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 #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 #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 __attribute__((unused)))
 128 {
 129         return 0;
 130 }
 131 #endif
 132
 133 static void futex_wait(int32_t *futex)
 134 {
 135         int ret;
 136
 137         /* Read condition before read futex */
 138         cmm_smp_mb();
 139         while (uatomic_read(futex) == -1) {
 140                 if (!futex_async(futex, FUTEX_WAIT, -1, NULL, NULL, 0)) {
 141                         /*
 142                          * Prior queued wakeups queued by unrelated code
 143                          * using the same address can cause futex wait to
 144                          * return 0 even through the futex value is still
 145                          * -1 (spurious wakeups). Check the value again
 146                          * in user-space to validate whether it really
 147                          * differs from -1.
 148                          */
 149                         continue;
 150                 }
 151                 switch (errno) {
 152                 case EAGAIN:
 153                         /* Value already changed. */
 154                         return;
 155                 case EINTR:
 156                         /* Retry if interrupted by signal. */
 157                         break;  /* Get out of switch. Check again. */
 158                 default:
 159                         /* Unexpected error. */
 160                         urcu_die(errno);
 161                 }
 162         }
 163 }
 164
 165 static void futex_wake_up(int32_t *futex)
 166 {
 167         /* Write to condition before reading/writing futex */
 168         cmm_smp_mb();
 169         if (caa_unlikely(uatomic_read(futex) == -1)) {
 170                 uatomic_set(futex, 0);
 171                 if (futex_async(futex, FUTEX_WAKE, 1,
 172                                 NULL, NULL, 0) < 0)
 173                         urcu_die(errno);
 174         }
 175 }
 176
 177 /* This is the code run by each worker thread. */
 178
 179 static void *workqueue_thread(void *arg)
 180 {
 181         unsigned long cbcount;
 182         struct urcu_workqueue *workqueue = (struct urcu_workqueue *) arg;
 183         int rt = !!(uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_RT);
 184
 185         if (set_thread_cpu_affinity(workqueue))
 186                 urcu_die(errno);
 187
 188         if (workqueue->initialize_worker_fct)
 189                 workqueue->initialize_worker_fct(workqueue, workqueue->priv);
 190
 191         if (!rt) {
 192                 uatomic_dec(&workqueue->futex);
 193                 /* Decrement futex before reading workqueue */
 194                 cmm_smp_mb();
 195         }
 196         for (;;) {
 197                 struct cds_wfcq_head cbs_tmp_head;
 198                 struct cds_wfcq_tail cbs_tmp_tail;
 199                 struct cds_wfcq_node *cbs, *cbs_tmp_n;
 200                 enum cds_wfcq_ret splice_ret;
 201
 202                 if (set_thread_cpu_affinity(workqueue))
 203                         urcu_die(errno);
 204
 205                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) {
 206                         /*
 207                          * Pause requested. Become quiescent: remove
 208                          * ourself from all global lists, and don't
 209                          * process any callback. The callback lists may
 210                          * still be non-empty though.
 211                          */
 212                         if (workqueue->worker_before_pause_fct)
 213                                 workqueue->worker_before_pause_fct(workqueue, workqueue->priv);
 214                         cmm_smp_mb__before_uatomic_or();
 215                         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSED);
 216                         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSE) != 0)
 217                                 (void) poll(NULL, 0, 1);
 218                         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSED);
 219                         cmm_smp_mb__after_uatomic_and();
 220                         if (workqueue->worker_after_resume_fct)
 221                                 workqueue->worker_after_resume_fct(workqueue, workqueue->priv);
 222                 }
 223
 224                 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
 225                 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
 226                         &cbs_tmp_tail, &workqueue->cbs_head, &workqueue->cbs_tail);
 227                 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
 228                 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
 229                 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
 230                         if (workqueue->grace_period_fct)
 231                                 workqueue->grace_period_fct(workqueue, workqueue->priv);
 232                         cbcount = 0;
 233                         __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
 234                                         &cbs_tmp_tail, cbs, cbs_tmp_n) {
 235                                 struct urcu_work *uwp;
 236
 237                                 uwp = caa_container_of(cbs,
 238                                         struct urcu_work, next);
 239                                 uwp->func(uwp);
 240                                 cbcount++;
 241                         }
 242                         uatomic_sub(&workqueue->qlen, cbcount);
 243                 }
 244                 if (uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_STOP)
 245                         break;
 246                 if (workqueue->worker_before_wait_fct)
 247                         workqueue->worker_before_wait_fct(workqueue, workqueue->priv);
 248                 if (!rt) {
 249                         if (cds_wfcq_empty(&workqueue->cbs_head,
 250                                         &workqueue->cbs_tail)) {
 251                                 futex_wait(&workqueue->futex);
 252                                 uatomic_dec(&workqueue->futex);
 253                                 /*
 254                                  * Decrement futex before reading
 255                                  * urcu_work list.
 256                                  */
 257                                 cmm_smp_mb();
 258                         }
 259                 } else {
 260                         if (cds_wfcq_empty(&workqueue->cbs_head,
 261                                         &workqueue->cbs_tail)) {
 262                                 (void) poll(NULL, 0, 10);
 263                         }
 264                 }
 265                 if (workqueue->worker_after_wake_up_fct)
 266                         workqueue->worker_after_wake_up_fct(workqueue, workqueue->priv);
 267         }
 268         if (!rt) {
 269                 /*
 270                  * Read urcu_work list before write futex.
 271                  */
 272                 cmm_smp_mb();
 273                 uatomic_set(&workqueue->futex, 0);
 274         }
 275         if (workqueue->finalize_worker_fct)
 276                 workqueue->finalize_worker_fct(workqueue, workqueue->priv);
 277         return NULL;
 278 }
 279
 280 struct urcu_workqueue *urcu_workqueue_create(unsigned long flags,
 281                 int cpu_affinity, void *priv,
 282                 void (*grace_period_fct)(struct urcu_workqueue *workqueue, void *priv),
 283                 void (*initialize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 284                 void (*finalize_worker_fct)(struct urcu_workqueue *workqueue, void *priv),
 285                 void (*worker_before_wait_fct)(struct urcu_workqueue *workqueue, void *priv),
 286                 void (*worker_after_wake_up_fct)(struct urcu_workqueue *workqueue, void *priv),
 287                 void (*worker_before_pause_fct)(struct urcu_workqueue *workqueue, void *priv),
 288                 void (*worker_after_resume_fct)(struct urcu_workqueue *workqueue, void *priv))
 289 {
 290         struct urcu_workqueue *workqueue;
 291         int ret;
 292
 293         workqueue = malloc(sizeof(*workqueue));
 294         if (workqueue == NULL)
 295                 urcu_die(errno);
 296         memset(workqueue, '\0', sizeof(*workqueue));
 297         cds_wfcq_init(&workqueue->cbs_head, &workqueue->cbs_tail);
 298         workqueue->qlen = 0;
 299         workqueue->futex = 0;
 300         workqueue->flags = flags;
 301         workqueue->priv = priv;
 302         workqueue->grace_period_fct = grace_period_fct;
 303         workqueue->initialize_worker_fct = initialize_worker_fct;
 304         workqueue->finalize_worker_fct = finalize_worker_fct;
 305         workqueue->worker_before_wait_fct = worker_before_wait_fct;
 306         workqueue->worker_after_wake_up_fct = worker_after_wake_up_fct;
 307         workqueue->worker_before_pause_fct = worker_before_pause_fct;
 308         workqueue->worker_after_resume_fct = worker_after_resume_fct;
 309         workqueue->cpu_affinity = cpu_affinity;
 310         workqueue->loop_count = 0;
 311         cmm_smp_mb();  /* Structure initialized before pointer is planted. */
 312         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 313         if (ret) {
 314                 urcu_die(ret);
 315         }
 316         return workqueue;
 317 }
 318
 319 static void wake_worker_thread(struct urcu_workqueue *workqueue)
 320 {
 321         if (!(_CMM_LOAD_SHARED(workqueue->flags) & URCU_WORKQUEUE_RT))
 322                 futex_wake_up(&workqueue->futex);
 323 }
 324
 325 static int urcu_workqueue_destroy_worker(struct urcu_workqueue *workqueue)
 326 {
 327         int ret;
 328         void *retval;
 329
 330         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_STOP);
 331         wake_worker_thread(workqueue);
 332
 333         ret = pthread_join(workqueue->tid, &retval);
 334         if (ret) {
 335                 urcu_die(ret);
 336         }
 337         if (retval != NULL) {
 338                 urcu_die(EINVAL);
 339         }
 340         workqueue->flags &= ~URCU_WORKQUEUE_STOP;
 341         workqueue->tid = 0;
 342         return 0;
 343 }
 344
 345 void urcu_workqueue_destroy(struct urcu_workqueue *workqueue)
 346 {
 347         if (workqueue == NULL) {
 348                 return;
 349         }
 350         if (urcu_workqueue_destroy_worker(workqueue)) {
 351                 urcu_die(errno);
 352         }
 353         assert(cds_wfcq_empty(&workqueue->cbs_head, &workqueue->cbs_tail));
 354         free(workqueue);
 355 }
 356
 357 void urcu_workqueue_queue_work(struct urcu_workqueue *workqueue,
 358                       struct urcu_work *work,
 359                       void (*func)(struct urcu_work *work))
 360 {
 361         cds_wfcq_node_init(&work->next);
 362         work->func = func;
 363         cds_wfcq_enqueue(&workqueue->cbs_head, &workqueue->cbs_tail, &work->next);
 364         uatomic_inc(&workqueue->qlen);
 365         wake_worker_thread(workqueue);
 366 }
 367
 368 static
 369 void free_completion(struct urcu_ref *ref)
 370 {
 371         struct urcu_workqueue_completion *completion;
 372
 373         completion = caa_container_of(ref, struct urcu_workqueue_completion, ref);
 374         free(completion);
 375 }
 376
 377 static
 378 void _urcu_workqueue_wait_complete(struct urcu_work *work)
 379 {
 380         struct urcu_workqueue_completion_work *completion_work;
 381         struct urcu_workqueue_completion *completion;
 382
 383         completion_work = caa_container_of(work, struct urcu_workqueue_completion_work, work);
 384         completion = completion_work->completion;
 385         if (!uatomic_sub_return(&completion->barrier_count, 1))
 386                 futex_wake_up(&completion->futex);
 387         urcu_ref_put(&completion->ref, free_completion);
 388         free(completion_work);
 389 }
 390
 391 struct urcu_workqueue_completion *urcu_workqueue_create_completion(void)
 392 {
 393         struct urcu_workqueue_completion *completion;
 394
 395         completion = calloc(sizeof(*completion), 1);
 396         if (!completion)
 397                 urcu_die(errno);
 398         urcu_ref_set(&completion->ref, 1);
 399         completion->barrier_count = 0;
 400         return completion;
 401 }
 402
 403 void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion *completion)
 404 {
 405         urcu_ref_put(&completion->ref, free_completion);
 406 }
 407
 408 void urcu_workqueue_wait_completion(struct urcu_workqueue_completion *completion)
 409 {
 410         /* Wait for them */
 411         for (;;) {
 412                 uatomic_dec(&completion->futex);
 413                 /* Decrement futex before reading barrier_count */
 414                 cmm_smp_mb();
 415                 if (!uatomic_read(&completion->barrier_count))
 416                         break;
 417                 futex_wait(&completion->futex);
 418         }
 419 }
 420
 421 void urcu_workqueue_queue_completion(struct urcu_workqueue *workqueue,
 422                 struct urcu_workqueue_completion *completion)
 423 {
 424         struct urcu_workqueue_completion_work *work;
 425
 426         work = calloc(sizeof(*work), 1);
 427         if (!work)
 428                 urcu_die(errno);
 429         work->completion = completion;
 430         urcu_ref_get(&completion->ref);
 431         uatomic_inc(&completion->barrier_count);
 432         urcu_workqueue_queue_work(workqueue, &work->work, _urcu_workqueue_wait_complete);
 433 }
 434
 435 /*
 436  * Wait for all in-flight work to complete execution.
 437  */
 438 void urcu_workqueue_flush_queued_work(struct urcu_workqueue *workqueue)
 439 {
 440         struct urcu_workqueue_completion *completion;
 441
 442         completion = urcu_workqueue_create_completion();
 443         if (!completion)
 444                 urcu_die(ENOMEM);
 445         urcu_workqueue_queue_completion(workqueue, completion);
 446         urcu_workqueue_wait_completion(completion);
 447         urcu_workqueue_destroy_completion(completion);
 448 }
 449
 450 /* To be used in before fork handler. */
 451 void urcu_workqueue_pause_worker(struct urcu_workqueue *workqueue)
 452 {
 453         uatomic_or(&workqueue->flags, URCU_WORKQUEUE_PAUSE);
 454         cmm_smp_mb__after_uatomic_or();
 455         wake_worker_thread(workqueue);
 456
 457         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) == 0)
 458                 (void) poll(NULL, 0, 1);
 459 }
 460
 461 /* To be used in after fork parent handler. */
 462 void urcu_workqueue_resume_worker(struct urcu_workqueue *workqueue)
 463 {
 464         uatomic_and(&workqueue->flags, ~URCU_WORKQUEUE_PAUSE);
 465         while ((uatomic_read(&workqueue->flags) & URCU_WORKQUEUE_PAUSED) != 0)
 466                 (void) poll(NULL, 0, 1);
 467 }
 468
 469 void urcu_workqueue_create_worker(struct urcu_workqueue *workqueue)
 470 {
 471         int ret;
 472
 473         /* Clear workqueue state from parent. */
 474         workqueue->flags &= ~URCU_WORKQUEUE_PAUSED;
 475         workqueue->flags &= ~URCU_WORKQUEUE_PAUSE;
 476         workqueue->tid = 0;
 477         ret = pthread_create(&workqueue->tid, NULL, workqueue_thread, workqueue);
 478         if (ret) {
 479                 urcu_die(ret);
 480         }
 481 }