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 <assert.h>
  29 #include <stdlib.h>
  30 #include <stdint.h>
  31 #include <string.h>
  32 #include <errno.h>
  33 #include <poll.h>
  34 #include <sys/time.h>
  35 #include <unistd.h>
  36 #include <sched.h>
  37
  38 #include "compat-getcpu.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 #if 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 #if SCHED_SETAFFINITY_ARGS == 2
 111         ret = sched_setaffinity(0, &mask);
 112 #else
 113         ret = sched_setaffinity(0, sizeof(mask), &mask);
 114 #endif
 115         /*
 116          * EINVAL is fine: can be caused by hotunplugged CPUs, or by
 117          * cpuset(7). This is why we should always retry if we detect
 118          * migration.
 119          */
 120         if (ret && errno == EINVAL) {
 121                 ret = 0;
 122                 errno = 0;
 123         }
 124         return ret;
 125 }
 126 #else
 127 static int set_thread_cpu_affinity(struct urcu_workqueue *workqueue)
 128 {
 129         return 0;
 130 }
 131 #endif
 132
 133 static void futex_wait(int32_t *futex)
 134 {
 135         /* Read condition before read futex */
 136         cmm_smp_mb();
 137         if (uatomic_read(futex) != -1)
 138                 return;
 139         while (futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 140                 switch (errno) {
 141                 case EWOULDBLOCK:
 142                         /* Value already changed. */
 143                         return;
 144                 case EINTR:
 145                         /* Retry if interrupted by signal. */
 146                         break;  /* Get out of switch. */
 147                 default:
 148                         /* Unexpected error. */
 149                         urcu_die(errno);
 150                 }
 151         }
 152 }
 153
 154 static void futex_wake_up(int32_t *futex)
 155 {
 156         /* Write to condition before reading/writing futex */
 157         cmm_smp_mb();
 158         if (caa_unlikely(uatomic_read(futex) == -1)) {
 159                 uatomic_set(futex, 0);
 160                 if (futex_async(futex, FUTEX_WAKE, 1,
 161                                 NULL, NULL, 0) < 0)
 162                         urcu_die(errno);
 163         }
 164 }
 165
 166 /* This is the code run by each worker thread. */
 167
 168 static void *workqueue_thread(void *arg)
 169 {
 170         unsigned long cbcount;
 171         struct urcu_workqueue *workqueue = (struct urcu_workqueue *) arg;
 172         int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);
 173
 174         if (set_thread_cpu_affinity(workqueue))
 175                 urcu_die(errno);
 176
 177         if (workqueue->initialize_worker_fct)
 178                 workqueue->initialize_worker_fct(workqueue, workqueue->priv);
 179
 180         if (!rt) {
 181                 uatomic_dec(&workqueue->futex);
 182                 /* Decrement futex before reading workqueue */
 183                 cmm_smp_mb();
 184         }
 185         for (;;) {
 186                 struct cds_wfcq_head cbs_tmp_head;
 187                 struct cds_wfcq_tail cbs_tmp_tail;
 188                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 189                 enum cds_wfcq_ret splice_ret;
 190
 191                 if (set_thread_cpu_affinity(workqueue))
 192                         urcu_die(errno);
 193
 194                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
 195                         /*
 196                          * Pause requested. Become quiescent: remove
 197                          * ourself from all global lists, and don't
 198                          * process any callback. The callback lists may
 199                          * still be non-empty though.
 200                          */
 201                         if (workqueue->worker_before_pause_fct)
 202                                 workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
 203                         cmm_smp_mb__before_uatomic_or();
 204                         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
 205                         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
 206                                 (void) poll(NULL, 0, 1);
 207                         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
 208                         cmm_smp_mb__after_uatomic_and();
 209                         if (workqueue->worker_after_resume_fct)
 210                                 workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
 211                 }
 212
 213                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 214                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 215                         &cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
 216                 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 217                 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 218                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 219                         if (workqueue->grace_period_fct)
 220                                 workqueue->grace_period_fct(workqueue, workqueue->priv);
 221                         cbcount = 0;
 222                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 223                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 224                                 struct urcu_work *uwp;
 225
 226                                 uwp = caa_container_of(cbs,
 227                                         struct urcu_work, next);
 228                                 uwp->func(uwp);
 229                                 cbcount++;
 230                         }
 231                         uatomic_sub(&workqueue->qlen, cbcount);
 232                 }
 233                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
 234                         break;
 235                 if (workqueue->worker_before_wait_fct)
 236                         workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
 237                 if (!rt) {
 238                         if (cds_wfcq_empty(&workqueue->cbs_head,
 239                                         &workqueue->cbs_tail)) {
 240                                 futex_wait(&workqueue->futex);
 241                                 uatomic_dec(&workqueue->futex);
 242                                 /*
 243                                  * Decrement futex before reading
 244                                  * urcu_work list.
 245                                  */
 246                                 cmm_smp_mb();
 247                         } else {
 248                                 (void) poll(NULL, 0, 10);
 249                         }
 250                 } else {
 251                         (void) poll(NULL, 0, 10);
 252                 }
 253                 if (workqueue->worker_after_wake_up_fct)
 254                         workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
 255         }
 256         if (!rt) {
 257                 /*
 258                  * Read urcu_work list before write futex.
 259                  */
 260                 cmm_smp_mb();
 261                 uatomic_set(&workqueue->futex, 0);
 262         }
 263         if (workqueue->finalize_worker_fct)
 264                 workqueue->finalize_worker_fct(workqueue, workqueue->priv);
 265         return NULL;
 266 }
 267
 268 struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
 269                 int cpu_affinity, void *priv,
 270                 void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv),
 271                 void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 272                 void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 273                 void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv),
 274                 void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv),
 275                 void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv),
 276                 void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv))
 277 {
 278         struct urcu_workqueue *workqueue;
 279         int ret;
 280
 281         workqueue = malloc(sizeof(*workqueue));
 282         if (workqueue == NULL)
 283                 urcu_die(errno);
 284         memset(workqueue, '\0', sizeof(*workqueue));
 285         cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
 286         workqueue->qlen = 0;
 287         workqueue->futex = 0;
 288         workqueue->flags = flags;
 289         workqueue->priv = priv;
 290         workqueue->grace_period_fct = grace_period_fct;
 291         workqueue->initialize_worker_fct = initialize_worker_fct;
 292         workqueue->finalize_worker_fct = finalize_worker_fct;
 293         workqueue->worker_before_wait_fct = worker_before_wait_fct;
 294         workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
 295         workqueue->worker_before_pause_fct = worker_before_pause_fct;
 296         workqueue->worker_after_resume_fct = worker_after_resume_fct;
 297         workqueue->cpu_affinity = cpu_affinity;
 298         workqueue->loop_count = 0;
 299         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 300         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 301         if (ret) {
 302                 urcu_die(ret);
 303         }
 304         return workqueue;
 305 }
 306
 307 static void wake_worker_thread(struct urcu_workqueue *workqueue)
 308 {
 309         if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_WORKQUEUE_RT))
 310                 futex_wake_up(&workqueue->futex);
 311 }
 312
 313 static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
 314 {
 315         int ret;
 316         void *retval;
 317
 318         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
 319         wake_worker_thread(workqueue);
 320
 321         ret = pthread_join(workqueue->tid, &retval);
 322         if (ret) {
 323                 urcu_die(ret);
 324         }
 325         if (retval != NULL) {
 326                 urcu_die(EINVAL);
 327         }
 328         workqueue->flags &= ~URCU_WORKQUEUE_STOP;
 329         workqueue->tid = 0;
 330         return 0;
 331 }
 332
 333 void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
 334 {
 335         if (workqueue == NULL) {
 336                 return;
 337         }
 338         if (urcu_workqueue_destroy_worker(workqueue)) {
 339                 urcu_die(errno);
 340         }
 341         assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
 342         free(workqueue);
 343 }
 344
 345 void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
 346                       struct urcu_work *work,
 347                       void (*func)(struct urcu_work *work))
 348 {
 349         cds_wfcq_node_init(&work->next);
 350         work->func = func;
 351         cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
 352         uatomic_inc(&workqueue->qlen);
 353         wake_worker_thread(workqueue);
 354 }
 355
 356 static
 357 void free_completion(struct urcu_ref *ref)
 358 {
 359         struct urcu_workqueue_completion *completion;
 360
 361         completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
 362         free(completion);
 363 }
 364
 365 static
 366 void _urcu_workqueue_wait_complete(struct urcu_work *work)
 367 {
 368         struct urcu_workqueue_completion_work *completion_work;
 369         struct urcu_workqueue_completion *completion;
 370
 371         completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
 372         completion = completion_work->completion;
 373         if (!uatomic_sub_return(&completion->barrier_count, 1))
 374                 futex_wake_up(&completion->futex);
 375         urcu_ref_put(&completion->ref, free_completion);
 376         free(completion_work);
 377 }
 378
 379 struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
 380 {
 381         struct urcu_workqueue_completion *completion;
 382
 383         completion = calloc(sizeof(*completion), 1);
 384         if (!completion)
 385                 urcu_die(errno);
 386         urcu_ref_set(&completion->ref, 1);
 387         completion->barrier_count = 0;
 388         return completion;
 389 }
 390
 391 void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
 392 {
 393         urcu_ref_put(&completion->ref, free_completion);
 394 }
 395
 396 void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
 397 {
 398         /* Wait for them */
 399         for (;;) {
 400                 uatomic_dec(&completion->futex);
 401                 /* Decrement futex before reading barrier_count */
 402                 cmm_smp_mb();
 403                 if (!uatomic_read(&completion->barrier_count))
 404                         break;
 405                 futex_wait(&completion->futex);
 406         }
 407 }
 408
 409 void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
 410                 struct urcu_workqueue_completion *completion)
 411 {
 412         struct urcu_workqueue_completion_work *work;
 413
 414         work = calloc(sizeof(*work), 1);
 415         if (!work)
 416                 urcu_die(errno);
 417         work->completion = completion;
 418         urcu_ref_get(&completion->ref);
 419         uatomic_inc(&completion->barrier_count);
 420         urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
 421 }
 422
 423 /*
 424  * Wait for all in-flight work to complete execution.
 425  */
 426 void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
 427 {
 428         struct urcu_workqueue_completion *completion;
 429
 430         completion = urcu_workqueue_create_completion();
 431         if (!completion)
 432                 urcu_die(ENOMEM);
 433         urcu_workqueue_queue_completion(workqueue, completion);
 434         urcu_workqueue_wait_completion(completion);
 435         urcu_workqueue_destroy_completion(completion);
 436 }
 437
 438 /* To be used in before fork handler. */
 439 void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
 440 {
 441         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
 442         cmm_smp_mb__after_uatomic_or();
 443         wake_worker_thread(workqueue);
 444
 445         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
 446                 (void) poll(NULL, 0, 1);
 447 }
 448
 449 /* To be used in after fork parent handler. */
 450 void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
 451 {
 452         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
 453         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
 454                 (void) poll(NULL, 0, 1);
 455 }
 456
 457 void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
 458 {
 459         int ret;
 460
 461         /* Clear workqueue state from parent. */
 462         workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
 463         workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
 464         workqueue->tid = 0;
 465         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 466         if (ret) {
 467                 urcu_die(ret);
 468         }
 469 }