workqueue: drop @bind from create_worker()
authorTejun Heo <tj@kernel.org>
Tue, 17 Jul 2012 19:39:27 +0000 (12:39 -0700)
committerTejun Heo <tj@kernel.org>
Tue, 17 Jul 2012 19:39:27 +0000 (12:39 -0700)
Currently, create_worker()'s callers are responsible for deciding
whether the newly created worker should be bound to the associated CPU
and create_worker() sets WORKER_UNBOUND only for the workers for the
unbound global_cwq.  Creation during normal operation is always via
maybe_create_worker() and @bind is true.  For workers created during
hotplug, @bind is false.

Normal operation path is planned to be used even while the CPU is
going through hotplug operations or offline and this static decision
won't work.

Drop @bind from create_worker() and decide whether to bind by looking
at GCWQ_DISASSOCIATED.  create_worker() will also set WORKER_UNBOUND
autmatically if disassociated.  To avoid flipping GCWQ_DISASSOCIATED
while create_worker() is in progress, the flag is now allowed to be
changed only while holding all manager_mutexes on the global_cwq.

This requires that GCWQ_DISASSOCIATED is not cleared behind trustee's
back.  CPU_ONLINE no longer clears DISASSOCIATED before flushing
trustee, which clears DISASSOCIATED before rebinding remaining workers
if asked to release.  For cases where trustee isn't around, CPU_ONLINE
clears DISASSOCIATED after flushing trustee.  Also, now, first_idle
has UNBOUND set on creation which is explicitly cleared by CPU_ONLINE
while binding it.  These convolutions will soon be removed by further
simplification of CPU hotplug path.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: "Rafael J. Wysocki" <rjw@sisk.pl>
kernel/workqueue.c

index f7a0069..e1d05e5 100644 (file)
 #include "workqueue_sched.h"
 
 enum {
-       /* global_cwq flags */
+       /*
+        * global_cwq flags
+        *
+        * A bound gcwq 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.
+        *
+        * Note that DISASSOCIATED can be flipped only while holding
+        * managership of all pools on the gcwq 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 */
 
@@ -1334,7 +1349,6 @@ static struct worker *alloc_worker(void)
 /**
  * create_worker - create a new workqueue worker
  * @pool: pool the new worker will belong to
- * @bind: whether to set affinity to @cpu or not
  *
  * Create a new worker which is bound to @pool.  The returned worker
  * can be started by calling start_worker() or destroyed using
@@ -1346,10 +1360,9 @@ static struct worker *alloc_worker(void)
  * RETURNS:
  * Pointer to the newly created worker.
  */
-static struct worker *create_worker(struct worker_pool *pool, bool bind)
+static struct worker *create_worker(struct worker_pool *pool)
 {
        struct global_cwq *gcwq = pool->gcwq;
-       bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
        const char *pri = worker_pool_pri(pool) ? "H" : "";
        struct worker *worker = NULL;
        int id = -1;
@@ -1370,7 +1383,7 @@ static struct worker *create_worker(struct worker_pool *pool, bool bind)
        worker->pool = pool;
        worker->id = id;
 
-       if (!on_unbound_cpu)
+       if (gcwq->cpu != WORK_CPU_UNBOUND)
                worker->task = kthread_create_on_node(worker_thread,
                                        worker, cpu_to_node(gcwq->cpu),
                                        "kworker/%u:%d%s", gcwq->cpu, id, pri);
@@ -1384,15 +1397,19 @@ static struct worker *create_worker(struct worker_pool *pool, bool bind)
                set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
 
        /*
-        * An unbound worker will become a regular one if CPU comes online
-        * later on.  Make sure every worker has PF_THREAD_BOUND set.
+        * Determine CPU binding of the new worker depending on
+        * %GCWQ_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 (bind && !on_unbound_cpu)
+       if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
                kthread_bind(worker->task, gcwq->cpu);
-       else {
+       else {
                worker->task->flags |= PF_THREAD_BOUND;
-               if (on_unbound_cpu)
-                       worker->flags |= WORKER_UNBOUND;
+               worker->flags |= WORKER_UNBOUND;
        }
 
        return worker;
@@ -1568,7 +1585,7 @@ restart:
        while (true) {
                struct worker *worker;
 
-               worker = create_worker(pool, true);
+               worker = create_worker(pool);
                if (worker) {
                        del_timer_sync(&pool->mayday_timer);
                        spin_lock_irq(&gcwq->lock);
@@ -3420,12 +3437,10 @@ static int __cpuinit trustee_thread(void *__gcwq)
 
                        if (need_to_create_worker(pool)) {
                                spin_unlock_irq(&gcwq->lock);
-                               worker = create_worker(pool, false);
+                               worker = create_worker(pool);
                                spin_lock_irq(&gcwq->lock);
-                               if (worker) {
-                                       worker->flags |= WORKER_UNBOUND;
+                               if (worker)
                                        start_worker(worker);
-                               }
                        }
                }
 
@@ -3463,6 +3478,10 @@ static int __cpuinit trustee_thread(void *__gcwq)
        for_each_worker_pool(pool, gcwq)
                WARN_ON(!list_empty(&pool->idle_list));
 
+       /* if we're reassociating, clear DISASSOCIATED */
+       if (gcwq->trustee_state == TRUSTEE_RELEASE)
+               gcwq->flags &= ~GCWQ_DISASSOCIATED;
+
        for_each_busy_worker(worker, i, pos, gcwq) {
                struct work_struct *rebind_work = &worker->rebind_work;
 
@@ -3546,7 +3565,7 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
                i = 0;
                for_each_worker_pool(pool, gcwq) {
                        BUG_ON(pool->first_idle);
-                       new_workers[i] = create_worker(pool, false);
+                       new_workers[i] = create_worker(pool);
                        if (!new_workers[i++])
                                goto err_destroy;
                }
@@ -3584,7 +3603,6 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
 
        case CPU_DOWN_FAILED:
        case CPU_ONLINE:
-               gcwq->flags &= ~GCWQ_DISASSOCIATED;
                if (gcwq->trustee_state != TRUSTEE_DONE) {
                        gcwq->trustee_state = TRUSTEE_RELEASE;
                        wake_up_process(gcwq->trustee);
@@ -3592,6 +3610,13 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
                }
 
                /*
+                * Either DISASSOCIATED is already cleared or no worker is
+                * left on the gcwq.  Safe to clear DISASSOCIATED without
+                * claiming managers.
+                */
+               gcwq->flags &= ~GCWQ_DISASSOCIATED;
+
+               /*
                 * Trustee is done and there might be no worker left.
                 * Put the first_idle in and request a real manager to
                 * take a look.
@@ -3601,6 +3626,7 @@ static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
                        kthread_bind(pool->first_idle->task, cpu);
                        spin_lock_irq(&gcwq->lock);
                        pool->flags |= POOL_MANAGE_WORKERS;
+                       pool->first_idle->flags &= ~WORKER_UNBOUND;
                        start_worker(pool->first_idle);
                        pool->first_idle = NULL;
                }
@@ -3899,7 +3925,7 @@ static int __init init_workqueues(void)
                for_each_worker_pool(pool, gcwq) {
                        struct worker *worker;
 
-                       worker = create_worker(pool, true);
+                       worker = create_worker(pool);
                        BUG_ON(!worker);
                        spin_lock_irq(&gcwq->lock);
                        start_worker(worker);