Invoke tear_down when exiting test_encode_tls_sct() prematurely
[openssl.git] / crypto / threads_win.c
1 /*
2  * Copyright 2016-2024 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 #if defined(_WIN32)
11 # include <windows.h>
12 # if defined(_WIN32_WINNT) && _WIN32_WINNT >= 0x600
13 #  define USE_RWLOCK
14 # endif
15 #endif
16 #include <assert.h>
17
18 /*
19  * VC++ 2008 or earlier x86 compilers do not have an inline implementation
20  * of InterlockedOr64 for 32bit and will fail to run on Windows XP 32bit.
21  * https://docs.microsoft.com/en-us/cpp/intrinsics/interlockedor-intrinsic-functions#requirements
22  * To work around this problem, we implement a manual locking mechanism for
23  * only VC++ 2008 or earlier x86 compilers.
24  */
25
26 #if (defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER <= 1600)
27 # define NO_INTERLOCKEDOR64
28 #endif
29
30 #include <openssl/crypto.h>
31 #include <crypto/cryptlib.h>
32 #include "internal/common.h"
33 #include "internal/thread_arch.h"
34 #include "internal/rcu.h"
35 #include "rcu_internal.h"
36
37 #if defined(OPENSSL_THREADS) && !defined(CRYPTO_TDEBUG) && defined(OPENSSL_SYS_WINDOWS)
38
39 # ifdef USE_RWLOCK
40 typedef struct {
41     SRWLOCK lock;
42     int exclusive;
43 } CRYPTO_win_rwlock;
44 # endif
45
46 # define READER_SHIFT 0
47 # define ID_SHIFT 32 
48 # define READER_SIZE 32 
49 # define ID_SIZE 32 
50
51 # define READER_MASK     (((LONG64)1 << READER_SIZE)-1)
52 # define ID_MASK         (((LONG64)1 << ID_SIZE)-1)
53 # define READER_COUNT(x) (((LONG64)(x) >> READER_SHIFT) & READER_MASK)
54 # define ID_VAL(x)       (((LONG64)(x) >> ID_SHIFT) & ID_MASK)
55 # define VAL_READER      ((LONG64)1 << READER_SHIFT)
56 # define VAL_ID(x)       ((LONG64)x << ID_SHIFT)
57
58 /*
59  * This defines a quescent point (qp)
60  * This is the barrier beyond which a writer
61  * must wait before freeing data that was
62  * atomically updated
63  */
64 struct rcu_qp {
65     volatile LONG64 users;
66 };
67
68 struct thread_qp {
69     struct rcu_qp *qp;
70     unsigned int depth;
71     CRYPTO_RCU_LOCK *lock;
72 };
73
74 #define MAX_QPS 10
75 /*
76  * This is the per thread tracking data
77  * that is assigned to each thread participating
78  * in an rcu qp
79  *
80  * qp points to the qp that it last acquired
81  *
82  */
83 struct rcu_thr_data {
84     struct thread_qp thread_qps[MAX_QPS];
85 };
86
87 /*
88  * This is the internal version of a CRYPTO_RCU_LOCK
89  * it is cast from CRYPTO_RCU_LOCK
90  */
91 struct rcu_lock_st {
92     struct rcu_cb_item *cb_items;
93     OSSL_LIB_CTX *ctx;
94     uint32_t id_ctr;
95     struct rcu_qp *qp_group;
96     size_t group_count;
97     uint32_t next_to_retire;
98     volatile long int reader_idx;
99     uint32_t current_alloc_idx;
100     uint32_t writers_alloced;
101     CRYPTO_MUTEX *write_lock;
102     CRYPTO_MUTEX *alloc_lock;
103     CRYPTO_CONDVAR *alloc_signal;
104     CRYPTO_MUTEX *prior_lock;
105     CRYPTO_CONDVAR *prior_signal;
106 };
107
108 static struct rcu_qp *allocate_new_qp_group(struct rcu_lock_st *lock,
109                                             int count)
110 {
111     struct rcu_qp *new =
112         OPENSSL_zalloc(sizeof(*new) * count);
113
114     lock->group_count = count;
115     return new;
116 }
117
118 CRYPTO_RCU_LOCK *ossl_rcu_lock_new(int num_writers, OSSL_LIB_CTX *ctx)
119 {
120     struct rcu_lock_st *new;
121
122     if (num_writers < 1)
123         num_writers = 1;
124
125     ctx = ossl_lib_ctx_get_concrete(ctx);
126     if (ctx == NULL)
127         return 0;
128
129     new = OPENSSL_zalloc(sizeof(*new));
130
131     if (new == NULL)
132         return NULL;
133
134     new->ctx = ctx;
135     new->write_lock = ossl_crypto_mutex_new();
136     new->alloc_signal = ossl_crypto_condvar_new();
137     new->prior_signal = ossl_crypto_condvar_new();
138     new->alloc_lock = ossl_crypto_mutex_new();
139     new->prior_lock = ossl_crypto_mutex_new();
140     new->qp_group = allocate_new_qp_group(new, num_writers + 1);
141     if (new->qp_group == NULL
142         || new->alloc_signal == NULL
143         || new->prior_signal == NULL
144         || new->write_lock == NULL
145         || new->alloc_lock == NULL
146         || new->prior_lock == NULL) {
147         OPENSSL_free(new->qp_group);
148         ossl_crypto_condvar_free(&new->alloc_signal);
149         ossl_crypto_condvar_free(&new->prior_signal);
150         ossl_crypto_mutex_free(&new->alloc_lock);
151         ossl_crypto_mutex_free(&new->prior_lock);
152         ossl_crypto_mutex_free(&new->write_lock);
153         OPENSSL_free(new);
154         new = NULL;
155     }
156     return new;
157
158 }
159
160 void ossl_rcu_lock_free(CRYPTO_RCU_LOCK *lock)
161 {
162     OPENSSL_free(lock->qp_group);
163     ossl_crypto_condvar_free(&lock->alloc_signal);
164     ossl_crypto_condvar_free(&lock->prior_signal);
165     ossl_crypto_mutex_free(&lock->alloc_lock);
166     ossl_crypto_mutex_free(&lock->prior_lock);
167     ossl_crypto_mutex_free(&lock->write_lock);
168     OPENSSL_free(lock);
169 }
170
171 static inline struct rcu_qp *get_hold_current_qp(CRYPTO_RCU_LOCK *lock)
172 {
173     uint32_t qp_idx;
174
175     /* get the current qp index */
176     for (;;) {
177         qp_idx = InterlockedOr(&lock->reader_idx, 0);
178         InterlockedAdd64(&lock->qp_group[qp_idx].users, VAL_READER);
179         if (qp_idx == InterlockedOr(&lock->reader_idx, 0))
180             break;
181         InterlockedAdd64(&lock->qp_group[qp_idx].users, -VAL_READER);
182     }
183
184     return &lock->qp_group[qp_idx];
185 }
186
187 static void ossl_rcu_free_local_data(void *arg)
188 {
189     OSSL_LIB_CTX *ctx = arg;
190     CRYPTO_THREAD_LOCAL *lkey = ossl_lib_ctx_get_rcukey(ctx);
191     struct rcu_thr_data *data = CRYPTO_THREAD_get_local(lkey);
192     OPENSSL_free(data);
193 }
194
195 void ossl_rcu_read_lock(CRYPTO_RCU_LOCK *lock)
196 {
197     struct rcu_thr_data *data;
198     int i;
199     int available_qp = -1;
200     CRYPTO_THREAD_LOCAL *lkey = ossl_lib_ctx_get_rcukey(lock->ctx);
201
202     /*
203      * we're going to access current_qp here so ask the
204      * processor to fetch it
205      */
206     data = CRYPTO_THREAD_get_local(lkey);
207
208     if (data == NULL) {
209         data = OPENSSL_zalloc(sizeof(*data));
210         OPENSSL_assert(data != NULL);
211         CRYPTO_THREAD_set_local(lkey, data);
212         ossl_init_thread_start(NULL, lock->ctx, ossl_rcu_free_local_data);
213     }
214
215     for (i = 0; i < MAX_QPS; i++) {
216         if (data->thread_qps[i].qp == NULL && available_qp == -1)
217             available_qp = i;
218         /* If we have a hold on this lock already, we're good */
219         if (data->thread_qps[i].lock == lock)
220             return;
221     }
222
223     /*
224      * if we get here, then we don't have a hold on this lock yet
225      */
226     assert(available_qp != -1);
227
228     data->thread_qps[available_qp].qp = get_hold_current_qp(lock);
229     data->thread_qps[available_qp].depth = 1;
230     data->thread_qps[available_qp].lock = lock;
231 }
232
233 void ossl_rcu_write_lock(CRYPTO_RCU_LOCK *lock)
234 {
235     ossl_crypto_mutex_lock(lock->write_lock);
236 }
237
238 void ossl_rcu_write_unlock(CRYPTO_RCU_LOCK *lock)
239 {
240     ossl_crypto_mutex_unlock(lock->write_lock);
241 }
242
243 void ossl_rcu_read_unlock(CRYPTO_RCU_LOCK *lock)
244 {
245     CRYPTO_THREAD_LOCAL *lkey = ossl_lib_ctx_get_rcukey(lock->ctx);
246     struct rcu_thr_data *data = CRYPTO_THREAD_get_local(lkey);
247     int i;
248     LONG64 ret;
249
250     assert(data != NULL);
251
252     for (i = 0; i < MAX_QPS; i++) {
253         if (data->thread_qps[i].lock == lock) {
254             data->thread_qps[i].depth--;
255             if (data->thread_qps[i].depth == 0) {
256                 ret = InterlockedAdd64(&data->thread_qps[i].qp->users, -VAL_READER);
257                 OPENSSL_assert(ret >= 0);
258                 data->thread_qps[i].qp = NULL;
259                 data->thread_qps[i].lock = NULL;
260             }
261             return;
262         }
263     }
264 }
265
266 static struct rcu_qp *update_qp(CRYPTO_RCU_LOCK *lock)
267 {
268     uint64_t new_id;
269     uint32_t current_idx;
270     uint32_t tmp;
271
272     ossl_crypto_mutex_lock(lock->alloc_lock);
273     /*
274      * we need at least one qp to be available with one
275      * left over, so that readers can start working on
276      * one that isn't yet being waited on
277      */
278     while (lock->group_count - lock->writers_alloced < 2)
279         ossl_crypto_condvar_wait(lock->alloc_signal, lock->alloc_lock);
280
281     current_idx = lock->current_alloc_idx;
282     /* Allocate the qp */
283     lock->writers_alloced++;
284
285     /* increment the allocation index */
286     lock->current_alloc_idx =
287         (lock->current_alloc_idx + 1) % lock->group_count;
288
289     /* get and insert a new id */
290     new_id = lock->id_ctr;
291     lock->id_ctr++;
292
293     new_id = VAL_ID(new_id);
294     InterlockedAnd64(&lock->qp_group[current_idx].users, ID_MASK);
295     InterlockedAdd64(&lock->qp_group[current_idx].users, new_id);
296
297     /* update the reader index to be the prior qp */
298     tmp = lock->current_alloc_idx;
299     InterlockedExchange(&lock->reader_idx, tmp);
300
301     /* wake up any waiters */
302     ossl_crypto_condvar_broadcast(lock->alloc_signal);
303     ossl_crypto_mutex_unlock(lock->alloc_lock);
304     return &lock->qp_group[current_idx];
305 }
306
307 static void retire_qp(CRYPTO_RCU_LOCK *lock,
308                       struct rcu_qp *qp)
309 {
310     ossl_crypto_mutex_lock(lock->alloc_lock);
311     lock->writers_alloced--;
312     ossl_crypto_condvar_broadcast(lock->alloc_signal);
313     ossl_crypto_mutex_unlock(lock->alloc_lock);
314 }
315
316
317 void ossl_synchronize_rcu(CRYPTO_RCU_LOCK *lock)
318 {
319     struct rcu_qp *qp;
320     uint64_t count;
321     struct rcu_cb_item *cb_items, *tmpcb;
322
323     /* before we do anything else, lets grab the cb list */
324     cb_items = InterlockedExchangePointer((void * volatile *)&lock->cb_items, NULL);
325
326     qp = update_qp(lock);
327
328     /* wait for the reader count to reach zero */
329     do {
330         count = InterlockedOr64(&qp->users, 0);
331     } while (READER_COUNT(count) != 0);
332
333     /* retire in order */
334     ossl_crypto_mutex_lock(lock->prior_lock);
335     while (lock->next_to_retire != ID_VAL(count))
336         ossl_crypto_condvar_wait(lock->prior_signal, lock->prior_lock);
337
338     lock->next_to_retire++;
339     ossl_crypto_condvar_broadcast(lock->prior_signal);
340     ossl_crypto_mutex_unlock(lock->prior_lock);
341
342     retire_qp(lock, qp);
343
344     /* handle any callbacks that we have */
345     while (cb_items != NULL) {
346         tmpcb = cb_items;
347         cb_items = cb_items->next;
348         tmpcb->fn(tmpcb->data);
349         OPENSSL_free(tmpcb);
350     }
351
352     /* and we're done */
353     return;
354
355 }
356
357 int ossl_rcu_call(CRYPTO_RCU_LOCK *lock, rcu_cb_fn cb, void *data)
358 {
359     struct rcu_cb_item *new;
360     struct rcu_cb_item *prev;
361
362     new = OPENSSL_zalloc(sizeof(struct rcu_cb_item));
363     if (new == NULL)
364         return 0;
365     prev = new;
366     new->data = data;
367     new->fn = cb;
368
369     InterlockedExchangePointer((void * volatile *)&lock->cb_items, prev);
370     new->next = prev;
371     return 1;
372 }
373
374 void *ossl_rcu_uptr_deref(void **p)
375 {
376     return (void *)*p;
377 }
378
379 void ossl_rcu_assign_uptr(void **p, void **v)
380 {
381     InterlockedExchangePointer((void * volatile *)p, (void *)*v);
382 }
383
384
385 CRYPTO_RWLOCK *CRYPTO_THREAD_lock_new(void)
386 {
387     CRYPTO_RWLOCK *lock;
388 # ifdef USE_RWLOCK
389     CRYPTO_win_rwlock *rwlock;
390
391     if ((lock = OPENSSL_zalloc(sizeof(CRYPTO_win_rwlock))) == NULL)
392         /* Don't set error, to avoid recursion blowup. */
393         return NULL;
394     rwlock = lock;
395     InitializeSRWLock(&rwlock->lock);
396 # else
397
398     if ((lock = OPENSSL_zalloc(sizeof(CRITICAL_SECTION))) == NULL)
399         /* Don't set error, to avoid recursion blowup. */
400         return NULL;
401
402 #  if !defined(_WIN32_WCE)
403     /* 0x400 is the spin count value suggested in the documentation */
404     if (!InitializeCriticalSectionAndSpinCount(lock, 0x400)) {
405         OPENSSL_free(lock);
406         return NULL;
407     }
408 #  else
409     InitializeCriticalSection(lock);
410 #  endif
411 # endif
412
413     return lock;
414 }
415
416 __owur int CRYPTO_THREAD_read_lock(CRYPTO_RWLOCK *lock)
417 {
418 # ifdef USE_RWLOCK
419     CRYPTO_win_rwlock *rwlock = lock;
420
421     AcquireSRWLockShared(&rwlock->lock);
422 # else
423     EnterCriticalSection(lock);
424 # endif
425     return 1;
426 }
427
428 __owur int CRYPTO_THREAD_write_lock(CRYPTO_RWLOCK *lock)
429 {
430 # ifdef USE_RWLOCK
431     CRYPTO_win_rwlock *rwlock = lock;
432
433     AcquireSRWLockExclusive(&rwlock->lock);
434     rwlock->exclusive = 1;
435 # else
436     EnterCriticalSection(lock);
437 # endif
438     return 1;
439 }
440
441 int CRYPTO_THREAD_unlock(CRYPTO_RWLOCK *lock)
442 {
443 # ifdef USE_RWLOCK
444     CRYPTO_win_rwlock *rwlock = lock;
445
446     if (rwlock->exclusive) {
447         rwlock->exclusive = 0;
448         ReleaseSRWLockExclusive(&rwlock->lock);
449     } else {
450         ReleaseSRWLockShared(&rwlock->lock);
451     }
452 # else
453     LeaveCriticalSection(lock);
454 # endif
455     return 1;
456 }
457
458 void CRYPTO_THREAD_lock_free(CRYPTO_RWLOCK *lock)
459 {
460     if (lock == NULL)
461         return;
462
463 # ifndef USE_RWLOCK
464     DeleteCriticalSection(lock);
465 # endif
466     OPENSSL_free(lock);
467
468     return;
469 }
470
471 # define ONCE_UNINITED     0
472 # define ONCE_ININIT       1
473 # define ONCE_DONE         2
474
475 /*
476  * We don't use InitOnceExecuteOnce because that isn't available in WinXP which
477  * we still have to support.
478  */
479 int CRYPTO_THREAD_run_once(CRYPTO_ONCE *once, void (*init)(void))
480 {
481     LONG volatile *lock = (LONG *)once;
482     LONG result;
483
484     if (*lock == ONCE_DONE)
485         return 1;
486
487     do {
488         result = InterlockedCompareExchange(lock, ONCE_ININIT, ONCE_UNINITED);
489         if (result == ONCE_UNINITED) {
490             init();
491             *lock = ONCE_DONE;
492             return 1;
493         }
494     } while (result == ONCE_ININIT);
495
496     return (*lock == ONCE_DONE);
497 }
498
499 int CRYPTO_THREAD_init_local(CRYPTO_THREAD_LOCAL *key, void (*cleanup)(void *))
500 {
501     *key = TlsAlloc();
502     if (*key == TLS_OUT_OF_INDEXES)
503         return 0;
504
505     return 1;
506 }
507
508 void *CRYPTO_THREAD_get_local(CRYPTO_THREAD_LOCAL *key)
509 {
510     DWORD last_error;
511     void *ret;
512
513     /*
514      * TlsGetValue clears the last error even on success, so that callers may
515      * distinguish it successfully returning NULL or failing. It is documented
516      * to never fail if the argument is a valid index from TlsAlloc, so we do
517      * not need to handle this.
518      *
519      * However, this error-mangling behavior interferes with the caller's use of
520      * GetLastError. In particular SSL_get_error queries the error queue to
521      * determine whether the caller should look at the OS's errors. To avoid
522      * destroying state, save and restore the Windows error.
523      *
524      * https://msdn.microsoft.com/en-us/library/windows/desktop/ms686812(v=vs.85).aspx
525      */
526     last_error = GetLastError();
527     ret = TlsGetValue(*key);
528     SetLastError(last_error);
529     return ret;
530 }
531
532 int CRYPTO_THREAD_set_local(CRYPTO_THREAD_LOCAL *key, void *val)
533 {
534     if (TlsSetValue(*key, val) == 0)
535         return 0;
536
537     return 1;
538 }
539
540 int CRYPTO_THREAD_cleanup_local(CRYPTO_THREAD_LOCAL *key)
541 {
542     if (TlsFree(*key) == 0)
543         return 0;
544
545     return 1;
546 }
547
548 CRYPTO_THREAD_ID CRYPTO_THREAD_get_current_id(void)
549 {
550     return GetCurrentThreadId();
551 }
552
553 int CRYPTO_THREAD_compare_id(CRYPTO_THREAD_ID a, CRYPTO_THREAD_ID b)
554 {
555     return (a == b);
556 }
557
558 int CRYPTO_atomic_add(int *val, int amount, int *ret, CRYPTO_RWLOCK *lock)
559 {
560     *ret = (int)InterlockedExchangeAdd((long volatile *)val, (long)amount) + amount;
561     return 1;
562 }
563
564 int CRYPTO_atomic_or(uint64_t *val, uint64_t op, uint64_t *ret,
565                      CRYPTO_RWLOCK *lock)
566 {
567 #if (defined(NO_INTERLOCKEDOR64))
568     if (lock == NULL || !CRYPTO_THREAD_write_lock(lock))
569         return 0;
570     *val |= op;
571     *ret = *val;
572
573     if (!CRYPTO_THREAD_unlock(lock))
574         return 0;
575
576     return 1;
577 #else
578     *ret = (uint64_t)InterlockedOr64((LONG64 volatile *)val, (LONG64)op) | op;
579     return 1;
580 #endif
581 }
582
583 int CRYPTO_atomic_load(uint64_t *val, uint64_t *ret, CRYPTO_RWLOCK *lock)
584 {
585 #if (defined(NO_INTERLOCKEDOR64))
586     if (lock == NULL || !CRYPTO_THREAD_read_lock(lock))
587         return 0;
588     *ret = *val;
589     if (!CRYPTO_THREAD_unlock(lock))
590         return 0;
591
592     return 1;
593 #else
594     *ret = (uint64_t)InterlockedOr64((LONG64 volatile *)val, 0);
595     return 1;
596 #endif
597 }
598
599 int CRYPTO_atomic_load_int(int *val, int *ret, CRYPTO_RWLOCK *lock)
600 {
601 #if (defined(NO_INTERLOCKEDOR64))
602     if (lock == NULL || !CRYPTO_THREAD_read_lock(lock))
603         return 0;
604     *ret = *val;
605     if (!CRYPTO_THREAD_unlock(lock))
606         return 0;
607
608     return 1;
609 #else
610     /* On Windows, LONG is always the same size as int. */
611     *ret = (int)InterlockedOr((LONG volatile *)val, 0);
612     return 1;
613 #endif
614 }
615
616 int openssl_init_fork_handlers(void)
617 {
618     return 0;
619 }
620
621 int openssl_get_fork_id(void)
622 {
623     return 0;
624 }
625 #endif