On 1/28/19 4:07 PM, Bart Van Assche wrote:

On 1/28/19 4:56 AM, Matias Bjorling wrote:

quoted

Damien and I would like to propose a couple of topics centering around
zoned block devices:

1) Zoned block devices require that writes to a zone are sequential. If
the writes are dispatched to the device out of order, the drive rejects
the write with a write failure.

So far it has been the responsibility the deadline I/O scheduler to
serialize writes to zones to avoid intra-zone write command reordering.
This I/O scheduler based approach has worked so far for HDDs, but we can
do better for multi-queue devices. NVMe has support for multiple queues,
and one could dedicate a single queue to writes alone. Furthermore, the
queue is processed in-order, enabling the host to serialize writes on
the queue, instead of issuing them one by one. We like to gather
feedback on this approach (new HCTX_TYPE_WRITE).

2) Adoption of Zone Append in file-systems and user-space applications.

A Zone Append command, together with Zoned Namespaces, is being defined
in the NVMe workgroup. The new command allows one to automatically
direct writes to a zone write pointer position, similarly to writing to
a file open with O_APPEND. With this write append command, the drive
returns where data was written in the zone. Providing two benefits:

(A) It moves the fine-grained logical block allocation in file-systems
to the device side. A file-system continues to do coarse-grained logical
block allocation, but the specific LBAs where data is written and
reported from the device. Thus improving file-system performance. The
current target is XFS but we would like to hear the feasibility of it
being used in other file-systems.

(B) It lets host issue multiple outstanding write I/Os to a zone,
without having to maintain I/O order. Thus, improving the performance of
the drive, but also reducing the need for zone locking on the host side.

Is there other use-cases for this, and will an interface like this be
valuable in the kernel? If the interface is successful, we would expect
the interface to move to ATA/SCSI for standardization as well.

Hi Matias,

This topic proposal sounds interesting to me, but I think it is 
incomplete. Shouldn't it also be discussed how user space applications 
are expected to submit "zone append" writes? Which system call should 
e.g. fio use to submit this new type of write request? How will the 
offset at which data has been written be communicated back to user space?

Thanks,

Bart.

Hi Bart,

That's a good point. Originally, we only looked into support for 
file-systems due to the complexity of exposing it to user-space (e.g., 
we do not have an easy way to support psync/libaio workloads). I would 
love for us to be able to combine this with liburing, such that an LBA 
can be returned on I/O completion. However, I'm not sure we have enough 
bits available on the completion entry.

-Matias