/src/openssl34/crypto/threads_pthread.c
Line | Count | Source |
1 | | /* |
2 | | * Copyright 2016-2025 The OpenSSL Project Authors. All Rights Reserved. |
3 | | * |
4 | | * Licensed under the Apache License 2.0 (the "License"). You may not use |
5 | | * this file except in compliance with the License. You can obtain a copy |
6 | | * in the file LICENSE in the source distribution or at |
7 | | * https://www.openssl.org/source/license.html |
8 | | */ |
9 | | |
10 | | /* We need to use the OPENSSL_fork_*() deprecated APIs */ |
11 | | #define OPENSSL_SUPPRESS_DEPRECATED |
12 | | |
13 | | #include <openssl/crypto.h> |
14 | | #include <crypto/cryptlib.h> |
15 | | #include "internal/cryptlib.h" |
16 | | #include "internal/rcu.h" |
17 | | #include "rcu_internal.h" |
18 | | |
19 | | #if defined(__clang__) && defined(__has_feature) |
20 | | # if __has_feature(thread_sanitizer) |
21 | | # define __SANITIZE_THREAD__ |
22 | | # endif |
23 | | #endif |
24 | | |
25 | | #if defined(__SANITIZE_THREAD__) |
26 | | # include <sanitizer/tsan_interface.h> |
27 | | # define TSAN_FAKE_UNLOCK(x) __tsan_mutex_pre_unlock((x), 0); \ |
28 | | __tsan_mutex_post_unlock((x), 0) |
29 | | |
30 | | # define TSAN_FAKE_LOCK(x) __tsan_mutex_pre_lock((x), 0); \ |
31 | | __tsan_mutex_post_lock((x), 0, 0) |
32 | | #else |
33 | | # define TSAN_FAKE_UNLOCK(x) |
34 | | # define TSAN_FAKE_LOCK(x) |
35 | | #endif |
36 | | |
37 | | #if defined(__sun) |
38 | | # include <atomic.h> |
39 | | #endif |
40 | | |
41 | | #if defined(__apple_build_version__) && __apple_build_version__ < 6000000 |
42 | | /* |
43 | | * OS/X 10.7 and 10.8 had a weird version of clang which has __ATOMIC_ACQUIRE and |
44 | | * __ATOMIC_ACQ_REL but which expects only one parameter for __atomic_is_lock_free() |
45 | | * rather than two which has signature __atomic_is_lock_free(sizeof(_Atomic(T))). |
46 | | * All of this makes impossible to use __atomic_is_lock_free here. |
47 | | * |
48 | | * See: https://github.com/llvm/llvm-project/commit/a4c2602b714e6c6edb98164550a5ae829b2de760 |
49 | | */ |
50 | | # define BROKEN_CLANG_ATOMICS |
51 | | #endif |
52 | | |
53 | | #if defined(OPENSSL_THREADS) && !defined(CRYPTO_TDEBUG) && !defined(OPENSSL_SYS_WINDOWS) |
54 | | |
55 | | # if defined(OPENSSL_SYS_UNIX) |
56 | | # include <sys/types.h> |
57 | | # include <unistd.h> |
58 | | # endif |
59 | | |
60 | | # include <assert.h> |
61 | | |
62 | | /* |
63 | | * The Non-Stop KLT thread model currently seems broken in its rwlock |
64 | | * implementation |
65 | | * Likewise is there a problem with the glibc implementation on riscv. |
66 | | */ |
67 | | # if defined(PTHREAD_RWLOCK_INITIALIZER) && !defined(_KLT_MODEL_) \ |
68 | | && !defined(__riscv) |
69 | | # define USE_RWLOCK |
70 | | # endif |
71 | | |
72 | | /* |
73 | | * For all GNU/clang atomic builtins, we also need fallbacks, to cover all |
74 | | * other compilers. |
75 | | |
76 | | * Unfortunately, we can't do that with some "generic type", because there's no |
77 | | * guarantee that the chosen generic type is large enough to cover all cases. |
78 | | * Therefore, we implement fallbacks for each applicable type, with composed |
79 | | * names that include the type they handle. |
80 | | * |
81 | | * (an anecdote: we previously tried to use |void *| as the generic type, with |
82 | | * the thought that the pointer itself is the largest type. However, this is |
83 | | * not true on 32-bit pointer platforms, as a |uint64_t| is twice as large) |
84 | | * |
85 | | * All applicable ATOMIC_ macros take the intended type as first parameter, so |
86 | | * they can map to the correct fallback function. In the GNU/clang case, that |
87 | | * parameter is simply ignored. |
88 | | */ |
89 | | |
90 | | /* |
91 | | * Internal types used with the ATOMIC_ macros, to make it possible to compose |
92 | | * fallback function names. |
93 | | */ |
94 | | typedef void *pvoid; |
95 | | |
96 | | # if defined(__GNUC__) && defined(__ATOMIC_ACQUIRE) && !defined(BROKEN_CLANG_ATOMICS) \ |
97 | | && !defined(USE_ATOMIC_FALLBACKS) |
98 | 54.9M | # define ATOMIC_LOAD_N(t, p, o) __atomic_load_n(p, o) |
99 | 916 | # define ATOMIC_STORE_N(t, p, v, o) __atomic_store_n(p, v, o) |
100 | 28.7k | # define ATOMIC_STORE(t, p, v, o) __atomic_store(p, v, o) |
101 | 961 | # define ATOMIC_ADD_FETCH(p, v, o) __atomic_add_fetch(p, v, o) |
102 | 45 | # define ATOMIC_SUB_FETCH(p, v, o) __atomic_sub_fetch(p, v, o) |
103 | | # else |
104 | | static pthread_mutex_t atomic_sim_lock = PTHREAD_MUTEX_INITIALIZER; |
105 | | |
106 | | # define IMPL_fallback_atomic_load_n(t) \ |
107 | | static ossl_inline t fallback_atomic_load_n_##t(t *p) \ |
108 | | { \ |
109 | | t ret; \ |
110 | | \ |
111 | | pthread_mutex_lock(&atomic_sim_lock); \ |
112 | | ret = *p; \ |
113 | | pthread_mutex_unlock(&atomic_sim_lock); \ |
114 | | return ret; \ |
115 | | } |
116 | | IMPL_fallback_atomic_load_n(uint32_t) |
117 | | IMPL_fallback_atomic_load_n(uint64_t) |
118 | | IMPL_fallback_atomic_load_n(pvoid) |
119 | | |
120 | | # define ATOMIC_LOAD_N(t, p, o) fallback_atomic_load_n_##t(p) |
121 | | |
122 | | # define IMPL_fallback_atomic_store_n(t) \ |
123 | | static ossl_inline t fallback_atomic_store_n_##t(t *p, t v) \ |
124 | | { \ |
125 | | t ret; \ |
126 | | \ |
127 | | pthread_mutex_lock(&atomic_sim_lock); \ |
128 | | ret = *p; \ |
129 | | *p = v; \ |
130 | | pthread_mutex_unlock(&atomic_sim_lock); \ |
131 | | return ret; \ |
132 | | } |
133 | | IMPL_fallback_atomic_store_n(uint32_t) |
134 | | |
135 | | # define ATOMIC_STORE_N(t, p, v, o) fallback_atomic_store_n_##t(p, v) |
136 | | |
137 | | # define IMPL_fallback_atomic_store(t) \ |
138 | | static ossl_inline void fallback_atomic_store_##t(t *p, t *v) \ |
139 | | { \ |
140 | | pthread_mutex_lock(&atomic_sim_lock); \ |
141 | | *p = *v; \ |
142 | | pthread_mutex_unlock(&atomic_sim_lock); \ |
143 | | } |
144 | | IMPL_fallback_atomic_store(pvoid) |
145 | | |
146 | | # define ATOMIC_STORE(t, p, v, o) fallback_atomic_store_##t(p, v) |
147 | | |
148 | | /* |
149 | | * The fallbacks that follow don't need any per type implementation, as |
150 | | * they are designed for uint64_t only. If there comes a time when multiple |
151 | | * types need to be covered, it's relatively easy to refactor them the same |
152 | | * way as the fallbacks above. |
153 | | */ |
154 | | |
155 | | static ossl_inline uint64_t fallback_atomic_add_fetch(uint64_t *p, uint64_t v) |
156 | | { |
157 | | uint64_t ret; |
158 | | |
159 | | pthread_mutex_lock(&atomic_sim_lock); |
160 | | *p += v; |
161 | | ret = *p; |
162 | | pthread_mutex_unlock(&atomic_sim_lock); |
163 | | return ret; |
164 | | } |
165 | | |
166 | | # define ATOMIC_ADD_FETCH(p, v, o) fallback_atomic_add_fetch(p, v) |
167 | | |
168 | | static ossl_inline uint64_t fallback_atomic_sub_fetch(uint64_t *p, uint64_t v) |
169 | | { |
170 | | uint64_t ret; |
171 | | |
172 | | pthread_mutex_lock(&atomic_sim_lock); |
173 | | *p -= v; |
174 | | ret = *p; |
175 | | pthread_mutex_unlock(&atomic_sim_lock); |
176 | | return ret; |
177 | | } |
178 | | |
179 | | # define ATOMIC_SUB_FETCH(p, v, o) fallback_atomic_sub_fetch(p, v) |
180 | | # endif |
181 | | |
182 | | /* |
183 | | * This is the core of an rcu lock. It tracks the readers and writers for the |
184 | | * current quiescence point for a given lock. Users is the 64 bit value that |
185 | | * stores the READERS/ID as defined above |
186 | | * |
187 | | */ |
188 | | struct rcu_qp { |
189 | | uint64_t users; |
190 | | }; |
191 | | |
192 | | struct thread_qp { |
193 | | struct rcu_qp *qp; |
194 | | unsigned int depth; |
195 | | CRYPTO_RCU_LOCK *lock; |
196 | | }; |
197 | | |
198 | 364 | # define MAX_QPS 10 |
199 | | /* |
200 | | * This is the per thread tracking data |
201 | | * that is assigned to each thread participating |
202 | | * in an rcu qp |
203 | | * |
204 | | * qp points to the qp that it last acquired |
205 | | * |
206 | | */ |
207 | | struct rcu_thr_data { |
208 | | struct thread_qp thread_qps[MAX_QPS]; |
209 | | }; |
210 | | |
211 | | /* |
212 | | * This is the internal version of a CRYPTO_RCU_LOCK |
213 | | * it is cast from CRYPTO_RCU_LOCK |
214 | | */ |
215 | | struct rcu_lock_st { |
216 | | /* Callbacks to call for next ossl_synchronize_rcu */ |
217 | | struct rcu_cb_item *cb_items; |
218 | | |
219 | | /* The context we are being created against */ |
220 | | OSSL_LIB_CTX *ctx; |
221 | | |
222 | | /* Array of quiescent points for synchronization */ |
223 | | struct rcu_qp *qp_group; |
224 | | |
225 | | /* rcu generation counter for in-order retirement */ |
226 | | uint32_t id_ctr; |
227 | | |
228 | | /* Number of elements in qp_group array */ |
229 | | uint32_t group_count; |
230 | | |
231 | | /* Index of the current qp in the qp_group array */ |
232 | | uint32_t reader_idx; |
233 | | |
234 | | /* value of the next id_ctr value to be retired */ |
235 | | uint32_t next_to_retire; |
236 | | |
237 | | /* index of the next free rcu_qp in the qp_group */ |
238 | | uint32_t current_alloc_idx; |
239 | | |
240 | | /* number of qp's in qp_group array currently being retired */ |
241 | | uint32_t writers_alloced; |
242 | | |
243 | | /* lock protecting write side operations */ |
244 | | pthread_mutex_t write_lock; |
245 | | |
246 | | /* lock protecting updates to writers_alloced/current_alloc_idx */ |
247 | | pthread_mutex_t alloc_lock; |
248 | | |
249 | | /* signal to wake threads waiting on alloc_lock */ |
250 | | pthread_cond_t alloc_signal; |
251 | | |
252 | | /* lock to enforce in-order retirement */ |
253 | | pthread_mutex_t prior_lock; |
254 | | |
255 | | /* signal to wake threads waiting on prior_lock */ |
256 | | pthread_cond_t prior_signal; |
257 | | }; |
258 | | |
259 | | /* Read side acquisition of the current qp */ |
260 | | static struct rcu_qp *get_hold_current_qp(struct rcu_lock_st *lock) |
261 | 45 | { |
262 | 45 | uint32_t qp_idx; |
263 | | |
264 | | /* get the current qp index */ |
265 | 45 | for (;;) { |
266 | 45 | qp_idx = ATOMIC_LOAD_N(uint32_t, &lock->reader_idx, __ATOMIC_RELAXED); |
267 | | |
268 | | /* |
269 | | * Notes on use of __ATOMIC_ACQUIRE |
270 | | * We need to ensure the following: |
271 | | * 1) That subsequent operations aren't optimized by hoisting them above |
272 | | * this operation. Specifically, we don't want the below re-load of |
273 | | * qp_idx to get optimized away |
274 | | * 2) We want to ensure that any updating of reader_idx on the write side |
275 | | * of the lock is flushed from a local cpu cache so that we see any |
276 | | * updates prior to the load. This is a non-issue on cache coherent |
277 | | * systems like x86, but is relevant on other arches |
278 | | */ |
279 | 45 | ATOMIC_ADD_FETCH(&lock->qp_group[qp_idx].users, (uint64_t)1, |
280 | 45 | __ATOMIC_ACQUIRE); |
281 | | |
282 | | /* if the idx hasn't changed, we're good, else try again */ |
283 | 45 | if (qp_idx == ATOMIC_LOAD_N(uint32_t, &lock->reader_idx, |
284 | 45 | __ATOMIC_ACQUIRE)) |
285 | 45 | break; |
286 | | |
287 | 0 | ATOMIC_SUB_FETCH(&lock->qp_group[qp_idx].users, (uint64_t)1, |
288 | 0 | __ATOMIC_RELAXED); |
289 | 0 | } |
290 | | |
291 | 45 | return &lock->qp_group[qp_idx]; |
292 | 45 | } |
293 | | |
294 | | static void ossl_rcu_free_local_data(void *arg) |
295 | 3 | { |
296 | 3 | OSSL_LIB_CTX *ctx = arg; |
297 | 3 | CRYPTO_THREAD_LOCAL *lkey = ossl_lib_ctx_get_rcukey(ctx); |
298 | 3 | struct rcu_thr_data *data = CRYPTO_THREAD_get_local(lkey); |
299 | | |
300 | 3 | OPENSSL_free(data); |
301 | 3 | CRYPTO_THREAD_set_local(lkey, NULL); |
302 | 3 | } |
303 | | |
304 | | void ossl_rcu_read_lock(CRYPTO_RCU_LOCK *lock) |
305 | 29 | { |
306 | 29 | struct rcu_thr_data *data; |
307 | 29 | int i, available_qp = -1; |
308 | 29 | CRYPTO_THREAD_LOCAL *lkey = ossl_lib_ctx_get_rcukey(lock->ctx); |
309 | | |
310 | | /* |
311 | | * we're going to access current_qp here so ask the |
312 | | * processor to fetch it |
313 | | */ |
314 | 29 | data = CRYPTO_THREAD_get_local(lkey); |
315 | | |
316 | 29 | if (data == NULL) { |
317 | 2 | data = OPENSSL_zalloc(sizeof(*data)); |
318 | 2 | OPENSSL_assert(data != NULL); |
319 | 2 | CRYPTO_THREAD_set_local(lkey, data); |
320 | 2 | ossl_init_thread_start(NULL, lock->ctx, ossl_rcu_free_local_data); |
321 | 2 | } |
322 | | |
323 | 319 | for (i = 0; i < MAX_QPS; i++) { |
324 | 290 | if (data->thread_qps[i].qp == NULL && available_qp == -1) |
325 | 29 | available_qp = i; |
326 | | /* If we have a hold on this lock already, we're good */ |
327 | 290 | if (data->thread_qps[i].lock == lock) { |
328 | 0 | data->thread_qps[i].depth++; |
329 | 0 | return; |
330 | 0 | } |
331 | 290 | } |
332 | | |
333 | | /* |
334 | | * if we get here, then we don't have a hold on this lock yet |
335 | | */ |
336 | 29 | assert(available_qp != -1); |
337 | | |
338 | 29 | data->thread_qps[available_qp].qp = get_hold_current_qp(lock); |
339 | 29 | data->thread_qps[available_qp].depth = 1; |
340 | 29 | data->thread_qps[available_qp].lock = lock; |
341 | 29 | } |
342 | | |
343 | | void ossl_rcu_read_unlock(CRYPTO_RCU_LOCK *lock) |
344 | 45 | { |
345 | 45 | int i; |
346 | 45 | CRYPTO_THREAD_LOCAL *lkey = ossl_lib_ctx_get_rcukey(lock->ctx); |
347 | 45 | struct rcu_thr_data *data = CRYPTO_THREAD_get_local(lkey); |
348 | 45 | uint64_t ret; |
349 | | |
350 | 45 | assert(data != NULL); |
351 | | |
352 | 45 | for (i = 0; i < MAX_QPS; i++) { |
353 | 45 | if (data->thread_qps[i].lock == lock) { |
354 | | /* |
355 | | * we have to use __ATOMIC_RELEASE here |
356 | | * to ensure that all preceding read instructions complete |
357 | | * before the decrement is visible to ossl_synchronize_rcu |
358 | | */ |
359 | 45 | data->thread_qps[i].depth--; |
360 | 45 | if (data->thread_qps[i].depth == 0) { |
361 | 45 | ret = ATOMIC_SUB_FETCH(&data->thread_qps[i].qp->users, |
362 | 45 | (uint64_t)1, __ATOMIC_RELEASE); |
363 | 45 | OPENSSL_assert(ret != UINT64_MAX); |
364 | 45 | data->thread_qps[i].qp = NULL; |
365 | 45 | data->thread_qps[i].lock = NULL; |
366 | 45 | } |
367 | 45 | return; |
368 | 45 | } |
369 | 45 | } |
370 | | /* |
371 | | * If we get here, we're trying to unlock a lock that we never acquired - |
372 | | * that's fatal. |
373 | | */ |
374 | 45 | assert(0); |
375 | 0 | } |
376 | | |
377 | | /* |
378 | | * Write side allocation routine to get the current qp |
379 | | * and replace it with a new one |
380 | | */ |
381 | | static struct rcu_qp *update_qp(CRYPTO_RCU_LOCK *lock, uint32_t *curr_id) |
382 | 916 | { |
383 | 916 | uint32_t current_idx; |
384 | | |
385 | 916 | pthread_mutex_lock(&lock->alloc_lock); |
386 | | |
387 | | /* |
388 | | * we need at least one qp to be available with one |
389 | | * left over, so that readers can start working on |
390 | | * one that isn't yet being waited on |
391 | | */ |
392 | 916 | while (lock->group_count - lock->writers_alloced < 2) |
393 | | /* we have to wait for one to be free */ |
394 | 0 | pthread_cond_wait(&lock->alloc_signal, &lock->alloc_lock); |
395 | | |
396 | 916 | current_idx = lock->current_alloc_idx; |
397 | | |
398 | | /* Allocate the qp */ |
399 | 916 | lock->writers_alloced++; |
400 | | |
401 | | /* increment the allocation index */ |
402 | 916 | lock->current_alloc_idx = |
403 | 916 | (lock->current_alloc_idx + 1) % lock->group_count; |
404 | | |
405 | 916 | *curr_id = lock->id_ctr; |
406 | 916 | lock->id_ctr++; |
407 | | |
408 | | /* |
409 | | * make the current state of everything visible by this release |
410 | | * when get_hold_current_qp acquires the next qp |
411 | | */ |
412 | 916 | ATOMIC_STORE_N(uint32_t, &lock->reader_idx, lock->current_alloc_idx, |
413 | 916 | __ATOMIC_RELEASE); |
414 | | |
415 | | /* |
416 | | * this should make sure that the new value of reader_idx is visible in |
417 | | * get_hold_current_qp, directly after incrementing the users count |
418 | | */ |
419 | 916 | ATOMIC_ADD_FETCH(&lock->qp_group[current_idx].users, (uint64_t)0, |
420 | 916 | __ATOMIC_RELEASE); |
421 | | |
422 | | /* wake up any waiters */ |
423 | 916 | pthread_cond_signal(&lock->alloc_signal); |
424 | 916 | pthread_mutex_unlock(&lock->alloc_lock); |
425 | 916 | return &lock->qp_group[current_idx]; |
426 | 916 | } |
427 | | |
428 | | static void retire_qp(CRYPTO_RCU_LOCK *lock, struct rcu_qp *qp) |
429 | 916 | { |
430 | 916 | pthread_mutex_lock(&lock->alloc_lock); |
431 | 916 | lock->writers_alloced--; |
432 | 916 | pthread_cond_signal(&lock->alloc_signal); |
433 | 916 | pthread_mutex_unlock(&lock->alloc_lock); |
434 | 916 | } |
435 | | |
436 | | static struct rcu_qp *allocate_new_qp_group(CRYPTO_RCU_LOCK *lock, |
437 | | uint32_t count) |
438 | 438 | { |
439 | 438 | struct rcu_qp *new = |
440 | 438 | OPENSSL_zalloc(sizeof(*new) * count); |
441 | | |
442 | 438 | lock->group_count = count; |
443 | 438 | return new; |
444 | 438 | } |
445 | | |
446 | | void ossl_rcu_write_lock(CRYPTO_RCU_LOCK *lock) |
447 | 666 | { |
448 | 666 | pthread_mutex_lock(&lock->write_lock); |
449 | 666 | TSAN_FAKE_UNLOCK(&lock->write_lock); |
450 | 666 | } |
451 | | |
452 | | void ossl_rcu_write_unlock(CRYPTO_RCU_LOCK *lock) |
453 | 666 | { |
454 | 666 | TSAN_FAKE_LOCK(&lock->write_lock); |
455 | 666 | pthread_mutex_unlock(&lock->write_lock); |
456 | 666 | } |
457 | | |
458 | | void ossl_synchronize_rcu(CRYPTO_RCU_LOCK *lock) |
459 | 916 | { |
460 | 916 | struct rcu_qp *qp; |
461 | 916 | uint64_t count; |
462 | 916 | uint32_t curr_id; |
463 | 916 | struct rcu_cb_item *cb_items, *tmpcb; |
464 | | |
465 | 916 | pthread_mutex_lock(&lock->write_lock); |
466 | 916 | cb_items = lock->cb_items; |
467 | 916 | lock->cb_items = NULL; |
468 | 916 | pthread_mutex_unlock(&lock->write_lock); |
469 | | |
470 | 916 | qp = update_qp(lock, &curr_id); |
471 | | |
472 | | /* retire in order */ |
473 | 916 | pthread_mutex_lock(&lock->prior_lock); |
474 | 916 | while (lock->next_to_retire != curr_id) |
475 | 0 | pthread_cond_wait(&lock->prior_signal, &lock->prior_lock); |
476 | | |
477 | | /* |
478 | | * wait for the reader count to reach zero |
479 | | * Note the use of __ATOMIC_ACQUIRE here to ensure that any |
480 | | * prior __ATOMIC_RELEASE write operation in ossl_rcu_read_unlock |
481 | | * is visible prior to our read |
482 | | * however this is likely just necessary to silence a tsan warning |
483 | | * because the read side should not do any write operation |
484 | | * outside the atomic itself |
485 | | */ |
486 | 916 | do { |
487 | 916 | count = ATOMIC_LOAD_N(uint64_t, &qp->users, __ATOMIC_ACQUIRE); |
488 | 916 | } while (count != (uint64_t)0); |
489 | | |
490 | 916 | lock->next_to_retire++; |
491 | 916 | pthread_cond_broadcast(&lock->prior_signal); |
492 | 916 | pthread_mutex_unlock(&lock->prior_lock); |
493 | | |
494 | 916 | retire_qp(lock, qp); |
495 | | |
496 | | /* handle any callbacks that we have */ |
497 | 1.09k | while (cb_items != NULL) { |
498 | 179 | tmpcb = cb_items; |
499 | 179 | cb_items = cb_items->next; |
500 | 179 | tmpcb->fn(tmpcb->data); |
501 | 179 | OPENSSL_free(tmpcb); |
502 | 179 | } |
503 | 916 | } |
504 | | |
505 | | /* |
506 | | * Note: This call assumes its made under the protection of |
507 | | * ossl_rcu_write_lock |
508 | | */ |
509 | | int ossl_rcu_call(CRYPTO_RCU_LOCK *lock, rcu_cb_fn cb, void *data) |
510 | 179 | { |
511 | 179 | struct rcu_cb_item *new = |
512 | 179 | OPENSSL_zalloc(sizeof(*new)); |
513 | | |
514 | 179 | if (new == NULL) |
515 | 0 | return 0; |
516 | | |
517 | 179 | new->data = data; |
518 | 179 | new->fn = cb; |
519 | | |
520 | 179 | new->next = lock->cb_items; |
521 | 179 | lock->cb_items = new; |
522 | | |
523 | 179 | return 1; |
524 | 179 | } |
525 | | |
526 | | void *ossl_rcu_uptr_deref(void **p) |
527 | 54.9M | { |
528 | 54.9M | return ATOMIC_LOAD_N(pvoid, p, __ATOMIC_ACQUIRE); |
529 | 54.9M | } |
530 | | |
531 | | void ossl_rcu_assign_uptr(void **p, void **v) |
532 | 28.7k | { |
533 | 28.7k | ATOMIC_STORE(pvoid, p, v, __ATOMIC_RELEASE); |
534 | 28.7k | } |
535 | | |
536 | | CRYPTO_RCU_LOCK *ossl_rcu_lock_new(int num_writers, OSSL_LIB_CTX *ctx) |
537 | 438 | { |
538 | 438 | struct rcu_lock_st *new; |
539 | | |
540 | | /* |
541 | | * We need a minimum of 2 qp's |
542 | | */ |
543 | 438 | if (num_writers < 2) |
544 | 438 | num_writers = 2; |
545 | | |
546 | 438 | ctx = ossl_lib_ctx_get_concrete(ctx); |
547 | 438 | if (ctx == NULL) |
548 | 0 | return 0; |
549 | | |
550 | 438 | new = OPENSSL_zalloc(sizeof(*new)); |
551 | 438 | if (new == NULL) |
552 | 0 | return NULL; |
553 | | |
554 | 438 | new->ctx = ctx; |
555 | 438 | pthread_mutex_init(&new->write_lock, NULL); |
556 | 438 | pthread_mutex_init(&new->prior_lock, NULL); |
557 | 438 | pthread_mutex_init(&new->alloc_lock, NULL); |
558 | 438 | pthread_cond_init(&new->prior_signal, NULL); |
559 | 438 | pthread_cond_init(&new->alloc_signal, NULL); |
560 | | |
561 | 438 | new->qp_group = allocate_new_qp_group(new, num_writers); |
562 | 438 | if (new->qp_group == NULL) { |
563 | 0 | OPENSSL_free(new); |
564 | 0 | new = NULL; |
565 | 0 | } |
566 | | |
567 | 438 | return new; |
568 | 438 | } |
569 | | |
570 | | void ossl_rcu_lock_free(CRYPTO_RCU_LOCK *lock) |
571 | 350 | { |
572 | 350 | struct rcu_lock_st *rlock = (struct rcu_lock_st *)lock; |
573 | | |
574 | 350 | if (lock == NULL) |
575 | 0 | return; |
576 | | |
577 | | /* make sure we're synchronized */ |
578 | 350 | ossl_synchronize_rcu(rlock); |
579 | | |
580 | 350 | OPENSSL_free(rlock->qp_group); |
581 | | /* There should only be a single qp left now */ |
582 | 350 | OPENSSL_free(rlock); |
583 | 350 | } |
584 | | |
585 | | CRYPTO_RWLOCK *CRYPTO_THREAD_lock_new(void) |
586 | 9.94M | { |
587 | 9.94M | # ifdef USE_RWLOCK |
588 | 9.94M | CRYPTO_RWLOCK *lock; |
589 | | |
590 | 9.94M | if ((lock = OPENSSL_zalloc(sizeof(pthread_rwlock_t))) == NULL) |
591 | | /* Don't set error, to avoid recursion blowup. */ |
592 | 0 | return NULL; |
593 | | |
594 | 9.94M | if (pthread_rwlock_init(lock, NULL) != 0) { |
595 | 0 | OPENSSL_free(lock); |
596 | 0 | return NULL; |
597 | 0 | } |
598 | | # else |
599 | | pthread_mutexattr_t attr; |
600 | | CRYPTO_RWLOCK *lock; |
601 | | |
602 | | if ((lock = OPENSSL_zalloc(sizeof(pthread_mutex_t))) == NULL) |
603 | | /* Don't set error, to avoid recursion blowup. */ |
604 | | return NULL; |
605 | | |
606 | | /* |
607 | | * We don't use recursive mutexes, but try to catch errors if we do. |
608 | | */ |
609 | | pthread_mutexattr_init(&attr); |
610 | | # if !defined (__TANDEM) && !defined (_SPT_MODEL_) |
611 | | # if !defined(NDEBUG) && !defined(OPENSSL_NO_MUTEX_ERRORCHECK) |
612 | | pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_ERRORCHECK); |
613 | | # endif |
614 | | # else |
615 | | /* The SPT Thread Library does not define MUTEX attributes. */ |
616 | | # endif |
617 | | |
618 | | if (pthread_mutex_init(lock, &attr) != 0) { |
619 | | pthread_mutexattr_destroy(&attr); |
620 | | OPENSSL_free(lock); |
621 | | return NULL; |
622 | | } |
623 | | |
624 | | pthread_mutexattr_destroy(&attr); |
625 | | # endif |
626 | | |
627 | 9.94M | return lock; |
628 | 9.94M | } |
629 | | |
630 | | __owur int CRYPTO_THREAD_read_lock(CRYPTO_RWLOCK *lock) |
631 | 1.09G | { |
632 | 1.09G | # ifdef USE_RWLOCK |
633 | 1.09G | if (pthread_rwlock_rdlock(lock) != 0) |
634 | 0 | return 0; |
635 | | # else |
636 | | if (pthread_mutex_lock(lock) != 0) { |
637 | | assert(errno != EDEADLK && errno != EBUSY); |
638 | | return 0; |
639 | | } |
640 | | # endif |
641 | | |
642 | 1.09G | return 1; |
643 | 1.09G | } |
644 | | |
645 | | __owur int CRYPTO_THREAD_write_lock(CRYPTO_RWLOCK *lock) |
646 | 77.7M | { |
647 | 77.7M | # ifdef USE_RWLOCK |
648 | 77.7M | if (pthread_rwlock_wrlock(lock) != 0) |
649 | 0 | return 0; |
650 | | # else |
651 | | if (pthread_mutex_lock(lock) != 0) { |
652 | | assert(errno != EDEADLK && errno != EBUSY); |
653 | | return 0; |
654 | | } |
655 | | # endif |
656 | | |
657 | 77.7M | return 1; |
658 | 77.7M | } |
659 | | |
660 | | int CRYPTO_THREAD_unlock(CRYPTO_RWLOCK *lock) |
661 | 1.26G | { |
662 | 1.26G | # ifdef USE_RWLOCK |
663 | 1.26G | if (pthread_rwlock_unlock(lock) != 0) |
664 | 0 | return 0; |
665 | | # else |
666 | | if (pthread_mutex_unlock(lock) != 0) { |
667 | | assert(errno != EPERM); |
668 | | return 0; |
669 | | } |
670 | | # endif |
671 | | |
672 | 1.26G | return 1; |
673 | 1.26G | } |
674 | | |
675 | | void CRYPTO_THREAD_lock_free(CRYPTO_RWLOCK *lock) |
676 | 9.94M | { |
677 | 9.94M | if (lock == NULL) |
678 | 2.20k | return; |
679 | | |
680 | 9.94M | # ifdef USE_RWLOCK |
681 | 9.94M | pthread_rwlock_destroy(lock); |
682 | | # else |
683 | | pthread_mutex_destroy(lock); |
684 | | # endif |
685 | 9.94M | OPENSSL_free(lock); |
686 | | |
687 | 9.94M | return; |
688 | 9.94M | } |
689 | | |
690 | | int CRYPTO_THREAD_run_once(CRYPTO_ONCE *once, void (*init)(void)) |
691 | 2.92G | { |
692 | 2.92G | if (pthread_once(once, init) != 0) |
693 | 0 | return 0; |
694 | | |
695 | 2.92G | return 1; |
696 | 2.92G | } |
697 | | |
698 | | int CRYPTO_THREAD_init_local(CRYPTO_THREAD_LOCAL *key, void (*cleanup)(void *)) |
699 | 1.89k | { |
700 | 1.89k | if (pthread_key_create(key, cleanup) != 0) |
701 | 0 | return 0; |
702 | | |
703 | 1.89k | return 1; |
704 | 1.89k | } |
705 | | |
706 | | void *CRYPTO_THREAD_get_local(CRYPTO_THREAD_LOCAL *key) |
707 | 1.83G | { |
708 | 1.83G | return pthread_getspecific(*key); |
709 | 1.83G | } |
710 | | |
711 | | int CRYPTO_THREAD_set_local(CRYPTO_THREAD_LOCAL *key, void *val) |
712 | 1.91k | { |
713 | 1.91k | if (pthread_setspecific(*key, val) != 0) |
714 | 0 | return 0; |
715 | | |
716 | 1.91k | return 1; |
717 | 1.91k | } |
718 | | |
719 | | int CRYPTO_THREAD_cleanup_local(CRYPTO_THREAD_LOCAL *key) |
720 | 1.57k | { |
721 | 1.57k | if (pthread_key_delete(*key) != 0) |
722 | 0 | return 0; |
723 | | |
724 | 1.57k | return 1; |
725 | 1.57k | } |
726 | | |
727 | | CRYPTO_THREAD_ID CRYPTO_THREAD_get_current_id(void) |
728 | 183k | { |
729 | 183k | return pthread_self(); |
730 | 183k | } |
731 | | |
732 | | int CRYPTO_THREAD_compare_id(CRYPTO_THREAD_ID a, CRYPTO_THREAD_ID b) |
733 | 11.8k | { |
734 | 11.8k | return pthread_equal(a, b); |
735 | 11.8k | } |
736 | | |
737 | | int CRYPTO_atomic_add(int *val, int amount, int *ret, CRYPTO_RWLOCK *lock) |
738 | 10.9M | { |
739 | 10.9M | # if defined(__GNUC__) && defined(__ATOMIC_ACQ_REL) && !defined(BROKEN_CLANG_ATOMICS) |
740 | 10.9M | if (__atomic_is_lock_free(sizeof(*val), val)) { |
741 | 10.9M | *ret = __atomic_add_fetch(val, amount, __ATOMIC_ACQ_REL); |
742 | 10.9M | return 1; |
743 | 10.9M | } |
744 | | # elif defined(__sun) && (defined(__SunOS_5_10) || defined(__SunOS_5_11)) |
745 | | /* This will work for all future Solaris versions. */ |
746 | | if (ret != NULL) { |
747 | | *ret = atomic_add_int_nv((volatile unsigned int *)val, amount); |
748 | | return 1; |
749 | | } |
750 | | # endif |
751 | 0 | if (lock == NULL || !CRYPTO_THREAD_write_lock(lock)) |
752 | 0 | return 0; |
753 | | |
754 | 0 | *val += amount; |
755 | 0 | *ret = *val; |
756 | |
|
757 | 0 | if (!CRYPTO_THREAD_unlock(lock)) |
758 | 0 | return 0; |
759 | | |
760 | 0 | return 1; |
761 | 0 | } |
762 | | |
763 | | int CRYPTO_atomic_add64(uint64_t *val, uint64_t op, uint64_t *ret, |
764 | | CRYPTO_RWLOCK *lock) |
765 | 0 | { |
766 | 0 | # if defined(__GNUC__) && defined(__ATOMIC_ACQ_REL) && !defined(BROKEN_CLANG_ATOMICS) |
767 | 0 | if (__atomic_is_lock_free(sizeof(*val), val)) { |
768 | 0 | *ret = __atomic_add_fetch(val, op, __ATOMIC_ACQ_REL); |
769 | 0 | return 1; |
770 | 0 | } |
771 | | # elif defined(__sun) && (defined(__SunOS_5_10) || defined(__SunOS_5_11)) |
772 | | /* This will work for all future Solaris versions. */ |
773 | | if (ret != NULL) { |
774 | | *ret = atomic_add_64_nv(val, op); |
775 | | return 1; |
776 | | } |
777 | | # endif |
778 | 0 | if (lock == NULL || !CRYPTO_THREAD_write_lock(lock)) |
779 | 0 | return 0; |
780 | 0 | *val += op; |
781 | 0 | *ret = *val; |
782 | |
|
783 | 0 | if (!CRYPTO_THREAD_unlock(lock)) |
784 | 0 | return 0; |
785 | | |
786 | 0 | return 1; |
787 | 0 | } |
788 | | |
789 | | int CRYPTO_atomic_and(uint64_t *val, uint64_t op, uint64_t *ret, |
790 | | CRYPTO_RWLOCK *lock) |
791 | 0 | { |
792 | 0 | # if defined(__GNUC__) && defined(__ATOMIC_ACQ_REL) && !defined(BROKEN_CLANG_ATOMICS) |
793 | 0 | if (__atomic_is_lock_free(sizeof(*val), val)) { |
794 | 0 | *ret = __atomic_and_fetch(val, op, __ATOMIC_ACQ_REL); |
795 | 0 | return 1; |
796 | 0 | } |
797 | | # elif defined(__sun) && (defined(__SunOS_5_10) || defined(__SunOS_5_11)) |
798 | | /* This will work for all future Solaris versions. */ |
799 | | if (ret != NULL) { |
800 | | *ret = atomic_and_64_nv(val, op); |
801 | | return 1; |
802 | | } |
803 | | # endif |
804 | 0 | if (lock == NULL || !CRYPTO_THREAD_write_lock(lock)) |
805 | 0 | return 0; |
806 | 0 | *val &= op; |
807 | 0 | *ret = *val; |
808 | |
|
809 | 0 | if (!CRYPTO_THREAD_unlock(lock)) |
810 | 0 | return 0; |
811 | | |
812 | 0 | return 1; |
813 | 0 | } |
814 | | |
815 | | int CRYPTO_atomic_or(uint64_t *val, uint64_t op, uint64_t *ret, |
816 | | CRYPTO_RWLOCK *lock) |
817 | 644 | { |
818 | 644 | # if defined(__GNUC__) && defined(__ATOMIC_ACQ_REL) && !defined(BROKEN_CLANG_ATOMICS) |
819 | 644 | if (__atomic_is_lock_free(sizeof(*val), val)) { |
820 | 644 | *ret = __atomic_or_fetch(val, op, __ATOMIC_ACQ_REL); |
821 | 644 | return 1; |
822 | 644 | } |
823 | | # elif defined(__sun) && (defined(__SunOS_5_10) || defined(__SunOS_5_11)) |
824 | | /* This will work for all future Solaris versions. */ |
825 | | if (ret != NULL) { |
826 | | *ret = atomic_or_64_nv(val, op); |
827 | | return 1; |
828 | | } |
829 | | # endif |
830 | 0 | if (lock == NULL || !CRYPTO_THREAD_write_lock(lock)) |
831 | 0 | return 0; |
832 | 0 | *val |= op; |
833 | 0 | *ret = *val; |
834 | |
|
835 | 0 | if (!CRYPTO_THREAD_unlock(lock)) |
836 | 0 | return 0; |
837 | | |
838 | 0 | return 1; |
839 | 0 | } |
840 | | |
841 | | int CRYPTO_atomic_load(uint64_t *val, uint64_t *ret, CRYPTO_RWLOCK *lock) |
842 | 2.43G | { |
843 | 2.43G | # if defined(__GNUC__) && defined(__ATOMIC_ACQ_REL) && !defined(BROKEN_CLANG_ATOMICS) |
844 | 2.43G | if (__atomic_is_lock_free(sizeof(*val), val)) { |
845 | 2.43G | __atomic_load(val, ret, __ATOMIC_ACQUIRE); |
846 | 2.43G | return 1; |
847 | 2.43G | } |
848 | | # elif defined(__sun) && (defined(__SunOS_5_10) || defined(__SunOS_5_11)) |
849 | | /* This will work for all future Solaris versions. */ |
850 | | if (ret != NULL) { |
851 | | *ret = atomic_or_64_nv(val, 0); |
852 | | return 1; |
853 | | } |
854 | | # endif |
855 | 0 | if (lock == NULL || !CRYPTO_THREAD_read_lock(lock)) |
856 | 0 | return 0; |
857 | 0 | *ret = *val; |
858 | 0 | if (!CRYPTO_THREAD_unlock(lock)) |
859 | 0 | return 0; |
860 | | |
861 | 0 | return 1; |
862 | 0 | } |
863 | | |
864 | | int CRYPTO_atomic_store(uint64_t *dst, uint64_t val, CRYPTO_RWLOCK *lock) |
865 | 28.2k | { |
866 | 28.2k | # if defined(__GNUC__) && defined(__ATOMIC_ACQ_REL) && !defined(BROKEN_CLANG_ATOMICS) |
867 | 28.2k | if (__atomic_is_lock_free(sizeof(*dst), dst)) { |
868 | 28.2k | __atomic_store(dst, &val, __ATOMIC_RELEASE); |
869 | 28.2k | return 1; |
870 | 28.2k | } |
871 | | # elif defined(__sun) && (defined(__SunOS_5_10) || defined(__SunOS_5_11)) |
872 | | /* This will work for all future Solaris versions. */ |
873 | | if (dst != NULL) { |
874 | | atomic_swap_64(dst, val); |
875 | | return 1; |
876 | | } |
877 | | # endif |
878 | 0 | if (lock == NULL || !CRYPTO_THREAD_write_lock(lock)) |
879 | 0 | return 0; |
880 | 0 | *dst = val; |
881 | 0 | if (!CRYPTO_THREAD_unlock(lock)) |
882 | 0 | return 0; |
883 | | |
884 | 0 | return 1; |
885 | 0 | } |
886 | | |
887 | | int CRYPTO_atomic_load_int(int *val, int *ret, CRYPTO_RWLOCK *lock) |
888 | 0 | { |
889 | 0 | # if defined(__GNUC__) && defined(__ATOMIC_ACQ_REL) && !defined(BROKEN_CLANG_ATOMICS) |
890 | 0 | if (__atomic_is_lock_free(sizeof(*val), val)) { |
891 | 0 | __atomic_load(val, ret, __ATOMIC_ACQUIRE); |
892 | 0 | return 1; |
893 | 0 | } |
894 | | # elif defined(__sun) && (defined(__SunOS_5_10) || defined(__SunOS_5_11)) |
895 | | /* This will work for all future Solaris versions. */ |
896 | | if (ret != NULL) { |
897 | | *ret = (int)atomic_or_uint_nv((unsigned int *)val, 0); |
898 | | return 1; |
899 | | } |
900 | | # endif |
901 | 0 | if (lock == NULL || !CRYPTO_THREAD_read_lock(lock)) |
902 | 0 | return 0; |
903 | 0 | *ret = *val; |
904 | 0 | if (!CRYPTO_THREAD_unlock(lock)) |
905 | 0 | return 0; |
906 | | |
907 | 0 | return 1; |
908 | 0 | } |
909 | | |
910 | | # ifndef FIPS_MODULE |
911 | | int openssl_init_fork_handlers(void) |
912 | 0 | { |
913 | 0 | return 1; |
914 | 0 | } |
915 | | # endif /* FIPS_MODULE */ |
916 | | |
917 | | int openssl_get_fork_id(void) |
918 | 74.8k | { |
919 | 74.8k | return getpid(); |
920 | 74.8k | } |
921 | | #endif |