1 <!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook V3.1//EN"[]>
3 <book id="lk-hacking-guide">
5 <title>Unreliable Guide To Hacking The Linux Kernel</title>
9 <firstname>Paul</firstname>
10 <othername>Rusty</othername>
11 <surname>Russell</surname>
14 <email>rusty@rustcorp.com.au</email>
22 <holder>Rusty Russell</holder>
27 This documentation is free software; you can redistribute
28 it and/or modify it under the terms of the GNU General Public
29 License as published by the Free Software Foundation; either
30 version 2 of the License, or (at your option) any later
35 This program is distributed in the hope that it will be
36 useful, but WITHOUT ANY WARRANTY; without even the implied
37 warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
38 See the GNU General Public License for more details.
42 You should have received a copy of the GNU General Public
43 License along with this program; if not, write to the Free
44 Software Foundation, Inc., 59 Temple Place, Suite 330, Boston,
49 For more details see the file COPYING in the source
50 distribution of Linux.
55 This is the first release of this document as part of the kernel tarball.
62 <chapter id="introduction">
63 <title>Introduction</title>
65 Welcome, gentle reader, to Rusty's Unreliable Guide to Linux
66 Kernel Hacking. This document describes the common routines and
67 general requirements for kernel code: its goal is to serve as a
68 primer for Linux kernel development for experienced C
69 programmers. I avoid implementation details: that's what the
70 code is for, and I ignore whole tracts of useful routines.
73 Before you read this, please understand that I never wanted to
74 write this document, being grossly under-qualified, but I always
75 wanted to read it, and this was the only way. I hope it will
76 grow into a compendium of best practice, common starting points
77 and random information.
81 <chapter id="basic-players">
82 <title>The Players</title>
85 At any time each of the CPUs in a system can be:
91 not associated with any process, serving a hardware interrupt;
97 not associated with any process, serving a softirq, tasklet or bh;
103 running in kernel space, associated with a process;
109 running a process in user space.
115 There is a strict ordering between these: other than the last
116 category (userspace) each can only be pre-empted by those above.
117 For example, while a softirq is running on a CPU, no other
118 softirq will pre-empt it, but a hardware interrupt can. However,
119 any other CPUs in the system execute independently.
123 We'll see a number of ways that the user context can block
124 interrupts, to become truly non-preemptable.
127 <sect1 id="basics-usercontext">
128 <title>User Context</title>
131 User context is when you are coming in from a system call or
132 other trap: you can sleep, and you own the CPU (except for
133 interrupts) until you call <function>schedule()</function>.
134 In other words, user context (unlike userspace) is not pre-emptable.
139 You are always in user context on module load and unload,
140 and on operations on the block device layer.
145 In user context, the <varname>current</varname> pointer (indicating
146 the task we are currently executing) is valid, and
147 <function>in_interrupt()</function>
148 (<filename>include/asm/hardirq.h</filename>) is <returnvalue>false
154 Beware that if you have interrupts or bottom halves disabled
155 (see below), <function>in_interrupt()</function> will return a
161 <sect1 id="basics-hardirqs">
162 <title>Hardware Interrupts (Hard IRQs)</title>
165 Timer ticks, <hardware>network cards</hardware> and
166 <hardware>keyboard</hardware> are examples of real
167 hardware which produce interrupts at any time. The kernel runs
168 interrupt handlers, which services the hardware. The kernel
169 guarantees that this handler is never re-entered: if another
170 interrupt arrives, it is queued (or dropped). Because it
171 disables interrupts, this handler has to be fast: frequently it
172 simply acknowledges the interrupt, marks a `software interrupt'
173 for execution and exits.
177 You can tell you are in a hardware interrupt, because
178 <function>in_irq()</function> returns <returnvalue>true</returnvalue>.
182 Beware that this will return a false positive if interrupts are disabled
188 <sect1 id="basics-softirqs">
189 <title>Software Interrupt Context: Bottom Halves, Tasklets, softirqs</title>
192 Whenever a system call is about to return to userspace, or a
193 hardware interrupt handler exits, any `software interrupts'
194 which are marked pending (usually by hardware interrupts) are
195 run (<filename>kernel/softirq.c</filename>).
199 Much of the real interrupt handling work is done here. Early in
200 the transition to <acronym>SMP</acronym>, there were only `bottom
201 halves' (BHs), which didn't take advantage of multiple CPUs. Shortly
202 after we switched from wind-up computers made of match-sticks and snot,
203 we abandoned this limitation.
207 <filename class="headerfile">include/linux/interrupt.h</filename> lists the
208 different BH's. No matter how many CPUs you have, no two BHs will run at
209 the same time. This made the transition to SMP simpler, but sucks hard for
210 scalable performance. A very important bottom half is the timer
211 BH (<filename class="headerfile">include/linux/timer.h</filename>): you
212 can register to have it call functions for you in a given length of time.
216 2.3.43 introduced softirqs, and re-implemented the (now
217 deprecated) BHs underneath them. Softirqs are fully-SMP
218 versions of BHs: they can run on as many CPUs at once as
219 required. This means they need to deal with any races in shared
220 data using their own locks. A bitmask is used to keep track of
221 which are enabled, so the 32 available softirqs should not be
222 used up lightly. (<emphasis>Yes</emphasis>, people will
227 tasklets (<filename class="headerfile">include/linux/interrupt.h</filename>)
228 are like softirqs, except they are dynamically-registrable (meaning you
229 can have as many as you want), and they also guarantee that any tasklet
230 will only run on one CPU at any time, although different tasklets can
231 run simultaneously (unlike different BHs).
235 The name `tasklet' is misleading: they have nothing to do with `tasks',
236 and probably more to do with some bad vodka Alexey Kuznetsov had at the
242 You can tell you are in a softirq (or bottom half, or tasklet)
243 using the <function>in_softirq()</function> macro
244 (<filename class="headerfile">include/asm/hardirq.h</filename>).
248 Beware that this will return a false positive if a bh lock (see below)
255 <chapter id="basic-rules">
256 <title>Some Basic Rules</title>
260 <term>No memory protection</term>
263 If you corrupt memory, whether in user context or
264 interrupt context, the whole machine will crash. Are you
265 sure you can't do what you want in userspace?
271 <term>No floating point or <acronym>MMX</acronym></term>
274 The <acronym>FPU</acronym> context is not saved; even in user
275 context the <acronym>FPU</acronym> state probably won't
276 correspond with the current process: you would mess with some
277 user process' <acronym>FPU</acronym> state. If you really want
278 to do this, you would have to explicitly save/restore the full
279 <acronym>FPU</acronym> state (and avoid context switches). It
280 is generally a bad idea; use fixed point arithmetic first.
286 <term>A rigid stack limit</term>
289 The kernel stack is about 6K in 2.2 (for most
290 architectures: it's about 14K on the Alpha), and shared
291 with interrupts so you can't use it all. Avoid deep
292 recursion and huge local arrays on the stack (allocate
293 them dynamically instead).
299 <term>The Linux kernel is portable</term>
302 Let's keep it that way. Your code should be 64-bit clean,
303 and endian-independent. You should also minimize CPU
304 specific stuff, e.g. inline assembly should be cleanly
305 encapsulated and minimized to ease porting. Generally it
306 should be restricted to the architecture-dependent part of
314 <chapter id="ioctls">
315 <title>ioctls: Not writing a new system call</title>
318 A system call generally looks like this
322 asmlinkage long sys_mycall(int arg)
329 First, in most cases you don't want to create a new system call.
330 You create a character device and implement an appropriate ioctl
331 for it. This is much more flexible than system calls, doesn't have
332 to be entered in every architecture's
333 <filename class="headerfile">include/asm/unistd.h</filename> and
334 <filename>arch/kernel/entry.S</filename> file, and is much more
335 likely to be accepted by Linus.
339 If all your routine does is read or write some parameter, consider
340 implementing a <function>sysctl</function> interface instead.
344 Inside the ioctl you're in user context to a process. When a
345 error occurs you return a negated errno (see
346 <filename class="headerfile">include/linux/errno.h</filename>),
347 otherwise you return <returnvalue>0</returnvalue>.
351 After you slept you should check if a signal occurred: the
352 Unix/Linux way of handling signals is to temporarily exit the
353 system call with the <constant>-ERESTARTSYS</constant> error. The
354 system call entry code will switch back to user context, process
355 the signal handler and then your system call will be restarted
356 (unless the user disabled that). So you should be prepared to
357 process the restart, e.g. if you're in the middle of manipulating
362 if (signal_pending())
367 If you're doing longer computations: first think userspace. If you
368 <emphasis>really</emphasis> want to do it in kernel you should
369 regularly check if you need to give up the CPU (remember there is
370 cooperative multitasking per CPU). Idiom:
374 cond_resched(); /* Will sleep */
378 A short note on interface design: the UNIX system call motto is
379 "Provide mechanism not policy".
383 <chapter id="deadlock-recipes">
384 <title>Recipes for Deadlock</title>
387 You cannot call any routines which may sleep, unless:
392 You are in user context.
398 You do not own any spinlocks.
404 You have interrupts enabled (actually, Andi Kleen says
405 that the scheduling code will enable them for you, but
406 that's probably not what you wanted).
412 Note that some functions may sleep implicitly: common ones are
413 the user space access functions (*_user) and memory allocation
414 functions without <symbol>GFP_ATOMIC</symbol>.
418 You will eventually lock up your box if you break these rules.
426 <chapter id="common-routines">
427 <title>Common Routines</title>
429 <sect1 id="routines-printk">
431 <function>printk()</function>
432 <filename class="headerfile">include/linux/kernel.h</filename>
436 <function>printk()</function> feeds kernel messages to the
437 console, dmesg, and the syslog daemon. It is useful for debugging
438 and reporting errors, and can be used inside interrupt context,
439 but use with caution: a machine which has its console flooded with
440 printk messages is unusable. It uses a format string mostly
441 compatible with ANSI C printf, and C string concatenation to give
442 it a first "priority" argument:
446 printk(KERN_INFO "i = %u\n", i);
450 See <filename class="headerfile">include/linux/kernel.h</filename>;
451 for other KERN_ values; these are interpreted by syslog as the
452 level. Special case: for printing an IP address use
457 printk(KERN_INFO "my ip: %d.%d.%d.%d\n", NIPQUAD(ipaddress));
461 <function>printk()</function> internally uses a 1K buffer and does
462 not catch overruns. Make sure that will be enough.
467 You will know when you are a real kernel hacker
468 when you start typoing printf as printk in your user programs :)
472 <!--- From the Lions book reader department -->
476 Another sidenote: the original Unix Version 6 sources had a
477 comment on top of its printf function: "Printf should not be
478 used for chit-chat". You should follow that advice.
483 <sect1 id="routines-copy">
485 <function>copy_[to/from]_user()</function>
487 <function>get_user()</function>
489 <function>put_user()</function>
490 <filename class="headerfile">include/asm/uaccess.h</filename>
494 <emphasis>[SLEEPS]</emphasis>
498 <function>put_user()</function> and <function>get_user()</function>
499 are used to get and put single values (such as an int, char, or
500 long) from and to userspace. A pointer into userspace should
501 never be simply dereferenced: data should be copied using these
502 routines. Both return <constant>-EFAULT</constant> or 0.
505 <function>copy_to_user()</function> and
506 <function>copy_from_user()</function> are more general: they copy
507 an arbitrary amount of data to and from userspace.
510 Unlike <function>put_user()</function> and
511 <function>get_user()</function>, they return the amount of
512 uncopied data (ie. <returnvalue>0</returnvalue> still means
516 [Yes, this moronic interface makes me cringe. Please submit a
517 patch and become my hero --RR.]
520 The functions may sleep implicitly. This should never be called
521 outside user context (it makes no sense), with interrupts
522 disabled, or a spinlock held.
526 <sect1 id="routines-kmalloc">
527 <title><function>kmalloc()</function>/<function>kfree()</function>
528 <filename class="headerfile">include/linux/slab.h</filename></title>
531 <emphasis>[MAY SLEEP: SEE BELOW]</emphasis>
535 These routines are used to dynamically request pointer-aligned
536 chunks of memory, like malloc and free do in userspace, but
537 <function>kmalloc()</function> takes an extra flag word.
550 May sleep and swap to free memory. Only allowed in user
551 context, but is the most reliable way to allocate memory.
564 Don't sleep. Less reliable than <constant>GFP_KERNEL</constant>,
565 but may be called from interrupt context. You should
566 <emphasis>really</emphasis> have a good out-of-memory
567 error-handling strategy.
580 Allocate ISA DMA lower than 16MB. If you don't know what that
581 is you don't need it. Very unreliable.
588 If you see a <errorname>kmem_grow: Called nonatomically from int
589 </errorname> warning message you called a memory allocation function
590 from interrupt context without <constant>GFP_ATOMIC</constant>.
591 You should really fix that. Run, don't walk.
595 If you are allocating at least <constant>PAGE_SIZE</constant>
596 (<filename class="headerfile">include/asm/page.h</filename>) bytes,
597 consider using <function>__get_free_pages()</function>
599 (<filename class="headerfile">include/linux/mm.h</filename>). It
600 takes an order argument (0 for page sized, 1 for double page, 2
601 for four pages etc.) and the same memory priority flag word as
606 If you are allocating more than a page worth of bytes you can use
607 <function>vmalloc()</function>. It'll allocate virtual memory in
608 the kernel map. This block is not contiguous in physical memory,
609 but the <acronym>MMU</acronym> makes it look like it is for you
610 (so it'll only look contiguous to the CPUs, not to external device
611 drivers). If you really need large physically contiguous memory
612 for some weird device, you have a problem: it is poorly supported
613 in Linux because after some time memory fragmentation in a running
614 kernel makes it hard. The best way is to allocate the block early
615 in the boot process via the <function>alloc_bootmem()</function>
620 Before inventing your own cache of often-used objects consider
621 using a slab cache in
622 <filename class="headerfile">include/linux/slab.h</filename>
626 <sect1 id="routines-current">
627 <title><function>current</function>
628 <filename class="headerfile">include/asm/current.h</filename></title>
631 This global variable (really a macro) contains a pointer to
632 the current task structure, so is only valid in user context.
633 For example, when a process makes a system call, this will
634 point to the task structure of the calling process. It is
635 <emphasis>not NULL</emphasis> in interrupt context.
639 <sect1 id="routines-udelay">
640 <title><function>udelay()</function>/<function>mdelay()</function>
641 <filename class="headerfile">include/asm/delay.h</filename>
642 <filename class="headerfile">include/linux/delay.h</filename>
646 The <function>udelay()</function> function can be used for small pauses.
647 Do not use large values with <function>udelay()</function> as you risk
648 overflow - the helper function <function>mdelay()</function> is useful
649 here, or even consider <function>schedule_timeout()</function>.
653 <sect1 id="routines-endian">
654 <title><function>cpu_to_be32()</function>/<function>be32_to_cpu()</function>/<function>cpu_to_le32()</function>/<function>le32_to_cpu()</function>
655 <filename class="headerfile">include/asm/byteorder.h</filename>
659 The <function>cpu_to_be32()</function> family (where the "32" can
660 be replaced by 64 or 16, and the "be" can be replaced by "le") are
661 the general way to do endian conversions in the kernel: they
662 return the converted value. All variations supply the reverse as
663 well: <function>be32_to_cpu()</function>, etc.
667 There are two major variations of these functions: the pointer
668 variation, such as <function>cpu_to_be32p()</function>, which take
669 a pointer to the given type, and return the converted value. The
670 other variation is the "in-situ" family, such as
671 <function>cpu_to_be32s()</function>, which convert value referred
672 to by the pointer, and return void.
676 <sect1 id="routines-local-irqs">
677 <title><function>local_irq_save()</function>/<function>local_irq_restore()</function>
678 <filename class="headerfile">include/asm/system.h</filename>
682 These routines disable hard interrupts on the local CPU, and
683 restore them. They are reentrant; saving the previous state in
684 their one <varname>unsigned long flags</varname> argument. If you
685 know that interrupts are enabled, you can simply use
686 <function>local_irq_disable()</function> and
687 <function>local_irq_enable()</function>.
691 <sect1 id="routines-softirqs">
692 <title><function>local_bh_disable()</function>/<function>local_bh_enable()</function>
693 <filename class="headerfile">include/linux/interrupt.h</filename></title>
696 These routines disable soft interrupts on the local CPU, and
697 restore them. They are reentrant; if soft interrupts were
698 disabled before, they will still be disabled after this pair
699 of functions has been called. They prevent softirqs, tasklets
700 and bottom halves from running on the current CPU.
704 <sect1 id="routines-processorids">
705 <title><function>smp_processor_id</function>()
706 <filename class="headerfile">include/asm/smp.h</filename></title>
709 <function>smp_processor_id()</function> returns the current
710 processor number, between 0 and <symbol>NR_CPUS</symbol> (the
711 maximum number of CPUs supported by Linux, currently 32). These
712 values are not necessarily continuous.
716 <sect1 id="routines-init">
717 <title><type>__init</type>/<type>__exit</type>/<type>__initdata</type>
718 <filename class="headerfile">include/linux/init.h</filename></title>
721 After boot, the kernel frees up a special section; functions
722 marked with <type>__init</type> and data structures marked with
723 <type>__initdata</type> are dropped after boot is complete (within
724 modules this directive is currently ignored). <type>__exit</type>
725 is used to declare a function which is only required on exit: the
726 function will be dropped if this file is not compiled as a module.
727 See the header file for use. Note that it makes no sense for a function
728 marked with <type>__init</type> to be exported to modules with
729 <function>EXPORT_SYMBOL()</function> - this will break.
732 Static data structures marked as <type>__initdata</type> must be initialised
733 (as opposed to ordinary static data which is zeroed BSS) and cannot be
739 <sect1 id="routines-init-again">
740 <title><function>__initcall()</function>/<function>module_init()</function>
741 <filename class="headerfile">include/linux/init.h</filename></title>
743 Many parts of the kernel are well served as a module
744 (dynamically-loadable parts of the kernel). Using the
745 <function>module_init()</function> and
746 <function>module_exit()</function> macros it is easy to write code
747 without #ifdefs which can operate both as a module or built into
752 The <function>module_init()</function> macro defines which
753 function is to be called at module insertion time (if the file is
754 compiled as a module), or at boot time: if the file is not
755 compiled as a module the <function>module_init()</function> macro
756 becomes equivalent to <function>__initcall()</function>, which
757 through linker magic ensures that the function is called on boot.
761 The function can return a negative error number to cause
762 module loading to fail (unfortunately, this has no effect if
763 the module is compiled into the kernel). For modules, this is
764 called in user context, with interrupts enabled, and the
765 kernel lock held, so it can sleep.
769 <sect1 id="routines-moduleexit">
770 <title> <function>module_exit()</function>
771 <filename class="headerfile">include/linux/init.h</filename> </title>
774 This macro defines the function to be called at module removal
775 time (or never, in the case of the file compiled into the
776 kernel). It will only be called if the module usage count has
777 reached zero. This function can also sleep, but cannot fail:
778 everything must be cleaned up by the time it returns.
782 <sect1 id="routines-module-use-counters">
783 <title> <function>MOD_INC_USE_COUNT</function>/<function>MOD_DEC_USE_COUNT</function>
784 <filename class="headerfile">include/linux/module.h</filename></title>
787 These manipulate the module usage count, to protect against
788 removal (a module also can't be removed if another module uses
789 one of its exported symbols: see below). Every reference to
790 the module from user context should be reflected by this
791 counter (e.g. for every data structure or socket) before the
792 function sleeps. To quote Tim Waugh:
802 sleep.. (might get unloaded here)
824 You can often avoid having to deal with these problems by using the
825 <structfield>owner</structfield> field of the
826 <structname>file_operations</structname> structure. Set this field
827 as the macro <symbol>THIS_MODULE</symbol>.
831 For more complicated module unload locking requirements, you can set the
832 <structfield>can_unload</structfield> function pointer to your own routine,
833 which should return <returnvalue>0</returnvalue> if the module is
834 unloadable, or <returnvalue>-EBUSY</returnvalue> otherwise.
840 <chapter id="queues">
842 <filename class="headerfile">include/linux/wait.h</filename>
845 <emphasis>[SLEEPS]</emphasis>
849 A wait queue is used to wait for someone to wake you up when a
850 certain condition is true. They must be used carefully to ensure
851 there is no race condition. You declare a
852 <type>wait_queue_head_t</type>, and then processes which want to
853 wait for that condition declare a <type>wait_queue_t</type>
854 referring to themselves, and place that in the queue.
857 <sect1 id="queue-declaring">
858 <title>Declaring</title>
861 You declare a <type>wait_queue_head_t</type> using the
862 <function>DECLARE_WAIT_QUEUE_HEAD()</function> macro, or using the
863 <function>init_waitqueue_head()</function> routine in your
868 <sect1 id="queue-waitqueue">
869 <title>Queuing</title>
872 Placing yourself in the waitqueue is fairly complex, because you
873 must put yourself in the queue before checking the condition.
874 There is a macro to do this:
875 <function>wait_event_interruptible()</function>
877 <filename class="headerfile">include/linux/sched.h</filename> The
878 first argument is the wait queue head, and the second is an
879 expression which is evaluated; the macro returns
880 <returnvalue>0</returnvalue> when this expression is true, or
881 <returnvalue>-ERESTARTSYS</returnvalue> if a signal is received.
882 The <function>wait_event()</function> version ignores signals.
885 Do not use the <function>sleep_on()</function> function family -
886 it is very easy to accidentally introduce races; almost certainly
887 one of the <function>wait_event()</function> family will do, or a
888 loop around <function>schedule_timeout()</function>. If you choose
889 to loop around <function>schedule_timeout()</function> remember
890 you must set the task state (with
891 <function>set_current_state()</function>) on each iteration to avoid
897 <sect1 id="queue-waking">
898 <title>Waking Up Queued Tasks</title>
901 Call <function>wake_up()</function>
903 <filename class="headerfile">include/linux/sched.h</filename>;,
904 which will wake up every process in the queue. The exception is
905 if one has <constant>TASK_EXCLUSIVE</constant> set, in which case
906 the remainder of the queue will not be woken.
911 <chapter id="atomic-ops">
912 <title>Atomic Operations</title>
915 Certain operations are guaranteed atomic on all platforms. The
916 first class of operations work on <type>atomic_t</type>
918 <filename class="headerfile">include/asm/atomic.h</filename>; this
919 contains a signed integer (at least 24 bits long), and you must use
920 these functions to manipulate or read atomic_t variables.
921 <function>atomic_read()</function> and
922 <function>atomic_set()</function> get and set the counter,
923 <function>atomic_add()</function>,
924 <function>atomic_sub()</function>,
925 <function>atomic_inc()</function>,
926 <function>atomic_dec()</function>, and
927 <function>atomic_dec_and_test()</function> (returns
928 <returnvalue>true</returnvalue> if it was decremented to zero).
932 Yes. It returns <returnvalue>true</returnvalue> (i.e. != 0) if the
933 atomic variable is zero.
937 Note that these functions are slower than normal arithmetic, and
938 so should not be used unnecessarily. On some platforms they
939 are much slower, like 32-bit Sparc where they use a spinlock.
943 The second class of atomic operations is atomic bit operations on a
944 <type>long</type>, defined in
946 <filename class="headerfile">include/asm/bitops.h</filename>. These
947 operations generally take a pointer to the bit pattern, and a bit
948 number: 0 is the least significant bit.
949 <function>set_bit()</function>, <function>clear_bit()</function>
950 and <function>change_bit()</function> set, clear, and flip the
951 given bit. <function>test_and_set_bit()</function>,
952 <function>test_and_clear_bit()</function> and
953 <function>test_and_change_bit()</function> do the same thing,
954 except return true if the bit was previously set; these are
955 particularly useful for very simple locking.
959 It is possible to call these operations with bit indices greater
960 than BITS_PER_LONG. The resulting behavior is strange on big-endian
961 platforms though so it is a good idea not to do this.
965 Note that the order of bits depends on the architecture, and in
966 particular, the bitfield passed to these operations must be at
967 least as large as a <type>long</type>.
971 <chapter id="symbols">
972 <title>Symbols</title>
975 Within the kernel proper, the normal linking rules apply
976 (ie. unless a symbol is declared to be file scope with the
977 <type>static</type> keyword, it can be used anywhere in the
978 kernel). However, for modules, a special exported symbol table is
979 kept which limits the entry points to the kernel proper. Modules
980 can also export symbols.
983 <sect1 id="sym-exportsymbols">
984 <title><function>EXPORT_SYMBOL()</function>
985 <filename class="headerfile">include/linux/module.h</filename></title>
988 This is the classic method of exporting a symbol, and it works
989 for both modules and non-modules. In the kernel all these
990 declarations are often bundled into a single file to help
991 genksyms (which searches source files for these declarations).
992 See the comment on genksyms and Makefiles below.
996 <sect1 id="sym-exportsymbols-gpl">
997 <title><function>EXPORT_SYMBOL_GPL()</function>
998 <filename class="headerfile">include/linux/module.h</filename></title>
1001 Similar to <function>EXPORT_SYMBOL()</function> except that the
1002 symbols exported by <function>EXPORT_SYMBOL_GPL()</function> can
1003 only be seen by modules with a
1004 <function>MODULE_LICENSE()</function> that specifies a GPL
1010 <chapter id="conventions">
1011 <title>Routines and Conventions</title>
1013 <sect1 id="conventions-doublelinkedlist">
1014 <title>Double-linked lists
1015 <filename class="headerfile">include/linux/list.h</filename></title>
1018 There are three sets of linked-list routines in the kernel
1019 headers, but this one seems to be winning out (and Linus has
1020 used it). If you don't have some particular pressing need for
1021 a single list, it's a good choice. In fact, I don't care
1022 whether it's a good choice or not, just use it so we can get
1027 <sect1 id="convention-returns">
1028 <title>Return Conventions</title>
1031 For code called in user context, it's very common to defy C
1032 convention, and return <returnvalue>0</returnvalue> for success,
1033 and a negative error number
1034 (eg. <returnvalue>-EFAULT</returnvalue>) for failure. This can be
1035 unintuitive at first, but it's fairly widespread in the networking
1040 The filesystem code uses <function>ERR_PTR()</function>
1042 <filename class="headerfile">include/linux/fs.h</filename>; to
1043 encode a negative error number into a pointer, and
1044 <function>IS_ERR()</function> and <function>PTR_ERR()</function>
1045 to get it back out again: avoids a separate pointer parameter for
1046 the error number. Icky, but in a good way.
1050 <sect1 id="conventions-borkedcompile">
1051 <title>Breaking Compilation</title>
1054 Linus and the other developers sometimes change function or
1055 structure names in development kernels; this is not done just to
1056 keep everyone on their toes: it reflects a fundamental change
1057 (eg. can no longer be called with interrupts on, or does extra
1058 checks, or doesn't do checks which were caught before). Usually
1059 this is accompanied by a fairly complete note to the linux-kernel
1060 mailing list; search the archive. Simply doing a global replace
1061 on the file usually makes things <emphasis>worse</emphasis>.
1065 <sect1 id="conventions-initialising">
1066 <title>Initializing structure members</title>
1069 The preferred method of initializing structures is to use
1070 designated initialisers, as defined by ISO C99, eg:
1073 static struct block_device_operations opt_fops = {
1075 .release = opt_release,
1077 .check_media_change = opt_media_change,
1081 This makes it easy to grep for, and makes it clear which
1082 structure fields are set. You should do this because it looks
1087 <sect1 id="conventions-gnu-extns">
1088 <title>GNU Extensions</title>
1091 GNU Extensions are explicitly allowed in the Linux kernel.
1092 Note that some of the more complex ones are not very well
1093 supported, due to lack of general use, but the following are
1094 considered standard (see the GCC info page section "C
1095 Extensions" for more details - Yes, really the info page, the
1096 man page is only a short summary of the stuff in info):
1106 Statement expressions (ie. the ({ and }) constructs).
1111 Declaring attributes of a function / variable / type
1132 Arithmetic on void pointers
1137 Non-Constant initializers
1142 Assembler Instructions (not outside arch/ and include/asm/)
1147 Function names as strings (__FUNCTION__)
1152 __builtin_constant_p()
1158 Be wary when using long long in the kernel, the code gcc generates for
1159 it is horrible and worse: division and multiplication does not work
1160 on i386 because the GCC runtime functions for it are missing from
1161 the kernel environment.
1164 <!-- FIXME: add a note about ANSI aliasing cleanness -->
1167 <sect1 id="conventions-cplusplus">
1171 Using C++ in the kernel is usually a bad idea, because the
1172 kernel does not provide the necessary runtime environment
1173 and the include files are not tested for it. It is still
1174 possible, but not recommended. If you really want to do
1175 this, forget about exceptions at least.
1179 <sect1 id="conventions-ifdef">
1180 <title>#if</title>
1183 It is generally considered cleaner to use macros in header files
1184 (or at the top of .c files) to abstract away functions rather than
1185 using `#if' pre-processor statements throughout the source code.
1190 <chapter id="submitting">
1191 <title>Putting Your Stuff in the Kernel</title>
1194 In order to get your stuff into shape for official inclusion, or
1195 even to make a neat patch, there's administrative work to be
1201 Figure out whose pond you've been pissing in. Look at the top of
1202 the source files, inside the <filename>MAINTAINERS</filename>
1203 file, and last of all in the <filename>CREDITS</filename> file.
1204 You should coordinate with this person to make sure you're not
1205 duplicating effort, or trying something that's already been
1210 Make sure you put your name and EMail address at the top of
1211 any files you create or mangle significantly. This is the
1212 first place people will look when they find a bug, or when
1213 <emphasis>they</emphasis> want to make a change.
1219 Usually you want a configuration option for your kernel hack.
1220 Edit <filename>Config.in</filename> in the appropriate directory
1221 (but under <filename>arch/</filename> it's called
1222 <filename>config.in</filename>). The Config Language used is not
1223 bash, even though it looks like bash; the safe way is to use only
1224 the constructs that you already see in
1225 <filename>Config.in</filename> files (see
1226 <filename>Documentation/kbuild/kconfig-language.txt</filename>).
1227 It's good to run "make xconfig" at least once to test (because
1228 it's the only one with a static parser).
1232 Variables which can be Y or N use <type>bool</type> followed by a
1233 tagline and the config define name (which must start with
1234 CONFIG_). The <type>tristate</type> function is the same, but
1235 allows the answer M (which defines
1236 <symbol>CONFIG_foo_MODULE</symbol> in your source, instead of
1237 <symbol>CONFIG_FOO</symbol>) if <symbol>CONFIG_MODULES</symbol>
1242 You may well want to make your CONFIG option only visible if
1243 <symbol>CONFIG_EXPERIMENTAL</symbol> is enabled: this serves as a
1244 warning to users. There many other fancy things you can do: see
1245 the various <filename>Config.in</filename> files for ideas.
1251 Edit the <filename>Makefile</filename>: the CONFIG variables are
1252 exported here so you can conditionalize compilation with `ifeq'.
1253 If your file exports symbols then add the names to
1254 <varname>export-objs</varname> so that genksyms will find them.
1257 There is a restriction on the kernel build system that objects
1258 which export symbols must have globally unique names.
1259 If your object does not have a globally unique name then the
1260 standard fix is to move the
1261 <function>EXPORT_SYMBOL()</function> statements to their own
1262 object with a unique name.
1263 This is why several systems have separate exporting objects,
1264 usually suffixed with ksyms.
1272 Document your option in Documentation/Configure.help. Mention
1273 incompatibilities and issues here. <emphasis> Definitely
1274 </emphasis> end your description with <quote> if in doubt, say N
1275 </quote> (or, occasionally, `Y'); this is for people who have no
1276 idea what you are talking about.
1282 Put yourself in <filename>CREDITS</filename> if you've done
1283 something noteworthy, usually beyond a single file (your name
1284 should be at the top of the source files anyway).
1285 <filename>MAINTAINERS</filename> means you want to be consulted
1286 when changes are made to a subsystem, and hear about bugs; it
1287 implies a more-than-passing commitment to some part of the code.
1293 Finally, don't forget to read <filename>Documentation/SubmittingPatches</filename>
1294 and possibly <filename>Documentation/SubmittingDrivers</filename>.
1300 <chapter id="cantrips">
1301 <title>Kernel Cantrips</title>
1304 Some favorites from browsing the source. Feel free to add to this
1309 <filename>include/linux/brlock.h:</filename>
1312 extern inline void br_read_lock (enum brlock_indices idx)
1315 * This causes a link-time bug message if an
1316 * invalid index is used:
1318 if (idx >= __BR_END)
1319 __br_lock_usage_bug();
1321 read_lock(&__brlock_array[smp_processor_id()][idx]);
1326 <filename>include/linux/fs.h</filename>:
1330 * Kernel pointers have redundant information, so we can use a
1331 * scheme where we can return either an error code or a dentry
1332 * pointer with the same return value.
1334 * This should be a per-architecture thing, to allow different
1335 * error and pointer decisions.
1337 #define ERR_PTR(err) ((void *)((long)(err)))
1338 #define PTR_ERR(ptr) ((long)(ptr))
1339 #define IS_ERR(ptr) ((unsigned long)(ptr) > (unsigned long)(-1000))
1343 <filename>include/asm-i386/uaccess.h:</filename>
1347 #define copy_to_user(to,from,n) \
1348 (__builtin_constant_p(n) ? \
1349 __constant_copy_to_user((to),(from),(n)) : \
1350 __generic_copy_to_user((to),(from),(n)))
1354 <filename>arch/sparc/kernel/head.S:</filename>
1359 * Sun people can't spell worth damn. "compatability" indeed.
1360 * At least we *know* we can't spell, and use a spell-checker.
1363 /* Uh, actually Linus it is I who cannot spell. Too much murky
1364 * Sparc assembly will do this to ya.
1367 .asciz "compatability"
1369 /* Tested on SS-5, SS-10. Probably someone at Sun applied a spell-checker. */
1371 C_LABEL(cputypvar_sun4m):
1376 <filename>arch/sparc/lib/checksum.S:</filename>
1380 /* Sun, you just can't beat me, you just can't. Stop trying,
1381 * give up. I'm serious, I am going to kick the living shit
1382 * out of you, game over, lights out.
1387 <chapter id="credits">
1388 <title>Thanks</title>
1391 Thanks to Andi Kleen for the idea, answering my questions, fixing
1392 my mistakes, filling content, etc. Philipp Rumpf for more spelling
1393 and clarity fixes, and some excellent non-obvious points. Werner
1394 Almesberger for giving me a great summary of
1395 <function>disable_irq()</function>, and Jes Sorensen and Andrea
1396 Arcangeli added caveats. Michael Elizabeth Chastain for checking
1397 and adding to the Configure section. <!-- Rusty insisted on this
1398 bit; I didn't do it! --> Telsa Gwynne for teaching me DocBook.