Writing to eventfd from kernel module

Question

I have created an eventfd instance in a userspace program using eventfd(). Is there a way in which I can pass some reference (a pointer to its struct or pid+fd pair) to this created instance of eventfd to a kernel module so that it can update the counter value?

Here is what I want to do: I am developing a userspace program which needs to exchange data and signals with a kernel space module which I have written. For transferring data, I am already using ioctl. But I want the kernel module to be able to signal the userspace program whenever new data is ready for it to consume over ioctl.

To do this, my userspace program will create a few eventfds in various threads. These threads will wait on these eventfds using select() and whenever the kernel module updates the counts on these eventfds, they will go on to consume the data by requesting for it over ioctl.

The problem is, how do I resolve the "struct file *" pointers to these eventfds from kernelspace? What kind of information bout the eventfds can I sent to kernel modules so that it can get the pointers to the eventfds? what functions would I use in the kernel module to get those pointers?

Is there better way to signal events to userspace from kernelspace? I cannot let go of using select().

Answer 1

I finally figured out how to do this. I realized that each open file on a system could be identified by the pid of one of the processes which opened it and the fd corresponding to that file (within that process's context). So if my kernel module knows the pid and fd, it can look up the struct * task_struct of the process and from that the struct * files and finally using the fd, it can acquire the pointer to the eventfd's struct * file. Then, using this last pointer, it can write to the eventfd's counter.

Here are the codes for the userspace program and the kernel module that I wrote up to demonstrate the concept (which now work):

Userspace C code (efd_us.c):

#include <unistd.h>
#include <stdlib.h>
#include <stdio.h>
#include <stdint.h>     //Definition of uint64_t
#include <sys/eventfd.h>

int efd; //Eventfd file descriptor
uint64_t eftd_ctr;

int retval;     //for select()
fd_set rfds;        //for select()

int s;

int main() {


    //Create eventfd
    efd = eventfd(0,0);
    if (efd == -1){
        printf("\nUnable to create eventfd! Exiting...\n");
        exit(EXIT_FAILURE);
    }

    printf("\nefd=%d pid=%d",efd,getpid());

    //Watch efd
    FD_ZERO(&rfds);
    FD_SET(efd, &rfds);

    printf("\nNow waiting on select()...");
    fflush(stdout);

    retval = select(efd+1, &rfds, NULL, NULL, NULL);

    if (retval == -1){
        printf("\nselect() error. Exiting...");
        exit(EXIT_FAILURE);
    } else if (retval > 0) {
        printf("\nselect() says data is available now. Exiting...");
        printf("\nreturned from select(), now executing read()...");
        s = read(efd, &eftd_ctr, sizeof(uint64_t));
        if (s != sizeof(uint64_t)){
            printf("\neventfd read error. Exiting...");
        } else {
            printf("\nReturned from read(), value read = %lld",eftd_ctr);
        }
    } else if (retval == 0) {
        printf("\nselect() says that no data was available");
    }

    printf("\nClosing eventfd. Exiting...");
    close(efd);
    printf("\n");
    exit(EXIT_SUCCESS);
}

Kernel Module C code (efd_lkm.c):

#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/pid.h>
#include <linux/sched.h>
#include <linux/fdtable.h>
#include <linux/rcupdate.h>
#include <linux/eventfd.h>

//Received from userspace. Process ID and eventfd's File descriptor are enough to uniquely identify an eventfd object.
int pid;
int efd;

//Resolved references...
struct task_struct * userspace_task = NULL; //...to userspace program's task struct
struct file * efd_file = NULL;          //...to eventfd's file struct
struct eventfd_ctx * efd_ctx = NULL;        //...and finally to eventfd context

//Increment Counter by 1
static uint64_t plus_one = 1;

int init_module(void) {
    printk(KERN_ALERT "~~~Received from userspace: pid=%d efd=%d\n",pid,efd);

    userspace_task = pid_task(find_vpid(pid), PIDTYPE_PID);
    printk(KERN_ALERT "~~~Resolved pointer to the userspace program's task struct: %p\n",userspace_task);

    printk(KERN_ALERT "~~~Resolved pointer to the userspace program's files struct: %p\n",userspace_task->files);

    rcu_read_lock();
    efd_file = fcheck_files(userspace_task->files, efd);
    rcu_read_unlock();
    printk(KERN_ALERT "~~~Resolved pointer to the userspace program's eventfd's file struct: %p\n",efd_file);


    efd_ctx = eventfd_ctx_fileget(efd_file);
    if (!efd_ctx) {
        printk(KERN_ALERT "~~~eventfd_ctx_fileget() Jhol, Bye.\n");
        return -1;
    }
    printk(KERN_ALERT "~~~Resolved pointer to the userspace program's eventfd's context: %p\n",efd_ctx);

    eventfd_signal(efd_ctx, plus_one);

    printk(KERN_ALERT "~~~Incremented userspace program's eventfd's counter by 1\n");

    eventfd_ctx_put(efd_ctx);

    return 0;
}


void cleanup_module(void) {
    printk(KERN_ALERT "~~~Module Exiting...\n");
}  

MODULE_LICENSE("GPL");
module_param(pid, int, 0);
module_param(efd, int, 0);

To run this, carry out the following steps:

1. Compile the userspace program (efd_us.out) and the kernel module (efd_lkm.ko)
2. Run the userspace program (./efd_us.out) and note the pid and efd values that it print. (for eg. "pid=2803 efd=3". The userspace program will wait endlessly on select()
3. Open a new terminal window and insert the kernel module passing the pid and efd as params: sudo insmod efd_lkm.ko pid=2803 efd=3
4. Switch back to the userspace program window and you will see that the userspace program has broken out of select and exited.

Answer 2

Consult the kernel source here:

http://lxr.free-electrons.com/source/fs/eventfd.c

Basically, send your userspace file descriptor, as produced by eventfd(), to your module via ioctl() or some other path. From the kernel, call eventfd_ctx_fdget() to get an eventfd context, then eventfd_signal() on the resulting context. Don't forget eventfd_ctx_put() when you're done with the context.

Answer 3

how do I resolve the "struct file *" pointers to these eventfds from kernelspace

You must resolve those pointers into data structures that this interface you've created has published (create new types and read the fields you want from struct file into it).

Is there better way to signal events to userspace from kernelspace?

Netlink sockets are another convenient way for the kernel to communicate with userspace. "Better" is in the eye of the beholder.