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);
328 static bool zbd_verify_bs(void)
330 struct thread_data *td;
336 for_each_file(td, f, j) {
339 zone_size = f->zbd_info->zone_size;
340 for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
341 if (td->o.verify != VERIFY_NONE &&
342 zone_size % td->o.bs[k] != 0) {
343 log_info("%s: block size %llu is not a divisor of the zone size %d\n",
344 f->file_name, td->o.bs[k],
354 static int ilog2(uint64_t i)
366 * Initialize f->zbd_info for devices that are not zoned block devices. This
367 * allows to execute a ZBD workload against a non-ZBD device.
369 static int init_zone_info(struct thread_data *td, struct fio_file *f)
372 struct fio_zone_info *p;
373 uint64_t zone_size = td->o.zone_size;
374 struct zoned_block_device_info *zbd_info = NULL;
377 if (zone_size == 0) {
378 log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n",
383 if (zone_size < 512) {
384 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
389 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
390 zbd_info = scalloc(1, sizeof(*zbd_info) +
391 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
395 mutex_init_pshared(&zbd_info->mutex);
396 zbd_info->refcount = 1;
397 p = &zbd_info->zone_info[0];
398 for (i = 0; i < nr_zones; i++, p++) {
399 mutex_init_pshared_with_type(&p->mutex,
400 PTHREAD_MUTEX_RECURSIVE);
401 p->start = i * zone_size;
403 p->type = ZBD_ZONE_TYPE_SWR;
404 p->cond = ZBD_ZONE_COND_EMPTY;
405 p->capacity = zone_size;
408 p->start = nr_zones * zone_size;
410 f->zbd_info = zbd_info;
411 f->zbd_info->zone_size = zone_size;
412 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
413 ilog2(zone_size) : 0;
414 f->zbd_info->nr_zones = nr_zones;
419 * Maximum number of zones to report in one operation.
421 #define ZBD_REPORT_MAX_ZONES 8192U
424 * Parse the device zone report and store it in f->zbd_info. Must be called
425 * only for devices that are zoned, namely those with a model != ZBD_NONE.
427 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
430 struct zbd_zone *zones, *z;
431 struct fio_zone_info *p;
432 uint64_t zone_size, offset;
433 struct zoned_block_device_info *zbd_info = NULL;
436 zones = calloc(ZBD_REPORT_MAX_ZONES, sizeof(struct zbd_zone));
440 nrz = zbd_report_zones(td, f, 0, zones, ZBD_REPORT_MAX_ZONES);
443 log_info("fio: report zones (offset 0) failed for %s (%d).\n",
448 zone_size = zones[0].len;
449 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
451 if (td->o.zone_size == 0) {
452 td->o.zone_size = zone_size;
453 } else if (td->o.zone_size != zone_size) {
454 log_err("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n",
455 f->file_name, (unsigned long long) td->o.zone_size,
456 (unsigned long long) zone_size);
461 dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name,
462 nr_zones, (unsigned long long) zone_size / 1024);
464 zbd_info = scalloc(1, sizeof(*zbd_info) +
465 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
469 mutex_init_pshared(&zbd_info->mutex);
470 zbd_info->refcount = 1;
471 p = &zbd_info->zone_info[0];
472 for (offset = 0, j = 0; j < nr_zones;) {
474 for (i = 0; i < nrz; i++, j++, z++, p++) {
475 mutex_init_pshared_with_type(&p->mutex,
476 PTHREAD_MUTEX_RECURSIVE);
478 p->capacity = z->capacity;
480 case ZBD_ZONE_COND_NOT_WP:
481 case ZBD_ZONE_COND_FULL:
482 p->wp = p->start + p->capacity;
485 assert(z->start <= z->wp);
486 assert(z->wp <= z->start + zone_size);
492 if (j > 0 && p->start != p[-1].start + zone_size) {
493 log_info("%s: invalid zone data\n",
500 offset = z->start + z->len;
503 nrz = zbd_report_zones(td, f, offset,
504 zones, ZBD_REPORT_MAX_ZONES);
507 log_info("fio: report zones (offset %llu) failed for %s (%d).\n",
508 (unsigned long long)offset,
515 zbd_info->zone_info[nr_zones].start = offset;
517 f->zbd_info = zbd_info;
518 f->zbd_info->zone_size = zone_size;
519 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
520 ilog2(zone_size) : 0;
521 f->zbd_info->nr_zones = nr_zones;
532 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
534 * Returns 0 upon success and a negative error code upon failure.
536 static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
538 enum zbd_zoned_model zbd_model;
541 assert(td->o.zone_mode == ZONE_MODE_ZBD);
543 ret = zbd_get_zoned_model(td, f, &zbd_model);
551 case ZBD_HOST_MANAGED:
552 ret = parse_zone_info(td, f);
555 ret = init_zone_info(td, f);
558 td_verror(td, EINVAL, "Unsupported zoned model");
559 log_err("Unsupported zoned model\n");
564 f->zbd_info->model = zbd_model;
565 f->zbd_info->max_open_zones = td->o.max_open_zones;
570 void zbd_free_zone_info(struct fio_file *f)
576 pthread_mutex_lock(&f->zbd_info->mutex);
577 refcount = --f->zbd_info->refcount;
578 pthread_mutex_unlock(&f->zbd_info->mutex);
580 assert((int32_t)refcount >= 0);
587 * Initialize f->zbd_info.
589 * Returns 0 upon success and a negative error code upon failure.
591 * Note: this function can only work correctly if it is called before the first
594 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
596 struct thread_data *td2;
600 for_each_td(td2, i) {
601 for_each_file(td2, f2, j) {
602 if (td2 == td && f2 == file)
605 strcmp(f2->file_name, file->file_name) != 0)
607 file->zbd_info = f2->zbd_info;
608 file->zbd_info->refcount++;
613 ret = zbd_create_zone_info(td, file);
615 td_verror(td, -ret, "zbd_create_zone_info() failed");
619 int zbd_setup_files(struct thread_data *td)
624 for_each_file(td, f, i) {
625 if (zbd_init_zone_info(td, f))
629 if (!zbd_using_direct_io()) {
630 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
634 if (!zbd_verify_sizes())
637 if (!zbd_verify_bs())
640 for_each_file(td, f, i) {
641 struct zoned_block_device_info *zbd = f->zbd_info;
646 zbd->max_open_zones = zbd->max_open_zones ?: ZBD_MAX_OPEN_ZONES;
648 if (td->o.max_open_zones > 0 &&
649 zbd->max_open_zones != td->o.max_open_zones) {
650 log_err("Different 'max_open_zones' values\n");
653 if (zbd->max_open_zones > ZBD_MAX_OPEN_ZONES) {
654 log_err("'max_open_zones' value is limited by %u\n", ZBD_MAX_OPEN_ZONES);
663 * zbd_reset_range - reset zones for a range of sectors
664 * @td: FIO thread data.
665 * @f: Fio file for which to reset zones
666 * @sector: Starting sector in units of 512 bytes
667 * @nr_sectors: Number of sectors in units of 512 bytes
669 * Returns 0 upon success and a negative error code upon failure.
671 static int zbd_reset_range(struct thread_data *td, struct fio_file *f,
672 uint64_t offset, uint64_t length)
674 uint32_t zone_idx_b, zone_idx_e;
675 struct fio_zone_info *zb, *ze, *z;
678 assert(is_valid_offset(f, offset + length - 1));
680 switch (f->zbd_info->model) {
682 case ZBD_HOST_MANAGED:
683 ret = zbd_reset_wp(td, f, offset, length);
691 zone_idx_b = zbd_zone_idx(f, offset);
692 zb = &f->zbd_info->zone_info[zone_idx_b];
693 zone_idx_e = zbd_zone_idx(f, offset + length);
694 ze = &f->zbd_info->zone_info[zone_idx_e];
695 for (z = zb; z < ze; z++) {
696 pthread_mutex_lock(&z->mutex);
697 pthread_mutex_lock(&f->zbd_info->mutex);
698 f->zbd_info->sectors_with_data -= z->wp - z->start;
699 pthread_mutex_unlock(&f->zbd_info->mutex);
702 pthread_mutex_unlock(&z->mutex);
705 td->ts.nr_zone_resets += ze - zb;
710 static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info,
711 struct fio_zone_info *zone)
713 return zone - zbd_info->zone_info;
717 * zbd_reset_zone - reset the write pointer of a single zone
718 * @td: FIO thread data.
719 * @f: FIO file associated with the disk for which to reset a write pointer.
722 * Returns 0 upon success and a negative error code upon failure.
724 static int zbd_reset_zone(struct thread_data *td, struct fio_file *f,
725 struct fio_zone_info *z)
727 dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
728 zbd_zone_nr(f->zbd_info, z));
730 return zbd_reset_range(td, f, z->start, zbd_zone_end(z) - z->start);
733 /* The caller must hold f->zbd_info->mutex */
734 static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
735 unsigned int open_zone_idx)
739 assert(open_zone_idx < f->zbd_info->num_open_zones);
740 zone_idx = f->zbd_info->open_zones[open_zone_idx];
741 memmove(f->zbd_info->open_zones + open_zone_idx,
742 f->zbd_info->open_zones + open_zone_idx + 1,
743 (ZBD_MAX_OPEN_ZONES - (open_zone_idx + 1)) *
744 sizeof(f->zbd_info->open_zones[0]));
745 f->zbd_info->num_open_zones--;
746 td->num_open_zones--;
747 f->zbd_info->zone_info[zone_idx].open = 0;
751 * Reset a range of zones. Returns 0 upon success and 1 upon failure.
752 * @td: fio thread data.
753 * @f: fio file for which to reset zones
754 * @zb: first zone to reset.
755 * @ze: first zone not to reset.
756 * @all_zones: whether to reset all zones or only those zones for which the
757 * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE].
759 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
760 struct fio_zone_info *const zb,
761 struct fio_zone_info *const ze, bool all_zones)
763 struct fio_zone_info *z;
764 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
770 dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name,
771 zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze));
772 for (z = zb; z < ze; z++) {
773 uint32_t nz = z - f->zbd_info->zone_info;
775 if (!zbd_zone_swr(z))
781 pthread_mutex_lock(&f->zbd_info->mutex);
782 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
783 if (f->zbd_info->open_zones[i] == nz)
784 zbd_close_zone(td, f, i);
786 pthread_mutex_unlock(&f->zbd_info->mutex);
788 reset_wp = z->wp != z->start;
790 reset_wp = z->wp % min_bs != 0;
793 dprint(FD_ZBD, "%s: resetting zone %u\n",
795 zbd_zone_nr(f->zbd_info, z));
796 if (zbd_reset_zone(td, f, z) < 0)
799 pthread_mutex_unlock(&z->mutex);
806 * Reset zbd_info.write_cnt, the counter that counts down towards the next
809 static void _zbd_reset_write_cnt(const struct thread_data *td,
810 const struct fio_file *f)
812 assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
814 f->zbd_info->write_cnt = td->o.zrf.u.f ?
815 min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
818 static void zbd_reset_write_cnt(const struct thread_data *td,
819 const struct fio_file *f)
821 pthread_mutex_lock(&f->zbd_info->mutex);
822 _zbd_reset_write_cnt(td, f);
823 pthread_mutex_unlock(&f->zbd_info->mutex);
826 static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
827 const struct fio_file *f)
829 uint32_t write_cnt = 0;
831 pthread_mutex_lock(&f->zbd_info->mutex);
832 assert(f->zbd_info->write_cnt);
833 if (f->zbd_info->write_cnt)
834 write_cnt = --f->zbd_info->write_cnt;
836 _zbd_reset_write_cnt(td, f);
837 pthread_mutex_unlock(&f->zbd_info->mutex);
839 return write_cnt == 0;
847 /* Calculate the number of sectors with data (swd) and perform action 'a' */
848 static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a)
850 struct fio_zone_info *zb, *ze, *z;
853 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
854 ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset +
856 for (z = zb; z < ze; z++) {
857 pthread_mutex_lock(&z->mutex);
858 swd += z->wp - z->start;
860 pthread_mutex_lock(&f->zbd_info->mutex);
863 assert(f->zbd_info->sectors_with_data == swd);
866 f->zbd_info->sectors_with_data = swd;
869 pthread_mutex_unlock(&f->zbd_info->mutex);
870 for (z = zb; z < ze; z++)
871 pthread_mutex_unlock(&z->mutex);
877 * The swd check is useful for debugging but takes too much time to leave
878 * it enabled all the time. Hence it is disabled by default.
880 static const bool enable_check_swd = false;
882 /* Check whether the value of zbd_info.sectors_with_data is correct. */
883 static void zbd_check_swd(const struct fio_file *f)
885 if (!enable_check_swd)
888 zbd_process_swd(f, CHECK_SWD);
891 static void zbd_init_swd(struct fio_file *f)
895 if (!enable_check_swd)
898 swd = zbd_process_swd(f, SET_SWD);
899 dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name,
903 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
905 struct fio_zone_info *zb, *ze;
907 if (!f->zbd_info || !td_write(td))
910 zb = &f->zbd_info->zone_info[f->min_zone];
911 ze = &f->zbd_info->zone_info[f->max_zone];
914 * If data verification is enabled reset the affected zones before
915 * writing any data to avoid that a zone reset has to be issued while
916 * writing data, which causes data loss.
918 zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE &&
919 td->runstate != TD_VERIFYING);
920 zbd_reset_write_cnt(td, f);
923 /* The caller must hold f->zbd_info->mutex. */
924 static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
925 unsigned int zone_idx)
927 struct zoned_block_device_info *zbdi = f->zbd_info;
930 assert(td->o.job_max_open_zones == 0 || td->num_open_zones <= td->o.job_max_open_zones);
931 assert(td->o.job_max_open_zones <= zbdi->max_open_zones);
932 assert(zbdi->num_open_zones <= zbdi->max_open_zones);
934 for (i = 0; i < zbdi->num_open_zones; i++)
935 if (zbdi->open_zones[i] == zone_idx)
942 * Open a ZBD zone if it was not yet open. Returns true if either the zone was
943 * already open or if opening a new zone is allowed. Returns false if the zone
944 * was not yet open and opening a new zone would cause the zone limit to be
947 static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
950 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
951 const struct fio_file *f = io_u->file;
952 struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
955 if (z->cond == ZBD_ZONE_COND_OFFLINE)
959 * Skip full zones with data verification enabled because resetting a
960 * zone causes data loss and hence causes verification to fail.
962 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
965 pthread_mutex_lock(&f->zbd_info->mutex);
966 if (is_zone_open(td, f, zone_idx))
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;
1010 assert(is_valid_offset(f, io_u->offset));
1012 if (td->o.max_open_zones || td->o.job_max_open_zones) {
1014 * This statement accesses f->zbd_info->open_zones[] on purpose
1017 zone_idx = f->zbd_info->open_zones[pick_random_zone_idx(f, io_u)];
1019 zone_idx = zbd_zone_idx(f, io_u->offset);
1021 if (zone_idx < f->min_zone)
1022 zone_idx = f->min_zone;
1023 else if (zone_idx >= f->max_zone)
1024 zone_idx = f->max_zone - 1;
1025 dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
1026 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
1029 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
1030 * lock it can happen that the state of the zone with index zone_idx
1031 * has changed after 'z' has been assigned and before f->zbd_info->mutex
1032 * has been obtained. Hence the loop.
1037 z = &f->zbd_info->zone_info[zone_idx];
1039 zone_lock(td, f, z);
1040 pthread_mutex_lock(&f->zbd_info->mutex);
1041 if (td->o.max_open_zones == 0 && td->o.job_max_open_zones == 0)
1043 if (f->zbd_info->num_open_zones == 0) {
1044 pthread_mutex_unlock(&f->zbd_info->mutex);
1045 pthread_mutex_unlock(&z->mutex);
1046 dprint(FD_ZBD, "%s(%s): no zones are open\n",
1047 __func__, f->file_name);
1052 * List of opened zones is per-device, shared across all threads.
1053 * Start with quasi-random candidate zone.
1054 * Ignore zones which don't belong to thread's offset/size area.
1056 open_zone_idx = pick_random_zone_idx(f, io_u);
1057 assert(open_zone_idx < f->zbd_info->num_open_zones);
1058 tmp_idx = open_zone_idx;
1059 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1062 if (tmp_idx >= f->zbd_info->num_open_zones)
1064 tmpz = f->zbd_info->open_zones[tmp_idx];
1065 if (f->min_zone <= tmpz && tmpz < f->max_zone) {
1066 open_zone_idx = tmp_idx;
1067 goto found_candidate_zone;
1073 dprint(FD_ZBD, "%s(%s): no candidate zone\n",
1074 __func__, f->file_name);
1075 pthread_mutex_unlock(&f->zbd_info->mutex);
1076 pthread_mutex_unlock(&z->mutex);
1079 found_candidate_zone:
1080 new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
1081 if (new_zone_idx == zone_idx)
1083 zone_idx = new_zone_idx;
1084 pthread_mutex_unlock(&f->zbd_info->mutex);
1085 pthread_mutex_unlock(&z->mutex);
1088 /* Both z->mutex and f->zbd_info->mutex are held. */
1091 if (z->wp + min_bs <= zbd_zone_capacity_end(z)) {
1092 pthread_mutex_unlock(&f->zbd_info->mutex);
1095 dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name,
1097 if (td->o.max_open_zones || td->o.job_max_open_zones)
1098 zbd_close_zone(td, f, open_zone_idx);
1099 pthread_mutex_unlock(&f->zbd_info->mutex);
1101 /* Only z->mutex is held. */
1103 /* Zone 'z' is full, so try to open a new zone. */
1104 for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
1106 pthread_mutex_unlock(&z->mutex);
1108 if (!is_valid_offset(f, z->start)) {
1110 zone_idx = f->min_zone;
1111 z = &f->zbd_info->zone_info[zone_idx];
1113 assert(is_valid_offset(f, z->start));
1114 zone_lock(td, f, z);
1117 if (zbd_open_zone(td, io_u, zone_idx))
1121 /* Only z->mutex is held. */
1123 /* Check whether the write fits in any of the already opened zones. */
1124 pthread_mutex_lock(&f->zbd_info->mutex);
1125 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1126 zone_idx = f->zbd_info->open_zones[i];
1127 if (zone_idx < f->min_zone || zone_idx >= f->max_zone)
1129 pthread_mutex_unlock(&f->zbd_info->mutex);
1130 pthread_mutex_unlock(&z->mutex);
1132 z = &f->zbd_info->zone_info[zone_idx];
1134 zone_lock(td, f, z);
1135 if (z->wp + min_bs <= zbd_zone_capacity_end(z))
1137 pthread_mutex_lock(&f->zbd_info->mutex);
1139 pthread_mutex_unlock(&f->zbd_info->mutex);
1140 pthread_mutex_unlock(&z->mutex);
1141 dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
1146 dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
1148 io_u->offset = z->start;
1152 /* The caller must hold z->mutex. */
1153 static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
1155 struct fio_zone_info *z)
1157 const struct fio_file *f = io_u->file;
1158 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
1160 if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
1161 pthread_mutex_unlock(&z->mutex);
1162 z = zbd_convert_to_open_zone(td, io_u);
1166 if (z->verify_block * min_bs >= z->capacity)
1167 log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
1168 min_bs, (unsigned long long)z->capacity);
1169 io_u->offset = z->start + z->verify_block++ * min_bs;
1174 * Find another zone for which @io_u fits below the write pointer. Start
1175 * searching in zones @zb + 1 .. @zl and continue searching in zones
1178 * Either returns NULL or returns a zone pointer and holds the mutex for that
1181 static struct fio_zone_info *
1182 zbd_find_zone(struct thread_data *td, struct io_u *io_u,
1183 struct fio_zone_info *zb, struct fio_zone_info *zl)
1185 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1186 struct fio_file *f = io_u->file;
1187 struct fio_zone_info *z1, *z2;
1188 const struct fio_zone_info *const zf =
1189 &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
1192 * Skip to the next non-empty zone in case of sequential I/O and to
1193 * the nearest non-empty zone in case of random I/O.
1195 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1196 if (z1 < zl && z1->cond != ZBD_ZONE_COND_OFFLINE) {
1197 zone_lock(td, f, z1);
1198 if (z1->start + min_bs <= z1->wp)
1200 pthread_mutex_unlock(&z1->mutex);
1201 } else if (!td_random(td)) {
1204 if (td_random(td) && z2 >= zf &&
1205 z2->cond != ZBD_ZONE_COND_OFFLINE) {
1206 zone_lock(td, f, z2);
1207 if (z2->start + min_bs <= z2->wp)
1209 pthread_mutex_unlock(&z2->mutex);
1212 dprint(FD_ZBD, "%s: adjusting random read offset failed\n",
1218 * zbd_queue_io - update the write pointer of a sequential zone
1220 * @success: Whether or not the I/O unit has been queued successfully
1221 * @q: queueing status (busy, completed or queued).
1223 * For write and trim operations, update the write pointer of the I/O unit
1226 static void zbd_queue_io(struct io_u *io_u, int q, bool success)
1228 const struct fio_file *f = io_u->file;
1229 struct zoned_block_device_info *zbd_info = f->zbd_info;
1230 struct fio_zone_info *z;
1237 zone_idx = zbd_zone_idx(f, io_u->offset);
1238 assert(zone_idx < zbd_info->nr_zones);
1239 z = &zbd_info->zone_info[zone_idx];
1241 if (!zbd_zone_swr(z))
1248 "%s: queued I/O (%lld, %llu) for zone %u\n",
1249 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1251 switch (io_u->ddir) {
1253 zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
1254 zbd_zone_capacity_end(z));
1255 pthread_mutex_lock(&zbd_info->mutex);
1257 * z->wp > zone_end means that one or more I/O errors
1260 if (z->wp <= zone_end)
1261 zbd_info->sectors_with_data += zone_end - z->wp;
1262 pthread_mutex_unlock(&zbd_info->mutex);
1266 assert(z->wp == z->start);
1273 if (!success || q != FIO_Q_QUEUED) {
1274 /* BUSY or COMPLETED: unlock the zone */
1275 pthread_mutex_unlock(&z->mutex);
1276 io_u->zbd_put_io = NULL;
1281 * zbd_put_io - Unlock an I/O unit target zone lock
1284 static void zbd_put_io(const struct io_u *io_u)
1286 const struct fio_file *f = io_u->file;
1287 struct zoned_block_device_info *zbd_info = f->zbd_info;
1288 struct fio_zone_info *z;
1295 zone_idx = zbd_zone_idx(f, io_u->offset);
1296 assert(zone_idx < zbd_info->nr_zones);
1297 z = &zbd_info->zone_info[zone_idx];
1299 if (!zbd_zone_swr(z))
1303 "%s: terminate I/O (%lld, %llu) for zone %u\n",
1304 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1306 ret = pthread_mutex_unlock(&z->mutex);
1312 * Windows and MacOS do not define this.
1315 #define EREMOTEIO 121 /* POSIX value */
1318 bool zbd_unaligned_write(int error_code)
1320 switch (error_code) {
1329 * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives
1330 * @td: FIO thread data.
1331 * @io_u: FIO I/O unit.
1333 * For sequential workloads, change the file offset to skip zoneskip bytes when
1334 * no more IO can be performed in the current zone.
1335 * - For read workloads, zoneskip is applied when the io has reached the end of
1336 * the zone or the zone write position (when td->o.read_beyond_wp is false).
1337 * - For write workloads, zoneskip is applied when the zone is full.
1338 * This applies only to read and write operations.
1340 void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
1342 struct fio_file *f = io_u->file;
1343 enum fio_ddir ddir = io_u->ddir;
1344 struct fio_zone_info *z;
1347 assert(td->o.zone_mode == ZONE_MODE_ZBD);
1348 assert(td->o.zone_size);
1350 zone_idx = zbd_zone_idx(f, f->last_pos[ddir]);
1351 z = &f->zbd_info->zone_info[zone_idx];
1354 * When the zone capacity is smaller than the zone size and the I/O is
1355 * sequential write, skip to zone end if the latest position is at the
1356 * zone capacity limit.
1358 if (z->capacity < f->zbd_info->zone_size && !td_random(td) &&
1359 ddir == DDIR_WRITE &&
1360 f->last_pos[ddir] >= zbd_zone_capacity_end(z)) {
1362 "%s: Jump from zone capacity limit to zone end:"
1363 " (%lu -> %lu) for zone %u (%ld)\n",
1364 f->file_name, f->last_pos[ddir], zbd_zone_end(z),
1365 zbd_zone_nr(f->zbd_info, z), z->capacity);
1366 td->io_skip_bytes += zbd_zone_end(z) - f->last_pos[ddir];
1367 f->last_pos[ddir] = zbd_zone_end(z);
1371 * zone_skip is valid only for sequential workloads.
1373 if (td_random(td) || !td->o.zone_skip)
1377 * It is time to switch to a new zone if:
1378 * - zone_bytes == zone_size bytes have already been accessed
1379 * - The last position reached the end of the current zone.
1380 * - For reads with td->o.read_beyond_wp == false, the last position
1381 * reached the zone write pointer.
1383 if (td->zone_bytes >= td->o.zone_size ||
1384 f->last_pos[ddir] >= zbd_zone_end(z) ||
1385 (ddir == DDIR_READ &&
1386 (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) {
1391 f->file_offset += td->o.zone_size + td->o.zone_skip;
1394 * Wrap from the beginning, if we exceed the file size
1396 if (f->file_offset >= f->real_file_size)
1397 f->file_offset = get_start_offset(td, f);
1399 f->last_pos[ddir] = f->file_offset;
1400 td->io_skip_bytes += td->o.zone_skip;
1405 * zbd_adjust_ddir - Adjust an I/O direction for zonemode=zbd.
1407 * @td: FIO thread data.
1408 * @io_u: FIO I/O unit.
1409 * @ddir: I/O direction before adjustment.
1411 * Return adjusted I/O direction.
1413 enum fio_ddir zbd_adjust_ddir(struct thread_data *td, struct io_u *io_u,
1417 * In case read direction is chosen for the first random I/O, fio with
1418 * zonemode=zbd stops because no data can be read from zoned block
1419 * devices with all empty zones. Overwrite the first I/O direction as
1420 * write to make sure data to read exists.
1422 if (ddir != DDIR_READ || !td_rw(td))
1425 if (io_u->file->zbd_info->sectors_with_data ||
1426 td->o.read_beyond_wp)
1433 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1434 * @td: FIO thread data.
1435 * @io_u: FIO I/O unit.
1437 * Locking strategy: returns with z->mutex locked if and only if z refers
1438 * to a sequential zone and if io_u_accept is returned. z is the zone that
1439 * corresponds to io_u->offset at the end of this function.
1441 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1443 struct fio_file *f = io_u->file;
1444 uint32_t zone_idx_b;
1445 struct fio_zone_info *zb, *zl, *orig_zb;
1446 uint32_t orig_len = io_u->buflen;
1447 uint32_t min_bs = td->o.min_bs[io_u->ddir];
1455 assert(is_valid_offset(f, io_u->offset));
1456 assert(io_u->buflen);
1457 zone_idx_b = zbd_zone_idx(f, io_u->offset);
1458 zb = &f->zbd_info->zone_info[zone_idx_b];
1461 /* Accept the I/O offset for conventional zones. */
1462 if (!zbd_zone_swr(zb))
1466 * Accept the I/O offset for reads if reading beyond the write pointer
1469 if (zb->cond != ZBD_ZONE_COND_OFFLINE &&
1470 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1475 zone_lock(td, f, zb);
1477 switch (io_u->ddir) {
1479 if (td->runstate == TD_VERIFYING) {
1481 zb = zbd_replay_write_order(td, io_u, zb);
1485 * Check that there is enough written data in the zone to do an
1486 * I/O of at least min_bs B. If there isn't, find a new zone for
1489 range = zb->cond != ZBD_ZONE_COND_OFFLINE ?
1490 zb->wp - zb->start : 0;
1491 if (range < min_bs ||
1492 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1493 pthread_mutex_unlock(&zb->mutex);
1494 zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
1495 f->file_offset + f->io_size)];
1496 zb = zbd_find_zone(td, io_u, zb, zl);
1499 "%s: zbd_find_zone(%lld, %llu) failed\n",
1500 f->file_name, io_u->offset,
1505 * zbd_find_zone() returned a zone with a range of at
1508 range = zb->wp - zb->start;
1509 assert(range >= min_bs);
1512 io_u->offset = zb->start;
1515 * Make sure the I/O is within the zone valid data range while
1516 * maximizing the I/O size and preserving randomness.
1518 if (range <= io_u->buflen)
1519 io_u->offset = zb->start;
1520 else if (td_random(td))
1521 io_u->offset = zb->start +
1522 ((io_u->offset - orig_zb->start) %
1523 (range - io_u->buflen)) / min_bs * min_bs;
1525 * Make sure the I/O does not cross over the zone wp position.
1527 new_len = min((unsigned long long)io_u->buflen,
1528 (unsigned long long)(zb->wp - io_u->offset));
1529 new_len = new_len / min_bs * min_bs;
1530 if (new_len < io_u->buflen) {
1531 io_u->buflen = new_len;
1532 dprint(FD_IO, "Changed length from %u into %llu\n",
1533 orig_len, io_u->buflen);
1535 assert(zb->start <= io_u->offset);
1536 assert(io_u->offset + io_u->buflen <= zb->wp);
1539 if (io_u->buflen > f->zbd_info->zone_size)
1541 if (!zbd_open_zone(td, io_u, zone_idx_b)) {
1542 pthread_mutex_unlock(&zb->mutex);
1543 zb = zbd_convert_to_open_zone(td, io_u);
1546 zone_idx_b = zb - f->zbd_info->zone_info;
1548 /* Check whether the zone reset threshold has been exceeded */
1549 if (td->o.zrf.u.f) {
1550 if (f->zbd_info->sectors_with_data >=
1551 f->io_size * td->o.zrt.u.f &&
1552 zbd_dec_and_reset_write_cnt(td, f)) {
1556 /* Reset the zone pointer if necessary */
1557 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
1558 assert(td->o.verify == VERIFY_NONE);
1560 * Since previous write requests may have been submitted
1561 * asynchronously and since we will submit the zone
1562 * reset synchronously, wait until previously submitted
1563 * write requests have completed before issuing a
1568 if (zbd_reset_zone(td, f, zb) < 0)
1571 if (zb->capacity < min_bs) {
1572 log_err("zone capacity %llu smaller than minimum block size %d\n",
1573 (unsigned long long)zb->capacity,
1578 /* Make writes occur at the write pointer */
1579 assert(!zbd_zone_full(f, zb, min_bs));
1580 io_u->offset = zb->wp;
1581 if (!is_valid_offset(f, io_u->offset)) {
1582 dprint(FD_ZBD, "Dropped request with offset %llu\n",
1587 * Make sure that the buflen is a multiple of the minimal
1588 * block size. Give up if shrinking would make the request too
1591 new_len = min((unsigned long long)io_u->buflen,
1592 zbd_zone_capacity_end(zb) - io_u->offset);
1593 new_len = new_len / min_bs * min_bs;
1594 if (new_len == io_u->buflen)
1596 if (new_len >= min_bs) {
1597 io_u->buflen = new_len;
1598 dprint(FD_IO, "Changed length from %u into %llu\n",
1599 orig_len, io_u->buflen);
1602 log_err("Zone remainder %lld smaller than minimum block size %d\n",
1603 (zbd_zone_capacity_end(zb) - io_u->offset),
1610 case DDIR_SYNC_FILE_RANGE:
1621 assert(zb->cond != ZBD_ZONE_COND_OFFLINE);
1622 assert(!io_u->zbd_queue_io);
1623 assert(!io_u->zbd_put_io);
1624 io_u->zbd_queue_io = zbd_queue_io;
1625 io_u->zbd_put_io = zbd_put_io;
1630 pthread_mutex_unlock(&zb->mutex);
1634 /* Return a string with ZBD statistics */
1635 char *zbd_write_status(const struct thread_stat *ts)
1639 if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0)