 pthread_mutex_unlock(&f->zbd_info->mutex);
863 for (z = zb; z < ze; z++)
864 pthread_mutex_unlock(&z->mutex);
870 * The swd check is useful for debugging but takes too much time to leave
871 * it enabled all the time. Hence it is disabled by default.
873 static const bool enable_check_swd = false;
875 /* Check whether the value of zbd_info.sectors_with_data is correct. */
876 static void zbd_check_swd(const struct fio_file *f)
878 if (!enable_check_swd)
881 zbd_process_swd(f, CHECK_SWD);
884 static void zbd_init_swd(struct fio_file *f)
888 if (!enable_check_swd)
891 swd = zbd_process_swd(f, SET_SWD);
892 dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name,
896 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
898 struct fio_zone_info *zb, *ze;
900 if (!f->zbd_info || !td_write(td))
903 zb = &f->zbd_info->zone_info[f->min_zone];
904 ze = &f->zbd_info->zone_info[f->max_zone];
907 * If data verification is enabled reset the affected zones before
908 * writing any data to avoid that a zone reset has to be issued while
909 * writing data, which causes data loss.
911 zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE &&
912 td->runstate != TD_VERIFYING);
913 zbd_reset_write_cnt(td, f);
916 /* The caller must hold f->zbd_info->mutex. */
917 static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
918 unsigned int zone_idx)
920 struct zoned_block_device_info *zbdi = f->zbd_info;
923 assert(td->o.job_max_open_zones == 0 || td->num_open_zones <= td->o.job_max_open_zones);
924 assert(td->o.job_max_open_zones <= zbdi->max_open_zones);
925 assert(zbdi->num_open_zones <= zbdi->max_open_zones);
927 for (i = 0; i < zbdi->num_open_zones; i++)
928 if (zbdi->open_zones[i] == zone_idx)
935 * Open a ZBD zone if it was not yet open. Returns true if either the zone was
936 * already open or if opening a new zone is allowed. Returns false if the zone
937 * was not yet open and opening a new zone would cause the zone limit to be
940 static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
943 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
944 const struct fio_file *f = io_u->file;
945 struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
948 if (z->cond == ZBD_ZONE_COND_OFFLINE)
952 * Skip full zones with data verification enabled because resetting a
953 * zone causes data loss and hence causes verification to fail.
955 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
958 pthread_mutex_lock(&f->zbd_info->mutex);
959 if (is_zone_open(td, f, zone_idx)) {
961 * If the zone is already open and going to be full by writes
962 * in-flight, handle it as a full zone instead of an open zone.
964 if (z->wp >= zbd_zone_capacity_end(z))
969 /* Zero means no limit */
970 if (td->o.job_max_open_zones > 0 &&
971 td->num_open_zones >= td->o.job_max_open_zones)
973 if (f->zbd_info->num_open_zones >= f->zbd_info->max_open_zones)
975 dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx);
976 f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx;
977 td->num_open_zones++;
982 pthread_mutex_unlock(&f->zbd_info->mutex);
986 /* Anything goes as long as it is not a constant. */
987 static uint32_t pick_random_zone_idx(const struct fio_file *f,
988 const struct io_u *io_u)
990 return io_u->offset * f->zbd_info->num_open_zones / f->real_file_size;
994 * Modify the offset of an I/O unit that does not refer to an open zone such
995 * that it refers to an open zone. Close an open zone and open a new zone if
996 * necessary. This algorithm can only work correctly if all write pointers are
997 * a multiple of the fio block size. The caller must neither hold z->mutex
998 * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
1000 static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
1003 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1004 struct fio_file *f = io_u->file;
1005 struct fio_zone_info *z;
1006 unsigned int open_zone_idx = -1;
1007 uint32_t zone_idx, new_zone_idx;
1009 bool wait_zone_close;
1011 assert(is_valid_offset(f, io_u->offset));
1013 if (td->o.max_open_zones || td->o.job_max_open_zones) {
1015 * This statement accesses f->zbd_info->open_zones[] on purpose
1018 zone_idx = f->zbd_info->open_zones[pick_random_zone_idx(f, io_u)];
1020 zone_idx = zbd_zone_idx(f, io_u->offset);
1022 if (zone_idx < f->min_zone)
1023 zone_idx = f->min_zone;
1024 else if (zone_idx >= f->max_zone)
1025 zone_idx = f->max_zone - 1;
1026 dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
1027 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
1030 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
1031 * lock it can happen that the state of the zone with index zone_idx
1032 * has changed after 'z' has been assigned and before f->zbd_info->mutex
1033 * has been obtained. Hence the loop.
1038 z = &f->zbd_info->zone_info[zone_idx];
1040 zone_lock(td, f, z);
1041 pthread_mutex_lock(&f->zbd_info->mutex);
1042 if (td->o.max_open_zones == 0 && td->o.job_max_open_zones == 0)
1044 if (f->zbd_info->num_open_zones == 0) {
1045 dprint(FD_ZBD, "%s(%s): no zones are open\n",
1046 __func__, f->file_name);
1047 goto open_other_zone;
1051 * List of opened zones is per-device, shared across all threads.
1052 * Start with quasi-random candidate zone.
1053 * Ignore zones which don't belong to thread's offset/size area.
1055 open_zone_idx = pick_random_zone_idx(f, io_u);
1056 assert(open_zone_idx < f->zbd_info->num_open_zones);
1057 tmp_idx = open_zone_idx;
1058 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1061 if (tmp_idx >= f->zbd_info->num_open_zones)
1063 tmpz = f->zbd_info->open_zones[tmp_idx];
1064 if (f->min_zone <= tmpz && tmpz < f->max_zone) {
1065 open_zone_idx = tmp_idx;
1066 goto found_candidate_zone;
1072 dprint(FD_ZBD, "%s(%s): no candidate zone\n",
1073 __func__, f->file_name);
1074 pthread_mutex_unlock(&f->zbd_info->mutex);
1075 pthread_mutex_unlock(&z->mutex);
1078 found_candidate_zone:
1079 new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
1080 if (new_zone_idx == zone_idx)
1082 zone_idx = new_zone_idx;
1083 pthread_mutex_unlock(&f->zbd_info->mutex);
1084 pthread_mutex_unlock(&z->mutex);
1087 /* Both z->mutex and f->zbd_info->mutex are held. */
1090 if (z->wp + min_bs <= zbd_zone_capacity_end(z)) {
1091 pthread_mutex_unlock(&f->zbd_info->mutex);
1096 /* Check if number of open zones reaches one of limits. */
1098 f->zbd_info->num_open_zones == f->max_zone - f->min_zone ||
1099 (td->o.max_open_zones &&
1100 f->zbd_info->num_open_zones == td->o.max_open_zones) ||
1101 (td->o.job_max_open_zones &&
1102 td->num_open_zones == td->o.job_max_open_zones);
1104 pthread_mutex_unlock(&f->zbd_info->mutex);
1106 /* Only z->mutex is held. */
1109 * When number of open zones reaches to one of limits, wait for
1110 * zone close before opening a new zone.
1112 if (wait_zone_close) {
1113 dprint(FD_ZBD, "%s(%s): quiesce to allow open zones to close\n",
1114 __func__, f->file_name);
1118 /* Zone 'z' is full, so try to open a new zone. */
1119 for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
1121 pthread_mutex_unlock(&z->mutex);
1123 if (!is_valid_offset(f, z->start)) {
1125 zone_idx = f->min_zone;
1126 z = &f->zbd_info->zone_info[zone_idx];
1128 assert(is_valid_offset(f, z->start));
1129 zone_lock(td, f, z);
1132 if (zbd_open_zone(td, io_u, zone_idx))
1136 /* Only z->mutex is held. */
1138 /* Check whether the write fits in any of the already opened zones. */
1139 pthread_mutex_lock(&f->zbd_info->mutex);
1140 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1141 zone_idx = f->zbd_info->open_zones[i];
1142 if (zone_idx < f->min_zone || zone_idx >= f->max_zone)
1144 pthread_mutex_unlock(&f->zbd_info->mutex);
1145 pthread_mutex_unlock(&z->mutex);
1147 z = &f->zbd_info->zone_info[zone_idx];
1149 zone_lock(td, f, z);
1150 if (z->wp + min_bs <= zbd_zone_capacity_end(z))
1152 pthread_mutex_lock(&f->zbd_info->mutex);
1154 pthread_mutex_unlock(&f->zbd_info->mutex);
1155 pthread_mutex_unlock(&z->mutex);
1156 dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
1161 dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
1163 io_u->offset = z->start;
1167 /* The caller must hold z->mutex. */
1168 static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
1170 struct fio_zone_info *z)
1172 const struct fio_file *f = io_u->file;
1173 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
1175 if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
1176 pthread_mutex_unlock(&z->mutex);
1177 z = zbd_convert_to_open_zone(td, io_u);
1181 if (z->verify_block * min_bs >= z->capacity)
1182 log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
1183 min_bs, (unsigned long long)z->capacity);
1184 io_u->offset = z->start + z->verify_block++ * min_bs;
1189 * Find another zone for which @io_u fits below the write pointer. Start
1190 * searching in zones @zb + 1 .. @zl and continue searching in zones
1193 * Either returns NULL or returns a zone pointer and holds the mutex for that
1196 static struct fio_zone_info *
1197 zbd_find_zone(struct thread_data *td, struct io_u *io_u,
1198 struct fio_zone_info *zb, struct fio_zone_info *zl)
1200 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1201 struct fio_file *f = io_u->file;
1202 struct fio_zone_info *z1, *z2;
1203 const struct fio_zone_info *const zf =
1204 &f->zbd_info->zone_info[f->min_zone];
1207 * Skip to the next non-empty zone in case of sequential I/O and to
1208 * the nearest non-empty zone in case of random I/O.
1210 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1211 if (z1 < zl && z1->cond != ZBD_ZONE_COND_OFFLINE) {
1212 zone_lock(td, f, z1);
1213 if (z1->start + min_bs <= z1->wp)
1215 pthread_mutex_unlock(&z1->mutex);
1216 } else if (!td_random(td)) {
1219 if (td_random(td) && z2 >= zf &&
1220 z2->cond != ZBD_ZONE_COND_OFFLINE) {
1221 zone_lock(td, f, z2);
1222 if (z2->start + min_bs <= z2->wp)
1224 pthread_mutex_unlock(&z2->mutex);
1227 dprint(FD_ZBD, "%s: adjusting random read offset failed\n",
1233 * zbd_end_zone_io - update zone status at command completion
1235 * @z: zone info pointer
1237 * If the write command made the zone full, close it.
1239 * The caller must hold z->mutex.
1241 static void zbd_end_zone_io(struct thread_data *td, const struct io_u *io_u,
1242 struct fio_zone_info *z)
1244 const struct fio_file *f = io_u->file;
1246 if (io_u->ddir == DDIR_WRITE &&
1247 io_u->offset + io_u->buflen >= zbd_zone_capacity_end(z)) {
1248 pthread_mutex_lock(&f->zbd_info->mutex);
1249 zbd_close_zone(td, f, z - f->zbd_info->zone_info);
1250 pthread_mutex_unlock(&f->zbd_info->mutex);
1255 * zbd_queue_io - update the write pointer of a sequential zone
1257 * @success: Whether or not the I/O unit has been queued successfully
1258 * @q: queueing status (busy, completed or queued).
1260 * For write and trim operations, update the write pointer of the I/O unit
1263 static void zbd_queue_io(struct thread_data *td, struct io_u *io_u, int q,
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))
1286 "%s: queued I/O (%lld, %llu) for zone %u\n",
1287 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1289 switch (io_u->ddir) {
1291 zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
1292 zbd_zone_capacity_end(z));
1293 pthread_mutex_lock(&zbd_info->mutex);
1295 * z->wp > zone_end means that one or more I/O errors
1298 if (z->wp <= zone_end)
1299 zbd_info->sectors_with_data += zone_end - z->wp;
1300 pthread_mutex_unlock(&zbd_info->mutex);
1304 assert(z->wp == z->start);
1310 if (q == FIO_Q_COMPLETED && !io_u->error)
1311 zbd_end_zone_io(td, io_u, z);
1314 if (!success || q != FIO_Q_QUEUED) {
1315 /* BUSY or COMPLETED: unlock the zone */
1316 pthread_mutex_unlock(&z->mutex);
1317 io_u->zbd_put_io = NULL;
1322 * zbd_put_io - Unlock an I/O unit target zone lock
1325 static void zbd_put_io(struct thread_data *td, const struct io_u *io_u)
1327 const struct fio_file *f = io_u->file;
1328 struct zoned_block_device_info *zbd_info = f->zbd_info;
1329 struct fio_zone_info *z;
1336 zone_idx = zbd_zone_idx(f, io_u->offset);
1337 assert(zone_idx < zbd_info->nr_zones);
1338 z = &zbd_info->zone_info[zone_idx];
1340 if (!zbd_zone_swr(z))
1344 "%s: terminate I/O (%lld, %llu) for zone %u\n",
1345 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1347 zbd_end_zone_io(td, io_u, z);
1349 ret = pthread_mutex_unlock(&z->mutex);
1355 * Windows and MacOS do not define this.
1358 #define EREMOTEIO 121 /* POSIX value */
1361 bool zbd_unaligned_write(int error_code)
1363 switch (error_code) {
1372 * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives
1373 * @td: FIO thread data.
1374 * @io_u: FIO I/O unit.
1376 * For sequential workloads, change the file offset to skip zoneskip bytes when
1377 * no more IO can be performed in the current zone.
1378 * - For read workloads, zoneskip is applied when the io has reached the end of
1379 * the zone or the zone write position (when td->o.read_beyond_wp is false).
1380 * - For write workloads, zoneskip is applied when the zone is full.
1381 * This applies only to read and write operations.
1383 void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
1385 struct fio_file *f = io_u->file;
1386 enum fio_ddir ddir = io_u->ddir;
1387 struct fio_zone_info *z;
1390 assert(td->o.zone_mode == ZONE_MODE_ZBD);
1391 assert(td->o.zone_size);
1393 zone_idx = zbd_zone_idx(f, f->last_pos[ddir]);
1394 z = &f->zbd_info->zone_info[zone_idx];
1397 * When the zone capacity is smaller than the zone size and the I/O is
1398 * sequential write, skip to zone end if the latest position is at the
1399 * zone capacity limit.
1401 if (z->capacity < f->zbd_info->zone_size && !td_random(td) &&
1402 ddir == DDIR_WRITE &&
1403 f->last_pos[ddir] >= zbd_zone_capacity_end(z)) {
1405 "%s: Jump from zone capacity limit to zone end:"
1406 " (%llu -> %llu) for zone %u (%llu)\n",
1407 f->file_name, (unsigned long long) f->last_pos[ddir],
1408 (unsigned long long) zbd_zone_end(z),
1409 zbd_zone_nr(f->zbd_info, z),
1410 (unsigned long long) z->capacity);
1411 td->io_skip_bytes += zbd_zone_end(z) - f->last_pos[ddir];
1412 f->last_pos[ddir] = zbd_zone_end(z);
1416 * zone_skip is valid only for sequential workloads.
1418 if (td_random(td) || !td->o.zone_skip)
1422 * It is time to switch to a new zone if:
1423 * - zone_bytes == zone_size bytes have already been accessed
1424 * - The last position reached the end of the current zone.
1425 * - For reads with td->o.read_beyond_wp == false, the last position
1426 * reached the zone write pointer.
1428 if (td->zone_bytes >= td->o.zone_size ||
1429 f->last_pos[ddir] >= zbd_zone_end(z) ||
1430 (ddir == DDIR_READ &&
1431 (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) {
1436 f->file_offset += td->o.zone_size + td->o.zone_skip;
1439 * Wrap from the beginning, if we exceed the file size
1441 if (f->file_offset >= f->real_file_size)
1442 f->file_offset = get_start_offset(td, f);
1444 f->last_pos[ddir] = f->file_offset;
1445 td->io_skip_bytes += td->o.zone_skip;
1450 * zbd_adjust_ddir - Adjust an I/O direction for zonemode=zbd.
1452 * @td: FIO thread data.
1453 * @io_u: FIO I/O unit.
1454 * @ddir: I/O direction before adjustment.
1456 * Return adjusted I/O direction.
1458 enum fio_ddir zbd_adjust_ddir(struct thread_data *td, struct io_u *io_u,
1462 * In case read direction is chosen for the first random I/O, fio with
1463 * zonemode=zbd stops because no data can be read from zoned block
1464 * devices with all empty zones. Overwrite the first I/O direction as
1465 * write to make sure data to read exists.
1467 if (ddir != DDIR_READ || !td_rw(td))
1470 if (io_u->file->zbd_info->sectors_with_data ||
1471 td->o.read_beyond_wp)
1478 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1479 * @td: FIO thread data.
1480 * @io_u: FIO I/O unit.
1482 * Locking strategy: returns with z->mutex locked if and only if z refers
1483 * to a sequential zone and if io_u_accept is returned. z is the zone that
1484 * corresponds to io_u->offset at the end of this function.
1486 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1488 struct fio_file *f = io_u->file;
1489 uint32_t zone_idx_b;
1490 struct fio_zone_info *zb, *zl, *orig_zb;
1491 uint32_t orig_len = io_u->buflen;
1492 uint32_t min_bs = td->o.min_bs[io_u->ddir];
1500 assert(is_valid_offset(f, io_u->offset));
1501 assert(io_u->buflen);
1502 zone_idx_b = zbd_zone_idx(f, io_u->offset);
1503 zb = &f->zbd_info->zone_info[zone_idx_b];
1506 /* Accept the I/O offset for conventional zones. */
1507 if (!zbd_zone_swr(zb))
1511 * Accept the I/O offset for reads if reading beyond the write pointer
1514 if (zb->cond != ZBD_ZONE_COND_OFFLINE &&
1515 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1520 zone_lock(td, f, zb);
1522 switch (io_u->ddir) {
1524 if (td->runstate == TD_VERIFYING && td_write(td)) {
1525 zb = zbd_replay_write_order(td, io_u, zb);
1529 * Check that there is enough written data in the zone to do an
1530 * I/O of at least min_bs B. If there isn't, find a new zone for
1533 range = zb->cond != ZBD_ZONE_COND_OFFLINE ?
1534 zb->wp - zb->start : 0;
1535 if (range < min_bs ||
1536 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1537 pthread_mutex_unlock(&zb->mutex);
1538 zl = &f->zbd_info->zone_info[f->max_zone];
1539 zb = zbd_find_zone(td, io_u, zb, zl);
1542 "%s: zbd_find_zone(%lld, %llu) failed\n",
1543 f->file_name, io_u->offset,
1548 * zbd_find_zone() returned a zone with a range of at
1551 range = zb->wp - zb->start;
1552 assert(range >= min_bs);
1555 io_u->offset = zb->start;
1558 * Make sure the I/O is within the zone valid data range while
1559 * maximizing the I/O size and preserving randomness.
1561 if (range <= io_u->buflen)
1562 io_u->offset = zb->start;
1563 else if (td_random(td))
1564 io_u->offset = zb->start +
1565 ((io_u->offset - orig_zb->start) %
1566 (range - io_u->buflen)) / min_bs * min_bs;
1568 * Make sure the I/O does not cross over the zone wp position.
1570 new_len = min((unsigned long long)io_u->buflen,
1571 (unsigned long long)(zb->wp - io_u->offset));
1572 new_len = new_len / min_bs * min_bs;
1573 if (new_len < io_u->buflen) {
1574 io_u->buflen = new_len;
1575 dprint(FD_IO, "Changed length from %u into %llu\n",
1576 orig_len, io_u->buflen);
1578 assert(zb->start <= io_u->offset);
1579 assert(io_u->offset + io_u->buflen <= zb->wp);
1582 if (io_u->buflen > f->zbd_info->zone_size)
1584 if (!zbd_open_zone(td, io_u, zone_idx_b)) {
1585 pthread_mutex_unlock(&zb->mutex);
1586 zb = zbd_convert_to_open_zone(td, io_u);
1589 zone_idx_b = zb - f->zbd_info->zone_info;
1591 /* Check whether the zone reset threshold has been exceeded */
1592 if (td->o.zrf.u.f) {
1593 if (f->zbd_info->sectors_with_data >=
1594 f->io_size * td->o.zrt.u.f &&
1595 zbd_dec_and_reset_write_cnt(td, f)) {
1599 /* Reset the zone pointer if necessary */
1600 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
1601 assert(td->o.verify == VERIFY_NONE);
1603 * Since previous write requests may have been submitted
1604 * asynchronously and since we will submit the zone
1605 * reset synchronously, wait until previously submitted
1606 * write requests have completed before issuing a
1611 if (zbd_reset_zone(td, f, zb) < 0)
1614 if (zb->capacity < min_bs) {
1615 log_err("zone capacity %llu smaller than minimum block size %d\n",
1616 (unsigned long long)zb->capacity,
1621 /* Make writes occur at the write pointer */
1622 assert(!zbd_zone_full(f, zb, min_bs));
1623 io_u->offset = zb->wp;
1624 if (!is_valid_offset(f, io_u->offset)) {
1625 dprint(FD_ZBD, "Dropped request with offset %llu\n",
1630 * Make sure that the buflen is a multiple of the minimal
1631 * block size. Give up if shrinking would make the request too
1634 new_len = min((unsigned long long)io_u->buflen,
1635 zbd_zone_capacity_end(zb) - io_u->offset);
1636 new_len = new_len / min_bs * min_bs;
1637 if (new_len == io_u->buflen)
1639 if (new_len >= min_bs) {
1640 io_u->buflen = new_len;
1641 dprint(FD_IO, "Changed length from %u into %llu\n",
1642 orig_len, io_u->buflen);
1645 log_err("Zone remainder %lld smaller than minimum block size %d\n",
1646 (zbd_zone_capacity_end(zb) - io_u->offset),
1653 case DDIR_SYNC_FILE_RANGE:
1664 assert(zb->cond != ZBD_ZONE_COND_OFFLINE);
1665 assert(!io_u->zbd_queue_io);
1666 assert(!io_u->zbd_put_io);
1667 io_u->zbd_queue_io = zbd_queue_io;
1668 io_u->zbd_put_io = zbd_put_io;
1673 pthread_mutex_unlock(&zb->mutex);
1677 /* Return a string with ZBD statistics */
1678 char *zbd_write_status(const struct thread_stat *ts)
1682 if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0)