1 // Copyright (c) 2012 The Chromium Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style license that can be 3 // found in the LICENSE file. 4 5 #include "base/threading/thread_local_storage.h" 6 7 #include <windows.h> 8 9 #include "base/logging.h" 10 11 12 namespace { 13 // In order to make TLS destructors work, we need to keep function 14 // pointers to the destructor for each TLS that we allocate. 15 // We make this work by allocating a single OS-level TLS, which 16 // contains an array of slots for the application to use. In 17 // parallel, we also allocate an array of destructors, which we 18 // keep track of and call when threads terminate. 19 20 // g_native_tls_key is the one native TLS that we use. It stores our table. 21 long g_native_tls_key = TLS_OUT_OF_INDEXES; 22 23 // g_last_used_tls_key is the high-water-mark of allocated thread local storage. 24 // Each allocation is an index into our g_tls_destructors[]. Each such index is 25 // assigned to the instance variable slot_ in a ThreadLocalStorage::Slot 26 // instance. We reserve the value slot_ == 0 to indicate that the corresponding 27 // instance of ThreadLocalStorage::Slot has been freed (i.e., destructor called, 28 // etc.). This reserved use of 0 is then stated as the initial value of 29 // g_last_used_tls_key, so that the first issued index will be 1. 30 long g_last_used_tls_key = 0; 31 32 // The maximum number of 'slots' in our thread local storage stack. 33 const int kThreadLocalStorageSize = 64; 34 35 // The maximum number of times to try to clear slots by calling destructors. 36 // Use pthread naming convention for clarity. 37 const int kMaxDestructorIterations = kThreadLocalStorageSize; 38 39 // An array of destructor function pointers for the slots. If a slot has a 40 // destructor, it will be stored in its corresponding entry in this array. 41 // The elements are volatile to ensure that when the compiler reads the value 42 // to potentially call the destructor, it does so once, and that value is tested 43 // for null-ness and then used. Yes, that would be a weird de-optimization, 44 // but I can imagine some register machines where it was just as easy to 45 // re-fetch an array element, and I want to be sure a call to free the key 46 // (i.e., null out the destructor entry) that happens on a separate thread can't 47 // hurt the racy calls to the destructors on another thread. 48 volatile base::ThreadLocalStorage::TLSDestructorFunc 49 g_tls_destructors[kThreadLocalStorageSize]; 50 51 void** ConstructTlsVector() { 52 if (g_native_tls_key == TLS_OUT_OF_INDEXES) { 53 long value = TlsAlloc(); 54 DCHECK(value != TLS_OUT_OF_INDEXES); 55 56 // Atomically test-and-set the tls_key. If the key is TLS_OUT_OF_INDEXES, 57 // go ahead and set it. Otherwise, do nothing, as another 58 // thread already did our dirty work. 59 if (TLS_OUT_OF_INDEXES != InterlockedCompareExchange( 60 &g_native_tls_key, value, TLS_OUT_OF_INDEXES)) { 61 // We've been shortcut. Another thread replaced g_native_tls_key first so 62 // we need to destroy our index and use the one the other thread got 63 // first. 64 TlsFree(value); 65 } 66 } 67 DCHECK(!TlsGetValue(g_native_tls_key)); 68 69 // Some allocators, such as TCMalloc, make use of thread local storage. 70 // As a result, any attempt to call new (or malloc) will lazily cause such a 71 // system to initialize, which will include registering for a TLS key. If we 72 // are not careful here, then that request to create a key will call new back, 73 // and we'll have an infinite loop. We avoid that as follows: 74 // Use a stack allocated vector, so that we don't have dependence on our 75 // allocator until our service is in place. (i.e., don't even call new until 76 // after we're setup) 77 void* stack_allocated_tls_data[kThreadLocalStorageSize]; 78 memset(stack_allocated_tls_data, 0, sizeof(stack_allocated_tls_data)); 79 // Ensure that any rentrant calls change the temp version. 80 TlsSetValue(g_native_tls_key, stack_allocated_tls_data); 81 82 // Allocate an array to store our data. 83 void** tls_data = new void*[kThreadLocalStorageSize]; 84 memcpy(tls_data, stack_allocated_tls_data, sizeof(stack_allocated_tls_data)); 85 TlsSetValue(g_native_tls_key, tls_data); 86 return tls_data; 87 } 88 89 // Called when we terminate a thread, this function calls any TLS destructors 90 // that are pending for this thread. 91 void WinThreadExit() { 92 if (g_native_tls_key == TLS_OUT_OF_INDEXES) 93 return; 94 95 void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key)); 96 // Maybe we have never initialized TLS for this thread. 97 if (!tls_data) 98 return; 99 100 // Some allocators, such as TCMalloc, use TLS. As a result, when a thread 101 // terminates, one of the destructor calls we make may be to shut down an 102 // allocator. We have to be careful that after we've shutdown all of the 103 // known destructors (perchance including an allocator), that we don't call 104 // the allocator and cause it to resurrect itself (with no possibly destructor 105 // call to follow). We handle this problem as follows: 106 // Switch to using a stack allocated vector, so that we don't have dependence 107 // on our allocator after we have called all g_tls_destructors. (i.e., don't 108 // even call delete[] after we're done with destructors.) 109 void* stack_allocated_tls_data[kThreadLocalStorageSize]; 110 memcpy(stack_allocated_tls_data, tls_data, sizeof(stack_allocated_tls_data)); 111 // Ensure that any re-entrant calls change the temp version. 112 TlsSetValue(g_native_tls_key, stack_allocated_tls_data); 113 delete[] tls_data; // Our last dependence on an allocator. 114 115 int remaining_attempts = kMaxDestructorIterations; 116 bool need_to_scan_destructors = true; 117 while (need_to_scan_destructors) { 118 need_to_scan_destructors = false; 119 // Try to destroy the first-created-slot (which is slot 1) in our last 120 // destructor call. That user was able to function, and define a slot with 121 // no other services running, so perhaps it is a basic service (like an 122 // allocator) and should also be destroyed last. If we get the order wrong, 123 // then we'll itterate several more times, so it is really not that 124 // critical (but it might help). 125 for (int slot = g_last_used_tls_key; slot > 0; --slot) { 126 void* value = stack_allocated_tls_data[slot]; 127 if (value == NULL) 128 continue; 129 base::ThreadLocalStorage::TLSDestructorFunc destructor = 130 g_tls_destructors[slot]; 131 if (destructor == NULL) 132 continue; 133 stack_allocated_tls_data[slot] = NULL; // pre-clear the slot. 134 destructor(value); 135 // Any destructor might have called a different service, which then set 136 // a different slot to a non-NULL value. Hence we need to check 137 // the whole vector again. This is a pthread standard. 138 need_to_scan_destructors = true; 139 } 140 if (--remaining_attempts <= 0) { 141 NOTREACHED(); // Destructors might not have been called. 142 break; 143 } 144 } 145 146 // Remove our stack allocated vector. 147 TlsSetValue(g_native_tls_key, NULL); 148 } 149 150 } // namespace 151 152 namespace base { 153 154 ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) { 155 initialized_ = false; 156 slot_ = 0; 157 Initialize(destructor); 158 } 159 160 bool ThreadLocalStorage::StaticSlot::Initialize(TLSDestructorFunc destructor) { 161 if (g_native_tls_key == TLS_OUT_OF_INDEXES || !TlsGetValue(g_native_tls_key)) 162 ConstructTlsVector(); 163 164 // Grab a new slot. 165 slot_ = InterlockedIncrement(&g_last_used_tls_key); 166 DCHECK_GT(slot_, 0); 167 if (slot_ >= kThreadLocalStorageSize) { 168 NOTREACHED(); 169 return false; 170 } 171 172 // Setup our destructor. 173 g_tls_destructors[slot_] = destructor; 174 initialized_ = true; 175 return true; 176 } 177 178 void ThreadLocalStorage::StaticSlot::Free() { 179 // At this time, we don't reclaim old indices for TLS slots. 180 // So all we need to do is wipe the destructor. 181 DCHECK_GT(slot_, 0); 182 DCHECK_LT(slot_, kThreadLocalStorageSize); 183 g_tls_destructors[slot_] = NULL; 184 slot_ = 0; 185 initialized_ = false; 186 } 187 188 void* ThreadLocalStorage::StaticSlot::Get() const { 189 void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key)); 190 if (!tls_data) 191 tls_data = ConstructTlsVector(); 192 DCHECK_GT(slot_, 0); 193 DCHECK_LT(slot_, kThreadLocalStorageSize); 194 return tls_data[slot_]; 195 } 196 197 void ThreadLocalStorage::StaticSlot::Set(void* value) { 198 void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key)); 199 if (!tls_data) 200 tls_data = ConstructTlsVector(); 201 DCHECK_GT(slot_, 0); 202 DCHECK_LT(slot_, kThreadLocalStorageSize); 203 tls_data[slot_] = value; 204 } 205 206 } // namespace base 207 208 // Thread Termination Callbacks. 209 // Windows doesn't support a per-thread destructor with its 210 // TLS primitives. So, we build it manually by inserting a 211 // function to be called on each thread's exit. 212 // This magic is from http://www.codeproject.com/threads/tls.asp 213 // and it works for VC++ 7.0 and later. 214 215 // Force a reference to _tls_used to make the linker create the TLS directory 216 // if it's not already there. (e.g. if __declspec(thread) is not used). 217 // Force a reference to p_thread_callback_base to prevent whole program 218 // optimization from discarding the variable. 219 #ifdef _WIN64 220 221 #pragma comment(linker, "/INCLUDE:_tls_used") 222 #pragma comment(linker, "/INCLUDE:p_thread_callback_base") 223 224 #else // _WIN64 225 226 #pragma comment(linker, "/INCLUDE:__tls_used") 227 #pragma comment(linker, "/INCLUDE:_p_thread_callback_base") 228 229 #endif // _WIN64 230 231 // Static callback function to call with each thread termination. 232 void NTAPI OnThreadExit(PVOID module, DWORD reason, PVOID reserved) { 233 // On XP SP0 & SP1, the DLL_PROCESS_ATTACH is never seen. It is sent on SP2+ 234 // and on W2K and W2K3. So don't assume it is sent. 235 if (DLL_THREAD_DETACH == reason || DLL_PROCESS_DETACH == reason) 236 WinThreadExit(); 237 } 238 239 // .CRT$XLA to .CRT$XLZ is an array of PIMAGE_TLS_CALLBACK pointers that are 240 // called automatically by the OS loader code (not the CRT) when the module is 241 // loaded and on thread creation. They are NOT called if the module has been 242 // loaded by a LoadLibrary() call. It must have implicitly been loaded at 243 // process startup. 244 // By implicitly loaded, I mean that it is directly referenced by the main EXE 245 // or by one of its dependent DLLs. Delay-loaded DLL doesn't count as being 246 // implicitly loaded. 247 // 248 // See VC\crt\src\tlssup.c for reference. 249 250 // extern "C" suppresses C++ name mangling so we know the symbol name for the 251 // linker /INCLUDE:symbol pragma above. 252 extern "C" { 253 // The linker must not discard p_thread_callback_base. (We force a reference 254 // to this variable with a linker /INCLUDE:symbol pragma to ensure that.) If 255 // this variable is discarded, the OnThreadExit function will never be called. 256 #ifdef _WIN64 257 258 // .CRT section is merged with .rdata on x64 so it must be constant data. 259 #pragma const_seg(".CRT$XLB") 260 // When defining a const variable, it must have external linkage to be sure the 261 // linker doesn't discard it. 262 extern const PIMAGE_TLS_CALLBACK p_thread_callback_base; 263 const PIMAGE_TLS_CALLBACK p_thread_callback_base = OnThreadExit; 264 265 // Reset the default section. 266 #pragma const_seg() 267 268 #else // _WIN64 269 270 #pragma data_seg(".CRT$XLB") 271 PIMAGE_TLS_CALLBACK p_thread_callback_base = OnThreadExit; 272 273 // Reset the default section. 274 #pragma data_seg() 275 276 #endif // _WIN64 277 } // extern "C" 278