On 2021/01/19 3:25, Christoph Hellwig wrote:

On Tue, Jan 12, 2021 at 07:52:27AM +0000, Damien Le Moal wrote:

quoted

quoted

I do not understand the logic here, given that NVMe does not have
conventional zones.

512e SAS & SATA SMR drives (512B logical, 4K physical) are a big thing, and for
these, all writes in sequential zones must be 4K aligned. So I suggested to
Chaitanya to simply use the physical block size as the LBA size for the target
to avoid weird IO errors that would not make sense in ZNS/NVMe world (e.g. 512B
aligned write requests failing).

But in NVMe the physical block size exposes the atomic write unit, which
could be way too large.  Іf we want to do this cleanly we need to expose
a minimum sequential zone write alignment value in the block layer.

OK, good point. I think I can cook something like that today.

-- 
Damien Le Moal
Western Digital Research