workqueue.c

   1 /*
   2  * Rated submission helpers
   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 #include "lib/getrusage.h"
  13
  14 enum {
  15         SW_F_IDLE       = 1 << 0,
  16         SW_F_RUNNING    = 1 << 1,
  17         SW_F_EXIT       = 1 << 2,
  18         SW_F_EXITED     = 1 << 3,
  19         SW_F_ACCOUNTED  = 1 << 4,
  20         SW_F_ERROR      = 1 << 5,
  21 };
  22
  23 static struct submit_worker *__get_submit_worker(struct workqueue *wq,
  24                                                  unsigned int start,
  25                                                  unsigned int end,
  26                                                  struct submit_worker **best)
  27 {
  28         struct submit_worker *sw = NULL;
  29
  30         while (start <= end) {
  31                 sw = &wq->workers[start];
  32                 if (sw->flags & SW_F_IDLE)
  33                         return sw;
  34                 if (!(*best) || sw->seq < (*best)->seq)
  35                         *best = sw;
  36                 start++;
  37         }
  38
  39         return NULL;
  40 }
  41
  42 static struct submit_worker *get_submit_worker(struct workqueue *wq)
  43 {
  44         unsigned int next = wq->next_free_worker;
  45         struct submit_worker *sw, *best = NULL;
  46
  47         assert(next < wq->max_workers);
  48
  49         sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
  50         if (!sw && next)
  51                 sw = __get_submit_worker(wq, 0, next - 1, &best);
  52
  53         /*
  54          * No truly idle found, use best match
  55          */
  56         if (!sw)
  57                 sw = best;
  58
  59         if (sw->index == wq->next_free_worker) {
  60                 if (sw->index + 1 < wq->max_workers)
  61                         wq->next_free_worker = sw->index + 1;
  62                 else
  63                         wq->next_free_worker = 0;
  64         }
  65
  66         return sw;
  67 }
  68
  69 static bool all_sw_idle(struct workqueue *wq)
  70 {
  71         int i;
  72
  73         for (i = 0; i < wq->max_workers; i++) {
  74                 struct submit_worker *sw = &wq->workers[i];
  75
  76                 if (!(sw->flags & SW_F_IDLE))
  77                         return false;
  78         }
  79
  80         return true;
  81 }
  82
  83 /*
  84  * Must be serialized wrt workqueue_enqueue() by caller
  85  */
  86 void workqueue_flush(struct workqueue *wq)
  87 {
  88         wq->wake_idle = 1;
  89
  90         while (!all_sw_idle(wq)) {
  91                 pthread_mutex_lock(&wq->flush_lock);
  92                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
  93                 pthread_mutex_unlock(&wq->flush_lock);
  94         }
  95
  96         wq->wake_idle = 0;
  97 }
  98
  99 /*
 100  * Must be serialized by caller. Returns true for queued, false for busy.
 101  */
 102 bool workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
 103 {
 104         struct submit_worker *sw;
 105
 106         sw = get_submit_worker(wq);
 107         if (sw) {
 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                 return true;
 116         }
 117
 118         return false;
 119 }
 120
 121 static void handle_list(struct submit_worker *sw, struct flist_head *list)
 122 {
 123         struct workqueue *wq = sw->wq;
 124         struct workqueue_work *work;
 125
 126         while (!flist_empty(list)) {
 127                 work = flist_first_entry(list, struct workqueue_work, list);
 128                 flist_del_init(&work->list);
 129                 wq->ops.fn(sw, work);
 130         }
 131 }
 132
 133 #ifdef CONFIG_SFAA
 134 static void sum_val(uint64_t *dst, uint64_t *src)
 135 {
 136         if (*src) {
 137                 __sync_fetch_and_add(dst, *src);
 138                 *src = 0;
 139         }
 140 }
 141 #else
 142 static void sum_val(uint64_t *dst, uint64_t *src)
 143 {
 144         if (*src) {
 145                 *dst += *src;
 146                 *src = 0;
 147         }
 148 }
 149 #endif
 150
 151 static void pthread_double_unlock(pthread_mutex_t *lock1,
 152                                   pthread_mutex_t *lock2)
 153 {
 154 #ifndef CONFIG_SFAA
 155         pthread_mutex_unlock(lock1);
 156         pthread_mutex_unlock(lock2);
 157 #endif
 158 }
 159
 160 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
 161 {
 162 #ifndef CONFIG_SFAA
 163         if (lock1 < lock2) {
 164                 pthread_mutex_lock(lock1);
 165                 pthread_mutex_lock(lock2);
 166         } else {
 167                 pthread_mutex_lock(lock2);
 168                 pthread_mutex_lock(lock1);
 169         }
 170 #endif
 171 }
 172
 173 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 174                      enum fio_ddir ddir)
 175 {
 176         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 177
 178         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 179         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 180         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 181         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 182         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 183
 184         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 185 }
 186
 187 static void update_accounting(struct submit_worker *sw)
 188 {
 189         struct thread_data *src = sw->private;
 190         struct thread_data *dst = sw->wq->td;
 191
 192         if (td_read(src))
 193                 sum_ddir(dst, src, DDIR_READ);
 194         if (td_write(src))
 195                 sum_ddir(dst, src, DDIR_WRITE);
 196         if (td_trim(src))
 197                 sum_ddir(dst, src, DDIR_TRIM);
 198 }
 199
 200 static void *worker_thread(void *data)
 201 {
 202         struct submit_worker *sw = data;
 203         struct workqueue *wq = sw->wq;
 204         unsigned int eflags = 0, ret;
 205         FLIST_HEAD(local_list);
 206
 207         ret = workqueue_init_worker(sw);
 208         pthread_mutex_lock(&sw->lock);
 209         sw->flags |= SW_F_RUNNING;
 210         if (ret)
 211                 sw->flags |= SW_F_ERROR;
 212         pthread_mutex_unlock(&sw->lock);
 213
 214         pthread_mutex_lock(&wq->flush_lock);
 215         pthread_cond_signal(&wq->flush_cond);
 216         pthread_mutex_unlock(&wq->flush_lock);
 217
 218         if (sw->flags & SW_F_ERROR)
 219                 goto done;
 220
 221         while (1) {
 222                 pthread_mutex_lock(&sw->lock);
 223
 224                 if (flist_empty(&sw->work_list)) {
 225                         if (sw->flags & SW_F_EXIT) {
 226                                 pthread_mutex_unlock(&sw->lock);
 227                                 break;
 228                         }
 229
 230                         if (workqueue_pre_sleep_check(sw)) {
 231                                 pthread_mutex_unlock(&sw->lock);
 232                                 workqueue_pre_sleep(sw);
 233                                 pthread_mutex_lock(&sw->lock);
 234                         }
 235
 236                         /*
 237                          * We dropped and reaquired the lock, check
 238                          * state again.
 239                          */
 240                         if (!flist_empty(&sw->work_list))
 241                                 goto handle_work;
 242
 243                         if (sw->flags & SW_F_EXIT) {
 244                                 pthread_mutex_unlock(&sw->lock);
 245                                 break;
 246                         } else if (!(sw->flags & SW_F_IDLE)) {
 247                                 sw->flags |= SW_F_IDLE;
 248                                 wq->next_free_worker = sw->index;
 249                                 if (wq->wake_idle)
 250                                         pthread_cond_signal(&wq->flush_cond);
 251                         }
 252                         update_accounting(sw);
 253                         pthread_cond_wait(&sw->cond, &sw->lock);
 254                 } else {
 255 handle_work:
 256                         flist_splice_init(&sw->work_list, &local_list);
 257                 }
 258                 pthread_mutex_unlock(&sw->lock);
 259                 handle_list(sw, &local_list);
 260         }
 261
 262         update_accounting(sw);
 263
 264 done:
 265         pthread_mutex_lock(&sw->lock);
 266         sw->flags |= (SW_F_EXITED | eflags);
 267         pthread_mutex_unlock(&sw->lock);
 268         return NULL;
 269 }
 270
 271 static void free_worker(struct submit_worker *sw)
 272 {
 273         struct workqueue *wq = sw->wq;
 274
 275         workqueue_exit_worker(sw);
 276
 277         pthread_cond_destroy(&sw->cond);
 278         pthread_mutex_destroy(&sw->lock);
 279
 280         if (wq->ops.free_worker_fn)
 281                 wq->ops.free_worker_fn(sw);
 282 }
 283
 284 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 285 {
 286         struct thread_data *parent = sw->wq->td;
 287         struct thread_data *td = sw->private;
 288
 289         pthread_join(sw->thread, NULL);
 290         (*sum_cnt)++;
 291         sum_thread_stats(&parent->ts, &td->ts, *sum_cnt == 1);
 292         free_worker(sw);
 293 }
 294
 295 void workqueue_exit(struct workqueue *wq)
 296 {
 297         unsigned int shutdown, sum_cnt = 0;
 298         struct submit_worker *sw;
 299         int i;
 300
 301         for (i = 0; i < wq->max_workers; i++) {
 302                 sw = &wq->workers[i];
 303
 304                 pthread_mutex_lock(&sw->lock);
 305                 sw->flags |= SW_F_EXIT;
 306                 pthread_cond_signal(&sw->cond);
 307                 pthread_mutex_unlock(&sw->lock);
 308         }
 309
 310         do {
 311                 shutdown = 0;
 312                 for (i = 0; i < wq->max_workers; i++) {
 313                         sw = &wq->workers[i];
 314                         if (sw->flags & SW_F_ACCOUNTED)
 315                                 continue;
 316                         pthread_mutex_lock(&sw->lock);
 317                         sw->flags |= SW_F_ACCOUNTED;
 318                         pthread_mutex_unlock(&sw->lock);
 319                         shutdown_worker(sw, &sum_cnt);
 320                         shutdown++;
 321                 }
 322         } while (shutdown && shutdown != wq->max_workers);
 323
 324         free(wq->workers);
 325         pthread_mutex_destroy(&wq->flush_lock);
 326         pthread_cond_destroy(&wq->flush_cond);
 327         pthread_mutex_destroy(&wq->stat_lock);
 328 }
 329
 330 static int start_worker(struct workqueue *wq, unsigned int index)
 331 {
 332         struct submit_worker *sw = &wq->workers[index];
 333         int ret;
 334
 335         INIT_FLIST_HEAD(&sw->work_list);
 336         pthread_cond_init(&sw->cond, NULL);
 337         pthread_mutex_init(&sw->lock, NULL);
 338         sw->wq = wq;
 339         sw->index = index;
 340
 341         if (wq->ops.alloc_worker_fn) {
 342                 ret = wq->ops.alloc_worker_fn(sw);
 343                 if (ret)
 344                         return ret;
 345         }
 346
 347         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 348         if (!ret) {
 349                 pthread_mutex_lock(&sw->lock);
 350                 sw->flags = SW_F_IDLE;
 351                 pthread_mutex_unlock(&sw->lock);
 352                 return 0;
 353         }
 354
 355         free_worker(sw);
 356         return 1;
 357 }
 358
 359 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 360                    struct workqueue_ops *ops, unsigned max_pending)
 361 {
 362         unsigned int running;
 363         int i, error;
 364
 365         wq->max_workers = max_pending;
 366         wq->td = td;
 367         wq->ops = *ops;
 368         wq->work_seq = 0;
 369         wq->next_free_worker = 0;
 370         pthread_cond_init(&wq->flush_cond, NULL);
 371         pthread_mutex_init(&wq->flush_lock, NULL);
 372         pthread_mutex_init(&wq->stat_lock, NULL);
 373
 374         wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
 375
 376         for (i = 0; i < wq->max_workers; i++)
 377                 if (start_worker(wq, i))
 378                         break;
 379
 380         wq->max_workers = i;
 381         if (!wq->max_workers)
 382                 goto err;
 383
 384         /*
 385          * Wait for them all to be started and initialized
 386          */
 387         error = 0;
 388         do {
 389                 struct submit_worker *sw;
 390
 391                 running = 0;
 392                 pthread_mutex_lock(&wq->flush_lock);
 393                 for (i = 0; i < wq->max_workers; i++) {
 394                         sw = &wq->workers[i];
 395                         pthread_mutex_lock(&sw->lock);
 396                         if (sw->flags & SW_F_RUNNING)
 397                                 running++;
 398                         if (sw->flags & SW_F_ERROR)
 399                                 error++;
 400                         pthread_mutex_unlock(&sw->lock);
 401                 }
 402
 403                 if (error || running == wq->max_workers) {
 404                         pthread_mutex_unlock(&wq->flush_lock);
 405                         break;
 406                 }
 407
 408                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 409                 pthread_mutex_unlock(&wq->flush_lock);
 410         } while (1);
 411
 412         if (!error)
 413                 return 0;
 414
 415 err:
 416         log_err("Can't create rate workqueue\n");
 417         td_verror(td, ESRCH, "workqueue_init");
 418         workqueue_exit(wq);
 419         return 1;
 420 }