workqueue.c

   1 /*
   2  * Generic workqueue offload mechanism
   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 "smalloc.h"
  13 #include "pshared.h"
  14
  15 enum {
  16         SW_F_IDLE       = 1 << 0,
  17         SW_F_RUNNING    = 1 << 1,
  18         SW_F_EXIT       = 1 << 2,
  19         SW_F_ACCOUNTED  = 1 << 3,
  20         SW_F_ERROR      = 1 << 4,
  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         pthread_mutex_lock(&wq->flush_lock);
  89         wq->wake_idle = 1;
  90
  91         while (!all_sw_idle(wq))
  92                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
  93
  94         wq->wake_idle = 0;
  95         pthread_mutex_unlock(&wq->flush_lock);
  96 }
  97
  98 /*
  99  * Must be serialized by caller.
 100  */
 101 void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
 102 {
 103         struct submit_worker *sw;
 104
 105         sw = get_submit_worker(wq);
 106         assert(sw);
 107
 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
 113         pthread_cond_signal(&sw->cond);
 114         pthread_mutex_unlock(&sw->lock);
 115 }
 116
 117 static void handle_list(struct submit_worker *sw, struct flist_head *list)
 118 {
 119         struct workqueue *wq = sw->wq;
 120         struct workqueue_work *work;
 121
 122         while (!flist_empty(list)) {
 123                 work = flist_first_entry(list, struct workqueue_work, list);
 124                 flist_del_init(&work->list);
 125                 wq->ops.fn(sw, work);
 126         }
 127 }
 128
 129 static void *worker_thread(void *data)
 130 {
 131         struct submit_worker *sw = data;
 132         struct workqueue *wq = sw->wq;
 133         unsigned int ret = 0;
 134         FLIST_HEAD(local_list);
 135
 136         sk_out_assign(sw->sk_out);
 137
 138         if (wq->ops.nice) {
 139                 if (nice(wq->ops.nice) < 0) {
 140                         log_err("workqueue: nice %s\n", strerror(errno));
 141                         ret = 1;
 142                 }
 143         }
 144
 145         if (!ret)
 146                 ret = workqueue_init_worker(sw);
 147
 148         pthread_mutex_lock(&sw->lock);
 149         sw->flags |= SW_F_RUNNING;
 150         if (ret)
 151                 sw->flags |= SW_F_ERROR;
 152         pthread_mutex_unlock(&sw->lock);
 153
 154         pthread_mutex_lock(&wq->flush_lock);
 155         pthread_cond_signal(&wq->flush_cond);
 156         pthread_mutex_unlock(&wq->flush_lock);
 157
 158         if (sw->flags & SW_F_ERROR)
 159                 goto done;
 160
 161         pthread_mutex_lock(&sw->lock);
 162         while (1) {
 163                 if (flist_empty(&sw->work_list)) {
 164                         if (sw->flags & SW_F_EXIT) {
 165                                 break;
 166                         }
 167
 168                         if (workqueue_pre_sleep_check(sw)) {
 169                                 pthread_mutex_unlock(&sw->lock);
 170                                 workqueue_pre_sleep(sw);
 171                                 pthread_mutex_lock(&sw->lock);
 172                         }
 173
 174                         /*
 175                          * We dropped and reaquired the lock, check
 176                          * state again.
 177                          */
 178                         if (!flist_empty(&sw->work_list))
 179                                 goto handle_work;
 180
 181                         if (sw->flags & SW_F_EXIT) {
 182                                 break;
 183                         } else if (!(sw->flags & SW_F_IDLE)) {
 184                                 sw->flags |= SW_F_IDLE;
 185                                 wq->next_free_worker = sw->index;
 186                                 if (wq->wake_idle)
 187                                         pthread_cond_signal(&wq->flush_cond);
 188                         }
 189
 190                         pthread_cond_wait(&sw->cond, &sw->lock);
 191                 } else {
 192 handle_work:
 193                         flist_splice_init(&sw->work_list, &local_list);
 194                 }
 195                 pthread_mutex_unlock(&sw->lock);
 196                 handle_list(sw, &local_list);
 197                 if (wq->ops.update_acct_fn)
 198                         wq->ops.update_acct_fn(sw);
 199                 pthread_mutex_lock(&sw->lock);
 200         }
 201         pthread_mutex_unlock(&sw->lock);
 202
 203 done:
 204         sk_out_drop();
 205         return NULL;
 206 }
 207
 208 static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 209 {
 210         struct workqueue *wq = sw->wq;
 211
 212         workqueue_exit_worker(sw, sum_cnt);
 213
 214         pthread_cond_destroy(&sw->cond);
 215         pthread_mutex_destroy(&sw->lock);
 216
 217         if (wq->ops.free_worker_fn)
 218                 wq->ops.free_worker_fn(sw);
 219 }
 220
 221 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 222 {
 223         pthread_join(sw->thread, NULL);
 224         free_worker(sw, sum_cnt);
 225 }
 226
 227 void workqueue_exit(struct workqueue *wq)
 228 {
 229         unsigned int shutdown, sum_cnt = 0;
 230         struct submit_worker *sw;
 231         int i;
 232
 233         if (!wq->workers)
 234                 return;
 235
 236         for (i = 0; i < wq->max_workers; i++) {
 237                 sw = &wq->workers[i];
 238
 239                 pthread_mutex_lock(&sw->lock);
 240                 sw->flags |= SW_F_EXIT;
 241                 pthread_cond_signal(&sw->cond);
 242                 pthread_mutex_unlock(&sw->lock);
 243         }
 244
 245         do {
 246                 shutdown = 0;
 247                 for (i = 0; i < wq->max_workers; i++) {
 248                         sw = &wq->workers[i];
 249                         if (sw->flags & SW_F_ACCOUNTED)
 250                                 continue;
 251                         pthread_mutex_lock(&sw->lock);
 252                         sw->flags |= SW_F_ACCOUNTED;
 253                         pthread_mutex_unlock(&sw->lock);
 254                         shutdown_worker(sw, &sum_cnt);
 255                         shutdown++;
 256                 }
 257         } while (shutdown && shutdown != wq->max_workers);
 258
 259         sfree(wq->workers);
 260         wq->workers = NULL;
 261         pthread_mutex_destroy(&wq->flush_lock);
 262         pthread_cond_destroy(&wq->flush_cond);
 263         pthread_mutex_destroy(&wq->stat_lock);
 264 }
 265
 266 static int start_worker(struct workqueue *wq, unsigned int index,
 267                         struct sk_out *sk_out)
 268 {
 269         struct submit_worker *sw = &wq->workers[index];
 270         int ret;
 271
 272         INIT_FLIST_HEAD(&sw->work_list);
 273
 274         ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
 275         if (ret)
 276                 return ret;
 277
 278         sw->wq = wq;
 279         sw->index = index;
 280         sw->sk_out = sk_out;
 281
 282         if (wq->ops.alloc_worker_fn) {
 283                 ret = wq->ops.alloc_worker_fn(sw);
 284                 if (ret)
 285                         return ret;
 286         }
 287
 288         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 289         if (!ret) {
 290                 pthread_mutex_lock(&sw->lock);
 291                 sw->flags = SW_F_IDLE;
 292                 pthread_mutex_unlock(&sw->lock);
 293                 return 0;
 294         }
 295
 296         free_worker(sw, NULL);
 297         return 1;
 298 }
 299
 300 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 301                    struct workqueue_ops *ops, unsigned int max_workers,
 302                    struct sk_out *sk_out)
 303 {
 304         unsigned int running;
 305         int i, error;
 306         int ret;
 307
 308         wq->max_workers = max_workers;
 309         wq->td = td;
 310         wq->ops = *ops;
 311         wq->work_seq = 0;
 312         wq->next_free_worker = 0;
 313
 314         ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
 315         if (ret)
 316                 goto err;
 317         ret = mutex_init_pshared(&wq->stat_lock);
 318         if (ret)
 319                 goto err;
 320
 321         wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
 322         if (!wq->workers)
 323                 goto err;
 324
 325         for (i = 0; i < wq->max_workers; i++)
 326                 if (start_worker(wq, i, sk_out))
 327                         break;
 328
 329         wq->max_workers = i;
 330         if (!wq->max_workers)
 331                 goto err;
 332
 333         /*
 334          * Wait for them all to be started and initialized
 335          */
 336         error = 0;
 337         pthread_mutex_lock(&wq->flush_lock);
 338         do {
 339                 struct submit_worker *sw;
 340
 341                 running = 0;
 342                 for (i = 0; i < wq->max_workers; i++) {
 343                         sw = &wq->workers[i];
 344                         pthread_mutex_lock(&sw->lock);
 345                         if (sw->flags & SW_F_RUNNING)
 346                                 running++;
 347                         if (sw->flags & SW_F_ERROR)
 348                                 error++;
 349                         pthread_mutex_unlock(&sw->lock);
 350                 }
 351
 352                 if (error || running == wq->max_workers)
 353                         break;
 354
 355                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 356         } while (1);
 357         pthread_mutex_unlock(&wq->flush_lock);
 358
 359         if (!error)
 360                 return 0;
 361
 362 err:
 363         log_err("Can't create rate workqueue\n");
 364         td_verror(td, ESRCH, "workqueue_init");
 365         workqueue_exit(wq);
 366         return 1;
 367 }