If thread bumps into "max_open_zones" limit, it tries to close/reopen some
other zone before issuing IO. This scan is done over full list of block device's
opened zones. It means that a zone which doesn't belong to thread's working
area can be altered or IO can be retargeted at such zone.
If IO is retargeted then it will be dropped by "is_valid_offset()" check.
What happens with null block device testing is that one thread monopolises
IO and others threads do basically nothing.
This config will reliably succeed now:
[global]
zonemode=zbd
zonesize=1M
rw=randwrite
...
thread
numjobs=2
offset_increment=128M
[j]
max_open_zones=2
size=2M
Starting 2 threads
zbd 7991 /dev/nullb0: zbd model string: host-managed
zbd 7991 Device /dev/nullb0 has 1024 zones of size 1024 KB
zbd 8009 /dev/nullb0: examining zones 0 .. 2
zbd 8010 /dev/nullb0: examining zones 128 .. 130
zbd 8009 /dev/nullb0: opening zone 0
zbd 8010 /dev/nullb0: opening zone 128
zbd 8009 /dev/nullb0: queued I/O (0, 4096) for zone 0
zbd 8009 zbd_convert_to_open_zone(/dev/nullb0): starting from zone 128 (offset 1552384, buflen 4096)
retargeted for other thread's zone (zone 0 => zone 128)
zbd 8010 /dev/nullb0: queued I/O (134217728, 4096) for zone 128
zbd 8009 zbd_convert_to_open_zone(/dev/nullb0): returning zone 128
zbd 8009 Dropped request with offset 134221824
and dropped
Note: quasi-randomness is kind of necessary to spread I/O. Imagine index 0
is picked all the time, zone living there will be reopened constantly and
get relatively little I/O.
Signed-off-by: Alexey Dobriyan (SK hynix) <adobriyan@xxxxxxxxx>
Reviewed-by: Damien Le Moal <damien.lemoal@xxxxxxx>
Tested-by: Damien Le Moal <damien.lemoal@xxxxxxx>
---
zbd.c | 43 ++++++++++++++++++++++++++++++++++++++-----
1 file changed, 38 insertions(+), 5 deletions(-)
--- a/zbd.c
+++ b/zbd.c
@@ -945,6 +945,13 @@ static void zone_lock(struct thread_data *td, struct fio_zone_info *z)
}
}
+/* Anything goes as long as it is not a constant. */
+static uint32_t pick_random_zone_idx(const struct fio_file *f,
+ const struct io_u *io_u)
+{
+ return io_u->offset * f->zbd_info->num_open_zones / f->real_file_size;
+}
+
/*
* Modify the offset of an I/O unit that does not refer to an open zone such
* that it refers to an open zone. Close an open zone and open a new zone if
@@ -969,9 +976,7 @@ static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
* This statement accesses f->zbd_info->open_zones[] on purpose
* without locking.
*/
- zone_idx = f->zbd_info->open_zones[(io_u->offset -
- f->file_offset) *
- f->zbd_info->num_open_zones / f->io_size];
+ zone_idx = f->zbd_info->open_zones[pick_random_zone_idx(f, io_u)];
} else {
zone_idx = zbd_zone_idx(f, io_u->offset);
}
@@ -985,6 +990,8 @@ static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
* has been obtained. Hence the loop.
*/
for (;;) {
+ uint32_t tmp_idx;
+
z = &f->zbd_info->zone_info[zone_idx];
zone_lock(td, z);
@@ -998,9 +1005,35 @@ static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td,
__func__, f->file_name);
return NULL;
}
- open_zone_idx = (io_u->offset - f->file_offset) *
- f->zbd_info->num_open_zones / f->io_size;
+
+ /*
+ * List of opened zones is per-device, shared across all threads.
+ * Start with quasi-random candidate zone.
+ * Ignore zones which don't belong to thread's offset/size area.
+ */
+ open_zone_idx = pick_random_zone_idx(f, io_u);
assert(open_zone_idx < f->zbd_info->num_open_zones);
+ tmp_idx = open_zone_idx;
+ for (i = 0; i < f->zbd_info->num_open_zones; i++) {
+ uint32_t tmpz;
+
+ if (tmp_idx >= f->zbd_info->num_open_zones)
+ tmp_idx = 0;
+ tmpz = f->zbd_info->open_zones[tmp_idx];
+
+ if (is_valid_offset(f, f->zbd_info->zone_info[tmpz].start)) {
+ open_zone_idx = tmp_idx;
+ goto found_candidate_zone;
+ }
+
+ tmp_idx++;
+ }
+
+ dprint(FD_ZBD, "%s(%s): no candidate zone\n",
+ __func__, f->file_name);
+ return NULL;
+
+found_candidate_zone:
new_zone_idx = f->zbd_info->open_zones[open_zone_idx];
if (new_zone_idx == zone_idx)
break;
