workqueue: Drain BH work items on hot-unplugged CPUs
authorTejun Heo <tj@kernel.org>
Tue, 27 Feb 2024 01:38:55 +0000 (15:38 -1000)
committerTejun Heo <tj@kernel.org>
Thu, 29 Feb 2024 21:51:24 +0000 (11:51 -1000)
Boqun pointed out that workqueues aren't handling BH work items on offlined
CPUs. Unlike tasklet which transfers out the pending tasks from
CPUHP_SOFTIRQ_DEAD, BH workqueue would just leave them pending which is
problematic. Note that this behavior is specific to BH workqueues as the
non-BH per-CPU workers just become unbound when the CPU goes offline.

This patch fixes the issue by draining the pending BH work items from an
offlined CPU from CPUHP_SOFTIRQ_DEAD. Because work items carry more context,
it's not as easy to transfer the pending work items from one pool to
another. Instead, run BH work items which execute the offlined pools on an
online CPU.

Note that this assumes that no further BH work items will be queued on the
offlined CPUs. This assumption is shared with tasklet and should be fine for
conversions. However, this issue also exists for per-CPU workqueues which
will just keep executing work items queued after CPU offline on unbound
workers and workqueue should reject per-CPU and BH work items queued on
offline CPUs. This will be addressed separately later.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-and-reviewed-by: Boqun Feng <boqun.feng@gmail.com>
Link: http://lkml.kernel.org/r/Zdvw0HdSXcU3JZ4g@boqun-archlinux
include/linux/workqueue.h
kernel/softirq.c
kernel/workqueue.c

index 64a60b9232d30f21437dada0e54f08e3d8ccc433..158784dd189ab2a424b02fca0760f95f0af7de2d 100644 (file)
@@ -458,6 +458,7 @@ extern struct workqueue_struct *system_bh_wq;
 extern struct workqueue_struct *system_bh_highpri_wq;
 
 void workqueue_softirq_action(bool highpri);
+void workqueue_softirq_dead(unsigned int cpu);
 
 /**
  * alloc_workqueue - allocate a workqueue
index 547d282548a8e3543763eed9ce0f01633ff8dcd7..b315b21fb28cd281fc38d1a8a00ed9f9527e2f53 100644 (file)
@@ -932,6 +932,8 @@ static void run_ksoftirqd(unsigned int cpu)
 #ifdef CONFIG_HOTPLUG_CPU
 static int takeover_tasklets(unsigned int cpu)
 {
+       workqueue_softirq_dead(cpu);
+
        /* CPU is dead, so no lock needed. */
        local_irq_disable();
 
index 38783e3a60bbfa627c34a308c5fb87b19615d533..a60eb65955e739108c878229ee3042dcf311363e 100644 (file)
@@ -81,6 +81,7 @@ enum worker_pool_flags {
        POOL_BH                 = 1 << 0,       /* is a BH pool */
        POOL_MANAGER_ACTIVE     = 1 << 1,       /* being managed */
        POOL_DISASSOCIATED      = 1 << 2,       /* cpu can't serve workers */
+       POOL_BH_DRAINING        = 1 << 3,       /* draining after CPU offline */
 };
 
 enum worker_flags {
@@ -1218,7 +1219,9 @@ static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
 static void kick_bh_pool(struct worker_pool *pool)
 {
 #ifdef CONFIG_SMP
-       if (unlikely(pool->cpu != smp_processor_id())) {
+       /* see drain_dead_softirq_workfn() for BH_DRAINING */
+       if (unlikely(pool->cpu != smp_processor_id() &&
+                    !(pool->flags & POOL_BH_DRAINING))) {
                irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
                return;
        }
@@ -3155,6 +3158,7 @@ __acquires(&pool->lock)
        struct worker_pool *pool = worker->pool;
        unsigned long work_data;
        int lockdep_start_depth, rcu_start_depth;
+       bool bh_draining = pool->flags & POOL_BH_DRAINING;
 #ifdef CONFIG_LOCKDEP
        /*
         * It is permissible to free the struct work_struct from
@@ -3220,7 +3224,9 @@ __acquires(&pool->lock)
 
        rcu_start_depth = rcu_preempt_depth();
        lockdep_start_depth = lockdep_depth(current);
-       lock_map_acquire(&pwq->wq->lockdep_map);
+       /* see drain_dead_softirq_workfn() */
+       if (!bh_draining)
+               lock_map_acquire(&pwq->wq->lockdep_map);
        lock_map_acquire(&lockdep_map);
        /*
         * Strictly speaking we should mark the invariant state without holding
@@ -3253,7 +3259,8 @@ __acquires(&pool->lock)
        trace_workqueue_execute_end(work, worker->current_func);
        pwq->stats[PWQ_STAT_COMPLETED]++;
        lock_map_release(&lockdep_map);
-       lock_map_release(&pwq->wq->lockdep_map);
+       if (!bh_draining)
+               lock_map_release(&pwq->wq->lockdep_map);
 
        if (unlikely((worker->task && in_atomic()) ||
                     lockdep_depth(current) != lockdep_start_depth ||
@@ -3615,6 +3622,84 @@ void workqueue_softirq_action(bool highpri)
                bh_worker(list_first_entry(&pool->workers, struct worker, node));
 }
 
+struct wq_drain_dead_softirq_work {
+       struct work_struct      work;
+       struct worker_pool      *pool;
+       struct completion       done;
+};
+
+static void drain_dead_softirq_workfn(struct work_struct *work)
+{
+       struct wq_drain_dead_softirq_work *dead_work =
+               container_of(work, struct wq_drain_dead_softirq_work, work);
+       struct worker_pool *pool = dead_work->pool;
+       bool repeat;
+
+       /*
+        * @pool's CPU is dead and we want to execute its still pending work
+        * items from this BH work item which is running on a different CPU. As
+        * its CPU is dead, @pool can't be kicked and, as work execution path
+        * will be nested, a lockdep annotation needs to be suppressed. Mark
+        * @pool with %POOL_BH_DRAINING for the special treatments.
+        */
+       raw_spin_lock_irq(&pool->lock);
+       pool->flags |= POOL_BH_DRAINING;
+       raw_spin_unlock_irq(&pool->lock);
+
+       bh_worker(list_first_entry(&pool->workers, struct worker, node));
+
+       raw_spin_lock_irq(&pool->lock);
+       pool->flags &= ~POOL_BH_DRAINING;
+       repeat = need_more_worker(pool);
+       raw_spin_unlock_irq(&pool->lock);
+
+       /*
+        * bh_worker() might hit consecutive execution limit and bail. If there
+        * still are pending work items, reschedule self and return so that we
+        * don't hog this CPU's BH.
+        */
+       if (repeat) {
+               if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
+                       queue_work(system_bh_highpri_wq, work);
+               else
+                       queue_work(system_bh_wq, work);
+       } else {
+               complete(&dead_work->done);
+       }
+}
+
+/*
+ * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
+ * possible to allocate dead_work per CPU and avoid flushing. However, then we
+ * have to worry about draining overlapping with CPU coming back online or
+ * nesting (one CPU's dead_work queued on another CPU which is also dead and so
+ * on). Let's keep it simple and drain them synchronously. These are BH work
+ * items which shouldn't be requeued on the same pool. Shouldn't take long.
+ */
+void workqueue_softirq_dead(unsigned int cpu)
+{
+       int i;
+
+       for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
+               struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
+               struct wq_drain_dead_softirq_work dead_work;
+
+               if (!need_more_worker(pool))
+                       continue;
+
+               INIT_WORK(&dead_work.work, drain_dead_softirq_workfn);
+               dead_work.pool = pool;
+               init_completion(&dead_work.done);
+
+               if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
+                       queue_work(system_bh_highpri_wq, &dead_work.work);
+               else
+                       queue_work(system_bh_wq, &dead_work.work);
+
+               wait_for_completion(&dead_work.done);
+       }
+}
+
 /**
  * check_flush_dependency - check for flush dependency sanity
  * @target_wq: workqueue being flushed