| 1 | // SPDX-License-Identifier: GPL-3.0-or-later |
| 2 | |
| 3 | #include "benchmark-rw.h" |
| 4 | |
| 5 | #define MAX_THREADS 64 |
| 6 | #define TEST_DURATION_SEC 1 |
| 7 | #define STOP_SIGNAL UINT64_MAX |
| 8 | #define MAX_CONFIGS 10 |
| 9 | |
| 10 | // Structure to store summary statistics |
| 11 | typedef struct { |
| 12 | double ops_per_sec[2][MAX_CONFIGS]; // [lock_type][config_index], Total ops/sec |
| 13 | double reader_ops_per_sec[2][MAX_CONFIGS]; // [lock_type][config_index], Reader ops/sec |
| 14 | double writer_ops_per_sec[2][MAX_CONFIGS]; // [lock_type][config_index], Writer ops/sec |
| 15 | int readers[MAX_CONFIGS]; // Number of readers for each config |
| 16 | int writers[MAX_CONFIGS]; // Number of writers for each config |
| 17 | int config_count; // Number of configurations tested |
| 18 | } summary_stats_t; |
| 19 | |
| 20 | typedef struct { |
| 21 | // Protected state to validate reader/writer mutual exclusion |
| 22 | volatile int readers; // Number of active readers |
| 23 | volatile int writers; // Number of active writers |
| 24 | volatile uint64_t violations; // Counter for reader/writer violations |
| 25 | |
| 26 | // Protected counter for actual work |
| 27 | uint64_t counter; |
| 28 | |
| 29 | // Statistics per thread |
| 30 | struct { |
| 31 | uint64_t operations; // Number of read/write operations |
| 32 | usec_t test_time; // Time spent in test |
| 33 | volatile int ready; // Thread completed flag |
| 34 | } stats[MAX_THREADS]; |
| 35 | |
| 36 | // Per-thread control |
| 37 | struct { |
| 38 | netdata_cond_t cond; // Thread start condition |
| 39 | netdata_mutex_t cond_mutex; // Mutex for condition |
| 40 | uint64_t run_flag; // Thread run control |
| 41 | } thread_controls[MAX_THREADS]; |
| 42 | } rwlock_control_t; |
| 43 | |
| 44 | typedef enum { |
| 45 | THREAD_READER, |
| 46 | THREAD_WRITER |
| 47 | } thread_type_t; |
| 48 | |
| 49 | typedef struct { |
| 50 | int thread_id; |
| 51 | thread_type_t type; |
| 52 | void *lock; // Points to either netdata_rwlock_t or RW_SPINLOCK |
| 53 | bool is_spinlock; // true for RW_SPINLOCK, false for netdata_rwlock_t |
| 54 | rwlock_control_t *control; |
| 55 | ND_THREAD *thread; |
| 56 | } thread_context_t; |
| 57 | |
| 58 | static inline void verify_no_violations(rwlock_control_t *control) { |
| 59 | if(__atomic_load_n(&control->violations, __ATOMIC_RELAXED) > 0) { |
| 60 | fprintf(stderr, "\nFATAL ERROR: Detected %"PRIu64" read/write violations!\n" |
| 61 | "This indicates readers and writers were concurrently inside the lock.\n", |
| 62 | control->violations); |
| 63 | exit(1); |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | static inline void check_access_safety(rwlock_control_t *control, thread_type_t type) { |
| 68 | if(type == THREAD_READER) { |
| 69 | // Reader entering critical section |
| 70 | __atomic_add_fetch(&control->readers, 1, __ATOMIC_RELAXED); |
| 71 | |
| 72 | // Check if we have any writers - this would be a violation |
| 73 | if(__atomic_load_n(&control->writers, __ATOMIC_RELAXED) > 0) { |
| 74 | __atomic_add_fetch(&control->violations, 1, __ATOMIC_RELAXED); |
| 75 | } |
| 76 | } |
| 77 | else { |
| 78 | // Writer entering critical section |
| 79 | int writers = __atomic_add_fetch(&control->writers, 1, __ATOMIC_RELAXED); |
| 80 | |
| 81 | // Check for other writers - violation! |
| 82 | if(writers > 1) { |
| 83 | __atomic_add_fetch(&control->violations, 1, __ATOMIC_RELAXED); |
| 84 | } |
| 85 | |
| 86 | // Check if we have any readers - this would be a violation |
| 87 | if(__atomic_load_n(&control->readers, __ATOMIC_RELAXED) > 0) { |
| 88 | __atomic_add_fetch(&control->violations, 1, __ATOMIC_RELAXED); |
| 89 | } |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | static void release_access(rwlock_control_t *control, thread_type_t type) { |
| 94 | if(type == THREAD_READER) { |
| 95 | __atomic_sub_fetch(&control->readers, 1, __ATOMIC_RELAXED); |
| 96 | } |
| 97 | else { |
| 98 | __atomic_sub_fetch(&control->writers, 1, __ATOMIC_RELAXED); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | static void wait_for_start(netdata_cond_t *cond, netdata_mutex_t *mutex, uint64_t *flag) { |
| 103 | netdata_mutex_lock(mutex); |
| 104 | while (*flag == 0) |
| 105 | netdata_cond_wait(cond, mutex); |
| 106 | netdata_mutex_unlock(mutex); |
| 107 | } |
| 108 | |
| 109 | static void benchmark_thread(void *arg) { |
| 110 | thread_context_t *ctx = (thread_context_t *)arg; |
| 111 | rwlock_control_t *control = ctx->control; |
| 112 | |
| 113 | while(1) { |
| 114 | // Wait for start signal |
| 115 | wait_for_start(&control->thread_controls[ctx->thread_id].cond, |
| 116 | &control->thread_controls[ctx->thread_id].cond_mutex, |
| 117 | &control->thread_controls[ctx->thread_id].run_flag); |
| 118 | |
| 119 | if (control->thread_controls[ctx->thread_id].run_flag == STOP_SIGNAL) |
| 120 | break; |
| 121 | |
| 122 | usec_t start = now_monotonic_high_precision_usec(); |
| 123 | uint64_t operations = 0; |
| 124 | |
| 125 | while (control->thread_controls[ctx->thread_id].run_flag) { |
| 126 | if(ctx->is_spinlock) { |
| 127 | RW_SPINLOCK *spinlock = ctx->lock; |
| 128 | if(ctx->type == THREAD_READER) { |
| 129 | rw_spinlock_read_lock(spinlock); |
| 130 | check_access_safety(control, THREAD_READER); |
| 131 | control->counter++; // Just to do some work |
| 132 | release_access(control, THREAD_READER); |
| 133 | rw_spinlock_read_unlock(spinlock); |
| 134 | } |
| 135 | else { |
| 136 | rw_spinlock_write_lock(spinlock); |
| 137 | check_access_safety(control, THREAD_WRITER); |
| 138 | control->counter++; |
| 139 | release_access(control, THREAD_WRITER); |
| 140 | rw_spinlock_write_unlock(spinlock); |
| 141 | } |
| 142 | } |
| 143 | else { |
| 144 | netdata_rwlock_t *rwlock = ctx->lock; |
| 145 | if(ctx->type == THREAD_READER) { |
| 146 | netdata_rwlock_rdlock(rwlock); |
| 147 | check_access_safety(control, THREAD_READER); |
| 148 | control->counter++; |
| 149 | release_access(control, THREAD_READER); |
| 150 | netdata_rwlock_rdunlock(rwlock); |
| 151 | } |
| 152 | else { |
| 153 | netdata_rwlock_wrlock(rwlock); |
| 154 | check_access_safety(control, THREAD_WRITER); |
| 155 | control->counter++; |
| 156 | release_access(control, THREAD_WRITER); |
| 157 | netdata_rwlock_rdunlock(rwlock); |
| 158 | } |
| 159 | } |
| 160 | operations++; |
| 161 | } |
| 162 | |
| 163 | // Store results |
| 164 | usec_t test_time = now_monotonic_high_precision_usec() - start; |
| 165 | __atomic_store_n(&control->stats[ctx->thread_id].test_time, test_time, __ATOMIC_RELEASE); |
| 166 | __atomic_store_n(&control->stats[ctx->thread_id].operations, operations, __ATOMIC_RELEASE); |
| 167 | __atomic_store_n(&control->stats[ctx->thread_id].ready, 1, __ATOMIC_RELEASE); |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | static void print_summary(const summary_stats_t *summary) { |
| 172 | fprintf(stderr, "\n=== Performance Summary (Million operations/sec) ===\n\n"); |
| 173 | fprintf(stderr, "%-16s %-8s %-8s %-16s %-16s\n", |
| 174 | "Lock Type", "Readers", "Writers", "Reader Ops/s", "Writer Ops/s"); |
| 175 | fprintf(stderr, "----------------------------------------------------------------------\n"); |
| 176 | |
| 177 | const char *lock_names[] = {"netdata_rwlock", "rw_spinlock"}; |
| 178 | |
| 179 | for (int config = 0; config < summary->config_count; config++) { |
| 180 | for (int lock_type = 0; lock_type < 2; lock_type++) { |
| 181 | // double total_ops = summary->ops_per_sec[lock_type][config]; |
| 182 | int readers = summary->readers[config]; |
| 183 | int writers = summary->writers[config]; |
| 184 | |
| 185 | // Get the actual reader and writer operations |
| 186 | double reader_ops = readers > 0 ? summary->reader_ops_per_sec[lock_type][config] : 0; |
| 187 | double writer_ops = writers > 0 ? summary->writer_ops_per_sec[lock_type][config] : 0; |
| 188 | |
| 189 | fprintf(stderr, "%-16s %-8d %-8d %-16.2f %-16.2f\n", |
| 190 | lock_names[lock_type], |
| 191 | readers, |
| 192 | writers, |
| 193 | reader_ops / 1000000.0, |
| 194 | writer_ops / 1000000.0); |
| 195 | } |
| 196 | // Add a separator between configurations |
| 197 | if (config < summary->config_count - 1) |
| 198 | fprintf(stderr, "----------------------------------------------------------------------\n"); |
| 199 | } |
| 200 | fprintf(stderr, "\n"); |
| 201 | } |
| 202 | |
| 203 | static void print_thread_stats(const char *test_name, int readers, int writers, |
| 204 | thread_context_t *contexts, rwlock_control_t *control, |
| 205 | summary_stats_t *summary, int config_idx, int lock_type) { |
| 206 | fprintf(stderr, "\n%-20s (readers: %d, writers: %d)\n", test_name, readers, writers); |
| 207 | fprintf(stderr, "%4s %8s %12s %12s %12s\n", |
| 208 | "THR", "TYPE", "OPS", "OPS/SEC", "TIME (ms)"); |
| 209 | |
| 210 | uint64_t total_ops = 0; |
| 211 | double total_ops_per_sec = 0; |
| 212 | double reader_ops_per_sec = 0; |
| 213 | double writer_ops_per_sec = 0; |
| 214 | |
| 215 | for(int i = 0; i < readers + writers; i++) { |
| 216 | uint64_t ops = __atomic_load_n(&control->stats[i].operations, __ATOMIC_RELAXED); |
| 217 | usec_t time = __atomic_load_n(&control->stats[i].test_time, __ATOMIC_RELAXED); |
| 218 | double ops_per_sec = (double)ops * USEC_PER_SEC / time; |
| 219 | |
| 220 | fprintf(stderr, "%4d %8s %12"PRIu64" %12.0f %12.2f\n", |
| 221 | i, |
| 222 | contexts[i].type == THREAD_READER ? "READER" : "WRITER", |
| 223 | ops, |
| 224 | ops_per_sec, |
| 225 | (double)time / 1000.0); |
| 226 | |
| 227 | total_ops += ops; |
| 228 | total_ops_per_sec += ops_per_sec; |
| 229 | |
| 230 | if (contexts[i].type == THREAD_READER) { |
| 231 | reader_ops_per_sec += ops_per_sec; |
| 232 | } else { |
| 233 | writer_ops_per_sec += ops_per_sec; |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | fprintf(stderr, "%4s %8s %12"PRIu64" %12.0f\n", |
| 238 | "TOT", "", total_ops, total_ops_per_sec); |
| 239 | |
| 240 | // Store in summary |
| 241 | summary->ops_per_sec[lock_type][config_idx] = total_ops_per_sec; |
| 242 | summary->reader_ops_per_sec[lock_type][config_idx] = reader_ops_per_sec; |
| 243 | summary->writer_ops_per_sec[lock_type][config_idx] = writer_ops_per_sec; |
| 244 | summary->readers[config_idx] = readers; |
| 245 | summary->writers[config_idx] = writers; |
| 246 | |
| 247 | verify_no_violations(control); |
| 248 | } |
| 249 | |
| 250 | |
| 251 | static void run_test(const char *name, int readers, int writers, |
| 252 | thread_context_t *contexts, rwlock_control_t *control, |
| 253 | summary_stats_t *summary, int config_idx, int lock_type) { |
| 254 | fprintf(stderr, "\nRunning test: %s with %d readers and %d writers...\n", |
| 255 | name, readers, writers); |
| 256 | |
| 257 | // Reset all stats and control |
| 258 | memset(&control->stats, 0, sizeof(control->stats)); |
| 259 | control->counter = 0; |
| 260 | control->readers = 0; |
| 261 | control->writers = 0; |
| 262 | control->violations = 0; |
| 263 | |
| 264 | int total_threads = readers + writers; |
| 265 | |
| 266 | // Signal threads to start |
| 267 | for(int i = 0; i < total_threads; i++) { |
| 268 | netdata_mutex_lock(&control->thread_controls[i].cond_mutex); |
| 269 | control->thread_controls[i].run_flag = 1; |
| 270 | netdata_cond_signal(&control->thread_controls[i].cond); |
| 271 | netdata_mutex_unlock(&control->thread_controls[i].cond_mutex); |
| 272 | } |
| 273 | |
| 274 | // Wait for test duration |
| 275 | sleep_usec(TEST_DURATION_SEC * USEC_PER_SEC); |
| 276 | |
| 277 | // Signal threads to stop |
| 278 | for(int i = 0; i < total_threads; i++) { |
| 279 | __atomic_store_n(&control->thread_controls[i].run_flag, 0, __ATOMIC_RELEASE); |
| 280 | } |
| 281 | |
| 282 | // Wait for threads to report results |
| 283 | for(int i = 0; i < total_threads; i++) { |
| 284 | while(!__atomic_load_n(&control->stats[i].ready, __ATOMIC_ACQUIRE)) |
| 285 | sleep_usec(10); |
| 286 | } |
| 287 | |
| 288 | print_thread_stats(name, readers, writers, contexts, control, summary, config_idx, lock_type); |
| 289 | } |
| 290 | |
| 291 | int rwlocks_stress_test(void) { |
| 292 | netdata_rwlock_t netdata_rwlock; |
| 293 | netdata_rwlock_init(&netdata_rwlock); |
| 294 | |
| 295 | RW_SPINLOCK rw_spinlock = RW_SPINLOCK_INITIALIZER; |
| 296 | summary_stats_t summary = {0}; |
| 297 | |
| 298 | // Initialize control structures |
| 299 | rwlock_control_t netdata_control = { 0 }; |
| 300 | rwlock_control_t spinlock_control = { 0 }; |
| 301 | |
| 302 | // Initialize per-thread controls for both locks |
| 303 | for(int i = 0; i < MAX_THREADS; i++) { |
| 304 | netdata_cond_init(&netdata_control.thread_controls[i].cond); |
| 305 | netdata_mutex_init(&netdata_control.thread_controls[i].cond_mutex); |
| 306 | netdata_control.thread_controls[i].run_flag = 0; |
| 307 | |
| 308 | netdata_cond_init(&spinlock_control.thread_controls[i].cond); |
| 309 | netdata_mutex_init(&spinlock_control.thread_controls[i].cond_mutex); |
| 310 | |
| 311 | spinlock_control.thread_controls[i].run_flag = 0; |
| 312 | } |
| 313 | |
| 314 | // Create thread contexts |
| 315 | thread_context_t netdata_contexts[MAX_THREADS]; |
| 316 | thread_context_t spinlock_contexts[MAX_THREADS]; |
| 317 | |
| 318 | fprintf(stderr, "\nStarting RW locks benchmark...\n"); |
| 319 | |
| 320 | // Test configurations: [readers, writers] |
| 321 | int configs[][2] = { |
| 322 | {1, 0}, // Single reader |
| 323 | {0, 1}, // Single writer |
| 324 | {1, 1}, // One reader + one writer |
| 325 | {2, 1}, // Two readers + one writer |
| 326 | {1, 2}, // One reader + two writers |
| 327 | {2, 2}, // Two readers + two writers |
| 328 | {4, 1}, // Four readers + one writer |
| 329 | {1, 4}, // One reader + four writers |
| 330 | {4, 4}, // Four readers + four writers |
| 331 | }; |
| 332 | |
| 333 | const int num_configs = sizeof(configs) / sizeof(configs[0]); |
| 334 | summary.config_count = num_configs; |
| 335 | |
| 336 | // Create all threads |
| 337 | for(int i = 0; i < MAX_THREADS; i++) { |
| 338 | char thr_name[32]; |
| 339 | |
| 340 | // Initialize pthread contexts |
| 341 | netdata_contexts[i] = (thread_context_t){ |
| 342 | .thread_id = i, |
| 343 | .type = i % 2 == 0 ? THREAD_READER :THREAD_WRITER, |
| 344 | .lock = &netdata_rwlock, |
| 345 | .is_spinlock = false, |
| 346 | .control = &netdata_control |
| 347 | }; |
| 348 | |
| 349 | snprintf(thr_name, sizeof(thr_name), "netdata_rw%d", i); |
| 350 | netdata_contexts[i].thread = |
| 351 | nd_thread_create(thr_name, NETDATA_THREAD_OPTION_DONT_LOG, benchmark_thread, &netdata_contexts[i]); |
| 352 | |
| 353 | // Initialize spinlock contexts |
| 354 | spinlock_contexts[i] = (thread_context_t){ |
| 355 | .thread_id = i, |
| 356 | .type = i % 2 == 0 ? THREAD_READER : THREAD_WRITER, |
| 357 | .lock = &rw_spinlock, |
| 358 | .is_spinlock = true, |
| 359 | .control = &spinlock_control |
| 360 | }; |
| 361 | |
| 362 | snprintf(thr_name, sizeof(thr_name), "spin_rw%d", i); |
| 363 | spinlock_contexts[i].thread = |
| 364 | nd_thread_create(thr_name, NETDATA_THREAD_OPTION_DONT_LOG, benchmark_thread, &spinlock_contexts[i]); |
| 365 | } |
| 366 | |
| 367 | // Run all configurations |
| 368 | for(int i = 0; i < num_configs; i++) { |
| 369 | int readers = configs[i][0]; |
| 370 | int writers = configs[i][1]; |
| 371 | |
| 372 | // Create all threads |
| 373 | int thread_idx = 0; |
| 374 | |
| 375 | // First assign reader threads |
| 376 | for(int r = 0; r < readers; r++) { |
| 377 | netdata_contexts[thread_idx].type = THREAD_READER; |
| 378 | spinlock_contexts[thread_idx].type = THREAD_READER; |
| 379 | thread_idx++; |
| 380 | } |
| 381 | |
| 382 | // Then assign writer threads |
| 383 | for(int w = 0; w < writers; w++) { |
| 384 | netdata_contexts[thread_idx].type = THREAD_WRITER; |
| 385 | spinlock_contexts[thread_idx].type = THREAD_WRITER; |
| 386 | thread_idx++; |
| 387 | } |
| 388 | |
| 389 | char test_name[64]; |
| 390 | snprintf(test_name, sizeof(test_name), "netdata_rwlock %dR/%dW", readers, writers); |
| 391 | run_test(test_name, readers, writers, netdata_contexts, &netdata_control, &summary, i, 0); |
| 392 | |
| 393 | snprintf(test_name, sizeof(test_name), "rw_spinlock %dR/%dW", readers, writers); |
| 394 | run_test(test_name, readers, writers, spinlock_contexts, &spinlock_control, &summary, i, 1); |
| 395 | } |
| 396 | |
| 397 | // Print the summary table |
| 398 | print_summary(&summary); |
| 399 | |
| 400 | // Stop all threads |
| 401 | fprintf(stderr, "\nStopping threads...\n"); |
| 402 | for(int i = 0; i < MAX_THREADS; i++) { |
| 403 | // Signal pthread threads |
| 404 | netdata_mutex_lock(&netdata_control.thread_controls[i].cond_mutex); |
| 405 | netdata_control.thread_controls[i].run_flag = STOP_SIGNAL; |
| 406 | netdata_cond_signal(&netdata_control.thread_controls[i].cond); |
| 407 | netdata_mutex_unlock(&netdata_control.thread_controls[i].cond_mutex); |
| 408 | |
| 409 | // Signal spinlock threads |
| 410 | netdata_mutex_lock(&spinlock_control.thread_controls[i].cond_mutex); |
| 411 | spinlock_control.thread_controls[i].run_flag = STOP_SIGNAL; |
| 412 | netdata_cond_signal(&spinlock_control.thread_controls[i].cond); |
| 413 | netdata_mutex_unlock(&spinlock_control.thread_controls[i].cond_mutex); |
| 414 | } |
| 415 | |
| 416 | // Join all threads |
| 417 | fprintf(stderr, "\nWaiting for threads to exit...\n"); |
| 418 | for(int i = 0; i < MAX_THREADS; i++) { |
| 419 | nd_thread_join(netdata_contexts[i].thread); |
| 420 | nd_thread_join(spinlock_contexts[i].thread); |
| 421 | } |
| 422 | |
| 423 | // Cleanup condition variables and mutexes |
| 424 | for(int i = 0; i < MAX_THREADS; i++) { |
| 425 | netdata_cond_destroy(&netdata_control.thread_controls[i].cond); |
| 426 | netdata_mutex_destroy(&netdata_control.thread_controls[i].cond_mutex); |
| 427 | netdata_cond_destroy(&spinlock_control.thread_controls[i].cond); |
| 428 | netdata_mutex_destroy(&spinlock_control.thread_controls[i].cond_mutex); |
| 429 | } |
| 430 | |
| 431 | netdata_rwlock_destroy(&netdata_rwlock); |
| 432 | |
| 433 | return 0; |
| 434 | } |