workqueue.c

   1 /*
   2  * Generic workqueue offload mechanism
   3  *
   4  * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
   5  *
   6  */
   7 #include <unistd.h>
   8
   9 #include "fio.h"
  10 #include "flist.h"
  11 #include "workqueue.h"
  12
  13 enum {
  14         SW_F_IDLE       = 1 << 0,
  15         SW_F_RUNNING    = 1 << 1,
  16         SW_F_EXIT       = 1 << 2,
  17         SW_F_EXITED     = 1 << 3,
  18         SW_F_ACCOUNTED  = 1 << 4,
  19         SW_F_ERROR      = 1 << 5,
  20 };
  21
  22 static struct submit_worker *__get_submit_worker(struct workqueue *wq,
  23                                                  unsigned int start,
  24                                                  unsigned int end,
  25                                                  struct submit_worker **best)
  26 {
  27         struct submit_worker *sw = NULL;
  28
  29         while (start <= end) {
  30                 sw = &wq->workers[start];
  31                 if (sw->flags & SW_F_IDLE)
  32                         return sw;
  33                 if (!(*best) || sw->seq < (*best)->seq)
  34                         *best = sw;
  35                 start++;
  36         }
  37
  38         return NULL;
  39 }
  40
  41 static struct submit_worker *get_submit_worker(struct workqueue *wq)
  42 {
  43         unsigned int next = wq->next_free_worker;
  44         struct submit_worker *sw, *best = NULL;
  45
  46         assert(next < wq->max_workers);
  47
  48         sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
  49         if (!sw && next)
  50                 sw = __get_submit_worker(wq, 0, next - 1, &best);
  51
  52         /*
  53          * No truly idle found, use best match
  54          */
  55         if (!sw)
  56                 sw = best;
  57
  58         if (sw->index == wq->next_free_worker) {
  59                 if (sw->index + 1 < wq->max_workers)
  60                         wq->next_free_worker = sw->index + 1;
  61                 else
  62                         wq->next_free_worker = 0;
  63         }
  64
  65         return sw;
  66 }
  67
  68 static bool all_sw_idle(struct workqueue *wq)
  69 {
  70         int i;
  71
  72         for (i = 0; i < wq->max_workers; i++) {
  73                 struct submit_worker *sw = &wq->workers[i];
  74
  75                 if (!(sw->flags & SW_F_IDLE))
  76                         return false;
  77         }
  78
  79         return true;
  80 }
  81
  82 /*
  83  * Must be serialized wrt workqueue_enqueue() by caller
  84  */
  85 void workqueue_flush(struct workqueue *wq)
  86 {
  87         wq->wake_idle = 1;
  88
  89         while (!all_sw_idle(wq)) {
  90                 pthread_mutex_lock(&wq->flush_lock);
  91                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
  92                 pthread_mutex_unlock(&wq->flush_lock);
  93         }
  94
  95         wq->wake_idle = 0;
  96 }
  97
  98 /*
  99  * Must be serialized by caller. Returns true for queued, false for busy.
 100  */
 101 void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
 102 {
 103         struct submit_worker *sw;
 104
 105         sw = get_submit_worker(wq);
 106         assert(sw);
 107
 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;
 112         pthread_mutex_unlock(&sw->lock);
 113
 114         pthread_cond_signal(&sw->cond);
 115 }
 116
 117 static void handle_list(struct submit_worker *sw, struct flist_head *list)
 118 {
 119         struct workqueue *wq = sw->wq;
 120         struct workqueue_work *work;
 121
 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);
 126         }
 127 }
 128
 129 static void *worker_thread(void *data)
 130 {
 131         struct submit_worker *sw = data;
 132         struct workqueue *wq = sw->wq;
 133         unsigned int eflags = 0, ret = 0;
 134         FLIST_HEAD(local_list);
 135
 136         if (wq->ops.nice) {
 137                 if (nice(wq->ops.nice) < 0) {
 138                         log_err("workqueue: nice %s\n", strerror(errno));
 139                         ret = 1;
 140                 }
 141         }
 142
 143         if (!ret)
 144                 ret = workqueue_init_worker(sw);
 145
 146         pthread_mutex_lock(&sw->lock);
 147         sw->flags |= SW_F_RUNNING;
 148         if (ret)
 149                 sw->flags |= SW_F_ERROR;
 150         pthread_mutex_unlock(&sw->lock);
 151
 152         pthread_mutex_lock(&wq->flush_lock);
 153         pthread_cond_signal(&wq->flush_cond);
 154         pthread_mutex_unlock(&wq->flush_lock);
 155
 156         if (sw->flags & SW_F_ERROR)
 157                 goto done;
 158
 159         while (1) {
 160                 pthread_mutex_lock(&sw->lock);
 161
 162                 if (flist_empty(&sw->work_list)) {
 163                         if (sw->flags & SW_F_EXIT) {
 164                                 pthread_mutex_unlock(&sw->lock);
 165                                 break;
 166                         }
 167
 168                         if (workqueue_pre_sleep_check(sw)) {
 169                                 pthread_mutex_unlock(&sw->lock);
 170                                 workqueue_pre_sleep(sw);
 171                                 pthread_mutex_lock(&sw->lock);
 172                         }
 173
 174                         /*
 175                          * We dropped and reaquired the lock, check
 176                          * state again.
 177                          */
 178                         if (!flist_empty(&sw->work_list))
 179                                 goto handle_work;
 180
 181                         if (sw->flags & SW_F_EXIT) {
 182                                 pthread_mutex_unlock(&sw->lock);
 183                                 break;
 184                         } else if (!(sw->flags & SW_F_IDLE)) {
 185                                 sw->flags |= SW_F_IDLE;
 186                                 wq->next_free_worker = sw->index;
 187                                 if (wq->wake_idle)
 188                                         pthread_cond_signal(&wq->flush_cond);
 189                         }
 190                         if (wq->ops.update_acct_fn)
 191                                 wq->ops.update_acct_fn(sw);
 192
 193                         pthread_cond_wait(&sw->cond, &sw->lock);
 194                 } else {
 195 handle_work:
 196                         flist_splice_init(&sw->work_list, &local_list);
 197                 }
 198                 pthread_mutex_unlock(&sw->lock);
 199                 handle_list(sw, &local_list);
 200         }
 201
 202         if (wq->ops.update_acct_fn)
 203                 wq->ops.update_acct_fn(sw);
 204
 205 done:
 206         pthread_mutex_lock(&sw->lock);
 207         sw->flags |= (SW_F_EXITED | eflags);
 208         pthread_mutex_unlock(&sw->lock);
 209         return NULL;
 210 }
 211
 212 static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 213 {
 214         struct workqueue *wq = sw->wq;
 215
 216         workqueue_exit_worker(sw, sum_cnt);
 217
 218         pthread_cond_destroy(&sw->cond);
 219         pthread_mutex_destroy(&sw->lock);
 220
 221         if (wq->ops.free_worker_fn)
 222                 wq->ops.free_worker_fn(sw);
 223 }
 224
 225 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 226 {
 227         pthread_join(sw->thread, NULL);
 228         free_worker(sw, sum_cnt);
 229 }
 230
 231 void workqueue_exit(struct workqueue *wq)
 232 {
 233         unsigned int shutdown, sum_cnt = 0;
 234         struct submit_worker *sw;
 235         int i;
 236
 237         if (!wq->workers)
 238                 return;
 239
 240         for (i = 0; i < wq->max_workers; i++) {
 241                 sw = &wq->workers[i];
 242
 243                 pthread_mutex_lock(&sw->lock);
 244                 sw->flags |= SW_F_EXIT;
 245                 pthread_cond_signal(&sw->cond);
 246                 pthread_mutex_unlock(&sw->lock);
 247         }
 248
 249         do {
 250                 shutdown = 0;
 251                 for (i = 0; i < wq->max_workers; i++) {
 252                         sw = &wq->workers[i];
 253                         if (sw->flags & SW_F_ACCOUNTED)
 254                                 continue;
 255                         pthread_mutex_lock(&sw->lock);
 256                         sw->flags |= SW_F_ACCOUNTED;
 257                         pthread_mutex_unlock(&sw->lock);
 258                         shutdown_worker(sw, &sum_cnt);
 259                         shutdown++;
 260                 }
 261         } while (shutdown && shutdown != wq->max_workers);
 262
 263         free(wq->workers);
 264         wq->workers = NULL;
 265         pthread_mutex_destroy(&wq->flush_lock);
 266         pthread_cond_destroy(&wq->flush_cond);
 267         pthread_mutex_destroy(&wq->stat_lock);
 268 }
 269
 270 static int start_worker(struct workqueue *wq, unsigned int index)
 271 {
 272         struct submit_worker *sw = &wq->workers[index];
 273         int ret;
 274
 275         INIT_FLIST_HEAD(&sw->work_list);
 276         pthread_cond_init(&sw->cond, NULL);
 277         pthread_mutex_init(&sw->lock, NULL);
 278         sw->wq = wq;
 279         sw->index = index;
 280
 281         if (wq->ops.alloc_worker_fn) {
 282                 ret = wq->ops.alloc_worker_fn(sw);
 283                 if (ret)
 284                         return ret;
 285         }
 286
 287         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 288         if (!ret) {
 289                 pthread_mutex_lock(&sw->lock);
 290                 sw->flags = SW_F_IDLE;
 291                 pthread_mutex_unlock(&sw->lock);
 292                 return 0;
 293         }
 294
 295         free_worker(sw, NULL);
 296         return 1;
 297 }
 298
 299 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 300                    struct workqueue_ops *ops, unsigned max_pending)
 301 {
 302         unsigned int running;
 303         int i, error;
 304
 305         wq->max_workers = max_pending;
 306         wq->td = td;
 307         wq->ops = *ops;
 308         wq->work_seq = 0;
 309         wq->next_free_worker = 0;
 310         pthread_cond_init(&wq->flush_cond, NULL);
 311         pthread_mutex_init(&wq->flush_lock, NULL);
 312         pthread_mutex_init(&wq->stat_lock, NULL);
 313
 314         wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
 315
 316         for (i = 0; i < wq->max_workers; i++)
 317                 if (start_worker(wq, i))
 318                         break;
 319
 320         wq->max_workers = i;
 321         if (!wq->max_workers)
 322                 goto err;
 323
 324         /*
 325          * Wait for them all to be started and initialized
 326          */
 327         error = 0;
 328         do {
 329                 struct submit_worker *sw;
 330
 331                 running = 0;
 332                 pthread_mutex_lock(&wq->flush_lock);
 333                 for (i = 0; i < wq->max_workers; i++) {
 334                         sw = &wq->workers[i];
 335                         pthread_mutex_lock(&sw->lock);
 336                         if (sw->flags & SW_F_RUNNING)
 337                                 running++;
 338                         if (sw->flags & SW_F_ERROR)
 339                                 error++;
 340                         pthread_mutex_unlock(&sw->lock);
 341                 }
 342
 343                 if (error || running == wq->max_workers) {
 344                         pthread_mutex_unlock(&wq->flush_lock);
 345                         break;
 346                 }
 347
 348                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 349                 pthread_mutex_unlock(&wq->flush_lock);
 350         } while (1);
 351
 352         if (!error)
 353                 return 0;
 354
 355 err:
 356         log_err("Can't create rate workqueue\n");
 357         td_verror(td, ESRCH, "workqueue_init");
 358         workqueue_exit(wq);
 359         return 1;
 360 }