Rusty Russell

rusty@rustcorp.com.au

2005 Rusty Russell 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. 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. 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 For more details see the file COPYING in the source distribution of Linux. This is the first release of this document as part of the kernel tarball. Welcome, gentle reader, to Rusty's Remarkably Unreliable Guide to Linux Kernel Hacking. This document describes the common routines and general requirements for kernel code: its goal is to serve as a primer for Linux kernel development for experienced C programmers. I avoid implementation details: that's what the code is for, and I ignore whole tracts of useful routines. Before you read this, please understand that I never wanted to write this document, being grossly under-qualified, but I always wanted to read it, and this was the only way. I hope it will grow into a compendium of best practice, common starting points and random information. At any time each of the CPUs in a system can be: not associated with any process, serving a hardware interrupt; not associated with any process, serving a softirq or tasklet; running in kernel space, associated with a process (user context); running a process in user space. There is an ordering between these. The bottom two can preempt each other, but above that is a strict hierarchy: each can only be preempted by the ones above it. For example, while a softirq is running on a CPU, no other softirq will preempt it, but a hardware interrupt can. However, any other CPUs in the system execute independently. We'll see a number of ways that the user context can block interrupts, to become truly non-preemptable. 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(). You are always in user context on module load and unload, and on operations on the block device layer. In user context, the current pointer (indicating the task we are currently executing) is valid, and in_interrupt() (include/linux/interrupt.h) is false . Beware that if you have preemption or softirqs disabled (see below), in_interrupt() will return a false positive. 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. Beware that this will return a false positive if interrupts are disabled (see below). 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. The name 'tasklet' is misleading: they have nothing to do with 'tasks', and probably more to do with some bad vodka Alexey Kuznetsov had at the time. You can tell you are in a softirq (or tasklet) using the in_softirq() macro (include/linux/interrupt.h). Beware that this will return a false positive if a bh lock (see below) is held. No memory protection If you corrupt memory, whether in user context or interrupt context, the whole machine will crash. Are you sure you can't do what you want in userspace? No floating point or MMX The FPU context is not saved; even in user context the FPU state probably won't correspond with the current process: you would mess with some user process' FPU state. If you really want to do this, you would have to explicitly save/restore the full FPU state (and avoid context switches). It is generally a bad idea; use fixed point arithmetic first. A rigid stack limit Depending on configuration options the kernel stack is about 3K to 6K for most 32-bit architectures: it's about 14K on most 64-bit archs, and often shared with interrupts so you can't use it all. Avoid deep recursion and huge local arrays on the stack (allocate them dynamically instead). The Linux kernel is portable Let's keep it that way. Your code should be 64-bit clean, and endian-independent. You should also minimize CPU specific stuff, e.g. inline assembly should be cleanly encapsulated and minimized to ease porting. Generally it should be restricted to the architecture-dependent part of the kernel tree. A system call generally looks like this asmlinkage long sys_mycall(int arg) { return 0; } First, in most cases you don't want to create a new system call. You create a character device and implement an appropriate ioctl for it. This is much more flexible than system calls, doesn't have to be entered in every architecture's include/asm/unistd.h and arch/kernel/entry.S file, and is much more likely to be accepted by Linus. If all your routine does is read or write some parameter, consider implementing a sysfs interface instead. Inside the ioctl you're in user context to a process. When a error occurs you return a negated errno (see include/linux/errno.h), otherwise you return 0. After you slept you should check if a signal occurred: the Unix/Linux way of handling signals is to temporarily exit the system call with the -ERESTARTSYS error. The system call entry code will switch back to user context, process the signal handler and then your system call will be restarted (unless the user disabled that). So you should be prepared to process the restart, e.g. if you're in the middle of manipulating some data structure. if (signal_pending(current)) return -ERESTARTSYS; If you're doing longer computations: first think userspace. If you really want to do it in kernel you should regularly check if you need to give up the CPU (remember there is cooperative multitasking per CPU). Idiom: cond_resched(); /* Will sleep */ A short note on interface design: the UNIX system call motto is "Provide mechanism not policy". You cannot call any routines which may sleep, unless: You are in user context. You do not own any spinlocks. You have interrupts enabled (actually, Andi Kleen says that the scheduling code will enable them for you, but that's probably not what you wanted). Note that some functions may sleep implicitly: common ones are the user space access functions (*_user) and memory allocation functions without GFP_ATOMIC. You should always compile your kernel CONFIG_DEBUG_ATOMIC_SLEEP on, and it will warn you if you break these rules. If you do break the rules, you will eventually lock up your box. Really. 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. You will know when you are a real kernel hacker when you start typoing printf as printk in your user programs :) Another sidenote: the original Unix Version 6 sources had a comment on top of its printf function: "Printf should not be used for chit-chat". You should follow that advice. [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. Unlike put_user() and get_user(), they return the amount of uncopied data (ie. 0 still means success). [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: GFP_KERNEL May sleep and swap to free memory. Only allowed in user context, but is the most reliable way to allocate memory. GFP_ATOMIC Don't sleep. Less reliable than GFP_KERNEL, but may be called from interrupt context. You should really have a good out-of-memory error-handling strategy. GFP_DMA Allocate ISA DMA lower than 16MB. If you don't know what that is you don't need it. Very unreliable. 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. [SLEEPS] A wait queue is used to wait for someone to wake you up when a certain condition is true. They must be used carefully to ensure there is no race condition. You declare a wait_queue_head_t, and then processes which want to wait for that condition declare a wait_queue_t referring to themselves, and place that in the queue. 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. Certain operations are guaranteed atomic on all platforms. The first class of operations work on atomic_t include/asm/atomic.h; this contains a signed integer (at least 32 bits long), and you must use these functions to manipulate or read atomic_t variables. atomic_read() and atomic_set() get and set the counter, atomic_add(), atomic_sub(), atomic_inc(), atomic_dec(), and atomic_dec_and_test() (returns true if it was decremented to zero). Yes. It returns true (i.e. != 0) if the atomic variable is zero. Note that these functions are slower than normal arithmetic, and so should not be used unnecessarily. The second class of atomic operations is atomic bit operations on an unsigned long, defined in include/linux/bitops.h. These operations generally take a pointer to the bit pattern, and a bit number: 0 is the least significant bit. set_bit(), clear_bit() and change_bit() set, clear, and flip the given bit. test_and_set_bit(), test_and_clear_bit() and test_and_change_bit() do the same thing, except return true if the bit was previously set; these are particularly useful for atomically setting flags. It is possible to call these operations with bit indices greater than BITS_PER_LONG. The resulting behavior is strange on big-endian platforms though so it is a good idea not to do this. Within the kernel proper, the normal linking rules apply (ie. unless a symbol is declared to be file scope with the static keyword, it can be used anywhere in the kernel). However, for modules, a special exported symbol table is kept which limits the entry points to the kernel proper. Modules can also export symbols. 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). Inline functions Statement expressions (ie. the ({ and }) constructs). Declaring attributes of a function / variable / type (__attribute__) typeof Zero length arrays Macro varargs Arithmetic on void pointers Non-Constant initializers Assembler Instructions (not outside arch/ and include/asm/) Function names as strings (__func__). __builtin_constant_p() 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. In order to get your stuff into shape for official inclusion, or even to make a neat patch, there's administrative work to be done: Figure out whose pond you've been pissing in. Look at the top of the source files, inside the MAINTAINERS file, and last of all in the CREDITS file. You should coordinate with this person to make sure you're not duplicating effort, or trying something that's already been rejected. Make sure you put your name and EMail address at the top of any files you create or mangle significantly. This is the first place people will look when they find a bug, or when they want to make a change. Usually you want a configuration option for your kernel hack. Edit Kconfig in the appropriate directory. The Config language is simple to use by cut and paste, and there's complete documentation in Documentation/kbuild/kconfig-language.txt. In your description of the option, make sure you address both the expert user and the user who knows nothing about your feature. Mention incompatibilities and issues here. Definitely end your description with if in doubt, say N (or, occasionally, `Y'); this is for people who have no idea what you are talking about. Edit the Makefile: the CONFIG variables are exported here so you can usually just add a "obj-$(CONFIG_xxx) += xxx.o" line. The syntax is documented in Documentation/kbuild/makefiles.txt. Put yourself in CREDITS if you've done something noteworthy, usually beyond a single file (your name should be at the top of the source files anyway). MAINTAINERS means you want to be consulted when changes are made to a subsystem, and hear about bugs; it implies a more-than-passing commitment to some part of the code. Finally, don't forget to read Documentation/SubmittingPatches and possibly Documentation/SubmittingDrivers. Some favorites from browsing the source. Feel free to add to this list. arch/x86/include/asm/delay.h: #define ndelay(n) (__builtin_constant_p(n) ? \ ((n) > 20000 ? __bad_ndelay() : __const_udelay((n) * 5ul)) : \ __ndelay(n)) include/linux/fs.h: /* * Kernel pointers have redundant information, so we can use a * scheme where we can return either an error code or a dentry * pointer with the same return value. * * This should be a per-architecture thing, to allow different * error and pointer decisions. */ #define ERR_PTR(err) ((void *)((long)(err))) #define PTR_ERR(ptr) ((long)(ptr)) #define IS_ERR(ptr) ((unsigned long)(ptr) > (unsigned long)(-1000)) arch/x86/include/asm/uaccess_32.h: #define copy_to_user(to,from,n) \ (__builtin_constant_p(n) ? \ __constant_copy_to_user((to),(from),(n)) : \ __generic_copy_to_user((to),(from),(n))) arch/sparc/kernel/head.S: /* * Sun people can't spell worth damn. "compatability" indeed. * At least we *know* we can't spell, and use a spell-checker. */ /* Uh, actually Linus it is I who cannot spell. Too much murky * Sparc assembly will do this to ya. */ C_LABEL(cputypvar): .asciz "compatibility" /* Tested on SS-5, SS-10. Probably someone at Sun applied a spell-checker. */ .align 4 C_LABEL(cputypvar_sun4m): .asciz "compatible" arch/sparc/lib/checksum.S: /* Sun, you just can't beat me, you just can't. Stop trying, * give up. I'm serious, I am going to kick the living shit * out of you, game over, lights out. */ Thanks to Andi Kleen for the idea, answering my questions, fixing my mistakes, filling content, etc. Philipp Rumpf for more spelling and clarity fixes, and some excellent non-obvious points. Werner Almesberger for giving me a great summary of disable_irq(), and Jes Sorensen and Andrea Arcangeli added caveats. Michael Elizabeth Chastain for checking and adding to the Configure section. Telsa Gwynne for teaching me DocBook.