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 new_offset = (z+1)->start;
267 if (new_offset >= f->file_offset + f->io_size) {
268 log_info("%s: io_size must be at least one zone\n",
272 log_info("%s: rounded up offset from %llu to %llu\n",
273 f->file_name, (unsigned long long) f->file_offset,
274 (unsigned long long) new_offset);
275 f->io_size -= (new_offset - f->file_offset);
276 f->file_offset = new_offset;
278 zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size);
279 z = &f->zbd_info->zone_info[zone_idx];
281 if (f->file_offset + f->io_size != new_end) {
282 if (new_end <= f->file_offset) {
283 log_info("%s: io_size must be at least one zone\n",
287 log_info("%s: rounded down io_size from %llu to %llu\n",
288 f->file_name, (unsigned long long) f->io_size,
289 (unsigned long long) new_end - f->file_offset);
290 f->io_size = new_end - f->file_offset;
298 static bool zbd_verify_bs(void)
300 struct thread_data *td;
306 for_each_file(td, f, j) {
309 zone_size = f->zbd_info->zone_size;
310 for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
311 if (td->o.verify != VERIFY_NONE &&
312 zone_size % td->o.bs[k] != 0) {
313 log_info("%s: block size %llu is not a divisor of the zone size %d\n",
314 f->file_name, td->o.bs[k],
324 static int ilog2(uint64_t i)
336 * Initialize f->zbd_info for devices that are not zoned block devices. This
337 * allows to execute a ZBD workload against a non-ZBD device.
339 static int init_zone_info(struct thread_data *td, struct fio_file *f)
342 struct fio_zone_info *p;
343 uint64_t zone_size = td->o.zone_size;
344 struct zoned_block_device_info *zbd_info = NULL;
345 pthread_mutexattr_t attr;
348 if (zone_size == 0) {
349 log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n",
354 if (zone_size < 512) {
355 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
360 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
361 zbd_info = scalloc(1, sizeof(*zbd_info) +
362 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
366 pthread_mutexattr_init(&attr);
367 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
368 pthread_mutexattr_setpshared(&attr, true);
369 pthread_mutex_init(&zbd_info->mutex, &attr);
370 zbd_info->refcount = 1;
371 p = &zbd_info->zone_info[0];
372 for (i = 0; i < nr_zones; i++, p++) {
373 pthread_mutex_init(&p->mutex, &attr);
374 p->start = i * zone_size;
375 p->wp = p->start + zone_size;
376 p->type = ZBD_ZONE_TYPE_SWR;
377 p->cond = ZBD_ZONE_COND_EMPTY;
380 p->start = nr_zones * zone_size;
382 f->zbd_info = zbd_info;
383 f->zbd_info->zone_size = zone_size;
384 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
385 ilog2(zone_size) : 0;
386 f->zbd_info->nr_zones = nr_zones;
387 pthread_mutexattr_destroy(&attr);
392 * Maximum number of zones to report in one operation.
394 #define ZBD_REPORT_MAX_ZONES 8192U
397 * Parse the device zone report and store it in f->zbd_info. Must be called
398 * only for devices that are zoned, namely those with a model != ZBD_NONE.
400 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
403 struct zbd_zone *zones, *z;
404 struct fio_zone_info *p;
405 uint64_t zone_size, offset;
406 struct zoned_block_device_info *zbd_info = NULL;
407 pthread_mutexattr_t attr;
410 pthread_mutexattr_init(&attr);
411 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
412 pthread_mutexattr_setpshared(&attr, true);
414 zones = calloc(ZBD_REPORT_MAX_ZONES, sizeof(struct zbd_zone));
418 nrz = zbd_report_zones(td, f, 0, zones, ZBD_REPORT_MAX_ZONES);
421 log_info("fio: report zones (offset 0) failed for %s (%d).\n",
426 zone_size = zones[0].len;
427 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
429 if (td->o.zone_size == 0) {
430 td->o.zone_size = zone_size;
431 } else if (td->o.zone_size != zone_size) {
432 log_err("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n",
433 f->file_name, (unsigned long long) td->o.zone_size,
434 (unsigned long long) zone_size);
439 dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name,
440 nr_zones, (unsigned long long) zone_size / 1024);
442 zbd_info = scalloc(1, sizeof(*zbd_info) +
443 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
447 pthread_mutex_init(&zbd_info->mutex, &attr);
448 zbd_info->refcount = 1;
449 p = &zbd_info->zone_info[0];
450 for (offset = 0, j = 0; j < nr_zones;) {
452 for (i = 0; i < nrz; i++, j++, z++, p++) {
453 pthread_mutex_init(&p->mutex, &attr);
456 case ZBD_ZONE_COND_NOT_WP:
457 case ZBD_ZONE_COND_FULL:
458 p->wp = p->start + zone_size;
461 assert(z->start <= z->wp);
462 assert(z->wp <= z->start + zone_size);
468 if (j > 0 && p->start != p[-1].start + zone_size) {
469 log_info("%s: invalid zone data\n",
476 offset = z->start + z->len;
479 nrz = zbd_report_zones(td, f, offset,
480 zones, ZBD_REPORT_MAX_ZONES);
483 log_info("fio: report zones (offset %llu) failed for %s (%d).\n",
484 (unsigned long long)offset,
491 zbd_info->zone_info[nr_zones].start = offset;
493 f->zbd_info = zbd_info;
494 f->zbd_info->zone_size = zone_size;
495 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
496 ilog2(zone_size) : 0;
497 f->zbd_info->nr_zones = nr_zones;
504 pthread_mutexattr_destroy(&attr);
509 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
511 * Returns 0 upon success and a negative error code upon failure.
513 static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
515 enum zbd_zoned_model zbd_model;
518 assert(td->o.zone_mode == ZONE_MODE_ZBD);
520 ret = zbd_get_zoned_model(td, f, &zbd_model);
528 case ZBD_HOST_MANAGED:
529 ret = parse_zone_info(td, f);
532 ret = init_zone_info(td, f);
535 td_verror(td, EINVAL, "Unsupported zoned model");
536 log_err("Unsupported zoned model\n");
541 f->zbd_info->model = zbd_model;
545 void zbd_free_zone_info(struct fio_file *f)
552 pthread_mutex_lock(&f->zbd_info->mutex);
553 refcount = --f->zbd_info->refcount;
554 pthread_mutex_unlock(&f->zbd_info->mutex);
556 assert((int32_t)refcount >= 0);
563 * Initialize f->zbd_info.
565 * Returns 0 upon success and a negative error code upon failure.
567 * Note: this function can only work correctly if it is called before the first
570 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
572 struct thread_data *td2;
576 for_each_td(td2, i) {
577 for_each_file(td2, f2, j) {
578 if (td2 == td && f2 == file)
581 strcmp(f2->file_name, file->file_name) != 0)
583 file->zbd_info = f2->zbd_info;
584 file->zbd_info->refcount++;
589 ret = zbd_create_zone_info(td, file);
591 td_verror(td, -ret, "zbd_create_zone_info() failed");
595 int zbd_init(struct thread_data *td)
600 for_each_file(td, f, i) {
601 if (zbd_init_zone_info(td, f))
605 if (!zbd_using_direct_io()) {
606 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
610 if (!zbd_verify_sizes())
613 if (!zbd_verify_bs())
620 * zbd_reset_range - reset zones for a range of sectors
621 * @td: FIO thread data.
622 * @f: Fio file for which to reset zones
623 * @sector: Starting sector in units of 512 bytes
624 * @nr_sectors: Number of sectors in units of 512 bytes
626 * Returns 0 upon success and a negative error code upon failure.
628 static int zbd_reset_range(struct thread_data *td, struct fio_file *f,
629 uint64_t offset, uint64_t length)
631 uint32_t zone_idx_b, zone_idx_e;
632 struct fio_zone_info *zb, *ze, *z;
635 assert(is_valid_offset(f, offset + length - 1));
637 switch (f->zbd_info->model) {
639 case ZBD_HOST_MANAGED:
640 ret = zbd_reset_wp(td, f, offset, length);
648 zone_idx_b = zbd_zone_idx(f, offset);
649 zb = &f->zbd_info->zone_info[zone_idx_b];
650 zone_idx_e = zbd_zone_idx(f, offset + length);
651 ze = &f->zbd_info->zone_info[zone_idx_e];
652 for (z = zb; z < ze; z++) {
653 pthread_mutex_lock(&z->mutex);
654 pthread_mutex_lock(&f->zbd_info->mutex);
655 f->zbd_info->sectors_with_data -= z->wp - z->start;
656 pthread_mutex_unlock(&f->zbd_info->mutex);
659 pthread_mutex_unlock(&z->mutex);
662 td->ts.nr_zone_resets += ze - zb;
667 static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info,
668 struct fio_zone_info *zone)
670 return zone - zbd_info->zone_info;
674 * zbd_reset_zone - reset the write pointer of a single zone
675 * @td: FIO thread data.
676 * @f: FIO file associated with the disk for which to reset a write pointer.
679 * Returns 0 upon success and a negative error code upon failure.
681 static int zbd_reset_zone(struct thread_data *td, struct fio_file *f,
682 struct fio_zone_info *z)
684 dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
685 zbd_zone_nr(f->zbd_info, z));
687 return zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
690 /* The caller must hold f->zbd_info->mutex */
691 static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
692 unsigned int open_zone_idx)
696 assert(open_zone_idx < f->zbd_info->num_open_zones);
697 zone_idx = f->zbd_info->open_zones[open_zone_idx];
698 memmove(f->zbd_info->open_zones + open_zone_idx,
699 f->zbd_info->open_zones + open_zone_idx + 1,
700 (ZBD_MAX_OPEN_ZONES - (open_zone_idx + 1)) *
701 sizeof(f->zbd_info->open_zones[0]));
702 f->zbd_info->num_open_zones--;
703 f->zbd_info->zone_info[zone_idx].open = 0;
707 * Reset a range of zones. Returns 0 upon success and 1 upon failure.
708 * @td: fio thread data.
709 * @f: fio file for which to reset zones
710 * @zb: first zone to reset.
711 * @ze: first zone not to reset.
712 * @all_zones: whether to reset all zones or only those zones for which the
713 * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE].
715 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
716 struct fio_zone_info *const zb,
717 struct fio_zone_info *const ze, bool all_zones)
719 struct fio_zone_info *z;
720 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
726 dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name,
727 zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze));
728 for (z = zb; z < ze; z++) {
729 uint32_t nz = z - f->zbd_info->zone_info;
731 if (!zbd_zone_swr(z))
737 pthread_mutex_lock(&f->zbd_info->mutex);
738 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
739 if (f->zbd_info->open_zones[i] == nz)
740 zbd_close_zone(td, f, i);
742 pthread_mutex_unlock(&f->zbd_info->mutex);
744 reset_wp = z->wp != z->start;
746 reset_wp = (td->o.td_ddir & TD_DDIR_WRITE) &&
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;
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->o.td_ddir & TD_DDIR_WRITE) &&
873 td->runstate != TD_VERIFYING);
874 zbd_reset_write_cnt(td, f);
877 /* The caller must hold f->zbd_info->mutex. */
878 static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
879 unsigned int zone_idx)
881 struct zoned_block_device_info *zbdi = f->zbd_info;
884 assert(td->o.max_open_zones <= ARRAY_SIZE(zbdi->open_zones));
885 assert(zbdi->num_open_zones <= td->o.max_open_zones);
887 for (i = 0; i < zbdi->num_open_zones; i++)
888 if (zbdi->open_zones[i] == zone_idx)
895 * Open a ZBD zone if it was not yet open. Returns true if either the zone was
896 * already open or if opening a new zone is allowed. Returns false if the zone
897 * was not yet open and opening a new zone would cause the zone limit to be
900 static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
903 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
904 const struct fio_file *f = io_u->file;
905 struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
908 if (z->cond == ZBD_ZONE_COND_OFFLINE)
912 * Skip full zones with data verification enabled because resetting a
913 * zone causes data loss and hence causes verification to fail.
915 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
918 /* Zero means no limit */
919 if (!td->o.max_open_zones)
922 pthread_mutex_lock(&f->zbd_info->mutex);
923 if (is_zone_open(td, f, zone_idx))
926 if (f->zbd_info->num_open_zones >= td->o.max_open_zones)
928 dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx);
929 f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx;
934 pthread_mutex_unlock(&f->zbd_info->mutex);
938 /* Anything goes as long as it is not a constant. */
939 static uint32_t pick_random_zone_idx(const struct fio_file *f,
940 const struct io_u *io_u)
942 return io_u->offset * f->zbd_info->num_open_zones / f->real_file_size;
946 * Modify the offset of an I/O unit that does not refer to an open zone such
947 * that it refers to an open zone. Close an open zone and open a new zone if
948 * necessary. This algorithm can only work correctly if all write pointers are
949 * a multiple of the fio block size. The caller must neither hold z->mutex
950 * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
952 static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
955 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
956 const struct fio_file *f = io_u->file;
957 struct fio_zone_info *z;
958 unsigned int open_zone_idx = -1;
959 uint32_t zone_idx, new_zone_idx;
962 assert(is_valid_offset(f, io_u->offset));
964 if (td->o.max_open_zones) {
966 * This statement accesses f->zbd_info->open_zones[] on purpose
969 zone_idx = f->zbd_info->open_zones[pick_random_zone_idx(f, io_u)];
971 zone_idx = zbd_zone_idx(f, io_u->offset);
973 dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
974 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
977 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
978 * lock it can happen that the state of the zone with index zone_idx
979 * has changed after 'z' has been assigned and before f->zbd_info->mutex
980 * has been obtained. Hence the loop.
985 z = &f->zbd_info->zone_info[zone_idx];
988 pthread_mutex_lock(&f->zbd_info->mutex);
989 if (td->o.max_open_zones == 0)
991 if (f->zbd_info->num_open_zones == 0) {
992 pthread_mutex_unlock(&f->zbd_info->mutex);
993 pthread_mutex_unlock(&z->mutex);
994 dprint(FD_ZBD, "%s(%s): no zones are open\n",
995 __func__, f->file_name);
1000 * List of opened zones is per-device, shared across all threads.
1001 * Start with quasi-random candidate zone.
1002 * Ignore zones which don't belong to thread's offset/size area.
1004 open_zone_idx = pick_random_zone_idx(f, io_u);
1005 assert(open_zone_idx < f->zbd_info->num_open_zones);
1006 tmp_idx = open_zone_idx;
1007 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1010 if (tmp_idx >= f->zbd_info->num_open_zones)
1012 tmpz = f->zbd_info->open_zones[tmp_idx];
1014 if (is_valid_offset(f, f->zbd_info->zone_info[tmpz].start)) {
1015 open_zone_idx = tmp_idx;
1016 goto found_candidate_zone;
1022 dprint(FD_ZBD, "%s(%s): no candidate zone\n",
1023 __func__, f->file_name);
1024 pthread_mutex_unlock(&f->zbd_info->mutex);
1025 pthread_mutex_unlock(&z->mutex);
1028 found_candidate_zone:
1029 new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
1030 if (new_zone_idx == zone_idx)
1032 zone_idx = new_zone_idx;
1033 pthread_mutex_unlock(&f->zbd_info->mutex);
1034 pthread_mutex_unlock(&z->mutex);
1037 /* Both z->mutex and f->zbd_info->mutex are held. */
1040 if (z->wp + min_bs <= (z+1)->start) {
1041 pthread_mutex_unlock(&f->zbd_info->mutex);
1044 dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name,
1046 if (td->o.max_open_zones)
1047 zbd_close_zone(td, f, open_zone_idx);
1048 pthread_mutex_unlock(&f->zbd_info->mutex);
1050 /* Only z->mutex is held. */
1052 /* Zone 'z' is full, so try to open a new zone. */
1053 for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
1055 pthread_mutex_unlock(&z->mutex);
1057 if (!is_valid_offset(f, z->start)) {
1059 zone_idx = zbd_zone_idx(f, f->file_offset);
1060 z = &f->zbd_info->zone_info[zone_idx];
1062 assert(is_valid_offset(f, z->start));
1066 if (zbd_open_zone(td, io_u, zone_idx))
1070 /* Only z->mutex is held. */
1072 /* Check whether the write fits in any of the already opened zones. */
1073 pthread_mutex_lock(&f->zbd_info->mutex);
1074 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
1075 zone_idx = f->zbd_info->open_zones[i];
1076 pthread_mutex_unlock(&f->zbd_info->mutex);
1077 pthread_mutex_unlock(&z->mutex);
1079 z = &f->zbd_info->zone_info[zone_idx];
1082 if (z->wp + min_bs <= (z+1)->start)
1084 pthread_mutex_lock(&f->zbd_info->mutex);
1086 pthread_mutex_unlock(&f->zbd_info->mutex);
1087 pthread_mutex_unlock(&z->mutex);
1088 dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
1093 dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
1095 io_u->offset = z->start;
1099 /* The caller must hold z->mutex. */
1100 static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
1102 struct fio_zone_info *z)
1104 const struct fio_file *f = io_u->file;
1105 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
1107 if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
1108 pthread_mutex_unlock(&z->mutex);
1109 z = zbd_convert_to_open_zone(td, io_u);
1113 if (z->verify_block * min_bs >= f->zbd_info->zone_size)
1114 log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
1115 min_bs, (unsigned long long) f->zbd_info->zone_size);
1116 io_u->offset = z->start + z->verify_block++ * min_bs;
1121 * Find another zone for which @io_u fits below the write pointer. Start
1122 * searching in zones @zb + 1 .. @zl and continue searching in zones
1125 * Either returns NULL or returns a zone pointer and holds the mutex for that
1128 static struct fio_zone_info *
1129 zbd_find_zone(struct thread_data *td, struct io_u *io_u,
1130 struct fio_zone_info *zb, struct fio_zone_info *zl)
1132 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1133 const struct fio_file *f = io_u->file;
1134 struct fio_zone_info *z1, *z2;
1135 const struct fio_zone_info *const zf =
1136 &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
1139 * Skip to the next non-empty zone in case of sequential I/O and to
1140 * the nearest non-empty zone in case of random I/O.
1142 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1143 if (z1 < zl && z1->cond != ZBD_ZONE_COND_OFFLINE) {
1144 pthread_mutex_lock(&z1->mutex);
1145 if (z1->start + min_bs <= z1->wp)
1147 pthread_mutex_unlock(&z1->mutex);
1148 } else if (!td_random(td)) {
1151 if (td_random(td) && z2 >= zf &&
1152 z2->cond != ZBD_ZONE_COND_OFFLINE) {
1153 pthread_mutex_lock(&z2->mutex);
1154 if (z2->start + min_bs <= z2->wp)
1156 pthread_mutex_unlock(&z2->mutex);
1159 dprint(FD_ZBD, "%s: adjusting random read offset failed\n",
1165 * zbd_queue_io - update the write pointer of a sequential zone
1167 * @success: Whether or not the I/O unit has been queued successfully
1168 * @q: queueing status (busy, completed or queued).
1170 * For write and trim operations, update the write pointer of the I/O unit
1173 static void zbd_queue_io(struct io_u *io_u, int q, bool success)
1175 const struct fio_file *f = io_u->file;
1176 struct zoned_block_device_info *zbd_info = f->zbd_info;
1177 struct fio_zone_info *z;
1184 zone_idx = zbd_zone_idx(f, io_u->offset);
1185 assert(zone_idx < zbd_info->nr_zones);
1186 z = &zbd_info->zone_info[zone_idx];
1188 if (!zbd_zone_swr(z))
1195 "%s: queued I/O (%lld, %llu) for zone %u\n",
1196 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1198 switch (io_u->ddir) {
1200 zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
1202 pthread_mutex_lock(&zbd_info->mutex);
1204 * z->wp > zone_end means that one or more I/O errors
1207 if (z->wp <= zone_end)
1208 zbd_info->sectors_with_data += zone_end - z->wp;
1209 pthread_mutex_unlock(&zbd_info->mutex);
1213 assert(z->wp == z->start);
1220 if (!success || q != FIO_Q_QUEUED) {
1221 /* BUSY or COMPLETED: unlock the zone */
1222 pthread_mutex_unlock(&z->mutex);
1223 io_u->zbd_put_io = NULL;
1228 * zbd_put_io - Unlock an I/O unit target zone lock
1231 static void zbd_put_io(const struct io_u *io_u)
1233 const struct fio_file *f = io_u->file;
1234 struct zoned_block_device_info *zbd_info = f->zbd_info;
1235 struct fio_zone_info *z;
1242 zone_idx = zbd_zone_idx(f, io_u->offset);
1243 assert(zone_idx < zbd_info->nr_zones);
1244 z = &zbd_info->zone_info[zone_idx];
1246 if (!zbd_zone_swr(z))
1250 "%s: terminate I/O (%lld, %llu) for zone %u\n",
1251 f->file_name, io_u->offset, io_u->buflen, zone_idx);
1253 ret = pthread_mutex_unlock(&z->mutex);
1259 * Windows and MacOS do not define this.
1262 #define EREMOTEIO 121 /* POSIX value */
1265 bool zbd_unaligned_write(int error_code)
1267 switch (error_code) {
1276 * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives
1277 * @td: FIO thread data.
1278 * @io_u: FIO I/O unit.
1280 * For sequential workloads, change the file offset to skip zoneskip bytes when
1281 * no more IO can be performed in the current zone.
1282 * - For read workloads, zoneskip is applied when the io has reached the end of
1283 * the zone or the zone write position (when td->o.read_beyond_wp is false).
1284 * - For write workloads, zoneskip is applied when the zone is full.
1285 * This applies only to read and write operations.
1287 void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
1289 struct fio_file *f = io_u->file;
1290 enum fio_ddir ddir = io_u->ddir;
1291 struct fio_zone_info *z;
1294 assert(td->o.zone_mode == ZONE_MODE_ZBD);
1295 assert(td->o.zone_size);
1298 * zone_skip is valid only for sequential workloads.
1300 if (td_random(td) || !td->o.zone_skip)
1304 * It is time to switch to a new zone if:
1305 * - zone_bytes == zone_size bytes have already been accessed
1306 * - The last position reached the end of the current zone.
1307 * - For reads with td->o.read_beyond_wp == false, the last position
1308 * reached the zone write pointer.
1310 zone_idx = zbd_zone_idx(f, f->last_pos[ddir]);
1311 z = &f->zbd_info->zone_info[zone_idx];
1313 if (td->zone_bytes >= td->o.zone_size ||
1314 f->last_pos[ddir] >= (z+1)->start ||
1315 (ddir == DDIR_READ &&
1316 (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) {
1321 f->file_offset += td->o.zone_size + td->o.zone_skip;
1324 * Wrap from the beginning, if we exceed the file size
1326 if (f->file_offset >= f->real_file_size)
1327 f->file_offset = get_start_offset(td, f);
1329 f->last_pos[ddir] = f->file_offset;
1330 td->io_skip_bytes += td->o.zone_skip;
1335 * zbd_adjust_ddir - Adjust an I/O direction for zonedmode=zbd.
1337 * @td: FIO thread data.
1338 * @io_u: FIO I/O unit.
1339 * @ddir: I/O direction before adjustment.
1341 * Return adjusted I/O direction.
1343 enum fio_ddir zbd_adjust_ddir(struct thread_data *td, struct io_u *io_u,
1347 * In case read direction is chosen for the first random I/O, fio with
1348 * zonemode=zbd stops because no data can be read from zoned block
1349 * devices with all empty zones. Overwrite the first I/O direction as
1350 * write to make sure data to read exists.
1352 if (td->o.zone_mode != ZONE_MODE_ZBD ||
1353 ddir != DDIR_READ ||
1357 if (io_u->file->zbd_info->sectors_with_data ||
1358 td->o.read_beyond_wp)
1365 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1366 * @td: FIO thread data.
1367 * @io_u: FIO I/O unit.
1369 * Locking strategy: returns with z->mutex locked if and only if z refers
1370 * to a sequential zone and if io_u_accept is returned. z is the zone that
1371 * corresponds to io_u->offset at the end of this function.
1373 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1375 struct fio_file *f = io_u->file;
1376 uint32_t zone_idx_b;
1377 struct fio_zone_info *zb, *zl, *orig_zb;
1378 uint32_t orig_len = io_u->buflen;
1379 uint32_t min_bs = td->o.min_bs[io_u->ddir];
1387 assert(is_valid_offset(f, io_u->offset));
1388 assert(io_u->buflen);
1389 zone_idx_b = zbd_zone_idx(f, io_u->offset);
1390 zb = &f->zbd_info->zone_info[zone_idx_b];
1393 /* Accept the I/O offset for conventional zones. */
1394 if (!zbd_zone_swr(zb))
1398 * Accept the I/O offset for reads if reading beyond the write pointer
1401 if (zb->cond != ZBD_ZONE_COND_OFFLINE &&
1402 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1409 switch (io_u->ddir) {
1411 if (td->runstate == TD_VERIFYING) {
1412 zb = zbd_replay_write_order(td, io_u, zb);
1416 * Check that there is enough written data in the zone to do an
1417 * I/O of at least min_bs B. If there isn't, find a new zone for
1420 range = zb->cond != ZBD_ZONE_COND_OFFLINE ?
1421 zb->wp - zb->start : 0;
1422 if (range < min_bs ||
1423 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1424 pthread_mutex_unlock(&zb->mutex);
1425 zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
1426 f->file_offset + f->io_size)];
1427 zb = zbd_find_zone(td, io_u, zb, zl);
1430 "%s: zbd_find_zone(%lld, %llu) failed\n",
1431 f->file_name, io_u->offset,
1436 * zbd_find_zone() returned a zone with a range of at
1439 range = zb->wp - zb->start;
1440 assert(range >= min_bs);
1443 io_u->offset = zb->start;
1446 * Make sure the I/O is within the zone valid data range while
1447 * maximizing the I/O size and preserving randomness.
1449 if (range <= io_u->buflen)
1450 io_u->offset = zb->start;
1451 else if (td_random(td))
1452 io_u->offset = zb->start +
1453 ((io_u->offset - orig_zb->start) %
1454 (range - io_u->buflen)) / min_bs * min_bs;
1456 * Make sure the I/O does not cross over the zone wp position.
1458 new_len = min((unsigned long long)io_u->buflen,
1459 (unsigned long long)(zb->wp - io_u->offset));
1460 new_len = new_len / min_bs * min_bs;
1461 if (new_len < io_u->buflen) {
1462 io_u->buflen = new_len;
1463 dprint(FD_IO, "Changed length from %u into %llu\n",
1464 orig_len, io_u->buflen);
1466 assert(zb->start <= io_u->offset);
1467 assert(io_u->offset + io_u->buflen <= zb->wp);
1470 if (io_u->buflen > f->zbd_info->zone_size)
1472 if (!zbd_open_zone(td, io_u, zone_idx_b)) {
1473 pthread_mutex_unlock(&zb->mutex);
1474 zb = zbd_convert_to_open_zone(td, io_u);
1477 zone_idx_b = zb - f->zbd_info->zone_info;
1479 /* Check whether the zone reset threshold has been exceeded */
1480 if (td->o.zrf.u.f) {
1481 if (f->zbd_info->sectors_with_data >=
1482 f->io_size * td->o.zrt.u.f &&
1483 zbd_dec_and_reset_write_cnt(td, f)) {
1487 /* Reset the zone pointer if necessary */
1488 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
1489 assert(td->o.verify == VERIFY_NONE);
1491 * Since previous write requests may have been submitted
1492 * asynchronously and since we will submit the zone
1493 * reset synchronously, wait until previously submitted
1494 * write requests have completed before issuing a
1499 if (zbd_reset_zone(td, f, zb) < 0)
1502 /* Make writes occur at the write pointer */
1503 assert(!zbd_zone_full(f, zb, min_bs));
1504 io_u->offset = zb->wp;
1505 if (!is_valid_offset(f, io_u->offset)) {
1506 dprint(FD_ZBD, "Dropped request with offset %llu\n",
1511 * Make sure that the buflen is a multiple of the minimal
1512 * block size. Give up if shrinking would make the request too
1515 new_len = min((unsigned long long)io_u->buflen,
1516 (zb + 1)->start - io_u->offset);
1517 new_len = new_len / min_bs * min_bs;
1518 if (new_len == io_u->buflen)
1520 if (new_len >= min_bs) {
1521 io_u->buflen = new_len;
1522 dprint(FD_IO, "Changed length from %u into %llu\n",
1523 orig_len, io_u->buflen);
1526 log_err("Zone remainder %lld smaller than minimum block size %d\n",
1527 ((zb + 1)->start - io_u->offset),
1534 case DDIR_SYNC_FILE_RANGE:
1545 assert(zb->cond != ZBD_ZONE_COND_OFFLINE);
1546 assert(!io_u->zbd_queue_io);
1547 assert(!io_u->zbd_put_io);
1548 io_u->zbd_queue_io = zbd_queue_io;
1549 io_u->zbd_put_io = zbd_put_io;
1554 pthread_mutex_unlock(&zb->mutex);
1558 /* Return a string with ZBD statistics */
1559 char *zbd_write_status(const struct thread_stat *ts)
1563 if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0)