Implement chromium's TLS.

base/threading/thread_local_storage.cc

Index: base/threading/thread_local_storage.cc
diff --git a/base/threading/thread_local_storage_win.cc b/base/threading/thread_local_storage.cc
similarity index 65%
copy from base/threading/thread_local_storage_win.cc
copy to base/threading/thread_local_storage.cc
index 0ae3cb4c8cd39471b9b532a5dc87e2b5d61647a8..1b42fe6a031f0a772934f7fd0328c1a3e35f5358 100644
--- a/base/threading/thread_local_storage_win.cc
+++ b/base/threading/thread_local_storage.cc
@@ -4,10 +4,10 @@
 
 #include "base/threading/thread_local_storage.h"
 
-#include <windows.h>
-
+#include "base/atomicops.h"
 #include "base/logging.h"
 
+using base::internal::PlatformThreadLocalStorage;
 
 namespace {
 // In order to make TLS destructors work, we need to keep function
@@ -18,7 +18,8 @@ namespace {
 // 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:

In order to make TLS destructors work, we need to keep around a function pointer
to the destructor for each slot. We keep this array of pointers in a global
(static) array.  
We use the single OS-level TLS slot (giving us one pointer per thread) to hold a
pointer to a per-thread array (table) of slots that we allocate to Chromium
consumers.

[michaelbai, 2013/11/27 00:29:26]

Done.

 
 // g_native_tls_key is the one native TLS that we use.  It stores our table.
-long g_native_tls_key = TLS_OUT_OF_INDEXES;
+PlatformThreadLocalStorage::TLSKey g_native_tls_key =
+    PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES;
 
 // g_last_used_tls_key is the high-water-mark of allocated thread local storage.
 // Each allocation is an index into our g_tls_destructors[].  Each such index is
@@ -27,7 +28,7 @@ long g_native_tls_key = TLS_OUT_OF_INDEXES;
 // instance of ThreadLocalStorage::Slot has been freed (i.e., destructor called,
 // etc.).  This reserved use of 0 is then stated as the initial value of
 // g_last_used_tls_key, so that the first issued index will be 1.
-long g_last_used_tls_key = 0;
+base::subtle::Atomic32 g_last_used_tls_key = 0;
 
 // The maximum number of 'slots' in our thread local storage stack.
 const int kThreadLocalStorageSize = 64;
@@ -49,22 +50,35 @@ volatile base::ThreadLocalStorage::TLSDestructorFunc
     g_tls_destructors[kThreadLocalStorageSize];
 
 void** ConstructTlsVector() {

[jar (doing other things), 2013/11/26 20:05:24]

Please add a comment:

This function is called to initialize our entire Chromium TLS system.
It may be called very early, and we need to complete most all of the setup
(initialization) before calling *any* memory allocator functions, which may
recursively depend on this initialization.
As a result, we use Atomics, and avoid anything (like a singleton) that might
require memory allocations.

[michaelbai, 2013/11/27 00:29:26]

Done.

-  if (g_native_tls_key == TLS_OUT_OF_INDEXES) {
-    long value = TlsAlloc();
-    DCHECK(value != TLS_OUT_OF_INDEXES);
-
+  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_key here.

[michaelbai, 2013/11/27 00:29:26]

Done.

+    PlatformThreadLocalStorage::TLSKey key;
+    CHECK(PlatformThreadLocalStorage::AllocTLS(&key));
+
+    // The TLS_KEY_OUT_OF_INDEXES is used to find out whether the key is set or
+    // not in NoBarrier_CompareAndSwap, but Posix doesn't have invalid key, we
+    // define an almost impossible value be it.
+    // If we really get TLS_KEY_OUT_OF_INDEXES as value of key, just alloc
+    // another TLS slot.
+    if (key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES) {
+      PlatformThreadLocalStorage::TLSKey tmp = key;
+      CHECK(PlatformThreadLocalStorage::AllocTLS(&key) &&
+            key != PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES);
+      PlatformThreadLocalStorage::FreeTLS(tmp);

[jar (doing other things), 2013/11/26 20:05:24]

This and line 78 are perhaps your only use of FreeTLS... and it is probably
better to leak this value.  I don't expect this code to be used... and I wonder
if we *do* free the value, will be be reallocated?  I suspect not... 

Maybe it is "nicer" to be clean and call FreeTLS.

Note that at process teardown, we do *not* FreeTLS... to we are fine with
leaking some of this.

[michaelbai, 2013/11/27 00:29:26]

Right, I want it be clean, it looks weird if there is no FreeTLS there, 


On 2013/11/26 20:05:24, jar wrote:

+    }
     // 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 have changed with your new code.

[michaelbai, 2013/11/27 00:29:26]

I changed TLS_OUT_OF_INDEXES to TLS_KEY_OUT_OF_INDEXES, otherwise I didn't do
anything different

On 2013/11/26 20:05:24, jar wrote:

     // go ahead and set it.  Otherwise, do nothing, as another
     // thread already did our dirty work.
-    if (TLS_OUT_OF_INDEXES != InterlockedCompareExchange(
-            &g_native_tls_key, value, TLS_OUT_OF_INDEXES)) {
+    if (PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES !=
+        base::subtle::NoBarrier_CompareAndSwap(
+            reinterpret_cast<base::subtle::Atomic32*>(&g_native_tls_key),
+            PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES, key)) {
       // We've been shortcut. Another thread replaced g_native_tls_key first so
       // we need to destroy our index and use the one the other thread got
       // first.
-      TlsFree(value);
+      PlatformThreadLocalStorage::FreeTLS(key);
     }
   }
-  DCHECK(!TlsGetValue(g_native_tls_key));
+  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, all accesses much use the
Atomic functions.  My guess is that you want to do an Atomic load... and then
keep it in a local for uses such as on line 94.

[michaelbai, 2013/11/27 00:29:26]

Thanks, Done.

 
   // Some allocators, such as TCMalloc, make use of thread local storage.
   // As a result, any attempt to call new (or malloc) will lazily cause such a
@@ -77,22 +91,28 @@ void** ConstructTlsVector() {
   void* stack_allocated_tls_data[kThreadLocalStorageSize];
   memset(stack_allocated_tls_data, 0, sizeof(stack_allocated_tls_data));
   // Ensure that any rentrant calls change the temp version.
-  TlsSetValue(g_native_tls_key, stack_allocated_tls_data);
+  PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key,
+                                          stack_allocated_tls_data);
 
   // Allocate an array to store our data.
   void** tls_data = new void*[kThreadLocalStorageSize];
   memcpy(tls_data, stack_allocated_tls_data, sizeof(stack_allocated_tls_data));
-  TlsSetValue(g_native_tls_key, tls_data);
+  PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key, tls_data);
   return tls_data;
 }
 
-// Called when we terminate a thread, this function calls any TLS destructors
-// that are pending for this thread.
-void WinThreadExit() {
-  if (g_native_tls_key == TLS_OUT_OF_INDEXES)
+}  // namespace
+
+namespace base {
+
+namespace internal {
+
+void PlatformThreadLocalStorage::OnThreadExit() {
+  if (g_native_tls_key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES)
     return;
 
-  void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key));
+  void** tls_data = static_cast<void**>(GetTLSValue(g_native_tls_key));
+
   // Maybe we have never initialized TLS for this thread.
   if (!tls_data)
     return;
@@ -109,7 +129,8 @@ void WinThreadExit() {
   void* stack_allocated_tls_data[kThreadLocalStorageSize];
   memcpy(stack_allocated_tls_data, tls_data, sizeof(stack_allocated_tls_data));
   // Ensure that any re-entrant calls change the temp version.
-  TlsSetValue(g_native_tls_key, stack_allocated_tls_data);
+  PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key,
+                                          stack_allocated_tls_data);
   delete[] tls_data;  // Our last dependence on an allocator.
 
   int remaining_attempts = kMaxDestructorIterations;
@@ -144,12 +165,11 @@ void WinThreadExit() {
   }
 
   // Remove our stack allocated vector.
-  TlsSetValue(g_native_tls_key, NULL);
+  PlatformThreadLocalStorage::SetTLSValue(g_native_tls_key, NULL);
 }
 
-}  // namespace
+}  // namespace internal
 
-namespace base {
 
 ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) {
   initialized_ = false;
@@ -158,11 +178,13 @@ ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) {
 }
 
 bool ThreadLocalStorage::StaticSlot::Initialize(TLSDestructorFunc destructor) {
-  if (g_native_tls_key == TLS_OUT_OF_INDEXES || !TlsGetValue(g_native_tls_key))
+  if (g_native_tls_key == PlatformThreadLocalStorage::TLS_KEY_OUT_OF_INDEXES ||
+      !PlatformThreadLocalStorage::GetTLSValue(g_native_tls_key))
     ConstructTlsVector();
 
   // Grab a new slot.
-  slot_ = InterlockedIncrement(&g_last_used_tls_key);
+  slot_ = base::subtle::NoBarrier_AtomicIncrement(
+      reinterpret_cast<base::subtle::Atomic32*>(&g_last_used_tls_key), 1);
   DCHECK_GT(slot_, 0);
   if (slot_ >= kThreadLocalStorageSize) {
     NOTREACHED();
@@ -186,7 +208,8 @@ void ThreadLocalStorage::StaticSlot::Free() {
 }
 
 void* ThreadLocalStorage::StaticSlot::Get() const {
-  void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key));
+  void** tls_data = static_cast<void**>(
+      PlatformThreadLocalStorage::GetTLSValue(g_native_tls_key));
   if (!tls_data)
     tls_data = ConstructTlsVector();
   DCHECK_GT(slot_, 0);
@@ -195,7 +218,8 @@ void* ThreadLocalStorage::StaticSlot::Get() const {
 }
 
 void ThreadLocalStorage::StaticSlot::Set(void* value) {
-  void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key));
+  void** tls_data = static_cast<void**>(
+      PlatformThreadLocalStorage::GetTLSValue(g_native_tls_key));
   if (!tls_data)
     tls_data = ConstructTlsVector();
   DCHECK_GT(slot_, 0);
@@ -204,74 +228,3 @@ void ThreadLocalStorage::StaticSlot::Set(void* value) {
 }
 
 }  // namespace base
-
-// Thread Termination Callbacks.
-// Windows doesn't support a per-thread destructor with its
-// TLS primitives.  So, we build it manually by inserting a
-// function to be called on each thread's exit.
-// This magic is from http://www.codeproject.com/threads/tls.asp
-// and it works for VC++ 7.0 and later.
-
-// Force a reference to _tls_used to make the linker create the TLS directory
-// if it's not already there.  (e.g. if __declspec(thread) is not used).
-// Force a reference to p_thread_callback_base to prevent whole program
-// optimization from discarding the variable.
-#ifdef _WIN64
-
-#pragma comment(linker, "/INCLUDE:_tls_used")
-#pragma comment(linker, "/INCLUDE:p_thread_callback_base")
-
-#else  // _WIN64
-
-#pragma comment(linker, "/INCLUDE:__tls_used")
-#pragma comment(linker, "/INCLUDE:_p_thread_callback_base")
-
-#endif  // _WIN64
-
-// Static callback function to call with each thread termination.
-void NTAPI OnThreadExit(PVOID module, DWORD reason, PVOID reserved) {
-  // On XP SP0 & SP1, the DLL_PROCESS_ATTACH is never seen. It is sent on SP2+
-  // and on W2K and W2K3. So don't assume it is sent.
-  if (DLL_THREAD_DETACH == reason || DLL_PROCESS_DETACH == reason)
-    WinThreadExit();
-}
-
-// .CRT$XLA to .CRT$XLZ is an array of PIMAGE_TLS_CALLBACK pointers that are
-// called automatically by the OS loader code (not the CRT) when the module is
-// loaded and on thread creation. They are NOT called if the module has been
-// loaded by a LoadLibrary() call. It must have implicitly been loaded at
-// process startup.
-// By implicitly loaded, I mean that it is directly referenced by the main EXE
-// or by one of its dependent DLLs. Delay-loaded DLL doesn't count as being
-// implicitly loaded.
-//
-// See VC\crt\src\tlssup.c for reference.
-
-// extern "C" suppresses C++ name mangling so we know the symbol name for the
-// linker /INCLUDE:symbol pragma above.
-extern "C" {
-// The linker must not discard p_thread_callback_base.  (We force a reference
-// to this variable with a linker /INCLUDE:symbol pragma to ensure that.) If
-// this variable is discarded, the OnThreadExit function will never be called.
-#ifdef _WIN64
-
-// .CRT section is merged with .rdata on x64 so it must be constant data.
-#pragma const_seg(".CRT$XLB")
-// When defining a const variable, it must have external linkage to be sure the
-// linker doesn't discard it.
-extern const PIMAGE_TLS_CALLBACK p_thread_callback_base;
-const PIMAGE_TLS_CALLBACK p_thread_callback_base = OnThreadExit;
-
-// Reset the default section.
-#pragma const_seg()
-
-#else  // _WIN64
-
-#pragma data_seg(".CRT$XLB")
-PIMAGE_TLS_CALLBACK p_thread_callback_base = OnThreadExit;
-
-// Reset the default section.
-#pragma data_seg()
-
-#endif  // _WIN64
-}  // extern "C"