The Universe of Discourse

Sun, 13 Mar 2022

Why no Unix error device?

Suppose you're writing some program that does file I/O. You'd like to include a unit test to make sure it properly handles the error when the disk fills up and the write can't complete. This is tough to simulate. The test itself obviously can't (or at least shouldn't) actually fill the disk.

A while back some Unix systems introduced a device called /dev/full. Reading from /dev/full returns zero bytes, just like /dev/zero. But all attempts to write to /dev/full fail with ENOSPC, the system error that indices a full disk. You can set up your tests to try to write to /dev/full and make sure they fail gracefully.

That's fun, but why not generalize it? Suppose there was a /dev/error device:

#include <sys/errdev.h>
error = open("/dev/error", O_RDWR);

ioctl(error, ERRDEV_SET, 23);

The device driver would remember the number 23 from this ioctl call, and the next time the process tried to read or write the error descriptor, the request would fail and set errno to 23, whatever that is. Of course you wouldn't hardwire the 23, you'd actually do

#include <sys/errno.h>

ioctl(error, ERRDEV_SET, EBUSY);

and then the next I/O attempt would fail with EBUSY.

Well, that's the way I always imagined it, but now that I think about it a little more, you don't need this to be a device driver. It would be better if instead of an ioctl it was an fcntl that you could do on any file descriptor at all.

Big drawback: the most common I/O errors are probably EACCESS and ENOENT, failures in the open, not in the actual I/O. This idea doesn't address that at all. But maybe some variation would work there. Maybe for those we go back to the original idea, have a /dev/openerror, and after you do ioctl(dev_openerror, ERRDEV_SET, EACCESS), the next call to open fails with EACCESS. That might be useful.

There are some security concerns with the fcntl version of the idea. Suppose I write malicious program that opens some file descriptor, dups it to standard input, does fcntl(1, ERRDEV_SET, ESOMEWEIRDERROR), then execs the target program t. Hapless t tries to read standard input, gets ESOMEWEIRDERROR, and then does something unexpected that it wasn't supposed to do. This particular attack is easily foiled: exec should reset all the file descriptor saved-error states. But there might be something more subtle that I haven't thought of and in OS security there usually is.

Eh, probably the right solution these days is to LD_PRELOAD a complete mock filesystem library that has any hooks you want in it. I don't know what the security implications of LD_PRELOAD are but I have to believe that someone figured them all out by now.

[ Addendum 20220314: Better solutions exist. ]

[Other articles in category /Unix] permanent link