workqueue: move busy_hash from global_cwq to worker_pool
authorTejun Heo <tj@kernel.org>
Thu, 24 Jan 2013 19:01:33 +0000 (11:01 -0800)
committerTejun Heo <tj@kernel.org>
Thu, 24 Jan 2013 19:01:33 +0000 (11:01 -0800)
There's no functional necessity for the two pools on the same CPU to
share the busy hash table.  It's also likely to be a bottleneck when
implementing pools with user-specified attributes.

This patch makes busy_hash per-pool.  The conversion is mostly
straight-forward.  Changes worth noting are,

* Large block of changes in rebind_workers() is moving the block
  inside for_each_worker_pool() as now there are separate hash tables
  for each pool.  This changes the order of operations but doesn't
  break anything.

* Thre for_each_worker_pool() loops in gcwq_unbind_fn() are combined
  into one.  This again changes the order of operaitons but doesn't
  break anything.

This is part of an effort to remove global_cwq and make worker_pool
the top level abstraction, which in turn will help implementing worker
pools with user-specified attributes.

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

index a4d7e3f0a87415abee8a3727a102a0ef463b8922..99c30116d2913ec37ab6e4384be5562a3c61c5cd 100644 (file)
@@ -137,6 +137,10 @@ struct worker_pool {
        struct timer_list       idle_timer;     /* L: worker idle timeout */
        struct timer_list       mayday_timer;   /* L: SOS timer for workers */
 
+       /* workers are chained either in busy_hash or idle_list */
+       DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
+                                               /* L: hash of busy workers */
+
        struct mutex            assoc_mutex;    /* protect POOL_DISASSOCIATED */
        struct ida              worker_ida;     /* L: for worker IDs */
 };
@@ -150,10 +154,6 @@ struct global_cwq {
        spinlock_t              lock;           /* the gcwq lock */
        unsigned int            cpu;            /* I: the associated cpu */
 
-       /* workers are chained either in busy_hash or pool idle_list */
-       DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
-                                               /* L: hash of busy workers */
-
        struct worker_pool      pools[NR_STD_WORKER_POOLS];
                                                /* normal and highpri pools */
 } ____cacheline_aligned_in_smp;
@@ -255,8 +255,8 @@ EXPORT_SYMBOL_GPL(system_freezable_wq);
        for ((pool) = &(gcwq)->pools[0];                                \
             (pool) < &(gcwq)->pools[NR_STD_WORKER_POOLS]; (pool)++)
 
-#define for_each_busy_worker(worker, i, pos, gcwq)                     \
-       hash_for_each(gcwq->busy_hash, i, pos, worker, hentry)
+#define for_each_busy_worker(worker, i, pos, pool)                     \
+       hash_for_each(pool->busy_hash, i, pos, worker, hentry)
 
 static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
                                  unsigned int sw)
@@ -892,11 +892,11 @@ static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 
 /**
  * find_worker_executing_work - find worker which is executing a work
- * @gcwq: gcwq of interest
+ * @pool: pool of interest
  * @work: work to find worker for
  *
- * Find a worker which is executing @work on @gcwq by searching
- * @gcwq->busy_hash which is keyed by the address of @work.  For a worker
+ * Find a worker which is executing @work on @pool by searching
+ * @pool->busy_hash which is keyed by the address of @work.  For a worker
  * to match, its current execution should match the address of @work and
  * its work function.  This is to avoid unwanted dependency between
  * unrelated work executions through a work item being recycled while still
@@ -924,13 +924,13 @@ static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
  * Pointer to worker which is executing @work if found, NULL
  * otherwise.
  */
-static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
+static struct worker *find_worker_executing_work(struct worker_pool *pool,
                                                 struct work_struct *work)
 {
        struct worker *worker;
        struct hlist_node *tmp;
 
-       hash_for_each_possible(gcwq->busy_hash, worker, tmp, hentry,
+       hash_for_each_possible(pool->busy_hash, worker, tmp, hentry,
                               (unsigned long)work)
                if (worker->current_work == work &&
                    worker->current_func == work->func)
@@ -1191,13 +1191,15 @@ static bool is_chained_work(struct workqueue_struct *wq)
        unsigned int cpu;
 
        for_each_gcwq_cpu(cpu) {
-               struct global_cwq *gcwq = get_gcwq(cpu);
+               struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
+               struct worker_pool *pool = cwq->pool;
+               struct global_cwq *gcwq = pool->gcwq;
                struct worker *worker;
                struct hlist_node *pos;
                int i;
 
                spin_lock_irqsave(&gcwq->lock, flags);
-               for_each_busy_worker(worker, i, pos, gcwq) {
+               for_each_busy_worker(worker, i, pos, pool) {
                        if (worker->task != current)
                                continue;
                        spin_unlock_irqrestore(&gcwq->lock, flags);
@@ -1238,7 +1240,7 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
 
        /* determine gcwq to use */
        if (!(wq->flags & WQ_UNBOUND)) {
-               struct global_cwq *last_gcwq;
+               struct worker_pool *last_pool;
 
                if (cpu == WORK_CPU_UNBOUND)
                        cpu = raw_smp_processor_id();
@@ -1250,14 +1252,15 @@ static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
                 * non-reentrancy.
                 */
                gcwq = get_gcwq(cpu);
-               last_gcwq = get_work_gcwq(work);
+               last_pool = get_work_pool(work);
 
-               if (last_gcwq && last_gcwq != gcwq) {
+               if (last_pool && last_pool->gcwq != gcwq) {
+                       struct global_cwq *last_gcwq = last_pool->gcwq;
                        struct worker *worker;
 
                        spin_lock(&last_gcwq->lock);
 
-                       worker = find_worker_executing_work(last_gcwq, work);
+                       worker = find_worker_executing_work(last_pool, work);
 
                        if (worker && worker->current_cwq->wq == wq)
                                gcwq = last_gcwq;
@@ -1722,31 +1725,32 @@ static void rebind_workers(struct global_cwq *gcwq)
                         */
                        wake_up_process(worker->task);
                }
-       }
 
-       /* rebind busy workers */
-       for_each_busy_worker(worker, i, pos, gcwq) {
-               struct work_struct *rebind_work = &worker->rebind_work;
-               struct workqueue_struct *wq;
+               /* rebind busy workers */
+               for_each_busy_worker(worker, i, pos, pool) {
+                       struct work_struct *rebind_work = &worker->rebind_work;
+                       struct workqueue_struct *wq;
 
-               if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
-                                    work_data_bits(rebind_work)))
-                       continue;
+                       if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
+                                            work_data_bits(rebind_work)))
+                               continue;
 
-               debug_work_activate(rebind_work);
+                       debug_work_activate(rebind_work);
 
-               /*
-                * wq doesn't really matter but let's keep @worker->pool
-                * and @cwq->pool consistent for sanity.
-                */
-               if (std_worker_pool_pri(worker->pool))
-                       wq = system_highpri_wq;
-               else
-                       wq = system_wq;
-
-               insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
-                       worker->scheduled.next,
-                       work_color_to_flags(WORK_NO_COLOR));
+                       /*
+                        * wq doesn't really matter but let's keep
+                        * @worker->pool and @cwq->pool consistent for
+                        * sanity.
+                        */
+                       if (std_worker_pool_pri(worker->pool))
+                               wq = system_highpri_wq;
+                       else
+                               wq = system_wq;
+
+                       insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
+                                   worker->scheduled.next,
+                                   work_color_to_flags(WORK_NO_COLOR));
+               }
        }
 }
 
@@ -2197,7 +2201,7 @@ __acquires(&gcwq->lock)
         * already processing the work.  If so, defer the work to the
         * currently executing one.
         */
-       collision = find_worker_executing_work(gcwq, work);
+       collision = find_worker_executing_work(pool, work);
        if (unlikely(collision)) {
                move_linked_works(work, &collision->scheduled, NULL);
                return;
@@ -2205,7 +2209,7 @@ __acquires(&gcwq->lock)
 
        /* claim and dequeue */
        debug_work_deactivate(work);
-       hash_add(gcwq->busy_hash, &worker->hentry, (unsigned long)work);
+       hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
        worker->current_work = work;
        worker->current_func = work->func;
        worker->current_cwq = cwq;
@@ -2833,13 +2837,15 @@ EXPORT_SYMBOL_GPL(drain_workqueue);
 static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
 {
        struct worker *worker = NULL;
+       struct worker_pool *pool;
        struct global_cwq *gcwq;
        struct cpu_workqueue_struct *cwq;
 
        might_sleep();
-       gcwq = get_work_gcwq(work);
-       if (!gcwq)
+       pool = get_work_pool(work);
+       if (!pool)
                return false;
+       gcwq = pool->gcwq;
 
        spin_lock_irq(&gcwq->lock);
        if (!list_empty(&work->entry)) {
@@ -2853,7 +2859,7 @@ static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
                if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
                        goto already_gone;
        } else {
-               worker = find_worker_executing_work(gcwq, work);
+               worker = find_worker_executing_work(pool, work);
                if (!worker)
                        goto already_gone;
                cwq = worker->current_cwq;
@@ -3482,18 +3488,20 @@ EXPORT_SYMBOL_GPL(workqueue_congested);
  */
 unsigned int work_busy(struct work_struct *work)
 {
-       struct global_cwq *gcwq = get_work_gcwq(work);
+       struct worker_pool *pool = get_work_pool(work);
+       struct global_cwq *gcwq;
        unsigned long flags;
        unsigned int ret = 0;
 
-       if (!gcwq)
+       if (!pool)
                return 0;
+       gcwq = pool->gcwq;
 
        spin_lock_irqsave(&gcwq->lock, flags);
 
        if (work_pending(work))
                ret |= WORK_BUSY_PENDING;
-       if (find_worker_executing_work(gcwq, work))
+       if (find_worker_executing_work(pool, work))
                ret |= WORK_BUSY_RUNNING;
 
        spin_unlock_irqrestore(&gcwq->lock, flags);
@@ -3555,15 +3563,15 @@ static void gcwq_unbind_fn(struct work_struct *work)
         * ones which are still executing works from before the last CPU
         * down must be on the cpu.  After this, they may become diasporas.
         */
-       for_each_worker_pool(pool, gcwq)
+       for_each_worker_pool(pool, gcwq) {
                list_for_each_entry(worker, &pool->idle_list, entry)
                        worker->flags |= WORKER_UNBOUND;
 
-       for_each_busy_worker(worker, i, pos, gcwq)
-               worker->flags |= WORKER_UNBOUND;
+               for_each_busy_worker(worker, i, pos, pool)
+                       worker->flags |= WORKER_UNBOUND;
 
-       for_each_worker_pool(pool, gcwq)
                pool->flags |= POOL_DISASSOCIATED;
+       }
 
        gcwq_release_assoc_and_unlock(gcwq);
 
@@ -3854,13 +3862,12 @@ static int __init init_workqueues(void)
                spin_lock_init(&gcwq->lock);
                gcwq->cpu = cpu;
 
-               hash_init(gcwq->busy_hash);
-
                for_each_worker_pool(pool, gcwq) {
                        pool->gcwq = gcwq;
                        pool->flags |= POOL_DISASSOCIATED;
                        INIT_LIST_HEAD(&pool->worklist);
                        INIT_LIST_HEAD(&pool->idle_list);
+                       hash_init(pool->busy_hash);
 
                        init_timer_deferrable(&pool->idle_timer);
                        pool->idle_timer.function = idle_worker_timeout;