(Edit: Please make sure that you check out this link and also this one and this. The gist of these discussions is that letting Rust automatically close a file descriptor without checking the return value of the close system call is not a good idea from the perspective of building reliable systems - I had not thought about it while writing this post).

This post is inspired by the excellent article Rust means never having to close a socket. So now you know where the title comes from!

Garbage Collection is one of the big ideas in the field of programming languages. Languages which employ some form of GC (Java, Python, Ruby, almost all modern languages) free the programmer from the tedious and error prone task of manually allocating and deallocating memory (the kind of stuff you do using malloc/free in C). But memory is not the only resource which a program has to manage - open file descriptors, network sockets, temporary files etc are also resources which have to be managed properly to prevent leaks. Unfortunately, GC is not a solution to this problem; it is easy to leak resources in a GC’d language. We will see that the compile time strategies adopted by Rust provide a more elegant solution to the general problem of acquiring and releasing all kinds of resources.

## A Python program which leaks file descriptors


# save this as leak.py

def use_file():
f = open('data.txt')

while True:
use_file()



The open function in Python returns a high-level file object (or a so-called file handle) - this object will have embedded in it a file descriptor. A file descriptor is simply an integer value which system calls like read, write etc use to access the file. The operating system (Linux, in this case) places a limit on the number of open file descriptors which a program can have at any given point in time (this limit can be seen and modified using the ulimit command - in my case, running ulimit shows me that my program can only have a maximum of 1024 open file descriptors).

In the case of the above program, each invocation of the open function in Python will result in a new file object (and an associated new file descriptor) getting created. If you call the close method of a file object:

f.close()


the associated file descriptor will get closed and the operating system can re-use this file descriptor the next time.

But we are NOT calling f.close() and our program is still working properly! This seems to be strange!

The answer lies in something which is purely an implementation detail of the CPython (the most commonly used implementation of Python) virtual machine.
CPython performs garbage collection using a strategy called reference counting. The idea behind refcounting is simple - all Python objects have a count associated with them which gets incremented when a new reference to the object is created and gets decremented when that reference is gone.


a = [1,2,3] # refcount = 1, only one reference "a"

b = a  # refcount = 2, both "a" and "b" point to the same list

c = a # refcount = 3

c = 0 # refcount is now 2

b = 0 # refcount is 1

a = 0 # refcount is 0, the list [1,2,3] now gets deallocated



In the case of our program which keeps on opening a file in an infinite loop, the only reference to the high-level file object, the variable f, disappears when the function use_file terminates - this will result in the memory associated with the file object getting deallocated. CPython will also close the associated file descriptor at the same time.

The key idea here is that this is an implementation detail of the CPython virtual machine. The Python language doesn’t guarantee that the file descriptor will get closed deterministically unless you call f.close().

The correct implementation of use_file should look like this:

def use_file():
f = open('data.txt')
# do something with "f"
f.close()


Or, even better:

def use_file():
with open('data.txt') as f:
#do something with f
pass


Python guarantees that f.close() is always called when the with block terminates.

### Experimenting with PyPy

PyPy is an alternative implementation of Python focused on speed. Let us run leak.py using PyPy:

pypy leak.py



The program will immediately crash; here is the error message I got:

Traceback (most recent call last):
File "leak.py", line 6, in <module>
File "leak.py", line 2, in use_file
IOError: [Errno 24] Too many open files: 'data.txt'


Why does the code work with CPython and crash under PyPy?

PyPy does not use reference counting - it uses some other garbage collection strategies which are not guaranteed to free up an object as soon as no references are pointing to it. The program basically uses up all available descriptors before the garbage collector steps in.

## Rewriting the code in Rust

Here is a program similar to the Python code in the previous section:

use std::fs::File;

fn use_file() {
let f = File::open("data.txt").unwrap();
}

fn main() {
loop {
use_file();
}
}


The file handle f is a Rust structure and it holds a file descriptor. The ownership rules of Rust guarantees that the structure is dropped at the point where the variable f goes out of scope. A destructor function gets called automatically and this function takes care of closing the file descriptor associated with the file handle.

There is no need to explicitly close a file in Rust! Or close a network socket, or clean up a temporary file! When a variable goes out of scope, all the resources associated with it are cleaned up.

The clean up happens deterministically. The code required to perform the clean up is inserted by the compiler into the executable precisely at the point where the variable goes out of scope - and this is completely determined at compile time.

### Using strace to watch the file getting closed

Let’s rewrite the code to eliminate the “loop”:

use std::fs::File;

fn use_file() {
let f = File::open("data.txt").unwrap();
}

fn main() {
use_file();
println!("good bye!");
}


Compile and run the executable using the strace command:

rustc leak.rs
strace ./leak



Strace will display the system calls invoked by our program. Here is a part of the output I got on my machine:

mmap(NULL, 8192, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0) = 0x7f3a46722000
sigaltstack({ss_sp=0x7f3a46722000, ss_flags=0, ss_size=8192}, NULL) = 0
open("data.txt", O_RDONLY|O_CLOEXEC)    = 3
ioctl(3, FIOCLEX)                       = 0
close(3)                                = 0
write(1, "good bye!\n", 10good bye!
)             = 10
sigaltstack({ss_sp=NULL, ss_flags=SS_DISABLE, ss_size=8192}, NULL) = 0



You can see the open and close system calls getting invoked!

### Moving a file handle

use std::fs::File;

fn use_file(f: File) {
// do something with f
}

fn main() {
let f = File::open("data.txt").unwrap();
use_file(f);
println!("good bye!");
}


Rust move semantics guarantees that f is no longer usable in main - the ownership of the file handle has effectively been transferred to f in the use_file function. At the point where this function ends, f goes out of scope and the associated file descriptor is closed.

Running strace on the executable generated by the above program shows that this is indeed the case:

sigaltstack({ss_sp=0x7f83c88e5000, ss_flags=0, ss_size=8192}, NULL) = 0
open("data.txt", O_RDONLY|O_CLOEXEC)    = 3
ioctl(3, FIOCLEX)                       = 0
close(3)                                = 0
write(1, "good bye!\n", 10good bye!
)             = 10
sigaltstack({ss_sp=NULL, ss_flags=SS_DISABLE, ss_size=8192}, NULL) = 0



### Cloning a file handle

What if you wish to share a file handle between two functions in such a way that both handles refer to exactly the same offset in exactly the same file?

A solution is to have one function borrow the file handle from the other one.

Here is another way to do this:

use std::fs::File;

fn use_file(f: File) {
// do something with f
}

fn main() {
let f = File::open("data.txt").unwrap();
use_file(f.try_clone().unwrap());
println!("good bye!");
}


The try_clone function creates a clone of the original file handle. So we now have two independent file handles: one in the main function and the other one in use_file.

Both the file handles will refer to exactly the same location in the file data.txt. If you read 5 bytes in use_file and then try to read 3 bytes in main, you are sure to get the next 3 bytes from the file.

This means the file descriptors embedded in both file handles should be the same.

But there is a problem. When use_file terminates, f gets dropped and the file descriptor is closed.

That means you will not be able to read from the file in main.

Also, when main terminates, the f in main gets dropped; this will result in an attempt to close an already closed descriptor.

How does Rust solve the problem? We have strace to our rescue. Let’s run the executable under strace and examine the output:

open("data.txt", O_RDONLY|O_CLOEXEC)    = 3
ioctl(3, FIOCLEX)                       = 0
fcntl(3, F_DUPFD_CLOEXEC, 0)            = 4
ioctl(4, FIOCLEX)                       = 0
close(4)                                = 0
write(1, "good bye!\n", 10good bye!
)             = 10
close(3)                                = 0
sigaltstack({ss_sp=NULL, ss_flags=SS_DISABLE, ss_size=8192}, NULL) = 0


This line is the key:

fcntl(3, F_DUPFD_CLOEXEC, 0)            = 4


The file descriptor 3 gets duplicated and we now have another file descriptor, 4, which refers to exactly the same offset of the file referred to by descriptor 3.

File descriptors 3 and 4 are two independent descriptors which share a single offset to the same file - both can be used independently and also closed independently. Closing 4 doesn’t affect 3. You can still access the file using 3 and close it once you are finished.

That is exactly what happens here. The cloned file handle has embedded in it a duped file descriptor!