On Tue, Jun 07, 2022 at 10:49:42AM +0900, Damien Le Moal wrote:
> When a write command to a sequential write required or sequential write
> preferred zone result in the zone write pointer reaching the end of the
> zone, the zone condition must be set to full AND the number of
> implicitly or explicitly open zones updated to have a correct accounting
> for zone resources. However, the function zbc_inc_wp() only sets the
> zone condition to full without updating the open zone counters,
> resulting in a zone state machine breakage.
>
> Factor out the correct code from zbc_finish_zone() to transition a zone
> to the full condition and introduce the helper zbc_set_zone_full(). Use
> this helper in zbc_finish_zone() and zbc_inc_wp() to correctly
> transition zones to the full condition.
>
> Fixes: 0d1cf9378bd4 ("scsi: scsi_debug: Add ZBC zone commands")
> Signed-off-by: Damien Le Moal <damien.lemoal@xxxxxxxxxxxxxxxxxx>
> ---
> drivers/scsi/scsi_debug.c | 27 +++++++++++++++++----------
> 1 file changed, 17 insertions(+), 10 deletions(-)
>
> diff --git a/drivers/scsi/scsi_debug.c b/drivers/scsi/scsi_debug.c
> index 1f423f723d06..6c2bb02a42d8 100644
> --- a/drivers/scsi/scsi_debug.c
> +++ b/drivers/scsi/scsi_debug.c
> @@ -2826,6 +2826,19 @@ static void zbc_open_zone(struct sdebug_dev_info *devip,
> }
> }
>
> +static inline void zbc_set_zone_full(struct sdebug_dev_info *devip,
> + struct sdeb_zone_state *zsp)
> +{
> + enum sdebug_z_cond zc = zsp->z_cond;
> +
> + if (zc == ZC2_IMPLICIT_OPEN || zc == ZC3_EXPLICIT_OPEN)
> + zbc_close_zone(devip, zsp);
> + if (zsp->z_cond == ZC4_CLOSED)
> + devip->nr_closed--;
> + zsp->z_wp = zsp->z_start + zsp->z_size;
> + zsp->z_cond = ZC5_FULL;
> +}
> +
> static void zbc_inc_wp(struct sdebug_dev_info *devip,
> unsigned long long lba, unsigned int num)
> {
> @@ -2838,7 +2851,7 @@ static void zbc_inc_wp(struct sdebug_dev_info *devip,
> if (zsp->z_type == ZBC_ZTYPE_SWR) {
> zsp->z_wp += num;
> if (zsp->z_wp >= zend)
> - zsp->z_cond = ZC5_FULL;
> + zbc_set_zone_full(devip, zsp);
> return;
> }
>
> @@ -2857,7 +2870,7 @@ static void zbc_inc_wp(struct sdebug_dev_info *devip,
> n = num;
> }
> if (zsp->z_wp >= zend)
> - zsp->z_cond = ZC5_FULL;
> + zbc_set_zone_full(devip, zsp);
Hello Damien,
In the equivalent function (null_zone_write()) in null_blk,
we instead do this:
if (zone->wp == zone->start + zone->capacity) {
null_lock_zone_res(dev);
if (zone->cond == BLK_ZONE_COND_EXP_OPEN)
dev->nr_zones_exp_open--;
else if (zone->cond == BLK_ZONE_COND_IMP_OPEN)
dev->nr_zones_imp_open--;
zone->cond = BLK_ZONE_COND_FULL;
null_unlock_zone_res(dev);
}
Isn't it more clear to do the same here?
i.e. set the state to FULL, like before, and simply decrease the
imp/exp open counters.
zbc_set_zone_full() does some things that are not applicable in
the write path, specifically this:
> + if (zc == ZC2_IMPLICIT_OPEN || zc == ZC3_EXPLICIT_OPEN)
> + zbc_close_zone(devip, zsp);
> + if (zsp->z_cond == ZC4_CLOSED)
> + devip->nr_closed--;
e.g. with this new helper, if we are in e.g. IMP OPEN, we will now
set the zone state first to CLOSED, increase the nr_closed counter,
decrease the nr_closed counter, and then set the zone state to FULL.
Isn't it more clear to just set it to FULL directly, like before,
and simply add:
if (zone->cond == BLK_ZONE_COND_EXP_OPEN)
dev->nr_zones_exp_open--;
else if (zone->cond == BLK_ZONE_COND_IMP_OPEN)
dev->nr_zones_imp_open--;
Just like the equivalent code in null_blk.
>
> num -= n;
> lba += n;
> @@ -4731,14 +4744,8 @@ static void zbc_finish_zone(struct sdebug_dev_info *devip,
> enum sdebug_z_cond zc = zsp->z_cond;
>
> if (zc == ZC4_CLOSED || zc == ZC2_IMPLICIT_OPEN ||
> - zc == ZC3_EXPLICIT_OPEN || (empty && zc == ZC1_EMPTY)) {
> - if (zc == ZC2_IMPLICIT_OPEN || zc == ZC3_EXPLICIT_OPEN)
> - zbc_close_zone(devip, zsp);
> - if (zsp->z_cond == ZC4_CLOSED)
> - devip->nr_closed--;
> - zsp->z_wp = zsp->z_start + zsp->z_size;
> - zsp->z_cond = ZC5_FULL;
> - }
> + zc == ZC3_EXPLICIT_OPEN || (empty && zc == ZC1_EMPTY))
> + zbc_set_zone_full(devip, zsp);
In the equivalent function (null_finish_zone()) in null_blk,
we instead do this:
switch (zone->cond) {
case BLK_ZONE_COND_FULL:
/* finish operation on full is not an error */
goto unlock;
case BLK_ZONE_COND_EMPTY:
ret = null_check_zone_resources(dev, zone);
if (ret != BLK_STS_OK)
goto unlock;
break;
case BLK_ZONE_COND_IMP_OPEN:
dev->nr_zones_imp_open--;
break;
case BLK_ZONE_COND_EXP_OPEN:
dev->nr_zones_exp_open--;
break;
case BLK_ZONE_COND_CLOSED:
ret = null_check_zone_resources(dev, zone);
if (ret != BLK_STS_OK)
goto unlock;
dev->nr_zones_closed--;
break;
default:
ret = BLK_STS_IOERR;
goto unlock;
}
This might be a bit more verbose, but isn't it more clear to implement it in
this way, since the spec (ZBC) defines the transition for each zone state.
That way, it is easier to follow that each zone transition follows the spec.
I realize that the equivalent for null_check_zone_resources() in scsi debug,
is check_zbc_access_params().
So for scsi debug, the equivalent call to null_check_zone_resources() has
already been done elsewhere, and can be skipped, but other than that, the
code should be able to look the same as null_blk.
Doing so also avoids the unnecessary temporary transition to CLOSED, and
increasing + decreasing the nr_closed counter, before transitioning to FULL.
The reason why I don't really like this, is that ZBC does not mention that
finishing a zone temporarily transitions to close, so why should the scsi
debug code do so, especially when the the null_blk code shows that it is
not needed.
Kind regards,
Niklas
