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 struct submit_worker {
  15         pthread_t thread;
  16         pthread_mutex_t lock;
  17         pthread_cond_t cond;
  18         struct flist_head work_list;
  19         unsigned int flags;
  20         unsigned int index;
  21         uint64_t seq;
  22         struct workqueue *wq;
  23         struct thread_data td;
  24 };
  25
  26 enum {
  27         SW_F_IDLE       = 1 << 0,
  28         SW_F_RUNNING    = 1 << 1,
  29         SW_F_EXIT       = 1 << 2,
  30         SW_F_EXITED     = 1 << 3,
  31         SW_F_ACCOUNTED  = 1 << 4,
  32         SW_F_ERROR      = 1 << 5,
  33 };
  34
  35 static struct submit_worker *__get_submit_worker(struct workqueue *wq,
  36                                                  unsigned int start,
  37                                                  unsigned int end,
  38                                                  struct submit_worker **best)
  39 {
  40         struct submit_worker *sw = NULL;
  41
  42         while (start <= end) {
  43                 sw = &wq->workers[start];
  44                 if (sw->flags & SW_F_IDLE)
  45                         return sw;
  46                 if (!(*best) || sw->seq < (*best)->seq)
  47                         *best = sw;
  48                 start++;
  49         }
  50
  51         return NULL;
  52 }
  53
  54 static struct submit_worker *get_submit_worker(struct workqueue *wq)
  55 {
  56         unsigned int next = wq->next_free_worker;
  57         struct submit_worker *sw, *best = NULL;
  58
  59         assert(next < wq->max_workers);
  60
  61         sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
  62         if (!sw && next)
  63                 sw = __get_submit_worker(wq, 0, next - 1, &best);
  64
  65         /*
  66          * No truly idle found, use best match
  67          */
  68         if (!sw)
  69                 sw = best;
  70
  71         if (sw->index == wq->next_free_worker) {
  72                 if (sw->index + 1 < wq->max_workers)
  73                         wq->next_free_worker = sw->index + 1;
  74                 else
  75                         wq->next_free_worker = 0;
  76         }
  77
  78         return sw;
  79 }
  80
  81 static bool all_sw_idle(struct workqueue *wq)
  82 {
  83         int i;
  84
  85         for (i = 0; i < wq->max_workers; i++) {
  86                 struct submit_worker *sw = &wq->workers[i];
  87
  88                 if (!(sw->flags & SW_F_IDLE))
  89                         return false;
  90         }
  91
  92         return true;
  93 }
  94
  95 /*
  96  * Must be serialized wrt workqueue_enqueue() by caller
  97  */
  98 void workqueue_flush(struct workqueue *wq)
  99 {
 100         wq->wake_idle = 1;
 101
 102         while (!all_sw_idle(wq)) {
 103                 pthread_mutex_lock(&wq->flush_lock);
 104                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 105                 pthread_mutex_unlock(&wq->flush_lock);
 106         }
 107
 108         wq->wake_idle = 0;
 109 }
 110
 111 /*
 112  * Must be serialized by caller. Returns true for queued, false for busy.
 113  */
 114 bool workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
 115 {
 116         struct submit_worker *sw;
 117
 118         sw = get_submit_worker(wq);
 119         if (sw) {
 120                 pthread_mutex_lock(&sw->lock);
 121                 flist_add_tail(&work->list, &sw->work_list);
 122                 sw->seq = ++wq->work_seq;
 123                 sw->flags &= ~SW_F_IDLE;
 124                 pthread_mutex_unlock(&sw->lock);
 125
 126                 pthread_cond_signal(&sw->cond);
 127                 return true;
 128         }
 129
 130         return false;
 131 }
 132
 133 static void handle_list(struct submit_worker *sw, struct flist_head *list)
 134 {
 135         struct workqueue *wq = sw->wq;
 136         struct workqueue_work *work;
 137
 138         while (!flist_empty(list)) {
 139                 work = flist_first_entry(list, struct workqueue_work, list);
 140                 flist_del_init(&work->list);
 141                 wq->fn(&sw->td, work);
 142         }
 143 }
 144
 145 static int init_submit_worker(struct submit_worker *sw)
 146 {
 147         struct thread_data *parent = sw->wq->td;
 148         struct thread_data *td = &sw->td;
 149         int fio_unused ret;
 150
 151         memcpy(&td->o, &parent->o, sizeof(td->o));
 152         memcpy(&td->ts, &parent->ts, sizeof(td->ts));
 153         td->o.uid = td->o.gid = -1U;
 154         dup_files(td, parent);
 155         td->eo = parent->eo;
 156         fio_options_mem_dupe(td);
 157
 158         if (ioengine_load(td))
 159                 goto err;
 160
 161         if (td->o.odirect)
 162                 td->io_ops->flags |= FIO_RAWIO;
 163
 164         td->pid = gettid();
 165
 166         INIT_FLIST_HEAD(&td->io_log_list);
 167         INIT_FLIST_HEAD(&td->io_hist_list);
 168         INIT_FLIST_HEAD(&td->verify_list);
 169         INIT_FLIST_HEAD(&td->trim_list);
 170         INIT_FLIST_HEAD(&td->next_rand_list);
 171         td->io_hist_tree = RB_ROOT;
 172
 173         td->o.iodepth = 1;
 174         if (td_io_init(td))
 175                 goto err_io_init;
 176
 177         fio_gettime(&td->epoch, NULL);
 178         fio_getrusage(&td->ru_start);
 179         clear_io_state(td, 1);
 180
 181         td_set_runstate(td, TD_RUNNING);
 182         td->flags |= TD_F_CHILD;
 183         td->parent = parent;
 184         return 0;
 185
 186 err_io_init:
 187         close_ioengine(td);
 188 err:
 189         return 1;
 190 }
 191
 192 #ifdef CONFIG_SFAA
 193 static void sum_val(uint64_t *dst, uint64_t *src)
 194 {
 195         if (*src) {
 196                 __sync_fetch_and_add(dst, *src);
 197                 *src = 0;
 198         }
 199 }
 200 #else
 201 static void sum_val(uint64_t *dst, uint64_t *src)
 202 {
 203         if (*src) {
 204                 *dst += *src;
 205                 *src = 0;
 206         }
 207 }
 208 #endif
 209
 210 static void pthread_double_unlock(pthread_mutex_t *lock1,
 211                                   pthread_mutex_t *lock2)
 212 {
 213 #ifndef CONFIG_SFAA
 214         pthread_mutex_unlock(lock1);
 215         pthread_mutex_unlock(lock2);
 216 #endif
 217 }
 218
 219 static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
 220 {
 221 #ifndef CONFIG_SFAA
 222         if (lock1 < lock2) {
 223                 pthread_mutex_lock(lock1);
 224                 pthread_mutex_lock(lock2);
 225         } else {
 226                 pthread_mutex_lock(lock2);
 227                 pthread_mutex_lock(lock1);
 228         }
 229 #endif
 230 }
 231
 232 static void sum_ddir(struct thread_data *dst, struct thread_data *src,
 233                      enum fio_ddir ddir)
 234 {
 235         pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 236
 237         sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
 238         sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
 239         sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
 240         sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
 241         sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
 242
 243         pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
 244 }
 245
 246 static void update_accounting(struct submit_worker *sw)
 247 {
 248         struct thread_data *src = &sw->td;
 249         struct thread_data *dst = sw->wq->td;
 250
 251         if (td_read(src))
 252                 sum_ddir(dst, src, DDIR_READ);
 253         if (td_write(src))
 254                 sum_ddir(dst, src, DDIR_WRITE);
 255         if (td_trim(src))
 256                 sum_ddir(dst, src, DDIR_TRIM);
 257 }
 258
 259 static void *worker_thread(void *data)
 260 {
 261         struct submit_worker *sw = data;
 262         struct workqueue *wq = sw->wq;
 263         struct thread_data *td = &sw->td;
 264         unsigned int eflags = 0, ret;
 265         FLIST_HEAD(local_list);
 266
 267         ret = init_submit_worker(sw);
 268         pthread_mutex_lock(&sw->lock);
 269         sw->flags |= SW_F_RUNNING;
 270         if (ret)
 271                 sw->flags |= SW_F_ERROR;
 272         pthread_mutex_unlock(&sw->lock);
 273
 274         pthread_mutex_lock(&wq->flush_lock);
 275         pthread_cond_signal(&wq->flush_cond);
 276         pthread_mutex_unlock(&wq->flush_lock);
 277
 278         if (sw->flags & SW_F_ERROR)
 279                 goto done;
 280
 281         while (1) {
 282                 pthread_mutex_lock(&sw->lock);
 283
 284                 if (flist_empty(&sw->work_list)) {
 285                         if (sw->flags & SW_F_EXIT) {
 286                                 pthread_mutex_unlock(&sw->lock);
 287                                 break;
 288                         }
 289
 290                         if (td->io_u_queued || td->cur_depth ||
 291                             td->io_u_in_flight) {
 292                                 int ret;
 293
 294                                 pthread_mutex_unlock(&sw->lock);
 295                                 ret = io_u_quiesce(td);
 296                                 if (ret > 0)
 297                                         td->cur_depth -= ret;
 298                                 pthread_mutex_lock(&sw->lock);
 299                         }
 300
 301                         /*
 302                          * We dropped and reaquired the lock, check
 303                          * state again.
 304                          */
 305                         if (!flist_empty(&sw->work_list))
 306                                 goto handle_work;
 307
 308                         if (sw->flags & SW_F_EXIT) {
 309                                 pthread_mutex_unlock(&sw->lock);
 310                                 break;
 311                         } else if (!(sw->flags & SW_F_IDLE)) {
 312                                 sw->flags |= SW_F_IDLE;
 313                                 wq->next_free_worker = sw->index;
 314                                 if (wq->wake_idle)
 315                                         pthread_cond_signal(&wq->flush_cond);
 316                         }
 317                         update_accounting(sw);
 318                         pthread_cond_wait(&sw->cond, &sw->lock);
 319                 } else {
 320 handle_work:
 321                         flist_splice_init(&sw->work_list, &local_list);
 322                 }
 323                 pthread_mutex_unlock(&sw->lock);
 324                 handle_list(sw, &local_list);
 325         }
 326
 327         update_accounting(sw);
 328
 329 done:
 330         pthread_mutex_lock(&sw->lock);
 331         sw->flags |= (SW_F_EXITED | eflags);
 332         pthread_mutex_unlock(&sw->lock);
 333         return NULL;
 334 }
 335
 336 static void free_worker(struct submit_worker *sw)
 337 {
 338         struct thread_data *td = &sw->td;
 339
 340         fio_options_free(td);
 341         close_and_free_files(td);
 342         if (td->io_ops)
 343                 close_ioengine(td);
 344         td_set_runstate(td, TD_EXITED);
 345
 346         pthread_cond_destroy(&sw->cond);
 347         pthread_mutex_destroy(&sw->lock);
 348 }
 349
 350 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 351 {
 352         struct thread_data *parent = sw->wq->td;
 353
 354         pthread_join(sw->thread, NULL);
 355         (*sum_cnt)++;
 356         sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt == 1);
 357         free_worker(sw);
 358 }
 359
 360 void workqueue_exit(struct workqueue *wq)
 361 {
 362         unsigned int shutdown, sum_cnt = 0;
 363         struct submit_worker *sw;
 364         int i;
 365
 366         for (i = 0; i < wq->max_workers; i++) {
 367                 sw = &wq->workers[i];
 368
 369                 pthread_mutex_lock(&sw->lock);
 370                 sw->flags |= SW_F_EXIT;
 371                 pthread_cond_signal(&sw->cond);
 372                 pthread_mutex_unlock(&sw->lock);
 373         }
 374
 375         do {
 376                 shutdown = 0;
 377                 for (i = 0; i < wq->max_workers; i++) {
 378                         sw = &wq->workers[i];
 379                         if (sw->flags & SW_F_ACCOUNTED)
 380                                 continue;
 381                         pthread_mutex_lock(&sw->lock);
 382                         sw->flags |= SW_F_ACCOUNTED;
 383                         pthread_mutex_unlock(&sw->lock);
 384                         shutdown_worker(sw, &sum_cnt);
 385                         shutdown++;
 386                 }
 387         } while (shutdown && shutdown != wq->max_workers);
 388
 389         free(wq->workers);
 390         pthread_mutex_destroy(&wq->flush_lock);
 391         pthread_cond_destroy(&wq->flush_cond);
 392         pthread_mutex_destroy(&wq->stat_lock);
 393 }
 394
 395 static int start_worker(struct workqueue *wq, unsigned int index)
 396 {
 397         struct submit_worker *sw = &wq->workers[index];
 398         int ret;
 399
 400         INIT_FLIST_HEAD(&sw->work_list);
 401         pthread_cond_init(&sw->cond, NULL);
 402         pthread_mutex_init(&sw->lock, NULL);
 403         sw->wq = wq;
 404         sw->index = index;
 405
 406         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 407         if (!ret) {
 408                 pthread_mutex_lock(&sw->lock);
 409                 sw->flags = SW_F_IDLE;
 410                 pthread_mutex_unlock(&sw->lock);
 411                 return 0;
 412         }
 413
 414         free_worker(sw);
 415         return 1;
 416 }
 417
 418 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 419                    workqueue_fn *fn, unsigned max_pending)
 420 {
 421         unsigned int running;
 422         int i, error;
 423
 424         wq->max_workers = max_pending;
 425         wq->td = td;
 426         wq->fn = fn;
 427         wq->work_seq = 0;
 428         wq->next_free_worker = 0;
 429         pthread_cond_init(&wq->flush_cond, NULL);
 430         pthread_mutex_init(&wq->flush_lock, NULL);
 431         pthread_mutex_init(&wq->stat_lock, NULL);
 432
 433         wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
 434
 435         for (i = 0; i < wq->max_workers; i++)
 436                 if (start_worker(wq, i))
 437                         break;
 438
 439         wq->max_workers = i;
 440         if (!wq->max_workers)
 441                 goto err;
 442
 443         /*
 444          * Wait for them all to be started and initialized
 445          */
 446         error = 0;
 447         do {
 448                 struct submit_worker *sw;
 449
 450                 running = 0;
 451                 pthread_mutex_lock(&wq->flush_lock);
 452                 for (i = 0; i < wq->max_workers; i++) {
 453                         sw = &wq->workers[i];
 454                         pthread_mutex_lock(&sw->lock);
 455                         if (sw->flags & SW_F_RUNNING)
 456                                 running++;
 457                         if (sw->flags & SW_F_ERROR)
 458                                 error++;
 459                         pthread_mutex_unlock(&sw->lock);
 460                 }
 461
 462                 if (error || running == wq->max_workers) {
 463                         pthread_mutex_unlock(&wq->flush_lock);
 464                         break;
 465                 }
 466
 467                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 468                 pthread_mutex_unlock(&wq->flush_lock);
 469         } while (1);
 470
 471         if (!error)
 472                 return 0;
 473
 474 err:
 475         log_err("Can't create rate workqueue\n");
 476         td_verror(td, ESRCH, "workqueue_init");
 477         workqueue_exit(wq);
 478         return 1;
 479 }