409 lines · c
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-exception6//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 deallocation17// to round 2. We need to delay deallocation because:18// - Android versions older than M lack __cxa_thread_atexit_impl, so apps19// 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 rounds22// after the final malloc/free call to free its thread-specific data (see23// https://reviews.llvm.org/D46978#1107507).24#define EMUTLS_SKIP_DESTRUCTOR_ROUNDS 125#else26#define EMUTLS_SKIP_DESTRUCTOR_ROUNDS 027#endif28 29#if defined(_MSC_VER) && !defined(__clang__)30// MSVC raises a warning about a nonstandard extension being used for the 031// sized element in this array. Disable this for warn-as-error builds.32#pragma warning(push)33#pragma warning(disable : 4200)34#endif35 36typedef struct emutls_address_array {37 uintptr_t skip_destructor_rounds;38 uintptr_t size; // number of elements in the 'data' array39 void *data[];40} emutls_address_array;41 42#if defined(_MSC_VER) && !defined(__clang__)43#pragma warning(pop)44#endif45 46static void emutls_shutdown(emutls_address_array *array);47 48#ifndef _WIN3249 50#include <pthread.h>51 52static pthread_mutex_t emutls_mutex = PTHREAD_MUTEX_INITIALIZER;53static pthread_key_t emutls_pthread_key;54static bool emutls_key_created = false;55 56typedef unsigned int gcc_word __attribute__((mode(word)));57typedef unsigned int gcc_pointer __attribute__((mode(pointer)));58 59// Default is not to use posix_memalign, so systems like Android60// can use thread local data without heavier POSIX memory allocators.61#ifndef EMUTLS_USE_POSIX_MEMALIGN62#define EMUTLS_USE_POSIX_MEMALIGN 063#endif64 65static __inline void *emutls_memalign_alloc(size_t align, size_t size) {66 void *base;67#if EMUTLS_USE_POSIX_MEMALIGN68 if (posix_memalign(&base, align, size) != 0)69 abort();70#else71#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#endif80 return base;81}82 83static __inline void emutls_memalign_free(void *base) {84#if EMUTLS_USE_POSIX_MEMALIGN85 free(base);86#else87 // The mallocated address is in ((void**)base)[-1]88 free(((void **)base)[-1]);89#endif90}91 92static __inline void emutls_setspecific(emutls_address_array *value) {93 pthread_setspecific(emutls_pthread_key, (void *)value);94}95 96static __inline emutls_address_array *emutls_getspecific(void) {97 return (emutls_address_array *)pthread_getspecific(emutls_pthread_key);98}99 100static 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. These104 // destructors are called in several rounds to accommodate destructor105 // functions that (re)initialize key values with pthread_setspecific.106 // Delay the emutls deallocation to accommodate other end-of-thread107 // cleanup tasks like calling thread_local destructors (e.g. the108 // __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 117static __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 123static __inline void emutls_init_once(void) {124 static pthread_once_t once = PTHREAD_ONCE_INIT;125 pthread_once(&once, emutls_init);126}127 128static __inline void emutls_lock(void) { pthread_mutex_lock(&emutls_mutex); }129 130static __inline void emutls_unlock(void) { pthread_mutex_unlock(&emutls_mutex); }131 132#else // _WIN32133 134#include <assert.h>135#include <malloc.h>136#include <stdio.h>137#include <windows.h>138 139static LPCRITICAL_SECTION emutls_mutex;140static DWORD emutls_tls_index = TLS_OUT_OF_INDEXES;141 142typedef uintptr_t gcc_word;143typedef void *gcc_pointer;144 145static 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, "Unknown Windows error: %s\n", hint);154 }155 LocalFree(buffer);156}157 158static __inline void win_abort(DWORD last_err, const char *hint) {159 win_error(last_err, hint);160 abort();161}162 163static __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 170static __inline void emutls_memalign_free(void *base) { _aligned_free(base); }171 172static 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 185static BOOL CALLBACK emutls_init(PINIT_ONCE p0, PVOID p1, PVOID *p2) {186 (void)p0;187 (void)p1;188 (void)p2;189 emutls_mutex =190 (LPCRITICAL_SECTION)_aligned_malloc(sizeof(CRITICAL_SECTION), 16);191 if (!emutls_mutex) {192 win_error(GetLastError(), "_aligned_malloc");193 return FALSE;194 }195 InitializeCriticalSection(emutls_mutex);196 197 emutls_tls_index = TlsAlloc();198 if (emutls_tls_index == TLS_OUT_OF_INDEXES) {199 emutls_exit();200 win_error(GetLastError(), "TlsAlloc");201 return FALSE;202 }203 atexit(&emutls_exit);204 return TRUE;205}206 207static __inline void emutls_init_once(void) {208 static INIT_ONCE once;209 InitOnceExecuteOnce(&once, emutls_init, NULL, NULL);210}211 212static __inline void emutls_lock(void) { EnterCriticalSection(emutls_mutex); }213 214static __inline void emutls_unlock(void) { LeaveCriticalSection(emutls_mutex); }215 216static __inline void emutls_setspecific(emutls_address_array *value) {217 if (TlsSetValue(emutls_tls_index, (LPVOID)value) == 0)218 win_abort(GetLastError(), "TlsSetValue");219}220 221static __inline emutls_address_array *emutls_getspecific(void) {222 LPVOID value = TlsGetValue(emutls_tls_index);223 if (value == NULL) {224 const DWORD err = GetLastError();225 if (err != ERROR_SUCCESS)226 win_abort(err, "TlsGetValue");227 }228 return (emutls_address_array *)value;229}230 231// Provide atomic load/store functions for emutls_get_index if built with MSVC.232#if !defined(__ATOMIC_RELEASE)233#include <intrin.h>234 235enum { __ATOMIC_ACQUIRE = 2, __ATOMIC_RELEASE = 3 };236 237static __inline uintptr_t __atomic_load_n(void *ptr, unsigned type) {238 assert(type == __ATOMIC_ACQUIRE);239 // These return the previous value - but since we do an OR with 0,240 // it's equivalent to a plain load.241#ifdef _WIN64242 return InterlockedOr64(ptr, 0);243#else244 return InterlockedOr(ptr, 0);245#endif246}247 248static __inline void __atomic_store_n(void *ptr, uintptr_t val, unsigned type) {249 assert(type == __ATOMIC_RELEASE);250 InterlockedExchangePointer((void *volatile *)ptr, (void *)val);251}252 253#endif // __ATOMIC_RELEASE254 255#endif // _WIN32256 257static size_t emutls_num_object = 0; // number of allocated TLS objects258 259// Free the allocated TLS data260static void emutls_shutdown(emutls_address_array *array) {261 if (array) {262 uintptr_t i;263 for (i = 0; i < array->size; ++i) {264 if (array->data[i])265 emutls_memalign_free(array->data[i]);266 }267 }268}269 270// For every TLS variable xyz,271// there is one __emutls_control variable named __emutls_v.xyz.272// If xyz has non-zero initial value, __emutls_v.xyz's "value"273// will point to __emutls_t.xyz, which has the initial value.274typedef struct __emutls_control {275 // Must use gcc_word here, instead of size_t, to match GCC. When276 // gcc_word is larger than size_t, the upper extra bits are all277 // zeros. We can use variables of size_t to operate on size and278 // align.279 gcc_word size; // size of the object in bytes280 gcc_word align; // alignment of the object in bytes281 union {282 uintptr_t index; // data[index-1] is the object address283 void *address; // object address, when in single thread env284 } object;285 void *value; // null or non-zero initial value for the object286} __emutls_control;287 288// Emulated TLS objects are always allocated at run-time.289static __inline void *emutls_allocate_object(__emutls_control *control) {290 // Use standard C types, check with gcc's emutls.o.291 COMPILE_TIME_ASSERT(sizeof(uintptr_t) == sizeof(gcc_pointer));292 COMPILE_TIME_ASSERT(sizeof(uintptr_t) == sizeof(void *));293 294 size_t size = control->size;295 size_t align = control->align;296 void *base;297 if (align < sizeof(void *))298 align = sizeof(void *);299 // Make sure that align is power of 2.300 if ((align & (align - 1)) != 0)301 abort();302 303 base = emutls_memalign_alloc(align, size);304 if (control->value)305 memcpy(base, control->value, size);306 else307 memset(base, 0, size);308 return base;309}310 311// Returns control->object.index; set index if not allocated yet.312static __inline uintptr_t emutls_get_index(__emutls_control *control) {313 uintptr_t index = __atomic_load_n(&control->object.index, __ATOMIC_ACQUIRE);314 if (!index) {315 emutls_init_once();316 emutls_lock();317 index = control->object.index;318 if (!index) {319 index = ++emutls_num_object;320 __atomic_store_n(&control->object.index, index, __ATOMIC_RELEASE);321 }322 emutls_unlock();323 }324 return index;325}326 327// Updates newly allocated thread local emutls_address_array.328static __inline void emutls_check_array_set_size(emutls_address_array *array,329 uintptr_t size) {330 if (array == NULL)331 abort();332 array->size = size;333 emutls_setspecific(array);334}335 336// Returns the new 'data' array size, number of elements,337// which must be no smaller than the given index.338static __inline uintptr_t emutls_new_data_array_size(uintptr_t index) {339 // Need to allocate emutls_address_array with extra slots340 // to store the header.341 // Round up the emutls_address_array size to multiple of 16.342 uintptr_t header_words = sizeof(emutls_address_array) / sizeof(void *);343 return ((index + header_words + 15) & ~((uintptr_t)15)) - header_words;344}345 346// Returns the size in bytes required for an emutls_address_array with347// N number of elements for data field.348static __inline uintptr_t emutls_asize(uintptr_t N) {349 return N * sizeof(void *) + sizeof(emutls_address_array);350}351 352// Returns the thread local emutls_address_array.353// Extends its size if necessary to hold address at index.354static __inline emutls_address_array *355emutls_get_address_array(uintptr_t index) {356 emutls_address_array *array = emutls_getspecific();357 if (array == NULL) {358 uintptr_t new_size = emutls_new_data_array_size(index);359 array = (emutls_address_array *)malloc(emutls_asize(new_size));360 if (array) {361 memset(array->data, 0, new_size * sizeof(void *));362 array->skip_destructor_rounds = EMUTLS_SKIP_DESTRUCTOR_ROUNDS;363 }364 emutls_check_array_set_size(array, new_size);365 } else if (index > array->size) {366 uintptr_t orig_size = array->size;367 uintptr_t new_size = emutls_new_data_array_size(index);368 array = (emutls_address_array *)realloc(array, emutls_asize(new_size));369 if (array)370 memset(array->data + orig_size, 0,371 (new_size - orig_size) * sizeof(void *));372 emutls_check_array_set_size(array, new_size);373 }374 return array;375}376 377#ifndef _WIN32378// Our emulated TLS implementation relies on local state (e.g. for the pthread379// key), and if we duplicate this state across different shared libraries,380// accesses to the same TLS variable from different shared libraries will yield381// different results (see https://github.com/android/ndk/issues/1551 for an382// example). __emutls_get_address is the only external entry point for emulated383// TLS, and by making it default visibility and weak, we can rely on the dynamic384// linker to coalesce multiple copies at runtime and ensure a single unique copy385// of TLS state. This is a best effort; it won't work if the user is linking386// with -Bsymbolic or -Bsymbolic-functions, and it also won't work on Windows,387// where the dynamic linker has no notion of coalescing weak symbols at runtime.388// A more robust solution would be to create a separate shared library for389// emulated TLS, to ensure a single copy of its state.390__attribute__((visibility("default"), weak))391#endif392void *__emutls_get_address(__emutls_control *control) {393 uintptr_t index = emutls_get_index(control);394 emutls_address_array *array = emutls_get_address_array(index--);395 if (array->data[index] == NULL)396 array->data[index] = emutls_allocate_object(control);397 return array->data[index];398}399 400#ifdef __BIONIC__401// Called by Bionic on dlclose to delete the emutls pthread key.402__attribute__((visibility("hidden"))) void __emutls_unregister_key(void) {403 if (emutls_key_created) {404 pthread_key_delete(emutls_pthread_key);405 emutls_key_created = false;406 }407}408#endif409