2 * Copyright (C) 2018 Western Digital Corporation or its affiliates.
4 * This file is released under the GPL.
12 #include <sys/ioctl.h>
15 #include <linux/blkzoned.h>
25 * zbd_zone_idx - convert an offset into a zone number
27 * @offset: offset in bytes. If this offset is in the first zone_size bytes
28 * past the disk size then the index of the sentinel is returned.
30 static uint32_t zbd_zone_idx(const struct fio_file *f, uint64_t offset)
34 if (f->zbd_info->zone_size_log2 > 0)
35 zone_idx = offset >> f->zbd_info->zone_size_log2;
37 zone_idx = offset / f->zbd_info->zone_size;
39 return min(zone_idx, f->zbd_info->nr_zones);
43 * zbd_zone_full - verify whether a minimum number of bytes remain in a zone
45 * @z: zone info pointer.
46 * @required: minimum number of bytes that must remain in a zone.
48 * The caller must hold z->mutex.
50 static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
53 assert((required & 511) == 0);
55 return z->type == BLK_ZONE_TYPE_SEQWRITE_REQ &&
56 z->wp + required > z->start + f->zbd_info->zone_size;
59 static bool is_valid_offset(const struct fio_file *f, uint64_t offset)
61 return (uint64_t)(offset - f->file_offset) < f->io_size;
64 /* Verify whether direct I/O is used for all host-managed zoned drives. */
65 static bool zbd_using_direct_io(void)
67 struct thread_data *td;
72 if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE))
74 for_each_file(td, f, j) {
76 f->zbd_info->model == ZBD_DM_HOST_MANAGED)
84 /* Whether or not the I/O range for f includes one or more sequential zones */
85 static bool zbd_is_seq_job(struct fio_file *f)
87 uint32_t zone_idx, zone_idx_b, zone_idx_e;
92 zone_idx_b = zbd_zone_idx(f, f->file_offset);
93 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size - 1);
94 for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++)
95 if (f->zbd_info->zone_info[zone_idx].type ==
96 BLK_ZONE_TYPE_SEQWRITE_REQ)
103 * Verify whether offset and size parameters are aligned with zone boundaries.
105 static bool zbd_verify_sizes(void)
107 const struct fio_zone_info *z;
108 struct thread_data *td;
110 uint64_t new_offset, new_end;
115 for_each_file(td, f, j) {
118 if (f->file_offset >= f->real_file_size)
120 if (!zbd_is_seq_job(f))
122 zone_idx = zbd_zone_idx(f, f->file_offset);
123 z = &f->zbd_info->zone_info[zone_idx];
124 if (f->file_offset != z->start) {
125 new_offset = (z+1)->start;
126 if (new_offset >= f->file_offset + f->io_size) {
127 log_info("%s: io_size must be at least one zone\n",
131 log_info("%s: rounded up offset from %lu to %lu\n",
132 f->file_name, f->file_offset,
134 f->io_size -= (new_offset - f->file_offset);
135 f->file_offset = new_offset;
137 zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size);
138 z = &f->zbd_info->zone_info[zone_idx];
140 if (f->file_offset + f->io_size != new_end) {
141 if (new_end <= f->file_offset) {
142 log_info("%s: io_size must be at least one zone\n",
146 log_info("%s: rounded down io_size from %lu to %lu\n",
147 f->file_name, f->io_size,
148 new_end - f->file_offset);
149 f->io_size = new_end - f->file_offset;
157 static bool zbd_verify_bs(void)
159 struct thread_data *td;
165 for_each_file(td, f, j) {
168 zone_size = f->zbd_info->zone_size;
169 for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) {
170 if (td->o.verify != VERIFY_NONE &&
171 zone_size % td->o.bs[k] != 0) {
172 log_info("%s: block size %llu is not a divisor of the zone size %d\n",
173 f->file_name, td->o.bs[k],
184 * Read zone information into @buf starting from sector @start_sector.
185 * @fd is a file descriptor that refers to a block device and @bufsz is the
188 * Returns 0 upon success and a negative error code upon failure.
190 static int read_zone_info(int fd, uint64_t start_sector,
191 void *buf, unsigned int bufsz)
193 struct blk_zone_report *hdr = buf;
195 if (bufsz < sizeof(*hdr))
198 memset(hdr, 0, sizeof(*hdr));
200 hdr->nr_zones = (bufsz - sizeof(*hdr)) / sizeof(struct blk_zone);
201 hdr->sector = start_sector;
202 return ioctl(fd, BLKREPORTZONE, hdr) >= 0 ? 0 : -errno;
206 * Read up to 255 characters from the first line of a file. Strip the trailing
209 static char *read_file(const char *path)
211 char line[256], *p = line;
214 f = fopen(path, "rb");
217 if (!fgets(line, sizeof(line), f))
225 static enum blk_zoned_model get_zbd_model(const char *file_name)
227 enum blk_zoned_model model = ZBD_DM_NONE;
228 char *zoned_attr_path = NULL;
229 char *model_str = NULL;
232 if (stat(file_name, &statbuf) < 0)
234 if (asprintf(&zoned_attr_path, "/sys/dev/block/%d:%d/queue/zoned",
235 major(statbuf.st_rdev), minor(statbuf.st_rdev)) < 0)
237 model_str = read_file(zoned_attr_path);
240 dprint(FD_ZBD, "%s: zbd model string: %s\n", file_name, model_str);
241 if (strcmp(model_str, "host-aware") == 0)
242 model = ZBD_DM_HOST_AWARE;
243 else if (strcmp(model_str, "host-managed") == 0)
244 model = ZBD_DM_HOST_MANAGED;
248 free(zoned_attr_path);
252 static int ilog2(uint64_t i)
264 * Initialize f->zbd_info for devices that are not zoned block devices. This
265 * allows to execute a ZBD workload against a non-ZBD device.
267 static int init_zone_info(struct thread_data *td, struct fio_file *f)
270 struct fio_zone_info *p;
272 struct zoned_block_device_info *zbd_info = NULL;
273 pthread_mutexattr_t attr;
276 zone_size = td->o.zone_size;
278 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
279 zbd_info = scalloc(1, sizeof(*zbd_info) +
280 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
284 pthread_mutexattr_init(&attr);
285 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
286 pthread_mutexattr_setpshared(&attr, true);
287 pthread_mutex_init(&zbd_info->mutex, &attr);
288 zbd_info->refcount = 1;
289 p = &zbd_info->zone_info[0];
290 for (i = 0; i < nr_zones; i++, p++) {
291 pthread_mutex_init(&p->mutex, &attr);
292 p->start = i * zone_size;
293 p->wp = p->start + zone_size;
294 p->type = BLK_ZONE_TYPE_SEQWRITE_REQ;
295 p->cond = BLK_ZONE_COND_EMPTY;
298 p->start = nr_zones * zone_size;
300 f->zbd_info = zbd_info;
301 f->zbd_info->zone_size = zone_size;
302 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
303 ilog2(zone_size) : -1;
304 f->zbd_info->nr_zones = nr_zones;
305 pthread_mutexattr_destroy(&attr);
310 * Parse the BLKREPORTZONE output and store it in f->zbd_info. Must be called
311 * only for devices that support this ioctl, namely zoned block devices.
313 static int parse_zone_info(struct thread_data *td, struct fio_file *f)
315 const unsigned int bufsz = sizeof(struct blk_zone_report) +
316 4096 * sizeof(struct blk_zone);
318 struct blk_zone_report *hdr;
319 const struct blk_zone *z;
320 struct fio_zone_info *p;
321 uint64_t zone_size, start_sector;
322 struct zoned_block_device_info *zbd_info = NULL;
323 pthread_mutexattr_t attr;
325 int fd, i, j, ret = 0;
327 pthread_mutexattr_init(&attr);
328 pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
329 pthread_mutexattr_setpshared(&attr, true);
335 fd = open(f->file_name, O_RDONLY | O_LARGEFILE);
341 ret = read_zone_info(fd, 0, buf, bufsz);
343 log_info("fio: BLKREPORTZONE(%lu) failed for %s (%d).\n",
344 0UL, f->file_name, -ret);
348 if (hdr->nr_zones < 1) {
349 log_info("fio: %s has invalid zone information.\n",
353 z = (void *)(hdr + 1);
354 zone_size = z->len << 9;
355 nr_zones = (f->real_file_size + zone_size - 1) / zone_size;
357 if (td->o.zone_size == 0) {
358 td->o.zone_size = zone_size;
359 } else if (td->o.zone_size != zone_size) {
360 log_info("fio: %s job parameter zonesize %lld does not match disk zone size %ld.\n",
361 f->file_name, td->o.zone_size, zone_size);
366 dprint(FD_ZBD, "Device %s has %d zones of size %lu KB\n", f->file_name,
367 nr_zones, zone_size / 1024);
369 zbd_info = scalloc(1, sizeof(*zbd_info) +
370 (nr_zones + 1) * sizeof(zbd_info->zone_info[0]));
374 pthread_mutex_init(&zbd_info->mutex, &attr);
375 zbd_info->refcount = 1;
376 p = &zbd_info->zone_info[0];
377 for (start_sector = 0, j = 0; j < nr_zones;) {
378 z = (void *)(hdr + 1);
379 for (i = 0; i < hdr->nr_zones; i++, j++, z++, p++) {
380 pthread_mutex_init(&p->mutex, &attr);
381 p->start = z->start << 9;
383 case BLK_ZONE_COND_NOT_WP:
386 case BLK_ZONE_COND_FULL:
387 p->wp = p->start + zone_size;
390 assert(z->start <= z->wp);
391 assert(z->wp <= z->start + (zone_size >> 9));
397 if (j > 0 && p->start != p[-1].start + zone_size) {
398 log_info("%s: invalid zone data\n",
405 start_sector = z->start + z->len;
408 ret = read_zone_info(fd, start_sector, buf, bufsz);
410 log_info("fio: BLKREPORTZONE(%lu) failed for %s (%d).\n",
411 start_sector, f->file_name, -ret);
416 zbd_info->zone_info[nr_zones].start = start_sector << 9;
418 f->zbd_info = zbd_info;
419 f->zbd_info->zone_size = zone_size;
420 f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ?
421 ilog2(zone_size) : -1;
422 f->zbd_info->nr_zones = nr_zones;
432 pthread_mutexattr_destroy(&attr);
437 * Allocate zone information and store it into f->zbd_info if zonemode=zbd.
439 * Returns 0 upon success and a negative error code upon failure.
441 int zbd_create_zone_info(struct thread_data *td, struct fio_file *f)
443 enum blk_zoned_model zbd_model;
446 assert(td->o.zone_mode == ZONE_MODE_ZBD);
448 zbd_model = get_zbd_model(f->file_name);
450 case ZBD_DM_HOST_AWARE:
451 case ZBD_DM_HOST_MANAGED:
452 ret = parse_zone_info(td, f);
455 ret = init_zone_info(td, f);
459 f->zbd_info->model = zbd_model;
463 void zbd_free_zone_info(struct fio_file *f)
470 pthread_mutex_lock(&f->zbd_info->mutex);
471 refcount = --f->zbd_info->refcount;
472 pthread_mutex_unlock(&f->zbd_info->mutex);
474 assert((int32_t)refcount >= 0);
481 * Initialize f->zbd_info.
483 * Returns 0 upon success and a negative error code upon failure.
485 * Note: this function can only work correctly if it is called before the first
488 static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file)
490 struct thread_data *td2;
494 for_each_td(td2, i) {
495 for_each_file(td2, f2, j) {
496 if (td2 == td && f2 == file)
499 strcmp(f2->file_name, file->file_name) != 0)
501 file->zbd_info = f2->zbd_info;
502 file->zbd_info->refcount++;
507 ret = zbd_create_zone_info(td, file);
509 td_verror(td, -ret, "BLKREPORTZONE failed");
513 int zbd_init(struct thread_data *td)
518 for_each_file(td, f, i) {
519 if (f->filetype != FIO_TYPE_BLOCK)
521 if (td->o.zone_size && td->o.zone_size < 512) {
522 log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n",
526 if (td->o.zone_size == 0 &&
527 get_zbd_model(f->file_name) == ZBD_DM_NONE) {
528 log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n",
532 zbd_init_zone_info(td, f);
535 if (!zbd_using_direct_io()) {
536 log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n");
540 if (!zbd_verify_sizes())
543 if (!zbd_verify_bs())
550 * zbd_reset_range - reset zones for a range of sectors
551 * @td: FIO thread data.
552 * @f: Fio file for which to reset zones
553 * @sector: Starting sector in units of 512 bytes
554 * @nr_sectors: Number of sectors in units of 512 bytes
556 * Returns 0 upon success and a negative error code upon failure.
558 static int zbd_reset_range(struct thread_data *td, const struct fio_file *f,
559 uint64_t offset, uint64_t length)
561 struct blk_zone_range zr = {
562 .sector = offset >> 9,
563 .nr_sectors = length >> 9,
565 uint32_t zone_idx_b, zone_idx_e;
566 struct fio_zone_info *zb, *ze, *z;
570 assert(is_valid_offset(f, offset + length - 1));
571 switch (f->zbd_info->model) {
572 case ZBD_DM_HOST_AWARE:
573 case ZBD_DM_HOST_MANAGED:
574 ret = ioctl(f->fd, BLKRESETZONE, &zr);
576 td_verror(td, errno, "resetting wp failed");
577 log_err("%s: resetting wp for %llu sectors at sector %llu failed (%d).\n",
578 f->file_name, zr.nr_sectors, zr.sector, errno);
586 zone_idx_b = zbd_zone_idx(f, offset);
587 zb = &f->zbd_info->zone_info[zone_idx_b];
588 zone_idx_e = zbd_zone_idx(f, offset + length);
589 ze = &f->zbd_info->zone_info[zone_idx_e];
590 for (z = zb; z < ze; z++) {
591 pthread_mutex_lock(&z->mutex);
592 pthread_mutex_lock(&f->zbd_info->mutex);
593 f->zbd_info->sectors_with_data -= z->wp - z->start;
594 pthread_mutex_unlock(&f->zbd_info->mutex);
597 pthread_mutex_unlock(&z->mutex);
600 td->ts.nr_zone_resets += ze - zb;
606 * zbd_reset_zone - reset the write pointer of a single zone
607 * @td: FIO thread data.
608 * @f: FIO file associated with the disk for which to reset a write pointer.
611 * Returns 0 upon success and a negative error code upon failure.
613 static int zbd_reset_zone(struct thread_data *td, const struct fio_file *f,
614 struct fio_zone_info *z)
618 dprint(FD_ZBD, "%s: resetting wp of zone %lu.\n", f->file_name,
619 z - f->zbd_info->zone_info);
620 ret = zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
625 * Reset a range of zones. Returns 0 upon success and 1 upon failure.
626 * @td: fio thread data.
627 * @f: fio file for which to reset zones
628 * @zb: first zone to reset.
629 * @ze: first zone not to reset.
630 * @all_zones: whether to reset all zones or only those zones for which the
631 * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE].
633 static int zbd_reset_zones(struct thread_data *td, struct fio_file *f,
634 struct fio_zone_info *const zb,
635 struct fio_zone_info *const ze, bool all_zones)
637 struct fio_zone_info *z, *start_z = ze;
638 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
642 dprint(FD_ZBD, "%s: examining zones %lu .. %lu\n", f->file_name,
643 zb - f->zbd_info->zone_info, ze - f->zbd_info->zone_info);
645 for (z = zb; z < ze; z++) {
646 pthread_mutex_lock(&z->mutex);
648 case BLK_ZONE_TYPE_SEQWRITE_REQ:
649 reset_wp = all_zones ? z->wp != z->start :
650 (td->o.td_ddir & TD_DDIR_WRITE) &&
652 if (start_z == ze && reset_wp) {
654 } else if (start_z < ze && !reset_wp) {
656 "%s: resetting zones %lu .. %lu\n",
658 start_z - f->zbd_info->zone_info,
659 z - f->zbd_info->zone_info);
660 if (zbd_reset_range(td, f, start_z->start,
661 z->start - start_z->start) < 0)
669 dprint(FD_ZBD, "%s: resetting zones %lu .. %lu\n",
670 f->file_name, start_z - f->zbd_info->zone_info,
671 z - f->zbd_info->zone_info);
672 if (zbd_reset_range(td, f, start_z->start,
673 z->start - start_z->start) < 0)
680 dprint(FD_ZBD, "%s: resetting zones %lu .. %lu\n", f->file_name,
681 start_z - f->zbd_info->zone_info,
682 z - f->zbd_info->zone_info);
683 if (zbd_reset_range(td, f, start_z->start,
684 z->start - start_z->start) < 0)
687 for (z = zb; z < ze; z++)
688 pthread_mutex_unlock(&z->mutex);
694 * Reset zbd_info.write_cnt, the counter that counts down towards the next
697 static void zbd_reset_write_cnt(const struct thread_data *td,
698 const struct fio_file *f)
700 assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1);
702 pthread_mutex_lock(&f->zbd_info->mutex);
703 f->zbd_info->write_cnt = td->o.zrf.u.f ?
704 min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX;
705 pthread_mutex_unlock(&f->zbd_info->mutex);
708 static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td,
709 const struct fio_file *f)
711 uint32_t write_cnt = 0;
713 pthread_mutex_lock(&f->zbd_info->mutex);
714 assert(f->zbd_info->write_cnt);
715 if (f->zbd_info->write_cnt)
716 write_cnt = --f->zbd_info->write_cnt;
718 zbd_reset_write_cnt(td, f);
719 pthread_mutex_unlock(&f->zbd_info->mutex);
721 return write_cnt == 0;
724 /* Check whether the value of zbd_info.sectors_with_data is correct. */
725 static void check_swd(const struct thread_data *td, const struct fio_file *f)
728 struct fio_zone_info *zb, *ze, *z;
731 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
732 ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset +
735 for (z = zb; z < ze; z++) {
736 pthread_mutex_lock(&z->mutex);
737 swd += z->wp - z->start;
739 pthread_mutex_lock(&f->zbd_info->mutex);
740 assert(f->zbd_info->sectors_with_data == swd);
741 pthread_mutex_unlock(&f->zbd_info->mutex);
742 for (z = zb; z < ze; z++)
743 pthread_mutex_unlock(&z->mutex);
747 void zbd_file_reset(struct thread_data *td, struct fio_file *f)
749 struct fio_zone_info *zb, *ze, *z;
756 zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
757 zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size);
758 ze = &f->zbd_info->zone_info[zone_idx_e];
759 for (z = zb ; z < ze; z++) {
760 pthread_mutex_lock(&z->mutex);
761 swd += z->wp - z->start;
763 pthread_mutex_lock(&f->zbd_info->mutex);
764 f->zbd_info->sectors_with_data = swd;
765 pthread_mutex_unlock(&f->zbd_info->mutex);
766 for (z = zb ; z < ze; z++)
767 pthread_mutex_unlock(&z->mutex);
768 dprint(FD_ZBD, "%s(%s): swd = %ld\n", __func__, f->file_name, swd);
770 * If data verification is enabled reset the affected zones before
771 * writing any data to avoid that a zone reset has to be issued while
772 * writing data, which causes data loss.
774 zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE &&
775 (td->o.td_ddir & TD_DDIR_WRITE) &&
776 td->runstate != TD_VERIFYING);
777 zbd_reset_write_cnt(td, f);
780 /* The caller must hold f->zbd_info->mutex. */
781 static bool is_zone_open(const struct thread_data *td, const struct fio_file *f,
782 unsigned int zone_idx)
784 struct zoned_block_device_info *zbdi = f->zbd_info;
787 assert(td->o.max_open_zones <= ARRAY_SIZE(zbdi->open_zones));
788 assert(zbdi->num_open_zones <= td->o.max_open_zones);
790 for (i = 0; i < zbdi->num_open_zones; i++)
791 if (zbdi->open_zones[i] == zone_idx)
798 * Open a ZBD zone if it was not yet open. Returns true if either the zone was
799 * already open or if opening a new zone is allowed. Returns false if the zone
800 * was not yet open and opening a new zone would cause the zone limit to be
803 static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u,
806 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
807 const struct fio_file *f = io_u->file;
808 struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx];
811 if (z->cond == BLK_ZONE_COND_OFFLINE)
815 * Skip full zones with data verification enabled because resetting a
816 * zone causes data loss and hence causes verification to fail.
818 if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs))
821 /* Zero means no limit */
822 if (!td->o.max_open_zones)
825 pthread_mutex_lock(&f->zbd_info->mutex);
826 if (is_zone_open(td, f, zone_idx))
829 if (f->zbd_info->num_open_zones >= td->o.max_open_zones)
831 dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx);
832 f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx;
837 pthread_mutex_unlock(&f->zbd_info->mutex);
841 /* The caller must hold f->zbd_info->mutex */
842 static void zbd_close_zone(struct thread_data *td, const struct fio_file *f,
843 unsigned int open_zone_idx)
847 assert(open_zone_idx < f->zbd_info->num_open_zones);
848 zone_idx = f->zbd_info->open_zones[open_zone_idx];
849 memmove(f->zbd_info->open_zones + open_zone_idx,
850 f->zbd_info->open_zones + open_zone_idx + 1,
851 (FIO_MAX_OPEN_ZBD_ZONES - (open_zone_idx + 1)) *
852 sizeof(f->zbd_info->open_zones[0]));
853 f->zbd_info->num_open_zones--;
854 f->zbd_info->zone_info[zone_idx].open = 0;
858 * Modify the offset of an I/O unit that does not refer to an open zone such
859 * that it refers to an open zone. Close an open zone and open a new zone if
860 * necessary. This algorithm can only work correctly if all write pointers are
861 * a multiple of the fio block size. The caller must neither hold z->mutex
862 * nor f->zbd_info->mutex. Returns with z->mutex held upon success.
864 struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
867 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
868 const struct fio_file *f = io_u->file;
869 struct fio_zone_info *z;
870 unsigned int open_zone_idx = -1;
871 uint32_t zone_idx, new_zone_idx;
874 assert(is_valid_offset(f, io_u->offset));
876 if (td->o.max_open_zones) {
878 * This statement accesses f->zbd_info->open_zones[] on purpose
881 zone_idx = f->zbd_info->open_zones[(io_u->offset -
883 f->zbd_info->num_open_zones / f->io_size];
885 zone_idx = zbd_zone_idx(f, io_u->offset);
887 dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n",
888 __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen);
891 * Since z->mutex is the outer lock and f->zbd_info->mutex the inner
892 * lock it can happen that the state of the zone with index zone_idx
893 * has changed after 'z' has been assigned and before f->zbd_info->mutex
894 * has been obtained. Hence the loop.
897 z = &f->zbd_info->zone_info[zone_idx];
899 pthread_mutex_lock(&z->mutex);
900 pthread_mutex_lock(&f->zbd_info->mutex);
901 if (td->o.max_open_zones == 0)
903 if (f->zbd_info->num_open_zones == 0) {
904 pthread_mutex_unlock(&f->zbd_info->mutex);
905 pthread_mutex_unlock(&z->mutex);
906 dprint(FD_ZBD, "%s(%s): no zones are open\n",
907 __func__, f->file_name);
910 open_zone_idx = (io_u->offset - f->file_offset) *
911 f->zbd_info->num_open_zones / f->io_size;
912 assert(open_zone_idx < f->zbd_info->num_open_zones);
913 new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
914 if (new_zone_idx == zone_idx)
916 zone_idx = new_zone_idx;
917 pthread_mutex_unlock(&f->zbd_info->mutex);
918 pthread_mutex_unlock(&z->mutex);
921 /* Both z->mutex and f->zbd_info->mutex are held. */
924 if (z->wp + min_bs <= (z+1)->start) {
925 pthread_mutex_unlock(&f->zbd_info->mutex);
928 dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name,
930 if (td->o.max_open_zones)
931 zbd_close_zone(td, f, open_zone_idx);
932 pthread_mutex_unlock(&f->zbd_info->mutex);
934 /* Only z->mutex is held. */
936 /* Zone 'z' is full, so try to open a new zone. */
937 for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) {
939 pthread_mutex_unlock(&z->mutex);
941 if (!is_valid_offset(f, z->start)) {
943 zone_idx = zbd_zone_idx(f, f->file_offset);
944 z = &f->zbd_info->zone_info[zone_idx];
946 assert(is_valid_offset(f, z->start));
947 pthread_mutex_lock(&z->mutex);
950 if (zbd_open_zone(td, io_u, zone_idx))
954 /* Only z->mutex is held. */
956 /* Check whether the write fits in any of the already opened zones. */
957 pthread_mutex_lock(&f->zbd_info->mutex);
958 for (i = 0; i < f->zbd_info->num_open_zones; i++) {
959 zone_idx = f->zbd_info->open_zones[i];
960 pthread_mutex_unlock(&f->zbd_info->mutex);
961 pthread_mutex_unlock(&z->mutex);
963 z = &f->zbd_info->zone_info[zone_idx];
965 pthread_mutex_lock(&z->mutex);
966 if (z->wp + min_bs <= (z+1)->start)
968 pthread_mutex_lock(&f->zbd_info->mutex);
970 pthread_mutex_unlock(&f->zbd_info->mutex);
971 pthread_mutex_unlock(&z->mutex);
972 dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__,
977 dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name,
979 io_u->offset = z->start;
983 /* The caller must hold z->mutex. */
984 static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
986 struct fio_zone_info *z)
988 const struct fio_file *f = io_u->file;
989 const uint32_t min_bs = td->o.min_bs[DDIR_WRITE];
991 if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) {
992 pthread_mutex_unlock(&z->mutex);
993 z = zbd_convert_to_open_zone(td, io_u);
997 if (z->verify_block * min_bs >= f->zbd_info->zone_size)
998 log_err("%s: %d * %d >= %ld\n", f->file_name, z->verify_block,
999 min_bs, f->zbd_info->zone_size);
1000 io_u->offset = z->start + z->verify_block++ * min_bs;
1005 * Find another zone for which @io_u fits below the write pointer. Start
1006 * searching in zones @zb + 1 .. @zl and continue searching in zones
1009 * Either returns NULL or returns a zone pointer and holds the mutex for that
1012 static struct fio_zone_info *
1013 zbd_find_zone(struct thread_data *td, struct io_u *io_u,
1014 struct fio_zone_info *zb, struct fio_zone_info *zl)
1016 const uint32_t min_bs = td->o.min_bs[io_u->ddir];
1017 const struct fio_file *f = io_u->file;
1018 struct fio_zone_info *z1, *z2;
1019 const struct fio_zone_info *const zf =
1020 &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)];
1023 * Skip to the next non-empty zone in case of sequential I/O and to
1024 * the nearest non-empty zone in case of random I/O.
1026 for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) {
1027 if (z1 < zl && z1->cond != BLK_ZONE_COND_OFFLINE) {
1028 pthread_mutex_lock(&z1->mutex);
1029 if (z1->start + min_bs <= z1->wp)
1031 pthread_mutex_unlock(&z1->mutex);
1032 } else if (!td_random(td)) {
1035 if (td_random(td) && z2 >= zf &&
1036 z2->cond != BLK_ZONE_COND_OFFLINE) {
1037 pthread_mutex_lock(&z2->mutex);
1038 if (z2->start + min_bs <= z2->wp)
1040 pthread_mutex_unlock(&z2->mutex);
1043 dprint(FD_ZBD, "%s: adjusting random read offset failed\n",
1050 * zbd_post_submit - update the write pointer and unlock the zone lock
1052 * @success: Whether or not the I/O unit has been executed successfully
1054 * For write and trim operations, update the write pointer of all affected
1057 static void zbd_post_submit(const struct io_u *io_u, bool success)
1059 struct zoned_block_device_info *zbd_info;
1060 struct fio_zone_info *z;
1062 uint64_t end, zone_end;
1064 zbd_info = io_u->file->zbd_info;
1068 zone_idx = zbd_zone_idx(io_u->file, io_u->offset);
1069 end = io_u->offset + io_u->buflen;
1070 z = &zbd_info->zone_info[zone_idx];
1071 assert(zone_idx < zbd_info->nr_zones);
1072 if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ)
1076 switch (io_u->ddir) {
1078 zone_end = min(end, (z + 1)->start);
1079 pthread_mutex_lock(&zbd_info->mutex);
1081 * z->wp > zone_end means that one or more I/O errors
1084 if (z->wp <= zone_end)
1085 zbd_info->sectors_with_data += zone_end - z->wp;
1086 pthread_mutex_unlock(&zbd_info->mutex);
1090 assert(z->wp == z->start);
1096 pthread_mutex_unlock(&z->mutex);
1099 bool zbd_unaligned_write(int error_code)
1101 switch (error_code) {
1110 * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives
1111 * @td: FIO thread data.
1112 * @io_u: FIO I/O unit.
1114 * Locking strategy: returns with z->mutex locked if and only if z refers
1115 * to a sequential zone and if io_u_accept is returned. z is the zone that
1116 * corresponds to io_u->offset at the end of this function.
1118 enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
1120 const struct fio_file *f = io_u->file;
1121 uint32_t zone_idx_b;
1122 struct fio_zone_info *zb, *zl, *orig_zb;
1123 uint32_t orig_len = io_u->buflen;
1124 uint32_t min_bs = td->o.min_bs[io_u->ddir];
1131 assert(is_valid_offset(f, io_u->offset));
1132 assert(io_u->buflen);
1133 zone_idx_b = zbd_zone_idx(f, io_u->offset);
1134 zb = &f->zbd_info->zone_info[zone_idx_b];
1137 /* Accept the I/O offset for conventional zones. */
1138 if (zb->type == BLK_ZONE_TYPE_CONVENTIONAL)
1142 * Accept the I/O offset for reads if reading beyond the write pointer
1145 if (zb->cond != BLK_ZONE_COND_OFFLINE &&
1146 io_u->ddir == DDIR_READ && td->o.read_beyond_wp)
1149 pthread_mutex_lock(&zb->mutex);
1150 switch (io_u->ddir) {
1152 if (td->runstate == TD_VERIFYING) {
1153 zb = zbd_replay_write_order(td, io_u, zb);
1157 * Check that there is enough written data in the zone to do an
1158 * I/O of at least min_bs B. If there isn't, find a new zone for
1161 range = zb->cond != BLK_ZONE_COND_OFFLINE ?
1162 zb->wp - zb->start : 0;
1163 if (range < min_bs ||
1164 ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) {
1165 pthread_mutex_unlock(&zb->mutex);
1166 zl = &f->zbd_info->zone_info[zbd_zone_idx(f,
1167 f->file_offset + f->io_size)];
1168 zb = zbd_find_zone(td, io_u, zb, zl);
1171 "%s: zbd_find_zone(%lld, %llu) failed\n",
1172 f->file_name, io_u->offset,
1177 * zbd_find_zone() returned a zone with a range of at
1180 range = zb->wp - zb->start;
1181 assert(range >= min_bs);
1184 io_u->offset = zb->start;
1187 * Make sure the I/O is within the zone valid data range while
1188 * maximizing the I/O size and preserving randomness.
1190 if (range <= io_u->buflen)
1191 io_u->offset = zb->start;
1192 else if (td_random(td))
1193 io_u->offset = zb->start +
1194 ((io_u->offset - orig_zb->start) %
1195 (range - io_u->buflen)) / min_bs * min_bs;
1197 * Make sure the I/O does not cross over the zone wp position.
1199 new_len = min((unsigned long long)io_u->buflen,
1200 (unsigned long long)(zb->wp - io_u->offset));
1201 new_len = new_len / min_bs * min_bs;
1202 if (new_len < io_u->buflen) {
1203 io_u->buflen = new_len;
1204 dprint(FD_IO, "Changed length from %u into %llu\n",
1205 orig_len, io_u->buflen);
1207 assert(zb->start <= io_u->offset);
1208 assert(io_u->offset + io_u->buflen <= zb->wp);
1211 if (io_u->buflen > f->zbd_info->zone_size)
1213 if (!zbd_open_zone(td, io_u, zone_idx_b)) {
1214 pthread_mutex_unlock(&zb->mutex);
1215 zb = zbd_convert_to_open_zone(td, io_u);
1218 zone_idx_b = zb - f->zbd_info->zone_info;
1220 /* Check whether the zone reset threshold has been exceeded */
1221 if (td->o.zrf.u.f) {
1223 if (f->zbd_info->sectors_with_data >=
1224 f->io_size * td->o.zrt.u.f &&
1225 zbd_dec_and_reset_write_cnt(td, f)) {
1229 /* Reset the zone pointer if necessary */
1230 if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) {
1231 assert(td->o.verify == VERIFY_NONE);
1233 * Since previous write requests may have been submitted
1234 * asynchronously and since we will submit the zone
1235 * reset synchronously, wait until previously submitted
1236 * write requests have completed before issuing a
1241 if (zbd_reset_zone(td, f, zb) < 0)
1245 /* Make writes occur at the write pointer */
1246 assert(!zbd_zone_full(f, zb, min_bs));
1247 io_u->offset = zb->wp;
1248 if (!is_valid_offset(f, io_u->offset)) {
1249 dprint(FD_ZBD, "Dropped request with offset %llu\n",
1254 * Make sure that the buflen is a multiple of the minimal
1255 * block size. Give up if shrinking would make the request too
1258 new_len = min((unsigned long long)io_u->buflen,
1259 (zb + 1)->start - io_u->offset);
1260 new_len = new_len / min_bs * min_bs;
1261 if (new_len == io_u->buflen)
1263 if (new_len >= min_bs) {
1264 io_u->buflen = new_len;
1265 dprint(FD_IO, "Changed length from %u into %llu\n",
1266 orig_len, io_u->buflen);
1269 log_err("Zone remainder %lld smaller than minimum block size %d\n",
1270 ((zb + 1)->start - io_u->offset),
1277 case DDIR_SYNC_FILE_RANGE:
1288 assert(zb->cond != BLK_ZONE_COND_OFFLINE);
1289 assert(!io_u->post_submit);
1290 io_u->post_submit = zbd_post_submit;
1295 pthread_mutex_unlock(&zb->mutex);
1299 /* Return a string with ZBD statistics */
1300 char *zbd_write_status(const struct thread_stat *ts)
1304 if (asprintf(&res, "; %ld zone resets", ts->nr_zone_resets) < 0)