src/workqueue.c

   1 /*
   2  * workqueue.c
   3  *
   4  * Userspace RCU library - Userspace workqeues
   5  *
   6  * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
   7  * Copyright (c) 2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   8  *
   9  * This library is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * License as published by the Free Software Foundation; either
  12  * version 2.1 of the License, or (at your option) any later version.
  13  *
  14  * This library is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU Lesser General Public
  20  * License along with this library; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  */
  23
  24 #define _LGPL_SOURCE
  25 #include <stdio.h>
  26 #include <pthread.h>
  27 #include <signal.h>
  28 #include <stdlib.h>
  29 #include <stdint.h>
  30 #include <string.h>
  31 #include <errno.h>
  32 #include <poll.h>
  33 #include <sys/time.h>
  34 #include <unistd.h>
  35 #include <sched.h>
  36
  37 #include "compat-getcpu.h"
  38 #include <urcu/assert.h>
  39 #include <urcu/wfcqueue.h>
  40 #include <urcu/pointer.h>
  41 #include <urcu/list.h>
  42 #include <urcu/futex.h>
  43 #include <urcu/tls-compat.h>
  44 #include <urcu/ref.h>
  45 #include "urcu-die.h"
  46
  47 #include "workqueue.h"
  48
  49 #define SET_AFFINITY_CHECK_PERIOD               (1U << 8)       /* 256 */
  50 #define SET_AFFINITY_CHECK_PERIOD_MASK          (SET_AFFINITY_CHECK_PERIOD - 1)
  51
  52 /* Data structure that identifies a workqueue. */
  53
  54 struct urcu_workqueue {
  55         /*
  56          * We do not align head on a different cache-line than tail
  57          * mainly because workqueue threads use batching ("splice") to
  58          * get an entire list of callbacks, which effectively empties
  59          * the queue, and requires to touch the tail anyway.
  60          */
  61         struct cds_wfcq_tail cbs_tail;
  62         struct cds_wfcq_head cbs_head;
  63         unsigned long flags;
  64         int32_t futex;
  65         unsigned long qlen; /* maintained for debugging. */
  66         pthread_t tid;
  67         int cpu_affinity;
  68         unsigned long loop_count;
  69         void *priv;
  70         void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv);
  71         void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv);
  72         void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv);
  73         void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv);
  74         void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv);
  75         void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv);
  76         void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv);
  77 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
  78
  79 struct urcu_workqueue_completion {
  80         int barrier_count;
  81         int32_t futex;
  82         struct urcu_ref ref;
  83 };
  84
  85 struct urcu_workqueue_completion_work {
  86         struct urcu_work work;
  87         struct urcu_workqueue_completion *completion;
  88 };
  89
  90 /*
  91  * Periodically retry setting CPU affinity if we migrate.
  92  * Losing affinity can be caused by CPU hotunplug/hotplug, or by
  93  * cpuset(7).
  94  */
  95 #ifdef HAVE_SCHED_SETAFFINITY
  96 static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue)
  97 {
  98         cpu_set_t mask;
  99         int ret;
 100
 101         if (workqueue->cpu_affinity < 0)
 102                 return 0;
 103         if (++workqueue->loop_count & SET_AFFINITY_CHECK_PERIOD_MASK)
 104                 return 0;
 105         if (urcu_sched_getcpu() == workqueue->cpu_affinity)
 106                 return 0;
 107
 108         CPU_ZERO(&mask);
 109         CPU_SET(workqueue->cpu_affinity, &mask);
 110         ret = sched_setaffinity(0, sizeof(mask), &mask);
 111
 112         /*
 113          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 114          * cpuset(7). This is why we should always retry if we detect
 115          * migration.
 116          */
 117         if (ret && errno == EINVAL) {
 118                 ret = 0;
 119                 errno = 0;
 120         }
 121         return ret;
 122 }
 123 #else
 124 static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue __attribute__((unused)))
 125 {
 126         return 0;
 127 }
 128 #endif
 129
 130 static void futex_wait(int32_t *futex)
 131 {
 132         /* Read condition before read futex */
 133         cmm_smp_mb();
 134         while (uatomic_read(futex) == -1) {
 135                 if (!futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 136                         /*
 137                          * Prior queued wakeups queued by unrelated code
 138                          * using the same address can cause futex wait to
 139                          * return 0 even through the futex value is still
 140                          * -1 (spurious wakeups). Check the value again
 141                          * in user-space to validate whether it really
 142                          * differs from -1.
 143                          */
 144                         continue;
 145                 }
 146                 switch (errno) {
 147                 case EAGAIN:
 148                         /* Value already changed. */
 149                         return;
 150                 case EINTR:
 151                         /* Retry if interrupted by signal. */
 152                         break;  /* Get out of switch. Check again. */
 153                 default:
 154                         /* Unexpected error. */
 155                         urcu_die(errno);
 156                 }
 157         }
 158 }
 159
 160 static void futex_wake_up(int32_t *futex)
 161 {
 162         /* Write to condition before reading/writing futex */
 163         cmm_smp_mb();
 164         if (caa_unlikely(uatomic_read(futex) == -1)) {
 165                 uatomic_set(futex, 0);
 166                 if (futex_async(futex, FUTEX_WAKE, 1,
 167                                 NULL, NULL, 0) < 0)
 168                         urcu_die(errno);
 169         }
 170 }
 171
 172 /* This is the code run by each worker thread. */
 173
 174 static void *workqueue_thread(void *arg)
 175 {
 176         unsigned long cbcount;
 177         struct urcu_workqueue *workqueue = (struct urcu_workqueue *) arg;
 178         int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);
 179
 180         if (set_thread_cpu_affinity(workqueue))
 181                 urcu_die(errno);
 182
 183         if (workqueue->initialize_worker_fct)
 184                 workqueue->initialize_worker_fct(workqueue, workqueue->priv);
 185
 186         if (!rt) {
 187                 uatomic_dec(&workqueue->futex);
 188                 /* Decrement futex before reading workqueue */
 189                 cmm_smp_mb();
 190         }
 191         for (;;) {
 192                 struct cds_wfcq_head cbs_tmp_head;
 193                 struct cds_wfcq_tail cbs_tmp_tail;
 194                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 195                 enum cds_wfcq_ret splice_ret;
 196
 197                 if (set_thread_cpu_affinity(workqueue))
 198                         urcu_die(errno);
 199
 200                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
 201                         /*
 202                          * Pause requested. Become quiescent: remove
 203                          * ourself from all global lists, and don't
 204                          * process any callback. The callback lists may
 205                          * still be non-empty though.
 206                          */
 207                         if (workqueue->worker_before_pause_fct)
 208                                 workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
 209                         cmm_smp_mb__before_uatomic_or();
 210                         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
 211                         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
 212                                 (void) poll(NULL, 0, 1);
 213                         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
 214                         cmm_smp_mb__after_uatomic_and();
 215                         if (workqueue->worker_after_resume_fct)
 216                                 workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
 217                 }
 218
 219                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 220                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 221                         &cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
 222                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 223                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 224                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 225                         if (workqueue->grace_period_fct)
 226                                 workqueue->grace_period_fct(workqueue, workqueue->priv);
 227                         cbcount = 0;
 228                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 229                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 230                                 struct urcu_work *uwp;
 231
 232                                 uwp = caa_container_of(cbs,
 233                                         struct urcu_work, next);
 234                                 uwp->func(uwp);
 235                                 cbcount++;
 236                         }
 237                         uatomic_sub(&workqueue->qlen, cbcount);
 238                 }
 239                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
 240                         break;
 241                 if (workqueue->worker_before_wait_fct)
 242                         workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
 243                 if (!rt) {
 244                         if (cds_wfcq_empty(&workqueue->cbs_head,
 245                                         &workqueue->cbs_tail)) {
 246                                 futex_wait(&workqueue->futex);
 247                                 uatomic_dec(&workqueue->futex);
 248                                 /*
 249                                  * Decrement futex before reading
 250                                  * urcu_work list.
 251                                  */
 252                                 cmm_smp_mb();
 253                         }
 254                 } else {
 255                         if (cds_wfcq_empty(&workqueue->cbs_head,
 256                                         &workqueue->cbs_tail)) {
 257                                 (void) poll(NULL, 0, 10);
 258                         }
 259                 }
 260                 if (workqueue->worker_after_wake_up_fct)
 261                         workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
 262         }
 263         if (!rt) {
 264                 /*
 265                  * Read urcu_work list before write futex.
 266                  */
 267                 cmm_smp_mb();
 268                 uatomic_set(&workqueue->futex, 0);
 269         }
 270         if (workqueue->finalize_worker_fct)
 271                 workqueue->finalize_worker_fct(workqueue, workqueue->priv);
 272         return NULL;
 273 }
 274
 275 struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
 276                 int cpu_affinity, void *priv,
 277                 void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv),
 278                 void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 279                 void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 280                 void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv),
 281                 void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv),
 282                 void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv),
 283                 void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv))
 284 {
 285         struct urcu_workqueue *workqueue;
 286         int ret;
 287         sigset_t newmask, oldmask;
 288
 289         workqueue = malloc(sizeof(*workqueue));
 290         if (workqueue == NULL)
 291                 urcu_die(errno);
 292         memset(workqueue, '\0', sizeof(*workqueue));
 293         cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
 294         workqueue->qlen = 0;
 295         workqueue->futex = 0;
 296         workqueue->flags = flags;
 297         workqueue->priv = priv;
 298         workqueue->grace_period_fct = grace_period_fct;
 299         workqueue->initialize_worker_fct = initialize_worker_fct;
 300         workqueue->finalize_worker_fct = finalize_worker_fct;
 301         workqueue->worker_before_wait_fct = worker_before_wait_fct;
 302         workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
 303         workqueue->worker_before_pause_fct = worker_before_pause_fct;
 304         workqueue->worker_after_resume_fct = worker_after_resume_fct;
 305         workqueue->cpu_affinity = cpu_affinity;
 306         workqueue->loop_count = 0;
 307         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 308
 309         ret = sigfillset(&newmask);
 310         urcu_posix_assert(!ret);
 311         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 312         urcu_posix_assert(!ret);
 313
 314         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 315         if (ret) {
 316                 urcu_die(ret);
 317         }
 318
 319         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 320         urcu_posix_assert(!ret);
 321
 322         return workqueue;
 323 }
 324
 325 static void wake_worker_thread(struct urcu_workqueue *workqueue)
 326 {
 327         if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_WORKQUEUE_RT))
 328                 futex_wake_up(&workqueue->futex);
 329 }
 330
 331 static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
 332 {
 333         int ret;
 334         void *retval;
 335
 336         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
 337         wake_worker_thread(workqueue);
 338
 339         ret = pthread_join(workqueue->tid, &retval);
 340         if (ret) {
 341                 urcu_die(ret);
 342         }
 343         if (retval != NULL) {
 344                 urcu_die(EINVAL);
 345         }
 346         workqueue->flags &= ~URCU_WORKQUEUE_STOP;
 347         workqueue->tid = 0;
 348         return 0;
 349 }
 350
 351 void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
 352 {
 353         if (workqueue == NULL) {
 354                 return;
 355         }
 356         if (urcu_workqueue_destroy_worker(workqueue)) {
 357                 urcu_die(errno);
 358         }
 359         urcu_posix_assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
 360         free(workqueue);
 361 }
 362
 363 void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
 364                       struct urcu_work *work,
 365                       void (*func)(struct urcu_work *work))
 366 {
 367         cds_wfcq_node_init(&work->next);
 368         work->func = func;
 369         cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
 370         uatomic_inc(&workqueue->qlen);
 371         wake_worker_thread(workqueue);
 372 }
 373
 374 static
 375 void free_completion(struct urcu_ref *ref)
 376 {
 377         struct urcu_workqueue_completion *completion;
 378
 379         completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
 380         free(completion);
 381 }
 382
 383 static
 384 void _urcu_workqueue_wait_complete(struct urcu_work *work)
 385 {
 386         struct urcu_workqueue_completion_work *completion_work;
 387         struct urcu_workqueue_completion *completion;
 388
 389         completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
 390         completion = completion_work->completion;
 391         if (!uatomic_sub_return(&completion->barrier_count, 1))
 392                 futex_wake_up(&completion->futex);
 393         urcu_ref_put(&completion->ref, free_completion);
 394         free(completion_work);
 395 }
 396
 397 struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
 398 {
 399         struct urcu_workqueue_completion *completion;
 400
 401         completion = calloc(sizeof(*completion), 1);
 402         if (!completion)
 403                 urcu_die(errno);
 404         urcu_ref_set(&completion->ref, 1);
 405         completion->barrier_count = 0;
 406         return completion;
 407 }
 408
 409 void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
 410 {
 411         urcu_ref_put(&completion->ref, free_completion);
 412 }
 413
 414 void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
 415 {
 416         /* Wait for them */
 417         for (;;) {
 418                 uatomic_dec(&completion->futex);
 419                 /* Decrement futex before reading barrier_count */
 420                 cmm_smp_mb();
 421                 if (!uatomic_read(&completion->barrier_count))
 422                         break;
 423                 futex_wait(&completion->futex);
 424         }
 425 }
 426
 427 void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
 428                 struct urcu_workqueue_completion *completion)
 429 {
 430         struct urcu_workqueue_completion_work *work;
 431
 432         work = calloc(sizeof(*work), 1);
 433         if (!work)
 434                 urcu_die(errno);
 435         work->completion = completion;
 436         urcu_ref_get(&completion->ref);
 437         uatomic_inc(&completion->barrier_count);
 438         urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
 439 }
 440
 441 /*
 442  * Wait for all in-flight work to complete execution.
 443  */
 444 void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
 445 {
 446         struct urcu_workqueue_completion *completion;
 447
 448         completion = urcu_workqueue_create_completion();
 449         if (!completion)
 450                 urcu_die(ENOMEM);
 451         urcu_workqueue_queue_completion(workqueue, completion);
 452         urcu_workqueue_wait_completion(completion);
 453         urcu_workqueue_destroy_completion(completion);
 454 }
 455
 456 /* To be used in before fork handler. */
 457 void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
 458 {
 459         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
 460         cmm_smp_mb__after_uatomic_or();
 461         wake_worker_thread(workqueue);
 462
 463         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
 464                 (void) poll(NULL, 0, 1);
 465 }
 466
 467 /* To be used in after fork parent handler. */
 468 void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
 469 {
 470         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
 471         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
 472                 (void) poll(NULL, 0, 1);
 473 }
 474
 475 void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
 476 {
 477         int ret;
 478         sigset_t newmask, oldmask;
 479
 480         /* Clear workqueue state from parent. */
 481         workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
 482         workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
 483         workqueue->tid = 0;
 484
 485         ret = sigfillset(&newmask);
 486         urcu_posix_assert(!ret);
 487         ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask);
 488         urcu_posix_assert(!ret);
 489
 490         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 491         if (ret) {
 492                 urcu_die(ret);
 493         }
 494
 495         ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL);
 496         urcu_posix_assert(!ret);
 497 }