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 static int init_submit_worker(struct submit_worker *sw)
 134 {
 135         struct thread_data *parent = sw->wq->td;
 136         struct thread_data *td = sw->private;
 137         int fio_unused ret;
 138
 139         memcpy(&td->o, &parent->o, sizeof(td->o));
 140         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
 141         td->o.uid = td->o.gid = -1U;
 142         dup_files(td, parent);
 143         td->eo = parent->eo;
 144         fio_options_mem_dupe(td);
 145
 146         if (ioengine_load(td))
 147                 goto err;
 148
 149         if (td->o.odirect)
 150                 td->io_ops->flags |= FIO_RAWIO;
 151
 152         td->pid = gettid();
 153
 154         INIT_FLIST_HEAD(&td->io_log_list);
 155         INIT_FLIST_HEAD(&td->io_hist_list);
 156         INIT_FLIST_HEAD(&td->verify_list);
 157         INIT_FLIST_HEAD(&td->trim_list);
 158         INIT_FLIST_HEAD(&td->next_rand_list);
 159         td->io_hist_tree = RB_ROOT;
 160
 161         td->o.iodepth = 1;
 162         if (td_io_init(td))
 163                 goto err_io_init;
 164
 165         fio_gettime(&td->epoch, NULL);
 166         fio_getrusage(&td->ru_start);
 167         clear_io_state(td, 1);
 168
 169         td_set_runstate(td, TD_RUNNING);
 170         td->flags |= TD_F_CHILD;
 171         td->parent = parent;
 172         return 0;
 173
 174 err_io_init:
 175         close_ioengine(td);
 176 err:
 177         return 1;
 178 }
 179
 180 #ifdef CONFIG_SFAA
 181 static void sum_val(uint64_t *dst, uint64_t *src)
 182 {
 183         if (*src) {
 184                 __sync_fetch_and_add(dst, *src);
 185                 *src = 0;
 186         }
 187 }
 188 #else
 189 static void sum_val(uint64_t *dst, uint64_t *src)
 190 {
 191         if (*src) {
 192                 *dst += *src;
 193                 *src = 0;
 194         }
 195 }
 196 #endif
 197
 198 static void pthread_double_unlock(pthread_mutex_t *lock1,
 199                                   pthread_mutex_t *lock2)
 200 {
 201 #ifndef CONFIG_SFAA
 202         pthread_mutex_unlock(lock1);
 203         pthread_mutex_unlock(lock2);
 204 #endif
 205 }
 206
 207 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
 208 {
 209 #ifndef CONFIG_SFAA
 210         if (lock1 < lock2) {
 211                 pthread_mutex_lock(lock1);
 212                 pthread_mutex_lock(lock2);
 213         } else {
 214                 pthread_mutex_lock(lock2);
 215                 pthread_mutex_lock(lock1);
 216         }
 217 #endif
 218 }
 219
 220 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 221                      enum fio_ddir ddir)
 222 {
 223         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 224
 225         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 226         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 227         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 228         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 229         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 230
 231         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 232 }
 233
 234 static void update_accounting(struct submit_worker *sw)
 235 {
 236         struct thread_data *src = sw->private;
 237         struct thread_data *dst = sw->wq->td;
 238
 239         if (td_read(src))
 240                 sum_ddir(dst, src, DDIR_READ);
 241         if (td_write(src))
 242                 sum_ddir(dst, src, DDIR_WRITE);
 243         if (td_trim(src))
 244                 sum_ddir(dst, src, DDIR_TRIM);
 245 }
 246
 247 static void *worker_thread(void *data)
 248 {
 249         struct submit_worker *sw = data;
 250         struct workqueue *wq = sw->wq;
 251         unsigned int eflags = 0, ret;
 252         FLIST_HEAD(local_list);
 253
 254         ret = init_submit_worker(sw);
 255         pthread_mutex_lock(&sw->lock);
 256         sw->flags |= SW_F_RUNNING;
 257         if (ret)
 258                 sw->flags |= SW_F_ERROR;
 259         pthread_mutex_unlock(&sw->lock);
 260
 261         pthread_mutex_lock(&wq->flush_lock);
 262         pthread_cond_signal(&wq->flush_cond);
 263         pthread_mutex_unlock(&wq->flush_lock);
 264
 265         if (sw->flags & SW_F_ERROR)
 266                 goto done;
 267
 268         while (1) {
 269                 pthread_mutex_lock(&sw->lock);
 270
 271                 if (flist_empty(&sw->work_list)) {
 272                         if (sw->flags & SW_F_EXIT) {
 273                                 pthread_mutex_unlock(&sw->lock);
 274                                 break;
 275                         }
 276
 277                         if (workqueue_pre_sleep_check(sw)) {
 278                                 pthread_mutex_unlock(&sw->lock);
 279                                 workqueue_pre_sleep(sw);
 280                                 pthread_mutex_lock(&sw->lock);
 281                         }
 282
 283                         /*
 284                          * We dropped and reaquired the lock, check
 285                          * state again.
 286                          */
 287                         if (!flist_empty(&sw->work_list))
 288                                 goto handle_work;
 289
 290                         if (sw->flags & SW_F_EXIT) {
 291                                 pthread_mutex_unlock(&sw->lock);
 292                                 break;
 293                         } else if (!(sw->flags & SW_F_IDLE)) {
 294                                 sw->flags |= SW_F_IDLE;
 295                                 wq->next_free_worker = sw->index;
 296                                 if (wq->wake_idle)
 297                                         pthread_cond_signal(&wq->flush_cond);
 298                         }
 299                         update_accounting(sw);
 300                         pthread_cond_wait(&sw->cond, &sw->lock);
 301                 } else {
 302 handle_work:
 303                         flist_splice_init(&sw->work_list, &local_list);
 304                 }
 305                 pthread_mutex_unlock(&sw->lock);
 306                 handle_list(sw, &local_list);
 307         }
 308
 309         update_accounting(sw);
 310
 311 done:
 312         pthread_mutex_lock(&sw->lock);
 313         sw->flags |= (SW_F_EXITED | eflags);
 314         pthread_mutex_unlock(&sw->lock);
 315         return NULL;
 316 }
 317
 318 static void free_worker(struct submit_worker *sw)
 319 {
 320         struct thread_data *td = sw->private;
 321         struct workqueue *wq = sw->wq;
 322
 323         fio_options_free(td);
 324         close_and_free_files(td);
 325         if (td->io_ops)
 326                 close_ioengine(td);
 327         td_set_runstate(td, TD_EXITED);
 328
 329         pthread_cond_destroy(&sw->cond);
 330         pthread_mutex_destroy(&sw->lock);
 331
 332         if (wq->ops.free_worker_fn)
 333                 wq->ops.free_worker_fn(sw);
 334 }
 335
 336 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 337 {
 338         struct thread_data *parent = sw->wq->td;
 339         struct thread_data *td = sw->private;
 340
 341         pthread_join(sw->thread, NULL);
 342         (*sum_cnt)++;
 343         sum_thread_stats(&parent->ts, &td->ts, *sum_cnt == 1);
 344         free_worker(sw);
 345 }
 346
 347 void workqueue_exit(struct workqueue *wq)
 348 {
 349         unsigned int shutdown, sum_cnt = 0;
 350         struct submit_worker *sw;
 351         int i;
 352
 353         for (i = 0; i < wq->max_workers; i++) {
 354                 sw = &wq->workers[i];
 355
 356                 pthread_mutex_lock(&sw->lock);
 357                 sw->flags |= SW_F_EXIT;
 358                 pthread_cond_signal(&sw->cond);
 359                 pthread_mutex_unlock(&sw->lock);
 360         }
 361
 362         do {
 363                 shutdown = 0;
 364                 for (i = 0; i < wq->max_workers; i++) {
 365                         sw = &wq->workers[i];
 366                         if (sw->flags & SW_F_ACCOUNTED)
 367                                 continue;
 368                         pthread_mutex_lock(&sw->lock);
 369                         sw->flags |= SW_F_ACCOUNTED;
 370                         pthread_mutex_unlock(&sw->lock);
 371                         shutdown_worker(sw, &sum_cnt);
 372                         shutdown++;
 373                 }
 374         } while (shutdown && shutdown != wq->max_workers);
 375
 376         free(wq->workers);
 377         pthread_mutex_destroy(&wq->flush_lock);
 378         pthread_cond_destroy(&wq->flush_cond);
 379         pthread_mutex_destroy(&wq->stat_lock);
 380 }
 381
 382 static int start_worker(struct workqueue *wq, unsigned int index)
 383 {
 384         struct submit_worker *sw = &wq->workers[index];
 385         int ret;
 386
 387         INIT_FLIST_HEAD(&sw->work_list);
 388         pthread_cond_init(&sw->cond, NULL);
 389         pthread_mutex_init(&sw->lock, NULL);
 390         sw->wq = wq;
 391         sw->index = index;
 392
 393         if (wq->ops.alloc_worker_fn) {
 394                 ret = wq->ops.alloc_worker_fn(sw);
 395                 if (ret)
 396                         return ret;
 397         }
 398
 399         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 400         if (!ret) {
 401                 pthread_mutex_lock(&sw->lock);
 402                 sw->flags = SW_F_IDLE;
 403                 pthread_mutex_unlock(&sw->lock);
 404                 return 0;
 405         }
 406
 407         free_worker(sw);
 408         return 1;
 409 }
 410
 411 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 412                    struct workqueue_ops *ops, unsigned max_pending)
 413 {
 414         unsigned int running;
 415         int i, error;
 416
 417         wq->max_workers = max_pending;
 418         wq->td = td;
 419         wq->ops = *ops;
 420         wq->work_seq = 0;
 421         wq->next_free_worker = 0;
 422         pthread_cond_init(&wq->flush_cond, NULL);
 423         pthread_mutex_init(&wq->flush_lock, NULL);
 424         pthread_mutex_init(&wq->stat_lock, NULL);
 425
 426         wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
 427
 428         for (i = 0; i < wq->max_workers; i++)
 429                 if (start_worker(wq, i))
 430                         break;
 431
 432         wq->max_workers = i;
 433         if (!wq->max_workers)
 434                 goto err;
 435
 436         /*
 437          * Wait for them all to be started and initialized
 438          */
 439         error = 0;
 440         do {
 441                 struct submit_worker *sw;
 442
 443                 running = 0;
 444                 pthread_mutex_lock(&wq->flush_lock);
 445                 for (i = 0; i < wq->max_workers; i++) {
 446                         sw = &wq->workers[i];
 447                         pthread_mutex_lock(&sw->lock);
 448                         if (sw->flags & SW_F_RUNNING)
 449                                 running++;
 450                         if (sw->flags & SW_F_ERROR)
 451                                 error++;
 452                         pthread_mutex_unlock(&sw->lock);
 453                 }
 454
 455                 if (error || running == wq->max_workers) {
 456                         pthread_mutex_unlock(&wq->flush_lock);
 457                         break;
 458                 }
 459
 460                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 461                 pthread_mutex_unlock(&wq->flush_lock);
 462         } while (1);
 463
 464         if (!error)
 465                 return 0;
 466
 467 err:
 468         log_err("Can't create rate workqueue\n");
 469         td_verror(td, ESRCH, "workqueue_init");
 470         workqueue_exit(wq);
 471         return 1;
 472 }