workqueue: map an unbound workqueues to multiple per-node pool_workqueues
authorTejun Heo <tj@kernel.org>
Mon, 1 Apr 2013 18:23:35 +0000 (11:23 -0700)
committerTejun Heo <tj@kernel.org>
Mon, 1 Apr 2013 18:23:35 +0000 (11:23 -0700)
Currently, an unbound workqueue has only one "current" pool_workqueue
associated with it.  It may have multple pool_workqueues but only the
first pool_workqueue servies new work items.  For NUMA affinity, we
want to change this so that there are multiple current pool_workqueues
serving different NUMA nodes.

Introduce workqueue->numa_pwq_tbl[] which is indexed by NUMA node and
points to the pool_workqueue to use for each possible node.  This
replaces first_pwq() in __queue_work() and workqueue_congested().

numa_pwq_tbl[] is currently initialized to point to the same
pool_workqueue as first_pwq() so this patch doesn't make any behavior
changes.

v2: Use rcu_dereference_raw() in unbound_pwq_by_node() as the function
    may be called only with wq->mutex held.

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

index 4c53fa216732a0f468b764cdeefc45bf736197ab..170226a24da89d040a42736649100ef11e78a535 100644 (file)
@@ -257,6 +257,7 @@ struct workqueue_struct {
        /* hot fields used during command issue, aligned to cacheline */
        unsigned int            flags ____cacheline_aligned; /* WQ: WQ_* flags */
        struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
+       struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
 };
 
 static struct kmem_cache *pwq_cache;
@@ -525,6 +526,22 @@ static struct pool_workqueue *first_pwq(struct workqueue_struct *wq)
                                      pwqs_node);
 }
 
+/**
+ * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
+ * @wq: the target workqueue
+ * @node: the node ID
+ *
+ * This must be called either with pwq_lock held or sched RCU read locked.
+ * If the pwq needs to be used beyond the locking in effect, the caller is
+ * responsible for guaranteeing that the pwq stays online.
+ */
+static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
+                                                 int node)
+{
+       assert_rcu_or_wq_mutex(wq);
+       return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
+}
+
 static unsigned int work_color_to_flags(int color)
 {
        return color << WORK_STRUCT_COLOR_SHIFT;
@@ -1278,14 +1295,14 @@ static void __queue_work(int cpu, struct workqueue_struct *wq,
            WARN_ON_ONCE(!is_chained_work(wq)))
                return;
 retry:
+       if (req_cpu == WORK_CPU_UNBOUND)
+               cpu = raw_smp_processor_id();
+
        /* pwq which will be used unless @work is executing elsewhere */
-       if (!(wq->flags & WQ_UNBOUND)) {
-               if (cpu == WORK_CPU_UNBOUND)
-                       cpu = raw_smp_processor_id();
+       if (!(wq->flags & WQ_UNBOUND))
                pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
-       } else {
-               pwq = first_pwq(wq);
-       }
+       else
+               pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
 
        /*
         * If @work was previously on a different pool, it might still be
@@ -1315,8 +1332,8 @@ retry:
         * pwq is determined and locked.  For unbound pools, we could have
         * raced with pwq release and it could already be dead.  If its
         * refcnt is zero, repeat pwq selection.  Note that pwqs never die
-        * without another pwq replacing it as the first pwq or while a
-        * work item is executing on it, so the retying is guaranteed to
+        * without another pwq replacing it in the numa_pwq_tbl or while
+        * work items are executing on it, so the retrying is guaranteed to
         * make forward-progress.
         */
        if (unlikely(!pwq->refcnt)) {
@@ -3614,6 +3631,8 @@ static void init_and_link_pwq(struct pool_workqueue *pwq,
                              struct worker_pool *pool,
                              struct pool_workqueue **p_last_pwq)
 {
+       int node;
+
        BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
 
        pwq->pool = pool;
@@ -3640,8 +3659,11 @@ static void init_and_link_pwq(struct pool_workqueue *pwq,
        /* link in @pwq */
        list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
 
-       if (wq->flags & WQ_UNBOUND)
+       if (wq->flags & WQ_UNBOUND) {
                copy_workqueue_attrs(wq->unbound_attrs, pool->attrs);
+               for_each_node(node)
+                       rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
+       }
 
        mutex_unlock(&wq->mutex);
 }
@@ -3761,12 +3783,16 @@ struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
                                               struct lock_class_key *key,
                                               const char *lock_name, ...)
 {
+       size_t tbl_size = 0;
        va_list args;
        struct workqueue_struct *wq;
        struct pool_workqueue *pwq;
 
        /* allocate wq and format name */
-       wq = kzalloc(sizeof(*wq), GFP_KERNEL);
+       if (flags & WQ_UNBOUND)
+               tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
+
+       wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
        if (!wq)
                return NULL;
 
@@ -3994,7 +4020,7 @@ bool workqueue_congested(int cpu, struct workqueue_struct *wq)
        if (!(wq->flags & WQ_UNBOUND))
                pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
        else
-               pwq = first_pwq(wq);
+               pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
 
        ret = !list_empty(&pwq->delayed_works);
        rcu_read_unlock_sched();