Subversion Repositories or1k

<!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook V3.1//EN"[]>

<book id="MouseGuide">
 <bookinfo>
  <title>Mouse Drivers</title>
  
  <authorgroup>
   <author>
    <firstname>Alan</firstname>
    <surname>Cox</surname>
    <affiliation>
     <address>
      <email>alan@redhat.com</email>
     </address>
    </affiliation>
   </author>
  </authorgroup>

  <copyright>
   <year>2000</year>
   <holder>Alan Cox</holder>
  </copyright>

  <legalnotice>
   <para>
     This documentation is free software; you can redistribute
     it and/or modify it under the terms of the GNU General Public
     License as published by the Free Software Foundation; either
     version 2 of the License, or (at your option) any later
     version.
   </para>
      
   <para>
     This program is distributed in the hope that it will be
     useful, but WITHOUT ANY WARRANTY; without even the implied
     warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
     See the GNU General Public License for more details.
   </para>
      
   <para>
     You should have received a copy of the GNU General Public
     License along with this program; if not, write to the Free
     Software Foundation, Inc., 59 Temple Place, Suite 330, Boston,
     MA 02111-1307 USA
   </para>
      
   <para>
     For more details see the file COPYING in the source
     distribution of Linux.
   </para>
  </legalnotice>
 </bookinfo>

 <toc></toc>

 <chapter id="intro">
  <title>Introduction</title>
  <note>
   <title>Earlier publication</title>
    <para>
      Parts of this document first appeared in Linux Magazine under a
      ninety day exclusivity.
   </para>
  </note> 

  <para>
    Mice are conceptually one of the simplest device interfaces in the 
    Linux operating system. Not all mice are handled by the kernel. 
    Instead there is a two layer abstraction.
  </para>

  <para>
    The kernel mouse drivers and userspace drivers for the serial mice are
    all managed by a system daemon called <application>gpm</application> 
    - the general purpose mouse driver. <application>gpm</application> 
    handles cutting and pasting on the text consoles. It provides a 
    general library for mouse-aware applications and it handles the 
    sharing of mouse services with the 
    <application>X Window System</application> user interface.
  </para>
  <para>
    Sometimes a mouse speaks a sufficiently convoluted protocol that the
    protocol is handled by <application>Gpm</application> itself. Most 
    of the mouse drivers follow a common interface called the bus mouse 
    protocol.
  </para>
  <para>
    Each read from a bus mouse interface device returns a block of data. 
    The first three bytes of each read are defined as follows: 

   <table frame=all>
    <title>Mouse Data Encoding</title>
    <tgroup cols=2 align=left>
     <tbody>
      <row>
       <entry>Byte 0</entry>
       <entry>0x80 + the buttons currently down.</entry>
      </row>
      <row>
       <entry>Byte 1</entry>
       <entry>A signed value for the shift in X position</entry>
      </row>
      <row>
       <entry>Byte 2</entry>
       <entry>A signed value for the shift in Y position</entry>
      </row>
     </tbody>
    </tgroup>
   </table>

    An application can choose to read more than 3 bytes. The rest of the 
    bytes will be zero, or may optionally return some additional 
    device-specific information.
  </para>
  <para>
    The position values are truncated if they exceed the 8bit range (that
    is -127 &lt;= delta &lt;= 127). While the value -128 does fit into a 
    byte is not allowed.
  </para>
  <para>
    The <mousebutton>buttons</mousebutton> are numbered left to right as 
    0, 1, 2, 3.. and each button sets the relevant bit. So a user pressing 
    the left and right button of a three button mouse will set bits 0 and 2.
  </para>
  <para>
    All mice are required to support the <function>poll</function> 
    operation. Indeed pretty much every user of a mouse device uses 
    <function>poll</function> to wait for mouse events to occur.
  </para>
  <para>
    Finally the mice support asynchronous I/O. This is a topic we have not 
    yet covered but which I will explain after looking at a simple mouse 
    driver.
  </para>
 </chapter>

 <chapter id="driver">
  <title>A simple mouse driver</title>
  <para>
    First we will need the set up functions for our mouse device. To keep 
    this simple our imaginary mouse device has three I/O ports fixed at I/O 
    address 0x300 and always lives on interrupt 5.  The ports will be the X 
    position, the Y position and the buttons in that order.
  </para>

  <programlisting>
#define OURMOUSE_BASE        0x300

static struct miscdevice our_mouse = {
        OURMOUSE_MINOR, "ourmouse", &amp;our_mouse_fops
};

__init ourmouse_init(void)
{

        if (request_region(OURMOUSE_BASE, 3, "ourmouse") < 0) {
                printk(KERN_ERR "ourmouse: request_region failed.\n");
                return -ENODEV;
        }

        if (misc_register(&amp;our_mouse) < 0) {
                printk(KERN_ERR "ourmouse: cannot register misc device.\n");
                release_region(OURMOUSE_BASE, 3);
                return -EBUSY;
        }

        return 0;
}
  </programlisting>

  <para>
    The <structname>miscdevice</structname> is new here. Linux normally 
    parcels devices out by major number, and each device has 256 units. 
    For things like mice this is extremely wasteful so a device exists 
    which is used to accumulate all the odd individual devices that 
    computers tend to have.
  </para>
  <para>
    Minor numbers in this space are allocated by a central source, although 
    you can look in the kernel <filename>Documentation/devices.txt</filename>
    file and pick a free one for development use. This kernel file also 
    carries instructions for registering a device. This may change over time 
    so it is a good idea to obtain a current copy of this file first.
  </para>
  <para>
    Our code then is fairly simple. We reserve our I/O address space with
    request_region, checking to make sure that it succeeded (i.e. the
    space wasn't reserved by anyone else). 
  </para>
  <para>
    Then we ask the misc driver to allocate our minor device number. We also
    hand it our name (which is used in 
    <filename class="directory">/proc/misc</filename>) and a set of file 
    operations that are to be used. The file operations work exactly like the 
    file operations you would register for a normal character device. The misc 
    device itself is simply acting as a redirector for requests.
    Since misc_register can fail, it is important to check for failure
    and act accordingly (which in the case of a mouse driver is to abort,
    since you can't use the mouse without a working device node).
  </para>
  <para>
    Next, in order to be able to use and test our code we need to add some 
    module code to support it. This too is fairly simple:
  </para>
  <programlisting>
#ifdef MODULE

int init_module(void)
{
        if(ourmouse_init()&lt;0)
                return -ENODEV:
        return 0;
}

void cleanup_module(void)
{
        misc_deregister(&amp;our_mouse);
        free_region(OURMOUSE_BASE, 3);
}


#endif
  </programlisting>

  <para>
    The module code provides the normal two functions. The 
    <function>init_module</function> function is called when the module is 
    loaded. In our case it simply calls the initialising function we wrote 
    and returns an error if this fails. This ensures the module will only 
    be loaded if it was successfully set up.
  </para>
  <para>
    The <function>cleanup_module</function> function is called when the 
    module is unloaded. We give the miscellaneous device entry back, and 
    then free our I/O resources. If we didn't free the I/O resources then 
    the next time the module loaded it would think someone else had its I/O 
    space.
  </para>
  <para>
    Once the <function>misc_deregister</function> has been called any 
    attempts to open the mouse device will fail with the error  
    <errorcode>ENODEV</errorcode> (<errorname>No such device</errorname>).
  </para>
  <para>
    Next we need to fill in our file operations. A mouse doesn't need many 
    of these. We need to provide open, release, read and poll. That makes 
    for a nice simple structure:
  </para>

  <programlisting>
struct file_operations our_mouse_fops = {
        owner: THIS_MODULE,            /* Automatic usage management */
        read:  read_mouse,             /* You can read a mouse */
        write: write_mouse,            /* This won't do a lot */
        poll:  poll_mouse,             /* Poll */
        open:  open_mouse,             /* Called on open */
        release: close_mouse,          /* Called on close */
};
  </programlisting>

  <para>
    There is nothing particularly special needed here. We provide functions 
    for all the relevant or required operations and little else. There is 
    nothing stopping us providing an ioctl function for this mouse. Indeed 
    if you have a configurable mouse it may be very appropriate to provide 
    configuration interfaces via ioctl calls.
  </para>
  <para>
    The syntax we use is not standard C as such. GCC provides the ability
    to initialise fields by name, and this generally makes the method table
    much easier to read than counting through NULL pointers and remembering
    the order by hand.
  </para>
  <para>
    The owner field is used to manage the locking of module load an
    unloading. It is obviously important that a module is not unloaded while
    in use. When your device is opened the module specified by "owner" is 
    locked. When it is finally released the module is unlocked.
  </para>
  <para>
    The open and close routines need to manage enabling and disabling the 
    interrupts for the mouse as well as stopping the mouse being unloaded
    when it is no longer required. 
  </para>

  <programlisting>
static int mouse_users = 0;                /* User count */
static int mouse_dx = 0;                   /* Position changes */
static int mouse_dy = 0;
static int mouse_event = 0;                /* Mouse has moved */

static int open_mouse(struct inode *inode, struct file *file)
{
        if(mouse_users++)
                return 0;

        if(request_irq(mouse_intr, OURMOUSE_IRQ, 0, "ourmouse", NULL))
        {
                mouse_users--;
                return -EBUSY;
        }
        mouse_dx = 0;
        mouse_dy = 0;
        mouse_event = 0;
        mouse_buttons = 0;
        return 0;
}
  </programlisting>
  <para>
    The open function has to do a small amount of housework. We keep a count 
    of the number of times the mouse is open. This is because we do not want 
    to request the interrupt multiple times. If the mouse has at least one 
    user then it is set up and we simply add to the user count and return
    <returnvalue>0</returnvalue> for success.
  </para>
  <para>
    We grab the interrupt and thus start mouse interrupts. If the interrupt 
    has been borrowed by some other driver then <function>request_irq</function>
    will fail and we will return an error. If we were capable of sharing an 
    interrupt line we would specify <constant>SA_SHIRQ</constant> instead of 
    <constant>zero</constant>. Provided that everyone claiming an interrupt 
    sets this flag, they get to share the line. <hardware>PCI</hardware> can 
    share interrupts, <hardware>ISA</hardware> normally however cannot. 
  </para>
  <para>
    We do the housekeeping. We make the current mouse position the starting
    point for accumulated changes and declare that nothing has happened
    since the mouse driver was opened.
  </para>
  <para>
    The release function needs to unwind all these:
  </para>
  <programlisting>
static int close_mouse(struct inode *inode, struct file *file)
{
        if(--mouse_users)
                return 0;
        free_irq(OURMOUSE_IRQ, NULL);
        return 0;
}
  </programlisting>
  <para>
    We count off a user and provided that there are still other users need 
    take no further action. The last person closing the mouse causes us to 
    free up the interrupt. This stops interrupts from the mouse from using 
    our CPU time, and ensures that the mouse can now be unloaded.
  </para>
  <para>
    We can fill in the write handler at this point as the write function for 
    our mouse simply declines to allow writes:
  </para>

  <programlisting>
static ssize_t write_mouse(struct file *file, const char *buffer, size_t
                                count, loff_t *ppos)
{
        return -EINVAL;
}
  </programlisting>

  <para>
    This is pretty much self-explanatory. Whenever you write you get told 
    it was an invalid function.
  </para>
  <para>
    To make the poll and read functions work we have to consider how we 
    handle the mouse interrupt. 
  </para>

  <programlisting>
static struct wait_queue *mouse_wait;
static spinlock_t mouse_lock = SPIN_LOCK_UNLOCKED;

static void ourmouse_interrupt(int irq, void *dev_id, struct pt_regs *regs)
{
        char delta_x;
        char delta_y;
        unsigned char new_buttons;

        delta_x = inb(OURMOUSE_BASE);
        delta_y = inb(OURMOUSE_BASE+1);
        new_buttons = inb(OURMOUSE_BASE+2);

        if(delta_x || delta_y || new_buttons != mouse_buttons)
        {
                /* Something happened */

                spin_lock(&amp;mouse_lock);
                mouse_event = 1;
                mouse_dx += delta_x;
                mouse_dy += delta_y;
                mouse_buttons = new_buttons;
                spin_unlock(&amp;mouse_lock);
                
                wake_up_interruptible(&amp;mouse_wait);
        }
}
  </programlisting>

  <para>
    The interrupt handler reads the mouse status. The next thing we do is 
    to check whether something has changed. If the mouse was smart it would
    only interrupt us if something had changed, but let's assume our mouse 
    is stupid as most mice actually tend to be. 
  </para>
  <para>
    If the mouse has changed we need to update the status variables. What we
    don't want is the mouse functions reading these variables to read them
    during a change. We add a spinlock that protects these variables while we
    play with them.
  </para>
  <para>
    If a change has occurred we also need to wake sleeping processes, so we 
    add a wakeup call and a <structname>wait_queue</structname> to use when 
    we wish to await a mouse event.
  </para>
  <para>
    Now we have the wait queue we can implement the poll function for the 
    mouse relatively easily:
  </para>

  <programlisting>
static unsigned int mouse_poll(struct file *file, poll_table *wait)
{
        poll_wait(file, &amp;mouse_wait, wait);
        if(mouse_event)
                return POLLIN | POLLRDNORM;
        return 0;
}
  </programlisting>

  <para>
    This is fairly standard poll code. First we add the wait queue to the 
    list of queues we want to monitor for an event. Secondly we check if an 
    event has occurred. We only have one kind of event - the 
    <varname>mouse_event</varname> flag tells us that something happened. 
    We know that this something can only be mouse data. We return the flags 
    indicating input and normal reading will succeed.
  </para>
  <para>
    You may be wondering what happens if the function returns saying 'no 
    event yet'. In this case the wake up from the wait queue we added to 
    the poll table will cause the function to be called again. Eventually 
    we will be woken up and have an event ready. At this point the 
    <function>poll</function> call will exit back to the user.
  </para>
  <para>
    After the poll completes the user will want to read the data. We now 
    need to think about how our <function>mouse_read</function> function 
    will work:
  </para>
  <programlisting>
static ssize_t mouse_read(struct file *file, char *buffer, 
                size_t count, loff_t *pos)
{
        int dx, dy;
        unsigned char button;
        unsigned long flags;
        int n;

        if(count&lt;3)
                return -EINVAL;

        /*
          *        Wait for an event
         */

        while(!mouse_event)
        {
                if(file-&gt;f_flags&amp;O_NDELAY)
                        return -EAGAIN;
                interruptible_sleep_on(&amp;mouse_wait);
                if(signal_pending(current))
                        return -ERESTARTSYS;
        }
  </programlisting>

  <para>
    We start by validating that the user is reading enough data. We could 
    handle partial reads if we wanted but it isn't terribly useful and the 
    mouse drivers don't bother to try.
  </para>
  <para>
    Next we wait for an event to occur. The loop is fairly standard event
    waiting in Linux. Having checked that the event has not yet occurred, we
    then check if an event is pending and if not we need to sleep. 
  </para>
  <para>
    A user process can set the <constant>O_NDELAY</constant> flag on a file 
    to indicate that it wishes to be told immediately if no event is 
    pending. We check this and give the appropriate error if so. 
  </para>
  <para>
    Next we sleep until the mouse or a signal awakens us. A signal will 
    awaken us as we have used <function>wakeup_interruptible</function>. 
    This is important as it means a user can kill processes waiting for 
    the mouse - clearly a desirable property. If we are interrupted we 
    exit the call and the kernel will then process signals and maybe 
    restart the call again - from the beginning.
  </para>
  <para>
    This code contains a classic Linux bug. All will be revealed later in this
    article as well as explanations for how to avoid it.
  </para>
  <programlisting>
        /* Grab the event */

        spinlock_irqsave(&amp;mouse_lock, flags);

        dx = mouse_dx;
        dy = mouse_dy;
        button = mouse_buttons;

        if(dx&lt;=-127)
                dx=-127;
        if(dx&gt;=127)
                dx=127;
        if(dy&lt;=-127)
                dy=-127;
        if(dy&gt;=127)
                dy=127;

        mouse_dx -= dx;
        mouse_dy -= dy;
        
        if(mouse_dx == 0 &amp;&amp; mouse_dy == 0)
                mouse_event = 0;

        spin_unlock_irqrestore(&amp;mouse_lock, flags);
  </programlisting>
  <para>
    This is the next stage. Having established that there is an event 
    going, we capture it. To be sure that the event is not being updated 
    as we capture it we also take the spinlock and thus prevent parallel 
    updates. Note here we use <function>spinlock_irqsave</function>. We 
    need to disable interrupts on the local processor otherwise bad things 
    will happen.
  </para>
  <para>
    What will occur is that we take the spinlock. While we hold the lock 
    an interrupt will occur. At this point our interrupt handler will try 
    and take the spinlock. It will sit in a loop waiting for the read 
    routine to release the lock. However because we are sitting in a loop 
    in the interrupt handler we will never release the lock. The machine 
    hangs and the user gets upset.
  </para>
  <para>
    By blocking the interrupt on this processor we ensure that the lock 
    holder will always give the lock back without deadlocking.
  </para>
  <para>
    There is a little cleverness in the reporting mechanism too. We can 
    only report a move of 127 per read. We don't however want to lose 
    information by throwing away further movement. Instead we keep 
    returning as much information as possible. Each time we return a 
    report we remove the amount from the pending movement in 
    <varname>mouse_dx</varname> and <varname>mouse_dy</varname>. Eventually 
    when these counts hit zero we clear the <varname>mouse_event</varname>
    flag as there is nothing else left to report.
  </para>

  <programlisting>
        if(put_user(button|0x80, buffer))
                return -EFAULT;
        if(put_user((char)dx, buffer+1))
                return -EFAULT;
        if(put_user((char)dy, buffer+2))
                return -EFAULT;

        for(n=3; n < count; n++)
                if(put_user(0x00, buffer+n))
                        return -EFAULT;

        return count;
}
  </programlisting>

  <para>
    Finally we must put the results in the user supplied buffer. We cannot 
    do this while holding the lock as a write to user memory may sleep. 
    For example the user memory may be residing on disk at this instant. 
    Thus we did our computation beforehand and now copy the data. Each 
    <function>put_user call</function> is filling in one byte of the buffer. 
    If it returns an error we inform the program that it passed us an 
    invalid buffer and abort.
  </para>
  <para>
    Having written the data we blank the rest of the buffer that was read 
    and report the read as being successful.
  </para>
 </chapter>

 <chapter id="debugging">
  <title>Debugging the mouse driver</title>

  <para>
    We now have an almost perfectly usable mouse driver. If you were to 
    actually try and use it however you would eventually find a couple of 
    problems with it. A few programs will also not work with as it does not 
    yet support asynchronous I/O.
  </para>
  <para>
    First let us look at the bugs. The most obvious one isn't really a driver
    bug but a failure to consider the consequences. Imagine you bumped the 
    mouse hard by accident and sent it skittering across the desk. The mouse 
    interrupt routine will add up all that movement and report it in steps of 
    127 until it has reported all of it. Clearly there is a point beyond 
    which mouse movement isn't worth reporting. We need to add this as a 
    limit to the interrupt handler:
  </para>

  <programlisting>
static void ourmouse_interrupt(int irq, void *dev_id, struct pt_regs *regs)
{
        char delta_x;
        char delta_y;
        unsigned char new_buttons;

        delta_x = inb(OURMOUSE_BASE);
        delta_y = inb(OURMOUSE_BASE+1);
        new_buttons = inb(OURMOUSE_BASE+2);

        if(delta_x || delta_y || new_buttons != mouse_buttons)
        {
                /* Something happened */

                spin_lock(&amp;mouse_lock);
                mouse_event = 1;
                mouse_dx += delta_x;
                mouse_dy += delta_y;

                if(mouse_dx &lt; -4096)
                        mouse_dx = -4096;
                if(mouse_dx &gt; 4096)
                        mouse_dx = 4096;

                if(mouse_dy &lt; -4096)
                        mouse_dy = -4096;
                if(mouse_dy &gt; 4096)
                        mouse_dy = 4096;

                mouse_buttons = new_buttons;
                spin_unlock(&amp;mouse_lock);
                
                wake_up_interruptible(&amp;mouse_wait);
        }
}
  </programlisting>

  <para>
    By adding these checks we limit the range of accumulated movement to
    something sensible. 
  </para>
  <para>
    The second bug is a bit more subtle, and that is perhaps why this is 
    such a common mistake. Remember, I said the waiting loop for the read 
    handler had a bug in it. Think about what happens when we execute:
  </para>

  <programlisting>
        while(!mouse_event)
        {
  </programlisting>

  <para>
    and an interrupt occurs at this point here. This causes a mouse movement
    and wakes up the queue. 
  </para>

  <programlisting>
                interruptible_sleep_on(&amp;mouse_wait);
  </programlisting>

  <para>
    Now we sleep on the queue. We missed the wake up and the application 
    will not see an event until the next mouse event occurs. This will 
    lead to just the odd instance when a mouse button gets delayed. The 
    consequences to the user will probably be almost undetectable with a 
    mouse driver. With other drivers this bug could be a lot more severe.
  </para>
  <para>
    There are two ways to solve this. The first is to disable interrupts 
    during the testing and the sleep. This works because when a task sleeps 
    it ceases to disable interrupts, and when it resumes it disables them 
    again. Our code thus becomes:
  </para>

  <programlisting>
        save_flags(flags);
        cli();

        while(!mouse_event)
        {
                if(file-&gt;f_flags&amp;O_NDELAY)
                {
                        restore_flags(flags);
                        return -EAGAIN;
                }
                interruptible_sleep_on(&amp;mouse_wait);
                if(signal_pending(current))
                {
                        restore_flags(flags);
                        return -ERESTARTSYS;
                }
        }
        restore_flags(flags);
  </programlisting>

  <para>
    This is the sledgehammer approach. It works but it means we spend a 
    lot more time turning interrupts on and off. It also affects 
    interrupts globally and has bad properties on multiprocessor machines 
    where turning interrupts off globally is not a simple operation, but 
    instead involves kicking each processor, waiting for them to disable 
    interrupts and reply.
  </para>
  <para>
    The real problem is the race between the event testing and the sleeping. 
    We can avoid that by using the scheduling functions more directly. 
    Indeed this is the way they generally should be used for an interrupt.
  </para>

  <programlisting>
        struct wait_queue wait = { current, NULL };

        add_wait_queue(&amp;mouse_wait, &amp;wait);
        set_current_state(TASK_INTERRUPTIBLE);
        
        while(!mouse_event)
        {
                if(file-&gt;f_flags&amp;O_NDELAY)
                {
                        remove_wait_queue(&amp;mouse_wait, &amp;wait);
                        set_current_state(TASK_RUNNING);
                        return -EWOULDBLOCK;
                }
                if(signal_pending(current))
                {
                        remove_wait_queue(&amp;mouse_wait, &amp;wait);
                        current-&gt;state = TASK_RUNNING;
                        return -ERESTARTSYS;
                }
                schedule();
                set_current_state(TASK_INTERRUPTIBLE);
        }
        
        remove_wait_wait(&amp;mouse_wait, &amp;wait);
        set_current_state(TASK_RUNNING);
  </programlisting>

  <para>
    At first sight this probably looks like deep magic. To understand how 
    this works you need to understand how scheduling and events work on 
    Linux. Having a good grasp of this is one of the keys to writing clean 
    efficient device drivers.
  </para>
  <para>
    <function>add_wait_queue</function> does what its name suggests. It adds 
    an entry to the <varname>mouse_wait</varname> list. The entry in this 
    case is the entry for our current process (<varname>current</varname>
    is the current task pointer). 
  </para>
  <para>
    So we start by adding an entry for ourself onto the 
    <varname>mouse_wait</varname> list. This does not put us to sleep 
    however. We are merely tagged onto the list. 
  </para>
  <para>
    Next we set our status to <constant>TASK_INTERRUPTIBLE</constant>. Again 
    this does not mean we are now asleep. This flag says what should happen 
    next time the process sleeps. <constant>TASK_INTERRUPTIBLE</constant> says 
    that the process should not be rescheduled. It will run from now until it 
    sleeps and then will need to be woken up.
  </para>
  <para>
    The <function>wakeup_interruptible</function> call in the interrupt 
    handler can now be explained in more detail. This function is also very 
    simple. It goes along the list of processes on the queue it is given and 
    any that are marked as <constant>TASK_INTERRUPTIBLE</constant> it changes 
    to <constant>TASK_RUNNING</constant> and tells the kernel that new 
    processes are runnable.
  </para>
  <para>
    Behind all the wrappers in the original code what is happening is this
  </para>

  <procedure>
   <step>
    <para>
      We add ourself to the mouse wait queue
    </para>
   </step>
   <step>
    <para>
      We mark ourself as sleeping
    </para>
   </step>
   <step>
    <para>
      We ask the kernel to schedule tasks again
    </para>
   </step>
   <step>
    <para>
      The kernel sees we are asleep and schedules someone else.
    </para>
   </step>
   <step>
    <para>
      The mouse interrupt sets our state to <constant>TASK_RUNNING</constant> 
      and makes a note that the kernel should reschedule tasks
    </para>
   </step>
   <step>
    <para>
      The kernel sees we are running again and continues our execution
    </para>
   </step>
  </procedure>
  <para>
    This is why the apparent magic works. Because we mark ourself as
    <constant>TASK_INTERRUPTIBLE</constant> and as we add ourselves 
    to the queue before we check if there are events pending, the race 
    condition is removed.
  </para>
  <para>
    Now if an interrupt occurs after we check the queue status and before 
    we call the <function>schedule</function> function in order to sleep, 
    things work out. Instead of missing an event, we are set back to 
    <constant>TASK_RUNNING</constant> by the mouse interrupt. We still call 
    <function>schedule</function> but it will continue running our task. 
    We go back around the loop and this time there may be an event.
  </para>
  <para>
    There will not always be an event. Thus we set ourselves back to
    <constant>TASK_INTERRUPTIBLE</constant> before resuming the loop. 
    Another process doing a read may already have cleared the event flag, 
    and if so we will need to go back to sleep again. Eventually we will 
    get our event and escape.
  </para>
  <para>
    Finally when we exit the loop we remove ourselves from the 
    <varname>mouse_wait</varname> queue as we are no longer interested
    in mouse events, and we set ourself back to 
    <constant>TASK_RUNNABLE</constant> as we do not wish to go to sleep 
    again just yet.
  </para>
  <note>
   <title>Note</title> 
   <para>
     This isn't an easy topic. Don't be afraid to reread the description a 
     few times and also look at other device drivers to see how it works. 
     Finally if you can't grasp it just yet, you can use the code as 
     boilerplate to write other drivers and trust me instead.
   </para>
  </note>
 </chapter>

 <chapter id="asyncio">
  <title>Asynchronous I/O</title>
  <para>
    This leaves the missing feature - Asynchronous I/O. Normally UNIX 
    programs use the <function>poll</function> call (or its variant form 
    <function>select</function>) to wait for an event to occur on one of 
    multiple input or output devices. This model works well for most tasks 
    but because <function>poll</function> and <function>select</function> 
    wait for an event isn't suitable for tasks that are also continually 
    doing computation work. Such programs really want the kernel to kick 
    them when something happens rather than watch for events.
  </para>
  <para>
    Poll is akin to having a row of lights in front of you. You can see at a
    glance which ones if any are lit. You cannot however get anything useful
    done while watching them. Asynchronous I/O uses signals which work more 
    like a door bell. Instead of you watching, it tells you that something 
    is up.
  </para>
  <para>
    Asynchronous I/O sends the signal SIGIO to a user process when the I/O 
    events occur. In this case that means when people move the mouse. The 
    SIGIO signal causes the user process to jump to its signal handler and 
    execute code in that handler before returning to whatever was going on 
    previously. It is the application equivalent of an interrupt handler.
  </para>
  <para>
    Most of the code needed for this operation is common to all its users. 
    The kernel provides a simple set of functions for managing asynchronous 
    I/O.
  </para>
  <para>
    Our first job is to allow users to set asynchronous I/O on file handles. 
    To do that we need to add a new function to the file operations table for 
    our mouse:
  </para>

  <programlisting>
struct file_operations our_mouse_fops = {
        owner: THIS_MODULE
        read:  read_mouse,      /* You can read a mouse */
        write: write_mouse,     /* This won't do a lot */
        poll:  poll_mouse,      /* Poll */
        open:  open_mouse,      /* Called on open */
        release: close_mouse,   /* Called on close */
        fasync: fasync_mouse,   /* Asynchronous I/O */
};
  </programlisting>

  <para>
    Once we have installed this entry the kernel knows we support 
    asynchronous I/O and will allow all the relevant operations on the 
    device. Whenever a user adds or removes asynchronous I/O notification 
    on a file handle it calls our <function>fasync_mouse</function> routine 
    we just added. This routine uses the helper functions to keep the queue 
    of handles up to date:
  </para>

  <programlisting>
static struct fasync_struct *mouse_fasync = NULL;

static int fasync_mouse(int fd, struct file *filp, int on)
{
         int retval = fasync_helper(fd, filp, on, &amp;mouse_fasync);

         if (retval &lt; 0)
                 return retval;
        return 0;
}
  </programlisting>

  <para>
    The fasync helper adds and deletes entries by managing the supplied 
    list. We also need to remove entries from this list when the file is 
    closed. This requires we add one line to our close function:
  </para>

  <programlisting>
static int close_mouse(struct inode *inode, struct file *file)
{
        fasync_mouse(-1, file, 0)
        if(--mouse_users)
                return 0;
        free_irq(OURMOUSE_IRQ, NULL);
        MOD_DEC_USE_COUNT;
        return 0;
}
  </programlisting>

  <para>
    When we close the file we now call our own fasync handler as if the 
    user had requested that this file cease to be used for asynchronous 
    I/O. This rather neatly cleans up any loose ends. We certainly don't 
    wait to deliver a signal for a file that no longer exists.
  </para>
  <para>
    At this point the mouse driver supports all the asynchronous I/O 
    operations, and applications using them will not error. They won't 
    however work yet. We need to actually send the signals. Again the 
    kernel provides a function for handling this.
  </para>
  <para>
    We update our interrupt handler a little:
  </para>

  <programlisting>
static void ourmouse_interrupt(int irq, void *dev_id, struct pt_regs *regs)
{
        char delta_x;
        char delta_y;
        unsigned char new_buttons;

        delta_x = inb(OURMOUSE_BASE);
        delta_y = inb(OURMOUSE_BASE+1);
        new_buttons = inb(OURMOUSE_BASE+2);

        if(delta_x || delta_y || new_buttons != mouse_buttons)
        {
                /* Something happened */

                spin_lock(&amp;mouse_lock);
                mouse_event = 1;
                mouse_dx += delta_x;
                mouse_dy += delta_y;

                if(mouse_dx &lt; -4096)
                        mouse_dx = -4096;
                if(mouse_dx &gt; 4096)
                        mouse_dx = 4096;

                if(mouse_dy &lt; -4096)
                        mouse_dy = -4096;
                if(mouse_dy &gt; 4096)
                        mouse_dy = 4096;

                mouse_buttons = new_buttons;
                spin_unlock(&amp;mouse_lock);

                /* Now we do asynchronous I/O */
                kill_fasync(&amp;mouse_fasync, SIGIO); 
                
                wake_up_interruptible(&amp;mouse_wait);
        }
}
  </programlisting>

  <para>
    The new code simply calls the <function>kill_fasync</function> routine
    provided by the kernel if the queue is non-empty. This sends the 
    required signal (SIGIO in this case) to the process each file handle 
    says should be informed about the exciting new mouse movement that 
    just happened.
  </para>
  <para>
    With this in place and the bugs in the original version fixed, you now 
    have a fully functional mouse driver using the bus mouse protocol. It 
    will work with the <application>X window system</application>, will work 
    with <application>GPM</application> and should work with every other 
    application you need. <application>Doom</application> is of course the 
    ideal way to test your new mouse driver is functioning properly. Be sure 
    to test it thoroughly.
  </para>
 </chapter>
</book>