On Tue, May 10, 2022 at 12:59 PM Tejun Heo [off-list ref] wrote:

Hello,

On Tue, May 10, 2022 at 12:34:42PM -0700, Yosry Ahmed wrote:

quoted

The rationale behind associating this work with cgroup_subsys is that
usually the stats are associated with a resource (e.g. memory, cpu,
etc). For example, if the memory controller is only enabled for a
subtree in a big hierarchy, it would be more efficient to only run BPF
rstat programs for those cgroups, not the entire hierarchy. It
provides a way to control what part of the hierarchy you want to
collect stats for. This is also semantically similar to the
css_rstat_flush() callback.

Hmm... one major point of rstat is not having to worry about these things
because we iterate what's been active rather than what exists. Now, this
isn't entirely true because we share the same updated list for all sources.
This is a trade-off which makes sense because 1. the number of cgroups to
iterate each cycle is generally really low anyway 2. different controllers
often get enabled together. If the balance tilts towards "we're walking too
many due to the sharing of updated list across different sources", the
solution would be splitting the updated list so that we make the walk finer
grained.

Note that the above doesn't really affect the conceptual model. It's purely
an optimization decision. Tying these things to a cgroup_subsys does affect
the conceptual model and, in this case, the userland API for a performance
consideration which can be solved otherwise.

So, let's please keep this simple and in the (unlikely) case that the
overhead becomes an issue, solve it from rstat operation side.

Thanks.

I assume if we do this optimization, and have separate updated lists
for controllers, we will still have a "core" updated list that is not
tied to any controller. Is this correct?

If yes, then we can make the interface controller-agnostic (a global
list of BPF flushers). If we do the optimization later, we tie BPF
stats to the "core" updated list. We can even extend the userland
interface then to allow for controller-specific BPF stats if found
useful.

If not, and there will only be controller-specific updated lists then,
then we might need to maintain a "core" updated list just for the sake
of BPF programs, which I don't think would be favorable.

What do you think? Either-way, I will try to document our discussion
outcome in the commit message (and maybe the code), so that
if-and-when this optimization is made, we can come back to it.

--
tejun