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 void *worker_thread(void *data)
 134 {
 135         struct submit_worker *sw = data;
 136         struct workqueue *wq = sw->wq;
 137         unsigned int eflags = 0, ret = 0;
 138         FLIST_HEAD(local_list);
 139
 140         if (wq->ops.nice) {
 141                 if (nice(wq->ops.nice) < 0) {
 142                         log_err("workqueue: nice %s\n", strerror(errno));
 143                         ret = 1;
 144                 }
 145         }
 146
 147         if (!ret)
 148                 ret = workqueue_init_worker(sw);
 149
 150         pthread_mutex_lock(&sw->lock);
 151         sw->flags |= SW_F_RUNNING;
 152         if (ret)
 153                 sw->flags |= SW_F_ERROR;
 154         pthread_mutex_unlock(&sw->lock);
 155
 156         pthread_mutex_lock(&wq->flush_lock);
 157         pthread_cond_signal(&wq->flush_cond);
 158         pthread_mutex_unlock(&wq->flush_lock);
 159
 160         if (sw->flags & SW_F_ERROR)
 161                 goto done;
 162
 163         while (1) {
 164                 pthread_mutex_lock(&sw->lock);
 165
 166                 if (flist_empty(&sw->work_list)) {
 167                         if (sw->flags & SW_F_EXIT) {
 168                                 pthread_mutex_unlock(&sw->lock);
 169                                 break;
 170                         }
 171
 172                         if (workqueue_pre_sleep_check(sw)) {
 173                                 pthread_mutex_unlock(&sw->lock);
 174                                 workqueue_pre_sleep(sw);
 175                                 pthread_mutex_lock(&sw->lock);
 176                         }
 177
 178                         /*
 179                          * We dropped and reaquired the lock, check
 180                          * state again.
 181                          */
 182                         if (!flist_empty(&sw->work_list))
 183                                 goto handle_work;
 184
 185                         if (sw->flags & SW_F_EXIT) {
 186                                 pthread_mutex_unlock(&sw->lock);
 187                                 break;
 188                         } else if (!(sw->flags & SW_F_IDLE)) {
 189                                 sw->flags |= SW_F_IDLE;
 190                                 wq->next_free_worker = sw->index;
 191                                 if (wq->wake_idle)
 192                                         pthread_cond_signal(&wq->flush_cond);
 193                         }
 194                         if (wq->ops.update_acct_fn)
 195                                 wq->ops.update_acct_fn(sw);
 196
 197                         pthread_cond_wait(&sw->cond, &sw->lock);
 198                 } else {
 199 handle_work:
 200                         flist_splice_init(&sw->work_list, &local_list);
 201                 }
 202                 pthread_mutex_unlock(&sw->lock);
 203                 handle_list(sw, &local_list);
 204         }
 205
 206         if (wq->ops.update_acct_fn)
 207                 wq->ops.update_acct_fn(sw);
 208
 209 done:
 210         pthread_mutex_lock(&sw->lock);
 211         sw->flags |= (SW_F_EXITED | eflags);
 212         pthread_mutex_unlock(&sw->lock);
 213         return NULL;
 214 }
 215
 216 static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 217 {
 218         struct workqueue *wq = sw->wq;
 219
 220         workqueue_exit_worker(sw, sum_cnt);
 221
 222         pthread_cond_destroy(&sw->cond);
 223         pthread_mutex_destroy(&sw->lock);
 224
 225         if (wq->ops.free_worker_fn)
 226                 wq->ops.free_worker_fn(sw);
 227 }
 228
 229 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 230 {
 231         pthread_join(sw->thread, NULL);
 232         free_worker(sw, sum_cnt);
 233 }
 234
 235 void workqueue_exit(struct workqueue *wq)
 236 {
 237         unsigned int shutdown, sum_cnt = 0;
 238         struct submit_worker *sw;
 239         int i;
 240
 241         for (i = 0; i < wq->max_workers; i++) {
 242                 sw = &wq->workers[i];
 243
 244                 pthread_mutex_lock(&sw->lock);
 245                 sw->flags |= SW_F_EXIT;
 246                 pthread_cond_signal(&sw->cond);
 247                 pthread_mutex_unlock(&sw->lock);
 248         }
 249
 250         do {
 251                 shutdown = 0;
 252                 for (i = 0; i < wq->max_workers; i++) {
 253                         sw = &wq->workers[i];
 254                         if (sw->flags & SW_F_ACCOUNTED)
 255                                 continue;
 256                         pthread_mutex_lock(&sw->lock);
 257                         sw->flags |= SW_F_ACCOUNTED;
 258                         pthread_mutex_unlock(&sw->lock);
 259                         shutdown_worker(sw, &sum_cnt);
 260                         shutdown++;
 261                 }
 262         } while (shutdown && shutdown != wq->max_workers);
 263
 264         free(wq->workers);
 265         pthread_mutex_destroy(&wq->flush_lock);
 266         pthread_cond_destroy(&wq->flush_cond);
 267         pthread_mutex_destroy(&wq->stat_lock);
 268 }
 269
 270 static int start_worker(struct workqueue *wq, unsigned int index)
 271 {
 272         struct submit_worker *sw = &wq->workers[index];
 273         int ret;
 274
 275         INIT_FLIST_HEAD(&sw->work_list);
 276         pthread_cond_init(&sw->cond, NULL);
 277         pthread_mutex_init(&sw->lock, NULL);
 278         sw->wq = wq;
 279         sw->index = index;
 280
 281         if (wq->ops.alloc_worker_fn) {
 282                 ret = wq->ops.alloc_worker_fn(sw);
 283                 if (ret)
 284                         return ret;
 285         }
 286
 287         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 288         if (!ret) {
 289                 pthread_mutex_lock(&sw->lock);
 290                 sw->flags = SW_F_IDLE;
 291                 pthread_mutex_unlock(&sw->lock);
 292                 return 0;
 293         }
 294
 295         free_worker(sw, NULL);
 296         return 1;
 297 }
 298
 299 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 300                    struct workqueue_ops *ops, unsigned max_pending)
 301 {
 302         unsigned int running;
 303         int i, error;
 304
 305         wq->max_workers = max_pending;
 306         wq->td = td;
 307         wq->ops = *ops;
 308         wq->work_seq = 0;
 309         wq->next_free_worker = 0;
 310         pthread_cond_init(&wq->flush_cond, NULL);
 311         pthread_mutex_init(&wq->flush_lock, NULL);
 312         pthread_mutex_init(&wq->stat_lock, NULL);
 313
 314         wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
 315
 316         for (i = 0; i < wq->max_workers; i++)
 317                 if (start_worker(wq, i))
 318                         break;
 319
 320         wq->max_workers = i;
 321         if (!wq->max_workers)
 322                 goto err;
 323
 324         /*
 325          * Wait for them all to be started and initialized
 326          */
 327         error = 0;
 328         do {
 329                 struct submit_worker *sw;
 330
 331                 running = 0;
 332                 pthread_mutex_lock(&wq->flush_lock);
 333                 for (i = 0; i < wq->max_workers; i++) {
 334                         sw = &wq->workers[i];
 335                         pthread_mutex_lock(&sw->lock);
 336                         if (sw->flags & SW_F_RUNNING)
 337                                 running++;
 338                         if (sw->flags & SW_F_ERROR)
 339                                 error++;
 340                         pthread_mutex_unlock(&sw->lock);
 341                 }
 342
 343                 if (error || running == wq->max_workers) {
 344                         pthread_mutex_unlock(&wq->flush_lock);
 345                         break;
 346                 }
 347
 348                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 349                 pthread_mutex_unlock(&wq->flush_lock);
 350         } while (1);
 351
 352         if (!error)
 353                 return 0;
 354
 355 err:
 356         log_err("Can't create rate workqueue\n");
 357         td_verror(td, ESRCH, "workqueue_init");
 358         workqueue_exit(wq);
 359         return 1;
 360 }