2 * Copyright (C) 2018 Western Digital Corporation or its affiliates.
4 * This file is released under the GPL.
14 #include "compiler/compiler.h"
20 #include "oslib/asprintf.h"
26 static bool is_valid_offset(const struct fio_file *f, uint64_t offset)
28 return (uint64_t)(offset - f->file_offset) < f->io_size;
31 static inline unsigned int zbd_zone_idx(const struct fio_file *f,
32 struct fio_zone_info *zone)
34 return zone - f->zbd_info->zone_info;
38 * zbd_offset_to_zone_idx - convert an offset into a zone number
40 * @offset: offset in bytes. If this offset is in the first zone_size bytes
41 * past the disk size then the index of the sentinel is returned.
43 static unsigned int zbd_offset_to_zone_idx(const struct fio_file *f,
48 if (f->zbd_info->zone_size_log2 > 0)
49 zone_idx = offset >> f->zbd_info->zone_size_log2;
51 zone_idx = offset / f->zbd_info->zone_size;
53 return min(zone_idx, f->zbd_info->nr_zones);
57 * zbd_zone_end - Return zone end location
58 * @z: zone info pointer.
60 static inline uint64_t zbd_zone_end(const struct fio_zone_info *z)
66 * zbd_zone_capacity_end - Return zone capacity limit end location
67 * @z: zone info pointer.
69 static inline uint64_t zbd_zone_capacity_end(const struct fio_zone_info *z)
71 return z->start + z->capacity;
75 * zbd_zone_remainder - Return the number of bytes that are still available for
76 * writing before the zone gets full
77 * @z: zone info pointer.
79 static inline uint64_t zbd_zone_remainder(struct fio_zone_info *z)
81 if (z->wp >= zbd_zone_capacity_end(z))
84 return zbd_zone_capacity_end(z) - z->wp;
88 * zbd_zone_full - verify whether a minimum number of bytes remain in a zone
90 * @z: zone info pointer.
91 * @required: minimum number of bytes that must remain in a zone.
93 * The caller must hold z->mutex.
95 static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
98 assert((required & 511) == 0);
100 return z->has_wp && required > zbd_zone_remainder(z);
103 static void zone_lock(struct thread_data *td, const struct fio_file *f,
104 struct fio_zone_info *z)
107 unsigned int const nz = zbd_zone_idx(f, z);
108 /* A thread should never lock zones outside its working area. */
109 assert(f->min_zone <= nz && nz < f->max_zone);
114 * Lock the io_u target zone. The zone will be unlocked if io_u offset
115 * is changed or when io_u completes and zbd_put_io() executed.
116 * To avoid multiple jobs doing asynchronous I/Os from deadlocking each
117 * other waiting for zone locks when building an io_u batch, first
118 * only trylock the zone. If the zone is already locked by another job,
119 * process the currently queued I/Os so that I/O progress is made and
122 if (pthread_mutex_trylock(&z->mutex) != 0) {
123 if (!td_ioengine_flagged(td, FIO_SYNCIO))
125 pthread_mutex_lock(&z->mutex);
129 static inline void zone_unlock(struct fio_zone_info *z)
132 pthread_mutex_unlock(&z->mutex);
135 static inline struct fio_zone_info *zbd_get_zone(const struct fio_file *f,
136 unsigned int zone_idx)
138 return &f->zbd_info->zone_info[zone_idx];
141 static inline struct fio_zone_info *
142 zbd_offset_to_zone(const struct fio_file *f, uint64_t offset)
144 return zbd_get_zone(f, zbd_offset_to_zone_idx(f, offset));
147 static bool accounting_vdb(struct thread_data *td, const struct fio_file *f)
149 return td->o.zrt.u.f && td_write(td);
153 * zbd_get_zoned_model - Get a device zoned model
154 * @td: FIO thread data
155 * @f: FIO file for which to get model information
157 static int zbd_get_zoned_model(struct thread_data *td, struct fio_file *f,
158 enum zbd_zoned_model *model)
162 if (f->filetype == FIO_TYPE_PIPE) {
163 log_err("zonemode=zbd does not support pipes\n");
167 /* If regular file, always emulate zones inside the file. */
168 if (f->filetype == FIO_TYPE_FILE) {
173 if (td->io_ops && td->io_ops->get_zoned_model)
174 ret = td->io_ops->get_zoned_model(td, f, model);
176 ret = blkzoned_get_zoned_model(td, f, model);
178 td_verror(td, errno, "get zoned model failed");
179 log_err("%s: get zoned model failed (%d).\n",
180 f->file_name, errno);
187 * zbd_report_zones - Get zone information
188 * @td: FIO thread data.
189 * @f: FIO file for which to get zone information
190 * @offset: offset from which to report zones
191 * @zones: Array of struct zbd_zone
192 * @nr_zones: Size of @zones array
194 * Get zone information into @zones starting from the zone at offset @offset
195 * for the device specified by @f.
197 * Returns the number of zones reported upon success and a negative error code
198 * upon failure. If the zone report is empty, always assume an error (device
199 * problem) and return -EIO.
201 static int zbd_report_zones(struct thread_data *td, struct fio_file *f,
202 uint64_t offset, struct zbd_zone *zones,
203 unsigned int nr_zones)
207 if (td->io_ops && td->io_ops->report_zones)
208 ret = td->io_ops->report_zones(td, f, offset, zones, nr_zones);
210 ret = blkzoned_report_zones(td, f, offset, zones, nr_zones);
212 td_verror(td, errno, "report zones failed");
213 log_err("%s: report zones from sector %"PRIu64" failed (nr_zones=%d; errno=%d).\n",
214 f->file_name, offset >> 9, nr_zones, errno);
215 } else if (ret == 0) {
216 td_verror(td, errno, "Empty zone report");
217 log_err("%s: report zones from sector %"PRIu64" is empty.\n",
218 f->file_name, offset >> 9);
226 * zbd_reset_wp - reset the write pointer of a range of zones
227 * @td: FIO thread data.
228 * @f: FIO file for which to reset zones
229 * @offset: Starting offset of the first zone to reset
230 * @length: Length of the range of zones to reset
232 * Reset the write pointer of all zones in the range @offset...@offset+@length.
233 * Returns 0 upon success and a negative error code upon failure.
235 static int zbd_reset_wp(struct thread_data *td, struct fio_file *f,
236 uint64_t offset, uint64_t length)
240 if (td->io_ops && td->io_ops->reset_wp)
241 ret = td->io_ops->reset_wp(td, f, offset, length);
243 ret = blkzoned_reset_wp(td, f, offset, length);
245 td_verror(td, errno, "resetting wp failed");
246 log_err("%s: resetting wp for %"PRIu64" sectors at sector %"PRIu64" failed (%d).\n",
247 f->file_name, length >> 9, offset >> 9, errno);
254 * __zbd_reset_zone - reset the write pointer of a single zone
255 * @td: FIO thread data.
256 * @f: FIO file associated with the disk for which to reset a write pointer.
259 * Returns 0 upon success and a negative error code upon failure.
261 * The caller must hold z->mutex.
263 static int __zbd_reset_zone(struct thread_data *td, struct fio_file *f,
264 struct fio_zone_info *z)
266 uint64_t offset = z->start;
267 uint64_t length = (z+1)->start - offset;
268 uint64_t data_in_zone = z->wp - z->start;
274 assert(is_valid_offset(f, offset + length - 1));
276 dprint(FD_ZBD, "%s: resetting wp of zone %u.\n",
277 f->file_name, zbd_zone_idx(f, z));
279 switch (f->zbd_info->model) {
281 case ZBD_HOST_MANAGED:
282 ret = zbd_reset_wp(td, f, offset, length);
290 if (accounting_vdb(td, f)) {
291 pthread_mutex_lock(&f->zbd_info->mutex);
292 f->zbd_info->wp_valid_data_bytes -= data_in_zone;
293 pthread_mutex_unlock(&f->zbd_info->mutex);
298 td->ts.nr_zone_resets++;
304 * zbd_write_zone_put - Remove a zone from the write target zones array.
305 * @td: FIO thread data.
306 * @f: FIO file that has the write zones array to remove.
307 * @zone_idx: Index of the zone to remove.
309 * The caller must hold f->zbd_info->mutex.
311 static void zbd_write_zone_put(struct thread_data *td, const struct fio_file *f,
312 struct fio_zone_info *z)
319 for (zi = 0; zi < f->zbd_info->num_write_zones; zi++) {
320 if (zbd_get_zone(f, f->zbd_info->write_zones[zi]) == z)
323 if (zi == f->zbd_info->num_write_zones)
326 dprint(FD_ZBD, "%s: removing zone %u from write zone array\n",
327 f->file_name, zbd_zone_idx(f, z));
329 memmove(f->zbd_info->write_zones + zi,
330 f->zbd_info->write_zones + zi + 1,
331 (ZBD_MAX_WRITE_ZONES - (zi + 1)) *
332 sizeof(f->zbd_info->write_zones[0]));
334 f->zbd_info->num_write_zones--;
335 td->num_write_zones--;
340 * zbd_reset_zone - reset the write pointer of a single zone and remove the zone
341 * from the array of write zones.
342 * @td: FIO thread data.
343 * @f: FIO file associated with the disk for which to reset a write pointer.
346 * Returns 0 upon success and a negative error code upon failure.
348 * The caller must hold z->mutex.
350 static int zbd_reset_zone(struct thread_data *td, struct fio_file *f,
351 struct fio_zone_info *z)
355 ret = __zbd_reset_zone(td, f, z);
359 pthread_mutex_lock(&f->zbd_info->mutex);
360 zbd_write_zone_put(td, f, z);
361 pthread_mutex_unlock(&f->zbd_info->mutex);
366 * zbd_finish_zone - finish the specified zone
367 * @td: FIO thread data.
368 * @f: FIO file for which to finish a zone
369 * @z: Zone to finish.
371 * Finish the zone at @offset with open or close status.
373 static int zbd_finish_zone(struct thread_data *td, struct fio_file *f,
374 struct fio_zone_info *z)
376 uint64_t offset = z->start;
377 uint64_t length = f->zbd_info->zone_size;
380 switch (f->zbd_info->model) {
382 case ZBD_HOST_MANAGED:
383 if (td->io_ops && td->io_ops->finish_zone)
384 ret = td->io_ops->finish_zone(td, f, offset, length);
386 ret = blkzoned_finish_zone(td, f, offset, length);
393 td_verror(td, errno, "finish zone failed");
394 log_err("%s: finish zone at sector %"PRIu64" failed (%d).\n",
395 f->file_name, offset >> 9, errno);
397 z->wp = (z+1)->start;
404 * zbd_reset_zones - Reset a range of zones.
405 * @td: fio thread data.
406 * @f: fio file for which to reset zones
407 * @zb: first zone to reset.
408 * @ze: first zone not to reset.
410 * Returns 0 upon success and 1 upon failure.
412 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
413 struct fio_zone_info *const zb,
414 struct fio_zone_info *const ze)
416 struct fio_zone_info *z;
417 const uint64_t min_bs = td->o.min_bs[DDIR_WRITE];
420 if (fio_unlikely(0 == min_bs))
423 dprint(FD_ZBD, "%s: examining zones %u .. %u\n",
424 f->file_name, zbd_zone_idx(f, zb), zbd_zone_idx(f, ze));
426 for (z = zb; z < ze; z++) {
432 if (z->wp != z->start) {
433 dprint(FD_ZBD, "%s: resetting zone %u\n",
434 f->file_name, zbd_zone_idx(f, z));
435 if (zbd_reset_zone(td, f, z) < 0)
446 * zbd_get_max_open_zones - Get the maximum number of open zones
447 * @td: FIO thread data
448 * @f: FIO file for which to get max open zones
449 * @max_open_zones: Upon success, result will be stored here.
451 * A @max_open_zones value set to zero means no limit.
453 * Returns 0 upon success and a negative error code upon failure.
455 static int zbd_get_max_open_zones(struct thread_data *td, struct fio_file *f,
456 unsigned int *max_open_zones)
460 if (td->io_ops && td->io_ops->get_max_open_zones)
461 ret = td->io_ops->get_max_open_zones(td, f, max_open_zones);
463 ret = blkzoned_get_max_open_zones(td, f, max_open_zones);
465 td_verror(td, errno, "get max open zones failed");
466 log_err("%s: get max open zones failed (%d).\n",
467 f->file_name, errno);
474 * zbd_get_max_active_zones - Get the maximum number of active zones
475 * @td: FIO thread data
476 * @f: FIO file for which to get max active zones
478 * Returns max_active_zones limit value of the target file if it is available.
479 * Otherwise return zero, which means no limit.
481 static unsigned int zbd_get_max_active_zones(struct thread_data *td,
484 unsigned int max_active_zones;
487 if (td->io_ops && td->io_ops->get_max_active_zones)
488 ret = td->io_ops->get_max_active_zones(td, f,
491 ret = blkzoned_get_max_active_zones(td, f, &max_active_zones);
493 dprint(FD_ZBD, "%s: max_active_zones is not available\n",
498 return max_active_zones;
502 * __zbd_write_zone_get - Add a zone to the array of write zones.
503 * @td: fio thread data.
504 * @f: fio file that has the write zones array to add.
505 * @zone_idx: Index of the zone to add.
507 * Do same operation as @zbd_write_zone_get, except it adds the zone at
508 * @zone_idx to write target zones array even when it does not have remainder
509 * space to write one block.
511 static bool __zbd_write_zone_get(struct thread_data *td,
512 const struct fio_file *f,
513 struct fio_zone_info *z)
515 struct zoned_block_device_info *zbdi = f->zbd_info;
516 uint32_t zone_idx = zbd_zone_idx(f, z);
519 if (z->cond == ZBD_ZONE_COND_OFFLINE)
523 * Skip full zones with data verification enabled because resetting a
524 * zone causes data loss and hence causes verification to fail.
526 if (td->o.verify != VERIFY_NONE && zbd_zone_remainder(z) == 0)
530 * zbdi->max_write_zones == 0 means that there is no limit on the
531 * maximum number of write target zones. In this case, do no track write
532 * target zones in zbdi->write_zones array.
534 if (!zbdi->max_write_zones)
537 pthread_mutex_lock(&zbdi->mutex);
541 * If the zone is going to be completely filled by writes
542 * already in-flight, handle it as a full zone instead of a
545 if (!zbd_zone_remainder(z))
551 /* Zero means no limit */
552 if (td->o.job_max_open_zones > 0 &&
553 td->num_write_zones >= td->o.job_max_open_zones)
555 if (zbdi->num_write_zones >= zbdi->max_write_zones)
558 dprint(FD_ZBD, "%s: adding zone %u to write zone array\n",
559 f->file_name, zone_idx);
561 zbdi->write_zones[zbdi->num_write_zones++] = zone_idx;
562 td->num_write_zones++;
567 pthread_mutex_unlock(&zbdi->mutex);
572 * zbd_write_zone_get - Add a zone to the array of write zones.
573 * @td: fio thread data.
574 * @f: fio file that has the open zones to add.
575 * @zone_idx: Index of the zone to add.
577 * Add a ZBD zone to write target zones array, if it is not yet added. Returns
578 * true if either the zone was already added or if the zone was successfully
579 * added to the array without exceeding the maximum number of write zones.
580 * Returns false if the zone was not already added and addition of the zone
581 * would cause the zone limit to be exceeded.
583 static bool zbd_write_zone_get(struct thread_data *td, const struct fio_file *f,
584 struct fio_zone_info *z)
586 const uint64_t min_bs = td->o.min_bs[DDIR_WRITE];
589 * Skip full zones with data verification enabled because resetting a
590 * zone causes data loss and hence causes verification to fail.
592 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
595 return __zbd_write_zone_get(td, f, z);
598 /* Verify whether direct I/O is used for all host-managed zoned block drives. */
599 static bool zbd_using_direct_io(void)
605 if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE))
607 for_each_file(td, f, j) {
608 if (f->zbd_info && f->filetype == FIO_TYPE_BLOCK &&
609 f->zbd_info->model == ZBD_HOST_MANAGED)
617 /* Whether or not the I/O range for f includes one or more sequential zones */
618 static bool zbd_is_seq_job(const struct fio_file *f)
620 uint32_t zone_idx, zone_idx_b, zone_idx_e;
627 zone_idx_b = zbd_offset_to_zone_idx(f, f->file_offset);
629 zbd_offset_to_zone_idx(f, f->file_offset + f->io_size - 1);
630 for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++)
631 if (zbd_get_zone(f, zone_idx)->has_wp)
638 * Verify whether the file offset and size parameters are aligned with zone
639 * boundaries. If the file offset is not aligned, align it down to the start of
640 * the zone containing the start offset and align up the file io_size parameter.
642 static bool zbd_zone_align_file_sizes(struct thread_data *td,
645 const struct fio_zone_info *z;
646 uint64_t new_offset, new_end;
650 if (f->file_offset >= f->real_file_size)
652 if (!zbd_is_seq_job(f))
655 if (!td->o.zone_size) {
656 td->o.zone_size = f->zbd_info->zone_size;
657 if (!td->o.zone_size) {
658 log_err("%s: invalid 0 zone size\n",
662 } else if (td->o.zone_size != f->zbd_info->zone_size) {
663 log_err("%s: zonesize %llu does not match the device zone size %"PRIu64".\n",
664 f->file_name, td->o.zone_size,
665 f->zbd_info->zone_size);
669 if (td->o.zone_skip % td->o.zone_size) {
670 log_err("%s: zoneskip %llu is not a multiple of the device zone size %llu.\n",
671 f->file_name, td->o.zone_skip,
676 if (td->o.td_ddir == TD_DDIR_READ) {
677 z = zbd_offset_to_zone(f, f->file_offset + f->io_size);
679 if (f->file_offset + f->io_size > new_end) {
680 log_info("%s: rounded io_size from %"PRIu64" to %"PRIu64"\n",
681 f->file_name, f->io_size,
682 new_end - f->file_offset);
683 f->io_size = new_end - f->file_offset;
688 z = zbd_offset_to_zone(f, f->file_offset);
689 if (f->file_offset != z->start) {
690 new_offset = zbd_zone_end(z);
691 if (new_offset >= f->file_offset + f->io_size) {
692 log_info("%s: io_size must be at least one zone\n",
696 log_info("%s: rounded up offset from %"PRIu64" to %"PRIu64"\n",
697 f->file_name, f->file_offset,
699 f->io_size -= (new_offset - f->file_offset);
700 f->file_offset = new_offset;
703 z = zbd_offset_to_zone(f, f->file_offset + f->io_size);
705 if (f->file_offset + f->io_size != new_end) {
706 if (new_end <= f->file_offset) {
707 log_info("%s: io_size must be at least one zone\n",
711 log_info("%s: rounded down io_size from %"PRIu64" to %"PRIu64"\n",
712 f->file_name, f->io_size,
713 new_end - f->file_offset);
714 f->io_size = new_end - f->file_offset;
721 * Verify whether offset and size parameters are aligned with zone boundaries.
723 static bool zbd_verify_sizes(void)
729 for_each_file(td, f, j) {
730 if (!zbd_zone_align_file_sizes(td, f))
738 static bool zbd_verify_bs(void)
745 (td->o.min_bs[DDIR_TRIM] != td->o.max_bs[DDIR_TRIM] ||
746 td->o.bssplit_nr[DDIR_TRIM])) {
747 log_info("bsrange and bssplit are not allowed for trim with zonemode=zbd\n");
750 for_each_file(td, f, j) {
756 zone_size = f->zbd_info->zone_size;
757 if (td_trim(td) && td->o.bs[DDIR_TRIM] != zone_size) {
758 log_info("%s: trim block size %llu is not the zone size %"PRIu64"\n",
759 f->file_name, td->o.bs[DDIR_TRIM],
768 static int ilog2(uint64_t i)
780 * Initialize f->zbd_info for devices that are not zoned block devices. This
781 * allows to execute a ZBD workload against a non-ZBD device.
783 static int init_zone_info(struct thread_data *td, struct fio_file *f)
786 struct fio_zone_info *p;
787 uint64_t zone_size = td->o.zone_size;
788 uint64_t zone_capacity = td->o.zone_capacity;
789 struct zoned_block_device_info *zbd_info = NULL;
792 if (zone_size == 0) {
793 log_err("%s: Specifying the zone size is mandatory for regular file/block device with --zonemode=zbd\n\n",
798 if (zone_size < 512) {
799 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
804 if (zone_capacity == 0)
805 zone_capacity = zone_size;
807 if (zone_capacity > zone_size) {
808 log_err("%s: job parameter zonecapacity %llu is larger than zone size %llu\n",
809 f->file_name, td->o.zone_capacity, td->o.zone_size);
813 if (f->real_file_size < zone_size) {
814 log_err("%s: file/device size %"PRIu64" is smaller than zone size %"PRIu64"\n",
815 f->file_name, f->real_file_size, zone_size);
819 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
820 zbd_info = scalloc(1, sizeof(*zbd_info) +
821 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
825 mutex_init_pshared(&zbd_info->mutex);
826 zbd_info->refcount = 1;
827 p = &zbd_info->zone_info[0];
828 for (i = 0; i < nr_zones; i++, p++) {
829 mutex_init_pshared_with_type(&p->mutex,
830 PTHREAD_MUTEX_RECURSIVE);
831 p->start = i * zone_size;
833 p->type = ZBD_ZONE_TYPE_SWR;
834 p->cond = ZBD_ZONE_COND_EMPTY;
835 p->capacity = zone_capacity;
839 p->start = nr_zones * zone_size;
841 f->zbd_info = zbd_info;
842 f->zbd_info->zone_size = zone_size;
843 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
844 ilog2(zone_size) : 0;
845 f->zbd_info->nr_zones = nr_zones;
850 * Maximum number of zones to report in one operation.
852 #define ZBD_REPORT_MAX_ZONES 8192U
855 * Parse the device zone report and store it in f->zbd_info. Must be called
856 * only for devices that are zoned, namely those with a model != ZBD_NONE.
858 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
861 struct zbd_zone *zones, *z;
862 struct fio_zone_info *p;
863 uint64_t zone_size, offset, capacity;
864 bool same_zone_cap = true;
865 struct zoned_block_device_info *zbd_info = NULL;
866 int i, j, ret = -ENOMEM;
868 zones = calloc(ZBD_REPORT_MAX_ZONES, sizeof(struct zbd_zone));
872 nrz = zbd_report_zones(td, f, 0, zones, ZBD_REPORT_MAX_ZONES);
875 log_info("fio: report zones (offset 0) failed for %s (%d).\n",
880 zone_size = zones[0].len;
881 capacity = zones[0].capacity;
882 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
884 if (td->o.zone_size == 0) {
885 td->o.zone_size = zone_size;
886 } else if (td->o.zone_size != zone_size) {
887 log_err("fio: %s job parameter zonesize %llu does not match disk zone size %"PRIu64".\n",
888 f->file_name, td->o.zone_size, zone_size);
893 dprint(FD_ZBD, "Device %s has %d zones of size %"PRIu64" KB\n",
894 f->file_name, nr_zones, zone_size / 1024);
896 zbd_info = scalloc(1, sizeof(*zbd_info) +
897 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
900 mutex_init_pshared(&zbd_info->mutex);
901 zbd_info->refcount = 1;
902 p = &zbd_info->zone_info[0];
903 for (offset = 0, j = 0; j < nr_zones;) {
905 for (i = 0; i < nrz; i++, j++, z++, p++) {
906 mutex_init_pshared_with_type(&p->mutex,
907 PTHREAD_MUTEX_RECURSIVE);
909 p->capacity = z->capacity;
910 if (capacity != z->capacity)
911 same_zone_cap = false;
914 case ZBD_ZONE_COND_NOT_WP:
915 case ZBD_ZONE_COND_FULL:
916 p->wp = p->start + p->capacity;
919 assert(z->start <= z->wp);
920 assert(z->wp <= z->start + zone_size);
926 case ZBD_ZONE_TYPE_SWR:
935 if (j > 0 && p->start != p[-1].start + zone_size) {
936 log_info("%s: invalid zone data [%d:%d]: %"PRIu64" + %"PRIu64" != %"PRIu64"\n",
938 p[-1].start, zone_size, p->start);
944 offset = z->start + z->len;
948 nrz = zbd_report_zones(td, f, offset, zones,
949 min((uint32_t)(nr_zones - j),
950 ZBD_REPORT_MAX_ZONES));
953 log_info("fio: report zones (offset %"PRIu64") failed for %s (%d).\n",
954 offset, f->file_name, -ret);
960 zbd_info->zone_info[nr_zones].start = offset;
962 f->zbd_info = zbd_info;
963 f->zbd_info->zone_size = zone_size;
964 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
965 ilog2(zone_size) : 0;
966 f->zbd_info->nr_zones = nr_zones;
967 f->zbd_info->max_active_zones = zbd_get_max_active_zones(td, f);
970 dprint(FD_ZBD, "Zone capacity = %"PRIu64" KB\n",
982 static int zbd_set_max_write_zones(struct thread_data *td, struct fio_file *f)
984 struct zoned_block_device_info *zbd = f->zbd_info;
985 unsigned int max_open_zones;
988 if (zbd->model != ZBD_HOST_MANAGED || td->o.ignore_zone_limits) {
989 /* Only host-managed devices have a max open limit */
990 zbd->max_write_zones = td->o.max_open_zones;
994 /* If host-managed, get the max open limit */
995 ret = zbd_get_max_open_zones(td, f, &max_open_zones);
999 if (!max_open_zones) {
1000 /* No device limit */
1001 zbd->max_write_zones = td->o.max_open_zones;
1002 } else if (!td->o.max_open_zones) {
1003 /* No user limit. Set limit to device limit */
1004 zbd->max_write_zones = max_open_zones;
1005 } else if (td->o.max_open_zones <= max_open_zones) {
1006 /* Both user limit and dev limit. User limit not too large */
1007 zbd->max_write_zones = td->o.max_open_zones;
1009 /* Both user limit and dev limit. User limit too large */
1010 td_verror(td, EINVAL,
1011 "Specified --max_open_zones is too large");
1012 log_err("Specified --max_open_zones (%d) is larger than max (%u)\n",
1013 td->o.max_open_zones, max_open_zones);
1018 /* Ensure that the limit is not larger than FIO's internal limit */
1019 if (zbd->max_write_zones > ZBD_MAX_WRITE_ZONES) {
1020 td_verror(td, EINVAL, "'max_open_zones' value is too large");
1021 log_err("'max_open_zones' value is larger than %u\n",
1022 ZBD_MAX_WRITE_ZONES);
1026 dprint(FD_ZBD, "%s: using max write zones limit: %"PRIu32"\n",
1027 f->file_name, zbd->max_write_zones);
1033 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
1035 * Returns 0 upon success and a negative error code upon failure.
1037 static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
1039 enum zbd_zoned_model zbd_model;
1042 assert(td->o.zone_mode == ZONE_MODE_ZBD);
1044 ret = zbd_get_zoned_model(td, f, &zbd_model);
1048 switch (zbd_model) {
1049 case ZBD_HOST_AWARE:
1050 case ZBD_HOST_MANAGED:
1051 ret = parse_zone_info(td, f);
1056 ret = init_zone_info(td, f);
1061 td_verror(td, EINVAL, "Unsupported zoned model");
1062 log_err("Unsupported zoned model\n");
1066 assert(f->zbd_info);
1067 f->zbd_info->model = zbd_model;
1069 ret = zbd_set_max_write_zones(td, f);
1071 zbd_free_zone_info(f);
1078 void zbd_free_zone_info(struct fio_file *f)
1082 assert(f->zbd_info);
1084 pthread_mutex_lock(&f->zbd_info->mutex);
1085 refcount = --f->zbd_info->refcount;
1086 pthread_mutex_unlock(&f->zbd_info->mutex);
1088 assert((int32_t)refcount >= 0);
1095 * Initialize f->zbd_info.
1097 * Returns 0 upon success and a negative error code upon failure.
1099 * Note: this function can only work correctly if it is called before the first
1102 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
1104 struct fio_file *f2;
1108 for_each_file(td2, f2, j) {
1109 if (td2 == td && f2 == file)
1111 if (!f2->zbd_info ||
1112 strcmp(f2->file_name, file->file_name) != 0)
1114 file->zbd_info = f2->zbd_info;
1115 file->zbd_info->refcount++;
1120 ret = zbd_create_zone_info(td, file);
1122 td_verror(td, -ret, "zbd_create_zone_info() failed");
1127 int zbd_init_files(struct thread_data *td)
1132 for_each_file(td, f, i) {
1133 if (zbd_init_zone_info(td, f))
1140 void zbd_recalc_options_with_zone_granularity(struct thread_data *td)
1145 for_each_file(td, f, i) {
1146 struct zoned_block_device_info *zbd = f->zbd_info;
1149 /* zonemode=strided doesn't get per-file zone size. */
1150 zone_size = zbd ? zbd->zone_size : td->o.zone_size;
1154 if (td->o.size_nz > 0)
1155 td->o.size = td->o.size_nz * zone_size;
1156 if (td->o.io_size_nz > 0)
1157 td->o.io_size = td->o.io_size_nz * zone_size;
1158 if (td->o.start_offset_nz > 0)
1159 td->o.start_offset = td->o.start_offset_nz * zone_size;
1160 if (td->o.offset_increment_nz > 0)
1161 td->o.offset_increment =
1162 td->o.offset_increment_nz * zone_size;
1163 if (td->o.zone_skip_nz > 0)
1164 td->o.zone_skip = td->o.zone_skip_nz * zone_size;
1168 static uint64_t zbd_verify_and_set_vdb(struct thread_data *td,
1169 const struct fio_file *f)
1171 struct fio_zone_info *zb, *ze, *z;
1172 uint64_t wp_vdb = 0;
1173 struct zoned_block_device_info *zbdi = f->zbd_info;
1175 assert(td->runstate < TD_RUNNING);
1178 if (!accounting_vdb(td, f))
1182 * Ensure that the I/O range includes one or more sequential zones so
1183 * that f->min_zone and f->max_zone have different values.
1185 if (!zbd_is_seq_job(f))
1188 if (zbdi->write_min_zone != zbdi->write_max_zone) {
1189 if (zbdi->write_min_zone != f->min_zone ||
1190 zbdi->write_max_zone != f->max_zone) {
1191 td_verror(td, EINVAL,
1192 "multi-jobs with different write ranges are "
1193 "not supported with zone_reset_threshold");
1194 log_err("multi-jobs with different write ranges are "
1195 "not supported with zone_reset_threshold\n");
1200 zbdi->write_min_zone = f->min_zone;
1201 zbdi->write_max_zone = f->max_zone;
1203 zb = zbd_get_zone(f, f->min_zone);
1204 ze = zbd_get_zone(f, f->max_zone);
1205 for (z = zb; z < ze; z++)
1207 wp_vdb += z->wp - z->start;
1209 zbdi->wp_valid_data_bytes = wp_vdb;
1214 int zbd_setup_files(struct thread_data *td)
1219 if (!zbd_using_direct_io()) {
1220 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
1224 if (!zbd_verify_sizes())
1227 if (!zbd_verify_bs())
1230 if (td->o.experimental_verify) {
1231 log_err("zonemode=zbd does not support experimental verify\n");
1235 for_each_file(td, f, i) {
1236 struct zoned_block_device_info *zbd = f->zbd_info;
1237 struct fio_zone_info *z;
1243 f->min_zone = zbd_offset_to_zone_idx(f, f->file_offset);
1245 zbd_offset_to_zone_idx(f, f->file_offset + f->io_size);
1247 vdb = zbd_verify_and_set_vdb(td, f);
1249 dprint(FD_ZBD, "%s(%s): valid data bytes = %" PRIu64 "\n",
1250 __func__, f->file_name, vdb);
1253 * When all zones in the I/O range are conventional, io_size
1254 * can be smaller than zone size, making min_zone the same
1255 * as max_zone. This is why the assert below needs to be made
1258 if (zbd_is_seq_job(f))
1259 assert(f->min_zone < f->max_zone);
1261 if (td->o.max_open_zones > 0 &&
1262 zbd->max_write_zones != td->o.max_open_zones) {
1263 log_err("Different 'max_open_zones' values\n");
1268 * The per job max open zones limit cannot be used without a
1269 * global max open zones limit. (As the tracking of open zones
1270 * is disabled when there is no global max open zones limit.)
1272 if (td->o.job_max_open_zones && !zbd->max_write_zones) {
1273 log_err("'job_max_open_zones' cannot be used without a global open zones limit\n");
1278 * zbd->max_write_zones is the global limit shared for all jobs
1279 * that target the same zoned block device. Force sync the per
1280 * thread global limit with the actual global limit. (The real
1281 * per thread/job limit is stored in td->o.job_max_open_zones).
1283 td->o.max_open_zones = zbd->max_write_zones;
1285 for (zi = f->min_zone; zi < f->max_zone; zi++) {
1286 z = &zbd->zone_info[zi];
1287 if (z->cond != ZBD_ZONE_COND_IMP_OPEN &&
1288 z->cond != ZBD_ZONE_COND_EXP_OPEN &&
1289 z->cond != ZBD_ZONE_COND_CLOSED)
1291 if (!zbd->max_active_zones &&
1292 z->cond == ZBD_ZONE_COND_CLOSED)
1294 if (__zbd_write_zone_get(td, f, z))
1297 * If the number of open zones exceeds specified limits,
1300 log_err("Number of open zones exceeds max_open_zones limit\n");
1309 * Reset zbd_info.write_cnt, the counter that counts down towards the next
1312 static void _zbd_reset_write_cnt(const struct thread_data *td,
1313 const struct fio_file *f)
1315 assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
1317 f->zbd_info->write_cnt = td->o.zrf.u.f ?
1318 min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
1321 static void zbd_reset_write_cnt(const struct thread_data *td,
1322 const struct fio_file *f)
1324 pthread_mutex_lock(&f->zbd_info->mutex);
1325 _zbd_reset_write_cnt(td, f);
1326 pthread_mutex_unlock(&f->zbd_info->mutex);
1329 static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
1330 const struct fio_file *f)
1332 uint32_t write_cnt = 0;
1334 pthread_mutex_lock(&f->zbd_info->mutex);
1335 assert(f->zbd_info->write_cnt);
1336 if (f->zbd_info->write_cnt)
1337 write_cnt = --f->zbd_info->write_cnt;
1339 _zbd_reset_write_cnt(td, f);
1340 pthread_mutex_unlock(&f->zbd_info->mutex);
1342 return write_cnt == 0;
1345 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
1347 struct fio_zone_info *zb, *ze;
1348 bool verify_data_left = false;
1350 if (!f->zbd_info || !td_write(td))
1353 zb = zbd_get_zone(f, f->min_zone);
1354 ze = zbd_get_zone(f, f->max_zone);
1357 * If data verification is enabled reset the affected zones before
1358 * writing any data to avoid that a zone reset has to be issued while
1359 * writing data, which causes data loss.
1361 if (td->o.verify != VERIFY_NONE) {
1362 verify_data_left = td->runstate == TD_VERIFYING ||
1363 td->io_hist_len || td->verify_batch;
1364 if (td->io_hist_len && td->o.verify_backlog)
1366 td->io_hist_len % td->o.verify_backlog;
1367 if (!verify_data_left)
1368 zbd_reset_zones(td, f, zb, ze);
1371 zbd_reset_write_cnt(td, f);
1374 /* Return random zone index for one of the write target zones. */
1375 static uint32_t pick_random_zone_idx(const struct fio_file *f,
1376 const struct io_u *io_u)
1378 return (io_u->offset - f->file_offset) *
1379 f->zbd_info->num_write_zones / f->io_size;
1382 static bool any_io_in_flight(void)
1385 if (td->io_u_in_flight)
1393 * Modify the offset of an I/O unit that does not refer to a zone such that
1394 * in write target zones array. Add a zone to or remove a zone from the lsit if
1395 * necessary. The write target zone is searched across sequential zones.
1396 * This algorithm can only work correctly if all write pointers are
1397 * a multiple of the fio block size. The caller must neither hold z->mutex
1398 * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
1400 static struct fio_zone_info *zbd_convert_to_write_zone(struct thread_data *td,
1403 const uint64_t min_bs = td->o.min_bs[io_u->ddir];
1404 struct fio_file *f = io_u->file;
1405 struct zoned_block_device_info *zbdi = f->zbd_info;
1406 struct fio_zone_info *z;
1407 unsigned int write_zone_idx = -1;
1408 uint32_t zone_idx, new_zone_idx;
1410 bool wait_zone_write;
1412 bool should_retry = true;
1414 assert(is_valid_offset(f, io_u->offset));
1416 if (zbdi->max_write_zones || td->o.job_max_open_zones) {
1418 * This statement accesses zbdi->write_zones[] on purpose
1421 zone_idx = zbdi->write_zones[pick_random_zone_idx(f, io_u)];
1423 zone_idx = zbd_offset_to_zone_idx(f, io_u->offset);
1425 if (zone_idx < f->min_zone)
1426 zone_idx = f->min_zone;
1427 else if (zone_idx >= f->max_zone)
1428 zone_idx = f->max_zone - 1;
1431 "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
1432 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
1435 * Since z->mutex is the outer lock and zbdi->mutex the inner
1436 * lock it can happen that the state of the zone with index zone_idx
1437 * has changed after 'z' has been assigned and before zbdi->mutex
1438 * has been obtained. Hence the loop.
1443 z = zbd_get_zone(f, zone_idx);
1445 zone_lock(td, f, z);
1447 pthread_mutex_lock(&zbdi->mutex);
1450 if (z->cond != ZBD_ZONE_COND_OFFLINE &&
1451 zbdi->max_write_zones == 0 &&
1452 td->o.job_max_open_zones == 0)
1454 if (zbdi->num_write_zones == 0) {
1455 dprint(FD_ZBD, "%s(%s): no zone is write target\n",
1456 __func__, f->file_name);
1457 goto choose_other_zone;
1462 * Array of write target zones is per-device, shared across all
1463 * threads. Start with quasi-random candidate zone. Ignore
1464 * zones which don't belong to thread's offset/size area.
1466 write_zone_idx = pick_random_zone_idx(f, io_u);
1467 assert(!write_zone_idx ||
1468 write_zone_idx < zbdi->num_write_zones);
1469 tmp_idx = write_zone_idx;
1471 for (i = 0; i < zbdi->num_write_zones; i++) {
1474 if (tmp_idx >= zbdi->num_write_zones)
1476 tmpz = zbdi->write_zones[tmp_idx];
1477 if (f->min_zone <= tmpz && tmpz < f->max_zone) {
1478 write_zone_idx = tmp_idx;
1479 goto found_candidate_zone;
1485 dprint(FD_ZBD, "%s(%s): no candidate zone\n",
1486 __func__, f->file_name);
1488 pthread_mutex_unlock(&zbdi->mutex);
1495 found_candidate_zone:
1496 new_zone_idx = zbdi->write_zones[write_zone_idx];
1497 if (new_zone_idx == zone_idx)
1499 zone_idx = new_zone_idx;
1501 pthread_mutex_unlock(&zbdi->mutex);
1507 /* Both z->mutex and zbdi->mutex are held. */
1510 if (zbd_zone_remainder(z) >= min_bs) {
1511 pthread_mutex_unlock(&zbdi->mutex);
1516 /* Check if number of write target zones reaches one of limits. */
1518 zbdi->num_write_zones == f->max_zone - f->min_zone ||
1519 (zbdi->max_write_zones &&
1520 zbdi->num_write_zones == zbdi->max_write_zones) ||
1521 (td->o.job_max_open_zones &&
1522 td->num_write_zones == td->o.job_max_open_zones);
1524 pthread_mutex_unlock(&zbdi->mutex);
1526 /* Only z->mutex is held. */
1529 * When number of write target zones reaches to one of limits, wait for
1530 * zone write completion to one of them before trying a new zone.
1532 if (wait_zone_write) {
1534 "%s(%s): quiesce to remove a zone from write target zones array\n",
1535 __func__, f->file_name);
1540 /* Zone 'z' is full, so try to choose a new zone. */
1541 for (i = f->io_size / zbdi->zone_size; i > 0; i--) {
1546 if (!is_valid_offset(f, z->start)) {
1548 zone_idx = f->min_zone;
1549 z = zbd_get_zone(f, zone_idx);
1551 assert(is_valid_offset(f, z->start));
1554 zone_lock(td, f, z);
1557 if (zbd_write_zone_get(td, f, z))
1561 /* Only z->mutex is held. */
1563 /* Check whether the write fits in any of the write target zones. */
1564 pthread_mutex_lock(&zbdi->mutex);
1565 for (i = 0; i < zbdi->num_write_zones; i++) {
1566 zone_idx = zbdi->write_zones[i];
1567 if (zone_idx < f->min_zone || zone_idx >= f->max_zone)
1569 pthread_mutex_unlock(&zbdi->mutex);
1572 z = zbd_get_zone(f, zone_idx);
1574 zone_lock(td, f, z);
1575 if (zbd_zone_remainder(z) >= min_bs)
1577 pthread_mutex_lock(&zbdi->mutex);
1581 * When any I/O is in-flight or when all I/Os in-flight get completed,
1582 * the I/Os might have removed zones from the write target array then
1583 * retry the steps to choose a zone. Before retry, call io_u_quiesce()
1584 * to complete in-flight writes.
1586 in_flight = any_io_in_flight();
1587 if (in_flight || should_retry) {
1589 "%s(%s): wait zone write and retry write target zone selection\n",
1590 __func__, f->file_name);
1591 should_retry = in_flight;
1592 pthread_mutex_unlock(&zbdi->mutex);
1595 zone_lock(td, f, z);
1599 pthread_mutex_unlock(&zbdi->mutex);
1603 dprint(FD_ZBD, "%s(%s): did not choose another write zone\n",
1604 __func__, f->file_name);
1609 dprint(FD_ZBD, "%s(%s): returning zone %d\n",
1610 __func__, f->file_name, zone_idx);
1612 io_u->offset = z->start;
1614 assert(z->cond != ZBD_ZONE_COND_OFFLINE);
1620 * Find another zone which has @min_bytes of readable data. Search in zones
1621 * @zb + 1 .. @zl. For random workload, also search in zones @zb - 1 .. @zf.
1623 * Either returns NULL or returns a zone pointer. When the zone has write
1624 * pointer, hold the mutex for the zone.
1626 static struct fio_zone_info *
1627 zbd_find_zone(struct thread_data *td, struct io_u *io_u, uint64_t min_bytes,
1628 struct fio_zone_info *zb, struct fio_zone_info *zl)
1630 struct fio_file *f = io_u->file;
1631 struct fio_zone_info *z1, *z2;
1632 const struct fio_zone_info *const zf = zbd_get_zone(f, f->min_zone);
1635 * Skip to the next non-empty zone in case of sequential I/O and to
1636 * the nearest non-empty zone in case of random I/O.
1638 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1639 if (z1 < zl && z1->cond != ZBD_ZONE_COND_OFFLINE) {
1641 zone_lock(td, f, z1);
1642 if (z1->start + min_bytes <= z1->wp)
1646 } else if (!td_random(td)) {
1650 if (td_random(td) && z2 >= zf &&
1651 z2->cond != ZBD_ZONE_COND_OFFLINE) {
1653 zone_lock(td, f, z2);
1654 if (z2->start + min_bytes <= z2->wp)
1662 "%s: no zone has %"PRIu64" bytes of readable data\n",
1663 f->file_name, min_bytes);
1669 * zbd_end_zone_io - update zone status at command completion
1671 * @z: zone info pointer
1673 * If the write command made the zone full, remove it from the write target
1676 * The caller must hold z->mutex.
1678 static void zbd_end_zone_io(struct thread_data *td, const struct io_u *io_u,
1679 struct fio_zone_info *z)
1681 const struct fio_file *f = io_u->file;
1683 if (io_u->ddir == DDIR_WRITE &&
1684 io_u->offset + io_u->buflen >= zbd_zone_capacity_end(z)) {
1685 pthread_mutex_lock(&f->zbd_info->mutex);
1686 zbd_write_zone_put(td, f, z);
1687 pthread_mutex_unlock(&f->zbd_info->mutex);
1692 * zbd_queue_io - update the write pointer of a sequential zone
1694 * @success: Whether or not the I/O unit has been queued successfully
1695 * @q: queueing status (busy, completed or queued).
1697 * For write and trim operations, update the write pointer of the I/O unit
1700 static void zbd_queue_io(struct thread_data *td, struct io_u *io_u, int q,
1703 const struct fio_file *f = io_u->file;
1704 struct zoned_block_device_info *zbd_info = f->zbd_info;
1705 struct fio_zone_info *z;
1710 z = zbd_offset_to_zone(f, io_u->offset);
1717 "%s: queued I/O (%lld, %llu) for zone %u\n",
1718 f->file_name, io_u->offset, io_u->buflen, zbd_zone_idx(f, z));
1720 switch (io_u->ddir) {
1722 zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
1723 zbd_zone_capacity_end(z));
1726 * z->wp > zone_end means that one or more I/O errors
1729 if (accounting_vdb(td, f) && z->wp <= zone_end) {
1730 pthread_mutex_lock(&zbd_info->mutex);
1731 zbd_info->wp_valid_data_bytes += zone_end - z->wp;
1732 pthread_mutex_unlock(&zbd_info->mutex);
1740 if (q == FIO_Q_COMPLETED && !io_u->error)
1741 zbd_end_zone_io(td, io_u, z);
1744 if (!success || q != FIO_Q_QUEUED) {
1745 /* BUSY or COMPLETED: unlock the zone */
1747 io_u->zbd_put_io = NULL;
1752 * zbd_put_io - Unlock an I/O unit target zone lock
1755 static void zbd_put_io(struct thread_data *td, const struct io_u *io_u)
1757 const struct fio_file *f = io_u->file;
1758 struct fio_zone_info *z;
1760 assert(f->zbd_info);
1762 z = zbd_offset_to_zone(f, io_u->offset);
1766 "%s: terminate I/O (%lld, %llu) for zone %u\n",
1767 f->file_name, io_u->offset, io_u->buflen, zbd_zone_idx(f, z));
1769 zbd_end_zone_io(td, io_u, z);
1775 * Windows and MacOS do not define this.
1778 #define EREMOTEIO 121 /* POSIX value */
1781 bool zbd_unaligned_write(int error_code)
1783 switch (error_code) {
1792 * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives
1793 * @td: FIO thread data.
1794 * @io_u: FIO I/O unit.
1796 * For sequential workloads, change the file offset to skip zoneskip bytes when
1797 * no more IO can be performed in the current zone.
1798 * - For read workloads, zoneskip is applied when the io has reached the end of
1799 * the zone or the zone write position (when td->o.read_beyond_wp is false).
1800 * - For write workloads, zoneskip is applied when the zone is full.
1801 * This applies only to read and write operations.
1803 void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
1805 struct fio_file *f = io_u->file;
1806 enum fio_ddir ddir = io_u->ddir;
1807 struct fio_zone_info *z;
1809 assert(td->o.zone_mode == ZONE_MODE_ZBD);
1810 assert(td->o.zone_size);
1811 assert(f->zbd_info);
1813 z = zbd_offset_to_zone(f, f->last_pos[ddir]);
1816 * When the zone capacity is smaller than the zone size and the I/O is
1817 * sequential write, skip to zone end if the latest position is at the
1818 * zone capacity limit.
1820 if (z->capacity < f->zbd_info->zone_size &&
1821 !td_random(td) && ddir == DDIR_WRITE &&
1822 f->last_pos[ddir] >= zbd_zone_capacity_end(z)) {
1824 "%s: Jump from zone capacity limit to zone end:"
1825 " (%"PRIu64" -> %"PRIu64") for zone %u (%"PRIu64")\n",
1826 f->file_name, f->last_pos[ddir],
1827 zbd_zone_end(z), zbd_zone_idx(f, z), z->capacity);
1828 td->io_skip_bytes += zbd_zone_end(z) - f->last_pos[ddir];
1829 f->last_pos[ddir] = zbd_zone_end(z);
1833 * zone_skip is valid only for sequential workloads.
1835 if (td_random(td) || !td->o.zone_skip)
1839 * It is time to switch to a new zone if:
1840 * - zone_bytes == zone_size bytes have already been accessed
1841 * - The last position reached the end of the current zone.
1842 * - For reads with td->o.read_beyond_wp == false, the last position
1843 * reached the zone write pointer.
1845 if (td->zone_bytes >= td->o.zone_size ||
1846 f->last_pos[ddir] >= zbd_zone_end(z) ||
1847 (ddir == DDIR_READ &&
1848 (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) {
1853 f->file_offset += td->o.zone_size + td->o.zone_skip;
1856 * Wrap from the beginning, if we exceed the file size
1858 if (f->file_offset >= f->real_file_size)
1859 f->file_offset = get_start_offset(td, f);
1861 f->last_pos[ddir] = f->file_offset;
1862 td->io_skip_bytes += td->o.zone_skip;
1867 * zbd_adjust_ddir - Adjust an I/O direction for zonemode=zbd.
1869 * @td: FIO thread data.
1870 * @io_u: FIO I/O unit.
1871 * @ddir: I/O direction before adjustment.
1873 * Return adjusted I/O direction.
1875 enum fio_ddir zbd_adjust_ddir(struct thread_data *td, struct io_u *io_u,
1879 * In case read direction is chosen for the first random I/O, fio with
1880 * zonemode=zbd stops because no data can be read from zoned block
1881 * devices with all empty zones. Overwrite the first I/O direction as
1882 * write to make sure data to read exists.
1884 assert(io_u->file->zbd_info);
1885 if (ddir != DDIR_READ || !td_rw(td))
1888 if (io_u->file->last_start[DDIR_WRITE] != -1ULL ||
1889 td->o.read_beyond_wp || td->o.rwmix[DDIR_WRITE] == 0)
1896 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1897 * @td: FIO thread data.
1898 * @io_u: FIO I/O unit.
1900 * Locking strategy: returns with z->mutex locked if and only if z refers
1901 * to a sequential zone and if io_u_accept is returned. z is the zone that
1902 * corresponds to io_u->offset at the end of this function.
1904 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1906 struct fio_file *f = io_u->file;
1907 struct zoned_block_device_info *zbdi = f->zbd_info;
1908 struct fio_zone_info *zb, *zl, *orig_zb;
1909 uint32_t orig_len = io_u->buflen;
1910 uint64_t min_bs = td->o.min_bs[io_u->ddir];
1916 assert(is_valid_offset(f, io_u->offset));
1917 assert(io_u->buflen);
1919 zb = zbd_offset_to_zone(f, io_u->offset);
1923 /* Accept non-write I/Os for conventional zones. */
1924 if (io_u->ddir != DDIR_WRITE)
1928 * Make sure that writes to conventional zones
1929 * don't cross over to any sequential zones.
1931 if (!(zb + 1)->has_wp ||
1932 io_u->offset + io_u->buflen <= (zb + 1)->start)
1935 if (io_u->offset + min_bs > (zb + 1)->start) {
1937 "%s: off=%llu + min_bs=%"PRIu64" > next zone %"PRIu64"\n",
1938 f->file_name, io_u->offset,
1939 min_bs, (zb + 1)->start);
1941 zb->start + (zb + 1)->start - io_u->offset;
1942 new_len = min(io_u->buflen,
1943 (zb + 1)->start - io_u->offset);
1945 new_len = (zb + 1)->start - io_u->offset;
1948 io_u->buflen = new_len / min_bs * min_bs;
1954 * Accept the I/O offset for reads if reading beyond the write pointer
1957 if (zb->cond != ZBD_ZONE_COND_OFFLINE &&
1958 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1961 zone_lock(td, f, zb);
1963 switch (io_u->ddir) {
1965 if (td->runstate == TD_VERIFYING && td_write(td))
1969 * Check that there is enough written data in the zone to do an
1970 * I/O of at least min_bs B. If there isn't, find a new zone for
1973 range = zb->cond != ZBD_ZONE_COND_OFFLINE ?
1974 zb->wp - zb->start : 0;
1975 if (range < min_bs ||
1976 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1978 zl = zbd_get_zone(f, f->max_zone);
1979 zb = zbd_find_zone(td, io_u, min_bs, zb, zl);
1982 "%s: zbd_find_zone(%lld, %llu) failed\n",
1983 f->file_name, io_u->offset,
1988 * zbd_find_zone() returned a zone with a range of at
1991 range = zb->wp - zb->start;
1992 assert(range >= min_bs);
1995 io_u->offset = zb->start;
1999 * Make sure the I/O is within the zone valid data range while
2000 * maximizing the I/O size and preserving randomness.
2002 if (range <= io_u->buflen)
2003 io_u->offset = zb->start;
2004 else if (td_random(td))
2005 io_u->offset = zb->start +
2006 ((io_u->offset - orig_zb->start) %
2007 (range - io_u->buflen)) / min_bs * min_bs;
2010 * When zbd_find_zone() returns a conventional zone,
2011 * we can simply accept the new i/o offset here.
2017 * Make sure the I/O does not cross over the zone wp position.
2019 new_len = min((unsigned long long)io_u->buflen,
2020 (unsigned long long)(zb->wp - io_u->offset));
2021 new_len = new_len / min_bs * min_bs;
2022 if (new_len < io_u->buflen) {
2023 io_u->buflen = new_len;
2024 dprint(FD_IO, "Changed length from %u into %llu\n",
2025 orig_len, io_u->buflen);
2028 assert(zb->start <= io_u->offset);
2029 assert(io_u->offset + io_u->buflen <= zb->wp);
2034 if (io_u->buflen > zbdi->zone_size) {
2035 td_verror(td, EINVAL, "I/O buflen exceeds zone size");
2037 "%s: I/O buflen %llu exceeds zone size %"PRIu64"\n",
2038 f->file_name, io_u->buflen, zbdi->zone_size);
2043 if (zbd_zone_remainder(zb) > 0 &&
2044 zbd_zone_remainder(zb) < min_bs) {
2045 pthread_mutex_lock(&f->zbd_info->mutex);
2046 zbd_write_zone_put(td, f, zb);
2047 pthread_mutex_unlock(&f->zbd_info->mutex);
2049 "%s: finish zone %d\n",
2050 f->file_name, zbd_zone_idx(f, zb));
2052 zbd_finish_zone(td, f, zb);
2053 if (zbd_zone_idx(f, zb) + 1 >= f->max_zone) {
2059 /* Find the next write pointer zone */
2062 if (zbd_zone_idx(f, zb) >= f->max_zone)
2063 zb = zbd_get_zone(f, f->min_zone);
2064 } while (!zb->has_wp);
2066 zone_lock(td, f, zb);
2069 if (!zbd_write_zone_get(td, f, zb)) {
2071 zb = zbd_convert_to_write_zone(td, io_u);
2073 dprint(FD_IO, "%s: can't convert to write target zone",
2079 if (zbd_zone_remainder(zb) > 0 &&
2080 zbd_zone_remainder(zb) < min_bs)
2083 /* Check whether the zone reset threshold has been exceeded */
2084 if (td->o.zrf.u.f) {
2085 if (zbdi->wp_valid_data_bytes >=
2086 f->io_size * td->o.zrt.u.f &&
2087 zbd_dec_and_reset_write_cnt(td, f))
2091 /* Reset the zone pointer if necessary */
2092 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
2093 if (td->o.verify != VERIFY_NONE) {
2095 * Unset io-u->file to tell get_next_verify()
2096 * that this IO is not requeue.
2099 if (!get_next_verify(td, io_u)) {
2107 * Since previous write requests may have been submitted
2108 * asynchronously and since we will submit the zone
2109 * reset synchronously, wait until previously submitted
2110 * write requests have completed before issuing a
2115 if (__zbd_reset_zone(td, f, zb) < 0)
2118 if (zb->capacity < min_bs) {
2119 td_verror(td, EINVAL, "ZCAP is less min_bs");
2120 log_err("zone capacity %"PRIu64" smaller than minimum block size %"PRIu64"\n",
2121 zb->capacity, min_bs);
2126 /* Make writes occur at the write pointer */
2127 assert(!zbd_zone_full(f, zb, min_bs));
2128 io_u->offset = zb->wp;
2129 if (!is_valid_offset(f, io_u->offset)) {
2130 td_verror(td, EINVAL, "invalid WP value");
2131 dprint(FD_ZBD, "%s: dropped request with offset %llu\n",
2132 f->file_name, io_u->offset);
2137 * Make sure that the buflen is a multiple of the minimal
2138 * block size. Give up if shrinking would make the request too
2141 new_len = min((unsigned long long)io_u->buflen,
2142 zbd_zone_capacity_end(zb) - io_u->offset);
2143 new_len = new_len / min_bs * min_bs;
2144 if (new_len == io_u->buflen)
2146 if (new_len >= min_bs) {
2147 io_u->buflen = new_len;
2148 dprint(FD_IO, "Changed length from %u into %llu\n",
2149 orig_len, io_u->buflen);
2153 td_verror(td, EIO, "zone remainder too small");
2154 log_err("zone remainder %lld smaller than min block size %"PRIu64"\n",
2155 (zbd_zone_capacity_end(zb) - io_u->offset), min_bs);
2160 /* Check random trim targets a non-empty zone */
2161 if (!td_random(td) || zb->wp > zb->start)
2164 /* Find out a non-empty zone to trim */
2166 zl = zbd_get_zone(f, f->max_zone);
2167 zb = zbd_find_zone(td, io_u, 1, zb, zl);
2169 io_u->offset = zb->start;
2170 dprint(FD_ZBD, "%s: found new zone(%lld) for trim\n",
2171 f->file_name, io_u->offset);
2180 case DDIR_SYNC_FILE_RANGE:
2192 assert(zb->cond != ZBD_ZONE_COND_OFFLINE);
2193 assert(!io_u->zbd_queue_io);
2194 assert(!io_u->zbd_put_io);
2196 io_u->zbd_queue_io = zbd_queue_io;
2197 io_u->zbd_put_io = zbd_put_io;
2200 * Since we return with the zone lock still held,
2201 * add an annotation to let Coverity know that it
2204 /* coverity[missing_unlock] */
2209 if (zb && zb->has_wp)
2215 /* Return a string with ZBD statistics */
2216 char *zbd_write_status(const struct thread_stat *ts)
2220 if (asprintf(&res, "; %"PRIu64" zone resets", ts->nr_zone_resets) < 0)
2226 * zbd_do_io_u_trim - If reset zone is applicable, do reset zone instead of trim
2228 * @td: FIO thread data.
2229 * @io_u: FIO I/O unit.
2231 * It is assumed that z->mutex is already locked.
2232 * Return io_u_completed when reset zone succeeds. Return 0 when the target zone
2233 * does not have write pointer. On error, return negative errno.
2235 int zbd_do_io_u_trim(struct thread_data *td, struct io_u *io_u)
2237 struct fio_file *f = io_u->file;
2238 struct fio_zone_info *z;
2241 z = zbd_offset_to_zone(f, io_u->offset);
2245 if (io_u->offset != z->start) {
2246 log_err("Trim offset not at zone start (%lld)\n",
2251 ret = zbd_reset_zone((struct thread_data *)td, f, z);
2255 return io_u_completed;
2258 void zbd_log_err(const struct thread_data *td, const struct io_u *io_u)
2260 const struct fio_file *f = io_u->file;
2262 if (td->o.zone_mode != ZONE_MODE_ZBD)
2265 if (io_u->error == EOVERFLOW)
2266 log_err("%s: Exceeded max_active_zones limit. Check conditions of zones out of I/O ranges.\n",