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                 }
 128
 129                 pthread_mutex_lock(&sw->lock);
 130                 flist_add_tail(&io_u->verify_list, &sw->work_list);
 131                 sw->seq = ++wq->work_seq;
 132                 sw->flags &= ~SW_F_IDLE;
 133                 pthread_mutex_unlock(&sw->lock);
 134
 135                 pthread_cond_signal(&sw->cond);
 136                 return FIO_Q_QUEUED;
 137         }
 138
 139         return FIO_Q_BUSY;
 140 }
 141
 142 static void handle_list(struct submit_worker *sw, struct flist_head *list)
 143 {
 144         struct workqueue *wq = sw->wq;
 145         struct io_u *io_u;
 146
 147         while (!flist_empty(list)) {
 148                 io_u = flist_first_entry(list, struct io_u, verify_list);
 149                 flist_del_init(&io_u->verify_list);
 150                 wq->fn(&sw->td, io_u);
 151         }
 152 }
 153
 154 static int init_submit_worker(struct submit_worker *sw)
 155 {
 156         struct thread_data *parent = sw->wq->td;
 157         struct thread_data *td = &sw->td;
 158         int fio_unused ret;
 159
 160         memcpy(&td->o, &parent->o, sizeof(td->o));
 161         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
 162         td->o.uid = td->o.gid = -1U;
 163         dup_files(td, parent);
 164         fio_options_mem_dupe(td);
 165
 166         if (ioengine_load(td))
 167                 goto err;
 168
 169         if (td->o.odirect)
 170                 td->io_ops->flags |= FIO_RAWIO;
 171
 172         td->pid = gettid();
 173
 174         INIT_FLIST_HEAD(&td->io_log_list);
 175         INIT_FLIST_HEAD(&td->io_hist_list);
 176         INIT_FLIST_HEAD(&td->verify_list);
 177         INIT_FLIST_HEAD(&td->trim_list);
 178         INIT_FLIST_HEAD(&td->next_rand_list);
 179         td->io_hist_tree = RB_ROOT;
 180
 181         td->o.iodepth = 1;
 182         if (td_io_init(td))
 183                 goto err_io_init;
 184
 185         fio_gettime(&td->epoch, NULL);
 186         fio_getrusage(&td->ru_start);
 187         clear_io_state(td);
 188
 189         td_set_runstate(td, TD_RUNNING);
 190         td->flags |= TD_F_CHILD;
 191         td->parent = parent;
 192         return 0;
 193
 194 err_io_init:
 195         close_ioengine(td);
 196 err:
 197         return 1;
 198 }
 199
 200 #ifdef CONFIG_SFAA
 201 static void sum_val(uint64_t *dst, uint64_t *src)
 202 {
 203         if (*src) {
 204                 __sync_fetch_and_add(dst, *src);
 205                 *src = 0;
 206         }
 207 }
 208 #else
 209 static void sum_val(uint64_t *dst, uint64_t *src)
 210 {
 211         if (*src) {
 212                 *dst += *src;
 213                 *src = 0;
 214         }
 215 }
 216 #endif
 217
 218 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 219                      enum fio_ddir ddir)
 220 {
 221 #ifndef CONFIG_SFAA
 222         pthread_mutex_lock(&dst->io_wq.stat_lock);
 223         pthread_mutex_lock(&src->io_wq.stat_lock);
 224 #endif
 225
 226         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 227         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 228         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 229         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 230         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 231
 232 #ifndef CONFIG_SFAA
 233         pthread_mutex_unlock(&src->io_wq.stat_lock);
 234         pthread_mutex_unlock(&dst->io_wq.stat_lock);
 235 #endif
 236 }
 237
 238 static void update_accounting(struct submit_worker *sw)
 239 {
 240         struct thread_data *src = &sw->td;
 241         struct thread_data *dst = sw->wq->td;
 242
 243         if (td_read(src))
 244                 sum_ddir(dst, src, DDIR_READ);
 245         if (td_write(src))
 246                 sum_ddir(dst, src, DDIR_WRITE);
 247         if (td_trim(src))
 248                 sum_ddir(dst, src, DDIR_TRIM);
 249 }
 250
 251 static void *worker_thread(void *data)
 252 {
 253         struct submit_worker *sw = data;
 254         struct workqueue *wq = sw->wq;
 255         struct thread_data *td = &sw->td;
 256         unsigned int eflags = 0, ret;
 257         FLIST_HEAD(local_list);
 258
 259         ret = init_submit_worker(sw);
 260         pthread_mutex_lock(&sw->lock);
 261         sw->flags |= SW_F_RUNNING;
 262         if (ret)
 263                 sw->flags |= SW_F_ERROR;
 264         pthread_mutex_unlock(&sw->lock);
 265
 266         pthread_mutex_lock(&wq->flush_lock);
 267         pthread_cond_signal(&wq->flush_cond);
 268         pthread_mutex_unlock(&wq->flush_lock);
 269
 270         if (sw->flags & SW_F_ERROR)
 271                 goto done;
 272
 273         while (1) {
 274                 pthread_mutex_lock(&sw->lock);
 275
 276                 if (flist_empty(&sw->work_list)) {
 277                         if (sw->flags & SW_F_EXIT) {
 278                                 pthread_mutex_unlock(&sw->lock);
 279                                 break;
 280                         }
 281
 282                         if (td->io_u_queued || td->cur_depth ||
 283                             td->io_u_in_flight) {
 284                                 pthread_mutex_unlock(&sw->lock);
 285                                 io_u_quiesce(td);
 286                                 pthread_mutex_lock(&sw->lock);
 287                         }
 288
 289                         /*
 290                          * We dropped and reaquired the lock, check
 291                          * state again.
 292                          */
 293                         if (!flist_empty(&sw->work_list))
 294                                 goto handle_work;
 295
 296                         if (sw->flags & SW_F_EXIT) {
 297                                 pthread_mutex_unlock(&sw->lock);
 298                                 break;
 299                         } else if (!(sw->flags & SW_F_IDLE)) {
 300                                 sw->flags |= SW_F_IDLE;
 301                                 wq->next_free_worker = sw->index;
 302                                 if (wq->wake_idle)
 303                                         pthread_cond_signal(&wq->flush_cond);
 304                         }
 305                         update_accounting(sw);
 306                         pthread_cond_wait(&sw->cond, &sw->lock);
 307                 } else {
 308 handle_work:
 309                         flist_splice_init(&sw->work_list, &local_list);
 310                 }
 311                 pthread_mutex_unlock(&sw->lock);
 312                 handle_list(sw, &local_list);
 313         }
 314
 315         update_accounting(sw);
 316
 317 done:
 318         pthread_mutex_lock(&sw->lock);
 319         sw->flags |= (SW_F_EXITED | eflags);
 320         pthread_mutex_unlock(&sw->lock);
 321         return NULL;
 322 }
 323
 324 static void free_worker(struct submit_worker *sw)
 325 {
 326         struct thread_data *td = &sw->td;
 327
 328         fio_options_free(td);
 329         close_and_free_files(td);
 330         if (td->io_ops)
 331                 close_ioengine(td);
 332         td_set_runstate(td, TD_EXITED);
 333
 334         pthread_cond_destroy(&sw->cond);
 335         pthread_mutex_destroy(&sw->lock);
 336 }
 337
 338 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 339 {
 340         struct thread_data *parent = sw->wq->td;
 341
 342         pthread_join(sw->thread, NULL);
 343         (*sum_cnt)++;
 344         sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt);
 345         free_worker(sw);
 346 }
 347
 348 void workqueue_exit(struct workqueue *wq)
 349 {
 350         unsigned int shutdown, sum_cnt = 0;
 351         struct submit_worker *sw;
 352         int i;
 353
 354         for (i = 0; i < wq->max_workers; i++) {
 355                 sw = &wq->workers[i];
 356
 357                 pthread_mutex_lock(&sw->lock);
 358                 sw->flags |= SW_F_EXIT;
 359                 pthread_cond_signal(&sw->cond);
 360                 pthread_mutex_unlock(&sw->lock);
 361         }
 362
 363         do {
 364                 shutdown = 0;
 365                 for (i = 0; i < wq->max_workers; i++) {
 366                         sw = &wq->workers[i];
 367                         if (sw->flags & SW_F_ACCOUNTED)
 368                                 continue;
 369                         sw->flags |= SW_F_ACCOUNTED;
 370                         shutdown_worker(sw, &sum_cnt);
 371                         shutdown++;
 372                 }
 373         } while (shutdown && shutdown != wq->max_workers);
 374
 375         free(wq->workers);
 376         pthread_mutex_destroy(&wq->flush_lock);
 377         pthread_cond_destroy(&wq->flush_cond);
 378         pthread_mutex_destroy(&wq->stat_lock);
 379 }
 380
 381 static int start_worker(struct workqueue *wq, unsigned int index)
 382 {
 383         struct submit_worker *sw = &wq->workers[index];
 384         int ret;
 385
 386         INIT_FLIST_HEAD(&sw->work_list);
 387         pthread_cond_init(&sw->cond, NULL);
 388         pthread_mutex_init(&sw->lock, NULL);
 389         sw->wq = wq;
 390         sw->index = index;
 391
 392         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 393         if (!ret) {
 394                 pthread_mutex_lock(&sw->lock);
 395                 sw->flags = SW_F_IDLE;
 396                 pthread_mutex_unlock(&sw->lock);
 397                 return 0;
 398         }
 399
 400         free_worker(sw);
 401         return 1;
 402 }
 403
 404 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 405                    workqueue_fn *fn, unsigned max_pending)
 406 {
 407         unsigned int running;
 408         int i, error;
 409
 410         wq->max_workers = max_pending;
 411         wq->td = td;
 412         wq->fn = fn;
 413         wq->work_seq = 0;
 414         wq->next_free_worker = 0;
 415         pthread_cond_init(&wq->flush_cond, NULL);
 416         pthread_mutex_init(&wq->flush_lock, NULL);
 417         pthread_mutex_init(&wq->stat_lock, NULL);
 418
 419         wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
 420
 421         for (i = 0; i < wq->max_workers; i++)
 422                 if (start_worker(wq, i))
 423                         break;
 424
 425         wq->max_workers = i;
 426         if (!wq->max_workers)
 427                 goto err;
 428
 429         /*
 430          * Wait for them all to be started and initialized
 431          */
 432         error = 0;
 433         do {
 434                 struct submit_worker *sw;
 435
 436                 running = 0;
 437                 pthread_mutex_lock(&wq->flush_lock);
 438                 for (i = 0; i < wq->max_workers; i++) {
 439                         sw = &wq->workers[i];
 440                         pthread_mutex_lock(&sw->lock);
 441                         if (sw->flags & SW_F_RUNNING)
 442                                 running++;
 443                         if (sw->flags & SW_F_ERROR)
 444                                 error++;
 445                         pthread_mutex_unlock(&sw->lock);
 446                 }
 447
 448                 if (error || running == wq->max_workers) {
 449                         pthread_mutex_unlock(&wq->flush_lock);
 450                         break;
 451                 }
 452
 453                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 454                 pthread_mutex_unlock(&wq->flush_lock);
 455         } while (1);
 456
 457         if (!error)
 458                 return 0;
 459
 460 err:
 461         log_err("Can't create rate workqueue\n");
 462         td_verror(td, ESRCH, "workqueue_init");
 463         workqueue_exit(wq);
 464         return 1;
 465 }