On 2020/04/08 2:05, Christoph Hellwig wrote:
[...]
>> + /* Only zone reset and zone finish need zone write locking */
>> + if (op != ZO_RESET_WRITE_POINTER && op != ZO_FINISH_ZONE)
>> + return BLK_STS_OK;
>> +
>> + if (all) {
>> + /* We do not write lock all zones for an all zone reset */
>> + if (op == ZO_RESET_WRITE_POINTER)
>> + return BLK_STS_OK;
>> +
>> + /* Finishing all zones is not supported */
>> + return BLK_STS_IOERR;
>
> I still find different locking rules for the all vs individual zone
> operations weird.
>
This is to avoid taking the zone write lock for 10s of thousands of zones, one
at a time. That would be too heavy.
As we discussed before, if the user is not well behaving and issuing writes and
zone reset/finish simultaneously for the same zone, errors will likely happen
with or without the zone write locking being used on the reset/finish side. So I
think we can safely remove the zone locking for reset and finish of a single
zone. If an error happens, the error recovery report zone will update the wp
offset. All we need is the spinlock for the wp_ofst update. So we can clean this
up further and have this locking difference you point out going away.
--
Damien Le Moal
Western Digital Research
