Fast path for ThreadSpecific for main thread on TLS-slow platforms

third_party/WebKit/Source/wtf/StackUtil.cpp

Index: third_party/WebKit/Source/wtf/StackUtil.cpp
diff --git a/third_party/WebKit/Source/platform/heap/StackFrameDepth.cpp b/third_party/WebKit/Source/wtf/StackUtil.cpp
similarity index 58%
copy from third_party/WebKit/Source/platform/heap/StackFrameDepth.cpp
copy to third_party/WebKit/Source/wtf/StackUtil.cpp
index 1d164f510a36de4ecb3328e6c80065f4913eca5d..52eff68b07ddf7919677e845896be26f54445f97 100644
--- a/third_party/WebKit/Source/platform/heap/StackFrameDepth.cpp
+++ b/third_party/WebKit/Source/wtf/StackUtil.cpp
@@ -1,10 +1,12 @@
-// Copyright 2015 The Chromium Authors. All rights reserved.
+// Copyright 2017 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
-#include "platform/heap/StackFrameDepth.h"
+#include "wtf/StackUtil.h"
 
-#include "public/platform/Platform.h"
+#include "wtf/Assertions.h"
+#include "wtf/Threading.h"
+#include "wtf/WTFThreadData.h"
 
 #if OS(WIN)
 #include <stddef.h>
@@ -14,51 +16,30 @@
 extern "C" void* __libc_stack_end;  // NOLINT
 #endif
 
-namespace blink {
+namespace WTF {
 
-static const char* s_avoidOptimization = nullptr;
+namespace {
 
-// NEVER_INLINE ensures that |dummy| array on configureLimit() is not optimized
-// away, and the stack frame base register is adjusted |kSafeStackFrameSize|.
-NEVER_INLINE static uintptr_t currentStackFrameBaseOnCallee(const char* dummy) {
-  s_avoidOptimization = dummy;
-  return StackFrameDepth::currentStackFrame();
+uintptr_t mainThreadStackStart() {
+  static uintptr_t s_mainThreadStackStart =
+      reinterpret_cast<uintptr_t>(WTF::getStackStart()) - sizeof(void*);

[haraken, 2017/01/18 05:29:42]

Add a comment why we're subtracting sizeof(void*).

[Charlie Harrison, 2017/01/18 16:29:55]

I had to look at historical blame to figure this one out. I looks like
getStackStart is exclusive, not inclusive (e.g. it points past
the last page of the stack in linear order), so we subtract sizeof(void*) here
and do a >= comparison with this value.

+  return s_mainThreadStackStart;
 }
 
-uintptr_t StackFrameDepth::getFallbackStackLimit() {
-  // Allocate an |kSafeStackFrameSize|-sized object on stack and query
-  // stack frame base after it.
-  char dummy[kSafeStackFrameSize];
-
-  // Check that the stack frame can be used.
-  dummy[sizeof(dummy) - 1] = 0;
-  return currentStackFrameBaseOnCallee(dummy);
-}
-
-void StackFrameDepth::enableStackLimit() {
-  // All supported platforms will currently return a non-zero estimate,
-  // except if ASan is enabled.
-  size_t stackSize = getUnderestimatedStackSize();
-  if (!stackSize) {
-    m_stackFrameLimit = getFallbackStackLimit();
-    return;
+uintptr_t mainThreadUnderestimatedStackSize() {
+  static uintptr_t s_underestimatedStackSize;
+  if (s_underestimatedStackSize)
+    return s_underestimatedStackSize;
+  size_t underestimatedStackSize = getUnderestimatedStackSize();
+  if (underestimatedStackSize > sizeof(void*)) {
+    s_underestimatedStackSize = underestimatedStackSize - sizeof(void*);

[haraken, 2017/01/18 05:29:42]

Ditto.

[Charlie Harrison, 2017/01/18 16:29:55]

Done.

   }
-
-  static const int kStackRoomSize = 1024;
-
-  Address stackBase = reinterpret_cast<Address>(getStackStart());
-  RELEASE_ASSERT(stackSize > static_cast<const size_t>(kStackRoomSize));
-  size_t stackRoom = stackSize - kStackRoomSize;
-  RELEASE_ASSERT(stackBase > reinterpret_cast<Address>(stackRoom));
-  m_stackFrameLimit = reinterpret_cast<uintptr_t>(stackBase - stackRoom);
-
-  // If current stack use is already exceeding estimated limit, mark as
-  // disabled.
-  if (!isSafeToRecurse())
-    disableStackLimit();
+  return s_underestimatedStackSize;
 }
 
-size_t StackFrameDepth::getUnderestimatedStackSize() {
+}  // namespace
+
+size_t getUnderestimatedStackSize() {
 // FIXME: ASAN bot uses a fake stack as a thread stack frame,
 // and its size is different from the value which APIs tells us.
 #if defined(ADDRESS_SANITIZER)
@@ -127,14 +108,14 @@ size_t StackFrameDepth::getUnderestimatedStackSize() {
   }
   return pthread_get_stacksize_np(pthread_self());
 #elif OS(WIN) && COMPILER(MSVC)
-  return ThreadState::current()->threadStackSize();
+  return WTFThreadData::threadStackSize();
 #else
 #error "Stack frame size estimation not supported on this platform."
   return 0;
 #endif
 }
 
-void* StackFrameDepth::getStackStart() {
+void* getStackStart() {
 #if defined(__GLIBC__) || OS(ANDROID) || OS(FREEBSD)
   pthread_attr_t attr;
   int error;
@@ -162,7 +143,7 @@ void* StackFrameDepth::getStackStart() {
   // See https://code.google.com/p/nativeclient/issues/detail?id=3431.
   return __libc_stack_end;
 #else
-  ASSERT_NOT_REACHED();
+  NOTREACHED();
   return nullptr;
 #endif
 #elif OS(MACOSX)
@@ -181,4 +162,51 @@ void* StackFrameDepth::getStackStart() {
 #endif
 }
 
-}  // namespace blink
+namespace internal {
+
+bool mayNotBeMainThread() {

[haraken, 2017/01/18 05:29:42]

Can we move this function (and all functions called in this function) to a
header file? It's important that the function is inlined.

[Charlie Harrison, 2017/01/18 16:29:55]

Done, though I haven't inlined the initializers to the static variables which
will only be called once.

+  uintptr_t dummy;
+  uintptr_t addressDiff =
+      mainThreadStackStart() - reinterpret_cast<uintptr_t>(&dummy);
+  // This is a fast way to judge if we are in the main thread.
+  // If |&dummy| is within |s_mainThreadUnderestimatedStackSize| byte from
+  // the stack start of the main thread, we judge that we are in
+  // the main thread.
+  return addressDiff >= mainThreadUnderestimatedStackSize();
+}
+
+#if OS(WIN) && COMPILER(MSVC)
+size_t threadStackSize() {
+  // Notice that we cannot use the TIB's StackLimit for the stack end, as i
+  // tracks the end of the committed range. We're after the end of the reserved
+  // stack area (most of which will be uncommitted, most times.)
+  MEMORY_BASIC_INFORMATION stackInfo;
+  memset(&stackInfo, 0, sizeof(MEMORY_BASIC_INFORMATION));
+  size_t resultSize =
+      VirtualQuery(&stackInfo, &stackInfo, sizeof(MEMORY_BASIC_INFORMATION));
+  DCHECK_GE(resultSize, sizeof(MEMORY_BASIC_INFORMATION));
+  uint8_t* stackEnd = reinterpret_cast<uint8_t*>(stackInfo.AllocationBase);
+
+  uint8_t* stackStart = reinterpret_cast<uint8_t*>(WTF::getStackStart());
+  RELEASE_ASSERT(stackStart && stackStart > stackEnd);
+  size_t s_threadStackSize = static_cast<size_t>(stackStart - stackEnd);
+  // When the third last page of the reserved stack is accessed as a
+  // guard page, the second last page will be committed (along with removing
+  // the guard bit on the third last) _and_ a stack overflow exception
+  // is raised.
+  //
+  // We have zero interest in running into stack overflow exceptions while
+  // marking objects, so simply consider the last three pages + one above
+  // as off-limits and adjust the reported stack size accordingly.
+  //
+  // http://blogs.msdn.com/b/satyem/archive/2012/08/13/thread-s-stack-memory-management.aspx
+  // explains the details.
+  RELEASE_ASSERT(s_threadStackSize > 4 * 0x1000);
+  s_threadStackSize -= 4 * 0x1000;
+  return s_threadStackSize;
+}
+#endif
+
+}  // namespace internal
+
+}  // namespace WTF