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