2 * Generic workqueue offload mechanism
4 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
11 #include "workqueue.h"
17 SW_F_RUNNING = 1 << 1,
19 SW_F_ACCOUNTED = 1 << 3,
23 static struct submit_worker *__get_submit_worker(struct workqueue *wq,
26 struct submit_worker **best)
28 struct submit_worker *sw = NULL;
30 while (start <= end) {
31 sw = &wq->workers[start];
32 if (sw->flags & SW_F_IDLE)
34 if (!(*best) || sw->seq < (*best)->seq)
42 static struct submit_worker *get_submit_worker(struct workqueue *wq)
44 unsigned int next = wq->next_free_worker;
45 struct submit_worker *sw, *best = NULL;
47 assert(next < wq->max_workers);
49 sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
51 sw = __get_submit_worker(wq, 0, next - 1, &best);
54 * No truly idle found, use best match
59 if (sw->index == wq->next_free_worker) {
60 if (sw->index + 1 < wq->max_workers)
61 wq->next_free_worker = sw->index + 1;
63 wq->next_free_worker = 0;
69 static bool all_sw_idle(struct workqueue *wq)
73 for (i = 0; i < wq->max_workers; i++) {
74 struct submit_worker *sw = &wq->workers[i];
76 if (!(sw->flags & SW_F_IDLE))
84 * Must be serialized wrt workqueue_enqueue() by caller
86 void workqueue_flush(struct workqueue *wq)
88 pthread_mutex_lock(&wq->flush_lock);
91 while (!all_sw_idle(wq))
92 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
95 pthread_mutex_unlock(&wq->flush_lock);
99 * Must be serialized by caller.
101 void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
103 struct submit_worker *sw;
105 sw = get_submit_worker(wq);
108 pthread_mutex_lock(&sw->lock);
109 flist_add_tail(&work->list, &sw->work_list);
110 sw->seq = ++wq->work_seq;
111 sw->flags &= ~SW_F_IDLE;
113 pthread_cond_signal(&sw->cond);
114 pthread_mutex_unlock(&sw->lock);
117 static void handle_list(struct submit_worker *sw, struct flist_head *list)
119 struct workqueue *wq = sw->wq;
120 struct workqueue_work *work;
122 while (!flist_empty(list)) {
123 work = flist_first_entry(list, struct workqueue_work, list);
124 flist_del_init(&work->list);
125 wq->ops.fn(sw, work);
129 static void *worker_thread(void *data)
131 struct submit_worker *sw = data;
132 struct workqueue *wq = sw->wq;
133 unsigned int ret = 0;
134 FLIST_HEAD(local_list);
136 sk_out_assign(sw->sk_out);
140 if (nice(wq->ops.nice) == -1 && errno != 0) {
141 log_err("workqueue: nice %s\n", strerror(errno));
147 ret = workqueue_init_worker(sw);
149 pthread_mutex_lock(&sw->lock);
150 sw->flags |= SW_F_RUNNING;
152 sw->flags |= SW_F_ERROR;
153 pthread_mutex_unlock(&sw->lock);
155 pthread_mutex_lock(&wq->flush_lock);
156 pthread_cond_signal(&wq->flush_cond);
157 pthread_mutex_unlock(&wq->flush_lock);
159 if (sw->flags & SW_F_ERROR)
162 pthread_mutex_lock(&sw->lock);
164 if (flist_empty(&sw->work_list)) {
165 if (sw->flags & SW_F_EXIT) {
169 if (workqueue_pre_sleep_check(sw)) {
170 pthread_mutex_unlock(&sw->lock);
171 workqueue_pre_sleep(sw);
172 pthread_mutex_lock(&sw->lock);
176 * We may have dropped and reaquired the lock, check state
179 if (flist_empty(&sw->work_list)) {
180 if (sw->flags & SW_F_EXIT) {
183 if (!(sw->flags & SW_F_IDLE)) {
184 sw->flags |= SW_F_IDLE;
185 wq->next_free_worker = sw->index;
186 pthread_mutex_unlock(&sw->lock);
187 pthread_mutex_lock(&wq->flush_lock);
189 pthread_cond_signal(&wq->flush_cond);
190 pthread_mutex_unlock(&wq->flush_lock);
191 pthread_mutex_lock(&sw->lock);
194 if (flist_empty(&sw->work_list)) {
195 if (sw->flags & SW_F_EXIT) {
198 pthread_cond_wait(&sw->cond, &sw->lock);
200 flist_splice_init(&sw->work_list, &local_list);
202 pthread_mutex_unlock(&sw->lock);
203 handle_list(sw, &local_list);
204 if (wq->ops.update_acct_fn)
205 wq->ops.update_acct_fn(sw);
206 pthread_mutex_lock(&sw->lock);
208 pthread_mutex_unlock(&sw->lock);
215 static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt)
217 struct workqueue *wq = sw->wq;
219 workqueue_exit_worker(sw, sum_cnt);
221 pthread_cond_destroy(&sw->cond);
222 pthread_mutex_destroy(&sw->lock);
224 if (wq->ops.free_worker_fn)
225 wq->ops.free_worker_fn(sw);
228 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
230 pthread_join(sw->thread, NULL);
231 free_worker(sw, sum_cnt);
234 void workqueue_exit(struct workqueue *wq)
236 unsigned int shutdown, sum_cnt = 0;
237 struct submit_worker *sw;
243 for (i = 0; i < wq->max_workers; i++) {
244 sw = &wq->workers[i];
246 pthread_mutex_lock(&sw->lock);
247 sw->flags |= SW_F_EXIT;
248 pthread_cond_signal(&sw->cond);
249 pthread_mutex_unlock(&sw->lock);
254 for (i = 0; i < wq->max_workers; i++) {
255 sw = &wq->workers[i];
256 if (sw->flags & SW_F_ACCOUNTED)
258 pthread_mutex_lock(&sw->lock);
259 sw->flags |= SW_F_ACCOUNTED;
260 pthread_mutex_unlock(&sw->lock);
261 shutdown_worker(sw, &sum_cnt);
264 } while (shutdown && shutdown != wq->max_workers);
268 pthread_mutex_destroy(&wq->flush_lock);
269 pthread_cond_destroy(&wq->flush_cond);
270 pthread_mutex_destroy(&wq->stat_lock);
273 static int start_worker(struct workqueue *wq, unsigned int index,
274 struct sk_out *sk_out)
276 struct submit_worker *sw = &wq->workers[index];
279 INIT_FLIST_HEAD(&sw->work_list);
281 ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
289 if (wq->ops.alloc_worker_fn) {
290 ret = wq->ops.alloc_worker_fn(sw);
295 ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
297 pthread_mutex_lock(&sw->lock);
298 sw->flags = SW_F_IDLE;
299 pthread_mutex_unlock(&sw->lock);
303 free_worker(sw, NULL);
307 int workqueue_init(struct thread_data *td, struct workqueue *wq,
308 struct workqueue_ops *ops, unsigned int max_workers,
309 struct sk_out *sk_out)
311 unsigned int running;
315 wq->max_workers = max_workers;
319 wq->next_free_worker = 0;
321 ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
324 ret = mutex_init_pshared(&wq->stat_lock);
328 wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
332 for (i = 0; i < wq->max_workers; i++)
333 if (start_worker(wq, i, sk_out))
337 if (!wq->max_workers)
341 * Wait for them all to be started and initialized
344 pthread_mutex_lock(&wq->flush_lock);
346 struct submit_worker *sw;
349 for (i = 0; i < wq->max_workers; i++) {
350 sw = &wq->workers[i];
351 pthread_mutex_lock(&sw->lock);
352 if (sw->flags & SW_F_RUNNING)
354 if (sw->flags & SW_F_ERROR)
356 pthread_mutex_unlock(&sw->lock);
359 if (error || running == wq->max_workers)
362 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
364 pthread_mutex_unlock(&wq->flush_lock);
370 log_err("Can't create rate workqueue\n");
371 td_verror(td, ESRCH, "workqueue_init");