History

workqueue: reimplement work flushing using linked works

A work is linked to the next one by having WORK_STRUCT_LINKED bit set and these links can be chained. When a linked work is dispatched to a worker, all linked works are dispatched to the worker's newly added ->scheduled queue and processed back-to-back. Currently, as there's only single worker per cwq, having linked works doesn't make any visible behavior difference. This change is to prepare for multiple shared workers per cpu.

Tejun Heo committed 16 years ago

Commit affee4b1 parent c34056a2 files · +134 −22Browse files

Showing the changes to kernel (1 of 2 files). Show all

Unified · Split

kernel/workqueue.c+131−21
@@ -51,6 +51,7 @@ struct cpu_workqueue_struct; · 50 unmodified lines
5151
5252 struct worker {
5353 struct work_struct *current_work; /* L: work being processed */
54+ struct list_head scheduled; /* L: scheduled works */
5455 struct task_struct *task; /* I: worker task */
5556 struct cpu_workqueue_struct *cwq; /* I: the associated cwq */
5657 int id; /* I: worker id */
@@ -445,6 +446,8 @@ static struct worker *alloc_worker(void) · 388 unmodified lines
445446 struct worker *worker;
446447
447448 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
449+ if (worker)
450+ INIT_LIST_HEAD(&worker->scheduled);
448451 return worker;
449452 }
450453
@@ -530,6 +533,7 @@ static void destroy_worker(struct worker *worker) · 79 unmodified lines
530533
531534 /* sanity check frenzy */
532535 BUG_ON(worker->current_work);
536+ BUG_ON(!list_empty(&worker->scheduled));
533537
534538 kthread_stop(worker->task);
535539 kfree(worker);
@@ -539,6 +543,47 @@ static void destroy_worker(struct worker *worker) · 3 unmodified lines
539543 spin_unlock(&workqueue_lock);
540544 }
541545
546+/**
547+ * move_linked_works - move linked works to a list
548+ * @work: start of series of works to be scheduled
549+ * @head: target list to append @work to
550+ * @nextp: out paramter for nested worklist walking
551+ *
552+ * Schedule linked works starting from @work to @head. Work series to
553+ * be scheduled starts at @work and includes any consecutive work with
554+ * WORK_STRUCT_LINKED set in its predecessor.
555+ *
556+ * If @nextp is not NULL, it's updated to point to the next work of
557+ * the last scheduled work. This allows move_linked_works() to be
558+ * nested inside outer list_for_each_entry_safe().
559+ *
560+ * CONTEXT:
561+ * spin_lock_irq(cwq->lock).
562+ */
563+static void move_linked_works(struct work_struct *work, struct list_head *head,
564+ struct work_struct **nextp)
565+{
566+ struct work_struct *n;
567+
568+ /*
569+ * Linked worklist will always end before the end of the list,
570+ * use NULL for list head.
571+ */
572+ list_for_each_entry_safe_from(work, n, NULL, entry) {
573+ list_move_tail(&work->entry, head);
574+ if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
575+ break;
576+ }
577+
578+ /*
579+ * If we're already inside safe list traversal and have moved
580+ * multiple works to the scheduled queue, the next position
581+ * needs to be updated.
582+ */
583+ if (nextp)
584+ *nextp = n;
585+}
586+
542587 /**
543588 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
544589 * @cwq: cwq of interest
@@ -639,17 +684,25 @@ static void process_one_work(struct worker *worker, struct work_struct *work) · 94 unmodified lines
639684 cwq_dec_nr_in_flight(cwq, work_color);
640685 }
641686
642−static void run_workqueue(struct worker *worker)
687+/**
688+ * process_scheduled_works - process scheduled works
689+ * @worker: self
690+ *
691+ * Process all scheduled works. Please note that the scheduled list
692+ * may change while processing a work, so this function repeatedly
693+ * fetches a work from the top and executes it.
694+ *
695+ * CONTEXT:
696+ * spin_lock_irq(cwq->lock) which may be released and regrabbed
697+ * multiple times.
698+ */
699+static void process_scheduled_works(struct worker *worker)
643700 {
644− struct cpu_workqueue_struct *cwq = worker->cwq;
645−
646− spin_lock_irq(&cwq->lock);
647− while (!list_empty(&cwq->worklist)) {
648− struct work_struct *work = list_entry(cwq->worklist.next,
701+ while (!list_empty(&worker->scheduled)) {
702+ struct work_struct *work = list_first_entry(&worker->scheduled,
649703 struct work_struct, entry);
650704 process_one_work(worker, work);
651705 }
652− spin_unlock_irq(&cwq->lock);
653706 }
654707
655708 /**
@@ -684,7 +737,28 @@ static int worker_thread(void *__worker) · 28 unmodified lines
684737 get_cpu_mask(cwq->cpu))))
685738 set_cpus_allowed_ptr(worker->task,
686739 get_cpu_mask(cwq->cpu));
687− run_workqueue(worker);
740+
741+ spin_lock_irq(&cwq->lock);
742+
743+ while (!list_empty(&cwq->worklist)) {
744+ struct work_struct *work =
745+ list_first_entry(&cwq->worklist,
746+ struct work_struct, entry);
747+
748+ if (likely(!(*work_data_bits(work) &
749+ WORK_STRUCT_LINKED))) {
750+ /* optimization path, not strictly necessary */
751+ process_one_work(worker, work);
752+ if (unlikely(!list_empty(&worker->scheduled)))
753+ process_scheduled_works(worker);
754+ } else {
755+ move_linked_works(work, &worker->scheduled,
756+ NULL);
757+ process_scheduled_works(worker);
758+ }
759+ }
760+
761+ spin_unlock_irq(&cwq->lock);
688762 }
689763
690764 return 0;
@@ -705,16 +779,33 @@ static void wq_barrier_func(struct work_struct *work) · 14 unmodified lines
705779 * insert_wq_barrier - insert a barrier work
706780 * @cwq: cwq to insert barrier into
707781 * @barr: wq_barrier to insert
708− * @head: insertion point
782+ * @target: target work to attach @barr to
783+ * @worker: worker currently executing @target, NULL if @target is not executing
709784 *
710− * Insert barrier @barr into @cwq before @head.
785+ * @barr is linked to @target such that @barr is completed only after
786+ * @target finishes execution. Please note that the ordering
787+ * guarantee is observed only with respect to @target and on the local
788+ * cpu.
789+ *
790+ * Currently, a queued barrier can't be canceled. This is because
791+ * try_to_grab_pending() can't determine whether the work to be
792+ * grabbed is at the head of the queue and thus can't clear LINKED
793+ * flag of the previous work while there must be a valid next work
794+ * after a work with LINKED flag set.
795+ *
796+ * Note that when @worker is non-NULL, @target may be modified
797+ * underneath us, so we can't reliably determine cwq from @target.
711798 *
712799 * CONTEXT:
713800 * spin_lock_irq(cwq->lock).
714801 */
715802 static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
716− struct wq_barrier *barr, struct list_head *head)
803+ struct wq_barrier *barr,
804+ struct work_struct *target, struct worker *worker)
717805 {
806+ struct list_head *head;
807+ unsigned int linked = 0;
808+
718809 /*
719810 * debugobject calls are safe here even with cwq->lock locked
720811 * as we know for sure that this will not trigger any of the
@@ -725,8 +816,24 @@ static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, · 4 unmodified lines
725816 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
726817 init_completion(&barr->done);
727818
819+ /*
820+ * If @target is currently being executed, schedule the
821+ * barrier to the worker; otherwise, put it after @target.
822+ */
823+ if (worker)
824+ head = worker->scheduled.next;
825+ else {
826+ unsigned long *bits = work_data_bits(target);
827+
828+ head = target->entry.next;
829+ /* there can already be other linked works, inherit and set */
830+ linked = *bits & WORK_STRUCT_LINKED;
831+ __set_bit(WORK_STRUCT_LINKED_BIT, bits);
832+ }
833+
728834 debug_work_activate(&barr->work);
729− insert_work(cwq, &barr->work, head, work_color_to_flags(WORK_NO_COLOR));
835+ insert_work(cwq, &barr->work, head,
836+ work_color_to_flags(WORK_NO_COLOR) | linked);
730837 }
731838
732839 /**
@@ -964,8 +1071,8 @@ EXPORT_SYMBOL_GPL(flush_workqueue); · 231 unmodified lines
9641071 */
9651072 int flush_work(struct work_struct *work)
9661073 {
1074+ struct worker *worker = NULL;
9671075 struct cpu_workqueue_struct *cwq;
968− struct list_head *prev;
9691076 struct wq_barrier barr;
9701077
9711078 might_sleep();
@@ -985,14 +1092,14 @@ int flush_work(struct work_struct *work) · 13 unmodified lines
9851092 smp_rmb();
9861093 if (unlikely(cwq != get_wq_data(work)))
9871094 goto already_gone;
988− prev = &work->entry;
9891095 } else {
990− if (!cwq->worker || cwq->worker->current_work != work)
1096+ if (cwq->worker && cwq->worker->current_work == work)
1097+ worker = cwq->worker;
1098+ if (!worker)
9911099 goto already_gone;
992− prev = &cwq->worklist;
9931100 }
994− insert_wq_barrier(cwq, &barr, prev->next);
9951101
1102+ insert_wq_barrier(cwq, &barr, work, worker);
9961103 spin_unlock_irq(&cwq->lock);
9971104 wait_for_completion(&barr.done);
9981105 destroy_work_on_stack(&barr.work);
@@ -1048,16 +1155,19 @@ static void wait_on_cpu_work(struct cpu_workqueue_struct *cwq, · 49 unmodified lines
10481155 struct work_struct *work)
10491156 {
10501157 struct wq_barrier barr;
1051− int running = 0;
1158+ struct worker *worker;
10521159
10531160 spin_lock_irq(&cwq->lock);
1161+
1162+ worker = NULL;
10541163 if (unlikely(cwq->worker && cwq->worker->current_work == work)) {
1055− insert_wq_barrier(cwq, &barr, cwq->worklist.next);
1056− running = 1;
1164+ worker = cwq->worker;
1165+ insert_wq_barrier(cwq, &barr, work, worker);
10571166 }
1167+
10581168 spin_unlock_irq(&cwq->lock);
10591169
1060− if (unlikely(running)) {
1170+ if (unlikely(worker)) {
10611171 wait_for_completion(&barr.done);
10621172 destroy_work_on_stack(&barr.work);
10631173 }