On 11/20/24 05:51, Bart Van Assche wrote:
>>> +/*
>>> + * Change the zone state to "error" if a request is requeued to postpone
>>> + * processing of requeued requests until all pending requests have either
>>> + * completed or have been requeued.
>>> + */
>>> +void blk_zone_write_plug_requeue_request(struct request *rq)
>>> +{
>>> + struct gendisk *disk = rq->q->disk;
>>> + struct blk_zone_wplug *zwplug;
>>> +
>>> + if (!disk->zone_wplugs_hash_bits || !blk_rq_is_seq_zoned_write(rq))
>>> + return;
>>
>> I think the disk->zone_wplugs_hash_bits check needs to go inside
>> disk_get_zone_wplug() as that will avoid a similar check in
>> blk_zone_write_plug_free_request() too. That said, I am not even convinced it
>> is needed at all since these functions should be called only for a zoned drive
>> which should have its zone wplug hash setup.
>
> Moving the disk->zone_wplugs_hash_bits check sounds good to me.
>
> I added this check after hitting an UBSAN report that indicates that
> disk->zone_wplugs_hash_bits was used before it was changed into a non-
> zero value. sd_revalidate_disk() submits a very substantial number of
> SCSI commands before it calls blk_revalidate_disk_zones(), the function
> that sets disk->zone_wplugs_hash_bits.
But non of the commands are writes to sequential zones, so the hash bits check
should not even be necessary, no ?
--
Damien Le Moal
Western Digital Research
