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         int ret;
 133
 134         /* Read condition before read futex */
 135         cmm_smp_mb();
 136         while (uatomic_read(futex) == -1) {
 137                 if (!futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 138                         /*
 139                          * Prior queued wakeups queued by unrelated code
 140                          * using the same address can cause futex wait to
 141                          * return 0 even through the futex value is still
 142                          * -1 (spurious wakeups). Check the value again
 143                          * in user-space to validate whether it really
 144                          * differs from -1.
 145                          */
 146                         continue;
 147                 }
 148                 switch (errno) {
 149                 case EAGAIN:
 150                         /* Value already changed. */
 151                         return;
 152                 case EINTR:
 153                         /* Retry if interrupted by signal. */
 154                         break;  /* Get out of switch. Check again. */
 155                 default:
 156                         /* Unexpected error. */
 157                         urcu_die(errno);
 158                 }
 159         }
 160 }
 161
 162 static void futex_wake_up(int32_t *futex)
 163 {
 164         /* Write to condition before reading/writing futex */
 165         cmm_smp_mb();
 166         if (caa_unlikely(uatomic_read(futex) == -1)) {
 167                 uatomic_set(futex, 0);
 168                 if (futex_async(futex, FUTEX_WAKE, 1,
 169                                 NULL, NULL, 0) < 0)
 170                         urcu_die(errno);
 171         }
 172 }
 173
 174 /* This is the code run by each worker thread. */
 175
 176 static void *workqueue_thread(void *arg)
 177 {
 178         unsigned long cbcount;
 179         struct urcu_workqueue *workqueue = (struct urcu_workqueue *) arg;
 180         int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);
 181
 182         if (set_thread_cpu_affinity(workqueue))
 183                 urcu_die(errno);
 184
 185         if (workqueue->initialize_worker_fct)
 186                 workqueue->initialize_worker_fct(workqueue, workqueue->priv);
 187
 188         if (!rt) {
 189                 uatomic_dec(&workqueue->futex);
 190                 /* Decrement futex before reading workqueue */
 191                 cmm_smp_mb();
 192         }
 193         for (;;) {
 194                 struct cds_wfcq_head cbs_tmp_head;
 195                 struct cds_wfcq_tail cbs_tmp_tail;
 196                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 197                 enum cds_wfcq_ret splice_ret;
 198
 199                 if (set_thread_cpu_affinity(workqueue))
 200                         urcu_die(errno);
 201
 202                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
 203                         /*
 204                          * Pause requested. Become quiescent: remove
 205                          * ourself from all global lists, and don't
 206                          * process any callback. The callback lists may
 207                          * still be non-empty though.
 208                          */
 209                         if (workqueue->worker_before_pause_fct)
 210                                 workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
 211                         cmm_smp_mb__before_uatomic_or();
 212                         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
 213                         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
 214                                 (void) poll(NULL, 0, 1);
 215                         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
 216                         cmm_smp_mb__after_uatomic_and();
 217                         if (workqueue->worker_after_resume_fct)
 218                                 workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
 219                 }
 220
 221                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 222                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 223                         &cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
 224                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 225                 urcu_posix_assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 226                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 227                         if (workqueue->grace_period_fct)
 228                                 workqueue->grace_period_fct(workqueue, workqueue->priv);
 229                         cbcount = 0;
 230                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 231                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 232                                 struct urcu_work *uwp;
 233
 234                                 uwp = caa_container_of(cbs,
 235                                         struct urcu_work, next);
 236                                 uwp->func(uwp);
 237                                 cbcount++;
 238                         }
 239                         uatomic_sub(&workqueue->qlen, cbcount);
 240                 }
 241                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
 242                         break;
 243                 if (workqueue->worker_before_wait_fct)
 244                         workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
 245                 if (!rt) {
 246                         if (cds_wfcq_empty(&workqueue->cbs_head,
 247                                         &workqueue->cbs_tail)) {
 248                                 futex_wait(&workqueue->futex);
 249                                 uatomic_dec(&workqueue->futex);
 250                                 /*
 251                                  * Decrement futex before reading
 252                                  * urcu_work list.
 253                                  */
 254                                 cmm_smp_mb();
 255                         }
 256                 } else {
 257                         if (cds_wfcq_empty(&workqueue->cbs_head,
 258                                         &workqueue->cbs_tail)) {
 259                                 (void) poll(NULL, 0, 10);
 260                         }
 261                 }
 262                 if (workqueue->worker_after_wake_up_fct)
 263                         workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
 264         }
 265         if (!rt) {
 266                 /*
 267                  * Read urcu_work list before write futex.
 268                  */
 269                 cmm_smp_mb();
 270                 uatomic_set(&workqueue->futex, 0);
 271         }
 272         if (workqueue->finalize_worker_fct)
 273                 workqueue->finalize_worker_fct(workqueue, workqueue->priv);
 274         return NULL;
 275 }
 276
 277 struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
 278                 int cpu_affinity, void *priv,
 279                 void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv),
 280                 void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 281                 void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 282                 void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv),
 283                 void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv),
 284                 void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv),
 285                 void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv))
 286 {
 287         struct urcu_workqueue *workqueue;
 288         int ret;
 289
 290         workqueue = malloc(sizeof(*workqueue));
 291         if (workqueue == NULL)
 292                 urcu_die(errno);
 293         memset(workqueue, '\0', sizeof(*workqueue));
 294         cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
 295         workqueue->qlen = 0;
 296         workqueue->futex = 0;
 297         workqueue->flags = flags;
 298         workqueue->priv = priv;
 299         workqueue->grace_period_fct = grace_period_fct;
 300         workqueue->initialize_worker_fct = initialize_worker_fct;
 301         workqueue->finalize_worker_fct = finalize_worker_fct;
 302         workqueue->worker_before_wait_fct = worker_before_wait_fct;
 303         workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
 304         workqueue->worker_before_pause_fct = worker_before_pause_fct;
 305         workqueue->worker_after_resume_fct = worker_after_resume_fct;
 306         workqueue->cpu_affinity = cpu_affinity;
 307         workqueue->loop_count = 0;
 308         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 309         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 310         if (ret) {
 311                 urcu_die(ret);
 312         }
 313         return workqueue;
 314 }
 315
 316 static void wake_worker_thread(struct urcu_workqueue *workqueue)
 317 {
 318         if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_WORKQUEUE_RT))
 319                 futex_wake_up(&workqueue->futex);
 320 }
 321
 322 static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
 323 {
 324         int ret;
 325         void *retval;
 326
 327         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
 328         wake_worker_thread(workqueue);
 329
 330         ret = pthread_join(workqueue->tid, &retval);
 331         if (ret) {
 332                 urcu_die(ret);
 333         }
 334         if (retval != NULL) {
 335                 urcu_die(EINVAL);
 336         }
 337         workqueue->flags &= ~URCU_WORKQUEUE_STOP;
 338         workqueue->tid = 0;
 339         return 0;
 340 }
 341
 342 void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
 343 {
 344         if (workqueue == NULL) {
 345                 return;
 346         }
 347         if (urcu_workqueue_destroy_worker(workqueue)) {
 348                 urcu_die(errno);
 349         }
 350         urcu_posix_assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
 351         free(workqueue);
 352 }
 353
 354 void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
 355                       struct urcu_work *work,
 356                       void (*func)(struct urcu_work *work))
 357 {
 358         cds_wfcq_node_init(&work->next);
 359         work->func = func;
 360         cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
 361         uatomic_inc(&workqueue->qlen);
 362         wake_worker_thread(workqueue);
 363 }
 364
 365 static
 366 void free_completion(struct urcu_ref *ref)
 367 {
 368         struct urcu_workqueue_completion *completion;
 369
 370         completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
 371         free(completion);
 372 }
 373
 374 static
 375 void _urcu_workqueue_wait_complete(struct urcu_work *work)
 376 {
 377         struct urcu_workqueue_completion_work *completion_work;
 378         struct urcu_workqueue_completion *completion;
 379
 380         completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
 381         completion = completion_work->completion;
 382         if (!uatomic_sub_return(&completion->barrier_count, 1))
 383                 futex_wake_up(&completion->futex);
 384         urcu_ref_put(&completion->ref, free_completion);
 385         free(completion_work);
 386 }
 387
 388 struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
 389 {
 390         struct urcu_workqueue_completion *completion;
 391
 392         completion = calloc(sizeof(*completion), 1);
 393         if (!completion)
 394                 urcu_die(errno);
 395         urcu_ref_set(&completion->ref, 1);
 396         completion->barrier_count = 0;
 397         return completion;
 398 }
 399
 400 void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
 401 {
 402         urcu_ref_put(&completion->ref, free_completion);
 403 }
 404
 405 void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
 406 {
 407         /* Wait for them */
 408         for (;;) {
 409                 uatomic_dec(&completion->futex);
 410                 /* Decrement futex before reading barrier_count */
 411                 cmm_smp_mb();
 412                 if (!uatomic_read(&completion->barrier_count))
 413                         break;
 414                 futex_wait(&completion->futex);
 415         }
 416 }
 417
 418 void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
 419                 struct urcu_workqueue_completion *completion)
 420 {
 421         struct urcu_workqueue_completion_work *work;
 422
 423         work = calloc(sizeof(*work), 1);
 424         if (!work)
 425                 urcu_die(errno);
 426         work->completion = completion;
 427         urcu_ref_get(&completion->ref);
 428         uatomic_inc(&completion->barrier_count);
 429         urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
 430 }
 431
 432 /*
 433  * Wait for all in-flight work to complete execution.
 434  */
 435 void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
 436 {
 437         struct urcu_workqueue_completion *completion;
 438
 439         completion = urcu_workqueue_create_completion();
 440         if (!completion)
 441                 urcu_die(ENOMEM);
 442         urcu_workqueue_queue_completion(workqueue, completion);
 443         urcu_workqueue_wait_completion(completion);
 444         urcu_workqueue_destroy_completion(completion);
 445 }
 446
 447 /* To be used in before fork handler. */
 448 void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
 449 {
 450         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
 451         cmm_smp_mb__after_uatomic_or();
 452         wake_worker_thread(workqueue);
 453
 454         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
 455                 (void) poll(NULL, 0, 1);
 456 }
 457
 458 /* To be used in after fork parent handler. */
 459 void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
 460 {
 461         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
 462         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
 463                 (void) poll(NULL, 0, 1);
 464 }
 465
 466 void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
 467 {
 468         int ret;
 469
 470         /* Clear workqueue state from parent. */
 471         workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
 472         workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
 473         workqueue->tid = 0;
 474         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 475         if (ret) {
 476                 urcu_die(ret);
 477         }
 478 }