Writing a NetBSD kernel module

Kernel modules are object files used to extend an operating system’s kernel functionality at run time.

In this post, we’ll look at implementing a simple character device driver as a kernel module in NetBSD. Once it is loaded, userspace processes will be able to write an arbitrary byte string to the device, and on every successive read expect a cryptographically-secure pseudorandom permutation of the original byte string.

Before we begin, compiling a kernel module requires the NetBSD source code to live in /usr/src. This explains how to get that.

Usually, most userspace interfaces to character or block devices are through special files that live in /dev. We’ll create one such special file through the command

$ mknod /dev/rperm c 420 0

The c indicates that this file is an interface to a character device, 420 indicates this device’s major number, and 0 indicates this device’s minor number. The major number is used by the kernel to uniquely identify each device, and the minor number is usually used internally by device drivers but we won’t be bothering with it.

Our device driver will specifically implement the open, read, write, and close I/O methods. To register our implementations of those methods with the kernel, we first prototype them in way that makes the compiler happy using the dev_type_* set of macros, and then put them into a struct cdevsw.


static struct cdevsw rperm_cdevsw = {
    .d_open = rperm_open,
    .d_close = rperm_close,
    .d_read = rperm_read,
    .d_write = rperm_write,

    .d_ioctl = noioctl,
    .d_stop = nostop,
    .d_tty = notty,
    .d_poll = nopoll,
    .d_mmap = nommap,
    .d_kqfilter = nokqfilter,
    .d_discard = nodiscard,
    .d_flag = D_OTHER

As we can see, there are plenty of functions we won’t be implementing. devsw stands for device switch.

Every kernel module is required to define it’s metadata through the C macro MODULE(class, name, required). Since our module is a device driver, named rperm, and won’t require another module being pre-loaded, we write


Every module is also required to implement a MODNAME_modcmd function, which the kernel calls to report important module-related events, like when the module loads or unloads. This is where we’ll register our struct cdevsw.

#define CMAJOR 420

static int
rperm_modcmd(modcmd_t cmd, void *args)
    devmajor_t bmajor, cmajor;

    bmajor = -1;
    cmajor = CMAJOR;
    switch(cmd) {
        case MODULE_CMD_INIT:
            devsw_attach("rperm", NULL, &bmajor, &rperm_cdevsw, &cmajor);
        case MODULE_CMD_FINI:
            devsw_detach(NULL, &rperm_cdevsw);
    return 0;

The NULL argument to the devsw_* functions is for a block device switch structure, which we aren’t bothered with. Similarly for bmajor, but the kernel ends up assigning an unused block device number for our driver anyway.

Now we turn to actually implementing the four device I/O methods.

On every write, we need to store the byte string somewhere. We use a static structure for that.

static struct rperm_softc {
    char *buf;
    int buf_len;
} sc;

sc.buf will end up pointing to a location in the kernel’s heap that contains the byte string. sc and softc stand for software context, which is just a convention followed in the NetBSD kernel for naming static structures in device driver code.

open is a required implementation, as it is always the first syscall in Unix I/O. But, there is nothing meaningful for us to do there, so we simply write a stub.

rperm_open(dev_t self, int flag, int mod, struct lwp *l)
    return 0;

In write, we allocate enough memory in the kernel’s heap to store the byte string, and then transfer the byte string from userspace to kernelspace.

rperm_write(dev_t self, struct uio *uio, int flags)
    if (sc.buf)
	kmem_free(sc.buf, sc.buf_len);
    sc.buf_len = uio->uio_iov->iov_len;
    sc.buf = (char *)kmem_alloc(sc.buf_len, KM_SLEEP);
    uiomove(sc.buf, sc.buf_len, uio);
    return 0;

First, let’s discuss the allocations.

kmem_alloc is similar to userspace malloc, in that it allocates some number of bytes of memory in the heap. Interestingly, this memory is wired, which means that during physical memory pressure, it is not paged out to a swap disk like userspace memory is. The KM_SLEEP flag to kmem_alloc tells the kernel that the current kernel thread should sleep until enough physical memory is avaiable for the request, if it already isn’t, as opposed to kmem_alloc simply returning NULL in such a situation. Hence, our allocation request never fails, and we don’t have to test for sc.buf == NULL.

kmem_free is similar to userspace free, except for a second argument that has to be the number of bytes allocated using kmem_alloc.

Next, we come to the transfer of the byte string from userspace to kernelspace. Generally, memory to be transfered, in either direction, comes in one or more non-contiguous chunks of memory (think scatter-gather I/O) along with some additional state variables like the amount of data remaining to be transfered in the current session, an offset into a block device, and some flags. All that information is encapulated in a struct uio data type. And uiomove performs the actual transfer by using that information. For example, here uio->uio_rw is set to UIO_WRITE, telling uiomove that data from uio should be transfered to sc.buf. uiomove also ends up updating uio->uio_resid, which is the total number of bytes left to transfer to uio.

Next we come to read.

rperm_read(dev_t self, struct uio *uio, int flags)
    if (sc.buf == NULL || uio->uio_resid < sc.buf_len)
	return EINVAL;

    char c;
    uint32_t i, n, r;

    for (i = 0; i < sc.buf_len-1; i++) {
	r = rand_n(i, sc.buf_len);
	c = sc.buf[r];
	sc.buf[r] = sc.buf[i];
	sc.buf[i] = c;
    uiomove(sc.buf, sc.buf_len, uio);
    return 0;

We first check if there is enough space in uio to transfer a permuted byte string, then use the Fisher–Yates shuffle to permute the original byte string using a random number generated by rand_n, and then copy the permuted string back to userspace. In this case, uio->uio_rw would be set to UIO_READ, telling uiomove that data from sc.buf should be transfered to uio.

The rand_n function, which we need to implement, returns a random integer n uniformly distributed over the range [low, high).

#define R32MAX 4294967295

uint32_t rand_n(uint32_t low, uint32_t high) {
    uint32_t limit, diff, r;

    diff = high - low;
    limit = diff * (R32MAX/diff);
    do {
	r = cprng_strong32();
    } while (r > limit);
    return (r % diff) + low;

For a source of randomness, we use cprng_strong32. The cprng_* family of functions supply cryptographically secure pseudorandom bytes (in this case, 4) to callers within the kernel.

Once we have it, we transform the range of our random number from [0, 2^32) to [low, high) by an iterative test that discards those values of r that are larger than the largest multiple of r less than 2^32, as using those numbers would result in numbers in a certain subrange of [low, high) being more likely to occur than those not.

In close, we free sc.buf if it was allocated before.

rperm_close(dev_t self, int flag, int mod, struct lwp *l)
    if (sc.buf != NULL) {
	kmem_free(sc.buf, sc.buf_len);
	sc.buf = NULL;
    return 0;

Lastly, we write a three line Makefile to build our module.

KMOD=   rperm
SRCS=   rperm.c

.include <bsd.kmodule.mk>

Now we compile and load the module.

$ make
$ modload ./rperm.kmod

The ./ has to be present for modload.

Unfortunately, we can’t simply do

$ echo 'bloop' > /dev/rperm
$ cat /dev/rperm

as that would end up copying the \0 along with the rest of the string into the driver, and the driver would end up shuffling the \0 as well, which we don’t want.

So we settle for a simple test program.

#define BUF_LEN  80
#define N_ITER   20

int main() {
    char buf[BUF_LEN] = "Hello NetBSD!";
    int i, fd, str_len;

    str_len = strlen(buf);
    fd = open("/dev/rperm", O_RDWR);
    write(fd, buf, str_len);
    for (i = 0; i < N_ITER; i++) {
	read(fd, buf, BUF_LEN);
	printf("%s\n", buf);
    return 0;
$ ./test
NDe eBlS!tHol
Hl!elDSNeo Bt
N!llD eSoHetB
tHelS!BlNeo D

All of the above code is available in it’s entirety on github. More examples of kernel modules can be found in the NetBSD source tree at src/sys/modules/examples/.

posted on Feb 2, 2017 by Saurav Sachidanand