This repository has been archived by the owner on Jun 18, 2024. It is now read-only.
scx: Fix hotplug #217
Merged
scx: Fix hotplug #217
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
The existing hotplug code is a bit brittle in that the sched_class->{online,
offline} callbacks can be invoked both from hotplug context, or from domain
rebuild context. A callback is only invoked in one of the two contexts, with
the context that runs the callback setting rq->online accordingly to avoid
doing the callback more times than necessary. Unfortuntaly, this causes commit
2125c00 ("cgroup/cpuset: Make cpuset hotplug processing synchronous") to
break hotplug for us because it makes the topology rq->online event happen
before the cpu hotplug rq->online event; thus preventing us from invoking the
cpu online callback in the scheduler.
This integration is fragile and hacky, so we'll instead call directly into
ext.c on the hotplug path. This will be added in a subsequent commit.
Signed-off-by: David Vernet <void@manifault.com>
htejun
reviewed
Jun 6, 2024
The rq_online() and rq_offline() sched_class callbacks may be invoked on either the domain creation path, or on the hotplug path. ext.c would previously only invoke its hotplug logic when those callbacks are invoked on the hotplug path. This turned out to be a bit brittle, as hotplug could break for sched_ext if we first invoked rq_online() from the domain path, as it would cause us to skip invoking it on the hotplug path due to hotplug implementation details. To avoid this frailty (and fix a hotplug regression for us after merging v6.10), let's instead just call directly into ext.c on the hotplug path. In doing so, we can also move the ops.cpu_online() and ops.cpu_offline() callbacks outside of the rq lock region, thus making them sleepable. A test for this will be added in a subsequent patch. Signed-off-by: David Vernet <void@manifault.com>
Now that ops.cpu_online() and ops.cpu_offline() can be sleepable, let's update their test BPF progs to be sleepable to validate. Signed-off-by: David Vernet <void@manifault.com>
Good, sorry for the oversight. I've added a call to a blockable kfunc in the testcase which would've noticed that. |
htejun
approved these changes
Jun 6, 2024
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The rq_online() and rq_offline() sched_class callbacks may be invoked on either
the domain creation path, or on the hotplug path. ext.c will only
invoke its hotplug logic when those callbacks are invoked on the hotplug path.
This turned out to be a bit brittle, as hotplug could break for sched_ext if we
first invoked rq_online() from the domain path, as it would cause us to skip
invoking it on the hotplug path due to hotplug implementation details. This in
fact happened after rebasing onto v6.10.
To avoid this frailty (and fix a hotplug regression for us after merging
v6.10), let's instead just call directly into ext.c on the hotplug path. This
could be made into a sched_class callback if that's what schedulers want, but
that's kind of what rq_online() and rq_offline() are supposed to be, so it's a
bit awkward as is. It would have been ideal if we could have just used a
hotplug event notifier, but as far as I can tell that doesn't exist.