On 18/06/2024 18:25, Keith Busch wrote:
On Tue, Jun 18, 2024 at 08:46:31AM +0100, John Garry wrote:
About NVMe, the spec says that NABSN and NOIOB may not be related to one
another (command set spec 1.0d 5.8.2.1), but I am wondering if people really
build HW which would have different NABSN/NABSPF and NOIOB. I don't know.
The history of NOIOB is from an nvme drive that had two back-end
controllers with their own isolated storage, and then striped together
on the front end for the host to see. A command crossing the stripe
boundary takes a slow path to split it for each backend controller's
portion and merge the results. Subsequent implementations may have
different reasons for advertising this boundary, but that was the
original.
In this case, I would expect NOIOB >= atomic write boundary.
Would it be sane to have a NOIOB < atomic write boundary in some other
config?
I can support these possibilities, but the code will just get more complex.
Anyway, there was an idea that the stripe size could be user
configurable, though that never shipped as far as I know. If it had,
then the optimal NOIOB could be made larger, but the atomic write size
doesn't change.
Thanks,
John
