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         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.
 101  */
 102 void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
 103 {
 104         struct submit_worker *sw;
 105
 106         sw = get_submit_worker(wq);
 107         assert(sw);
 108
 109         pthread_mutex_lock(&sw->lock);
 110         flist_add_tail(&work->list, &sw->work_list);
 111         sw->seq = ++wq->work_seq;
 112         sw->flags &= ~SW_F_IDLE;
 113
 114         pthread_cond_signal(&sw->cond);
 115         pthread_mutex_unlock(&sw->lock);
 116 }
 117
 118 static void handle_list(struct submit_worker *sw, struct flist_head *list)
 119 {
 120         struct workqueue *wq = sw->wq;
 121         struct workqueue_work *work;
 122
 123         while (!flist_empty(list)) {
 124                 work = flist_first_entry(list, struct workqueue_work, list);
 125                 flist_del_init(&work->list);
 126                 wq->ops.fn(sw, work);
 127         }
 128 }
 129
 130 static void *worker_thread(void *data)
 131 {
 132         struct submit_worker *sw = data;
 133         struct workqueue *wq = sw->wq;
 134         unsigned int ret = 0;
 135         FLIST_HEAD(local_list);
 136
 137         sk_out_assign(sw->sk_out);
 138
 139         if (wq->ops.nice) {
 140                 if (nice(wq->ops.nice) < 0) {
 141                         log_err("workqueue: nice %s\n", strerror(errno));
 142                         ret = 1;
 143                 }
 144         }
 145
 146         if (!ret)
 147                 ret = workqueue_init_worker(sw);
 148
 149         pthread_mutex_lock(&sw->lock);
 150         sw->flags |= SW_F_RUNNING;
 151         if (ret)
 152                 sw->flags |= SW_F_ERROR;
 153         pthread_mutex_unlock(&sw->lock);
 154
 155         pthread_mutex_lock(&wq->flush_lock);
 156         pthread_cond_signal(&wq->flush_cond);
 157         pthread_mutex_unlock(&wq->flush_lock);
 158
 159         if (sw->flags & SW_F_ERROR)
 160                 goto done;
 161
 162         while (1) {
 163                 pthread_mutex_lock(&sw->lock);
 164
 165                 if (flist_empty(&sw->work_list)) {
 166                         if (sw->flags & SW_F_EXIT) {
 167                                 pthread_mutex_unlock(&sw->lock);
 168                                 break;
 169                         }
 170
 171                         if (workqueue_pre_sleep_check(sw)) {
 172                                 pthread_mutex_unlock(&sw->lock);
 173                                 workqueue_pre_sleep(sw);
 174                                 pthread_mutex_lock(&sw->lock);
 175                         }
 176
 177                         /*
 178                          * We dropped and reaquired the lock, check
 179                          * state again.
 180                          */
 181                         if (!flist_empty(&sw->work_list))
 182                                 goto handle_work;
 183
 184                         if (sw->flags & SW_F_EXIT) {
 185                                 pthread_mutex_unlock(&sw->lock);
 186                                 break;
 187                         } else if (!(sw->flags & SW_F_IDLE)) {
 188                                 sw->flags |= SW_F_IDLE;
 189                                 wq->next_free_worker = sw->index;
 190                                 if (wq->wake_idle)
 191                                         pthread_cond_signal(&wq->flush_cond);
 192                         }
 193
 194                         pthread_cond_wait(&sw->cond, &sw->lock);
 195                 } else {
 196 handle_work:
 197                         flist_splice_init(&sw->work_list, &local_list);
 198                 }
 199                 pthread_mutex_unlock(&sw->lock);
 200                 handle_list(sw, &local_list);
 201                 if (wq->ops.update_acct_fn)
 202                         wq->ops.update_acct_fn(sw);
 203         }
 204
 205 done:
 206         sk_out_drop();
 207         return NULL;
 208 }
 209
 210 static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 211 {
 212         struct workqueue *wq = sw->wq;
 213
 214         workqueue_exit_worker(sw, sum_cnt);
 215
 216         pthread_cond_destroy(&sw->cond);
 217         pthread_mutex_destroy(&sw->lock);
 218
 219         if (wq->ops.free_worker_fn)
 220                 wq->ops.free_worker_fn(sw);
 221 }
 222
 223 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
 224 {
 225         pthread_join(sw->thread, NULL);
 226         free_worker(sw, sum_cnt);
 227 }
 228
 229 void workqueue_exit(struct workqueue *wq)
 230 {
 231         unsigned int shutdown, sum_cnt = 0;
 232         struct submit_worker *sw;
 233         int i;
 234
 235         if (!wq->workers)
 236                 return;
 237
 238         for (i = 0; i < wq->max_workers; i++) {
 239                 sw = &wq->workers[i];
 240
 241                 pthread_mutex_lock(&sw->lock);
 242                 sw->flags |= SW_F_EXIT;
 243                 pthread_cond_signal(&sw->cond);
 244                 pthread_mutex_unlock(&sw->lock);
 245         }
 246
 247         do {
 248                 shutdown = 0;
 249                 for (i = 0; i < wq->max_workers; i++) {
 250                         sw = &wq->workers[i];
 251                         if (sw->flags & SW_F_ACCOUNTED)
 252                                 continue;
 253                         pthread_mutex_lock(&sw->lock);
 254                         sw->flags |= SW_F_ACCOUNTED;
 255                         pthread_mutex_unlock(&sw->lock);
 256                         shutdown_worker(sw, &sum_cnt);
 257                         shutdown++;
 258                 }
 259         } while (shutdown && shutdown != wq->max_workers);
 260
 261         sfree(wq->workers);
 262         wq->workers = NULL;
 263         pthread_mutex_destroy(&wq->flush_lock);
 264         pthread_cond_destroy(&wq->flush_cond);
 265         pthread_mutex_destroy(&wq->stat_lock);
 266 }
 267
 268 static int start_worker(struct workqueue *wq, unsigned int index,
 269                         struct sk_out *sk_out)
 270 {
 271         struct submit_worker *sw = &wq->workers[index];
 272         int ret;
 273
 274         INIT_FLIST_HEAD(&sw->work_list);
 275
 276         ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
 277         if (ret)
 278                 return ret;
 279
 280         sw->wq = wq;
 281         sw->index = index;
 282         sw->sk_out = sk_out;
 283
 284         if (wq->ops.alloc_worker_fn) {
 285                 ret = wq->ops.alloc_worker_fn(sw);
 286                 if (ret)
 287                         return ret;
 288         }
 289
 290         ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
 291         if (!ret) {
 292                 pthread_mutex_lock(&sw->lock);
 293                 sw->flags = SW_F_IDLE;
 294                 pthread_mutex_unlock(&sw->lock);
 295                 return 0;
 296         }
 297
 298         free_worker(sw, NULL);
 299         return 1;
 300 }
 301
 302 int workqueue_init(struct thread_data *td, struct workqueue *wq,
 303                    struct workqueue_ops *ops, unsigned int max_workers,
 304                    struct sk_out *sk_out)
 305 {
 306         unsigned int running;
 307         int i, error;
 308         int ret;
 309
 310         wq->max_workers = max_workers;
 311         wq->td = td;
 312         wq->ops = *ops;
 313         wq->work_seq = 0;
 314         wq->next_free_worker = 0;
 315
 316         ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
 317         if (ret)
 318                 goto err;
 319         ret = mutex_init_pshared(&wq->stat_lock);
 320         if (ret)
 321                 goto err;
 322
 323         wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
 324         if (!wq->workers)
 325                 goto err;
 326
 327         for (i = 0; i < wq->max_workers; i++)
 328                 if (start_worker(wq, i, sk_out))
 329                         break;
 330
 331         wq->max_workers = i;
 332         if (!wq->max_workers)
 333                 goto err;
 334
 335         /*
 336          * Wait for them all to be started and initialized
 337          */
 338         error = 0;
 339         do {
 340                 struct submit_worker *sw;
 341
 342                 running = 0;
 343                 pthread_mutex_lock(&wq->flush_lock);
 344                 for (i = 0; i < wq->max_workers; i++) {
 345                         sw = &wq->workers[i];
 346                         pthread_mutex_lock(&sw->lock);
 347                         if (sw->flags & SW_F_RUNNING)
 348                                 running++;
 349                         if (sw->flags & SW_F_ERROR)
 350                                 error++;
 351                         pthread_mutex_unlock(&sw->lock);
 352                 }
 353
 354                 if (error || running == wq->max_workers) {
 355                         pthread_mutex_unlock(&wq->flush_lock);
 356                         break;
 357                 }
 358
 359                 pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
 360                 pthread_mutex_unlock(&wq->flush_lock);
 361         } while (1);
 362
 363         if (!error)
 364                 return 0;
 365
 366 err:
 367         log_err("Can't create rate workqueue\n");
 368         td_verror(td, ESRCH, "workqueue_init");
 369         workqueue_exit(wq);
 370         return 1;
 371 }