urcu-defer.c

   1 /*
   2  * urcu-defer.c
   3  *
   4  * Userspace RCU library - batch memory reclamation
   5  *
   6  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   7  *
   8  * This library is free software; you can redistribute it and/or
   9  * modify it under the terms of the GNU Lesser General Public
  10  * License as published by the Free Software Foundation; either
  11  * version 2.1 of the License, or (at your option) any later version.
  12  *
  13  * This library is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * Lesser General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU Lesser General Public
  19  * License along with this library; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  21  */
  22
  23 #include <stdio.h>
  24 #include <pthread.h>
  25 #include <signal.h>
  26 #include <assert.h>
  27 #include <stdlib.h>
  28 #include <string.h>
  29 #include <errno.h>
  30 #include <poll.h>
  31 #include <sys/time.h>
  32 #include <syscall.h>
  33 #include <unistd.h>
  34
  35 #include "urcu/urcu-futex.h"
  36 #include "urcu-defer-static.h"
  37 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
  38 #include "urcu-defer.h"
  39
  40 void __attribute__((destructor)) rcu_defer_exit(void);
  41
  42 extern void synchronize_rcu(void);
  43
  44 /*
  45  * rcu_defer_mutex nests inside defer_thread_mutex.
  46  */
  47 static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
  48 static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
  49
  50 static int defer_thread_futex;
  51
  52 /*
  53  * Written to only by each individual deferer. Read by both the deferer and
  54  * the reclamation tread.
  55  */
  56 static struct defer_queue __thread defer_queue;
  57 static LIST_HEAD(registry);
  58 static pthread_t tid_defer;
  59
  60 static void mutex_lock(pthread_mutex_t *mutex)
  61 {
  62         int ret;
  63
  64 #ifndef DISTRUST_SIGNALS_EXTREME
  65         ret = pthread_mutex_lock(mutex);
  66         if (ret) {
  67                 perror("Error in pthread mutex lock");
  68                 exit(-1);
  69         }
  70 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
  71         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
  72                 if (ret != EBUSY && ret != EINTR) {
  73                         printf("ret = %d, errno = %d\n", ret, errno);
  74                         perror("Error in pthread mutex lock");
  75                         exit(-1);
  76                 }
  77                 pthread_testcancel();
  78                 poll(NULL,0,10);
  79         }
  80 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
  81 }
  82
  83 static void mutex_unlock(pthread_mutex_t *mutex)
  84 {
  85         int ret;
  86
  87         ret = pthread_mutex_unlock(mutex);
  88         if (ret) {
  89                 perror("Error in pthread mutex unlock");
  90                 exit(-1);
  91         }
  92 }
  93
  94 /*
  95  * Wake-up any waiting defer thread. Called from many concurrent threads.
  96  */
  97 static void wake_up_defer(void)
  98 {
  99         if (unlikely(uatomic_read(&defer_thread_futex) == -1)) {
 100                 uatomic_set(&defer_thread_futex, 0);
 101                 futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
 102                       NULL, NULL, 0);
 103         }
 104 }
 105
 106 static unsigned long rcu_defer_num_callbacks(void)
 107 {
 108         unsigned long num_items = 0, head;
 109         struct defer_queue *index;
 110
 111         mutex_lock(&rcu_defer_mutex);
 112         list_for_each_entry(index, &registry, list) {
 113                 head = CAA_LOAD_SHARED(index->head);
 114                 num_items += head - index->tail;
 115         }
 116         mutex_unlock(&rcu_defer_mutex);
 117         return num_items;
 118 }
 119
 120 /*
 121  * Defer thread waiting. Single thread.
 122  */
 123 static void wait_defer(void)
 124 {
 125         uatomic_dec(&defer_thread_futex);
 126         cmm_smp_mb();   /* Write futex before read queue */
 127         if (rcu_defer_num_callbacks()) {
 128                 cmm_smp_mb();   /* Read queue before write futex */
 129                 /* Callbacks are queued, don't wait. */
 130                 uatomic_set(&defer_thread_futex, 0);
 131         } else {
 132                 cmm_smp_rmb();  /* Read queue before read futex */
 133                 if (uatomic_read(&defer_thread_futex) == -1)
 134                         futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
 135                               NULL, NULL, 0);
 136         }
 137 }
 138
 139 /*
 140  * Must be called after Q.S. is reached.
 141  */
 142 static void rcu_defer_barrier_queue(struct defer_queue *queue,
 143                                     unsigned long head)
 144 {
 145         unsigned long i;
 146         void (*fct)(void *p);
 147         void *p;
 148
 149         /*
 150          * Tail is only modified when lock is held.
 151          * Head is only modified by owner thread.
 152          */
 153
 154         for (i = queue->tail; i != head;) {
 155                 cmm_smp_rmb();       /* read head before q[]. */
 156                 p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 157                 if (unlikely(DQ_IS_FCT_BIT(p))) {
 158                         DQ_CLEAR_FCT_BIT(p);
 159                         queue->last_fct_out = p;
 160                         p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 161                 } else if (unlikely(p == DQ_FCT_MARK)) {
 162                         p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 163                         queue->last_fct_out = p;
 164                         p = CAA_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 165                 }
 166                 fct = queue->last_fct_out;
 167                 fct(p);
 168         }
 169         cmm_smp_mb();   /* push tail after having used q[] */
 170         CAA_STORE_SHARED(queue->tail, i);
 171 }
 172
 173 static void _rcu_defer_barrier_thread(void)
 174 {
 175         unsigned long head, num_items;
 176
 177         head = defer_queue.head;
 178         num_items = head - defer_queue.tail;
 179         if (unlikely(!num_items))
 180                 return;
 181         synchronize_rcu();
 182         rcu_defer_barrier_queue(&defer_queue, head);
 183 }
 184
 185 void rcu_defer_barrier_thread(void)
 186 {
 187         mutex_lock(&rcu_defer_mutex);
 188         _rcu_defer_barrier_thread();
 189         mutex_unlock(&rcu_defer_mutex);
 190 }
 191
 192 /*
 193  * rcu_defer_barrier - Execute all queued rcu callbacks.
 194  *
 195  * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
 196  * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
 197  * are guaranteed to be executed.
 198  * Callbacks queued by other threads concurrently with rcu_defer_barrier()
 199  * execution are not guaranteed to be executed in the current batch (could
 200  * be left for the next batch). These callbacks queued by other threads are only
 201  * guaranteed to be executed if there is explicit synchronization between
 202  * the thread adding to the queue and the thread issuing the defer_barrier call.
 203  */
 204
 205 void rcu_defer_barrier(void)
 206 {
 207         struct defer_queue *index;
 208         unsigned long num_items = 0;
 209
 210         if (list_empty(&registry))
 211                 return;
 212
 213         mutex_lock(&rcu_defer_mutex);
 214         list_for_each_entry(index, &registry, list) {
 215                 index->last_head = CAA_LOAD_SHARED(index->head);
 216                 num_items += index->last_head - index->tail;
 217         }
 218         if (likely(!num_items)) {
 219                 /*
 220                  * We skip the grace period because there are no queued
 221                  * callbacks to execute.
 222                  */
 223                 goto end;
 224         }
 225         synchronize_rcu();
 226         list_for_each_entry(index, &registry, list)
 227                 rcu_defer_barrier_queue(index, index->last_head);
 228 end:
 229         mutex_unlock(&rcu_defer_mutex);
 230 }
 231
 232 /*
 233  * _defer_rcu - Queue a RCU callback.
 234  */
 235 void _defer_rcu(void (*fct)(void *p), void *p)
 236 {
 237         unsigned long head, tail;
 238
 239         /*
 240          * Head is only modified by ourself. Tail can be modified by reclamation
 241          * thread.
 242          */
 243         head = defer_queue.head;
 244         tail = CAA_LOAD_SHARED(defer_queue.tail);
 245
 246         /*
 247          * If queue is full, or reached threshold. Empty queue ourself.
 248          * Worse-case: must allow 2 supplementary entries for fct pointer.
 249          */
 250         if (unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) {
 251                 assert(head - tail <= DEFER_QUEUE_SIZE);
 252                 rcu_defer_barrier_thread();
 253                 assert(head - CAA_LOAD_SHARED(defer_queue.tail) == 0);
 254         }
 255
 256         if (unlikely(defer_queue.last_fct_in != fct)) {
 257                 defer_queue.last_fct_in = fct;
 258                 if (unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) {
 259                         /*
 260                          * If the function to encode is not aligned or the
 261                          * marker, write DQ_FCT_MARK followed by the function
 262                          * pointer.
 263                          */
 264                         _CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 265                                       DQ_FCT_MARK);
 266                         _CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 267                                       fct);
 268                 } else {
 269                         DQ_SET_FCT_BIT(fct);
 270                         _CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 271                                       fct);
 272                 }
 273         } else {
 274                 if (unlikely(DQ_IS_FCT_BIT(p) || p == DQ_FCT_MARK)) {
 275                         /*
 276                          * If the data to encode is not aligned or the marker,
 277                          * write DQ_FCT_MARK followed by the function pointer.
 278                          */
 279                         _CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 280                                       DQ_FCT_MARK);
 281                         _CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 282                                       fct);
 283                 }
 284         }
 285         _CAA_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
 286         cmm_smp_wmb();  /* Publish new pointer before head */
 287                         /* Write q[] before head. */
 288         CAA_STORE_SHARED(defer_queue.head, head);
 289         cmm_smp_mb();   /* Write queue head before read futex */
 290         /*
 291          * Wake-up any waiting defer thread.
 292          */
 293         wake_up_defer();
 294 }
 295
 296 void *thr_defer(void *args)
 297 {
 298         for (;;) {
 299                 pthread_testcancel();
 300                 /*
 301                  * "Be green". Don't wake up the CPU if there is no RCU work
 302                  * to perform whatsoever. Aims at saving laptop battery life by
 303                  * leaving the processor in sleep state when idle.
 304                  */
 305                 wait_defer();
 306                 /* Sleeping after wait_defer to let many callbacks enqueue */
 307                 poll(NULL,0,100);       /* wait for 100ms */
 308                 rcu_defer_barrier();
 309         }
 310
 311         return NULL;
 312 }
 313
 314 /*
 315  * library wrappers to be used by non-LGPL compatible source code.
 316  */
 317
 318 void defer_rcu(void (*fct)(void *p), void *p)
 319 {
 320         _defer_rcu(fct, p);
 321 }
 322
 323 static void start_defer_thread(void)
 324 {
 325         int ret;
 326
 327         ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
 328         assert(!ret);
 329 }
 330
 331 static void stop_defer_thread(void)
 332 {
 333         int ret;
 334         void *tret;
 335
 336         pthread_cancel(tid_defer);
 337         wake_up_defer();
 338         ret = pthread_join(tid_defer, &tret);
 339         assert(!ret);
 340 }
 341
 342 void rcu_defer_register_thread(void)
 343 {
 344         int was_empty;
 345
 346         assert(defer_queue.last_head == 0);
 347         assert(defer_queue.q == NULL);
 348         defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
 349
 350         mutex_lock(&defer_thread_mutex);
 351         mutex_lock(&rcu_defer_mutex);
 352         was_empty = list_empty(&registry);
 353         list_add(&defer_queue.list, &registry);
 354         mutex_unlock(&rcu_defer_mutex);
 355
 356         if (was_empty)
 357                 start_defer_thread();
 358         mutex_unlock(&defer_thread_mutex);
 359 }
 360
 361 void rcu_defer_unregister_thread(void)
 362 {
 363         int is_empty;
 364
 365         mutex_lock(&defer_thread_mutex);
 366         mutex_lock(&rcu_defer_mutex);
 367         list_del(&defer_queue.list);
 368         _rcu_defer_barrier_thread();
 369         free(defer_queue.q);
 370         defer_queue.q = NULL;
 371         is_empty = list_empty(&registry);
 372         mutex_unlock(&rcu_defer_mutex);
 373
 374         if (is_empty)
 375                 stop_defer_thread();
 376         mutex_unlock(&defer_thread_mutex);
 377 }
 378
 379 void rcu_defer_exit(void)
 380 {
 381         assert(list_empty(&registry));
 382 }