Re: mm: deadlock between get_online_cpus/pcpu_alloc

[Mel Gorman]

On Tue, Feb 07, 2017 at 11:35:52AM +0100, Michal Hocko wrote:
> On Tue 07-02-17 10:28:09, Mel Gorman wrote:
> > On Tue, Feb 07, 2017 at 10:49:28AM +0100, Vlastimil Babka wrote:
> > > On 02/07/2017 10:43 AM, Mel Gorman wrote:
> > > > If I'm reading this right, a hot-remove will set the pool POOL_DISASSOCIATED
> > > > and unbound. A workqueue queued for draining get migrated during hot-remove
> > > > and a drain operation will execute twice on a CPU -- one for what was
> > > > queued and a second time for the CPU it was migrated from. It should still
> > > > work with flush_work which doesn't appear to block forever if an item
> > > > got migrated to another workqueue. The actual drain workqueue function is
> > > > using the CPU ID it's currently running on so it shouldn't get confused.
> > > 
> > > Is the worker that will process this migrated workqueue also guaranteed
> > > to be pinned to a cpu for the whole work, though? drain_local_pages()
> > > needs that guarantee.
> > > 
> > 
> > It should be by running on a workqueue handler bound to that CPU (queued
> > on wq->cpu_pwqs in __queue_work)
> 
> Are you sure? The comment in kernel/workqueue.c says
>          * While DISASSOCIATED, the cpu may be offline and all workers have
>          * %WORKER_UNBOUND set and concurrency management disabled, and may
>          * be executing on any CPU.  The pool behaves as an unbound one.
> 
> I might be misreadig but an unbound pool can be handled by workers which
> are not pinned on any cpu AFAIU.

Right. The unbind operation can set a mask that is any allowable CPU and
the final process_work is not done in a context that prevents
preemption.

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 3b93879990fd..7af165d308c4 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -2342,7 +2342,14 @@ void drain_local_pages(struct zone *zone)
 
 static void drain_local_pages_wq(struct work_struct *work)
 {
+	/*
+	 * Ordinarily a drain operation is bound to a CPU but may be unbound
+	 * after a CPU hotplug operation so it's necessary to disable
+	 * preemption for the drain to stabilise the CPU ID.
+	 */
+	preempt_disable();
 	drain_local_pages(NULL);
+	preempt_enable_no_resched();
 }
 
 /*
@@ -2377,13 +2384,10 @@ void drain_all_pages(struct zone *zone)
 		mutex_lock(&pcpu_drain_mutex);
 	}
 
-	get_online_cpus();
-
 	/*
-	 * We don't care about racing with CPU hotplug event
-	 * as offline notification will cause the notified
-	 * cpu to drain that CPU pcps and on_each_cpu_mask
-	 * disables preemption as part of its processing
+	 * We don't care about racing with CPU hotplug event as offline
+	 * notification will cause the notified cpu to drain that CPU pcps
+	 * and it is serialised against here via pcpu_drain_mutex.
 	 */
 	for_each_online_cpu(cpu) {
 		struct per_cpu_pageset *pcp;
@@ -2418,7 +2422,6 @@ void drain_all_pages(struct zone *zone)
 	for_each_cpu(cpu, &cpus_with_pcps)
 		flush_work(per_cpu_ptr(&pcpu_drain, cpu));
 
-	put_online_cpus();
 	mutex_unlock(&pcpu_drain_mutex);
 }
 
@@ -6711,7 +6714,16 @@ static int page_alloc_cpu_dead(unsigned int cpu)
 {
 
 	lru_add_drain_cpu(cpu);
+
+	/*
+	 * A per-cpu drain via a workqueue from drain_all_pages can be
+	 * rescheduled onto an unrelated CPU. That allows the hotplug
+	 * operation and the drain to potentially race on the same
+	 * CPU. Serialise hotplug versus drain using pcpu_drain_mutex
+	 */
+	mutex_lock(&pcpu_drain_mutex);
 	drain_pages(cpu);
+	mutex_unlock(&pcpu_drain_mutex);
 
 	/*
 	 * Spill the event counters of the dead processor

-- 
Mel Gorman
SUSE Labs