On 14/02/2024 08:00, Christoph Hellwig wrote:
On Tue, Feb 13, 2024 at 02:21:25PM +0000, John Garry wrote:
Please also read through TP4098(a) and look at the MAM field.
It's not public, AFAIK.
Oracle is a member, so you can take a look at it easily. If we need
it for Linux I can also work with the NVMe Board to release it.
What I really meant was that I prefer not to quote private TPs in public
domain. I have the doc.
And I don't think a feature which allows us to straddle boundaries is too
interesting really.
Without MAM=1 NVMe can't support atomic writes larger than
AWUPF/NAWUPF, which is typically set to the indirection table
size. You're leaving a lot of potential unused with that.
atomic_write_unit_max would always be dictated by min of boundary and
AWUPF/NAWUPF. With MAM=1, we could increase atomic_write_max_bytes - but
does it really help us? I mean, atomic_write_max_bytes only comes into
play for merging - do we get much merging for NVMe transports? I am not
sure.
Thanks,
John
