On 12/15/21 09:44, Michal Koutný wrote:

On Mon, Dec 13, 2021 at 11:00:17AM -1000, Tejun Heo [off-list ref] wrote:

quoted

* When a valid partition turns invalid, now we have a reliable way of
   discovering what exactly caused the transition. However, when a user now
   fails to turn a member into partition, all they get is -EINVAL and there's
   no way to discover why it failed and the failure conditions that -EINVAL
   represents aren't simple.

* In an automated configuration scenarios, this operation mode may be
   difficult to make reliable and lead to sporadic failures which can be
   tricky to track down. The core problem is that whether a given operation
   succeeds or not may depend on external states (CPU on/offline) which may
   change asynchronously in a way that the configuring entity doesn't have
   any control over.

It's true that both are existing problems with the current partition
interface and given that this is a pretty spcialized feature, this can be
okay. Michal, what are your thoughts?

Because of asynchronous changes, the return value should not be that
important and the user should watch cpuset.partitions for the result
(end state) anyway.
Furthermore, the reasons should be IMO just informative (i.e. I like
they're not explicitly documented) and not API.

But I see there could be a distinction between -EINVAL (the supplied
input makes no sense) and -EAGAIN(?) denoting that the switch to
partition root could not happen (due to outer constraints).

You seem to propose to replace the -EAGAIN above with a success code and
allow the switch to an invalid root.
The action of the configuring entity would be different: retry (when?)
vs wait till transition happens (notification) (although the immediate
effect (the change did not happen) is same).
I considered the two variants equal but the clear information about when
the change can happen I'd favor the variant allowing the switch to
invalid root now.

Allowing direct transition from member to invalid partition doesn't feel 
right for me. A casual user may assume a partition is correctly formed 
without double checking the "cpuset.partition" value. Returning an error 
will prevent this kind of issue. If returning more information about the 
failure is the main reason for allowing the invalid partition 
transition, we can extend the "cpuset.partition" read syntax to also 
show the reason for the previous failure.

Cheers,
Longman