fix: remove TLS destruction (#19365)
Building with dchecks_always_on=true in release configuration seems to introduce flakiness because the TLS is double-freed. Amending the check seems to fix the flakiness.
This commit is contained in:
parent
b7f554f1dc
commit
9a7276dffc
1 changed files with 0 additions and 17 deletions
|
@ -134,23 +134,6 @@ int APIENTRY wWinMain(HINSTANCE instance, HINSTANCE, wchar_t* cmd, int) {
|
|||
if (run_as_node || !IsEnvSet("ELECTRON_NO_ATTACH_CONSOLE"))
|
||||
base::RouteStdioToConsole(false);
|
||||
|
||||
#ifndef DEBUG
|
||||
// Chromium has its own TLS subsystem which supports automatic destruction
|
||||
// of thread-local data, and also depends on memory allocation routines
|
||||
// provided by the CRT. The problem is that the auto-destruction mechanism
|
||||
// uses a hidden feature of the OS loader which calls a callback on thread
|
||||
// exit, but only after all loaded DLLs have been detached. Since the CRT is
|
||||
// also a DLL, it happens that by the time Chromium's `OnThreadExit` function
|
||||
// is called, the heap functions, though still in memory, no longer perform
|
||||
// their duties, and when Chromium calls `free` on its buffer, it triggers
|
||||
// an access violation error.
|
||||
// We work around this problem by invoking Chromium's `OnThreadExit` in time
|
||||
// from within the CRT's atexit facility, ensuring the heap functions are
|
||||
// still active. The second invocation from the OS loader will be a no-op.
|
||||
extern void NTAPI OnThreadExit(PVOID module, DWORD reason, PVOID reserved);
|
||||
atexit([]() { OnThreadExit(nullptr, DLL_THREAD_DETACH, nullptr); });
|
||||
#endif
|
||||
|
||||
std::vector<char*> argv(arguments.argc);
|
||||
std::transform(arguments.argv, arguments.argv + arguments.argc, argv.begin(),
|
||||
[](auto& a) { return _strdup(base::WideToUTF8(a).c_str()); });
|
||||
|
|
Loading…
Reference in a new issue