tests/benchmark/test_urcu_gc.c

   1 // SPDX-FileCopyrightText: 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   2 //
   3 // SPDX-License-Identifier: GPL-2.0-or-later
   4
   5 /*
   6  * Userspace RCU library - test program (with batch reclamation)
   7  */
   8
   9 #include <stdio.h>
  10 #include <pthread.h>
  11 #include <stdlib.h>
  12 #include <string.h>
  13 #include <sys/types.h>
  14 #include <sys/wait.h>
  15 #include <unistd.h>
  16 #include <stdio.h>
  17 #include <errno.h>
  18
  19 #include <urcu/arch.h>
  20 #include <urcu/assert.h>
  21 #include <urcu/tls-compat.h>
  22 #include "thread-id.h"
  23 #include "../common/debug-yield.h"
  24
  25 /* hardcoded number of CPUs */
  26 #define NR_CPUS 16384
  27
  28 #ifndef DYNAMIC_LINK_TEST
  29 #define _LGPL_SOURCE
  30 #endif
  31 #include <urcu.h>
  32
  33 struct test_array {
  34         int a;
  35 };
  36
  37 static volatile int test_go, test_stop;
  38
  39 static unsigned long wdelay;
  40
  41 static struct test_array *test_rcu_pointer;
  42
  43 static long reclaim_batch = 1;
  44
  45 struct reclaim_queue {
  46         void **queue;   /* Beginning of queue */
  47         void **head;    /* Insert position */
  48 };
  49
  50 static struct reclaim_queue *pending_reclaims;
  51
  52 static unsigned long duration;
  53
  54 /* read-side C.S. duration, in loops */
  55 static unsigned long rduration;
  56
  57 /* write-side C.S. duration, in loops */
  58 static unsigned long wduration;
  59
  60 static inline void loop_sleep(unsigned long loops)
  61 {
  62         while (loops-- != 0)
  63                 caa_cpu_relax();
  64 }
  65
  66 static int verbose_mode;
  67
  68 #define printf_verbose(fmt, args...)            \
  69         do {                                    \
  70                 if (verbose_mode)               \
  71                         printf(fmt, args);      \
  72         } while (0)
  73
  74 static unsigned int cpu_affinities[NR_CPUS];
  75 static unsigned int next_aff = 0;
  76 static int use_affinity = 0;
  77
  78 pthread_mutex_t affinity_mutex = PTHREAD_MUTEX_INITIALIZER;
  79
  80 static void set_affinity(void)
  81 {
  82 #ifdef HAVE_SCHED_SETAFFINITY
  83         cpu_set_t mask;
  84         int cpu, ret;
  85 #endif /* HAVE_SCHED_SETAFFINITY */
  86
  87         if (!use_affinity)
  88                 return;
  89
  90 #ifdef HAVE_SCHED_SETAFFINITY
  91         ret = pthread_mutex_lock(&affinity_mutex);
  92         if (ret) {
  93                 perror("Error in pthread mutex lock");
  94                 exit(-1);
  95         }
  96         cpu = cpu_affinities[next_aff++];
  97         ret = pthread_mutex_unlock(&affinity_mutex);
  98         if (ret) {
  99                 perror("Error in pthread mutex unlock");
 100                 exit(-1);
 101         }
 102
 103         CPU_ZERO(&mask);
 104         CPU_SET(cpu, &mask);
 105         sched_setaffinity(0, sizeof(mask), &mask);
 106 #endif /* HAVE_SCHED_SETAFFINITY */
 107 }
 108
 109 /*
 110  * returns 0 if test should end.
 111  */
 112 static int test_duration_write(void)
 113 {
 114         return !test_stop;
 115 }
 116
 117 static int test_duration_read(void)
 118 {
 119         return !test_stop;
 120 }
 121
 122 static DEFINE_URCU_TLS(unsigned long long, nr_writes);
 123 static DEFINE_URCU_TLS(unsigned long long, nr_reads);
 124
 125 static
 126 unsigned long long __attribute__((aligned(CAA_CACHE_LINE_SIZE))) *tot_nr_writes;
 127
 128 static unsigned int nr_readers;
 129 static unsigned int nr_writers;
 130
 131 pthread_mutex_t rcu_copy_mutex = PTHREAD_MUTEX_INITIALIZER;
 132
 133 static
 134 void *thr_reader(void *_count)
 135 {
 136         unsigned long long *count = _count;
 137         struct test_array *local_ptr;
 138
 139         printf_verbose("thread_begin %s, tid %lu\n",
 140                         "reader", urcu_get_thread_id());
 141
 142         set_affinity();
 143
 144         rcu_register_thread();
 145
 146         while (!test_go)
 147         {
 148         }
 149         cmm_smp_mb();
 150
 151         for (;;) {
 152                 rcu_read_lock();
 153                 local_ptr = rcu_dereference(test_rcu_pointer);
 154                 rcu_debug_yield_read();
 155                 if (local_ptr)
 156                         urcu_posix_assert(local_ptr->a == 8);
 157                 if (caa_unlikely(rduration))
 158                         loop_sleep(rduration);
 159                 rcu_read_unlock();
 160                 URCU_TLS(nr_reads)++;
 161                 if (caa_unlikely(!test_duration_read()))
 162                         break;
 163         }
 164
 165         rcu_unregister_thread();
 166
 167         *count = URCU_TLS(nr_reads);
 168         printf_verbose("thread_end %s, tid %lu\n",
 169                         "reader", urcu_get_thread_id());
 170         return ((void*)1);
 171
 172 }
 173
 174 static void rcu_gc_clear_queue(unsigned long wtidx)
 175 {
 176         void **p;
 177
 178         /* Wait for Q.S and empty queue */
 179         synchronize_rcu();
 180
 181         for (p = pending_reclaims[wtidx].queue;
 182                         p < pending_reclaims[wtidx].head; p++) {
 183                 /* poison */
 184                 if (*p)
 185                         ((struct test_array *)*p)->a = 0;
 186                 free(*p);
 187         }
 188         pending_reclaims[wtidx].head = pending_reclaims[wtidx].queue;
 189 }
 190
 191 /* Using per-thread queue */
 192 static void rcu_gc_reclaim(unsigned long wtidx, void *old)
 193 {
 194         /* Queue pointer */
 195         *pending_reclaims[wtidx].head = old;
 196         pending_reclaims[wtidx].head++;
 197
 198         if (caa_likely(pending_reclaims[wtidx].head - pending_reclaims[wtidx].queue
 199                         < reclaim_batch))
 200                 return;
 201
 202         rcu_gc_clear_queue(wtidx);
 203 }
 204
 205 static
 206 void *thr_writer(void *data)
 207 {
 208         unsigned long wtidx = (unsigned long)data;
 209 #ifdef TEST_LOCAL_GC
 210         struct test_array *old = NULL;
 211 #else
 212         struct test_array *new, *old;
 213 #endif
 214
 215         printf_verbose("thread_begin %s, tid %lu\n",
 216                         "writer", urcu_get_thread_id());
 217
 218         set_affinity();
 219
 220         while (!test_go)
 221         {
 222         }
 223         cmm_smp_mb();
 224
 225         for (;;) {
 226 #ifndef TEST_LOCAL_GC
 227                 new = malloc(sizeof(*new));
 228                 new->a = 8;
 229                 old = rcu_xchg_pointer(&test_rcu_pointer, new);
 230 #endif
 231                 if (caa_unlikely(wduration))
 232                         loop_sleep(wduration);
 233                 rcu_gc_reclaim(wtidx, old);
 234                 URCU_TLS(nr_writes)++;
 235                 if (caa_unlikely(!test_duration_write()))
 236                         break;
 237                 if (caa_unlikely(wdelay))
 238                         loop_sleep(wdelay);
 239         }
 240
 241         printf_verbose("thread_end %s, tid %lu\n",
 242                         "writer", urcu_get_thread_id());
 243         tot_nr_writes[wtidx] = URCU_TLS(nr_writes);
 244         return ((void*)2);
 245 }
 246
 247 static
 248 void show_usage(char **argv)
 249 {
 250         printf("Usage : %s nr_readers nr_writers duration (s) <OPTIONS>\n",
 251                 argv[0]);
 252         printf("OPTIONS:\n");
 253         printf("        [-r] [-w] (yield reader and/or writer)\n");
 254         printf("        [-d delay] (writer period (us))\n");
 255         printf("        [-c duration] (reader C.S. duration (in loops))\n");
 256         printf("        [-e duration] (writer C.S. duration (in loops))\n");
 257         printf("        [-v] (verbose output)\n");
 258         printf("        [-a cpu#] [-a cpu#]... (affinity)\n");
 259         printf("\n");
 260 }
 261
 262 int main(int argc, char **argv)
 263 {
 264         int err;
 265         pthread_t *tid_reader, *tid_writer;
 266         void *tret;
 267         unsigned long long *count_reader;
 268         unsigned long long tot_reads = 0, tot_writes = 0;
 269         int i, a;
 270         unsigned int i_thr;
 271
 272         if (argc < 4) {
 273                 show_usage(argv);
 274                 return -1;
 275         }
 276
 277         err = sscanf(argv[1], "%u", &nr_readers);
 278         if (err != 1) {
 279                 show_usage(argv);
 280                 return -1;
 281         }
 282
 283         err = sscanf(argv[2], "%u", &nr_writers);
 284         if (err != 1) {
 285                 show_usage(argv);
 286                 return -1;
 287         }
 288
 289         err = sscanf(argv[3], "%lu", &duration);
 290         if (err != 1) {
 291                 show_usage(argv);
 292                 return -1;
 293         }
 294
 295         for (i = 4; i < argc; i++) {
 296                 if (argv[i][0] != '-')
 297                         continue;
 298                 switch (argv[i][1]) {
 299                 case 'r':
 300                         rcu_debug_yield_enable(RCU_YIELD_READ);
 301                         break;
 302                 case 'w':
 303                         rcu_debug_yield_enable(RCU_YIELD_WRITE);
 304                         break;
 305                 case 'a':
 306                         if (argc < i + 2) {
 307                                 show_usage(argv);
 308                                 return -1;
 309                         }
 310                         a = atoi(argv[++i]);
 311                         cpu_affinities[next_aff++] = a;
 312                         use_affinity = 1;
 313                         printf_verbose("Adding CPU %d affinity\n", a);
 314                         break;
 315                 case 'b':
 316                         if (argc < i + 2) {
 317                                 show_usage(argv);
 318                                 return -1;
 319                         }
 320                         reclaim_batch = atol(argv[++i]);
 321                         break;
 322                 case 'c':
 323                         if (argc < i + 2) {
 324                                 show_usage(argv);
 325                                 return -1;
 326                         }
 327                         rduration = atol(argv[++i]);
 328                         break;
 329                 case 'd':
 330                         if (argc < i + 2) {
 331                                 show_usage(argv);
 332                                 return -1;
 333                         }
 334                         wdelay = atol(argv[++i]);
 335                         break;
 336                 case 'e':
 337                         if (argc < i + 2) {
 338                                 show_usage(argv);
 339                                 return -1;
 340                         }
 341                         wduration = atol(argv[++i]);
 342                         break;
 343                 case 'v':
 344                         verbose_mode = 1;
 345                         break;
 346                 }
 347         }
 348
 349         printf_verbose("running test for %lu seconds, %u readers, %u writers.\n",
 350                 duration, nr_readers, nr_writers);
 351         printf_verbose("Writer delay : %lu loops.\n", wdelay);
 352         printf_verbose("Reader duration : %lu loops.\n", rduration);
 353         printf_verbose("thread %-6s, tid %lu\n",
 354                         "main", urcu_get_thread_id());
 355
 356         tid_reader = calloc(nr_readers, sizeof(*tid_reader));
 357         tid_writer = calloc(nr_writers, sizeof(*tid_writer));
 358         count_reader = calloc(nr_readers, sizeof(*count_reader));
 359         tot_nr_writes = calloc(nr_writers, sizeof(*tot_nr_writes));
 360         pending_reclaims = calloc(nr_writers, sizeof(*pending_reclaims));
 361
 362         if (reclaim_batch * sizeof(*pending_reclaims[0].queue)
 363                         < CAA_CACHE_LINE_SIZE)
 364                 for (i_thr = 0; i_thr < nr_writers; i_thr++)
 365                         pending_reclaims[i_thr].queue = calloc(1, CAA_CACHE_LINE_SIZE);
 366         else
 367                 for (i_thr = 0; i_thr < nr_writers; i_thr++)
 368                         pending_reclaims[i_thr].queue = calloc(reclaim_batch,
 369                                         sizeof(*pending_reclaims[i_thr].queue));
 370         for (i_thr = 0; i_thr < nr_writers; i_thr++)
 371                 pending_reclaims[i_thr].head = pending_reclaims[i_thr].queue;
 372
 373         next_aff = 0;
 374
 375         for (i_thr = 0; i_thr < nr_readers; i_thr++) {
 376                 err = pthread_create(&tid_reader[i_thr], NULL, thr_reader,
 377                                      &count_reader[i_thr]);
 378                 if (err != 0)
 379                         exit(1);
 380         }
 381         for (i_thr = 0; i_thr < nr_writers; i_thr++) {
 382                 err = pthread_create(&tid_writer[i_thr], NULL, thr_writer,
 383                                      (void *)(long)i_thr);
 384                 if (err != 0)
 385                         exit(1);
 386         }
 387
 388         cmm_smp_mb();
 389
 390         test_go = 1;
 391
 392         sleep(duration);
 393
 394         test_stop = 1;
 395
 396         for (i_thr = 0; i_thr < nr_readers; i_thr++) {
 397                 err = pthread_join(tid_reader[i_thr], &tret);
 398                 if (err != 0)
 399                         exit(1);
 400                 tot_reads += count_reader[i_thr];
 401         }
 402         for (i_thr = 0; i_thr < nr_writers; i_thr++) {
 403                 err = pthread_join(tid_writer[i_thr], &tret);
 404                 if (err != 0)
 405                         exit(1);
 406                 tot_writes += tot_nr_writes[i_thr];
 407                 rcu_gc_clear_queue(i_thr);
 408         }
 409
 410         printf_verbose("total number of reads : %llu, writes %llu\n", tot_reads,
 411                tot_writes);
 412         printf("SUMMARY %-25s testdur %4lu nr_readers %3u rdur %6lu wdur %6lu "
 413                 "nr_writers %3u "
 414                 "wdelay %6lu nr_reads %12llu nr_writes %12llu nr_ops %12llu "
 415                 "batch %ld\n",
 416                 argv[0], duration, nr_readers, rduration, wduration,
 417                 nr_writers, wdelay, tot_reads, tot_writes,
 418                 tot_reads + tot_writes, reclaim_batch);
 419
 420         free(tid_reader);
 421         free(tid_writer);
 422         free(count_reader);
 423         free(tot_nr_writes);
 424
 425         for (i_thr = 0; i_thr < nr_writers; i_thr++)
 426                 free(pending_reclaims[i_thr].queue);
 427         free(pending_reclaims);
 428
 429         return 0;
 430 }