How can a Linux character device driver detect when a program using it exits abnormally?

Question

I have a Linux character device driver that creates a /dev/mything entry, and then a C++/Qt program that opens the device and uses it. If that program exits correctly, with exit(), the device is closed and the driver properly resets itself. But if the program exits abnormally, via segfault or SIGINT or something, the device is not properly closed.

My current workaround is to reload the driver if it gets stuck in the "open" state.

This line in the driver tries to prevent multiple programs using the device simultaneously:

int mything_open( struct inode* inode, struct file* filp ) {
    ...
    if ( port->rings[bufcount].virt_addr ) return -EBUSY;
    ...
}

Then this cleans it up:

int mything_release( struct inode* inode, struct file* filp ) {
    ...
    port->rings[bufcount].virt_addr = NULL;
    ...
}

I think exit() is causing mything_release to be called but SIGINT is not. How can I make the driver more robust to this sort of situation?

EDIT:

Here are the operations I have implemented. Maybe I'm missing something?

static struct file_operations fatpipe_fops = {
    .owner =    THIS_MODULE,
    .open =     mything_open,
    .release =  mything_release,
    .read =     mything_read,
    .write =    mything_write,
    .ioctl =    mything_ioctl
};

Answer 1

There is no need for this test; the problem is not abnormal program termination (which, from your driver's standpoint, is exactly like a normal close on the device) but instead a problem in the state keeping of your device. In other words, if you inserted close(dev_fd) or even exit(0) at the exact point where your program is crashing, you'd have the same problem.

You should figure out what part of your driver's behavior is causing it to remain in a busy state and fix that.

Answer 2

The problem boiled down to this line in mything_release, put in to wait for some memory writes to complete:

if (wait_event_interruptible_timeout(port->inq, false, 10)) return -ERESTARTSYS;

With a normal program exit, this would spin for 10 jiffies and continue along. But with an abnormal exit from SIGINT or something, I think the interruptible timeout got interrupted and it returned -ERESTARTSYS, causing my if to return the same.

The thing that worked for me was to just get rid of the if and just wait:

wait_event_interruptible_timeout(port->inq, false, 10);

This patch from years ago made me believe that returning ERESTARTSYS from a close/_release function is not a good idea: http://us.generation-nt.com/answer/patch-fix-wrong-error-code-interrupted-close-syscalls-help-181191441.html