workqueue: add delayed_work->wq to simplify reentrancy handling
authorLai Jiangshan <laijs@cn.fujitsu.com>
Thu, 7 Feb 2013 02:04:53 +0000 (18:04 -0800)
committerTejun Heo <tj@kernel.org>
Thu, 7 Feb 2013 02:04:53 +0000 (18:04 -0800)
To avoid executing the same work item from multiple CPUs concurrently,
a work_struct records the last pool it was on in its ->data so that,
on the next queueing, the pool can be queried to determine whether the
work item is still executing or not.

A delayed_work goes through timer before actually being queued on the
target workqueue and the timer needs to know the target workqueue and
CPU.  This is currently achieved by modifying delayed_work->work.data
such that it points to the cwq which points to the target workqueue
and the last CPU the work item was on.  __queue_delayed_work()
extracts the last CPU from delayed_work->work.data and then combines
it with the target workqueue to create new work.data.

The only thing this rather ugly hack achieves is encoding the target
workqueue into delayed_work->work.data without using a separate field,
which could be a trade off one can make; unfortunately, this entangles
work->data management between regular workqueue and delayed_work code
by setting cwq pointer before the work item is actually queued and
becomes a hindrance for further improvements of work->data handling.

This can be easily made sane by adding a target workqueue field to
delayed_work.  While delayed_work is used widely in the kernel and
this does make it a bit larger (<5%), I think this is the right
trade-off especially given the prospect of much saner handling of
work->data which currently involves quite tricky memory barrier
dancing, and don't expect to see any measureable effect.

Add delayed_work->wq and drop the delayed_work->work.data overloading.

tj: Rewrote the description.

Signed-off-by: Lai Jiangshan <laijs@cn.fujitsu.com>
Signed-off-by: Tejun Heo <tj@kernel.org>
include/linux/workqueue.h
kernel/workqueue.c

index 426c39c..a3d7556 100644 (file)
@@ -109,6 +109,9 @@ struct work_struct {
 struct delayed_work {
        struct work_struct work;
        struct timer_list timer;
+
+       /* target workqueue and CPU ->timer uses to queue ->work */
+       struct workqueue_struct *wq;
        int cpu;
 };
 
index a229a56..41a502c 100644 (file)
@@ -1339,10 +1339,9 @@ EXPORT_SYMBOL_GPL(queue_work);
 void delayed_work_timer_fn(unsigned long __data)
 {
        struct delayed_work *dwork = (struct delayed_work *)__data;
-       struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
 
        /* should have been called from irqsafe timer with irq already off */
-       __queue_work(dwork->cpu, cwq->wq, &dwork->work);
+       __queue_work(dwork->cpu, dwork->wq, &dwork->work);
 }
 EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
 
@@ -1351,7 +1350,6 @@ static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
 {
        struct timer_list *timer = &dwork->timer;
        struct work_struct *work = &dwork->work;
-       unsigned int lcpu;
 
        WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
                     timer->data != (unsigned long)dwork);
@@ -1371,30 +1369,7 @@ static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
 
        timer_stats_timer_set_start_info(&dwork->timer);
 
-       /*
-        * This stores cwq for the moment, for the timer_fn.  Note that the
-        * work's pool is preserved to allow reentrance detection for
-        * delayed works.
-        */
-       if (!(wq->flags & WQ_UNBOUND)) {
-               struct worker_pool *pool = get_work_pool(work);
-
-               /*
-                * If we cannot get the last pool from @work directly,
-                * select the last CPU such that it avoids unnecessarily
-                * triggering non-reentrancy check in __queue_work().
-                */
-               lcpu = cpu;
-               if (pool)
-                       lcpu = pool->cpu;
-               if (lcpu == WORK_CPU_UNBOUND)
-                       lcpu = raw_smp_processor_id();
-       } else {
-               lcpu = WORK_CPU_UNBOUND;
-       }
-
-       set_work_cwq(work, get_cwq(lcpu, wq), 0);
-
+       dwork->wq = wq;
        dwork->cpu = cpu;
        timer->expires = jiffies + delay;
 
@@ -2944,8 +2919,7 @@ bool flush_delayed_work(struct delayed_work *dwork)
 {
        local_irq_disable();
        if (del_timer_sync(&dwork->timer))
-               __queue_work(dwork->cpu,
-                            get_work_cwq(&dwork->work)->wq, &dwork->work);
+               __queue_work(dwork->cpu, dwork->wq, &dwork->work);
        local_irq_enable();
        return flush_work(&dwork->work);
 }