On Thu, Dec 3, 2020 at 2:04 PM Mike Kravetz mike.kravetz@oracle.com wrote:
Adrian Moreno was ruuning a kubernetes 1.19 + containerd/docker workload using hugetlbfs. In this environment the issue is reproduced by: 1 - Start a simple pod that uses the recently added HugePages medium feature (pod yaml attached) 2 - Start a DPDK app. It doesn't need to run successfully (as in transfer packets) nor interact with real hardware. It seems just initializing the EAL layer (which handles hugepage reservation and locking) is enough to trigger the issue 3 - Delete the Pod (or let it "Complete").
This would result in a kworker thread going into a tight loop (top output): 1425 root 20 0 0 0 0 R 99.7 0.0 5:22.45 kworker/28:7+cgroup_destroy
'perf top -g' reports:
- 63.28% 0.01% [kernel] [k] worker_thread
- 49.97% worker_thread
- 52.64% process_one_work
- 62.08% css_killed_work_fn
- hugetlb_cgroup_css_offline 41.52% _raw_spin_lock
- 2.82% _cond_resched rcu_all_qs 2.66% PageHuge
- 0.57% schedule
- 0.57% __schedule
We are spinning in the do-while loop in hugetlb_cgroup_css_offline. Worse yet, we are holding the master cgroup lock (cgroup_mutex) while infinitely spinning. Little else can be done on the system as the cgroup_mutex can not be acquired.
Do note that the issue can be reproduced by simply offlining a hugetlb cgroup containing pages with reservation counts.
The loop in hugetlb_cgroup_css_offline is moving page counts from the cgroup being offlined to the parent cgroup. This is done for each hstate, and is repeated until hugetlb_cgroup_have_usage returns false. The routine moving counts (hugetlb_cgroup_move_parent) is only moving 'usage' counts. The routine hugetlb_cgroup_have_usage is checking for both 'usage' and 'reservation' counts. Discussion about what to do with reservation counts when reparenting was discussed here:
https://lore.kernel.org/linux-kselftest/CAHS8izMFAYTgxym-Hzb_JmkTK1N_S9tGN71...
The decision was made to leave a zombie cgroup for with reservation counts. Unfortunately, the code checking reservation counts was incorrectly added to hugetlb_cgroup_have_usage.
To fix the issue, simply remove the check for reservation counts. While fixing this issue, a related bug in hugetlb_cgroup_css_offline was noticed. The hstate index is not reinitialized each time through the do-while loop. Fix this as well.
Fixes: 1adc4d419aa2 ("hugetlb_cgroup: add interface for charge/uncharge hugetlb reservations") Cc: stable@vger.kernel.org Reported-by: Adrian Moreno amorenoz@redhat.com Tested-by: Adrian Moreno amorenoz@redhat.com Signed-off-by: Mike Kravetz mike.kravetz@oracle.com
Reviewed-by: Shakeel Butt shakeelb@google.com