The Devil's in the Details

by Georg V. Zezschwitz

Starting from the clean code environment of the two previous articles, we now turn to all the nasty interrupt stuff. Astonishingly, Linux hides most of this from us, so we do not need a single line of assembler...

Reading and writing

Right now, our magic skel-machine driver can load and even unload (painlessly, unlike in DOS), but it has neither read nor written a single character. So we will start fleshing out the skel_read() and skel_write() functions introduced in the previous article (under fops and filp). Both functions take four arguments:

Static int skel_read (struct inode *inode,
                      struct file *file,
                      char *buf, int count)
Static int skel_write (struct inode *inode,
                       struct file *file,
                       const char *buf,
                       int count)

The inode structure supplies the functions with information used already during the skel_open() call. For example, we determined from inode->i_rdev which board the user wants to open, and transferred this data—along with the board's base address and interrupt to the private_data entry of the file descriptor. We might ignore this information now, but if we did not use this hack, inode is our only chance to find out to which board we are talking.

The file structure contains data that is more valuable. You can explore all the elements in its definition in <linux/fs.h>. If you use the private_data entry, you find it here, and you should also make use of the f_flags entry—revealing to you, for instance, if the user wants blocking or non-blocking mode. (We explain this topic in more detail later on.)

The buf argument tells us where to put the bytes read (or where to find the bytes written) and count specifies how many bytes there are. But you must remember that every process has its own private address space. In kernel code, there is an address space common to all processes. When system calls execute on behalf of a specific process, they run in kernel address space, but are still able to access the user space. Historically, this was done through assembler code using the fs register; current Linux kernels hide the specific code within functions called get_user_byte() for reading a byte from user address space, put_user_byte() for writing one, and so on. They were formerly known as get_fs_byte, and only memcpy_tofs() and memcpy_fromfs() reveal these old days even on a DEC Alpha. If you want to explore, look in <asm/segment.h>.

Let us imagine ideal hardware that is always hungry to receive data, reads and writes quickly, and is accessed through a simple 8-bit data-port at the base address of our interface. Although this example is unrealistic, if you are impatient you might try the following code:

Static int skel_read (struct inode *inode,
                      struct file *file,
                      char *buf, int count) {
    int n = count;
    char *port = PORT0 ((struct Skel_Hw*)
    while (n--) {
        Wait till device is ready
        put_user_byte (inb_p (port), buf);
    return count;

Notice the inb_p() function call, which is the actual I/O read from the hardware. You might decide to use its fast equivalent, inb(), which omits a minimal delay some slow hardware might need, but I prefer the safe way.

The equivalent skel_write() function is nearly the same. Just replace the put_user_byte() line by the following:

        outb_p (get_user_byte (buf), port);

However, these lines have a lot of disadvantages. What using them causes Linux to loop infinitely while waiting for a device that never becomes ready? Our driver should dedicate the time in the waiting loop to other processes, making use of all the resources in our expensive CPU, and it should have an input and output buffer for bytes arriving while we are not in skel_read() and corresponding skel_write() calls. It should also contain a time-out test in case of errors, and it should support blocking and non-blocking modes.

Blocking and Non-Blocking Modes

Imagine a process that reads 256 bytes at a time. Unfortunately, our input buffer is empty when skel_read() is called. So what should it do—return and say that there is no data yet, or wait until at least some bytes have arrived?

The answer is both. Blocking mode means the user wants the driver to wait till some bytes are read. Non-blocking mode means to return as soon as possible—just read all the bytes that are available. Similar rules apply to writing: Blocking mode means “Don't return till you can accept some data,” while non-blocking mode means: “Return even if nothing is accepted.” The read() and write()calls usually return the number of data bytes successfully read or written. If, however, the device is non-blocking and no bytes can be transferred, -EAGAIN is typically returned (meaning: “Play it again, Sam”). occasionally, old code may return -EWOULDBLOCK, which is the same as -EAGAIN under Linux.

Maybe now you are smiling as happily as I did when I first heard about these two modes. If these concepts are new for you, you might find the following hints helpful. Every device is opened by default in blocking mode, but you may choose non-blocking mode by setting the O_NONBLOCK flag in the open() call. You can even change the behaviour of your files later on with the fcntl() call. The fcntl() call is an easy one, and the man page will be sufficient for any programmer.

Sleeping Beauty

Once upon a time, a beautiful princess was sent by a witch into a long, deep sleep, lasting for a hundred years. The world nearly forgot her and her castle, twined about by roses, until one day, a handsome prince came, kissed her, and awakened her --and all the other nice things happened that you hear about in fairy tales.

Our driver should do what the princess did while it is waiting for data: sleep, leaving the world spinning around. Linux provides a mechanism for that, called interruptible_sleep_on(). Every process reaching this call will fall asleep and contribute its time slices to the rest of the world. It will stay in this function till another process calls wake_up_interruptible(), and this “prince” usually takes the form of an interrupt handler that has successfully received or sent data, or Linux itself, if a time-out condition has occurred.

Installing an Interrupt Handler

The previous article in this series showed a minimal interrupt handler, which was called skel_trial_fn(), but its workings were not explained. Here, we introduce a “complete” interrupt handler, which will handle both input to and output from the actual hardware device. Figure 1 shows a simple version of its concept: When the driver is waiting for the device to get ready (blocking), it goes to sleep by calling interruptible_sleep_on(). A valid interrupt ends this sleep, restarting skel_write().

Figure 1 does not include the double-nested loop structure we need when working with an internal output buffer. The reason is that if we can perform only writing within the skel_write() function there is no need for an internal output buffer. But our driver should catch data even while not in skel_read() and should write the data in the background even when not in skel_write(). Therefore, we will change the hardware writing in skel_write() to write to an output buffer and let the interrupt handler perform the real writing to the hardware. The interrupt and skel_write() will now be linked by the “Sleeping Beauty” mechanism and the output buffer.

The interrupt handler is installed and uninstalled during the open() and close() calls to the device, as suggested in the previous article. This task is handled by the following kernel calls:

#include <linux/sched.h>
int request_irq(unsigned int irq,
                void (*handler)
                     (int, struct pt_regs *),
                unsigned long flags,
                const char *device);
void free_irq(unsigned int irq);

The handler argument is the actual interrupt handler we wish to install. The role of the flags argument is to set a few features of the handler, the most important being its behaviour as a fast handler (SA_INTERRUPT is set in flags) or as a slow handler ( SA_INTERRUPT is not set). A fast handler is run with all interrupts disabled, while a slow one is executed with all interrupts except itself enabled.

Finally, the device argument is used to identify the handler when looking at /proc/interrupts.

The handler function installed by request_irq() is passed only the interrupt number and the (often useless) contents of the processor registers.

Therefore, we'll first determine which board the calling interrupt belongs to. If we can't find any boards, a situation called a spurious interrupt has occurred, and we should ignore it. Typically interrupts are used to tell whether the device is ready either for reading or writing, so we have to find out by one or more hardware tests what the device wants us to do.

Of course, we should leave our interrupt handler quickly. Strangely enough, printk() (and thus the PDEBUG line) is allowed even within fast interrupt handlers. This is a very useful feature of the linux implementation. If you look at kernel/printk.c you'll discover that its implementation is based on wait queues, as the actual delivery of messages to log files is handled by an external process (usually klogd).

As shown in figure 2, Linux can handle a timeout when in interruptible_sleep_on(). For example, if you have are using a device to which you send an answer, and it is expected to reply within a limited time, causing a time-out to signal an I/O error (-EIO) in the return value to the user process might be a good choice.

Certainly the user process could care for this, too, using the alarm mechanism. But it is definitely easier to handle this in the driver itself. The timeout criteria is specified by SKEL_TIMEOUT, which is counted in jiffies. Jiffies are the steady heartbeat of a Linux system, a steady timer incremented every few milliseconds. The frequency, or number of jiffies per second, is defined by HZ in <asm/param.h> (included in <linux/sched.h>) and varies on different architectures (100 Hz Intel, 1 kHz Alpha). You simply have to set

#define SKEL_TIMEOUT timeout_seconds * HZ
/* ... */
current->timeout = jiffies + SKEL_TIMEOUT

and if interruptible_sleep_on timed out, current->timeout will be cleared after return.

Be aware that interrupts might happen within skel_read() and skel_write(). Variables that might be changed within the interrupt should be declared as volatile. They also need to be protected to avoid race conditions. The classic code sequence to protect a critical region is the following:

unsigned long flags;
save_flags (flags);
cli ();
critical region
restore_flags (flags);

Finally, the code for the “complete” error handler:

#define SKEL_IBUFSIZ 512
#define SKEL_OBUFSIZ 512
/* for 5 seconds timeout */
#define SKEL_TIMEOUT (5*HZ)
/* This should be inserted in the Skel_Hw-structure */
typedef struct Skel_Hw {
    /* write position in input-buffer */
    volatile int ibuf_wpos;
    /* read position in input-buffer */
    int ibuf_rpos;
    /* the input-buffer itself */
    char *ibuf;
    /* write position in output-buffer */
    int obuf_wpos;
    /* read position in output-buffer */
    volatile int buf_rpos;
    char *obuf;
    struct wait_queue *skel_wait_iq;
    struct wait_queue *skel_wait_oq;
#define SKEL_IBUF_EMPTY(b) \
#define SKEL_OBUF_EMPTY(b) \
#define SKEL_IBUF_FULL(b) \
#define SKEL_OBUF_FULL(b) \
Static int skel_open (struct inode *inode,
                      struct file *filp) {
    /* .... */
    /* First we allocate the buffers */
    board->ibuf = (char*) kmalloc (SKEL_IBUFSIZ,
    if (board->ibuf == NULL)
        return -ENOMEM;
    board->obuf = (char*) kmalloc (SKEL_OBUFSIZ,
    if (board->obuf == NULL) {
        kfree_s (board->ibuf, SKEL_IBUFSIZ);
        return -ENOMEM;
    /* Now we clear them */
    ibuf_wpos = ibuf_rpos = 0;
    obuf_wpos = obuf_rpos = 0;
    board->irq = board->hwirq;
    if ((err=request_irq(board->irq>
                         SA_INTERRUPT, "skel")))
        return err;
Static void skel_interrupt(int irq,
                    struct pt_regs *unused) {
    int i;
    Skel_Hw *board;
    for (i=0, board=skel_hw; i<skel_boards;
         board++, i++)
         /* spurious */
        if (board->irq==irq) break;
    if (i==skel_boards) return;
    if (board_is_ready_for_input)
        skel_hw_write (board);
    if (board_is_ready_for_output)
        skel_hw_read (board);
Static inline void skel_hw_write (Skel_Hw *board){
    int rpos;
    char c;
    while (! SKEL_OBUF_EMPTY (board) &&
        board_ready_for_writing) {
        c = board->obuf [board->obuf_rpos++];
        board->obuf_rpos %= SKEL_OBUF_SIZ;
    /* Sleeping Beauty */
    wake_up_interruptible (board->skel_wait_oq);
Static inline void skel_hw_read (Skel_Hw *board) {
    char c;
    /* If space left in the input buffer & device ready: */
    while (! SKEL_IBUF_FULL (board) &&
        board_ready_for_reading) {
        board->ibuf [board->ibuf_wpos++] = c;
        board->ibuf_wpos %= SKEL_IBUFSIZ;
    wake_up_interruptible (board->skel_wait_iq);
Static int skel_write (struct inode *inode,
                       struct file *file,
                       char *buf, int count) {
    int n;
    int written=0;
    Skel_Hw board =
        (Skel_Hw*) (file->private_data);
    for (;;) {
        while (written<count &&
               ! SKEL_OBUF_FULL (board)) {
            board->obuf [board->obuf_wpos] =
                get_user_byte (buf);
            buf++; board->obuf_wpos++;
            board->obuf_wpos %= SKEL_OBUFSIZ;
        if (written) return written;
        if (file->f_flags & O_NONBLOCK)
            return -EAGAIN;
        current->timeout = jiffies + SKEL_TIMEOUT;
        interruptible_sleep_on (
        /* Why did we return? */
        if (current->signal & ~current->blocked)
        /* If the signal is not not being
           blocked */
            return -ERESTARTSYS;
        if (!current->timeout)
            /* no write till timout: i/o-error */
            return -EIO;
Static int skel_read (struct inode *inode,
                      struct file *file,
                      char *buf, int count) {
    Skel_Hw board =
        (Skel_Hw*) (file->private_data);
    int bytes_read = 0;
    if (!count) return 0;
    if (SKEL_IBUF_EMPTY (board)) {
        if (file->f_flags & O_NONBLOCK)
            /* Non-blocking */
            return -EAGAIN;
        current->time_out = jiffies+SKEL_TIMEOUT;
        for (;;) {
            interruptible_sleep_on (
            if (current->signal
                & ~current->blocked)
                return -ERESTARTSYS;
            if (! SKEL_IBUF_EMPTY (board))
            if (!current->timeout)
                /* Got timeout: return -EIO */
                return -EIO;
    /* if some bytes are here, return them */
    while (! SKEL_IBUF_EMPTY (board)) {
        put_user_byte (board->ibuf
        buf++; board->ibuf_rpos++;
        board->ibuf_rpos %= SKEL_IBUFSIZ;
        if (--count == 0) break;
    if (count) /* still looking for some bytes */
    return bytes_read;
Handling select()

The last important I/O function to be shown is select(), one of the most interesting parts of Unix, in our opinion.

The select() call is used to wait for a device to become ready, and is one of the most scary functions for the novice C programmer. While its use from within an application is not shown here, the driver-specific part of the system call is shown, and its most impressive feature is its compactness.

Here's the full code:

Static int skel_select(struct inode *inode,
                       struct file *file,
                       int sel_type,
                       select_table *wait) {
    Skel_Clientdata *data=filp->private_data;
    Skel_Board *board=data->board;
    if (sel_type==SEL_IN) {
        if (! SKEL_IBUF_EMPTY (board))
            /* readable */
            return 1;
        select_wait(&(hwp->skel_wait_iq), wait);
        /* not readable */
        return 0;
    if (sel_type==SEL_OUT) {
        if (! SKEL_OBUF_FULL (board))
            return 1;  /* writable */
        /* hw knows */
        select_wait (&(hwp->skel_wait_oq), wait);
        return 0;
    /* exception condition: cannot happen */
    return 0;

As you can see, the kernel takes care of the hassle of managing wait queues, and you have only to check for readiness.

When we first wrote a select() call for a driver, we didn't understand the wait_queue implementation, and you don't need to either. You only have to know that the code works. wait_queues are challenging, and usually when you write a driver you have no time to accept the challenge.

Actually, select is better understood in its relationships with read and write: if select() says that the file is readable, the next read must not block (independently of O_NONBLOCK), and this means you have to tell the hardware to return data. The interrupt will collect data, and awaken the queue. If the user is selecting for writing, the situation is similar: the driver must tell if write() will block or not. If the buffer is full it will block, but you don't need to tell the hardware about it, since write() has already told it (when it filled the buffer). If the buffer is not full, the write won't block, so you return 1.

This way to think of selecting for write may appear strange, as there are times when you need to write synchronously, and you may expect that a device is writable when it has already accepted pending input. Unfortunately, this way of doing things will break the blocking/nonblocking machinery, and thus an extra call is provided: if you need to write synchronously, the driver must offer (within its fops) the fsync()call. The application invokes fops->fsync through the fsync() system call, and if the driver doesn't support it, -EINVAL is returned.

ioctl()--Passing Control Information

Imagine that you want to change the baud-rate of a serial multiport card you have built. Or tell your frame grabber to change the resolution of an image. Or whatever else... You could wrap these instructions into a series of escape sequences, such as, for example, the screen positioning in ANSI emulation. But, the normal method for this is to make an ioctl() call.

ioctl() calls as defined in <sys/ioctl.h> have the form

ioctl (int file_handle, int command, ...)

where ... is considered to be one argument of the type char * (according to the ioctl man page). Strange as it may be, the kernel receives these arguments in fs/ioctl.c in the form:

int sys_ioctl (unsigned int fd, unsigned int cmd,
               unsigned long arg);

To add to the confusion, <linux/ioctl.h> gives detailed rules how the commands in the second parameter should be built, but nobody in all the drivers is actually following these ideas yet.

In any case, rather than cleaning up the Linux source tree, let's concentrate on the general idea of ioctl() calls. As the user, you pass the file handle and a command in the first two arguments and pass as the third parameter a pointer to a data structure the driver should read and/or write.

A few commands are interpreted by the kernel itself—for example, FIONBIO, which changes the blocking/non-blocking flag of the file. The rest is passed to our own, driver-specific ioctl() call, and arrives in the form:

int skel_ioctl (struct inode *inode,
                struct file *file,
                unsigned int cmd,
                unsigned long arg)

Before we show a small example of a skel_ioctl() implementation, the commands you define should obey the following rules:

  1. Pick up a free MAGIC number from /usr/src/linux/MAGIC and make this number the upper eight bits of the 16-bit command word.

  2. Enumerate commands in the lower eight bits.

Why this? Imagine “Silly Billy” starts his favorite terminal program minicom to connect to his mailbox. “Silly Billy” accidentally changed the serial line minicom uses from /dev/ttyS0 to /dev/skel0 (he is quite silly). The next thing minicom does is initialize the “serial line” with an ioctl() using TCGETA as command. Unfortunately, your device driver, hidden behind /dev/skel0, uses that number to control the voltage for a long-term experiment in the lab...

If the upper eight bits in the commands for ioctl() differ from driver to driver, every ioctl() to an inappropriate device will result in an -EINVAL return, protecting us from extremely unexpected results.

Now, to finish this section, we will implement an ioctl() call reading or changing the timeout delay in our driver. If you want to use it, you have to introduce a new variable

unsigned long skel_timeout = SKEL_TIMEOUT;

right after the definition of SKEL_TIMEOUT and replace every later occurrence of SKEL_TIMEOUT with skel_timeout.

We choose the MAGIC '4' (the ASCII character 4) and define two commands:

# define SKEL_GET_TIMEOUT 0x3401
# define SKEL_SET_TIMEOUT 0x3402

In our user process, these lines will double the time-out value:

/* ... */
unsigned long timeout;
if (ioctl (skel_hd, SKEL_GET_TIMEOUT,
           &timeout) < 0) {
    /* an error occurred (Silly billy?) */
    /* ... */
timeout *= 2;
if (ioctl (skel_hd, SKEL_SET_TIMEOUT,
           &timeout) < 0) {
    /* another error */
    /* ... */

And in our driver, these lines will do the work:

int skel_ioctl (struct inode *inode,
                struct file *file,
                unsigned int cmd,
                unsigned long arg) {
    switch (cmd) {
        put_user_long(skel_timeout, (long*) arg);
        return 0;
        skel_timeout = get_user_long((long*) arg);
        return 0;
        return -EINVAL; /* for Silly Billy */

Georg V. Zezschwitz a 27-year-old Linuxer with a taste for the practical side of Computer Science and a tendency to avoid sleep.

XXXXXXXXXXXXXXXX (XXXXXXXXXXXXXX) Like Georg, is also 27-years-old and has the same interest in the practical side of Computer Science and the same tendency to avoid sleep.

Load Disqus comments

Firstwave Cloud