mbox series

[v2,0/2] cgroup/cpuset: Make cpuset hotplug processing synchronous

Message ID 20240404134749.2857852-1-longman@redhat.com (mailing list archive)
Headers show
Series cgroup/cpuset: Make cpuset hotplug processing synchronous | expand

Message

Waiman Long April 4, 2024, 1:47 p.m. UTC
v2:
  - Found that rebuild_sched_domains() has external callers, revert its
    change and introduce rebuild_sched_domains_cpuslocked() instead.

As discussed in the LKML thread [1], the asynchronous nature of cpuset
hotplug handling code is causing problem with RCU testing. With recent
changes in the way locking is being handled in the cpuset code, it is
now possible to make the cpuset hotplug code synchronous again without
major changes.

This series enables the hotplug code to call directly into cpuset hotplug
core without indirection with the exception of the special case of v1
cpuset becoming empty still being handled indirectly with workqueue.

A new simple test case was also written to test this special v1 cpuset
case. The test_cpuset_prs.sh script was also run with LOCKDEP on to
verify that there is no regression.

[1] https://lore.kernel.org/lkml/ZgYikMb5kZ7rxPp6@slm.duckdns.org/

Waiman Long (2):
  cgroup/cpuset: Make cpuset hotplug processing synchronous
  cgroup/cpuset: Add test_cpuset_v1_hp.sh

 include/linux/cpuset.h                        |   3 -
 kernel/cgroup/cpuset.c                        | 141 +++++++-----------
 kernel/cpu.c                                  |  48 ------
 kernel/power/process.c                        |   2 -
 tools/testing/selftests/cgroup/Makefile       |   2 +-
 .../selftests/cgroup/test_cpuset_v1_hp.sh     |  46 ++++++
 6 files changed, 103 insertions(+), 139 deletions(-)
 create mode 100755 tools/testing/selftests/cgroup/test_cpuset_v1_hp.sh

Comments

Tejun Heo April 8, 2024, 5:40 p.m. UTC | #1
On Thu, Apr 04, 2024 at 09:47:47AM -0400, Waiman Long wrote:
>  v2:
>   - Found that rebuild_sched_domains() has external callers, revert its
>     change and introduce rebuild_sched_domains_cpuslocked() instead.
> 
> As discussed in the LKML thread [1], the asynchronous nature of cpuset
> hotplug handling code is causing problem with RCU testing. With recent
> changes in the way locking is being handled in the cpuset code, it is
> now possible to make the cpuset hotplug code synchronous again without
> major changes.
> 
> This series enables the hotplug code to call directly into cpuset hotplug
> core without indirection with the exception of the special case of v1
> cpuset becoming empty still being handled indirectly with workqueue.
> 
> A new simple test case was also written to test this special v1 cpuset
> case. The test_cpuset_prs.sh script was also run with LOCKDEP on to
> verify that there is no regression.

Applied to cgroup/for-6.10. Hopefully, this one will work out. Thanks for
working on this.