workqueue: make GCWQ_DISASSOCIATED a pool flag
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)
Make GCWQ_DISASSOCIATED a pool flag POOL_DISASSOCIATED.  This patch
doesn't change locking - DISASSOCIATED on both pools of a CPU are set
or clear together while holding gcwq->lock.  It shouldn't cause any
functional difference.

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 6342515..1b8af92 100644 (file)
 enum {
        /*
         * global_cwq flags
+        */
+       GCWQ_FREEZING           = 1 << 1,       /* freeze in progress */
+
+       /*
+        * worker_pool flags
         *
-        * A bound gcwq is either associated or disassociated with its CPU.
+        * A bound pool is either associated or disassociated with its CPU.
         * While associated (!DISASSOCIATED), all workers are bound to the
         * CPU and none has %WORKER_UNBOUND set and concurrency management
         * is in effect.
         *
         * 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 gcwq behaves as an unbound one.
+        * be executing on any CPU.  The pool behaves as an unbound one.
         *
         * Note that DISASSOCIATED can be flipped only while holding
-        * assoc_mutex of all pools on the gcwq to avoid changing binding
-        * state while create_worker() is in progress.
+        * assoc_mutex to avoid changing binding state while
+        * create_worker() is in progress.
         */
-       GCWQ_DISASSOCIATED      = 1 << 0,       /* cpu can't serve workers */
-       GCWQ_FREEZING           = 1 << 1,       /* freeze in progress */
-
-       /* pool flags */
        POOL_MANAGE_WORKERS     = 1 << 0,       /* need to manage workers */
        POOL_MANAGING_WORKERS   = 1 << 1,       /* managing workers */
+       POOL_DISASSOCIATED      = 1 << 2,       /* cpu can't serve workers */
 
        /* worker flags */
        WORKER_STARTED          = 1 << 0,       /* started */
@@ -115,7 +117,7 @@ enum {
  * X: During normal operation, modification requires gcwq->lock and
  *    should be done only from local cpu.  Either disabling preemption
  *    on local cpu or grabbing gcwq->lock is enough for read access.
- *    If GCWQ_DISASSOCIATED is set, it's identical to L.
+ *    If POOL_DISASSOCIATED is set, it's identical to L.
  *
  * F: wq->flush_mutex protected.
  *
@@ -138,7 +140,7 @@ struct worker_pool {
        struct timer_list       idle_timer;     /* L: worker idle timeout */
        struct timer_list       mayday_timer;   /* L: SOS timer for workers */
 
-       struct mutex            assoc_mutex;    /* protect GCWQ_DISASSOCIATED */
+       struct mutex            assoc_mutex;    /* protect POOL_DISASSOCIATED */
        struct ida              worker_ida;     /* L: for worker IDs */
 };
 
@@ -439,9 +441,9 @@ static DEFINE_PER_CPU(struct global_cwq, global_cwq);
 static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_STD_WORKER_POOLS]);
 
 /*
- * Global cpu workqueue and nr_running counter for unbound gcwq.  The
- * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
- * workers have WORKER_UNBOUND set.
+ * Global cpu workqueue and nr_running counter for unbound gcwq.  The pools
+ * for online CPUs have POOL_DISASSOCIATED set, and all their workers have
+ * WORKER_UNBOUND set.
  */
 static struct global_cwq unbound_global_cwq;
 static atomic_t unbound_pool_nr_running[NR_STD_WORKER_POOLS] = {
@@ -1474,7 +1476,6 @@ EXPORT_SYMBOL_GPL(mod_delayed_work);
 static void worker_enter_idle(struct worker *worker)
 {
        struct worker_pool *pool = worker->pool;
-       struct global_cwq *gcwq = pool->gcwq;
 
        BUG_ON(worker->flags & WORKER_IDLE);
        BUG_ON(!list_empty(&worker->entry) &&
@@ -1497,7 +1498,7 @@ static void worker_enter_idle(struct worker *worker)
         * nr_running, the warning may trigger spuriously.  Check iff
         * unbind is not in progress.
         */
-       WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
+       WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
                     pool->nr_workers == pool->nr_idle &&
                     atomic_read(get_pool_nr_running(pool)));
 }
@@ -1538,7 +1539,7 @@ static void worker_leave_idle(struct worker *worker)
  * [dis]associated in the meantime.
  *
  * This function tries set_cpus_allowed() and locks gcwq and verifies the
- * binding against %GCWQ_DISASSOCIATED which is set during
+ * binding against %POOL_DISASSOCIATED which is set during
  * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
  * enters idle state or fetches works without dropping lock, it can
  * guarantee the scheduling requirement described in the first paragraph.
@@ -1554,7 +1555,8 @@ static void worker_leave_idle(struct worker *worker)
 static bool worker_maybe_bind_and_lock(struct worker *worker)
 __acquires(&gcwq->lock)
 {
-       struct global_cwq *gcwq = worker->pool->gcwq;
+       struct worker_pool *pool = worker->pool;
+       struct global_cwq *gcwq = pool->gcwq;
        struct task_struct *task = worker->task;
 
        while (true) {
@@ -1562,13 +1564,13 @@ __acquires(&gcwq->lock)
                 * The following call may fail, succeed or succeed
                 * without actually migrating the task to the cpu if
                 * it races with cpu hotunplug operation.  Verify
-                * against GCWQ_DISASSOCIATED.
+                * against POOL_DISASSOCIATED.
                 */
-               if (!(gcwq->flags & GCWQ_DISASSOCIATED))
+               if (!(pool->flags & POOL_DISASSOCIATED))
                        set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
 
                spin_lock_irq(&gcwq->lock);
-               if (gcwq->flags & GCWQ_DISASSOCIATED)
+               if (pool->flags & POOL_DISASSOCIATED)
                        return false;
                if (task_cpu(task) == gcwq->cpu &&
                    cpumask_equal(&current->cpus_allowed,
@@ -1766,14 +1768,14 @@ static struct worker *create_worker(struct worker_pool *pool)
 
        /*
         * Determine CPU binding of the new worker depending on
-        * %GCWQ_DISASSOCIATED.  The caller is responsible for ensuring the
+        * %POOL_DISASSOCIATED.  The caller is responsible for ensuring the
         * flag remains stable across this function.  See the comments
         * above the flag definition for details.
         *
         * As an unbound worker may later become a regular one if CPU comes
         * online, make sure every worker has %PF_THREAD_BOUND set.
         */
-       if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
+       if (!(pool->flags & POOL_DISASSOCIATED)) {
                kthread_bind(worker->task, gcwq->cpu);
        } else {
                worker->task->flags |= PF_THREAD_BOUND;
@@ -2134,10 +2136,10 @@ __acquires(&gcwq->lock)
        /*
         * Ensure we're on the correct CPU.  DISASSOCIATED test is
         * necessary to avoid spurious warnings from rescuers servicing the
-        * unbound or a disassociated gcwq.
+        * unbound or a disassociated pool.
         */
        WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
-                    !(gcwq->flags & GCWQ_DISASSOCIATED) &&
+                    !(pool->flags & POOL_DISASSOCIATED) &&
                     raw_smp_processor_id() != gcwq->cpu);
 
        /*
@@ -3472,7 +3474,7 @@ EXPORT_SYMBOL_GPL(work_busy);
  * gcwqs serve mix of short, long and very long running works making
  * blocked draining impractical.
  *
- * This is solved by allowing a gcwq to be disassociated from the CPU
+ * This is solved by allowing the pools to be disassociated from the CPU
  * running as an unbound one and allowing it to be reattached later if the
  * cpu comes back online.
  */
@@ -3522,7 +3524,8 @@ static void gcwq_unbind_fn(struct work_struct *work)
        for_each_busy_worker(worker, i, pos, gcwq)
                worker->flags |= WORKER_UNBOUND;
 
-       gcwq->flags |= GCWQ_DISASSOCIATED;
+       for_each_worker_pool(pool, gcwq)
+               pool->flags |= POOL_DISASSOCIATED;
 
        gcwq_release_assoc_and_unlock(gcwq);
 
@@ -3581,7 +3584,8 @@ static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
        case CPU_DOWN_FAILED:
        case CPU_ONLINE:
                gcwq_claim_assoc_and_lock(gcwq);
-               gcwq->flags &= ~GCWQ_DISASSOCIATED;
+               for_each_worker_pool(pool, gcwq)
+                       pool->flags &= ~POOL_DISASSOCIATED;
                rebind_workers(gcwq);
                gcwq_release_assoc_and_unlock(gcwq);
                break;
@@ -3806,12 +3810,12 @@ static int __init init_workqueues(void)
 
                spin_lock_init(&gcwq->lock);
                gcwq->cpu = cpu;
-               gcwq->flags |= GCWQ_DISASSOCIATED;
 
                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);
 
@@ -3832,12 +3836,12 @@ static int __init init_workqueues(void)
                struct global_cwq *gcwq = get_gcwq(cpu);
                struct worker_pool *pool;
 
-               if (cpu != WORK_CPU_UNBOUND)
-                       gcwq->flags &= ~GCWQ_DISASSOCIATED;
-
                for_each_worker_pool(pool, gcwq) {
                        struct worker *worker;
 
+                       if (cpu != WORK_CPU_UNBOUND)
+                               pool->flags &= ~POOL_DISASSOCIATED;
+
                        worker = create_worker(pool);
                        BUG_ON(!worker);
                        spin_lock_irq(&gcwq->lock);