xref: /freebsd/contrib/llvm-project/compiler-rt/lib/builtins/emutls.c (revision d500a85e640d1cd270747c12e17c511b53864436)
1 //===---------- emutls.c - Implements __emutls_get_address ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <string.h>
12 
13 #include "int_lib.h"
14 
15 #ifdef __BIONIC__
16 // There are 4 pthread key cleanup rounds on Bionic. Delay emutls deallocation
17 // to round 2. We need to delay deallocation because:
18 //  - Android versions older than M lack __cxa_thread_atexit_impl, so apps
19 //    use a pthread key destructor to call C++ destructors.
20 //  - Apps might use __thread/thread_local variables in pthread destructors.
21 // We can't wait until the final two rounds, because jemalloc needs two rounds
22 // after the final malloc/free call to free its thread-specific data (see
23 // https://reviews.llvm.org/D46978#1107507).
24 #define EMUTLS_SKIP_DESTRUCTOR_ROUNDS 1
25 #else
26 #define EMUTLS_SKIP_DESTRUCTOR_ROUNDS 0
27 #endif
28 
29 #if defined(_MSC_VER) && !defined(__clang__)
30 // MSVC raises a warning about a nonstandard extension being used for the 0
31 // sized element in this array. Disable this for warn-as-error builds.
32 #pragma warning(push)
33 #pragma warning(disable : 4206)
34 #endif
35 
36 typedef struct emutls_address_array {
37   uintptr_t skip_destructor_rounds;
38   uintptr_t size; // number of elements in the 'data' array
39   void *data[];
40 } emutls_address_array;
41 
42 #if defined(_MSC_VER) && !defined(__clang__)
43 #pragma warning(pop)
44 #endif
45 
46 static void emutls_shutdown(emutls_address_array *array);
47 
48 #ifndef _WIN32
49 
50 #include <pthread.h>
51 
52 static pthread_mutex_t emutls_mutex = PTHREAD_MUTEX_INITIALIZER;
53 static pthread_key_t emutls_pthread_key;
54 static bool emutls_key_created = false;
55 
56 typedef unsigned int gcc_word __attribute__((mode(word)));
57 typedef unsigned int gcc_pointer __attribute__((mode(pointer)));
58 
59 // Default is not to use posix_memalign, so systems like Android
60 // can use thread local data without heavier POSIX memory allocators.
61 #ifndef EMUTLS_USE_POSIX_MEMALIGN
62 #define EMUTLS_USE_POSIX_MEMALIGN 0
63 #endif
64 
65 static __inline void *emutls_memalign_alloc(size_t align, size_t size) {
66   void *base;
67 #if EMUTLS_USE_POSIX_MEMALIGN
68   if (posix_memalign(&base, align, size) != 0)
69     abort();
70 #else
71 #define EXTRA_ALIGN_PTR_BYTES (align - 1 + sizeof(void *))
72   char *object;
73   if ((object = (char *)malloc(EXTRA_ALIGN_PTR_BYTES + size)) == NULL)
74     abort();
75   base = (void *)(((uintptr_t)(object + EXTRA_ALIGN_PTR_BYTES)) &
76                   ~(uintptr_t)(align - 1));
77 
78   ((void **)base)[-1] = object;
79 #endif
80   return base;
81 }
82 
83 static __inline void emutls_memalign_free(void *base) {
84 #if EMUTLS_USE_POSIX_MEMALIGN
85   free(base);
86 #else
87   // The mallocated address is in ((void**)base)[-1]
88   free(((void **)base)[-1]);
89 #endif
90 }
91 
92 static __inline void emutls_setspecific(emutls_address_array *value) {
93   pthread_setspecific(emutls_pthread_key, (void *)value);
94 }
95 
96 static __inline emutls_address_array *emutls_getspecific() {
97   return (emutls_address_array *)pthread_getspecific(emutls_pthread_key);
98 }
99 
100 static void emutls_key_destructor(void *ptr) {
101   emutls_address_array *array = (emutls_address_array *)ptr;
102   if (array->skip_destructor_rounds > 0) {
103     // emutls is deallocated using a pthread key destructor. These
104     // destructors are called in several rounds to accommodate destructor
105     // functions that (re)initialize key values with pthread_setspecific.
106     // Delay the emutls deallocation to accommodate other end-of-thread
107     // cleanup tasks like calling thread_local destructors (e.g. the
108     // __cxa_thread_atexit fallback in libc++abi).
109     array->skip_destructor_rounds--;
110     emutls_setspecific(array);
111   } else {
112     emutls_shutdown(array);
113     free(ptr);
114   }
115 }
116 
117 static __inline void emutls_init(void) {
118   if (pthread_key_create(&emutls_pthread_key, emutls_key_destructor) != 0)
119     abort();
120   emutls_key_created = true;
121 }
122 
123 static __inline void emutls_init_once(void) {
124   static pthread_once_t once = PTHREAD_ONCE_INIT;
125   pthread_once(&once, emutls_init);
126 }
127 
128 static __inline void emutls_lock() { pthread_mutex_lock(&emutls_mutex); }
129 
130 static __inline void emutls_unlock() { pthread_mutex_unlock(&emutls_mutex); }
131 
132 #else // _WIN32
133 
134 #include <assert.h>
135 #include <malloc.h>
136 #include <stdio.h>
137 #include <windows.h>
138 
139 static LPCRITICAL_SECTION emutls_mutex;
140 static DWORD emutls_tls_index = TLS_OUT_OF_INDEXES;
141 
142 typedef uintptr_t gcc_word;
143 typedef void *gcc_pointer;
144 
145 static void win_error(DWORD last_err, const char *hint) {
146   char *buffer = NULL;
147   if (FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER |
148                          FORMAT_MESSAGE_FROM_SYSTEM |
149                          FORMAT_MESSAGE_MAX_WIDTH_MASK,
150                      NULL, last_err, 0, (LPSTR)&buffer, 1, NULL)) {
151     fprintf(stderr, "Windows error: %s\n", buffer);
152   } else {
153     fprintf(stderr, "Unkown Windows error: %s\n", hint);
154   }
155   LocalFree(buffer);
156 }
157 
158 static __inline void win_abort(DWORD last_err, const char *hint) {
159   win_error(last_err, hint);
160   abort();
161 }
162 
163 static __inline void *emutls_memalign_alloc(size_t align, size_t size) {
164   void *base = _aligned_malloc(size, align);
165   if (!base)
166     win_abort(GetLastError(), "_aligned_malloc");
167   return base;
168 }
169 
170 static __inline void emutls_memalign_free(void *base) { _aligned_free(base); }
171 
172 static void emutls_exit(void) {
173   if (emutls_mutex) {
174     DeleteCriticalSection(emutls_mutex);
175     _aligned_free(emutls_mutex);
176     emutls_mutex = NULL;
177   }
178   if (emutls_tls_index != TLS_OUT_OF_INDEXES) {
179     emutls_shutdown((emutls_address_array *)TlsGetValue(emutls_tls_index));
180     TlsFree(emutls_tls_index);
181     emutls_tls_index = TLS_OUT_OF_INDEXES;
182   }
183 }
184 
185 #pragma warning(push)
186 #pragma warning(disable : 4100)
187 static BOOL CALLBACK emutls_init(PINIT_ONCE p0, PVOID p1, PVOID *p2) {
188   emutls_mutex =
189       (LPCRITICAL_SECTION)_aligned_malloc(sizeof(CRITICAL_SECTION), 16);
190   if (!emutls_mutex) {
191     win_error(GetLastError(), "_aligned_malloc");
192     return FALSE;
193   }
194   InitializeCriticalSection(emutls_mutex);
195 
196   emutls_tls_index = TlsAlloc();
197   if (emutls_tls_index == TLS_OUT_OF_INDEXES) {
198     emutls_exit();
199     win_error(GetLastError(), "TlsAlloc");
200     return FALSE;
201   }
202   atexit(&emutls_exit);
203   return TRUE;
204 }
205 
206 static __inline void emutls_init_once(void) {
207   static INIT_ONCE once;
208   InitOnceExecuteOnce(&once, emutls_init, NULL, NULL);
209 }
210 
211 static __inline void emutls_lock() { EnterCriticalSection(emutls_mutex); }
212 
213 static __inline void emutls_unlock() { LeaveCriticalSection(emutls_mutex); }
214 
215 static __inline void emutls_setspecific(emutls_address_array *value) {
216   if (TlsSetValue(emutls_tls_index, (LPVOID)value) == 0)
217     win_abort(GetLastError(), "TlsSetValue");
218 }
219 
220 static __inline emutls_address_array *emutls_getspecific() {
221   LPVOID value = TlsGetValue(emutls_tls_index);
222   if (value == NULL) {
223     const DWORD err = GetLastError();
224     if (err != ERROR_SUCCESS)
225       win_abort(err, "TlsGetValue");
226   }
227   return (emutls_address_array *)value;
228 }
229 
230 // Provide atomic load/store functions for emutls_get_index if built with MSVC.
231 #if !defined(__ATOMIC_RELEASE)
232 #include <intrin.h>
233 
234 enum { __ATOMIC_ACQUIRE = 2, __ATOMIC_RELEASE = 3 };
235 
236 static __inline uintptr_t __atomic_load_n(void *ptr, unsigned type) {
237   assert(type == __ATOMIC_ACQUIRE);
238   // These return the previous value - but since we do an OR with 0,
239   // it's equivalent to a plain load.
240 #ifdef _WIN64
241   return InterlockedOr64(ptr, 0);
242 #else
243   return InterlockedOr(ptr, 0);
244 #endif
245 }
246 
247 static __inline void __atomic_store_n(void *ptr, uintptr_t val, unsigned type) {
248   assert(type == __ATOMIC_RELEASE);
249   InterlockedExchangePointer((void *volatile *)ptr, (void *)val);
250 }
251 
252 #endif // __ATOMIC_RELEASE
253 
254 #pragma warning(pop)
255 
256 #endif // _WIN32
257 
258 static size_t emutls_num_object = 0; // number of allocated TLS objects
259 
260 // Free the allocated TLS data
261 static void emutls_shutdown(emutls_address_array *array) {
262   if (array) {
263     uintptr_t i;
264     for (i = 0; i < array->size; ++i) {
265       if (array->data[i])
266         emutls_memalign_free(array->data[i]);
267     }
268   }
269 }
270 
271 // For every TLS variable xyz,
272 // there is one __emutls_control variable named __emutls_v.xyz.
273 // If xyz has non-zero initial value, __emutls_v.xyz's "value"
274 // will point to __emutls_t.xyz, which has the initial value.
275 typedef struct __emutls_control {
276   // Must use gcc_word here, instead of size_t, to match GCC.  When
277   // gcc_word is larger than size_t, the upper extra bits are all
278   // zeros.  We can use variables of size_t to operate on size and
279   // align.
280   gcc_word size;  // size of the object in bytes
281   gcc_word align; // alignment of the object in bytes
282   union {
283     uintptr_t index; // data[index-1] is the object address
284     void *address;   // object address, when in single thread env
285   } object;
286   void *value; // null or non-zero initial value for the object
287 } __emutls_control;
288 
289 // Emulated TLS objects are always allocated at run-time.
290 static __inline void *emutls_allocate_object(__emutls_control *control) {
291   // Use standard C types, check with gcc's emutls.o.
292   COMPILE_TIME_ASSERT(sizeof(uintptr_t) == sizeof(gcc_pointer));
293   COMPILE_TIME_ASSERT(sizeof(uintptr_t) == sizeof(void *));
294 
295   size_t size = control->size;
296   size_t align = control->align;
297   void *base;
298   if (align < sizeof(void *))
299     align = sizeof(void *);
300   // Make sure that align is power of 2.
301   if ((align & (align - 1)) != 0)
302     abort();
303 
304   base = emutls_memalign_alloc(align, size);
305   if (control->value)
306     memcpy(base, control->value, size);
307   else
308     memset(base, 0, size);
309   return base;
310 }
311 
312 // Returns control->object.index; set index if not allocated yet.
313 static __inline uintptr_t emutls_get_index(__emutls_control *control) {
314   uintptr_t index = __atomic_load_n(&control->object.index, __ATOMIC_ACQUIRE);
315   if (!index) {
316     emutls_init_once();
317     emutls_lock();
318     index = control->object.index;
319     if (!index) {
320       index = ++emutls_num_object;
321       __atomic_store_n(&control->object.index, index, __ATOMIC_RELEASE);
322     }
323     emutls_unlock();
324   }
325   return index;
326 }
327 
328 // Updates newly allocated thread local emutls_address_array.
329 static __inline void emutls_check_array_set_size(emutls_address_array *array,
330                                                  uintptr_t size) {
331   if (array == NULL)
332     abort();
333   array->size = size;
334   emutls_setspecific(array);
335 }
336 
337 // Returns the new 'data' array size, number of elements,
338 // which must be no smaller than the given index.
339 static __inline uintptr_t emutls_new_data_array_size(uintptr_t index) {
340   // Need to allocate emutls_address_array with extra slots
341   // to store the header.
342   // Round up the emutls_address_array size to multiple of 16.
343   uintptr_t header_words = sizeof(emutls_address_array) / sizeof(void *);
344   return ((index + header_words + 15) & ~((uintptr_t)15)) - header_words;
345 }
346 
347 // Returns the size in bytes required for an emutls_address_array with
348 // N number of elements for data field.
349 static __inline uintptr_t emutls_asize(uintptr_t N) {
350   return N * sizeof(void *) + sizeof(emutls_address_array);
351 }
352 
353 // Returns the thread local emutls_address_array.
354 // Extends its size if necessary to hold address at index.
355 static __inline emutls_address_array *
356 emutls_get_address_array(uintptr_t index) {
357   emutls_address_array *array = emutls_getspecific();
358   if (array == NULL) {
359     uintptr_t new_size = emutls_new_data_array_size(index);
360     array = (emutls_address_array *)malloc(emutls_asize(new_size));
361     if (array) {
362       memset(array->data, 0, new_size * sizeof(void *));
363       array->skip_destructor_rounds = EMUTLS_SKIP_DESTRUCTOR_ROUNDS;
364     }
365     emutls_check_array_set_size(array, new_size);
366   } else if (index > array->size) {
367     uintptr_t orig_size = array->size;
368     uintptr_t new_size = emutls_new_data_array_size(index);
369     array = (emutls_address_array *)realloc(array, emutls_asize(new_size));
370     if (array)
371       memset(array->data + orig_size, 0,
372              (new_size - orig_size) * sizeof(void *));
373     emutls_check_array_set_size(array, new_size);
374   }
375   return array;
376 }
377 
378 void *__emutls_get_address(__emutls_control *control) {
379   uintptr_t index = emutls_get_index(control);
380   emutls_address_array *array = emutls_get_address_array(index--);
381   if (array->data[index] == NULL)
382     array->data[index] = emutls_allocate_object(control);
383   return array->data[index];
384 }
385 
386 #ifdef __BIONIC__
387 // Called by Bionic on dlclose to delete the emutls pthread key.
388 __attribute__((visibility("hidden"))) void __emutls_unregister_key(void) {
389   if (emutls_key_created) {
390     pthread_key_delete(emutls_pthread_key);
391     emutls_key_created = false;
392   }
393 }
394 #endif
395