urcu-defer-impl.h

   1 #ifndef _URCU_DEFER_IMPL_H
   2 #define _URCU_DEFER_IMPL_H
   3
   4 /*
   5  * urcu-defer-impl.h
   6  *
   7  * Userspace RCU header - memory reclamation.
   8  *
   9  * TO BE INCLUDED ONLY FROM URCU LIBRARY CODE. See urcu-defer.h for linking
  10  * dynamically with the userspace rcu reclamation library.
  11  *
  12  * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  13  * Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
  14  *
  15  * This library is free software; you can redistribute it and/or
  16  * modify it under the terms of the GNU Lesser General Public
  17  * License as published by the Free Software Foundation; either
  18  * version 2.1 of the License, or (at your option) any later version.
  19  *
  20  * This library is distributed in the hope that it will be useful,
  21  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  22  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  23  * Lesser General Public License for more details.
  24  *
  25  * You should have received a copy of the GNU Lesser General Public
  26  * License along with this library; if not, write to the Free Software
  27  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  28  *
  29  * IBM's contributions to this file may be relicensed under LGPLv2 or later.
  30  */
  31
  32 #include <stdlib.h>
  33 #include <pthread.h>
  34 #include <stdio.h>
  35 #include <signal.h>
  36 #include <assert.h>
  37 #include <string.h>
  38 #include <errno.h>
  39 #include <poll.h>
  40 #include <sys/time.h>
  41 #include <unistd.h>
  42 #include <stdint.h>
  43
  44 #include "urcu/futex.h"
  45
  46 #include <urcu/compiler.h>
  47 #include <urcu/arch.h>
  48 #include <urcu/uatomic.h>
  49 #include <urcu/list.h>
  50 #include <urcu/system.h>
  51
  52 /*
  53  * Number of entries in the per-thread defer queue. Must be power of 2.
  54  */
  55 #define DEFER_QUEUE_SIZE        (1 << 12)
  56 #define DEFER_QUEUE_MASK        (DEFER_QUEUE_SIZE - 1)
  57
  58 /*
  59  * Typically, data is aligned at least on the architecture size.
  60  * Use lowest bit to indicate that the current callback is changing.
  61  * Assumes that (void *)-2L is not used often. Used to encode non-aligned
  62  * functions and non-aligned data using extra space.
  63  * We encode the (void *)-2L fct as: -2L, fct, data.
  64  * We encode the (void *)-2L data as: -2L, fct, data.
  65  * Here, DQ_FCT_MARK == ~DQ_FCT_BIT. Required for the test order.
  66  */
  67 #define DQ_FCT_BIT              (1 << 0)
  68 #define DQ_IS_FCT_BIT(x)        ((unsigned long)(x) & DQ_FCT_BIT)
  69 #define DQ_SET_FCT_BIT(x)       \
  70         (x = (void *)((unsigned long)(x) | DQ_FCT_BIT))
  71 #define DQ_CLEAR_FCT_BIT(x)     \
  72         (x = (void *)((unsigned long)(x) & ~DQ_FCT_BIT))
  73 #define DQ_FCT_MARK             ((void *)(~DQ_FCT_BIT))
  74
  75 /*
  76  * This code section can only be included in LGPL 2.1 compatible source code.
  77  * See below for the function call wrappers which can be used in code meant to
  78  * be only linked with the Userspace RCU library. This comes with a small
  79  * performance degradation on the read-side due to the added function calls.
  80  * This is required to permit relinking with newer versions of the library.
  81  */
  82
  83 #ifdef DEBUG_RCU
  84 #define rcu_assert(args...)     assert(args)
  85 #else
  86 #define rcu_assert(args...)
  87 #endif
  88
  89 /*
  90  * defer queue.
  91  * Contains pointers. Encoded to save space when same callback is often used.
  92  * When looking up the next item:
  93  * - if DQ_FCT_BIT is set, set the current callback to DQ_CLEAR_FCT_BIT(ptr)
  94  *   - next element contains pointer to data.
  95  * - else if item == DQ_FCT_MARK
  96  *   - set the current callback to next element ptr
  97  *   - following next element contains pointer to data.
  98  * - else current element contains data
  99  */
 100 struct defer_queue {
 101         unsigned long head;     /* add element at head */
 102         void *last_fct_in;      /* last fct pointer encoded */
 103         unsigned long tail;     /* next element to remove at tail */
 104         void *last_fct_out;     /* last fct pointer encoded */
 105         void **q;
 106         /* registry information */
 107         unsigned long last_head;
 108         struct cds_list_head list;      /* list of thread queues */
 109 };
 110
 111 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 112 #include "urcu-defer.h"
 113
 114 void __attribute__((destructor)) rcu_defer_exit(void);
 115
 116 extern void synchronize_rcu(void);
 117
 118 /*
 119  * rcu_defer_mutex nests inside defer_thread_mutex.
 120  */
 121 static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
 122 static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
 123
 124 static int32_t defer_thread_futex;
 125 static int32_t defer_thread_stop;
 126
 127 /*
 128  * Written to only by each individual deferer. Read by both the deferer and
 129  * the reclamation tread.
 130  */
 131 static struct defer_queue __thread defer_queue;
 132 static CDS_LIST_HEAD(registry_defer);
 133 static pthread_t tid_defer;
 134
 135 static void mutex_lock_defer(pthread_mutex_t *mutex)
 136 {
 137         int ret;
 138
 139 #ifndef DISTRUST_SIGNALS_EXTREME
 140         ret = pthread_mutex_lock(mutex);
 141         if (ret) {
 142                 perror("Error in pthread mutex lock");
 143                 exit(-1);
 144         }
 145 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 146         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 147                 if (ret != EBUSY && ret != EINTR) {
 148                         printf("ret = %d, errno = %d\n", ret, errno);
 149                         perror("Error in pthread mutex lock");
 150                         exit(-1);
 151                 }
 152                 poll(NULL,0,10);
 153         }
 154 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 155 }
 156
 157 /*
 158  * Wake-up any waiting defer thread. Called from many concurrent threads.
 159  */
 160 static void wake_up_defer(void)
 161 {
 162         if (unlikely(uatomic_read(&defer_thread_futex) == -1)) {
 163                 uatomic_set(&defer_thread_futex, 0);
 164                 futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
 165                       NULL, NULL, 0);
 166         }
 167 }
 168
 169 static unsigned long rcu_defer_num_callbacks(void)
 170 {
 171         unsigned long num_items = 0, head;
 172         struct defer_queue *index;
 173
 174         mutex_lock_defer(&rcu_defer_mutex);
 175         cds_list_for_each_entry(index, &registry_defer, list) {
 176                 head = CMM_LOAD_SHARED(index->head);
 177                 num_items += head - index->tail;
 178         }
 179         mutex_unlock(&rcu_defer_mutex);
 180         return num_items;
 181 }
 182
 183 /*
 184  * Defer thread waiting. Single thread.
 185  */
 186 static void wait_defer(void)
 187 {
 188         uatomic_dec(&defer_thread_futex);
 189         /* Write futex before read queue */
 190         /* Write futex before read defer_thread_stop */
 191         cmm_smp_mb();
 192         if (_CMM_LOAD_SHARED(defer_thread_stop)) {
 193                 uatomic_set(&defer_thread_futex, 0);
 194                 pthread_exit(0);
 195         }
 196         if (rcu_defer_num_callbacks()) {
 197                 cmm_smp_mb();   /* Read queue before write futex */
 198                 /* Callbacks are queued, don't wait. */
 199                 uatomic_set(&defer_thread_futex, 0);
 200         } else {
 201                 cmm_smp_rmb();  /* Read queue before read futex */
 202                 if (uatomic_read(&defer_thread_futex) == -1)
 203                         futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
 204                               NULL, NULL, 0);
 205         }
 206 }
 207
 208 /*
 209  * Must be called after Q.S. is reached.
 210  */
 211 static void rcu_defer_barrier_queue(struct defer_queue *queue,
 212                                     unsigned long head)
 213 {
 214         unsigned long i;
 215         void (*fct)(void *p);
 216         void *p;
 217
 218         /*
 219          * Tail is only modified when lock is held.
 220          * Head is only modified by owner thread.
 221          */
 222
 223         for (i = queue->tail; i != head;) {
 224                 cmm_smp_rmb();       /* read head before q[]. */
 225                 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 226                 if (unlikely(DQ_IS_FCT_BIT(p))) {
 227                         DQ_CLEAR_FCT_BIT(p);
 228                         queue->last_fct_out = p;
 229                         p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 230                 } else if (unlikely(p == DQ_FCT_MARK)) {
 231                         p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 232                         queue->last_fct_out = p;
 233                         p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 234                 }
 235                 fct = queue->last_fct_out;
 236                 fct(p);
 237         }
 238         cmm_smp_mb();   /* push tail after having used q[] */
 239         CMM_STORE_SHARED(queue->tail, i);
 240 }
 241
 242 static void _rcu_defer_barrier_thread(void)
 243 {
 244         unsigned long head, num_items;
 245
 246         head = defer_queue.head;
 247         num_items = head - defer_queue.tail;
 248         if (unlikely(!num_items))
 249                 return;
 250         synchronize_rcu();
 251         rcu_defer_barrier_queue(&defer_queue, head);
 252 }
 253
 254 void rcu_defer_barrier_thread(void)
 255 {
 256         mutex_lock_defer(&rcu_defer_mutex);
 257         _rcu_defer_barrier_thread();
 258         mutex_unlock(&rcu_defer_mutex);
 259 }
 260
 261 /*
 262  * rcu_defer_barrier - Execute all queued rcu callbacks.
 263  *
 264  * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
 265  * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
 266  * are guaranteed to be executed.
 267  * Callbacks queued by other threads concurrently with rcu_defer_barrier()
 268  * execution are not guaranteed to be executed in the current batch (could
 269  * be left for the next batch). These callbacks queued by other threads are only
 270  * guaranteed to be executed if there is explicit synchronization between
 271  * the thread adding to the queue and the thread issuing the defer_barrier call.
 272  */
 273
 274 void rcu_defer_barrier(void)
 275 {
 276         struct defer_queue *index;
 277         unsigned long num_items = 0;
 278
 279         if (cds_list_empty(&registry_defer))
 280                 return;
 281
 282         mutex_lock_defer(&rcu_defer_mutex);
 283         cds_list_for_each_entry(index, &registry_defer, list) {
 284                 index->last_head = CMM_LOAD_SHARED(index->head);
 285                 num_items += index->last_head - index->tail;
 286         }
 287         if (likely(!num_items)) {
 288                 /*
 289                  * We skip the grace period because there are no queued
 290                  * callbacks to execute.
 291                  */
 292                 goto end;
 293         }
 294         synchronize_rcu();
 295         cds_list_for_each_entry(index, &registry_defer, list)
 296                 rcu_defer_barrier_queue(index, index->last_head);
 297 end:
 298         mutex_unlock(&rcu_defer_mutex);
 299 }
 300
 301 /*
 302  * _defer_rcu - Queue a RCU callback.
 303  */
 304 void _defer_rcu(void (*fct)(void *p), void *p)
 305 {
 306         unsigned long head, tail;
 307
 308         /*
 309          * Head is only modified by ourself. Tail can be modified by reclamation
 310          * thread.
 311          */
 312         head = defer_queue.head;
 313         tail = CMM_LOAD_SHARED(defer_queue.tail);
 314
 315         /*
 316          * If queue is full, or reached threshold. Empty queue ourself.
 317          * Worse-case: must allow 2 supplementary entries for fct pointer.
 318          */
 319         if (unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) {
 320                 assert(head - tail <= DEFER_QUEUE_SIZE);
 321                 rcu_defer_barrier_thread();
 322                 assert(head - CMM_LOAD_SHARED(defer_queue.tail) == 0);
 323         }
 324
 325         if (unlikely(defer_queue.last_fct_in != fct)) {
 326                 defer_queue.last_fct_in = fct;
 327                 if (unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) {
 328                         /*
 329                          * If the function to encode is not aligned or the
 330                          * marker, write DQ_FCT_MARK followed by the function
 331                          * pointer.
 332                          */
 333                         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 334                                       DQ_FCT_MARK);
 335                         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 336                                       fct);
 337                 } else {
 338                         DQ_SET_FCT_BIT(fct);
 339                         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 340                                       fct);
 341                 }
 342         } else {
 343                 if (unlikely(DQ_IS_FCT_BIT(p) || p == DQ_FCT_MARK)) {
 344                         /*
 345                          * If the data to encode is not aligned or the marker,
 346                          * write DQ_FCT_MARK followed by the function pointer.
 347                          */
 348                         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 349                                       DQ_FCT_MARK);
 350                         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 351                                       fct);
 352                 }
 353         }
 354         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
 355         cmm_smp_wmb();  /* Publish new pointer before head */
 356                         /* Write q[] before head. */
 357         CMM_STORE_SHARED(defer_queue.head, head);
 358         cmm_smp_mb();   /* Write queue head before read futex */
 359         /*
 360          * Wake-up any waiting defer thread.
 361          */
 362         wake_up_defer();
 363 }
 364
 365 void *thr_defer(void *args)
 366 {
 367         for (;;) {
 368                 /*
 369                  * "Be green". Don't wake up the CPU if there is no RCU work
 370                  * to perform whatsoever. Aims at saving laptop battery life by
 371                  * leaving the processor in sleep state when idle.
 372                  */
 373                 wait_defer();
 374                 /* Sleeping after wait_defer to let many callbacks enqueue */
 375                 poll(NULL,0,100);       /* wait for 100ms */
 376                 rcu_defer_barrier();
 377         }
 378
 379         return NULL;
 380 }
 381
 382 /*
 383  * library wrappers to be used by non-LGPL compatible source code.
 384  */
 385
 386 void defer_rcu(void (*fct)(void *p), void *p)
 387 {
 388         _defer_rcu(fct, p);
 389 }
 390
 391 static void start_defer_thread(void)
 392 {
 393         int ret;
 394
 395         ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
 396         assert(!ret);
 397 }
 398
 399 static void stop_defer_thread(void)
 400 {
 401         int ret;
 402         void *tret;
 403
 404         _CMM_STORE_SHARED(defer_thread_stop, 1);
 405         /* Store defer_thread_stop before testing futex */
 406         cmm_smp_mb();
 407         wake_up_defer();
 408
 409         ret = pthread_join(tid_defer, &tret);
 410         assert(!ret);
 411
 412         CMM_STORE_SHARED(defer_thread_stop, 0);
 413         /* defer thread should always exit when futex value is 0 */
 414         assert(uatomic_read(&defer_thread_futex) == 0);
 415 }
 416
 417 int rcu_defer_register_thread(void)
 418 {
 419         int was_empty;
 420
 421         assert(defer_queue.last_head == 0);
 422         assert(defer_queue.q == NULL);
 423         defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
 424         if (!defer_queue.q)
 425                 return -ENOMEM;
 426
 427         mutex_lock_defer(&defer_thread_mutex);
 428         mutex_lock_defer(&rcu_defer_mutex);
 429         was_empty = cds_list_empty(&registry_defer);
 430         cds_list_add(&defer_queue.list, &registry_defer);
 431         mutex_unlock(&rcu_defer_mutex);
 432
 433         if (was_empty)
 434                 start_defer_thread();
 435         mutex_unlock(&defer_thread_mutex);
 436         return 0;
 437 }
 438
 439 void rcu_defer_unregister_thread(void)
 440 {
 441         int is_empty;
 442
 443         mutex_lock_defer(&defer_thread_mutex);
 444         mutex_lock_defer(&rcu_defer_mutex);
 445         cds_list_del(&defer_queue.list);
 446         _rcu_defer_barrier_thread();
 447         free(defer_queue.q);
 448         defer_queue.q = NULL;
 449         is_empty = cds_list_empty(&registry_defer);
 450         mutex_unlock(&rcu_defer_mutex);
 451
 452         if (is_empty)
 453                 stop_defer_thread();
 454         mutex_unlock(&defer_thread_mutex);
 455 }
 456
 457 void rcu_defer_exit(void)
 458 {
 459         assert(cds_list_empty(&registry_defer));
 460 }
 461
 462 #endif /* _URCU_DEFER_IMPL_H */