urcu/workqueue-fifo.h

   1 #ifndef _URCU_WORKQUEUE_FIFO_H
   2 #define _URCU_WORKQUEUE_FIFO_H
   3
   4 /*
   5  * urcu/workqueue-fifo.h
   6  *
   7  * Userspace RCU library - work queue scheme with FIFO semantic
   8  *
   9  * Copyright (c) 2014 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  10  *
  11  * This library is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License as published by the Free Software Foundation; either
  14  * version 2.1 of the License, or (at your option) any later version.
  15  *
  16  * This library is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with this library; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  24  */
  25
  26 #include <urcu/uatomic.h>
  27 #include <urcu/lfstack.h>
  28 #include <urcu/waitqueue-lifo.h>
  29 #include <urcu/wfcqueue.h>
  30 #include <urcu/rculist.h>
  31 #include <pthread.h>
  32 #include <assert.h>
  33
  34 enum urcu_accept_ret {
  35         URCU_ACCEPT_WORK        = 0,
  36         URCU_ACCEPT_SHUTDOWN    = 1,
  37 };
  38
  39 /*
  40  * We use RCU to steal work from siblings. Therefore, one of RCU flavors
  41  * need to be included before this header. All worker that participate
  42  * in stealing (initialized with the URCU_WORKER_STEAL flag) need to be
  43  * registered RCU readers threads.
  44  */
  45
  46 struct urcu_work {
  47         struct cds_wfcq_node node;
  48 };
  49
  50 struct urcu_workqueue {
  51         /* FIFO work queue */
  52         struct __cds_wfcq_head head;
  53         struct cds_wfcq_tail tail;
  54
  55         /* Associated wait queue for LIFO wait/wakeup */
  56         struct urcu_wait_queue waitqueue;
  57
  58         /* RCU linked list head of siblings for work stealing. */
  59         struct cds_list_head sibling_head;
  60         pthread_mutex_t sibling_lock;   /* Protect sibling list updates */
  61
  62         bool shutdown;                  /* Shutdown performed */
  63 };
  64
  65 struct urcu_worker {
  66         /* Workqueue which can be either used by worker, or stolen. */
  67         struct cds_wfcq_head head;
  68         struct cds_wfcq_tail tail;
  69
  70         /* Work belonging to worker. Cannot be stolen. */
  71         struct urcu_work *own;
  72
  73         struct urcu_wait_node wait_node;
  74         /* RCU linked list node of siblings for work stealing. */
  75         struct cds_list_head sibling_node;
  76         int flags;      /* enum urcu_worker_flags */
  77 };
  78
  79 enum urcu_worker_flags {
  80         URCU_WORKER_STEAL       = (1 << 0),
  81 };
  82
  83 static inline
  84 void urcu_workqueue_init(struct urcu_workqueue *queue)
  85 {
  86         __cds_wfcq_init(&queue->head, &queue->tail);
  87         urcu_wait_queue_init(&queue->waitqueue);
  88         CDS_INIT_LIST_HEAD(&queue->sibling_head);
  89         pthread_mutex_init(&queue->sibling_lock, NULL);
  90         queue->shutdown = false;
  91 }
  92
  93 static inline
  94 void urcu_queue_work(struct urcu_workqueue *queue, struct urcu_work *work)
  95 {
  96         bool was_empty;
  97
  98         cds_wfcq_node_init(&work->node);
  99
 100         /* Enqueue work. */
 101         was_empty = !cds_wfcq_enqueue(&queue->head, &queue->tail,
 102                         &work->node);
 103         /*
 104          * If workqueue was previously empty, wakeup one worker thread.
 105          * It will eventually grab the entire content of the work-queue
 106          * (therefore grabbing a "work batch"). After having grabbed the
 107          * work batch, while that thread is running and taking care of
 108          * that work batch, when we enqueue more work, we will wake
 109          * another thread (if there is one waiting), which will
 110          * eventually grab the new batch, and so on. This scheme ensures
 111          * that contiguous batch of work are handled by the same thread
 112          * (for locality), and also ensures that we scale work to many
 113          * worker threads when threads are busy enough to still be
 114          * running when work is enqueued.
 115          */
 116         if (was_empty) {
 117                 rcu_read_lock();        /* Protect stack dequeue */
 118                 (void) urcu_dequeue_wake_single(&queue->waitqueue);
 119                 rcu_read_unlock();      /* Protect stack dequeue */
 120         }
 121 }
 122
 123 static inline
 124 void __urcu_workqueue_wakeup_all(struct urcu_workqueue *queue)
 125 {
 126         struct urcu_waiters waiters;
 127
 128         rcu_read_lock();        /* Protect stack dequeue */
 129         urcu_move_waiters(&waiters, &queue->waitqueue);
 130         rcu_read_unlock();      /* Protect stack dequeue */
 131
 132         (void) urcu_wake_all_waiters(&waiters);
 133 }
 134
 135 static inline
 136 void urcu_worker_init(struct urcu_worker *worker, int flags)
 137 {
 138         cds_wfcq_init(&worker->head, &worker->tail);
 139         worker->flags = flags;
 140         urcu_wait_node_init(&worker->wait_node, URCU_WAIT_RUNNING);
 141         worker->own = NULL;
 142 }
 143
 144 static inline
 145 void urcu_worker_register(struct urcu_workqueue *queue,
 146                 struct urcu_worker *worker)
 147 {
 148         if (worker->flags & URCU_WORKER_STEAL) {
 149                 pthread_mutex_lock(&queue->sibling_lock);
 150                 cds_list_add_rcu(&worker->sibling_node, &queue->sibling_head);
 151                 pthread_mutex_unlock(&queue->sibling_lock);
 152         }
 153 }
 154
 155 static inline
 156 void urcu_worker_unregister(struct urcu_workqueue *queue,
 157                 struct urcu_worker *worker)
 158 {
 159         enum cds_wfcq_ret wfcq_ret;
 160
 161         if (worker->flags & URCU_WORKER_STEAL) {
 162                 pthread_mutex_lock(&queue->sibling_lock);
 163                 cds_list_del_rcu(&worker->sibling_node);
 164                 pthread_mutex_unlock(&queue->sibling_lock);
 165         }
 166
 167         /*
 168          * Make sure we are removed from waitqueue.
 169          */
 170         if (CMM_LOAD_SHARED(worker->wait_node.node.next))
 171                 __urcu_workqueue_wakeup_all(queue);
 172
 173         /*
 174          * Put any local work we still have back into the workqueue.
 175          */
 176         wfcq_ret = __cds_wfcq_splice_blocking(&queue->head,
 177                         &queue->tail,
 178                         &worker->head,
 179                         &worker->tail);
 180         if (wfcq_ret != CDS_WFCQ_RET_SRC_EMPTY
 181                         && wfcq_ret == CDS_WFCQ_RET_DEST_EMPTY) {
 182                 /*
 183                  * Wakeup worker thread if we have put work back into
 184                  * workqueue that was previously empty.
 185                  */
 186                 rcu_read_lock();        /* Protect stack dequeue */
 187                 (void) urcu_dequeue_wake_single(&queue->waitqueue);
 188                 rcu_read_unlock();      /* Protect stack dequeue */
 189         }
 190
 191         /*
 192          * Wait for grace period before freeing or reusing
 193          * "worker" because used by RCU linked list.
 194          * Also prevents ABA for waitqueue stack dequeue: matches RCU
 195          * read-side critical sections around dequeue and move all
 196          * operations on waitqueue).
 197          */
 198         synchronize_rcu();
 199 }
 200
 201 static inline
 202 bool ___urcu_grab_work(struct urcu_worker *worker,
 203                 cds_wfcq_head_ptr_t src_head,
 204                 struct cds_wfcq_tail *src_tail,
 205                 bool steal)
 206 {
 207         enum cds_wfcq_ret splice_ret;
 208         struct __cds_wfcq_head tmp_head;
 209         struct cds_wfcq_tail tmp_tail;
 210         struct cds_wfcq_node *node;
 211
 212         /*
 213          * Don't bother grabbing the src queue lock if it is empty.
 214          */
 215         if (cds_wfcq_empty(src_head, src_tail))
 216                 return false;
 217         __cds_wfcq_init(&tmp_head, &tmp_tail);
 218
 219         /* Ensure that we preserve FIFO work order. */
 220         assert(!steal || worker->own == NULL);
 221
 222         /* Splice to temporary queue. */
 223         if (steal)
 224                 cds_wfcq_dequeue_lock(src_head.h, src_tail);
 225         splice_ret = __cds_wfcq_splice_blocking(&tmp_head,
 226                         &tmp_tail,
 227                         src_head,
 228                         src_tail);
 229         if (steal)
 230                 cds_wfcq_dequeue_unlock(src_head.h, src_tail);
 231         if (splice_ret == CDS_WFCQ_RET_SRC_EMPTY)
 232                 return false;
 233
 234         /*
 235          * Keep one work entry for ourself. This ensures forward
 236          * progress amongst stealing co-workers. This also ensures that
 237          * when a worker grab some work from the global workqueue, it
 238          * will have at least one work item to deal with.
 239          */
 240         if (worker->own == NULL) {
 241                 if (!steal) {
 242                         /*
 243                          * Try to grab own work from worker workqueue to
 244                          * preserve FIFO order.
 245                          */
 246                         node = cds_wfcq_dequeue_blocking(&worker->head,
 247                                 &worker->tail);
 248                         if (node)
 249                                 goto got_node;
 250                 }
 251                 node = __cds_wfcq_dequeue_blocking(&tmp_head, &tmp_tail);
 252                 assert(node != NULL);
 253 got_node:
 254                 worker->own = caa_container_of(node, struct urcu_work, node);
 255         }
 256
 257         /* Splice into worker workqueue. */
 258         splice_ret = __cds_wfcq_splice_blocking(&worker->head,
 259                         &worker->tail,
 260                         &tmp_head,
 261                         &tmp_tail);
 262         /* Ensure that we preserve FIFO work order. */
 263         assert(!steal || splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 264         return true;
 265 }
 266
 267 /*
 268  * Try stealing work from siblings when we have nothing to do.
 269  */
 270 static inline
 271 bool ___urcu_steal_work(struct urcu_worker *worker,
 272                 struct urcu_worker *sibling)
 273 {
 274         return ___urcu_grab_work(worker, &sibling->head, &sibling->tail, 1);
 275 }
 276
 277 static inline
 278 bool __urcu_steal_work(struct urcu_workqueue *queue,
 279                 struct urcu_worker *worker)
 280 {
 281         struct urcu_worker *sibling_prev, *sibling_next;
 282         struct cds_list_head *sibling_node;
 283         bool steal_performed = 0;
 284
 285         if (!(worker->flags & URCU_WORKER_STEAL))
 286                 return false;
 287
 288         rcu_read_lock();
 289
 290         sibling_node = rcu_dereference(worker->sibling_node.next);
 291         if (sibling_node == &queue->sibling_head)
 292                 sibling_node = rcu_dereference(sibling_node->next);
 293         sibling_next = caa_container_of(sibling_node, struct urcu_worker,
 294                         sibling_node);
 295         if (sibling_next != worker)
 296                 steal_performed = ___urcu_steal_work(worker, sibling_next);
 297         if (steal_performed)
 298                 goto end;
 299
 300         sibling_node = rcu_dereference(worker->sibling_node.prev);
 301         if (sibling_node == &queue->sibling_head)
 302                 sibling_node = rcu_dereference(sibling_node->prev);
 303         sibling_prev = caa_container_of(sibling_node, struct urcu_worker,
 304                         sibling_node);
 305         if (sibling_prev != worker && sibling_prev != sibling_next)
 306                 steal_performed =  ___urcu_steal_work(worker, sibling_prev);
 307 end:
 308         rcu_read_unlock();
 309
 310         return steal_performed;
 311 }
 312
 313 static inline
 314 bool ___urcu_wakeup_sibling(struct urcu_worker *sibling)
 315 {
 316         return urcu_adaptative_wake_up(&sibling->wait_node);
 317 }
 318
 319 static inline
 320 bool __urcu_wakeup_siblings(struct urcu_workqueue *queue,
 321                 struct urcu_worker *worker)
 322 {
 323         struct urcu_worker *sibling_prev, *sibling_next;
 324         struct cds_list_head *sibling_node;
 325         bool wakeup_performed = 0;
 326
 327         if (!(worker->flags & URCU_WORKER_STEAL))
 328                 return;
 329
 330         /* Only wakeup siblings if we have work in our own queue. */
 331         if (cds_wfcq_empty(&worker->head, &worker->tail))
 332                 return;
 333
 334         rcu_read_lock();
 335
 336         sibling_node = rcu_dereference(worker->sibling_node.next);
 337         if (sibling_node == &queue->sibling_head)
 338                 sibling_node = rcu_dereference(sibling_node->next);
 339         sibling_next = caa_container_of(sibling_node, struct urcu_worker,
 340                         sibling_node);
 341         if (sibling_next != worker)
 342                 wakeup_performed = ___urcu_wakeup_sibling(sibling_next);
 343         if (wakeup_performed)
 344                 goto end;
 345
 346         sibling_node = rcu_dereference(worker->sibling_node.prev);
 347         if (sibling_node == &queue->sibling_head)
 348                 sibling_node = rcu_dereference(sibling_node->prev);
 349         sibling_prev = caa_container_of(sibling_node, struct urcu_worker,
 350                         sibling_node);
 351         if (sibling_prev != worker && sibling_prev != sibling_next)
 352                 wakeup_performed = ___urcu_wakeup_sibling(sibling_prev);
 353 end:
 354         rcu_read_unlock();
 355
 356         return wakeup_performed;
 357 }
 358
 359 static inline
 360 enum urcu_accept_ret urcu_accept_work(struct urcu_workqueue *queue,
 361                 struct urcu_worker *worker)
 362 {
 363         enum cds_wfcq_ret wfcq_ret;
 364         bool has_work;
 365
 366         has_work = ___urcu_grab_work(worker, &queue->head, &queue->tail, 0);
 367         /* Don't wait if we have work to do. */
 368         if (has_work || !cds_wfcq_empty(&worker->head, &worker->tail))
 369                 goto do_work;
 370         /* Try to steal work from sibling instead of blocking */
 371         if (__urcu_steal_work(queue, worker))
 372                 goto do_work;
 373         /* No more work to do, check shutdown state */
 374         if (CMM_LOAD_SHARED(queue->shutdown))
 375                 return URCU_ACCEPT_SHUTDOWN;
 376         urcu_wait_set_state(&worker->wait_node,
 377                         URCU_WAIT_WAITING);
 378         if (!CMM_LOAD_SHARED(worker->wait_node.node.next)) {
 379                 int was_empty;
 380
 381                 /*
 382                  * NULL next pointer. We are therefore not in
 383                  * the queue.
 384                  */
 385                 cds_lfs_node_init(&worker->wait_node.node);
 386                 /* Protect stack dequeue against ABA */
 387                 synchronize_rcu();
 388                 was_empty = !urcu_wait_add(&queue->waitqueue,
 389                                 &worker->wait_node);
 390                 /*
 391                  * If the wait queue was empty, it means we are the
 392                  * first thread to be put back into an otherwise empty
 393                  * wait queue. Re-check if work queue is empty after
 394                  * adding ourself to wait queue, so we can wakeup the
 395                  * top of wait queue since new work have appeared, and
 396                  * work enqueuer may not have seen that it needed to do
 397                  * a wake up.
 398                  */
 399                 if (was_empty && !cds_wfcq_empty(&queue->head,
 400                                                 &queue->tail)) {
 401                         rcu_read_lock();        /* Protect stack dequeue */
 402                         (void) urcu_dequeue_wake_single(&queue->waitqueue);
 403                         rcu_read_unlock();      /* Protect stack dequeue */
 404                 }
 405         } else {
 406                 /*
 407                  * Non-NULL next pointer. We are therefore in
 408                  * the queue, or the dispatcher just removed us
 409                  * from it (after we read the next pointer), and
 410                  * is therefore awakening us. The state will
 411                  * therefore have been changed from WAITING to
 412                  * some other state, which will let the busy
 413                  * wait pass through.
 414                  */
 415         }
 416         urcu_adaptative_busy_wait(&worker->wait_node);
 417         return;
 418
 419 do_work:
 420         /*
 421          * We will be busy handling the work batch, awaken siblings so
 422          * they can steal from us.
 423          */
 424         (void) __urcu_wakeup_siblings(queue, worker);
 425         return URCU_ACCEPT_WORK;
 426 }
 427
 428 static inline
 429 struct urcu_work *urcu_dequeue_work(struct urcu_worker *worker)
 430 {
 431         struct cds_wfcq_node *node;
 432
 433         if (worker->own) {
 434                 struct urcu_work *work;
 435
 436                 /* Process our own work entry. */
 437                 work = worker->own;
 438                 worker->own = NULL;
 439                 return work;
 440         }
 441         /*
 442          * If we are registered for work stealing, we need to dequeue
 443          * safely against siblings.
 444          */
 445         if (worker->flags & URCU_WORKER_STEAL) {
 446                 /*
 447                  * Don't bother grabbing the worker queue lock if it is
 448                  * empty.
 449                  */
 450                 if (cds_wfcq_empty(&worker->head, &worker->tail))
 451                         return NULL;
 452                 node = cds_wfcq_dequeue_blocking(&worker->head,
 453                                 &worker->tail);
 454         } else {
 455                 node = ___cds_wfcq_dequeue_with_state(&worker->head,
 456                                 &worker->tail, NULL, 1, 0);
 457         }
 458         if (!node)
 459                 return NULL;
 460         return caa_container_of(node, struct urcu_work, node);
 461 }
 462
 463 static inline
 464 void urcu_workqueue_shutdown(struct urcu_workqueue *queue)
 465 {
 466         /* Set shutdown */
 467         CMM_STORE_SHARED(queue->shutdown, true);
 468         /* Wakeup all workers */
 469         __urcu_workqueue_wakeup_all(queue);
 470 }
 471
 472 #endif /* _URCU_WORKQUEUE_FIFO_H */