Skip to main content

Julia Evans

I think I found a Mac kernel bug?

I’ve been working on Mac support for rbspy, and I accidentally found something that looks like a Mac kernel bug! Basically I managed to write a very short (17 line) C program that reliably causes ps to stop working. Without allocating any memory or anything like that!

This seems to be a kernel bug on High Sierra, but not Sierra – someone tried to reproduce on Sierra and couldn’t. So looks like it’s a new bug.

mysterious freezes on mac

This past week I’ve been building Mac support for rbspy (my Ruby profiler). I put a Mac support branch on github yesterday. Some amazing early adopters tried out the branch and reported that it froze their computer (Activity Monitor stopped working, they couldn’t open new terminals, etc). This was extremely surprising – how could that happen?!? There’s a very detailed report at https://github.com/rbspy/rbspy/issues/70.

I was really mystified by this – how could a program not running as root freeze someone’s computer? Somebody suggested that maybe the program was allocating a lot of memory, but I didn’t think it was. And it turns out that memory allocations weren’t the problem.

A 17-line C program that freezes (parts of) my Mac

@parkr was incredibly helpful and managed to narrow down the problem to right before calling the task_for_pid Mach function. Since task_for_pid seemed to be the issue, I worked on reproducing the issue in a small C program.

Here’s the program that shows the bug. It runs in a loop because it’s a bit race-y and sometimes it needs to try 10 times before it’ll freeze.

#include <mach/mach_init.h>
#include <mach/port.h>
#include <sys/wait.h>
#include <unistd.h>
#include <stdio.h>

int main() {
    for (;;) {
        pid_t pid = fork();
        if (pid == 0) execv("/usr/bin/true", NULL); 
        mach_port_name_t task = MACH_PORT_NULL;
        task_for_pid(mach_task_self(), pid, &task);
        printf("Ran task_for_pid\n");
        int wait_status;
        wait(&wait_status);
    }
}

Here’s the behaviour I see:

  • Run ./freeze-mac. (the example program, above) It hangs.
  • Run ps in another tab. ps hangs, and doesn’t display any output. also trying to Ctrl+C to stop ps doesn’t work.

This happens on the Mac I’ve been developing on (running High Sierra, 10.13.2).

The core issue seems to be – if I exec a new program (like /usr/bin/true), and then immediately run task_for_pid on that program, then the process I ran task_for_pid from will hang and ps will stop working (though I can terminate it with Ctrl+C and then things seem to go back to normal).

If you want to try it out on your Mac, you can run the program yourself! (at your own risk – for me it freezes some things but everything goes back to normal if I press ctrl+c, I don’t know what will happen if you run it :) )

wget https://gist.githubusercontent.com/jvns/16c1ea69352a81658d6d8e9c5a289f2a/raw/ea11fa0a16bfcd4fd019666b790c6c8fe624f9f0/freeze-mac.c
cc -o freeze-mac freeze-mac.c
./freeze-mac
# try running `ps` / starting Activity Monitor, they don't work!

this bug is affecting htop too

Somebody on twitter pointed me to this github issue on htop for mac where people are reporting that htop sometimes sporadically freezes their computer in the same way (terminals don’t start, Activity Monitor doesn’t work). In the comments @grrrrrrrrr says this is a race condition in task_for_pid and they have what looks to be a POC that’s similar to what I have here.

I have a workaround!

This is weird and I don’t really understand the underlying kernel issue. Anyway, I have a workaround for this – don’t run task_for_pid on processes immediately after exec’ing them, instead sleep for a few milliseconds first. So hopefully with this new knowledge I can get rbspy to work on Mac without freezing any users’ computers.

Systems programming is weird and exciting though!

if you have thoughts or questions, here’s the twitter thread for this post

How do you spy on a program running in a container? Profiler week 4: callgrind support, containers, Mac progress!