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