On Feb 07, 2023 / 14:06, Niklas Cassel wrote:
> On Tue, Feb 07, 2023 at 03:37:37PM +0900, Shin'ichiro Kawasaki wrote:
> > The valid data bytes accounting is used for zone_reset_threshold option.
> > This accounting usage has two issues. The first issue is unexpected
> > zone reset due to different IO ranges. The valid data bytes accounting
> > is done for all IO ranges per device, and shared by all jobs. On the
> > other hand, the zone_reset_threshold option is defined as the ratio to
> > each job's IO range. When a job refers to the accounting value, it
> > includes writes to IO ranges out of the job's IO range. Then zone reset
> > is triggered earlier than expected.
> >
> > The second issue is accounting value initialization. The initialization
> > of the accounting field is repeated for each job, then the value
> > initialized by the first job is overwritten by other jobs. This works as
> > expected for single job or multiple jobs with same write range. However,
> > when multiple jobs have different write ranges, the overwritten value is
> > wrong except for the last job.
> >
> > To ensure that the accounting works as expected for the option, check
> > that write ranges of all jobs are same. If jobs have different write
> > ranges, report it as an error. Initialize the accounting field only once
> > for the first job. All jobs have same write range, then one time
> > initialization is enough. Update man page to clarify this limitation of
> > the option.
> >
> > Signed-off-by: Shin'ichiro Kawasaki <shinichiro.kawasaki@xxxxxxx>
> > ---
> > HOWTO.rst | 3 ++-
> > fio.1 | 3 ++-
> > zbd.c | 22 +++++++++++++++++++---
> > zbd.h | 4 ++++
> > 4 files changed, 27 insertions(+), 5 deletions(-)
> >
> > diff --git a/HOWTO.rst b/HOWTO.rst
> > index d08f8a18..c71c3949 100644
> > --- a/HOWTO.rst
> > +++ b/HOWTO.rst
> > @@ -1088,7 +1088,8 @@ Target file/device
> > A number between zero and one that indicates the ratio of written bytes
> > to the total size of the zones with write pointers in the IO range. When
> > current ratio is above this ratio, zones are reset periodically as
> > - :option:`zone_reset_frequency` specifies.
> > + :option:`zone_reset_frequency` specifies. If there are multiple jobs when
> > + using this option, the IO range for all write jobs shall be the same.
>
> nit: perhaps replace "shall be the same" with "has to be the same"
Will replace. I find "has to" is more consitent in the document than "shall".
>
> >
> > .. option:: zone_reset_frequency=float
> >
> > diff --git a/fio.1 b/fio.1
> > index 54d2c403..80cd23b5 100644
> > --- a/fio.1
> > +++ b/fio.1
> > @@ -857,7 +857,8 @@ value to be larger than the device reported limit. Default: false.
> > A number between zero and one that indicates the ratio of written bytes to the
> > total size of the zones with write pointers in the IO range. When current ratio
> > is above this ratio, zones are reset periodically as \fBzone_reset_frequency\fR
> > -specifies.
> > +specifies. If there are multiple jobs when using this option, the IO range for
> > +all write jobs shall be the same.
> > .TP
> > .BI zone_reset_frequency \fR=\fPfloat
> > A number between zero and one that indicates how often a zone reset should be
> > diff --git a/zbd.c b/zbd.c
> > index 6783acf9..ca6816b9 100644
> > --- a/zbd.c
> > +++ b/zbd.c
> > @@ -1201,10 +1201,26 @@ static uint64_t zbd_set_vdb(struct thread_data *td, const struct fio_file *f)
> > {
> > struct fio_zone_info *zb, *ze, *z;
> > uint64_t wp_vdb = 0;
> > + struct zoned_block_device_info *zbdi = f->zbd_info;
> >
> > if (!accounting_vdb(td, f))
> > return 0;
> >
> > + if (zbdi->wp_write_min_zone != zbdi->wp_write_max_zone) {
> > + if (zbdi->wp_write_min_zone != f->min_zone ||
> > + zbdi->wp_write_max_zone != f->max_zone) {
> > + td_verror(td, EINVAL,
> > + "multi-jobs with different write ranges are "
> > + "not supported with zone_reset_threshold");
> > + log_err("multi-jobs with different write ranges are "
> > + "not supported with zone_reset_threshold\n");
> > + }
> > + return 0;
> > + }
>
> This check looks a bit weird.
> What if the first job specifies a f->file_offset and an f->io_size that is
> within the same zone? Then f->min_zone and f->max_zone will have the same value.
>
> Which should cause wp_write_min_zone and wp_write_max_zone to have the same
> value.
>
> The check will thus not be performed for the second job, after wp_write_min_zone
> and wp_write_min_zone have been initialized.
Thanks for catching this. In most cases, f->min_zone and f->max_zone have
different values, since f->io_size must be equal to or larger than zone size.
However, still f->min_zone and f->max_zone can be same when the IO range fits
within a conventional zone. The commit f952800a0b1a is a reference.
I suggest following 3 more changes below:
1) Before IO range check in zbd_set_vdb(), ensure that the IO range includes
zones with write pointer. This also ensures that f->min_zone and f->max_zone
are different. The code diff will be as follows:
diff --git a/zbd.c b/zbd.c
index a464d6df..f0a0856b 100644
--- a/zbd.c
+++ b/zbd.c
@@ -1086,6 +1086,13 @@ static uint64_t zbd_set_vdb(struct thread_data *td, const struct fio_file *f)
if (!accounting_vdb(td, f))
return 0;
+ /*
+ * Ensure that all zones in the I/O range are not conventional so that
+ * f->min_zone and f->max_zone have different values.
+ */
+ if (!zbd_is_seq_job(f))
+ return 0;
+
if (zbdi->wp_write_min_zone != zbdi->wp_write_max_zone) {
if (zbdi->wp_write_min_zone != f->min_zone ||
zbdi->wp_write_max_zone != f->max_zone) {
2) The names of new fields wp_write_min_zone and wp_write_max_zone are not
accurate, since the IO range may cover both coventional zones and zones
with write pointers. I will rename them to write_min_zone and write_max_zone
without the wp_ prefix.
3) The 4th patch needs modification also. The description of the
zone_reset_threshold option should cover the case that IO range has both
conventional zones and zones with write pointers. It will be like this:
.BI zone_reset_threshold \fR=\fPfloat
A number between zero and one that indicates the ratio of written bytes in the
zones with write pointers in the IO range to the size of the IO range.
--
Shin'ichiro Kawasaki
