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