User-space device drivers

The HyperNews Linux KHG Discussion Pages

It is not always necessary to write a device driver for a device, especially in applications where no two applications will compete for the device. The most useful example of this is a memory-mapped device, but you can also do this with devices in I/O space (devices accessed with inb() and outb(), etc.). If your process is running as superuser (root), you can use the mmap() call to map some of your process memory to actual memory locations, by mmap()'ing a section of /dev/mem. When you have done this mapping, it is pretty easy to write and read from real memory addresses just as you would read and write any variables.

If your driver needs to respond to interrupts, then you really need to be working in kernel space, and need to write a real device driver, as there is no good way at this time to deliver interrupts to user processes. Although the DOSEMU project has created something called the SIG (Silly Interrupt Generator) which allows interrupts to be posted to user processes (I believe through the use of signals), the SIG is not particularly fast, and should be thought of as a last resort for things like DOSEMU.

An interrupt is an asyncronous notification posted by the hardware to alert the device driver of some condition. You have likely dealt with `IRQ's when setting up your hardware; an IRQ is an ``Interrupt ReQuest line,'' which is triggered when the device wants to talk to the driver. This may be because it has data to give to the drive, or because it is now ready to receive data, or because of some other ``exceptional condition'' that the driver needs to know about. It is similar to user-level processes receiving a signal, so similar that the same sigaction structure is used in the kernel to deal with interrupts as is used in user-level programs to deal with signals. Where the user-level has its signals delivered to it by the kernel, the kernel has interrupt delivered to it by hardware.

If your driver must be accessible to multiple processes at once, and/or manage contention for a resource, then you also need to write a real device driver at the kernel level, and a user-space device driver will not be sufficient or even possible.

Example: `vgalib`

A good example of a user-space driver is the vgalib library. The standard read() and write() calls are really inadequate for writing a really fast graphics driver, and so instead there is a library which acts conceptually like a device driver, but runs in user space. Any processes which use it must run setuid root, because it uses the ioperm() system call. It is possible for a process that is not setuid root to write to /dev/mem if you have a group mem or kmem which is allowed write permission to /dev/mem and the process is properly setgid, but only a process running as root can execute the ioperm() call.

There are several I/O ports associated with VGA graphics. vgalib creates symbolic names for this with #define statements, and then issues the ioperm() call like this to make it possible for the process to read and write directly from and to those ports:

    if (ioperm(CRT_IC, 1, 1)) {
        printf("VGAlib: can't get I/O permissions \n");
        exit (-1);
    }
    ioperm(CRT_IM,  1, 1);
    ioperm(ATT_IW, 1, 1);
[...]

It only needs to do error checking once, because the only reason for the ioperm() call to fail is that it is not being called by the superuser, and this status is not going to change.

[Careful!] After making this call, the process is allowed to use inb and outb machine instructions, but only on the specified ports. These instructions can be accessed without writing directly in assembly by including , but will only work if you compile with optimization on, by giving the -O? to gcc. Read <linux/asm.h> for details.

After arranging for port I/O, vgalib arranges for writing directly to kernel memory with the following code:

    /* open /dev/mem */
    if ((mem_fd = open("/dev/mem", O_RDWR) ) < 0) {
        printf("VGAlib: can't open /dev/mem \n");
        exit (-1);
    }

    /* mmap graphics memory */
    if ((graph_mem = malloc(GRAPH_SIZE + (PAGE_SIZE-1))) == NULL) {
        printf("VGAlib: allocation error \n");
        exit (-1);
    }
    if ((unsigned long)graph_mem % PAGE_SIZE)
        graph_mem += PAGE_SIZE - ((unsigned long)graph_mem % PAGE_SIZE);
    graph_mem = (unsigned char *)mmap(
        (caddr_t)graph_mem,
        GRAPH_SIZE,
        PROT_READ|PROT_WRITE,
        MAP_SHARED|MAP_FIXED,
        mem_fd,
        GRAPH_BASE
    );
    if ((long)graph_mem < 0) {
        printf("VGAlib: mmap error \n");
        exit (-1);
    }

It first opens /dev/mem, then allocates memory enough so that the mapping can be done on a page (4 KB) boundary, and then attempts the map. GRAPH_SIZE is the size of VGA memory, and GRAPH_BASE is the first address of VGA memory in /dev/mem. Then by writing to the address that is returned by mmap(), the process is actually writing to screen memory.

Example: mouse conversion

If you want a driver that acts a bit more like a kernel-level driver, but does not live in kernel space, you can also make a fifo, or named pipe. This usually lives in the /dev/ directory (although it doesn't need to) and acts substantially like a device once set up. However, fifo's are one-directional only--they have one reader and one writer.

For instance, it used to be that if you had a PS/2-style mouse, and wanted to run XFree86, you had to create a fifo called /dev/mouse, and run a program called mconv which read PS/2 mouse ``droppings'' from /dev/psaux, and wrote the equivalent microsoft-style ``droppings'' to /dev/mouse. Then XFree86 would read the ``droppings'' from /dev/mouse, and it would be as if there were a microsoft mouse connected to /dev/mouse. Even though XFree86 is now able to read PS/2 style ``droppings'', the concepts in this example still stand. (If you have a better example, I'd be glad to see it.)

The evil instruction

Don't use the cli() instruction. It's possible to use it as root to disable interrupts, and one particular program used to used to use it--the clock program. However, this kills SMP machines. If you need to use cli(), you need a kernel-space driver, and a user-space driver will only cause grief as more and more Linux users use SMP machines.

Messages

1. What is SMP?
-> SMP: Two Definitions? by Reinhold J. Gerharz
-> Only one definition for Linux... by Michael K. Johnson