OLD | NEW |
---|---|
1 // Copyright (c) 2012 The Chromium Authors. All rights reserved. | 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 | 2 // Use of this source code is governed by a BSD-style license that can be |
3 // found in the LICENSE file. | 3 // found in the LICENSE file. |
4 | 4 |
5 #include "base/threading/thread_local_storage.h" | 5 #include "base/threading/thread_local_storage.h" |
6 | 6 |
7 #include <windows.h> | 7 #include "base/atomicops.h" |
8 | |
9 #include "base/logging.h" | 8 #include "base/logging.h" |
10 | 9 |
10 using base::internal::PlatformThreadLocalStorage; | |
11 | 11 |
12 namespace { | 12 namespace { |
13 // In order to make TLS destructors work, we need to keep function | 13 // In order to make TLS destructors work, we need to keep function |
14 // pointers to the destructor for each TLS that we allocate. | 14 // pointers to the destructor for each TLS that we allocate. |
15 // We make this work by allocating a single OS-level TLS, which | 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 | 16 // contains an array of slots for the application to use. In |
17 // parallel, we also allocate an array of destructors, which we | 17 // parallel, we also allocate an array of destructors, which we |
18 // keep track of and call when threads terminate. | 18 // keep track of and call when threads terminate. |
jar (doing other things)
2013/11/26 20:05:24
nit: Starting on line 13, better comment would be:
michaelbai
2013/11/27 00:29:26
Done.
| |
19 | 19 |
20 // g_native_tls_key is the one native TLS that we use. It stores our table. | 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; | 21 PlatformThreadLocalStorage::TLSKey g_native_tls_key = |
22 PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES; | |
22 | 23 |
23 // g_last_used_tls_key is the high-water-mark of allocated thread local storage. | 24 // 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 // 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 // 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. We reserve the value slot_ == 0 to indicate that the corresponding |
27 // instance of ThreadLocalStorage::Slot has been freed (i.e., destructor called, | 28 // 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 // 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 // g_last_used_tls_key, so that the first issued index will be 1. |
30 long g_last_used_tls_key = 0; | 31 base::subtle::Atomic32 g_last_used_tls_key = 0; |
31 | 32 |
32 // The maximum number of 'slots' in our thread local storage stack. | 33 // The maximum number of 'slots' in our thread local storage stack. |
33 const int kThreadLocalStorageSize = 64; | 34 const int kThreadLocalStorageSize = 64; |
34 | 35 |
35 // The maximum number of times to try to clear slots by calling destructors. | 36 // The maximum number of times to try to clear slots by calling destructors. |
36 // Use pthread naming convention for clarity. | 37 // Use pthread naming convention for clarity. |
37 const int kMaxDestructorIterations = kThreadLocalStorageSize; | 38 const int kMaxDestructorIterations = kThreadLocalStorageSize; |
38 | 39 |
39 // An array of destructor function pointers for the slots. If a slot has a | 40 // 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 // 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 // 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 // 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 // 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 // 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 // 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 // (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 // hurt the racy calls to the destructors on another thread. |
48 volatile base::ThreadLocalStorage::TLSDestructorFunc | 49 volatile base::ThreadLocalStorage::TLSDestructorFunc |
49 g_tls_destructors[kThreadLocalStorageSize]; | 50 g_tls_destructors[kThreadLocalStorageSize]; |
50 | 51 |
51 void** ConstructTlsVector() { | 52 void** ConstructTlsVector() { |
jar (doing other things)
2013/11/26 20:05:24
Please add a comment:
This function is called to
michaelbai
2013/11/27 00:29:26
Done.
| |
52 if (g_native_tls_key == TLS_OUT_OF_INDEXES) { | 53 if (g_native_tls_key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES) { |
jar (doing other things)
2013/11/26 20:05:24
I think you need an atomic access ofr g_native_tls
michaelbai
2013/11/27 00:29:26
Done.
| |
53 long value = TlsAlloc(); | 54 PlatformThreadLocalStorage::TLSKey key; |
54 DCHECK(value != TLS_OUT_OF_INDEXES); | 55 CHECK(PlatformThreadLocalStorage::AllocTLS(&key)); |
55 | 56 |
57 // The TLS_KEY_OUT_OF_INDEXES is used to find out whether the key is set or | |
58 // not in NoBarrier_CompareAndSwap, but Posix doesn't have invalid key, we | |
59 // define an almost impossible value be it. | |
60 // If we really get TLS_KEY_OUT_OF_INDEXES as value of key, just alloc | |
61 // another TLS slot. | |
62 if (key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES) { | |
63 PlatformThreadLocalStorage::TLSKey tmp = key; | |
64 CHECK(PlatformThreadLocalStorage::AllocTLS(&key) && | |
65 key != PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES); | |
66 PlatformThreadLocalStorage::FreeTLS(tmp); | |
jar (doing other things)
2013/11/26 20:05:24
This and line 78 are perhaps your only use of Free
michaelbai
2013/11/27 00:29:26
Right, I want it be clean, it looks weird if there
| |
67 } | |
56 // Atomically test-and-set the tls_key. If the key is TLS_OUT_OF_INDEXES, | 68 // Atomically test-and-set the tls_key. If the key is TLS_OUT_OF_INDEXES, |
jar (doing other things)
2013/11/26 20:05:24
This comment about TLS_OUT_OF_INDEXES seems to hav
michaelbai
2013/11/27 00:29:26
I changed TLS_OUT_OF_INDEXES to TLS_KEY_OUT_OF_IND
| |
57 // go ahead and set it. Otherwise, do nothing, as another | 69 // go ahead and set it. Otherwise, do nothing, as another |
58 // thread already did our dirty work. | 70 // thread already did our dirty work. |
59 if (TLS_OUT_OF_INDEXES != InterlockedCompareExchange( | 71 if (PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES != |
60 &g_native_tls_key, value, TLS_OUT_OF_INDEXES)) { | 72 base::subtle::NoBarrier_CompareAndSwap( |
73 reinterpret_cast<base::subtle::Atomic32*>(&g_native_tls_key), | |
74 PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES, key)) { | |
61 // We've been shortcut. Another thread replaced g_native_tls_key first so | 75 // 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 | 76 // we need to destroy our index and use the one the other thread got |
63 // first. | 77 // first. |
64 TlsFree(value); | 78 PlatformThreadLocalStorage::FreeTLS(key); |
65 } | 79 } |
66 } | 80 } |
67 DCHECK(!TlsGetValue(g_native_tls_key)); | 81 DCHECK(!PlatformThreadLocalStorage::GetTLSValue(g_native_tls_key)); |
jar (doing other things)
2013/11/26 20:05:24
I think that once you declare a value to be Atomic
michaelbai
2013/11/27 00:29:26
Thanks, Done.
| |
68 | 82 |
69 // Some allocators, such as TCMalloc, make use of thread local storage. | 83 // 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 | 84 // 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 | 85 // 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, | 86 // 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: | 87 // 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 | 88 // 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 | 89 // allocator until our service is in place. (i.e., don't even call new until |
76 // after we're setup) | 90 // after we're setup) |
77 void* stack_allocated_tls_data[kThreadLocalStorageSize]; | 91 void* stack_allocated_tls_data[kThreadLocalStorageSize]; |
78 memset(stack_allocated_tls_data, 0, sizeof(stack_allocated_tls_data)); | 92 memset(stack_allocated_tls_data, 0, sizeof(stack_allocated_tls_data)); |
79 // Ensure that any rentrant calls change the temp version. | 93 // Ensure that any rentrant calls change the temp version. |
80 TlsSetValue(g_native_tls_key, stack_allocated_tls_data); | 94 PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key, |
95 stack_allocated_tls_data); | |
81 | 96 |
82 // Allocate an array to store our data. | 97 // Allocate an array to store our data. |
83 void** tls_data = new void*[kThreadLocalStorageSize]; | 98 void** tls_data = new void*[kThreadLocalStorageSize]; |
84 memcpy(tls_data, stack_allocated_tls_data, sizeof(stack_allocated_tls_data)); | 99 memcpy(tls_data, stack_allocated_tls_data, sizeof(stack_allocated_tls_data)); |
85 TlsSetValue(g_native_tls_key, tls_data); | 100 PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key, tls_data); |
86 return tls_data; | 101 return tls_data; |
87 } | 102 } |
88 | 103 |
89 // Called when we terminate a thread, this function calls any TLS destructors | 104 } // namespace |
90 // that are pending for this thread. | 105 |
91 void WinThreadExit() { | 106 namespace base { |
92 if (g_native_tls_key == TLS_OUT_OF_INDEXES) | 107 |
108 namespace internal { | |
109 | |
110 void PlatformThreadLocalStorage::OnThreadExit() { | |
111 if (g_native_tls_key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES) | |
93 return; | 112 return; |
94 | 113 |
95 void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key)); | 114 void** tls_data = static_cast<void**>(GetTLSValue(g_native_tls_key)); |
115 | |
96 // Maybe we have never initialized TLS for this thread. | 116 // Maybe we have never initialized TLS for this thread. |
97 if (!tls_data) | 117 if (!tls_data) |
98 return; | 118 return; |
99 | 119 |
100 // Some allocators, such as TCMalloc, use TLS. As a result, when a thread | 120 // 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 | 121 // 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 | 122 // 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 | 123 // known destructors (perchance including an allocator), that we don't call |
104 // the allocator and cause it to resurrect itself (with no possibly destructor | 124 // the allocator and cause it to resurrect itself (with no possibly destructor |
105 // call to follow). We handle this problem as follows: | 125 // call to follow). We handle this problem as follows: |
106 // Switch to using a stack allocated vector, so that we don't have dependence | 126 // 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 | 127 // 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.) | 128 // even call delete[] after we're done with destructors.) |
109 void* stack_allocated_tls_data[kThreadLocalStorageSize]; | 129 void* stack_allocated_tls_data[kThreadLocalStorageSize]; |
110 memcpy(stack_allocated_tls_data, tls_data, sizeof(stack_allocated_tls_data)); | 130 memcpy(stack_allocated_tls_data, tls_data, sizeof(stack_allocated_tls_data)); |
111 // Ensure that any re-entrant calls change the temp version. | 131 // Ensure that any re-entrant calls change the temp version. |
112 TlsSetValue(g_native_tls_key, stack_allocated_tls_data); | 132 PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key, |
133 stack_allocated_tls_data); | |
113 delete[] tls_data; // Our last dependence on an allocator. | 134 delete[] tls_data; // Our last dependence on an allocator. |
114 | 135 |
115 int remaining_attempts = kMaxDestructorIterations; | 136 int remaining_attempts = kMaxDestructorIterations; |
116 bool need_to_scan_destructors = true; | 137 bool need_to_scan_destructors = true; |
117 while (need_to_scan_destructors) { | 138 while (need_to_scan_destructors) { |
118 need_to_scan_destructors = false; | 139 need_to_scan_destructors = false; |
119 // Try to destroy the first-created-slot (which is slot 1) in our last | 140 // 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 | 141 // 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 | 142 // 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, | 143 // allocator) and should also be destroyed last. If we get the order wrong, |
(...skipping 14 matching lines...) Expand all Loading... | |
137 // the whole vector again. This is a pthread standard. | 158 // the whole vector again. This is a pthread standard. |
138 need_to_scan_destructors = true; | 159 need_to_scan_destructors = true; |
139 } | 160 } |
140 if (--remaining_attempts <= 0) { | 161 if (--remaining_attempts <= 0) { |
141 NOTREACHED(); // Destructors might not have been called. | 162 NOTREACHED(); // Destructors might not have been called. |
142 break; | 163 break; |
143 } | 164 } |
144 } | 165 } |
145 | 166 |
146 // Remove our stack allocated vector. | 167 // Remove our stack allocated vector. |
147 TlsSetValue(g_native_tls_key, NULL); | 168 PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key, NULL); |
148 } | 169 } |
149 | 170 |
150 } // namespace | 171 } // namespace internal |
151 | 172 |
152 namespace base { | |
153 | 173 |
154 ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) { | 174 ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) { |
155 initialized_ = false; | 175 initialized_ = false; |
156 slot_ = 0; | 176 slot_ = 0; |
157 Initialize(destructor); | 177 Initialize(destructor); |
158 } | 178 } |
159 | 179 |
160 bool ThreadLocalStorage::StaticSlot::Initialize(TLSDestructorFunc destructor) { | 180 bool ThreadLocalStorage::StaticSlot::Initialize(TLSDestructorFunc destructor) { |
161 if (g_native_tls_key == TLS_OUT_OF_INDEXES || !TlsGetValue(g_native_tls_key)) | 181 if (g_native_tls_key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES || |
182 !PlatformThreadLocalStorage::GetTLSValue(g_native_tls_key)) | |
162 ConstructTlsVector(); | 183 ConstructTlsVector(); |
163 | 184 |
164 // Grab a new slot. | 185 // Grab a new slot. |
165 slot_ = InterlockedIncrement(&g_last_used_tls_key); | 186 slot_ = base::subtle::NoBarrier_AtomicIncrement( |
187 reinterpret_cast<base::subtle::Atomic32*>(&g_last_used_tls_key), 1); | |
166 DCHECK_GT(slot_, 0); | 188 DCHECK_GT(slot_, 0); |
167 if (slot_ >= kThreadLocalStorageSize) { | 189 if (slot_ >= kThreadLocalStorageSize) { |
168 NOTREACHED(); | 190 NOTREACHED(); |
169 return false; | 191 return false; |
170 } | 192 } |
171 | 193 |
172 // Setup our destructor. | 194 // Setup our destructor. |
173 g_tls_destructors[slot_] = destructor; | 195 g_tls_destructors[slot_] = destructor; |
174 initialized_ = true; | 196 initialized_ = true; |
175 return true; | 197 return true; |
176 } | 198 } |
177 | 199 |
178 void ThreadLocalStorage::StaticSlot::Free() { | 200 void ThreadLocalStorage::StaticSlot::Free() { |
179 // At this time, we don't reclaim old indices for TLS slots. | 201 // At this time, we don't reclaim old indices for TLS slots. |
180 // So all we need to do is wipe the destructor. | 202 // So all we need to do is wipe the destructor. |
181 DCHECK_GT(slot_, 0); | 203 DCHECK_GT(slot_, 0); |
182 DCHECK_LT(slot_, kThreadLocalStorageSize); | 204 DCHECK_LT(slot_, kThreadLocalStorageSize); |
183 g_tls_destructors[slot_] = NULL; | 205 g_tls_destructors[slot_] = NULL; |
184 slot_ = 0; | 206 slot_ = 0; |
185 initialized_ = false; | 207 initialized_ = false; |
186 } | 208 } |
187 | 209 |
188 void* ThreadLocalStorage::StaticSlot::Get() const { | 210 void* ThreadLocalStorage::StaticSlot::Get() const { |
189 void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key)); | 211 void** tls_data = static_cast<void**>( |
212 PlatformThreadLocalStorage::GetTLSValue(g_native_tls_key)); | |
190 if (!tls_data) | 213 if (!tls_data) |
191 tls_data = ConstructTlsVector(); | 214 tls_data = ConstructTlsVector(); |
192 DCHECK_GT(slot_, 0); | 215 DCHECK_GT(slot_, 0); |
193 DCHECK_LT(slot_, kThreadLocalStorageSize); | 216 DCHECK_LT(slot_, kThreadLocalStorageSize); |
194 return tls_data[slot_]; | 217 return tls_data[slot_]; |
195 } | 218 } |
196 | 219 |
197 void ThreadLocalStorage::StaticSlot::Set(void* value) { | 220 void ThreadLocalStorage::StaticSlot::Set(void* value) { |
198 void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key)); | 221 void** tls_data = static_cast<void**>( |
222 PlatformThreadLocalStorage::GetTLSValue(g_native_tls_key)); | |
199 if (!tls_data) | 223 if (!tls_data) |
200 tls_data = ConstructTlsVector(); | 224 tls_data = ConstructTlsVector(); |
201 DCHECK_GT(slot_, 0); | 225 DCHECK_GT(slot_, 0); |
202 DCHECK_LT(slot_, kThreadLocalStorageSize); | 226 DCHECK_LT(slot_, kThreadLocalStorageSize); |
203 tls_data[slot_] = value; | 227 tls_data[slot_] = value; |
204 } | 228 } |
205 | 229 |
206 } // namespace base | 230 } // 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" | |
OLD | NEW |