Question

I need to investigate/test the behavior of some code on Linux under conditions where close might be interrupted by signal handlers (either with or without SA_RESTART). What is the most convenient setup to make the close syscall sleep for a measurable window of time during which I could try to hit the process with a signal? Some ideas:

• Intentionally slow/non-responsive NFS mount
• Custom FUSE driver

But since these are a bit of a pain to setup, I’m wondering if there’s anything more off-the-shelf I could use that could give the desired behavior.

Answer 1

If nobody else has a better idea…

You could implement your own character device driver. Start with the template from Chapter 3 in Linux Device Drivers (3rd edition), and tweak it to do nothing except block for a while on close(). (You can use msleep or msleep_interruptible from Chapter 7 to do the blocking.)

Actually, if nobody else suggests something else, I can probably whip this up pretty quickly by adapting some existing code I have. How soon do you need it?

[edit]

OK, try this…

Makefile:

ifneq ($(KERNELRELEASE),)
        obj-m := closer.o

else
        KERNELDIR ?= /lib/modules/$(shell uname -r)/build
        PWD := $(shell pwd)

default: modules

%:
        $(MAKE) -C $(KERNELDIR) M=$(PWD) "$@"

.PHONY: default
endif

closer.c:

#include <linux/init.h>
#include <linux/module.h>
#include <linux/miscdevice.h>
#include <linux/delay.h>
#include <linux/fs.h>

MODULE_DESCRIPTION("Block-on-close driver");
MODULE_AUTHOR("Nemo <[e-mail]>");
MODULE_LICENSE("GPL");
#define VERSION "20110705"
MODULE_VERSION(VERSION);

#define MY_NAME "closer"

int my_open(struct inode *, struct file *);
int my_release(struct inode *, struct file *);
ssize_t my_read(struct file *, char __user *, size_t, loff_t *);
ssize_t my_write(struct file *, const char __user *, size_t, loff_t *);

static struct file_operations my_fops = {
    .owner = THIS_MODULE,
    .open = my_open,
    .read = my_read,
    .write = my_write,
    .release = my_release,
};

static struct miscdevice my_dev;

int __init
my_init(void)
{
    int err = 0;

    printk(KERN_INFO "%s: loading version %s\n", MY_NAME, VERSION);

    my_dev.minor = MISC_DYNAMIC_MINOR;
    my_dev.name = MY_NAME;
    my_dev.fops = &my_fops;
    err = misc_register(&my_dev);

    if (err)
        printk(KERN_ERR "%s: misc_register failed, error %d\n", MY_NAME, err);

    return err;
}

int
my_open(struct inode *inode, struct file *filp)
{
    return 0;
}

ssize_t
my_read(struct file *file, char __user *p, size_t n, loff_t *off) {
    return 0;
}

ssize_t
my_write(struct file *file, const char __user *p, size_t n, loff_t *off) {
    return n;
}

int
my_release(struct inode *inode, struct file *filp)
{
    int err = 0;
    /* 10 second sleep, interruptible. */
    if (msleep_interruptible(10 * 1000) > 0)
        err = -EINTR;

    return err;
}

void __exit
my_exit(void)
{
    misc_deregister(&my_dev);
    printk(KERN_INFO "%s: unloaded\n", MY_NAME);
}

module_init(my_init);
module_exit(my_exit);

Load the module using “insmod closer.o”. If you have a reasonably modern/complete Linux environment, udev will wake up and generate /dev/closer automatically. If not, you can create the device node yourself:

mknod /dev/closer c `tr : ' ' </sys/class/misc/closer/dev`

(That is, /sys/class/misc/closer/dev indicates the major:minor to use.)

Reads and writes work like /dev/null; i.e., EOF on any read, success on any write.

I have verified that “cat < /dev/closer” blocks in close() for 10 seconds. I have not created a test to catch SIGINT (or whatever) and verify that it actually results in EINTR.

Built against a 2.6.32 kernel. Let me know how it works for you.