| 1 |
// C Extension module to test pycore_lock.h API |
| 2 |
|
| 3 |
#include "parts.h" |
| 4 |
#include "pycore_lock.h" |
| 5 |
#include "pycore_pythread.h" // PyThread_get_thread_ident_ex() |
| 6 |
|
| 7 |
#include "clinic/test_lock.c.h" |
| 8 |
|
| 9 |
#ifdef MS_WINDOWS |
| 10 |
#define WIN32_LEAN_AND_MEAN |
| 11 |
#include <windows.h> |
| 12 |
#else |
| 13 |
#include <unistd.h> // usleep() |
| 14 |
#endif |
| 15 |
|
| 16 |
/*[clinic input] |
| 17 |
module _testinternalcapi |
| 18 |
[clinic start generated code]*/ |
| 19 |
/*[clinic end generated code: output=da39a3ee5e6b4b0d input=7bb583d8c9eb9a78]*/ |
| 20 |
|
| 21 |
|
| 22 |
static void |
| 23 |
pysleep(int ms) |
| 24 |
{ |
| 25 |
#ifdef MS_WINDOWS |
| 26 |
Sleep(ms); |
| 27 |
#else |
| 28 |
usleep(ms * 1000); |
| 29 |
#endif |
| 30 |
} |
| 31 |
|
| 32 |
static PyObject * |
| 33 |
test_lock_basic(PyObject *self, PyObject *obj) |
| 34 |
{ |
| 35 |
PyMutex m = (PyMutex){0}; |
| 36 |
|
| 37 |
// uncontended lock and unlock |
| 38 |
PyMutex_Lock(&m); |
| 39 |
assert(m._bits == 1); |
| 40 |
PyMutex_Unlock(&m); |
| 41 |
assert(m._bits == 0); |
| 42 |
|
| 43 |
Py_RETURN_NONE; |
| 44 |
} |
| 45 |
|
| 46 |
struct test_lock2_data { |
| 47 |
PyMutex m; |
| 48 |
PyEvent done; |
| 49 |
int started; |
| 50 |
}; |
| 51 |
|
| 52 |
static void |
| 53 |
lock_thread(void *arg) |
| 54 |
{ |
| 55 |
struct test_lock2_data *test_data = arg; |
| 56 |
PyMutex *m = &test_data->m; |
| 57 |
_Py_atomic_store_int(&test_data->started, 1); |
| 58 |
|
| 59 |
PyMutex_Lock(m); |
| 60 |
// gh-135641: in rare cases the lock may still have `_Py_HAS_PARKED` set |
| 61 |
// (m->_bits == 3) due to bucket collisions in the parking lot hash table |
| 62 |
// between this mutex and the `test_data.done` event. |
| 63 |
assert(m->_bits == 1 || m->_bits == 3); |
| 64 |
|
| 65 |
PyMutex_Unlock(m); |
| 66 |
assert(m->_bits == 0); |
| 67 |
|
| 68 |
_PyEvent_Notify(&test_data->done); |
| 69 |
} |
| 70 |
|
| 71 |
static PyObject * |
| 72 |
test_lock_two_threads(PyObject *self, PyObject *obj) |
| 73 |
{ |
| 74 |
// lock attempt by two threads |
| 75 |
struct test_lock2_data test_data; |
| 76 |
memset(&test_data, 0, sizeof(test_data)); |
| 77 |
|
| 78 |
PyMutex_Lock(&test_data.m); |
| 79 |
assert(test_data.m._bits == 1); |
| 80 |
|
| 81 |
PyThread_start_new_thread(lock_thread, &test_data); |
| 82 |
|
| 83 |
// wait up to two seconds for the lock_thread to attempt to lock "m" |
| 84 |
int iters = 0; |
| 85 |
uint8_t v; |
| 86 |
do { |
| 87 |
pysleep(10); // allow some time for the other thread to try to lock |
| 88 |
v = _Py_atomic_load_uint8_relaxed(&test_data.m._bits); |
| 89 |
assert(v == 1 || v == 3); |
| 90 |
iters++; |
| 91 |
} while (v != 3 && iters < 200); |
| 92 |
|
| 93 |
// both the "locked" and the "has parked" bits should be set |
| 94 |
v = _Py_atomic_load_uint8_relaxed(&test_data.m._bits); |
| 95 |
assert(v == 3); |
| 96 |
|
| 97 |
PyMutex_Unlock(&test_data.m); |
| 98 |
PyEvent_Wait(&test_data.done); |
| 99 |
assert(test_data.m._bits == 0); |
| 100 |
|
| 101 |
Py_RETURN_NONE; |
| 102 |
} |
| 103 |
|
| 104 |
#define COUNTER_THREADS 5 |
| 105 |
#define COUNTER_ITERS 10000 |
| 106 |
|
| 107 |
struct test_data_counter { |
| 108 |
PyMutex m; |
| 109 |
Py_ssize_t counter; |
| 110 |
}; |
| 111 |
|
| 112 |
struct thread_data_counter { |
| 113 |
struct test_data_counter *test_data; |
| 114 |
PyEvent done_event; |
| 115 |
}; |
| 116 |
|
| 117 |
static void |
| 118 |
counter_thread(void *arg) |
| 119 |
{ |
| 120 |
struct thread_data_counter *thread_data = arg; |
| 121 |
struct test_data_counter *test_data = thread_data->test_data; |
| 122 |
|
| 123 |
for (Py_ssize_t i = 0; i < COUNTER_ITERS; i++) { |
| 124 |
PyMutex_Lock(&test_data->m); |
| 125 |
test_data->counter++; |
| 126 |
PyMutex_Unlock(&test_data->m); |
| 127 |
} |
| 128 |
_PyEvent_Notify(&thread_data->done_event); |
| 129 |
} |
| 130 |
|
| 131 |
static PyObject * |
| 132 |
test_lock_counter(PyObject *self, PyObject *obj) |
| 133 |
{ |
| 134 |
// Test with rapidly locking and unlocking mutex |
| 135 |
struct test_data_counter test_data; |
| 136 |
memset(&test_data, 0, sizeof(test_data)); |
| 137 |
|
| 138 |
struct thread_data_counter thread_data[COUNTER_THREADS]; |
| 139 |
memset(&thread_data, 0, sizeof(thread_data)); |
| 140 |
|
| 141 |
for (Py_ssize_t i = 0; i < COUNTER_THREADS; i++) { |
| 142 |
thread_data[i].test_data = &test_data; |
| 143 |
PyThread_start_new_thread(counter_thread, &thread_data[i]); |
| 144 |
} |
| 145 |
|
| 146 |
for (Py_ssize_t i = 0; i < COUNTER_THREADS; i++) { |
| 147 |
PyEvent_Wait(&thread_data[i].done_event); |
| 148 |
} |
| 149 |
|
| 150 |
assert(test_data.counter == COUNTER_THREADS * COUNTER_ITERS); |
| 151 |
Py_RETURN_NONE; |
| 152 |
} |
| 153 |
|
| 154 |
#define SLOW_COUNTER_ITERS 100 |
| 155 |
|
| 156 |
static void |
| 157 |
slow_counter_thread(void *arg) |
| 158 |
{ |
| 159 |
struct thread_data_counter *thread_data = arg; |
| 160 |
struct test_data_counter *test_data = thread_data->test_data; |
| 161 |
|
| 162 |
for (Py_ssize_t i = 0; i < SLOW_COUNTER_ITERS; i++) { |
| 163 |
PyMutex_Lock(&test_data->m); |
| 164 |
if (i % 7 == 0) { |
| 165 |
pysleep(2); |
| 166 |
} |
| 167 |
test_data->counter++; |
| 168 |
PyMutex_Unlock(&test_data->m); |
| 169 |
} |
| 170 |
_PyEvent_Notify(&thread_data->done_event); |
| 171 |
} |
| 172 |
|
| 173 |
static PyObject * |
| 174 |
test_lock_counter_slow(PyObject *self, PyObject *obj) |
| 175 |
{ |
| 176 |
// Test lock/unlock with occasional "long" critical section, which will |
| 177 |
// trigger handoff of the lock. |
| 178 |
struct test_data_counter test_data; |
| 179 |
memset(&test_data, 0, sizeof(test_data)); |
| 180 |
|
| 181 |
struct thread_data_counter thread_data[COUNTER_THREADS]; |
| 182 |
memset(&thread_data, 0, sizeof(thread_data)); |
| 183 |
|
| 184 |
for (Py_ssize_t i = 0; i < COUNTER_THREADS; i++) { |
| 185 |
thread_data[i].test_data = &test_data; |
| 186 |
PyThread_start_new_thread(slow_counter_thread, &thread_data[i]); |
| 187 |
} |
| 188 |
|
| 189 |
for (Py_ssize_t i = 0; i < COUNTER_THREADS; i++) { |
| 190 |
PyEvent_Wait(&thread_data[i].done_event); |
| 191 |
} |
| 192 |
|
| 193 |
assert(test_data.counter == COUNTER_THREADS * SLOW_COUNTER_ITERS); |
| 194 |
Py_RETURN_NONE; |
| 195 |
} |
| 196 |
|
| 197 |
struct bench_data_locks { |
| 198 |
int stop; |
| 199 |
int use_pymutex; |
| 200 |
int critical_section_length; |
| 201 |
char padding[200]; |
| 202 |
PyThread_type_lock lock; |
| 203 |
PyMutex m; |
| 204 |
double value; |
| 205 |
Py_ssize_t total_iters; |
| 206 |
}; |
| 207 |
|
| 208 |
struct bench_thread_data { |
| 209 |
struct bench_data_locks *bench_data; |
| 210 |
Py_ssize_t iters; |
| 211 |
PyEvent done; |
| 212 |
}; |
| 213 |
|
| 214 |
static void |
| 215 |
thread_benchmark_locks(void *arg) |
| 216 |
{ |
| 217 |
struct bench_thread_data *thread_data = arg; |
| 218 |
struct bench_data_locks *bench_data = thread_data->bench_data; |
| 219 |
int use_pymutex = bench_data->use_pymutex; |
| 220 |
int critical_section_length = bench_data->critical_section_length; |
| 221 |
|
| 222 |
double my_value = 1.0; |
| 223 |
Py_ssize_t iters = 0; |
| 224 |
while (!_Py_atomic_load_int_relaxed(&bench_data->stop)) { |
| 225 |
if (use_pymutex) { |
| 226 |
PyMutex_Lock(&bench_data->m); |
| 227 |
for (int i = 0; i < critical_section_length; i++) { |
| 228 |
bench_data->value += my_value; |
| 229 |
my_value = bench_data->value; |
| 230 |
} |
| 231 |
PyMutex_Unlock(&bench_data->m); |
| 232 |
} |
| 233 |
else { |
| 234 |
PyThread_acquire_lock(bench_data->lock, 1); |
| 235 |
for (int i = 0; i < critical_section_length; i++) { |
| 236 |
bench_data->value += my_value; |
| 237 |
my_value = bench_data->value; |
| 238 |
} |
| 239 |
PyThread_release_lock(bench_data->lock); |
| 240 |
} |
| 241 |
iters++; |
| 242 |
} |
| 243 |
|
| 244 |
thread_data->iters = iters; |
| 245 |
_Py_atomic_add_ssize(&bench_data->total_iters, iters); |
| 246 |
_PyEvent_Notify(&thread_data->done); |
| 247 |
} |
| 248 |
|
| 249 |
/*[clinic input] |
| 250 |
_testinternalcapi.benchmark_locks |
| 251 |
|
| 252 |
num_threads: Py_ssize_t |
| 253 |
use_pymutex: bool = True |
| 254 |
critical_section_length: int = 1 |
| 255 |
time_ms: int = 1000 |
| 256 |
/ |
| 257 |
|
| 258 |
[clinic start generated code]*/ |
| 259 |
|
| 260 |
static PyObject * |
| 261 |
_testinternalcapi_benchmark_locks_impl(PyObject *module, |
| 262 |
Py_ssize_t num_threads, |
| 263 |
int use_pymutex, |
| 264 |
int critical_section_length, |
| 265 |
int time_ms) |
| 266 |
/*[clinic end generated code: output=381df8d7e9a74f18 input=f3aeaf688738c121]*/ |
| 267 |
{ |
| 268 |
// Run from Tools/lockbench/lockbench.py |
| 269 |
// Based on the WebKit lock benchmarks: |
| 270 |
// https://github.com/WebKit/WebKit/blob/main/Source/WTF/benchmarks/LockSpeedTest.cpp |
| 271 |
// See also https://webkit.org/blog/6161/locking-in-webkit/ |
| 272 |
PyObject *thread_iters = NULL; |
| 273 |
PyObject *res = NULL; |
| 274 |
|
| 275 |
struct bench_data_locks bench_data; |
| 276 |
memset(&bench_data, 0, sizeof(bench_data)); |
| 277 |
bench_data.use_pymutex = use_pymutex; |
| 278 |
bench_data.critical_section_length = critical_section_length; |
| 279 |
|
| 280 |
bench_data.lock = PyThread_allocate_lock(); |
| 281 |
if (bench_data.lock == NULL) { |
| 282 |
return PyErr_NoMemory(); |
| 283 |
} |
| 284 |
|
| 285 |
struct bench_thread_data *thread_data = NULL; |
| 286 |
thread_data = PyMem_Calloc(num_threads, sizeof(*thread_data)); |
| 287 |
if (thread_data == NULL) { |
| 288 |
PyErr_NoMemory(); |
| 289 |
goto exit; |
| 290 |
} |
| 291 |
|
| 292 |
thread_iters = PyList_New(num_threads); |
| 293 |
if (thread_iters == NULL) { |
| 294 |
goto exit; |
| 295 |
} |
| 296 |
|
| 297 |
PyTime_t start, end; |
| 298 |
if (PyTime_PerfCounter(&start) < 0) { |
| 299 |
goto exit; |
| 300 |
} |
| 301 |
|
| 302 |
for (Py_ssize_t i = 0; i < num_threads; i++) { |
| 303 |
thread_data[i].bench_data = &bench_data; |
| 304 |
PyThread_start_new_thread(thread_benchmark_locks, &thread_data[i]); |
| 305 |
} |
| 306 |
|
| 307 |
// Let the threads run for `time_ms` milliseconds |
| 308 |
pysleep(time_ms); |
| 309 |
_Py_atomic_store_int(&bench_data.stop, 1); |
| 310 |
|
| 311 |
// Wait for the threads to finish |
| 312 |
for (Py_ssize_t i = 0; i < num_threads; i++) { |
| 313 |
PyEvent_Wait(&thread_data[i].done); |
| 314 |
} |
| 315 |
|
| 316 |
Py_ssize_t total_iters = bench_data.total_iters; |
| 317 |
if (PyTime_PerfCounter(&end) < 0) { |
| 318 |
goto exit; |
| 319 |
} |
| 320 |
|
| 321 |
// Return the total number of acquisitions and the number of acquisitions |
| 322 |
// for each thread. |
| 323 |
for (Py_ssize_t i = 0; i < num_threads; i++) { |
| 324 |
PyObject *iter = PyLong_FromSsize_t(thread_data[i].iters); |
| 325 |
if (iter == NULL) { |
| 326 |
goto exit; |
| 327 |
} |
| 328 |
PyList_SET_ITEM(thread_iters, i, iter); |
| 329 |
} |
| 330 |
|
| 331 |
assert(end != start); |
| 332 |
double rate = total_iters * 1e9 / (end - start); |
| 333 |
res = Py_BuildValue("(dO)", rate, thread_iters); |
| 334 |
|
| 335 |
exit: |
| 336 |
PyThread_free_lock(bench_data.lock); |
| 337 |
PyMem_Free(thread_data); |
| 338 |
Py_XDECREF(thread_iters); |
| 339 |
return res; |
| 340 |
} |
| 341 |
|
| 342 |
static PyObject * |
| 343 |
test_lock_benchmark(PyObject *module, PyObject *obj) |
| 344 |
{ |
| 345 |
// Just make sure the benchmark runs without crashing |
| 346 |
PyObject *res = _testinternalcapi_benchmark_locks_impl( |
| 347 |
module, 1, 1, 1, 100); |
| 348 |
if (res == NULL) { |
| 349 |
return NULL; |
| 350 |
} |
| 351 |
Py_DECREF(res); |
| 352 |
Py_RETURN_NONE; |
| 353 |
} |
| 354 |
|
| 355 |
static int |
| 356 |
init_maybe_fail(void *arg) |
| 357 |
{ |
| 358 |
int *counter = (int *)arg; |
| 359 |
(*counter)++; |
| 360 |
if (*counter < 5) { |
| 361 |
// failure |
| 362 |
return -1; |
| 363 |
} |
| 364 |
assert(*counter == 5); |
| 365 |
return 0; |
| 366 |
} |
| 367 |
|
| 368 |
static PyObject * |
| 369 |
test_lock_once(PyObject *self, PyObject *obj) |
| 370 |
{ |
| 371 |
_PyOnceFlag once = {0}; |
| 372 |
int counter = 0; |
| 373 |
for (int i = 0; i < 10; i++) { |
| 374 |
int res = _PyOnceFlag_CallOnce(&once, init_maybe_fail, &counter); |
| 375 |
if (i < 4) { |
| 376 |
assert(res == -1); |
| 377 |
} |
| 378 |
else { |
| 379 |
assert(res == 0); |
| 380 |
assert(counter == 5); |
| 381 |
} |
| 382 |
} |
| 383 |
Py_RETURN_NONE; |
| 384 |
} |
| 385 |
|
| 386 |
struct test_rwlock_data { |
| 387 |
Py_ssize_t nthreads; |
| 388 |
_PyRWMutex rw; |
| 389 |
PyEvent step1; |
| 390 |
PyEvent step2; |
| 391 |
PyEvent step3; |
| 392 |
PyEvent done; |
| 393 |
}; |
| 394 |
|
| 395 |
static void |
| 396 |
rdlock_thread(void *arg) |
| 397 |
{ |
| 398 |
struct test_rwlock_data *test_data = arg; |
| 399 |
|
| 400 |
// Acquire the lock in read mode |
| 401 |
_PyRWMutex_RLock(&test_data->rw); |
| 402 |
PyEvent_Wait(&test_data->step1); |
| 403 |
_PyRWMutex_RUnlock(&test_data->rw); |
| 404 |
|
| 405 |
_PyRWMutex_RLock(&test_data->rw); |
| 406 |
PyEvent_Wait(&test_data->step3); |
| 407 |
_PyRWMutex_RUnlock(&test_data->rw); |
| 408 |
|
| 409 |
if (_Py_atomic_add_ssize(&test_data->nthreads, -1) == 1) { |
| 410 |
_PyEvent_Notify(&test_data->done); |
| 411 |
} |
| 412 |
} |
| 413 |
static void |
| 414 |
wrlock_thread(void *arg) |
| 415 |
{ |
| 416 |
struct test_rwlock_data *test_data = arg; |
| 417 |
|
| 418 |
// First acquire the lock in write mode |
| 419 |
_PyRWMutex_Lock(&test_data->rw); |
| 420 |
PyEvent_Wait(&test_data->step2); |
| 421 |
_PyRWMutex_Unlock(&test_data->rw); |
| 422 |
|
| 423 |
if (_Py_atomic_add_ssize(&test_data->nthreads, -1) == 1) { |
| 424 |
_PyEvent_Notify(&test_data->done); |
| 425 |
} |
| 426 |
} |
| 427 |
|
| 428 |
static void |
| 429 |
wait_until(uintptr_t *ptr, uintptr_t value) |
| 430 |
{ |
| 431 |
// wait up to two seconds for *ptr == value |
| 432 |
int iters = 0; |
| 433 |
uintptr_t bits; |
| 434 |
do { |
| 435 |
pysleep(10); |
| 436 |
bits = _Py_atomic_load_uintptr(ptr); |
| 437 |
iters++; |
| 438 |
} while (bits != value && iters < 200); |
| 439 |
} |
| 440 |
|
| 441 |
static PyObject * |
| 442 |
test_lock_rwlock(PyObject *self, PyObject *obj) |
| 443 |
{ |
| 444 |
struct test_rwlock_data test_data = {.nthreads = 3}; |
| 445 |
|
| 446 |
_PyRWMutex_Lock(&test_data.rw); |
| 447 |
assert(test_data.rw.bits == 1); |
| 448 |
|
| 449 |
_PyRWMutex_Unlock(&test_data.rw); |
| 450 |
assert(test_data.rw.bits == 0); |
| 451 |
|
| 452 |
// Start two readers |
| 453 |
PyThread_start_new_thread(rdlock_thread, &test_data); |
| 454 |
PyThread_start_new_thread(rdlock_thread, &test_data); |
| 455 |
|
| 456 |
// wait up to two seconds for the threads to attempt to read-lock "rw" |
| 457 |
wait_until(&test_data.rw.bits, 8); |
| 458 |
assert(test_data.rw.bits == 8); |
| 459 |
|
| 460 |
// start writer (while readers hold lock) |
| 461 |
PyThread_start_new_thread(wrlock_thread, &test_data); |
| 462 |
wait_until(&test_data.rw.bits, 10); |
| 463 |
assert(test_data.rw.bits == 10); |
| 464 |
|
| 465 |
// readers release lock, writer should acquire it |
| 466 |
_PyEvent_Notify(&test_data.step1); |
| 467 |
wait_until(&test_data.rw.bits, 3); |
| 468 |
assert(test_data.rw.bits == 3); |
| 469 |
|
| 470 |
// writer releases lock, readers acquire it |
| 471 |
_PyEvent_Notify(&test_data.step2); |
| 472 |
wait_until(&test_data.rw.bits, 8); |
| 473 |
assert(test_data.rw.bits == 8); |
| 474 |
|
| 475 |
// readers release lock again |
| 476 |
_PyEvent_Notify(&test_data.step3); |
| 477 |
wait_until(&test_data.rw.bits, 0); |
| 478 |
assert(test_data.rw.bits == 0); |
| 479 |
|
| 480 |
PyEvent_Wait(&test_data.done); |
| 481 |
Py_RETURN_NONE; |
| 482 |
} |
| 483 |
|
| 484 |
static PyObject * |
| 485 |
test_lock_recursive(PyObject *self, PyObject *obj) |
| 486 |
{ |
| 487 |
_PyRecursiveMutex m = (_PyRecursiveMutex){0}; |
| 488 |
assert(!_PyRecursiveMutex_IsLockedByCurrentThread(&m)); |
| 489 |
|
| 490 |
_PyRecursiveMutex_Lock(&m); |
| 491 |
assert(m.thread == PyThread_get_thread_ident_ex()); |
| 492 |
assert(PyMutex_IsLocked(&m.mutex)); |
| 493 |
assert(m.level == 0); |
| 494 |
|
| 495 |
_PyRecursiveMutex_Lock(&m); |
| 496 |
assert(m.level == 1); |
| 497 |
_PyRecursiveMutex_Unlock(&m); |
| 498 |
|
| 499 |
_PyRecursiveMutex_Unlock(&m); |
| 500 |
assert(m.thread == 0); |
| 501 |
assert(!PyMutex_IsLocked(&m.mutex)); |
| 502 |
assert(m.level == 0); |
| 503 |
|
| 504 |
Py_RETURN_NONE; |
| 505 |
} |
| 506 |
|
| 507 |
static PyMethodDef test_methods[] = { |
| 508 |
{"test_lock_basic", test_lock_basic, METH_NOARGS}, |
| 509 |
{"test_lock_two_threads", test_lock_two_threads, METH_NOARGS}, |
| 510 |
{"test_lock_counter", test_lock_counter, METH_NOARGS}, |
| 511 |
{"test_lock_counter_slow", test_lock_counter_slow, METH_NOARGS}, |
| 512 |
_TESTINTERNALCAPI_BENCHMARK_LOCKS_METHODDEF |
| 513 |
{"test_lock_benchmark", test_lock_benchmark, METH_NOARGS}, |
| 514 |
{"test_lock_once", test_lock_once, METH_NOARGS}, |
| 515 |
{"test_lock_rwlock", test_lock_rwlock, METH_NOARGS}, |
| 516 |
{"test_lock_recursive", test_lock_recursive, METH_NOARGS}, |
| 517 |
{NULL, NULL} /* sentinel */ |
| 518 |
}; |
| 519 |
|
| 520 |
int |
| 521 |
_PyTestInternalCapi_Init_Lock(PyObject *mod) |
| 522 |
{ |
| 523 |
if (PyModule_AddFunctions(mod, test_methods) < 0) { |
| 524 |
return -1; |
| 525 |
} |
| 526 |
return 0; |
| 527 |
} |