workqueue: remove WQ_SINGLE_CPU and use WQ_UNBOUND instead
Tejun Heo [Fri, 2 Jul 2010 08:03:51 +0000 (10:03 +0200)]
WQ_SINGLE_CPU combined with @max_active of 1 is used to achieve full
ordering among works queued to a workqueue.  The same can be achieved
using WQ_UNBOUND as unbound workqueues always use the gcwq for
WORK_CPU_UNBOUND.  As @max_active is always one and benefits from cpu
locality isn't accessible anyway, serving them with unbound workqueues
should be fine.

Drop WQ_SINGLE_CPU support and use WQ_UNBOUND instead.  Note that most
single thread workqueue users will be converted to use multithread or
non-reentrant instead and only the ones which require strict ordering
will keep using WQ_UNBOUND + @max_active of 1.

Signed-off-by: Tejun Heo <tj@kernel.org>

include/linux/workqueue.h
kernel/workqueue.c

index 67ce734..d74a529 100644 (file)
@@ -233,12 +233,11 @@ static inline unsigned int work_static(struct work_struct *work) { return 0; }
 
 enum {
        WQ_NON_REENTRANT        = 1 << 0, /* guarantee non-reentrance */
-       WQ_SINGLE_CPU           = 1 << 1, /* only single cpu at a time */
+       WQ_UNBOUND              = 1 << 1, /* not bound to any cpu */
        WQ_FREEZEABLE           = 1 << 2, /* freeze during suspend */
        WQ_RESCUER              = 1 << 3, /* has an rescue worker */
        WQ_HIGHPRI              = 1 << 4, /* high priority */
        WQ_CPU_INTENSIVE        = 1 << 5, /* cpu instensive workqueue */
-       WQ_UNBOUND              = 1 << 6, /* not bound to any cpu */
 
        WQ_MAX_ACTIVE           = 512,    /* I like 512, better ideas? */
        WQ_MAX_UNBOUND_PER_CPU  = 4,      /* 4 * #cpus for unbound wq */
@@ -300,9 +299,9 @@ __alloc_workqueue_key(const char *name, unsigned int flags, int max_active,
 #define create_workqueue(name)                                 \
        alloc_workqueue((name), WQ_RESCUER, 1)
 #define create_freezeable_workqueue(name)                      \
-       alloc_workqueue((name), WQ_FREEZEABLE | WQ_SINGLE_CPU | WQ_RESCUER, 1)
+       alloc_workqueue((name), WQ_FREEZEABLE | WQ_UNBOUND | WQ_RESCUER, 1)
 #define create_singlethread_workqueue(name)                    \
-       alloc_workqueue((name), WQ_SINGLE_CPU | WQ_RESCUER, 1)
+       alloc_workqueue((name), WQ_UNBOUND | WQ_RESCUER, 1)
 
 extern void destroy_workqueue(struct workqueue_struct *wq);
 
index 4608563..20d6237 100644 (file)
@@ -206,8 +206,6 @@ struct workqueue_struct {
        struct list_head        flusher_queue;  /* F: flush waiters */
        struct list_head        flusher_overflow; /* F: flush overflow list */
 
-       unsigned long           single_cpu;     /* cpu for single cpu wq */
-
        cpumask_var_t           mayday_mask;    /* cpus requesting rescue */
        struct worker           *rescuer;       /* I: rescue worker */
 
@@ -889,34 +887,6 @@ static void insert_work(struct cpu_workqueue_struct *cwq,
                wake_up_worker(gcwq);
 }
 
-/**
- * cwq_unbind_single_cpu - unbind cwq from single cpu workqueue processing
- * @cwq: cwq to unbind
- *
- * Try to unbind @cwq from single cpu workqueue processing.  If
- * @cwq->wq is frozen, unbind is delayed till the workqueue is thawed.
- *
- * CONTEXT:
- * spin_lock_irq(gcwq->lock).
- */
-static void cwq_unbind_single_cpu(struct cpu_workqueue_struct *cwq)
-{
-       struct workqueue_struct *wq = cwq->wq;
-       struct global_cwq *gcwq = cwq->gcwq;
-
-       BUG_ON(wq->single_cpu != gcwq->cpu);
-       /*
-        * Unbind from workqueue if @cwq is not frozen.  If frozen,
-        * thaw_workqueues() will either restart processing on this
-        * cpu or unbind if empty.  This keeps works queued while
-        * frozen fully ordered and flushable.
-        */
-       if (likely(!(gcwq->flags & GCWQ_FREEZING))) {
-               smp_wmb();      /* paired with cmpxchg() in __queue_work() */
-               wq->single_cpu = WORK_CPU_NONE;
-       }
-}
-
 static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
                         struct work_struct *work)
 {
@@ -924,20 +894,16 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
        struct cpu_workqueue_struct *cwq;
        struct list_head *worklist;
        unsigned long flags;
-       bool arbitrate;
 
        debug_work_activate(work);
 
-       if (unlikely(cpu == WORK_CPU_UNBOUND))
-               cpu = raw_smp_processor_id();
-
-       /*
-        * Determine gcwq to use.  SINGLE_CPU is inherently
-        * NON_REENTRANT, so test it first.
-        */
-       if (!(wq->flags & (WQ_SINGLE_CPU | WQ_UNBOUND))) {
+       /* determine gcwq to use */
+       if (!(wq->flags & WQ_UNBOUND)) {
                struct global_cwq *last_gcwq;
 
+               if (unlikely(cpu == WORK_CPU_UNBOUND))
+                       cpu = raw_smp_processor_id();
+
                /*
                 * It's multi cpu.  If @wq is non-reentrant and @work
                 * was previously on a different cpu, it might still
@@ -962,38 +928,6 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
                        }
                } else
                        spin_lock_irqsave(&gcwq->lock, flags);
-       } else if (!(wq->flags & WQ_UNBOUND)) {
-               unsigned int req_cpu = cpu;
-
-               /*
-                * It's a bit more complex for single cpu workqueues.
-                * We first need to determine which cpu is going to be
-                * used.  If no cpu is currently serving this
-                * workqueue, arbitrate using atomic accesses to
-                * wq->single_cpu; otherwise, use the current one.
-                */
-       retry:
-               cpu = wq->single_cpu;
-               arbitrate = cpu == WORK_CPU_NONE;
-               if (arbitrate)
-                       cpu = req_cpu;
-
-               gcwq = get_gcwq(cpu);
-               spin_lock_irqsave(&gcwq->lock, flags);
-
-               /*
-                * The following cmpxchg() is a full barrier paired
-                * with smp_wmb() in cwq_unbind_single_cpu() and
-                * guarantees that all changes to wq->st_* fields are
-                * visible on the new cpu after this point.
-                */
-               if (arbitrate)
-                       cmpxchg(&wq->single_cpu, WORK_CPU_NONE, cpu);
-
-               if (unlikely(wq->single_cpu != cpu)) {
-                       spin_unlock_irqrestore(&gcwq->lock, flags);
-                       goto retry;
-               }
        } else {
                gcwq = get_gcwq(WORK_CPU_UNBOUND);
                spin_lock_irqsave(&gcwq->lock, flags);
@@ -1105,19 +1039,30 @@ int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
        struct work_struct *work = &dwork->work;
 
        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
-               struct global_cwq *gcwq = get_work_gcwq(work);
-               unsigned int lcpu = gcwq ? gcwq->cpu : raw_smp_processor_id();
+               unsigned int lcpu;
 
                BUG_ON(timer_pending(timer));
                BUG_ON(!list_empty(&work->entry));
 
                timer_stats_timer_set_start_info(&dwork->timer);
+
                /*
                 * This stores cwq for the moment, for the timer_fn.
                 * Note that the work's gcwq is preserved to allow
                 * reentrance detection for delayed works.
                 */
+               if (!(wq->flags & WQ_UNBOUND)) {
+                       struct global_cwq *gcwq = get_work_gcwq(work);
+
+                       if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
+                               lcpu = gcwq->cpu;
+                       else
+                               lcpu = raw_smp_processor_id();
+               } else
+                       lcpu = WORK_CPU_UNBOUND;
+
                set_work_cwq(work, get_cwq(lcpu, wq), 0);
+
                timer->expires = jiffies + delay;
                timer->data = (unsigned long)dwork;
                timer->function = delayed_work_timer_fn;
@@ -1696,9 +1641,6 @@ static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
                /* one down, submit a delayed one */
                if (cwq->nr_active < cwq->max_active)
                        cwq_activate_first_delayed(cwq);
-       } else if (!cwq->nr_active && cwq->wq->flags & WQ_SINGLE_CPU) {
-               /* this was the last work, unbind from single cpu */
-               cwq_unbind_single_cpu(cwq);
        }
 
        /* is flush in progress and are we at the flushing tip? */
@@ -2751,7 +2693,6 @@ struct workqueue_struct *__alloc_workqueue_key(const char *name,
        atomic_set(&wq->nr_cwqs_to_flush, 0);
        INIT_LIST_HEAD(&wq->flusher_queue);
        INIT_LIST_HEAD(&wq->flusher_overflow);
-       wq->single_cpu = WORK_CPU_NONE;
 
        wq->name = name;
        lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
@@ -3513,11 +3454,6 @@ void thaw_workqueues(void)
                        while (!list_empty(&cwq->delayed_works) &&
                               cwq->nr_active < cwq->max_active)
                                cwq_activate_first_delayed(cwq);
-
-                       /* perform delayed unbind from single cpu if empty */
-                       if (wq->single_cpu == gcwq->cpu &&
-                           !cwq->nr_active && list_empty(&cwq->delayed_works))
-                               cwq_unbind_single_cpu(cwq);
                }
 
                wake_up_worker(gcwq);