Hi,
On 09/11/24 21:50, Waiman Long wrote:
> Since commit ff0ce721ec21 ("cgroup/cpuset: Eliminate unncessary
> sched domains rebuilds in hotplug"), there is only one
> rebuild_sched_domains_locked() call per hotplug operation. However,
> writing to the various cpuset control files may still casue more than
> one rebuild_sched_domains_locked() call to happen in some cases.
>
> Juri had found that two rebuild_sched_domains_locked() calls in
> update_prstate(), one from update_cpumasks_hier() and another one from
> update_partition_sd_lb() could cause cpuset partition to be created
> with null total_bw for DL tasks. IOW, DL tasks may not be scheduled
> correctly in such a partition.
>
> A sample command sequence that can reproduce null total_bw is as
> follows.
>
> # echo Y >/sys/kernel/debug/sched/verbose
> # echo +cpuset >/sys/fs/cgroup/cgroup.subtree_control
> # mkdir /sys/fs/cgroup/test
> # echo 0-7 > /sys/fs/cgroup/test/cpuset.cpus
> # echo 6-7 > /sys/fs/cgroup/test/cpuset.cpus.exclusive
> # echo root >/sys/fs/cgroup/test/cpuset.cpus.partition
>
> Fix this double rebuild_sched_domains_locked() calls problem
> by replacing existing calls with cpuset_force_rebuild() except
> the rebuild_sched_domains_cpuslocked() call at the end of
> cpuset_handle_hotplug(). Checking of the force_sd_rebuild flag is
> now done at the end of cpuset_write_resmask() and update_prstate()
> to determine if rebuild_sched_domains_locked() should be called or not.
>
> The cpuset v1 code can still call rebuild_sched_domains_locked()
> directly as double rebuild_sched_domains_locked() calls is not possible.
>
> Reported-by: Juri Lelli <juri.lelli@xxxxxxxxxx>
> Closes: https://lore.kernel.org/lkml/ZyuUcJDPBln1BK1Y@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/
> Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
This indeed works for me and fixes things with the test above (on v2).
Tested-by: Juri Lelli <juri.lelli@xxxxxxxxxx>
Thanks!
Juri
