undefined behaviors

Race Conditions in Signal Handlers

tar,signals,interrupt,handler,process,unix,race-condition

Signals offer a unique, low-level way of communicating with processes. But under certain circumstances, they can kill processes, even when they should work.

This article is a deep dive on a classic race condition issue. If you’re hoping for an elegant and interesting article on how I identified a critical vulnerability in tar, I’m sorry to say - there’s no such vulnerability. It all boils down to a simple race condition issue.

Signals are a special, but very primitive way for processes to communicate functionality. Signals are useful as they are a standardized interface available to 99.99% of programs run on UNIX systems (in existence). Interaction can be done with just the kill command.

While the signals API can be quite bare bones and simple, it’s technically much less complex compared to a network interface, usage of STDIN & STDOUT, a file, or even a shared memory segment.

These other options might have a lot more features, but none of them are perfectly standardized, completely secure, or simple to use.

If you’re looking to allow basic communication with your program for very specific use cases and don’t need complexity or I/O, signals can be a great way to go.

The tar command

This section is a bit of a tangent, but it’s a great example of how signals can be used in practice, as well as how I came across this issue. Skip to the next section if you just want to hear the error & solution.

The tar command is a ubiquitous tool for creating and extracting archives. It’s a very simple tool, but it’s extremely powerful. It’s also a great example of a program that uses signals.

A couple months ago, I was writing software to help bootstrap embedded devices. The software would use tar to extract a filesystem onto the device’s eMMC. Due to the size of the filesystem and the speed of the device, this process could take some time - I wanted to add a progress bar to confirm that the process was still running & progress was being made.

Unfortunately, tar doesn’t emit progress information under normal circumstances, and no alternatives were available in my language of choice that maintained the speed of tar. But looking into the documentation, tar could receive specific, designated signals to emit progress information for both archival and extraction operations.

By starting tar with the --totals flag, it would emit statistics upon completion. But to request information during the operation, a signal must be chosen, like so: tar -x -f archive.tar --totals=SIGUSR1.

$ tar -xf image.tar --totals
Total bytes read: 43048960 (42MiB, 23MiB/s)

Emitting a signal can be done with the kill command, like so: kill -USR1 <pid>. This will send the USR1 signal to the process with the given PID. The USR1 signal is a user-defined signal, and is not used by the system.

And so, my plan was to start a tar process as usual with the --totals flag, and then send the USR1 signal to the process occasionally to query an extraction operation’s progress. In Python, I used the subprocess module to start and manage the process.

import os
import subprocess
import signal
import time
import sys
 
# Define the command to execute
command = ["tar", "-xpf", sys.argv[2], "-C", sys.argv[1], "--totals=SIGUSR1"]
 
# Start the subprocess
print(' '.join(command))
process = subprocess.Popen(command, preexec_fn=os.setsid, stderr=subprocess.PIPE)
 
try:
    while True:
        # Both of these don't work! Why?
        # process.send_signal(signal.SIGUSR1)
        # os.killpg(os.getpgid(process.pid), signal.SIGUSR1)
 
        # Ping the subprocess with SIGUSR1 signal
        subprocess.Popen(["kill", "-SIGUSR1", str(process.pid)])
 
        print(process.stderr.readline().decode("utf-8").strip())
 
        # Wait for a specified interval
        time.sleep(1.9)  # Adjust the interval as needed
 
except KeyboardInterrupt:
    # Handle Ctrl+C to gracefully terminate the script
    process.terminate()
 
# Wait for the subprocess to complete
process.wait()

You’ll notice I have three different ways to send signals shown, but only one of them is working. Moreover, instead of the signal not working like expected, the signal actually kills the process. When checked the exit code, one will find that the status code is the same as the signal number, but negated.

For example, SIGUSR1 exits with -10, SIGUSR2 exits with -12, and SIGHUP exits with -2. In fact, when you look into signals, this is the default behavior for processes exited by signals.

Signal Handlers Aren’t Instant

To my surprise, the handlers that programs like tar provide aren’t available instantly - so much so that even Python can send a signal before they’re registered.

I am still not sure as to how signal handlers are implemented - I would’ve assumed they are static, unchanging, and registered at program start, but that doesn’t seem to be the case - or at least, Python can beat them to the punch.

Whatever the case, the issue with my implementation is that the signal is sent before the handler is registered, and the default behavior of the signal takes over. For many signals (including the one[s] I was using), this is to terminate the process.

How to wait for Signal Handlers

Besides just waiting for a second, there’s a way to wait for signal handlers to be registered. Or rather, there’s a way to check whether signal handlers have been provided or a process.

On Unix systems (which is the only place you’re going to find Unix signals), there’s a special pseudo-filesystem that provides intimate details on a process. This includes things like the process’s name, state, PID, memory usage, threads, and of course: signal handlers.

See below, the contents of /proc/<pid>/status for a process:

/proc/100162/status
 1   │ Name:   Isolated Web Co
 2   │ Umask:  0002
 3   │ State:  S (sleeping)
 4   │ Tgid:   100162
 5   │ Ngid:   0
 6   │ Pid:    100162
 7   │ PPid:   6225
 8   │ TracerPid:  0
 9   │ Uid:    1000    1000    1000    1000
10   │ Gid:    1000    1000    1000    1000
11   │ FDSize: 512
   
...
   
34   │ THP_enabled:    1
35   │ Threads:    27
36   │ SigQ:   0/62382
37   │ SigPnd: 0000000000000000
38   │ ShdPnd: 0000000000000000
39   │ SigBlk: 0000000000000000
40   │ SigIgn: 0000000001011002
41   │ SigCgt: 0000000f40800ef8
42   │ CapInh: 0000000000000000
43   │ CapPrm: 0000000000000000

It’s quite a long file (line numbers added for reference), but it contains a lot of useful information about a given process.

We’re interested in SigCgt, which is a bitmask of signals that are caught by the process. The specific bit depends on the platform, but in Python, this can be found in the signal module:

>>> from signal import SIGUSR1
>>> SIGUSR1
10

We can parse the SigCgt value using the the int function and setting the radix to 16 (hexadecimal).

>>> int("0000000f40800ef8", 16)
65506643704

Checking whether or not the Nth bit is set can be done with the bitwise AND operator (&) and a bitshift (<<).

>>> sigcgt = int("0000000f40800ef8", 16)
>>> mask = 1 << (SIGUSR1 - 1)
>>> sigcgt & mask
512

If the result is non-zero, the bit is set. If the result is zero, the bit is not set.

By simply polling the process’s signal handlers, we can wait for the signal handler to be registered before sending the SIGUSR1 signal.

Credits

Credit to Eryk Sun for explaining the issue and providing an immaculate solution to signal handlers in Python.