Coherent Musings - Implementing mmap for transferring data from user space to kernel space

I was recently working on an application where streams from four multiplexed analog video channels had to be displayed in four windows. I was trying to do this using OpenCV while the analog video decoder/multiplexer in question was the ADV7180. For switching the channels, I was using an ioctl() call to switch the channels in a while loop with a certain time interval, while capturing the frames and putting them in a separate queues as per the channel selected. This was done in the main thread, while separate threads pulled the frames from the queue and rendered them. The capturing and rendering part was being done with OpenCV. I was not able to achieve a decent enough frame rate with this. I had put delays in certain places to avoid frame glitches in the multiple windows displaying the frames.

Thinking that may be the ioctl() call and the context switch is the reason I have to use delays and this is going slow, I decided to look into ways of transferring data faster between the user and kernel space, instead of using an ioctl() call. An mmap() implementation in a driver or a memory mapping between user and kernel space is the fastest way to transfer data. This approach doesn’t incur a context switch nor a memory buffer copying. Below is a sample code showing how a mmap() implementation for a driver is done.

Below is the driver code.

#include "linux/module.h"
#include "linux/kernel.h"
#include "linux/init.h"
#include "linux/fs.h"
#include "linux/debugfs.h"
#include "linux/slab.h"
#include "linux/mm.h"

#ifndef VM_RESERVED
# define  VM_RESERVED   (VM_DONTEXPAND | VM_DONTDUMP)
#endif

struct dentry  *file;

struct mmap_info
{
    char *data;            
    int reference;      
};

void mmap_open(struct vm_area_struct *vma)
{
    struct mmap_info *info = (struct mmap_info *)vma->vm_private_data;
    info->reference++;
}

void mmap_close(struct vm_area_struct *vma)
{
    struct mmap_info *info = (struct mmap_info *)vma->vm_private_data;
    info->reference--;
}

static int mmap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
{
    struct page *page;
    struct mmap_info *info;    
    
    info = (struct mmap_info *)vma->vm_private_data;
    if (!info->data)
    {
        printk("No data\n");
        return 0;    
    }
    
    page = virt_to_page(info->data);    
    
    get_page(page);
    vmf->page = page;            
    
    return 0;
}

struct vm_operations_struct mmap_vm_ops =
{
    .open =     mmap_open,
    .close =    mmap_close,
    .fault =    mmap_fault,    
};

int op_mmap(struct file *filp, struct vm_area_struct *vma)
{
    vma->vm_ops = &mmap_vm_ops;
    vma->vm_flags |= VM_RESERVED;    
    vma->vm_private_data = filp->private_data;
    mmap_open(vma);
    return 0;
}

int mmapfop_close(struct inode *inode, struct file *filp)
{
    struct mmap_info *info = filp->private_data;
    
    free_page((unsigned long)info->data);
    kfree(info);
    filp->private_data = NULL;
    return 0;
}

int mmapfop_open(struct inode *inode, struct file *filp)
{
    struct mmap_info *info = kmalloc(sizeof(struct mmap_info), GFP_KERNEL);    
    info->data = (char *)get_zeroed_page(GFP_KERNEL);
    memcpy(info->data, "hello from kernel this is file: ", 32);
    memcpy(info->data + 32, filp->f_dentry->d_name.name, strlen(filp->f_dentry->d_name.name));
    /* assign this info struct to the file */
    filp->private_data = info;
    return 0;
}

static const struct file_operations mmap_fops = {
    .open = mmapfop_open,
    .release = mmapfop_close,
    .mmap = op_mmap,
};

static int __init mmapexample_module_init(void)
{
    file = debugfs_create_file("mmap_example", 0644, NULL, NULL, &mmap_fops);
    return 0;
}

static void __exit mmapexample_module_exit(void)
{
    debugfs_remove(file);
}

module_init(mmapexample_module_init);
module_exit(mmapexample_module_exit);
MODULE_LICENSE("GPL");

Below is the user space application showing it’s use in an application.

#include <stdio.h>
#include <string.h>
#include <fcntl.h>
#include <sys/mman.h>

#define PAGE_SIZE     4096

int main ( int argc, char **argv )
{
    int configfd;
    char * address = NULL;

    configfd = open("/sys/kernel/debug/mmap_example", O_RDWR);
    if(configfd < 0)
    {
        perror("Open call failed");
        return -1;
    }
    
    address = mmap(NULL, PAGE_SIZE, PROT_READ|PROT_WRITE, MAP_SHARED, configfd, 0);
    if (address == MAP_FAILED)
    {
        perror("mmap operation failed");
        return -1;
    }

    printf("Initial message: %s\n", address);
    memcpy(address + 11 , "*user*", 6);
    printf("Changed message: %s\n", address);
    close(configfd);    
    return 0;
}

The above code should run well for both desktop or embedded Linux. For learning more, refer to Chapter 15 of the Linux Device Drivers book.

Ultimately though, I didn’t do a mmap() but decided to do the switching channel work in the driver itself using kernel timers and workqueues. Not that this improved the frame rate much either, may be a 1 frame per second improvement. May be someone will give me some feedback some day with this task. Till then, hope you find this post useful.