Unreliable Guide To Hacking The Linux Kernel

User context is when you are coming in from a system call or other trap: like userspace, you can be preempted by more important tasks and by interrupts. You can sleep, by calling schedule().

In user context, the current pointer (indicating the task we are currently executing) is valid, and in_interrupt() (include/linux/interrupt.h) is false .

Timer ticks, network cards and keyboard are examples of real hardware which produce interrupts at any time. The kernel runs interrupt handlers, which services the hardware. The kernel guarantees that this handler is never re-entered: if the same interrupt arrives, it is queued (or dropped). Because it disables interrupts, this handler has to be fast: frequently it simply acknowledges the interrupt, marks a 'software interrupt' for execution and exits.

You can tell you are in a hardware interrupt, because in_irq() returns true.

Whenever a system call is about to return to userspace, or a hardware interrupt handler exits, any 'software interrupts' which are marked pending (usually by hardware interrupts) are run (kernel/softirq.c).

Much of the real interrupt handling work is done here. Early in the transition to SMP, there were only 'bottom halves' (BHs), which didn't take advantage of multiple CPUs. Shortly after we switched from wind-up computers made of match-sticks and snot, we abandoned this limitation and switched to 'softirqs'.

include/linux/interrupt.h lists the different softirqs. A very important softirq is the timer softirq (include/linux/timer.h): you can register to have it call functions for you in a given length of time.

Softirqs are often a pain to deal with, since the same softirq will run simultaneously on more than one CPU. For this reason, tasklets (include/linux/interrupt.h) are more often used: they are dynamically-registrable (meaning you can have as many as you want), and they also guarantee that any tasklet will only run on one CPU at any time, although different tasklets can run simultaneously.

You can tell you are in a softirq (or tasklet) using the in_softirq() macro (include/linux/interrupt.h).

printk() feeds kernel messages to the console, dmesg, and the syslog daemon. It is useful for debugging and reporting errors, and can be used inside interrupt context, but use with caution: a machine which has its console flooded with printk messages is unusable. It uses a format string mostly compatible with ANSI C printf, and C string concatenation to give it a first "priority" argument:

printk(KERN_INFO "i = %u\n", i);
   

See include/linux/kernel.h; for other KERN_ values; these are interpreted by syslog as the level. Special case: for printing an IP address use

__be32 ipaddress;
printk(KERN_INFO "my ip: %pI4\n", &ipaddress);
   

printk() internally uses a 1K buffer and does not catch overruns. Make sure that will be enough.

[SLEEPS]

put_user() and get_user() are used to get and put single values (such as an int, char, or long) from and to userspace. A pointer into userspace should never be simply dereferenced: data should be copied using these routines. Both return -EFAULT or 0.

copy_to_user() and copy_from_user() are more general: they copy an arbitrary amount of data to and from userspace.

[Yes, this moronic interface makes me cringe. The flamewar comes up every year or so. --RR.]

The functions may sleep implicitly. This should never be called outside user context (it makes no sense), with interrupts disabled, or a spinlock held.

[MAY SLEEP: SEE BELOW]

These routines are used to dynamically request pointer-aligned chunks of memory, like malloc and free do in userspace, but kmalloc() takes an extra flag word. Important values:

If you see a sleeping function called from invalid context warning message, then maybe you called a sleeping allocation function from interrupt context without GFP_ATOMIC. You should really fix that. Run, don't walk.

If you are allocating at least PAGE_SIZE (include/asm/page.h) bytes, consider using __get_free_pages() (include/linux/mm.h). It takes an order argument (0 for page sized, 1 for double page, 2 for four pages etc.) and the same memory priority flag word as above.

If you are allocating more than a page worth of bytes you can use vmalloc(). It'll allocate virtual memory in the kernel map. This block is not contiguous in physical memory, but the MMU makes it look like it is for you (so it'll only look contiguous to the CPUs, not to external device drivers). If you really need large physically contiguous memory for some weird device, you have a problem: it is poorly supported in Linux because after some time memory fragmentation in a running kernel makes it hard. The best way is to allocate the block early in the boot process via the alloc_bootmem() routine.

Before inventing your own cache of often-used objects consider using a slab cache in include/linux/slab.h

This global variable (really a macro) contains a pointer to the current task structure, so is only valid in user context. For example, when a process makes a system call, this will point to the task structure of the calling process. It is not NULL in interrupt context.

The udelay() and ndelay() functions can be used for small pauses. Do not use large values with them as you risk overflow - the helper function mdelay() is useful here, or consider msleep().

The cpu_to_be32() family (where the "32" can be replaced by 64 or 16, and the "be" can be replaced by "le") are the general way to do endian conversions in the kernel: they return the converted value. All variations supply the reverse as well: be32_to_cpu(), etc.

There are two major variations of these functions: the pointer variation, such as cpu_to_be32p(), which take a pointer to the given type, and return the converted value. The other variation is the "in-situ" family, such as cpu_to_be32s(), which convert value referred to by the pointer, and return void.

These routines disable hard interrupts on the local CPU, and restore them. They are reentrant; saving the previous state in their one unsigned long flags argument. If you know that interrupts are enabled, you can simply use local_irq_disable() and local_irq_enable().

These routines disable soft interrupts on the local CPU, and restore them. They are reentrant; if soft interrupts were disabled before, they will still be disabled after this pair of functions has been called. They prevent softirqs and tasklets from running on the current CPU.

get_cpu() disables preemption (so you won't suddenly get moved to another CPU) and returns the current processor number, between 0 and NR_CPUS. Note that the CPU numbers are not necessarily continuous. You return it again with put_cpu() when you are done.

If you know you cannot be preempted by another task (ie. you are in interrupt context, or have preemption disabled) you can use smp_processor_id().

After boot, the kernel frees up a special section; functions marked with __init and data structures marked with __initdata are dropped after boot is complete: similarly modules discard this memory after initialization. __exit is used to declare a function which is only required on exit: the function will be dropped if this file is not compiled as a module. See the header file for use. Note that it makes no sense for a function marked with __init to be exported to modules with EXPORT_SYMBOL() - this will break.

Many parts of the kernel are well served as a module (dynamically-loadable parts of the kernel). Using the module_init() and module_exit() macros it is easy to write code without #ifdefs which can operate both as a module or built into the kernel.

The module_init() macro defines which function is to be called at module insertion time (if the file is compiled as a module), or at boot time: if the file is not compiled as a module the module_init() macro becomes equivalent to __initcall(), which through linker magic ensures that the function is called on boot.

The function can return a negative error number to cause module loading to fail (unfortunately, this has no effect if the module is compiled into the kernel). This function is called in user context with interrupts enabled, so it can sleep.

This macro defines the function to be called at module removal time (or never, in the case of the file compiled into the kernel). It will only be called if the module usage count has reached zero. This function can also sleep, but cannot fail: everything must be cleaned up by the time it returns.

Note that this macro is optional: if it is not present, your module will not be removable (except for 'rmmod -f').

These manipulate the module usage count, to protect against removal (a module also can't be removed if another module uses one of its exported symbols: see below). Before calling into module code, you should call try_module_get() on that module: if it fails, then the module is being removed and you should act as if it wasn't there. Otherwise, you can safely enter the module, and call module_put() when you're finished.

Most registerable structures have an owner field, such as in the file_operations structure. Set this field to the macro THIS_MODULE.

You declare a wait_queue_head_t using the DECLARE_WAIT_QUEUE_HEAD() macro, or using the init_waitqueue_head() routine in your initialization code.

Placing yourself in the waitqueue is fairly complex, because you must put yourself in the queue before checking the condition. There is a macro to do this: wait_event_interruptible() include/linux/wait.h The first argument is the wait queue head, and the second is an expression which is evaluated; the macro returns 0 when this expression is true, or -ERESTARTSYS if a signal is received. The wait_event() version ignores signals.

Do not use the sleep_on() function family - it is very easy to accidentally introduce races; almost certainly one of the wait_event() family will do, or a loop around schedule_timeout(). If you choose to loop around schedule_timeout() remember you must set the task state (with set_current_state()) on each iteration to avoid busy-looping.

Call wake_up() include/linux/wait.h;, which will wake up every process in the queue. The exception is if one has TASK_EXCLUSIVE set, in which case the remainder of the queue will not be woken. There are other variants of this basic function available in the same header.

This is the classic method of exporting a symbol: dynamically loaded modules will be able to use the symbol as normal.

Similar to EXPORT_SYMBOL() except that the symbols exported by EXPORT_SYMBOL_GPL() can only be seen by modules with a MODULE_LICENSE() that specifies a GPL compatible license. It implies that the function is considered an internal implementation issue, and not really an interface.

There used to be three sets of linked-list routines in the kernel headers, but this one is the winner. If you don't have some particular pressing need for a single list, it's a good choice.

In particular, list_for_each_entry is useful.

For code called in user context, it's very common to defy C convention, and return 0 for success, and a negative error number (eg. -EFAULT) for failure. This can be unintuitive at first, but it's fairly widespread in the kernel.

Using ERR_PTR() include/linux/err.h; to encode a negative error number into a pointer, and IS_ERR() and PTR_ERR() to get it back out again: avoids a separate pointer parameter for the error number. Icky, but in a good way.

Linus and the other developers sometimes change function or structure names in development kernels; this is not done just to keep everyone on their toes: it reflects a fundamental change (eg. can no longer be called with interrupts on, or does extra checks, or doesn't do checks which were caught before). Usually this is accompanied by a fairly complete note to the linux-kernel mailing list; search the archive. Simply doing a global replace on the file usually makes things worse.

The preferred method of initializing structures is to use designated initialisers, as defined by ISO C99, eg:

static struct block_device_operations opt_fops = {
        .open               = opt_open,
        .release            = opt_release,
        .ioctl              = opt_ioctl,
        .check_media_change = opt_media_change,
};
   

This makes it easy to grep for, and makes it clear which structure fields are set. You should do this because it looks cool.

GNU Extensions are explicitly allowed in the Linux kernel. Note that some of the more complex ones are not very well supported, due to lack of general use, but the following are considered standard (see the GCC info page section "C Extensions" for more details - Yes, really the info page, the man page is only a short summary of the stuff in info).

Be wary when using long long in the kernel, the code gcc generates for it is horrible and worse: division and multiplication does not work on i386 because the GCC runtime functions for it are missing from the kernel environment.

Using C++ in the kernel is usually a bad idea, because the kernel does not provide the necessary runtime environment and the include files are not tested for it. It is still possible, but not recommended. If you really want to do this, forget about exceptions at least.

It is generally considered cleaner to use macros in header files (or at the top of .c files) to abstract away functions rather than using `#if' pre-processor statements throughout the source code.