softirq: Fix unplug deadlock
PATCH 1/1 softirq: Fix unplug deadlock
Subject: [RT] softirq: Fix unplug deadlock
From: Peter Zijlstra <[email protected]>
Date: Fri Sep 30 15:52:14 CEST 2011
If ksoftirqd gets woken during hot-unplug, __thread_do_softirq() will call pin_current_cpu() which will block on the held cpu_hotplug.lock. Moving the offline check in __thread_do_softirq() before the pin_current_cpu() call doesn't work, since the wakeup can happen before we mark the cpu offline.
So here we have the ksoftirq thread stuck until hotplug finishes, but then the ksoftirq CPU_DOWN notifier issues kthread_stop() which will wait for the ksoftirq thread to go away -- while holding the hotplug lock.
Sort this by delaying the kthread_stop() until CPU_POST_DEAD, which is outside of the cpu_hotplug.lock, but still serialized by the cpu_add_remove_lock.
--- kernel/softirq.c | 8 ++------ 1 file changed, 2 insertions(+), 6 deletions(-)
kernel/softirq.c +2 -6
@@ -1104,9 +1104,8 @@ static int __cpuinit cpu_callback(structint hotcpu = (unsigned long )hcpu;struct task_struct *p; -switch (action) { +switch (action & ~CPU_TASKS_FROZEN ) {case CPU_UP_PREPARE : -case CPU_UP_PREPARE_FROZEN : p = kthread_create_on_node(run_ksoftirqd, hcpu, cpu_to_node(hotcpu),
@@ -1119,19 +1118,16 @@ static int __cpuinit cpu_callback(structper_cpu (ksoftirqd, hotcpu) = p;break ;case CPU_ONLINE : -case CPU_ONLINE_FROZEN :wake_up_process (per_cpu (ksoftirqd, hotcpu));break ;#ifdef CONFIG_HOTPLUG_CPU case CPU_UP_CANCELED : -case CPU_UP_CANCELED_FROZEN :if (!per_cpu (ksoftirqd, hotcpu))break ;/* Unbind so it can run. Fall thru. */ kthread_bind (per_cpu (ksoftirqd, hotcpu),cpumask_any (cpu_online_mask)); -case CPU_DEAD : -case CPU_DEAD_FROZEN : { +case CPU_POST_DEAD : {static const struct sched_param param = { .sched_priority =MAX_RT_PRIO -1 };