On Wed, Feb 19, 2020 at 03:38:00PM +0900, Damien Le Moal wrote:
> The block layer generic blk_revalidate_disk_zones() checks the validity
> of zone descriptors reported by a disk using the
> blk_revalidate_zone_cb() callback function executed for each zone
> descriptor. If a ZBC disk reports invalid zone descriptors,
> blk_revalidate_disk_zones() returns an error and sd_zbc_read_zones()
> changes the disk capacity to 0, which in turn results in the gendisk
> structure capacity to be set to 0. This all works well for the first
> revalidate pass on a disk and the block layer detects the capactiy
> change.
>
> On the second revalidate pass, blk_revalidate_disk_zones() is called
> again and sd_zbc_report_zones() executed to check the zones a second
> time. However, for this second pass, the gendisk capacity is now 0,
> which results in sd_zbc_report_zones() to do nothing and to report
> success and no zones. blk_revalidate_disk_zones() in turn returns
> success and sets the disk queue chunk_sectors limit with zero as
> no zones were checked, causing a oops to trigger on the
> BUG_ON(!is_power_of_2(chunk_sectors)) in blk_queue_chunk_sectors().
>
> Fix this by using the sdkp capacity field rather than the gendisk
> capacity for the report zones loop in sd_zbc_report_zones(). Also add a
> check to return immediately an error if the sdkp capacity is 0.
> With this fix, invalid/buggy ZBC disk scan does not trigger a oops and
> are exposed with a 0 capacity. This change also preserve the chance for
> the disk to be correctly revalidated on the second revalidate pass as
> the scsi disk structure capacity field is always set to the disk
> reported value when sd_zbc_report_zones() is called.
>
> Fixes: d41003513e61 ("block: rework zone reporting")
> Cc: Cc: <stable@xxxxxxxxxxxxxxx> # v5.5
> Signed-off-by: Damien Le Moal <damien.lemoal@xxxxxxx>
Looks good,
Reviewed-by: Christoph Hellwig <hch@xxxxxx>
