std::thread creation spuriously fails

[dvyukov]

While running a program I got a spurious std::system_error exception.
The problem is that std::thread does very simple error handling for pthread_create:

llvm-project/libcxx/include/__thread/thread.h

Lines 257 to 258 in 96ae085

	int __ec = _VSTD::__libcpp_thread_create(&__t_, &__thread_proxy_cxx03<_InvokePair>, __pp.get());
	if (__ec == 0)

llvm-project/libcxx/include/__threading_support

Line 374 in 96ae085

return pthread_create(__t, nullptr, __func, __arg);

But on Linux clone system call can spuriously fail with EAGAIN:
torvalds/linux@498052b

I think std::thread should do what Go runtime does to avoid spuroius exceptions:
https://github.com/golang/go/blob/2ebe77a2fda1ee9ff6fd9a3e08933ad1ebaea039/src/runtime/cgo/gcc_libinit.c#L99-L111
Unfortunately it's not a good idea to loop infinitely since EAGAIN can be returned for other reasons (no resources, task limit reached).

[llvmbot]

@llvm/issue-subscribers-bug

[philnik777]

This isn't a bug. The library is free to throw the error.

[llvmbot]

@llvm/issue-subscribers-bug

[EricWF]

Actually, rereading the man page for pthread_create, I don't think there's much we can meaningfully do if we encounter EAGAIN (which I had initially thought). Remarking as enhancement.

[dvyukov]

I think we should retry as much as we can (until it's reasonable to assume the error is due to something else). Failing spuriously at random is not useful. Users will need to wrap std::thread with the same logic.
Imagine all command line programs you have (clang, cat, less, cd, grep, ...) would be just randomly failing from time to time.