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 either:
  65  *   fct | DQ_FCT_BIT, data (if fct is aligned), or
  66  *   -2L, fct, data (if fct is not aligned).
  67  * Here, DQ_FCT_MARK == ~DQ_FCT_BIT. Required for the test order.
  68  */
  69 #define DQ_FCT_BIT              (1 << 0)
  70 #define DQ_IS_FCT_BIT(x)        ((unsigned long)(x) & DQ_FCT_BIT)
  71 #define DQ_SET_FCT_BIT(x)       \
  72         (x = (void *)((unsigned long)(x) | DQ_FCT_BIT))
  73 #define DQ_CLEAR_FCT_BIT(x)     \
  74         (x = (void *)((unsigned long)(x) & ~DQ_FCT_BIT))
  75 #define DQ_FCT_MARK             ((void *)(~DQ_FCT_BIT))
  76
  77 /*
  78  * This code section can only be included in LGPL 2.1 compatible source code.
  79  * See below for the function call wrappers which can be used in code meant to
  80  * be only linked with the Userspace RCU library. This comes with a small
  81  * performance degradation on the read-side due to the added function calls.
  82  * This is required to permit relinking with newer versions of the library.
  83  */
  84
  85 #ifdef DEBUG_RCU
  86 #define rcu_assert(args...)     assert(args)
  87 #else
  88 #define rcu_assert(args...)
  89 #endif
  90
  91 /*
  92  * defer queue.
  93  * Contains pointers. Encoded to save space when same callback is often used.
  94  * When looking up the next item:
  95  * - if DQ_FCT_BIT is set, set the current callback to DQ_CLEAR_FCT_BIT(ptr)
  96  *   - next element contains pointer to data.
  97  * - else if item == DQ_FCT_MARK
  98  *   - set the current callback to next element ptr
  99  *   - following next element contains pointer to data.
 100  * - else current element contains data
 101  */
 102 struct defer_queue {
 103         unsigned long head;     /* add element at head */
 104         void *last_fct_in;      /* last fct pointer encoded */
 105         unsigned long tail;     /* next element to remove at tail */
 106         void *last_fct_out;     /* last fct pointer encoded */
 107         void **q;
 108         /* registry information */
 109         unsigned long last_head;
 110         struct cds_list_head list;      /* list of thread queues */
 111 };
 112
 113 /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */
 114 #include "urcu-defer.h"
 115
 116 void __attribute__((destructor)) rcu_defer_exit(void);
 117
 118 extern void synchronize_rcu(void);
 119
 120 /*
 121  * rcu_defer_mutex nests inside defer_thread_mutex.
 122  */
 123 static pthread_mutex_t rcu_defer_mutex = PTHREAD_MUTEX_INITIALIZER;
 124 static pthread_mutex_t defer_thread_mutex = PTHREAD_MUTEX_INITIALIZER;
 125
 126 static int32_t defer_thread_futex;
 127 static int32_t defer_thread_stop;
 128
 129 /*
 130  * Written to only by each individual deferer. Read by both the deferer and
 131  * the reclamation tread.
 132  */
 133 static struct defer_queue __thread defer_queue;
 134 static CDS_LIST_HEAD(registry_defer);
 135 static pthread_t tid_defer;
 136
 137 static void mutex_lock_defer(pthread_mutex_t *mutex)
 138 {
 139         int ret;
 140
 141 #ifndef DISTRUST_SIGNALS_EXTREME
 142         ret = pthread_mutex_lock(mutex);
 143         if (ret) {
 144                 perror("Error in pthread mutex lock");
 145                 exit(-1);
 146         }
 147 #else /* #ifndef DISTRUST_SIGNALS_EXTREME */
 148         while ((ret = pthread_mutex_trylock(mutex)) != 0) {
 149                 if (ret != EBUSY && ret != EINTR) {
 150                         printf("ret = %d, errno = %d\n", ret, errno);
 151                         perror("Error in pthread mutex lock");
 152                         exit(-1);
 153                 }
 154                 poll(NULL,0,10);
 155         }
 156 #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */
 157 }
 158
 159 /*
 160  * Wake-up any waiting defer thread. Called from many concurrent threads.
 161  */
 162 static void wake_up_defer(void)
 163 {
 164         if (caa_unlikely(uatomic_read(&defer_thread_futex) == -1)) {
 165                 uatomic_set(&defer_thread_futex, 0);
 166                 futex_noasync(&defer_thread_futex, FUTEX_WAKE, 1,
 167                       NULL, NULL, 0);
 168         }
 169 }
 170
 171 static unsigned long rcu_defer_num_callbacks(void)
 172 {
 173         unsigned long num_items = 0, head;
 174         struct defer_queue *index;
 175
 176         mutex_lock_defer(&rcu_defer_mutex);
 177         cds_list_for_each_entry(index, &registry_defer, list) {
 178                 head = CMM_LOAD_SHARED(index->head);
 179                 num_items += head - index->tail;
 180         }
 181         mutex_unlock(&rcu_defer_mutex);
 182         return num_items;
 183 }
 184
 185 /*
 186  * Defer thread waiting. Single thread.
 187  */
 188 static void wait_defer(void)
 189 {
 190         uatomic_dec(&defer_thread_futex);
 191         /* Write futex before read queue */
 192         /* Write futex before read defer_thread_stop */
 193         cmm_smp_mb();
 194         if (_CMM_LOAD_SHARED(defer_thread_stop)) {
 195                 uatomic_set(&defer_thread_futex, 0);
 196                 pthread_exit(0);
 197         }
 198         if (rcu_defer_num_callbacks()) {
 199                 cmm_smp_mb();   /* Read queue before write futex */
 200                 /* Callbacks are queued, don't wait. */
 201                 uatomic_set(&defer_thread_futex, 0);
 202         } else {
 203                 cmm_smp_rmb();  /* Read queue before read futex */
 204                 if (uatomic_read(&defer_thread_futex) == -1)
 205                         futex_noasync(&defer_thread_futex, FUTEX_WAIT, -1,
 206                               NULL, NULL, 0);
 207         }
 208 }
 209
 210 /*
 211  * Must be called after Q.S. is reached.
 212  */
 213 static void rcu_defer_barrier_queue(struct defer_queue *queue,
 214                                     unsigned long head)
 215 {
 216         unsigned long i;
 217         void (*fct)(void *p);
 218         void *p;
 219
 220         /*
 221          * Tail is only modified when lock is held.
 222          * Head is only modified by owner thread.
 223          */
 224
 225         for (i = queue->tail; i != head;) {
 226                 cmm_smp_rmb();       /* read head before q[]. */
 227                 p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 228                 if (caa_unlikely(DQ_IS_FCT_BIT(p))) {
 229                         DQ_CLEAR_FCT_BIT(p);
 230                         queue->last_fct_out = p;
 231                         p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 232                 } else if (caa_unlikely(p == DQ_FCT_MARK)) {
 233                         p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 234                         queue->last_fct_out = p;
 235                         p = CMM_LOAD_SHARED(queue->q[i++ & DEFER_QUEUE_MASK]);
 236                 }
 237                 fct = queue->last_fct_out;
 238                 fct(p);
 239         }
 240         cmm_smp_mb();   /* push tail after having used q[] */
 241         CMM_STORE_SHARED(queue->tail, i);
 242 }
 243
 244 static void _rcu_defer_barrier_thread(void)
 245 {
 246         unsigned long head, num_items;
 247
 248         head = defer_queue.head;
 249         num_items = head - defer_queue.tail;
 250         if (caa_unlikely(!num_items))
 251                 return;
 252         synchronize_rcu();
 253         rcu_defer_barrier_queue(&defer_queue, head);
 254 }
 255
 256 void rcu_defer_barrier_thread(void)
 257 {
 258         mutex_lock_defer(&rcu_defer_mutex);
 259         _rcu_defer_barrier_thread();
 260         mutex_unlock(&rcu_defer_mutex);
 261 }
 262
 263 /*
 264  * rcu_defer_barrier - Execute all queued rcu callbacks.
 265  *
 266  * Execute all RCU callbacks queued before rcu_defer_barrier() execution.
 267  * All callbacks queued on the local thread prior to a rcu_defer_barrier() call
 268  * are guaranteed to be executed.
 269  * Callbacks queued by other threads concurrently with rcu_defer_barrier()
 270  * execution are not guaranteed to be executed in the current batch (could
 271  * be left for the next batch). These callbacks queued by other threads are only
 272  * guaranteed to be executed if there is explicit synchronization between
 273  * the thread adding to the queue and the thread issuing the defer_barrier call.
 274  */
 275
 276 void rcu_defer_barrier(void)
 277 {
 278         struct defer_queue *index;
 279         unsigned long num_items = 0;
 280
 281         if (cds_list_empty(&registry_defer))
 282                 return;
 283
 284         mutex_lock_defer(&rcu_defer_mutex);
 285         cds_list_for_each_entry(index, &registry_defer, list) {
 286                 index->last_head = CMM_LOAD_SHARED(index->head);
 287                 num_items += index->last_head - index->tail;
 288         }
 289         if (caa_likely(!num_items)) {
 290                 /*
 291                  * We skip the grace period because there are no queued
 292                  * callbacks to execute.
 293                  */
 294                 goto end;
 295         }
 296         synchronize_rcu();
 297         cds_list_for_each_entry(index, &registry_defer, list)
 298                 rcu_defer_barrier_queue(index, index->last_head);
 299 end:
 300         mutex_unlock(&rcu_defer_mutex);
 301 }
 302
 303 /*
 304  * _defer_rcu - Queue a RCU callback.
 305  */
 306 void _defer_rcu(void (*fct)(void *p), void *p)
 307 {
 308         unsigned long head, tail;
 309
 310         /*
 311          * Head is only modified by ourself. Tail can be modified by reclamation
 312          * thread.
 313          */
 314         head = defer_queue.head;
 315         tail = CMM_LOAD_SHARED(defer_queue.tail);
 316
 317         /*
 318          * If queue is full, or reached threshold. Empty queue ourself.
 319          * Worse-case: must allow 2 supplementary entries for fct pointer.
 320          */
 321         if (caa_unlikely(head - tail >= DEFER_QUEUE_SIZE - 2)) {
 322                 assert(head - tail <= DEFER_QUEUE_SIZE);
 323                 rcu_defer_barrier_thread();
 324                 assert(head - CMM_LOAD_SHARED(defer_queue.tail) == 0);
 325         }
 326
 327         /*
 328          * Encode:
 329          * if the function is not changed and the data is aligned and it is
 330          * not the marker:
 331          *      store the data
 332          * otherwise if the function is aligned and its not the marker:
 333          *      store the function with DQ_FCT_BIT
 334          *      store the data
 335          * otherwise:
 336          *      store the marker (DQ_FCT_MARK)
 337          *      store the function
 338          *      store the data
 339          *
 340          * Decode: see the comments before 'struct defer_queue'
 341          *         or the code in rcu_defer_barrier_queue().
 342          */
 343         if (caa_unlikely(defer_queue.last_fct_in != fct
 344                         || DQ_IS_FCT_BIT(p)
 345                         || p == DQ_FCT_MARK)) {
 346                 defer_queue.last_fct_in = fct;
 347                 if (caa_unlikely(DQ_IS_FCT_BIT(fct) || fct == DQ_FCT_MARK)) {
 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                 } else {
 353                         DQ_SET_FCT_BIT(fct);
 354                         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK],
 355                                       fct);
 356                 }
 357         }
 358         _CMM_STORE_SHARED(defer_queue.q[head++ & DEFER_QUEUE_MASK], p);
 359         cmm_smp_wmb();  /* Publish new pointer before head */
 360                         /* Write q[] before head. */
 361         CMM_STORE_SHARED(defer_queue.head, head);
 362         cmm_smp_mb();   /* Write queue head before read futex */
 363         /*
 364          * Wake-up any waiting defer thread.
 365          */
 366         wake_up_defer();
 367 }
 368
 369 void *thr_defer(void *args)
 370 {
 371         for (;;) {
 372                 /*
 373                  * "Be green". Don't wake up the CPU if there is no RCU work
 374                  * to perform whatsoever. Aims at saving laptop battery life by
 375                  * leaving the processor in sleep state when idle.
 376                  */
 377                 wait_defer();
 378                 /* Sleeping after wait_defer to let many callbacks enqueue */
 379                 poll(NULL,0,100);       /* wait for 100ms */
 380                 rcu_defer_barrier();
 381         }
 382
 383         return NULL;
 384 }
 385
 386 /*
 387  * library wrappers to be used by non-LGPL compatible source code.
 388  */
 389
 390 void defer_rcu(void (*fct)(void *p), void *p)
 391 {
 392         _defer_rcu(fct, p);
 393 }
 394
 395 static void start_defer_thread(void)
 396 {
 397         int ret;
 398
 399         ret = pthread_create(&tid_defer, NULL, thr_defer, NULL);
 400         assert(!ret);
 401 }
 402
 403 static void stop_defer_thread(void)
 404 {
 405         int ret;
 406         void *tret;
 407
 408         _CMM_STORE_SHARED(defer_thread_stop, 1);
 409         /* Store defer_thread_stop before testing futex */
 410         cmm_smp_mb();
 411         wake_up_defer();
 412
 413         ret = pthread_join(tid_defer, &tret);
 414         assert(!ret);
 415
 416         CMM_STORE_SHARED(defer_thread_stop, 0);
 417         /* defer thread should always exit when futex value is 0 */
 418         assert(uatomic_read(&defer_thread_futex) == 0);
 419 }
 420
 421 int rcu_defer_register_thread(void)
 422 {
 423         int was_empty;
 424
 425         assert(defer_queue.last_head == 0);
 426         assert(defer_queue.q == NULL);
 427         defer_queue.q = malloc(sizeof(void *) * DEFER_QUEUE_SIZE);
 428         if (!defer_queue.q)
 429                 return -ENOMEM;
 430
 431         mutex_lock_defer(&defer_thread_mutex);
 432         mutex_lock_defer(&rcu_defer_mutex);
 433         was_empty = cds_list_empty(&registry_defer);
 434         cds_list_add(&defer_queue.list, &registry_defer);
 435         mutex_unlock(&rcu_defer_mutex);
 436
 437         if (was_empty)
 438                 start_defer_thread();
 439         mutex_unlock(&defer_thread_mutex);
 440         return 0;
 441 }
 442
 443 void rcu_defer_unregister_thread(void)
 444 {
 445         int is_empty;
 446
 447         mutex_lock_defer(&defer_thread_mutex);
 448         mutex_lock_defer(&rcu_defer_mutex);
 449         cds_list_del(&defer_queue.list);
 450         _rcu_defer_barrier_thread();
 451         free(defer_queue.q);
 452         defer_queue.q = NULL;
 453         is_empty = cds_list_empty(&registry_defer);
 454         mutex_unlock(&rcu_defer_mutex);
 455
 456         if (is_empty)
 457                 stop_defer_thread();
 458         mutex_unlock(&defer_thread_mutex);
 459 }
 460
 461 void rcu_defer_exit(void)
 462 {
 463         assert(cds_list_empty(&registry_defer));
 464 }
 465
 466 #endif /* _URCU_DEFER_IMPL_H */