Thanks for the clarification. I have a much better understanding now.
Updating the ndctl-create-namespace man page to clarify what '--region=all' does
and does not do in combination with other arguments and options would be beneficial.
This would actually provide a more immediate solution to the problem. It sounds like the
general statement used in the other man pages can remain? Or are there other
With the proposed implementation that bails out on first error, `--continue` seems
reasonable vs. `--greedy`. What do you think about `--all-regions` instead? It is more
meaningful towards its intended action or use-case, but would it cause confusion with
`--regions=all`? Documentation could provide the solution here. We would have to choose
an arbitrary short option since `-a` and `-r` are already taken.
Whatever the final decision, we should test for and return a usage error for a mutually
exclusive set of inputs, i.e. `ndctl create-namespace --all-regions --region=region0`
doesn't make sense - either you want to perform the create operation on all regions or
only in region0.
From: Dan Williams [mailto:email@example.com]
Sent: Wednesday, August 28, 2019 3:32 PM
To: Verma, Vishal L <vishal.l.verma(a)intel.com>
Cc: Scargall, Steve <steve.scargall(a)intel.com>; jmoyer(a)redhat.com;
Subject: Re: [ndctl RFC PATCH] ndctl/namespace: create namespaces greedily
On Wed, Aug 28, 2019 at 2:22 PM Verma, Vishal L <vishal.l.verma(a)intel.com> wrote:
On Wed, 2019-08-28 at 21:16 +0000, Verma, Vishal L wrote:
> On Wed, 2019-08-28 at 13:47 -0700, Scargall, Steve wrote:
> > Hi Jeff,
> > The issue is more of repetition. On an 8-socket system, should a
> > user really be expected to type 'ndctl create-namespace' eight times
> > vs. running 'ndctl create-namespace --region=all' once? SAP HANA is
> > an example app the requires one namespace per region. Scripting
> > is a viable solution, but that requires somebody to write the
> > script and do all the error checking & handling. Each
> > OEM/ISV/SysAdmin would have their own script. Pushing the logic
> > to ndctl seems to be a reasonable approach and let the user handle any errors
returned by ndctl.
> A scripted solution can indeed be really simple - e.g.:
> # while read -r region; do ndctl create-namespace
> --region="$region"; done < <(ndctl list --bus=nfit_test.0 -R | jq
> -r '..dev')
> "size":"62.00 MiB (65.01 MB)",
> "size":"30.00 MiB (31.46 MB)",
> > The ndctl-man-page implies the 'ndctl create-namespace --region=all'
> > feature exists today:
> > -r, --region=
> > A regionX device name, or a region id number. The
> > keyword all can be specified to carry out the operation on every
> > region in the system, optionally filtered by bus id (see --bus= option).
> This is true, but unfortunately, the implementation has treated
> create- namespace as an exception to this since the start, and I
> agree with Jeff that changing its behavior now can cause other
> Hyrum's law-esqe  breakage.
> I think however it should be easy to make a compromise, and retain
> the legacy behavior of create-namespace, while creating a new
> create- namespace-greedy command with the new semantics.
.. And it doesn't even need to be a new command, a simple --greedy
option to create-namespace should be sufficient.
Perhaps "--continue" to proceed with creating namespace after the first
successful invocation. I agree with Jeff that changing the default semantics would be
surprising. The man page can also be fixed up to make it clear that it's the
"singleton region capacity search", not "all possible namespaces" that
"-r all" implies.