On 2021/01/20 19:10, Christoph Hellwig wrote:
> On Tue, Jan 19, 2021 at 10:17:22PM +0900, Damien Le Moal wrote:
>> Per ZBC and ZAC specifications, host-managed SMR hard-disks mandate that
>> all writes into sequential write required zones be aligned to the device
>> physical block size. However, NVMe ZNS does not have this constraint and
>> allows write operations into sequential zones to be logical block size
>> aligned. This inconsistency does not help with portability of software
>> across device types.
>> To solve this, introduce the zone_write_granularity queue limit to
>> indicate the alignment constraint, in bytes, of write operations into
>> zones of a zoned block device. This new limit is exported as a
>> read-only sysfs queue attribute and the helper
>> blk_queue_zone_write_granularity() introduced for drivers to set this
>> limit. The scsi disk driver is modified to use this helper to set
>> host-managed SMR disk zone write granularity to the disk physical block
>> size. The nvme driver zns support use this helper to set the new limit
>> to the logical block size of the zoned namespace.
>>
>> Signed-off-by: Damien Le Moal <damien.lemoal@xxxxxxx>
>> ---
>> Documentation/block/queue-sysfs.rst | 7 +++++++
>> block/blk-settings.c | 28 ++++++++++++++++++++++++++++
>> block/blk-sysfs.c | 7 +++++++
>> drivers/nvme/host/zns.c | 1 +
>> drivers/scsi/sd_zbc.c | 10 ++++++++++
>> include/linux/blkdev.h | 3 +++
>> 6 files changed, 56 insertions(+)
>>
>> diff --git a/Documentation/block/queue-sysfs.rst b/Documentation/block/queue-sysfs.rst
>> index 2638d3446b79..c8bf8bc3c03a 100644
>> --- a/Documentation/block/queue-sysfs.rst
>> +++ b/Documentation/block/queue-sysfs.rst
>> @@ -273,4 +273,11 @@ devices are described in the ZBC (Zoned Block Commands) and ZAC
>> do not support zone commands, they will be treated as regular block devices
>> and zoned will report "none".
>>
>> +zone_write_granularity (RO)
>> +---------------------------
>> +This indicates the alignment constraint, in bytes, for write operations in
>> +sequential zones of zoned block devices (devices with a zoned attributed
>> +that reports "host-managed" or "host-aware"). This value is always 0 for
>> +regular block devices.
>> +
>> Jens Axboe <jens.axboe@xxxxxxxxxx>, February 2009
>> diff --git a/block/blk-settings.c b/block/blk-settings.c
>> index 43990b1d148b..6be6ed9485e3 100644
>> --- a/block/blk-settings.c
>> +++ b/block/blk-settings.c
>> @@ -60,6 +60,7 @@ void blk_set_default_limits(struct queue_limits *lim)
>> lim->io_opt = 0;
>> lim->misaligned = 0;
>> lim->zoned = BLK_ZONED_NONE;
>> + lim->zone_write_granularity = 0;
>
> I think this should default to 512 just like the logic and physical
> block size.
Hmm. I wanted to keep this limit to 0 for regular devices. If we default to 512,
regular devices will see that value. They can ignore it of course, but having it
as 0 makes it clear that it should be ignored.
>
>> }
>> EXPORT_SYMBOL(blk_set_default_limits);
>>
>> @@ -366,6 +367,31 @@ void blk_queue_physical_block_size(struct request_queue *q, unsigned int size)
>> }
>> EXPORT_SYMBOL(blk_queue_physical_block_size);
>>
>> +/**
>> + * blk_queue_zone_write_granularity - set zone write granularity for the queue
>> + * @q: the request queue for the zoned device
>> + * @size: the zone write granularity size, in bytes
>> + *
>> + * Description:
>> + * This should be set to the lowest possible size allowing to write in
>> + * sequential zones of a zoned block device.
>> + */
>> +void blk_queue_zone_write_granularity(struct request_queue *q,
>> + unsigned int size)
>> +{
>> + if (WARN_ON(!blk_queue_is_zoned(q)))
>> + return;
>> +
>> + q->limits.zone_write_granularity = size;
>> +
>> + if (q->limits.zone_write_granularity < q->limits.logical_block_size)
>> + q->limits.zone_write_granularity = q->limits.logical_block_size;
>
> I think this should be a WARN_ON_ONCE.
OK.
>
>> + if (q->limits.zone_write_granularity < q->limits.io_min)
>> + q->limits.zone_write_granularity = q->limits.io_min;
>
> I don't think this makes sense at all.
Arg ! Yes, that is a hint only ! I misread the comments for blk_limits_io_min().
Will remove this.
>
>> +static ssize_t queue_zone_write_granularity_show(struct request_queue *q, char *page)
>
> Overly long line.
>
>> + /*
>> + * Per ZBC and ZAC specifications, writes in sequential write required
>> + * zones of host-managed devices must be aligned to the device physical
>> + * block size.
>> + */
>> + if (blk_queue_zoned_model(q) == BLK_ZONED_HM)
>> + blk_queue_zone_write_granularity(q, sdkp->physical_block_size);
>> + else
>> + blk_queue_zone_write_granularity(q, sdkp->device->sector_size);
>
> Do we really want to special case HA drives here? I though we generally
> either treat them as drive managed (if they have partitions) or else like
> host managed ones.
If the HA drive is treated as drive-managed (i.e. it has a partition), then the
model here will be seen as BLK_ZONED_NONE and we should ignore it, or better,
set the zone_write_granularity to 0. If the drive is actually used as a zoned
drive, then we will see BLK_ZONED_HA here and we can use the logical block size
since HA drives do not have that restriction on physical block alignment.
So my code above is wrong. The else should be
else if (blk_queue_zoned_model(q) == BLK_ZONED_HA)
And I think we also need to add "q->limit.zone_write_granularity = 0" in
blk_queue_set_zoned() if model == BLK_ZONED_NONE at the end of that function.
Will send a v3 with these fixes.
--
Damien Le Moal
Western Digital Research
