Implement chromium's TLS.

Using one system TLS to implement multiple chrome's TLS slots.

BUG=264406

Review URL: https://codereview.chromium.org/60743004

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@241144 0039d316-1c4b-4281-b951-d872f2087c98

base/base.gypi

@@ -589,6 +589,7 @@
           'threading/thread_id_name_manager.h',
           'threading/thread_local.h',
           'threading/thread_local_posix.cc',
+          'threading/thread_local_storage.cc',
           'threading/thread_local_storage.h',
           'threading/thread_local_storage_posix.cc',
           'threading/thread_local_storage_win.cc',

base/threading/thread_local_storage.h

@@ -8,12 +8,71 @@
 #include "base/base_export.h"
 #include "base/basictypes.h"
 
-#if defined(OS_POSIX)
+#if defined(OS_WIN)
+#include <windows.h>
+#elif defined(OS_POSIX)
 #include <pthread.h>
 #endif
 
 namespace base {
 
+namespace internal {
+
+// WARNING: You should *NOT* be using this class directly.
+// PlatformThreadLocalStorage is low-level abstraction to the OS's TLS
+// interface, you should instead be using ThreadLocalStorage::StaticSlot/Slot.
+class BASE_EXPORT PlatformThreadLocalStorage {
+ public:
+
+#if defined(OS_WIN)
+  typedef unsigned long TLSKey;
+  enum { TLS_KEY_OUT_OF_INDEXES = TLS_OUT_OF_INDEXES };
+#elif defined(OS_POSIX)
+  typedef pthread_key_t TLSKey;
+  // The following is a "reserved key" which is used in our generic Chromium
+  // ThreadLocalStorage implementation.  We expect that an OS will not return
+  // such a key, but if it is returned (i.e., the OS tries to allocate it) we
+  // will just request another key.
+  enum { TLS_KEY_OUT_OF_INDEXES = 0x7FFFFFFF };
+#endif
+
+  // The following methods need to be supported on each OS platform, so that
+  // the Chromium ThreadLocalStore functionality can be constructed.
+  // Chromium will use these methods to acquire a single OS slot, and then use
+  // that to support a much larger number of Chromium slots (independent of the
+  // OS restrictions).
+  // The following returns true if it successfully is able to return an OS
+  // key in |key|.
+  static bool AllocTLS(TLSKey* key);
+  // Note: FreeTLS() doesn't have to be called, it is fine with this leak, OS
+  // might not reuse released slot, you might just reset the TLS value with
+  // SetTLSValue().
+  static void FreeTLS(TLSKey key);
+  static void SetTLSValue(TLSKey key, void* value);
+  static void* GetTLSValue(TLSKey key);
+
+  // Each platform (OS implementation) is required to call this method on each
+  // terminating thread when the thread is about to terminate.  This method
+  // will then call all registered destructors for slots in Chromium
+  // ThreadLocalStorage, until there are no slot values remaining as having
+  // been set on this thread.
+  // Destructors may end up being called multiple times on a terminating
+  // thread, as other destructors may re-set slots that were previously
+  // destroyed.
+#if defined(OS_WIN)
+  // Since Windows which doesn't support TLS destructor, the implementation
+  // should use GetTLSValue() to retrieve the value of TLS slot.
+  static void OnThreadExit();
+#elif defined(OS_POSIX)
+  // |Value| is the data stored in TLS slot, The implementation can't use
+  // GetTLSValue() to retrieve the value of slot as it has already been reset
+  // in Posix.
+  static void OnThreadExit(void* value);
+#endif
+};
+
+}  // namespace internal
+
 // Wrapper for thread local storage.  This class doesn't do much except provide
 // an API for portability.
 class BASE_EXPORT ThreadLocalStorage {
@@ -60,12 +119,7 @@ class BASE_EXPORT ThreadLocalStorage {
 
     // The internals of this struct should be considered private.
     bool initialized_;
-#if defined(OS_WIN)
     int slot_;
-#elif defined(OS_POSIX)
-    pthread_key_t key_;
-#endif
-
   };
 
   // A convenience wrapper around StaticSlot with a constructor. Can be used
@@ -77,11 +131,8 @@ class BASE_EXPORT ThreadLocalStorage {
 
    private:
     using StaticSlot::initialized_;
-#if defined(OS_WIN)
     using StaticSlot::slot_;
-#elif defined(OS_POSIX)
-    using StaticSlot::key_;
-#endif
+
     DISALLOW_COPY_AND_ASSIGN(Slot);
   };
 

base/threading/thread_local_storage_posix.cc

@@ -8,42 +8,25 @@
 
 namespace base {
 
-ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) {
-  initialized_ = false;
-  key_ = 0;
-  Initialize(destructor);
-}
-
-bool ThreadLocalStorage::StaticSlot::Initialize(TLSDestructorFunc destructor) {
-  DCHECK(!initialized_);
-  int error = pthread_key_create(&key_, destructor);
-  if (error) {
-    NOTREACHED();
-    return false;
-  }
+namespace internal {
 
-  initialized_ = true;
-  return true;
+bool PlatformThreadLocalStorage::AllocTLS(TLSKey* key) {
+  return !pthread_key_create(key,
+      base::internal::PlatformThreadLocalStorage::OnThreadExit);
 }
 
-void ThreadLocalStorage::StaticSlot::Free() {
-  DCHECK(initialized_);
-  int error = pthread_key_delete(key_);
-  if (error)
-    NOTREACHED();
-  initialized_ = false;
+void PlatformThreadLocalStorage::FreeTLS(TLSKey key) {
+  DCHECK_EQ(pthread_key_delete(key), 0);
 }
 
-void* ThreadLocalStorage::StaticSlot::Get() const {
-  DCHECK(initialized_);
-  return pthread_getspecific(key_);
+void* PlatformThreadLocalStorage::GetTLSValue(TLSKey key) {
+  return pthread_getspecific(key);
 }
 
-void ThreadLocalStorage::StaticSlot::Set(void* value) {
-  DCHECK(initialized_);
-  int error = pthread_setspecific(key_, value);
-  if (error)
-    NOTREACHED();
+void PlatformThreadLocalStorage::SetTLSValue(TLSKey key, void* value) {
+  DCHECK_EQ(pthread_setspecific(key, value), 0);
 }
 
+}  // namespace internal
+
 }  // namespace base

base/threading/thread_local_storage_win.cc

@@ -8,201 +8,33 @@
 
 #include "base/logging.h"
 
-
-namespace {
-// In order to make TLS destructors work, we need to keep function
-// pointers to the destructor for each TLS that we allocate.
-// We make this work by allocating a single OS-level TLS, which
-// contains an array of slots for the application to use.  In
-// parallel, we also allocate an array of destructors, which we
-// keep track of and call when threads terminate.
-
-// 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;
-
-// 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
-// assigned to the instance variable slot_ in a ThreadLocalStorage::Slot
-// instance.  We reserve the value slot_ == 0 to indicate that the corresponding
-// 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;
-
-// The maximum number of 'slots' in our thread local storage stack.
-const int kThreadLocalStorageSize = 64;
-
-// The maximum number of times to try to clear slots by calling destructors.
-// Use pthread naming convention for clarity.
-const int kMaxDestructorIterations = kThreadLocalStorageSize;
-
-// An array of destructor function pointers for the slots.  If a slot has a
-// destructor, it will be stored in its corresponding entry in this array.
-// The elements are volatile to ensure that when the compiler reads the value
-// to potentially call the destructor, it does so once, and that value is tested
-// for null-ness and then used. Yes, that would be a weird de-optimization,
-// but I can imagine some register machines where it was just as easy to
-// re-fetch an array element, and I want to be sure a call to free the key
-// (i.e., null out the destructor entry) that happens on a separate thread can't
-// hurt the racy calls to the destructors on another thread.
-volatile base::ThreadLocalStorage::TLSDestructorFunc
-    g_tls_destructors[kThreadLocalStorageSize];
-
-void** ConstructTlsVector() {
-  if (g_native_tls_key == TLS_OUT_OF_INDEXES) {
-    long value = TlsAlloc();
-    DCHECK(value != TLS_OUT_OF_INDEXES);
-
-    // Atomically test-and-set the tls_key.  If the key is TLS_OUT_OF_INDEXES,
-    // 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)) {
-      // 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);
-    }
-  }
-  DCHECK(!TlsGetValue(g_native_tls_key));
-
-  // 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
-  // system to initialize, which will include registering for a TLS key.  If we
-  // are not careful here, then that request to create a key will call new back,
-  // and we'll have an infinite loop.  We avoid that as follows:
-  // Use a stack allocated vector, so that we don't have dependence on our
-  // allocator until our service is in place.  (i.e., don't even call new until
-  // after we're setup)
-  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);
-
-  // 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);
-  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)
-    return;
-
-  void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key));
-  // Maybe we have never initialized TLS for this thread.
-  if (!tls_data)
-    return;
-
-  // Some allocators, such as TCMalloc, use TLS.  As a result, when a thread
-  // terminates, one of the destructor calls we make may be to shut down an
-  // allocator.  We have to be careful that after we've shutdown all of the
-  // known destructors (perchance including an allocator), that we don't call
-  // the allocator and cause it to resurrect itself (with no possibly destructor
-  // call to follow).  We handle this problem as follows:
-  // Switch to using a stack allocated vector, so that we don't have dependence
-  // on our allocator after we have called all g_tls_destructors.  (i.e., don't
-  // even call delete[] after we're done with destructors.)
-  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);
-  delete[] tls_data;  // Our last dependence on an allocator.
-
-  int remaining_attempts = kMaxDestructorIterations;
-  bool need_to_scan_destructors = true;
-  while (need_to_scan_destructors) {
-    need_to_scan_destructors = false;
-    // Try to destroy the first-created-slot (which is slot 1) in our last
-    // destructor call.  That user was able to function, and define a slot with
-    // no other services running, so perhaps it is a basic service (like an
-    // allocator) and should also be destroyed last.  If we get the order wrong,
-    // then we'll itterate several more times, so it is really not that
-    // critical (but it might help).
-    for (int slot = g_last_used_tls_key; slot > 0; --slot) {
-      void* value = stack_allocated_tls_data[slot];
-      if (value == NULL)
-        continue;
-      base::ThreadLocalStorage::TLSDestructorFunc destructor =
-          g_tls_destructors[slot];
-      if (destructor == NULL)
-        continue;
-      stack_allocated_tls_data[slot] = NULL;  // pre-clear the slot.
-      destructor(value);
-      // Any destructor might have called a different service, which then set
-      // a different slot to a non-NULL value.  Hence we need to check
-      // the whole vector again.  This is a pthread standard.
-      need_to_scan_destructors = true;
-    }
-    if (--remaining_attempts <= 0) {
-      NOTREACHED();  // Destructors might not have been called.
-      break;
-    }
-  }
-
-  // Remove our stack allocated vector.
-  TlsSetValue(g_native_tls_key, NULL);
-}
-
-}  // namespace
-
 namespace base {
 
-ThreadLocalStorage::Slot::Slot(TLSDestructorFunc destructor) {
-  initialized_ = false;
-  slot_ = 0;
-  Initialize(destructor);
-}
-
-bool ThreadLocalStorage::StaticSlot::Initialize(TLSDestructorFunc destructor) {
-  if (g_native_tls_key == TLS_OUT_OF_INDEXES || !TlsGetValue(g_native_tls_key))
-    ConstructTlsVector();
+namespace internal {
 
-  // Grab a new slot.
-  slot_ = InterlockedIncrement(&g_last_used_tls_key);
-  DCHECK_GT(slot_, 0);
-  if (slot_ >= kThreadLocalStorageSize) {
-    NOTREACHED();
-    return false;
+bool PlatformThreadLocalStorage::AllocTLS(TLSKey* key) {
+  TLSKey value = TlsAlloc();
+  if (value != TLS_OUT_OF_INDEXES) {
+    *key = value;
+    return true;
   }
-
-  // Setup our destructor.
-  g_tls_destructors[slot_] = destructor;
-  initialized_ = true;
-  return true;
+  return false;
 }
 
-void ThreadLocalStorage::StaticSlot::Free() {
-  // At this time, we don't reclaim old indices for TLS slots.
-  // So all we need to do is wipe the destructor.
-  DCHECK_GT(slot_, 0);
-  DCHECK_LT(slot_, kThreadLocalStorageSize);
-  g_tls_destructors[slot_] = NULL;
-  slot_ = 0;
-  initialized_ = false;
+void PlatformThreadLocalStorage::FreeTLS(TLSKey key) {
+  DCHECK(TlsFree(key));
 }
 
-void* ThreadLocalStorage::StaticSlot::Get() const {
-  void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key));
-  if (!tls_data)
-    tls_data = ConstructTlsVector();
-  DCHECK_GT(slot_, 0);
-  DCHECK_LT(slot_, kThreadLocalStorageSize);
-  return tls_data[slot_];
+void* PlatformThreadLocalStorage::GetTLSValue(TLSKey key) {
+  return TlsGetValue(key);
 }
 
-void ThreadLocalStorage::StaticSlot::Set(void* value) {
-  void** tls_data = static_cast<void**>(TlsGetValue(g_native_tls_key));
-  if (!tls_data)
-    tls_data = ConstructTlsVector();
-  DCHECK_GT(slot_, 0);
-  DCHECK_LT(slot_, kThreadLocalStorageSize);
-  tls_data[slot_] = value;
+void PlatformThreadLocalStorage::SetTLSValue(TLSKey key, void* value) {
+  DCHECK(TlsSetValue(key, value));
 }
 
+}  // namespace internal
+
 }  // namespace base
 
 // Thread Termination Callbacks.
@@ -233,7 +65,7 @@ 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();
+    base::internal::PlatformThreadLocalStorage::OnThreadExit();
 }
 
 // .CRT$XLA to .CRT$XLZ is an array of PIMAGE_TLS_CALLBACK pointers that are