When looking up the pool_workqueue to use for an unbound workqueue, workqueue assumes that the target CPU is always bound to a valid NUMA node. However, currently, when a CPU goes offline, the mapping is destroyed and cpu_to_node() returns NUMA_NO_NODE. This has always been broken but hasn't triggered often enough before 874bbfe600a6 ("workqueue: make sure delayed work run in local cpu"). After the commit, workqueue forcifully assigns the local CPU for delayed work items without explicit target CPU to fix a different issue. This widens the window where CPU can go offline while a delayed work item is pending causing delayed work items dispatched with target CPU set to an already offlined CPU. The resulting NUMA_NO_NODE mapping makes workqueue try to queue the work item on a NULL pool_workqueue and thus crash. While 874bbfe600a6 has been reverted for a different reason making the bug less visible again, it can still happen. Fix it by mapping NUMA_NO_NODE to the default pool_workqueue from unbound_pwq_by_node(). This is a temporary workaround. The long term solution is keeping CPU -> NODE mapping stable across CPU off/online cycles which is being worked on.
Showing the changes to kernel/workqueue.c (1 of 1 file). Show all
Unified · Split
| @@ -586,6 +586,16 @@ static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq, · 585 unmodified lines | ||
| 586 | 586 | int node) |
| 587 | 587 | { |
| 588 | 588 | assert_rcu_or_wq_mutex_or_pool_mutex(wq); |
| 589 | + | |
| 590 | + /* | |
| 591 | + * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a | |
| 592 | + * delayed item is pending. The plan is to keep CPU -> NODE | |
| 593 | + * mapping valid and stable across CPU on/offlines. Once that | |
| 594 | + * happens, this workaround can be removed. | |
| 595 | + */ | |
| 596 | + if (unlikely(node == NUMA_NO_NODE)) | |
| 597 | + return wq->dfl_pwq; | |
| 598 | + | |
| 589 | 599 | return rcu_dereference_raw(wq->numa_pwq_tbl[node]); |
| 590 | 600 | } |
| 591 | 601 | |