[Bug 199557] Hang on sysconf(_SC_OPEN_MAX)

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=199557

--- Comment #5 from ikosarev at accesssoftek.com ---
(In reply to Ed Maste from comment #2)
> Pointed out by kib, when invoking the fork syscall directly the child
> inherits potentially locked libc mutexes, and it's probably the printf
> that hangs, not the sysconf.

Note that the code doesn't leverage libc's printf(), but rather uses its own
version of it.

> What was the original underlying issue here?

The original problem is random Tsan tests hanging on high loaded machines like
the FreeBSD sanitizers buildbot:

http://lab.llvm.org:8011/builders/sanitizer_x86_64-freebsd/builds/4978/steps/make-check-tsan/logs/stdio

It mostly work fine when the tests run in about 4 threads, but tends to fail
constantly when run in about 64 threads--doesn't matter how much cores the
hardware has.

This is how it leads to the point of hang: during execution of a Tsan tests the
Tsan STL catches a data race or other condition triggering a run-time report.
That report contains references to related object files and associated
addresses. These file names together with the addresses are then translated
into source file name, line numbers and function names by calling the
llvm-symbolizer or, if the RTL is unable to locate it, the addr2line command.
To catch the output of the symbolizing tool the RTL fork() the process and
redirects the streams. This is where the whole thing hangs. Note is that if I
comment the sysconf(_SC_OPEN_MAX) call, then it hangs on one of the other
subsequent system calls, like read() so it appears to not to be a
sysconf()-specific issue.

Another note is that the RTL sources refer to a specific reason about why it
tries to use the syscall fork() in place of the libc's intercepted one:

    // Real fork() may call user callbacks registered with pthread_atfork().
    pid = internal_fork();

The internal_fork() function is just the syscall:

int internal_fork() {
#if SANITIZER_USES_CANONICAL_LINUX_SYSCALLS
  return internal_syscall(SYSCALL(clone), SIGCHLD, 0);
#else
  return internal_syscall(SYSCALL(fork));
#endif
}

-- 
You are receiving this mail because:
You are the assignee for the bug.