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 "ioengine.h"
  11 #include "flist.h"
  12 #include "workqueue.h"
  13 #include "lib/getrusage.h"
  14
  15 struct submit_worker {
  16         pthread_t thread;
  17         pthread_mutex_t lock;
  18         pthread_cond_t cond;
  19         struct flist_head work_list;
  20         unsigned int flags;
  21         unsigned int index;
  22         uint64_t seq;
  23         struct workqueue *wq;
  24         struct thread_data td;
  25 };
  26
  27 enum {
  28         SW_F_IDLE       = 1 << 0,
  29         SW_F_RUNNING    = 1 << 1,
  30         SW_F_EXIT       = 1 << 2,
  31         SW_F_EXITED     = 1 << 3,
  32         SW_F_ACCOUNTED  = 1 << 4,
  33         SW_F_ERROR      = 1 << 5,
  34 };
  35
  36 static struct submit_worker *__get_submit_worker(struct workqueue *wq,
  37                                                  unsigned int start,
  38                                                  unsigned int end,
  39                                                  struct submit_worker **best)
  40 {
  41         struct submit_worker *sw = NULL;
  42
  43         while (start <= end) {
  44                 sw = &wq->workers[start];
  45                 if (sw->flags & SW_F_IDLE)
  46                         return sw;
  47                 if (!(*best) || sw->seq < (*best)->seq)
  48                         *best = sw;
  49                 start++;
  50         }
  51
  52         return NULL;
  53 }
  54
  55 static struct submit_worker *get_submit_worker(struct workqueue *wq)
  56 {
  57         unsigned int next = wq->next_free_worker;
  58         struct submit_worker *sw, *best = NULL;
  59
  60         assert(next < wq->max_workers);
  61
  62         sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
  63         if (!sw && next)
  64                 sw = __get_submit_worker(wq, 0, next - 1, &best);
  65
  66         /*
  67          * No truly idle found, use best match
  68          */
  69         if (!sw)
  70                 sw = best;
  71
  72         if (sw->index == wq->next_free_worker) {
  73                 if (sw->index + 1 < wq->max_workers)
  74                         wq->next_free_worker = sw->index + 1;
  75                 else
  76                         wq->next_free_worker = 0;
  77         }
  78
  79         return sw;
  80 }
  81
  82 static int all_sw_idle(struct workqueue *wq)
  83 {
  84         int i;
  85
  86         for (i = 0; i < wq->max_workers; i++) {
  87                 struct submit_worker *sw = &wq->workers[i];
  88
  89                 if (!(sw->flags & SW_F_IDLE))
  90                         return 0;
  91         }
  92
  93         return 1;
  94 }
  95
  96 /*
  97  * Must be serialized wrt workqueue_enqueue() by caller
  98  */
  99 void workqueue_flush(struct workqueue *wq)
 100 {
 101         wq->wake_idle = 1;
 102
 103         while (!all_sw_idle(wq)) {
 104                 pthread_mutex_lock(&wq->flush_lock);
 105                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 106                 pthread_mutex_unlock(&wq->flush_lock);
 107         }
 108
 109         wq->wake_idle = 0;
 110 }
 111
 112 /*
 113  * Must be serialized by caller.
 114  */
 115 int workqueue_enqueue(struct workqueue *wq, struct io_u *io_u)
 116 {
 117         struct submit_worker *sw;
 118
 119         sw = get_submit_worker(wq);
 120         if (sw) {
 121                 const enum fio_ddir ddir = acct_ddir(io_u);
 122                 struct thread_data *parent = wq->td;
 123
 124                 if (ddir_rw(ddir)) {
 125                         parent->io_issues[ddir]++;
 126                         parent->io_issue_bytes[ddir] += io_u->xfer_buflen;
 127                         parent->rate_io_issue_bytes[ddir] += io_u->xfer_buflen;
 128                 }
 129
 130                 pthread_mutex_lock(&sw->lock);
 131                 flist_add_tail(&io_u->verify_list, &sw->work_list);
 132                 sw->seq = ++wq->work_seq;
 133                 sw->flags &= ~SW_F_IDLE;
 134                 pthread_mutex_unlock(&sw->lock);
 135
 136                 pthread_cond_signal(&sw->cond);
 137                 return FIO_Q_QUEUED;
 138         }
 139
 140         return FIO_Q_BUSY;
 141 }
 142
 143 static void handle_list(struct submit_worker *sw, struct flist_head *list)
 144 {
 145         struct workqueue *wq = sw->wq;
 146         struct io_u *io_u;
 147
 148         while (!flist_empty(list)) {
 149                 io_u = flist_first_entry(list, struct io_u, verify_list);
 150                 flist_del_init(&io_u->verify_list);
 151                 wq->fn(&sw->td, io_u);
 152         }
 153 }
 154
 155 static int init_submit_worker(struct submit_worker *sw)
 156 {
 157         struct thread_data *parent = sw->wq->td;
 158         struct thread_data *td = &sw->td;
 159         int fio_unused ret;
 160
 161         memcpy(&td->o, &parent->o, sizeof(td->o));
 162         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
 163         td->o.uid = td->o.gid = -1U;
 164         dup_files(td, parent);
 165         td->eo = parent->eo;
 166         fio_options_mem_dupe(td);
 167
 168         if (ioengine_load(td))
 169                 goto err;
 170
 171         if (td->o.odirect)
 172                 td->io_ops->flags |= FIO_RAWIO;
 173
 174         td->pid = gettid();
 175
 176         INIT_FLIST_HEAD(&td->io_log_list);
 177         INIT_FLIST_HEAD(&td->io_hist_list);
 178         INIT_FLIST_HEAD(&td->verify_list);
 179         INIT_FLIST_HEAD(&td->trim_list);
 180         INIT_FLIST_HEAD(&td->next_rand_list);
 181         td->io_hist_tree = RB_ROOT;
 182
 183         td->o.iodepth = 1;
 184         if (td_io_init(td))
 185                 goto err_io_init;
 186
 187         fio_gettime(&td->epoch, NULL);
 188         fio_getrusage(&td->ru_start);
 189         clear_io_state(td);
 190
 191         td_set_runstate(td, TD_RUNNING);
 192         td->flags |= TD_F_CHILD;
 193         td->parent = parent;
 194         return 0;
 195
 196 err_io_init:
 197         close_ioengine(td);
 198 err:
 199         return 1;
 200 }
 201
 202 #ifdef CONFIG_SFAA
 203 static void sum_val(uint64_t *dst, uint64_t *src)
 204 {
 205         if (*src) {
 206                 __sync_fetch_and_add(dst, *src);
 207                 *src = 0;
 208         }
 209 }
 210 #else
 211 static void sum_val(uint64_t *dst, uint64_t *src)
 212 {
 213         if (*src) {
 214                 *dst += *src;
 215                 *src = 0;
 216         }
 217 }
 218 #endif
 219
 220 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 221                      enum fio_ddir ddir)
 222 {
 223 #ifndef CONFIG_SFAA
 224         pthread_mutex_lock(&dst->io_wq.stat_lock);
 225         pthread_mutex_lock(&src->io_wq.stat_lock);
 226 #endif
 227
 228         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 229         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 230         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 231         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 232         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 233
 234 #ifndef CONFIG_SFAA
 235         pthread_mutex_unlock(&src->io_wq.stat_lock);
 236         pthread_mutex_unlock(&dst->io_wq.stat_lock);
 237 #endif
 238 }
 239
 240 static void update_accounting(struct submit_worker *sw)
 241 {
 242         struct thread_data *src = &sw->td;
 243         struct thread_data *dst = sw->wq->td;
 244
 245         if (td_read(src))
 246                 sum_ddir(dst, src, DDIR_READ);
 247         if (td_write(src))
 248                 sum_ddir(dst, src, DDIR_WRITE);
 249         if (td_trim(src))
 250                 sum_ddir(dst, src, DDIR_TRIM);
 251 }
 252
 253 static void *worker_thread(void *data)
 254 {
 255         struct submit_worker *sw = data;
 256         struct workqueue *wq = sw->wq;
 257         struct thread_data *td = &sw->td;
 258         unsigned int eflags = 0, ret;
 259         FLIST_HEAD(local_list);
 260
 261         ret = init_submit_worker(sw);
 262         pthread_mutex_lock(&sw->lock);
 263         sw->flags |= SW_F_RUNNING;
 264         if (ret)
 265                 sw->flags |= SW_F_ERROR;
 266         pthread_mutex_unlock(&sw->lock);
 267
 268         pthread_mutex_lock(&wq->flush_lock);
 269         pthread_cond_signal(&wq->flush_cond);
 270         pthread_mutex_unlock(&wq->flush_lock);
 271
 272         if (sw->flags & SW_F_ERROR)
 273                 goto done;
 274
 275         while (1) {
 276                 pthread_mutex_lock(&sw->lock);
 277
 278                 if (flist_empty(&sw->work_list)) {
 279                         if (sw->flags & SW_F_EXIT) {
 280                                 pthread_mutex_unlock(&sw->lock);
 281                                 break;
 282                         }
 283
 284                         if (td->io_u_queued || td->cur_depth ||
 285                             td->io_u_in_flight) {
 286                                 pthread_mutex_unlock(&sw->lock);
 287                                 io_u_quiesce(td);
 288                                 pthread_mutex_lock(&sw->lock);
 289                         }
 290
 291                         /*
 292                          * We dropped and reaquired the lock, check
 293                          * state again.
 294                          */
 295                         if (!flist_empty(&sw->work_list))
 296                                 goto handle_work;
 297
 298                         if (sw->flags & SW_F_EXIT) {
 299                                 pthread_mutex_unlock(&sw->lock);
 300                                 break;
 301                         } else if (!(sw->flags & SW_F_IDLE)) {
 302                                 sw->flags |= SW_F_IDLE;
 303                                 wq->next_free_worker = sw->index;
 304                                 if (wq->wake_idle)
 305                                         pthread_cond_signal(&wq->flush_cond);
 306                         }
 307                         update_accounting(sw);
 308                         pthread_cond_wait(&sw->cond, &sw->lock);
 309                 } else {
 310 handle_work:
 311                         flist_splice_init(&sw->work_list, &local_list);
 312                 }
 313                 pthread_mutex_unlock(&sw->lock);
 314                 handle_list(sw, &local_list);
 315         }
 316
 317         update_accounting(sw);
 318
 319 done:
 320         pthread_mutex_lock(&sw->lock);
 321         sw->flags |= (SW_F_EXITED | eflags);
 322         pthread_mutex_unlock(&sw->lock);
 323         return NULL;
 324 }
 325
 326 static void free_worker(struct submit_worker *sw)
 327 {
 328         struct thread_data *td = &sw->td;
 329
 330         fio_options_free(td);
 331         close_and_free_files(td);
 332         if (td->io_ops)
 333                 close_ioengine(td);
 334         td_set_runstate(td, TD_EXITED);
 335
 336         pthread_cond_destroy(&sw->cond);
 337         pthread_mutex_destroy(&sw->lock);
 338 }
 339
 340 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 341 {
 342         struct thread_data *parent = sw->wq->td;
 343
 344         pthread_join(sw->thread, NULL);
 345         (*sum_cnt)++;
 346         sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt);
 347         free_worker(sw);
 348 }
 349
 350 void workqueue_exit(struct workqueue *wq)
 351 {
 352         unsigned int shutdown, sum_cnt = 0;
 353         struct submit_worker *sw;
 354         int i;
 355
 356         for (i = 0; i < wq->max_workers; i++) {
 357                 sw = &wq->workers[i];
 358
 359                 pthread_mutex_lock(&sw->lock);
 360                 sw->flags |= SW_F_EXIT;
 361                 pthread_cond_signal(&sw->cond);
 362                 pthread_mutex_unlock(&sw->lock);
 363         }
 364
 365         do {
 366                 shutdown = 0;
 367                 for (i = 0; i < wq->max_workers; i++) {
 368                         sw = &wq->workers[i];
 369                         if (sw->flags & SW_F_ACCOUNTED)
 370                                 continue;
 371                         sw->flags |= SW_F_ACCOUNTED;
 372                         shutdown_worker(sw, &sum_cnt);
 373                         shutdown++;
 374                 }
 375         } while (shutdown && shutdown != wq->max_workers);
 376
 377         free(wq->workers);
 378         pthread_mutex_destroy(&wq->flush_lock);
 379         pthread_cond_destroy(&wq->flush_cond);
 380         pthread_mutex_destroy(&wq->stat_lock);
 381 }
 382
 383 static int start_worker(struct workqueue *wq, unsigned int index)
 384 {
 385         struct submit_worker *sw = &wq->workers[index];
 386         int ret;
 387
 388         INIT_FLIST_HEAD(&sw->work_list);
 389         pthread_cond_init(&sw->cond, NULL);
 390         pthread_mutex_init(&sw->lock, NULL);
 391         sw->wq = wq;
 392         sw->index = index;
 393
 394         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 395         if (!ret) {
 396                 pthread_mutex_lock(&sw->lock);
 397                 sw->flags = SW_F_IDLE;
 398                 pthread_mutex_unlock(&sw->lock);
 399                 return 0;
 400         }
 401
 402         free_worker(sw);
 403         return 1;
 404 }
 405
 406 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 407                    workqueue_fn *fn, unsigned max_pending)
 408 {
 409         unsigned int running;
 410         int i, error;
 411
 412         wq->max_workers = max_pending;
 413         wq->td = td;
 414         wq->fn = fn;
 415         wq->work_seq = 0;
 416         wq->next_free_worker = 0;
 417         pthread_cond_init(&wq->flush_cond, NULL);
 418         pthread_mutex_init(&wq->flush_lock, NULL);
 419         pthread_mutex_init(&wq->stat_lock, NULL);
 420
 421         wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
 422
 423         for (i = 0; i < wq->max_workers; i++)
 424                 if (start_worker(wq, i))
 425                         break;
 426
 427         wq->max_workers = i;
 428         if (!wq->max_workers)
 429                 goto err;
 430
 431         /*
 432          * Wait for them all to be started and initialized
 433          */
 434         error = 0;
 435         do {
 436                 struct submit_worker *sw;
 437
 438                 running = 0;
 439                 pthread_mutex_lock(&wq->flush_lock);
 440                 for (i = 0; i < wq->max_workers; i++) {
 441                         sw = &wq->workers[i];
 442                         pthread_mutex_lock(&sw->lock);
 443                         if (sw->flags & SW_F_RUNNING)
 444                                 running++;
 445                         if (sw->flags & SW_F_ERROR)
 446                                 error++;
 447                         pthread_mutex_unlock(&sw->lock);
 448                 }
 449
 450                 if (error || running == wq->max_workers) {
 451                         pthread_mutex_unlock(&wq->flush_lock);
 452                         break;
 453                 }
 454
 455                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 456                 pthread_mutex_unlock(&wq->flush_lock);
 457         } while (1);
 458
 459         if (!error)
 460                 return 0;
 461
 462 err:
 463         log_err("Can't create rate workqueue\n");
 464         td_verror(td, ESRCH, "workqueue_init");
 465         workqueue_exit(wq);
 466         return 1;
 467 }