Bart, Mikulas

On Tue, Aug 17, 2021 at 10:44 PM Bart Van Assche [off-list ref] wrote:

On 8/17/21 3:14 AM, SelvaKumar S wrote:

quoted

Introduce REQ_OP_COPY, a no-merge copy offload operation. Create
bio with control information as payload and submit to the device.
Larger copy operation may be divided if necessary by looking at device
limits. REQ_OP_COPY(19) is a write op and takes zone_write_lock when
submitted to zoned device.
Native copy offload is not supported for stacked devices.

Using a single operation for copy-offloading instead of separate
operations for reading and writing is fundamentally incompatible with
the device mapper. I think we need a copy-offloading implementation that
is compatible with the device mapper.

While each read/write command is for a single contiguous range of
device, with simple-copy we get to operate on multiple discontiguous
ranges, with a single command.
That seemed like a good opportunity to reduce control-plane traffic
(compared to read/write operations) as well.

With a separate read-and-write bio approach, each source-range will
spawn at least one read, one write and eventually one SCC command. And
it only gets worse as there could be many such discontiguous ranges (for
GC use-case at least) coming from user-space in a single payload.
Overall sequence will be
- Receive a payload from user-space
- Disassemble into many read-write pair bios at block-layer
- Assemble those (somehow) in NVMe to reduce simple-copy commands
- Send commands to device

We thought payload could be a good way to reduce the
disassembly/assembly work and traffic between block-layer to nvme.
How do you see this tradeoff?  What seems necessary for device-mapper
usecase, appears to be a cost when device-mapper isn't used.
Especially for SCC (since copy is within single ns), device-mappers
may not be too compelling anyway.

Must device-mapper support be a requirement for the initial support atop SCC?
Or do you think it will still be a progress if we finalize the
user-space interface to cover all that is foreseeable.And for
device-mapper compatible transport between block-layer and NVMe - we
do it in the later stage when NVMe too comes up with better copy
capabilities?


-- 
Joshi