| 1 | // SPDX-License-Identifier: GPL-3.0-or-later |
| 2 | |
| 3 | #include "rrdset-slots.h" |
| 4 | #include "rrdset-pluginsd-array.h" |
| 5 | |
| 6 | void rrdset_stream_send_chart_slot_assign(RRDSET *st) { |
| 7 | RRDHOST *host = st->rrdhost; |
| 8 | spinlock_lock(&host->stream.snd.pluginsd_chart_slots.available.spinlock); |
| 9 | |
| 10 | if(host->stream.snd.pluginsd_chart_slots.available.used > 0) |
| 11 | st->stream.snd.chart_slot = |
| 12 | host->stream.snd.pluginsd_chart_slots.available.array[--host->stream.snd.pluginsd_chart_slots.available.used]; |
| 13 | else |
| 14 | st->stream.snd.chart_slot = ++host->stream.snd.pluginsd_chart_slots.last_used; |
| 15 | |
| 16 | spinlock_unlock(&host->stream.snd.pluginsd_chart_slots.available.spinlock); |
| 17 | } |
| 18 | |
| 19 | void rrdset_stream_send_chart_slot_release(RRDSET *st) { |
| 20 | if(!st->stream.snd.chart_slot || st->rrdhost->stream.snd.pluginsd_chart_slots.available.ignore) |
| 21 | return; |
| 22 | |
| 23 | RRDHOST *host = st->rrdhost; |
| 24 | spinlock_lock(&host->stream.snd.pluginsd_chart_slots.available.spinlock); |
| 25 | |
| 26 | if(host->stream.snd.pluginsd_chart_slots.available.used >= host->stream.snd.pluginsd_chart_slots.available.size) { |
| 27 | uint32_t old_slots = host->stream.snd.pluginsd_chart_slots.available.size; |
| 28 | uint32_t new_slots = (old_slots > 0) ? (old_slots * 2) : 1024; |
| 29 | |
| 30 | host->stream.snd.pluginsd_chart_slots.available.array = |
| 31 | reallocz(host->stream.snd.pluginsd_chart_slots.available.array, new_slots * sizeof(uint32_t)); |
| 32 | |
| 33 | host->stream.snd.pluginsd_chart_slots.available.size = new_slots; |
| 34 | |
| 35 | rrd_slot_memory_added((new_slots - old_slots) * sizeof(uint32_t)); |
| 36 | } |
| 37 | |
| 38 | host->stream.snd.pluginsd_chart_slots.available.array[host->stream.snd.pluginsd_chart_slots.available.used++] = |
| 39 | st->stream.snd.chart_slot; |
| 40 | |
| 41 | st->stream.snd.chart_slot = 0; |
| 42 | spinlock_unlock(&host->stream.snd.pluginsd_chart_slots.available.spinlock); |
| 43 | } |
| 44 | |
| 45 | // -------------------------------------------------------------------------------------------------------------------- |
| 46 | // Helper function to release RRDDIM_ACQUIRED references in array entries |
| 47 | // This must be called before the final prd_array_release when cleaning up |
| 48 | |
| 49 | static void prd_array_release_entries(PRD_ARRAY *arr) { |
| 50 | if (!arr) |
| 51 | return; |
| 52 | |
| 53 | for (size_t i = 0; i < arr->size; i++) { |
| 54 | rrddim_acquired_release(arr->entries[i].rda); // safe with NULL |
| 55 | arr->entries[i].rda = NULL; |
| 56 | arr->entries[i].rd = NULL; |
| 57 | arr->entries[i].id = NULL; |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | static inline void rrdset_clear_host_chart_slot_mapping(RRDSET *st, int32_t last_slot) { |
| 62 | if(last_slot < 0) |
| 63 | return; |
| 64 | |
| 65 | RRDHOST *host = st->rrdhost; |
| 66 | spinlock_lock(&host->stream.rcv.pluginsd_chart_slots.spinlock); |
| 67 | if((uint32_t)last_slot < host->stream.rcv.pluginsd_chart_slots.size && |
| 68 | host->stream.rcv.pluginsd_chart_slots.array[last_slot] == st) { |
| 69 | host->stream.rcv.pluginsd_chart_slots.array[last_slot] = NULL; |
| 70 | } |
| 71 | spinlock_unlock(&host->stream.rcv.pluginsd_chart_slots.spinlock); |
| 72 | } |
| 73 | |
| 74 | // -------------------------------------------------------------------------------------------------------------------- |
| 75 | // Unslot a chart - releases dimension references but keeps the array for reuse |
| 76 | // This is called when switching charts, marking them obsolete, or during cleanup. |
| 77 | // |
| 78 | // Safe to call from: |
| 79 | // - The collector thread itself (collector_tid == gettid_cached()): uses lock-free access |
| 80 | // - Any thread when the collector is fully stopped (collector_tid == 0): uses refcount |
| 81 | // Skips with a warning if a DIFFERENT thread's collector is active. |
| 82 | |
| 83 | void rrdset_pluginsd_receive_unslot(RRDSET *st) { |
| 84 | if(!st) |
| 85 | return; |
| 86 | |
| 87 | RRDDIM_ACQUIRED **detached_rdas = NULL; |
| 88 | size_t detached_capacity = 0; |
| 89 | size_t detached_entries = 0; |
| 90 | bool we_are_collector = false; |
| 91 | PRD_ARRAY *arr = NULL; |
| 92 | int32_t last_slot = -1; |
| 93 | |
| 94 | while(true) { |
| 95 | spinlock_lock(&st->pluginsd.spinlock); |
| 96 | |
| 97 | // Check collector_tid inside spinlock |
| 98 | pid_t collector_tid = __atomic_load_n(&st->pluginsd.collector_tid, __ATOMIC_ACQUIRE); |
| 99 | we_are_collector = (collector_tid == gettid_cached()); |
| 100 | bool different_collector_active = (collector_tid != 0 && !we_are_collector); |
| 101 | |
| 102 | last_slot = st->pluginsd.last_slot; |
| 103 | |
| 104 | if(different_collector_active) { |
| 105 | // Another thread is the active collector - we cannot safely touch the array. |
| 106 | // Keep pluginsd state unchanged in this path: the active collector may |
| 107 | // still read last_slot / dims_with_slots lock-free. |
| 108 | // Clear only the host slot mapping and bail out. |
| 109 | nd_log_limit_static_global_var(erl, 1, 0); |
| 110 | nd_log_limit(&erl, NDLS_DAEMON, NDLP_WARNING, |
| 111 | "PLUGINSD: rrdset_pluginsd_receive_unslot called while collector (tid %d) is active, skipping", |
| 112 | collector_tid); |
| 113 | |
| 114 | spinlock_unlock(&st->pluginsd.spinlock); |
| 115 | freez(detached_rdas); |
| 116 | rrdset_clear_host_chart_slot_mapping(st, last_slot); |
| 117 | return; |
| 118 | } |
| 119 | |
| 120 | // Either collector_tid == 0 (collector stopped) or collector_tid == our tid |
| 121 | // (we ARE the collector). In both cases, it's safe to detach dimension references. |
| 122 | arr = we_are_collector ? |
| 123 | prd_array_get_unsafe(&st->pluginsd.prd_array) : |
| 124 | prd_array_acquire_locked(&st->pluginsd.prd_array); |
| 125 | |
| 126 | if(arr) { |
| 127 | if(!we_are_collector) { |
| 128 | // Verify no other thread holds an extra reference before clearing entries. |
| 129 | // After acquire_locked, refcount should be 2 (original + ours). |
| 130 | int32_t rc = __atomic_load_n(&arr->refcount, __ATOMIC_ACQUIRE); |
| 131 | internal_fatal(rc != 2, |
| 132 | "PRD_ARRAY: expected refcount 2 after acquire, got %d - concurrent reference leak", rc); |
| 133 | |
| 134 | if(unlikely(rc != 2)) { |
| 135 | // Production guard: skip detachment when another reference is active. |
| 136 | // Clearing entries in this state can race and double-release references. |
| 137 | nd_log_limit_static_global_var(erl_rc, 1, 0); |
| 138 | nd_log_limit(&erl_rc, NDLS_DAEMON, NDLP_WARNING, |
| 139 | "PLUGINSD: unslot skipped for chart with unexpected PRD_ARRAY refcount %d (expected 2)", |
| 140 | rc); |
| 141 | |
| 142 | prd_array_release(arr); |
| 143 | spinlock_unlock(&st->pluginsd.spinlock); |
| 144 | freez(detached_rdas); |
| 145 | rrdset_clear_host_chart_slot_mapping(st, last_slot); |
| 146 | return; |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | detached_entries = arr->size; |
| 151 | if(detached_entries > detached_capacity) { |
| 152 | // Allocate outside the spinlock to avoid allocator latency while other |
| 153 | // threads are spinning on this lock. |
| 154 | if(!we_are_collector) |
| 155 | prd_array_release(arr); |
| 156 | |
| 157 | spinlock_unlock(&st->pluginsd.spinlock); |
| 158 | |
| 159 | freez(detached_rdas); |
| 160 | detached_rdas = callocz(detached_entries, sizeof(*detached_rdas)); |
| 161 | detached_capacity = detached_entries; |
| 162 | continue; |
| 163 | } |
| 164 | |
| 165 | // Detach entries while holding st->pluginsd.spinlock so concurrent unslot/cleanup |
| 166 | // cannot race and release the same RRDDIM_ACQUIRED pointers twice. |
| 167 | for(size_t i = 0; i < detached_entries; i++) { |
| 168 | detached_rdas[i] = arr->entries[i].rda; |
| 169 | arr->entries[i].rda = NULL; |
| 170 | arr->entries[i].rd = NULL; |
| 171 | arr->entries[i].id = NULL; |
| 172 | } |
| 173 | } |
| 174 | else |
| 175 | detached_entries = 0; |
| 176 | |
| 177 | st->pluginsd.last_slot = -1; |
| 178 | st->pluginsd.dims_with_slots = false; |
| 179 | |
| 180 | spinlock_unlock(&st->pluginsd.spinlock); |
| 181 | break; |
| 182 | } |
| 183 | |
| 184 | // Release detached references outside the spinlock. |
| 185 | if(detached_rdas) { |
| 186 | for(size_t i = 0; i < detached_entries; i++) |
| 187 | rrddim_acquired_release(detached_rdas[i]); // safe with NULL |
| 188 | |
| 189 | freez(detached_rdas); |
| 190 | } |
| 191 | |
| 192 | if(arr && !we_are_collector) { |
| 193 | // Release our acquired reference (keeps the struct's reference alive for reuse) |
| 194 | prd_array_release(arr); |
| 195 | } |
| 196 | |
| 197 | rrdset_clear_host_chart_slot_mapping(st, last_slot); |
| 198 | } |
| 199 | |
| 200 | // -------------------------------------------------------------------------------------------------------------------- |
| 201 | // Full cleanup - unslots and frees the array |
| 202 | // This is called during chart finalization or host cleanup |
| 203 | // Thread-safe: uses spinlock for cleanup coordination and reference counting for array lifetime |
| 204 | |
| 205 | void rrdset_pluginsd_receive_unslot_and_cleanup(RRDSET *st) { |
| 206 | if(!st) |
| 207 | return; |
| 208 | |
| 209 | spinlock_lock(&st->pluginsd.spinlock); |
| 210 | |
| 211 | // Check if collector is still active. |
| 212 | pid_t collector_tid = __atomic_load_n(&st->pluginsd.collector_tid, __ATOMIC_ACQUIRE); |
| 213 | pid_t current_tid = gettid_cached(); |
| 214 | if(collector_tid != 0) { |
| 215 | if(collector_tid != current_tid) { |
| 216 | // A different thread owns this chart as collector. |
| 217 | // We must not release PRD dimension references while that thread may |
| 218 | // still be using cached rd pointers from the array. |
| 219 | // RRDSET_FLAG_COLLECTION_FINISHED is not a reliable signal here: |
| 220 | // it can be set by the cleanup caller (service thread) rather than |
| 221 | // by the collector itself, creating a race where we free dimensions |
| 222 | // that the collector is actively dereferencing. |
| 223 | // Legitimate teardown clears collector_tid via |
| 224 | // rrdhost_pluginsd_receive_chart_slots_free() after the receiver |
| 225 | // is fully stopped, before invoking cleanup. |
| 226 | nd_log_limit_static_global_var(erl, 1, 0); |
| 227 | nd_log_limit(&erl, NDLS_DAEMON, NDLP_WARNING, |
| 228 | "PLUGINSD: attempted cleanup while collector (tid %d) is still active on chart, skipping", |
| 229 | collector_tid); |
| 230 | spinlock_unlock(&st->pluginsd.spinlock); |
| 231 | return; |
| 232 | } |
| 233 | |
| 234 | // We ARE the collector thread - safe to proceed with cleanup. |
| 235 | // This should not normally happen; keep it explicit so we don't |
| 236 | // mask it as a stale tid case. |
| 237 | #ifdef NETDATA_INTERNAL_CHECKS |
| 238 | internal_fatal(true, |
| 239 | "PRD_ARRAY: cleanup called from collector thread (tid %d) - lifecycle violation", |
| 240 | collector_tid); |
| 241 | #endif |
| 242 | |
| 243 | nd_log_limit_static_global_var(erl_collector, 1, 0); |
| 244 | nd_log_limit(&erl_collector, NDLS_DAEMON, NDLP_WARNING, |
| 245 | "PLUGINSD: cleanup called from collector thread (tid %d), forcing collector_tid=0", |
| 246 | collector_tid); |
| 247 | |
| 248 | __atomic_store_n(&st->pluginsd.collector_tid, 0, __ATOMIC_RELEASE); |
| 249 | } |
| 250 | |
| 251 | // Replace the array with NULL - this prevents new references from being acquired |
| 252 | PRD_ARRAY *old_arr = prd_array_replace(&st->pluginsd.prd_array, NULL); |
| 253 | |
| 254 | // Capture last_slot before resetting - we need it to clear the host mapping |
| 255 | int32_t last_slot = st->pluginsd.last_slot; |
| 256 | |
| 257 | // Reset state while holding the lock |
| 258 | __atomic_store_n(&st->pluginsd.pos, 0, __ATOMIC_RELAXED); |
| 259 | st->pluginsd.set = false; |
| 260 | st->pluginsd.last_slot = -1; |
| 261 | st->pluginsd.dims_with_slots = false; |
| 262 | |
| 263 | spinlock_unlock(&st->pluginsd.spinlock); |
| 264 | |
| 265 | // Clear the chart slot mapping using the captured last_slot value |
| 266 | rrdset_clear_host_chart_slot_mapping(st, last_slot); |
| 267 | |
| 268 | // Now handle the old array outside the lock |
| 269 | if (old_arr) { |
| 270 | // After prd_array_replace, we hold the only reference (refcount should be 1). |
| 271 | // It's safe to release entries only when we're the sole owner, to avoid clearing |
| 272 | // entries that another thread might still be reading through its own reference. |
| 273 | int32_t rc = __atomic_load_n(&old_arr->refcount, __ATOMIC_ACQUIRE); |
| 274 | internal_fatal(rc != 1, |
| 275 | "PRD_ARRAY: expected refcount 1 after replace, got %d - concurrent reference leak", rc); |
| 276 | |
| 277 | if(unlikely(rc != 1)) { |
| 278 | // Production guard: another reference still exists, so clearing entries |
| 279 | // here could race with readers and double-release RRDDIM_ACQUIRED. |
| 280 | nd_log_limit_static_global_var(erl_cleanup_rc, 1, 0); |
| 281 | nd_log_limit(&erl_cleanup_rc, NDLS_DAEMON, NDLP_WARNING, |
| 282 | "PLUGINSD: cleanup deferred for chart with unexpected PRD_ARRAY refcount %d (expected 1)", |
| 283 | rc); |
| 284 | |
| 285 | // Drop our reference only; remaining owners will eventually release. |
| 286 | prd_array_release(old_arr); |
| 287 | return; |
| 288 | } |
| 289 | |
| 290 | // Release all dimension references (safe - we're the sole owner). |
| 291 | prd_array_release_entries(old_arr); |
| 292 | |
| 293 | // Release our reference - this will free the array (refcount 1 -> 0) |
| 294 | prd_array_release(old_arr); |
| 295 | } |
| 296 | } |
| 297 | |
| 298 | // -------------------------------------------------------------------------------------------------------------------- |
| 299 | // Initialize the pluginsd slots for a chart |
| 300 | |
| 301 | void rrdset_pluginsd_receive_slots_initialize(RRDSET *st) { |
| 302 | spinlock_init(&st->pluginsd.spinlock); |
| 303 | st->pluginsd.last_slot = -1; |
| 304 | st->pluginsd.prd_array = NULL; // Explicitly initialize to NULL |
| 305 | } |
| 306 | |
| 307 | // -------------------------------------------------------------------------------------------------------------------- |
| 308 | // Stress test for PRD_ARRAY lifecycle separation model |
| 309 | // Run with: netdata -W prd-array-stress |
| 310 | // |
| 311 | // This test validates the lifecycle separation model used in production: |
| 312 | // - In production, the collector is FULLY STOPPED before cleanup runs |
| 313 | // - The collector_tid check is a safety mechanism, but the real protection comes from lifecycle separation |
| 314 | // - This test simulates that by running the writer and cleaner in non-overlapping phases |
| 315 | // |
| 316 | // The test runs in cycles: |
| 317 | // 1. Writer phase: collector runs multiple iterations (collector_tid set) |
| 318 | // 2. Handoff: collector fully stops (collector_tid cleared, writer_done signaled) |
| 319 | // 3. Cleanup phase: cleaner runs (only when writer is fully stopped) |
| 320 | // 4. Repeat |
| 321 | // -------------------------------------------------------------------------------------------------------------------- |
| 322 | |
| 323 | #define PRD_STRESS_TEST_DURATION_SEC 5 |
| 324 | #define PRD_STRESS_ITERATIONS_PER_PHASE 50 |
| 325 | |
| 326 | typedef struct { |
| 327 | PRD_ARRAY *prd_array; |
| 328 | pid_t collector_tid; |
| 329 | SPINLOCK spinlock; |
| 330 | |
| 331 | // Lifecycle coordination (simulates stream receiver stop/start) |
| 332 | bool test_running; // Overall test is running |
| 333 | bool writer_should_run; // Writer is allowed to run |
| 334 | bool writer_is_running; // Writer is currently in a phase |
| 335 | |
| 336 | // Counters |
| 337 | uint64_t grow_count; |
| 338 | uint64_t cleanup_count; |
| 339 | uint64_t phase_count; |
| 340 | } prd_stress_state_t; |
| 341 | |
| 342 | static prd_stress_state_t prd_stress_state; |
| 343 | |
| 344 | static void prd_stress_writer_thread(void *arg __maybe_unused) { |
| 345 | while (__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE)) { |
| 346 | |
| 347 | // Wait until we're allowed to run (simulates stream receiver starting) |
| 348 | while (__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE) && |
| 349 | !__atomic_load_n(&prd_stress_state.writer_should_run, __ATOMIC_ACQUIRE)) { |
| 350 | tinysleep(); |
| 351 | } |
| 352 | |
| 353 | if (!__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE)) |
| 354 | break; |
| 355 | |
| 356 | // Signal that writer is now running |
| 357 | __atomic_store_n(&prd_stress_state.writer_is_running, true, __ATOMIC_RELEASE); |
| 358 | |
| 359 | // Simulate collector_tid being set (like pluginsd_set_scope_chart does) |
| 360 | __atomic_store_n(&prd_stress_state.collector_tid, gettid_cached(), __ATOMIC_RELEASE); |
| 361 | |
| 362 | // Run multiple iterations in this phase (simulates collecting data) |
| 363 | for (int iter = 0; iter < PRD_STRESS_ITERATIONS_PER_PHASE; iter++) { |
| 364 | if (!__atomic_load_n(&prd_stress_state.writer_should_run, __ATOMIC_ACQUIRE)) |
| 365 | break; |
| 366 | |
| 367 | PRD_ARRAY *current_arr = prd_array_get_unsafe(&prd_stress_state.prd_array); |
| 368 | |
| 369 | size_t current_size = current_arr ? current_arr->size : 0; |
| 370 | size_t new_size = current_size + 10; |
| 371 | |
| 372 | if (new_size > 500) |
| 373 | new_size = 10; |
| 374 | |
| 375 | PRD_ARRAY *new_arr = prd_array_create(new_size); |
| 376 | |
| 377 | if (current_arr && current_size > 0) { |
| 378 | size_t copy_count = (current_size < new_size) ? current_size : new_size; |
| 379 | for(size_t i = 0; i < copy_count; i++) { |
| 380 | new_arr->entries[i].rd = current_arr->entries[i].rd; |
| 381 | new_arr->entries[i].id = current_arr->entries[i].id; |
| 382 | new_arr->entries[i].rda = NULL; |
| 383 | } |
| 384 | } |
| 385 | |
| 386 | for (size_t i = (current_size < new_size ? current_size : 0); i < new_size; i++) { |
| 387 | new_arr->entries[i].rd = (void *)(uintptr_t)(i + 1); |
| 388 | new_arr->entries[i].id = "test"; |
| 389 | } |
| 390 | |
| 391 | PRD_ARRAY *old_arr = prd_array_replace(&prd_stress_state.prd_array, new_arr); |
| 392 | |
| 393 | if (old_arr) |
| 394 | prd_array_release(old_arr); |
| 395 | |
| 396 | __atomic_fetch_add(&prd_stress_state.grow_count, 1, __ATOMIC_RELAXED); |
| 397 | |
| 398 | tinysleep(); |
| 399 | } |
| 400 | |
| 401 | // Clear collector_tid (like pluginsd_set_scope_chart does when switching away) |
| 402 | __atomic_store_n(&prd_stress_state.collector_tid, 0, __ATOMIC_RELEASE); |
| 403 | |
| 404 | // Signal that writer phase is complete |
| 405 | __atomic_store_n(&prd_stress_state.writer_is_running, false, __ATOMIC_RELEASE); |
| 406 | |
| 407 | // Wait until controller signals us to run again |
| 408 | while (__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE) && |
| 409 | !__atomic_load_n(&prd_stress_state.writer_should_run, __ATOMIC_ACQUIRE)) { |
| 410 | tinysleep(); |
| 411 | } |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | static void prd_stress_cleanup_thread(void *arg __maybe_unused) { |
| 416 | while (__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE)) { |
| 417 | |
| 418 | // Wait until writer is fully stopped (simulates stream_receiver_signal_to_stop_and_wait) |
| 419 | while (__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE) && |
| 420 | __atomic_load_n(&prd_stress_state.writer_is_running, __ATOMIC_ACQUIRE)) { |
| 421 | tinysleep(); |
| 422 | } |
| 423 | |
| 424 | if (!__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE)) |
| 425 | break; |
| 426 | |
| 427 | // Now safe to cleanup - writer is fully stopped |
| 428 | spinlock_lock(&prd_stress_state.spinlock); |
| 429 | |
| 430 | // Double-check collector_tid (should be 0 since writer stopped) |
| 431 | pid_t collector_tid = __atomic_load_n(&prd_stress_state.collector_tid, __ATOMIC_ACQUIRE); |
| 432 | if (collector_tid != 0) { |
| 433 | // This shouldn't happen if lifecycle is correct |
| 434 | spinlock_unlock(&prd_stress_state.spinlock); |
| 435 | continue; |
| 436 | } |
| 437 | |
| 438 | PRD_ARRAY *old_arr = prd_array_replace(&prd_stress_state.prd_array, NULL); |
| 439 | |
| 440 | spinlock_unlock(&prd_stress_state.spinlock); |
| 441 | |
| 442 | if (old_arr) { |
| 443 | for (size_t i = 0; i < old_arr->size; i++) { |
| 444 | old_arr->entries[i].rda = NULL; |
| 445 | old_arr->entries[i].rd = NULL; |
| 446 | old_arr->entries[i].id = NULL; |
| 447 | } |
| 448 | |
| 449 | prd_array_release(old_arr); |
| 450 | __atomic_fetch_add(&prd_stress_state.cleanup_count, 1, __ATOMIC_RELAXED); |
| 451 | } |
| 452 | |
| 453 | tinysleep(); |
| 454 | } |
| 455 | } |
| 456 | |
| 457 | // Controller thread - orchestrates the lifecycle phases |
| 458 | static void prd_stress_controller_thread(void *arg __maybe_unused) { |
| 459 | while (__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE)) { |
| 460 | |
| 461 | // Start writer phase |
| 462 | __atomic_store_n(&prd_stress_state.writer_should_run, true, __ATOMIC_RELEASE); |
| 463 | |
| 464 | // Wait for writer to start and run |
| 465 | sleep_usec(10000); // 10ms - let writer run |
| 466 | |
| 467 | // Signal writer to stop (simulates stream receiver stopping) |
| 468 | __atomic_store_n(&prd_stress_state.writer_should_run, false, __ATOMIC_RELEASE); |
| 469 | |
| 470 | // Wait for writer to fully stop |
| 471 | while (__atomic_load_n(&prd_stress_state.test_running, __ATOMIC_ACQUIRE) && |
| 472 | __atomic_load_n(&prd_stress_state.writer_is_running, __ATOMIC_ACQUIRE)) { |
| 473 | tinysleep(); |
| 474 | } |
| 475 | |
| 476 | // Cleanup phase - cleaner will run now that writer is stopped |
| 477 | sleep_usec(5000); // 5ms - let cleanup run |
| 478 | |
| 479 | __atomic_fetch_add(&prd_stress_state.phase_count, 1, __ATOMIC_RELAXED); |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | int prd_array_stress_test(void) { |
| 484 | int duration_secs = PRD_STRESS_TEST_DURATION_SEC; |
| 485 | |
| 486 | fprintf(stderr, "\nPRD_ARRAY Lifecycle Stress Test\n"); |
| 487 | fprintf(stderr, "================================\n"); |
| 488 | fprintf(stderr, "Duration: %d seconds\n", duration_secs); |
| 489 | fprintf(stderr, "This test simulates production lifecycle:\n"); |
| 490 | fprintf(stderr, " 1. Writer (collector) runs with collector_tid set\n"); |
| 491 | fprintf(stderr, " 2. Writer fully stops (collector_tid cleared)\n"); |
| 492 | fprintf(stderr, " 3. Cleaner runs cleanup\n"); |
| 493 | fprintf(stderr, " 4. Repeat\n\n"); |
| 494 | |
| 495 | // Initialize state |
| 496 | memset(&prd_stress_state, 0, sizeof(prd_stress_state)); |
| 497 | prd_stress_state.prd_array = prd_array_create(10); |
| 498 | spinlock_init(&prd_stress_state.spinlock); |
| 499 | __atomic_store_n(&prd_stress_state.test_running, true, __ATOMIC_RELEASE); |
| 500 | |
| 501 | // Start threads |
| 502 | char thread_name[32]; |
| 503 | |
| 504 | snprintfz(thread_name, sizeof(thread_name), "PRDSTRESS_W"); |
| 505 | ND_THREAD *writer_thread = nd_thread_create(thread_name, NETDATA_THREAD_OPTION_DEFAULT, |
| 506 | prd_stress_writer_thread, NULL); |
| 507 | |
| 508 | snprintfz(thread_name, sizeof(thread_name), "PRDSTRESS_C"); |
| 509 | ND_THREAD *cleanup_thread = nd_thread_create(thread_name, NETDATA_THREAD_OPTION_DEFAULT, |
| 510 | prd_stress_cleanup_thread, NULL); |
| 511 | |
| 512 | snprintfz(thread_name, sizeof(thread_name), "PRDSTRESS_CTRL"); |
| 513 | ND_THREAD *controller_thread = nd_thread_create(thread_name, NETDATA_THREAD_OPTION_DEFAULT, |
| 514 | prd_stress_controller_thread, NULL); |
| 515 | |
| 516 | // Run the test |
| 517 | fprintf(stderr, "Running stress test...\n"); |
| 518 | for (int i = 0; i < duration_secs; i++) { |
| 519 | sleep_usec(USEC_PER_SEC); |
| 520 | fprintf(stderr, " %d/%d sec - phases: %"PRIu64", grows: %"PRIu64", cleanups: %"PRIu64"\n", |
| 521 | i + 1, duration_secs, |
| 522 | __atomic_load_n(&prd_stress_state.phase_count, __ATOMIC_RELAXED), |
| 523 | __atomic_load_n(&prd_stress_state.grow_count, __ATOMIC_RELAXED), |
| 524 | __atomic_load_n(&prd_stress_state.cleanup_count, __ATOMIC_RELAXED)); |
| 525 | } |
| 526 | |
| 527 | // Stop all threads |
| 528 | __atomic_store_n(&prd_stress_state.test_running, false, __ATOMIC_RELEASE); |
| 529 | __atomic_store_n(&prd_stress_state.writer_should_run, true, __ATOMIC_RELEASE); // Unblock writer |
| 530 | |
| 531 | nd_thread_join(controller_thread); |
| 532 | nd_thread_join(writer_thread); |
| 533 | nd_thread_join(cleanup_thread); |
| 534 | |
| 535 | // Final cleanup |
| 536 | PRD_ARRAY *final_arr = prd_array_replace(&prd_stress_state.prd_array, NULL); |
| 537 | if (final_arr) |
| 538 | prd_array_release(final_arr); |
| 539 | |
| 540 | // Print results |
| 541 | fprintf(stderr, "\nTest completed!\n"); |
| 542 | fprintf(stderr, "===============\n"); |
| 543 | fprintf(stderr, "Total phases: %"PRIu64"\n", prd_stress_state.phase_count); |
| 544 | fprintf(stderr, "Total grows: %"PRIu64"\n", prd_stress_state.grow_count); |
| 545 | fprintf(stderr, "Total cleanups: %"PRIu64"\n", prd_stress_state.cleanup_count); |
| 546 | |
| 547 | if (prd_stress_state.cleanup_count > 0 && prd_stress_state.grow_count > 0) { |
| 548 | fprintf(stderr, "\nSUCCESS: Lifecycle separation validated\n"); |
| 549 | fprintf(stderr, "- Writer and cleaner ran in non-overlapping phases\n"); |
| 550 | fprintf(stderr, "- No concurrent access to the array\n"); |
| 551 | fprintf(stderr, "- Reference counting worked correctly\n"); |
| 552 | return 0; |
| 553 | } else { |
| 554 | fprintf(stderr, "\nWARNING: Low activity - increase test duration\n"); |
| 555 | return 1; |
| 556 | } |
| 557 | } |