Re: [PATCH v3] Weighted interleave auto-tuning

[Gregory Price <gourry@gourry.net>]

On Tue, Jan 21, 2025 at 08:27:17PM +0900, Honggyu Kim wrote:
> Hi Ying and Joshua,
> > IMHO, this interface is somewhat hard to be used.  Users need to know
> > which value is legal.  So, this will become something like,
> > 
> > $ cat mode
> > auto [manual]
> > $ echo auto > mode
> > $ cat mode
> > [auto] manual
> 
> This is exactly I internally proposed to Hyeonggon, but couldn't share
> the idea directly here.
> 
> That also makes sense, but I feel like somewhat vague what "auto" false
> means. The "auto" might be better to be "use_hmat" instead and this
> makes "use_hmat" false more meaningful. "use_hmat_weight" or
> "use_hmat_info" might be another candidates.
> 

I don't think we want to encode hmat-ism into the uapi. In fact,
mempolicy doesn't even know about hmat.  It just gets source information
from *somewhere* and applies it accordingly.

I think what you might be asking for is

auto -> [true, false]

if auto=true
  mode -> [default,
           read_bw, write_bw, combined_bw,
           read_ltc, write_ltc, combined_ltc]
if auto=false
  mode -> [disabled]

Where default mode is the kernel selection of whatever combination of
read/write bw/ltc data, and user could switch the attribute.

HOWEVER, such `mode` would require us to cache the attribute structure
per-node, and maybe some thought on what's reasonable, so that I would
prefer that to be a completely different feature / discussion.

~Gregory