kernel-hacking.tmpl

linux 内核源代码

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook XML V4.1.2//EN"
	"http://www.oasis-open.org/docbook/xml/4.1.2/docbookx.dtd" []>

<book id="lk-hacking-guide">
 <bookinfo>
  <title>Unreliable Guide To Hacking The Linux Kernel</title>
  
  <authorgroup>
   <author>
    <firstname>Rusty</firstname>
    <surname>Russell</surname>
    <affiliation>
     <address>
      <email>[e-mail]</email>
     </address>
    </affiliation>
   </author>
  </authorgroup>

  <copyright>
   <year>2005</year>
   <holder>Rusty Russell</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>

  <releaseinfo>
   This is the first release of this document as part of the kernel tarball.
  </releaseinfo>

 </bookinfo>

 <toc></toc>

 <chapter id="introduction">
  <title>Introduction</title>
  <para>
   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.
  </para>
  <para>
   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.
  </para>
 </chapter>

 <chapter id="basic-players">
  <title>The Players</title>

  <para>
   At any time each of the CPUs in a system can be:
  </para>

  <itemizedlist>
   <listitem>
    <para>
     not associated with any process, serving a hardware interrupt;
    </para>
   </listitem>

   <listitem>
    <para>
     not associated with any process, serving a softirq or tasklet;
    </para>
   </listitem>

   <listitem>
    <para>
     running in kernel space, associated with a process (user context);
    </para>
   </listitem>

   <listitem>
    <para>
     running a process in user space.
    </para>
   </listitem>
  </itemizedlist>

  <para>
   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.
  </para>

  <para>
   We'll see a number of ways that the user context can block
   interrupts, to become truly non-preemptable.
  </para>
  
  <sect1 id="basics-usercontext">
   <title>User Context</title>

   <para>
    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
    <function>schedule()</function>.
   </para>

   <note>
    <para>
     You are always in user context on module load and unload,
     and on operations on the block device layer.
    </para>
   </note>

   <para>
    In user context, the <varname>current</varname> pointer (indicating 
    the task we are currently executing) is valid, and
    <function>in_interrupt()</function>
    (<filename>include/linux/interrupt.h</filename>) is <returnvalue>false
    </returnvalue>.  
   </para>

   <caution>
    <para>
     Beware that if you have preemption or softirqs disabled
     (see below), <function>in_interrupt()</function> will return a 
     false positive.
    </para>
   </caution>
  </sect1>

  <sect1 id="basics-hardirqs">
   <title>Hardware Interrupts (Hard IRQs)</title>

   <para>
    Timer ticks, <hardware>network cards</hardware> and 
    <hardware>keyboard</hardware> 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.
   </para>

   <para>
    You can tell you are in a hardware interrupt, because 
    <function>in_irq()</function> returns <returnvalue>true</returnvalue>.  
   </para>
   <caution>
    <para>
     Beware that this will return a false positive if interrupts are disabled 
     (see below).
    </para>
   </caution>
  </sect1>

  <sect1 id="basics-softirqs">
   <title>Software Interrupt Context: Softirqs and Tasklets</title>

   <para>
    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 (<filename>kernel/softirq.c</filename>).
   </para>

   <para>
    Much of the real interrupt handling work is done here.  Early in
    the transition to <acronym>SMP</acronym>, 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'.
   </para>

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

   <para>
    Softirqs are often a pain to deal with, since the same softirq
    will run simultaneously on more than one CPU.  For this reason,
    tasklets (<filename
    class="headerfile">include/linux/interrupt.h</filename>) 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.
   </para>
   <caution>
    <para>
     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.
    </para>
   </caution>

   <para>
    You can tell you are in a softirq (or tasklet)
    using the <function>in_softirq()</function> macro 
    (<filename class="headerfile">include/linux/interrupt.h</filename>).
   </para>
   <caution>
    <para>
     Beware that this will return a false positive if a bh lock (see below)
     is held.
    </para>
   </caution>
  </sect1>
 </chapter>

 <chapter id="basic-rules">
  <title>Some Basic Rules</title>

  <variablelist>
   <varlistentry>
    <term>No memory protection</term>
    <listitem>
     <para>
      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?
     </para>
    </listitem>
   </varlistentry>

   <varlistentry>
    <term>No floating point or <acronym>MMX</acronym></term>
    <listitem>
     <para>
      The <acronym>FPU</acronym> context is not saved; even in user
      context the <acronym>FPU</acronym> state probably won't
      correspond with the current process: you would mess with some
      user process' <acronym>FPU</acronym> state.  If you really want
      to do this, you would have to explicitly save/restore the full
      <acronym>FPU</acronym> state (and avoid context switches).  It
      is generally a bad idea; use fixed point arithmetic first.
     </para>
    </listitem>
   </varlistentry>

   <varlistentry>
    <term>A rigid stack limit</term>
    <listitem>
     <para>
      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).
     </para>
    </listitem>
   </varlistentry>

   <varlistentry>
    <term>The Linux kernel is portable</term>
    <listitem>
     <para>
      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.
     </para>
    </listitem>
   </varlistentry>
  </variablelist>
 </chapter>

 <chapter id="ioctls">
  <title>ioctls: Not writing a new system call</title>

  <para>
   A system call generally looks like this
  </para>

  <programlisting>
asmlinkage long sys_mycall(int arg)
{
        return 0; 
}
  </programlisting>

  <para>
   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
   <filename class="headerfile">include/asm/unistd.h</filename> and
   <filename>arch/kernel/entry.S</filename> file, and is much more
   likely to be accepted by Linus.
  </para>

  <para>
   If all your routine does is read or write some parameter, consider
   implementing a <function>sysfs</function> interface instead.
  </para>

  <para>
   Inside the ioctl you're in user context to a process.  When a
   error occurs you return a negated errno (see
   <filename class="headerfile">include/linux/errno.h</filename>),
   otherwise you return <returnvalue>0</returnvalue>.
  </para>

  <para>
   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 <constant>-ERESTARTSYS</constant> 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.
  </para>

  <programlisting>
if (signal_pending()) 
        return -ERESTARTSYS;
  </programlisting>

  <para>
   If you're doing longer computations: first think userspace. If you
   <emphasis>really</emphasis> 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:
  </para>

  <programlisting>
cond_resched(); /* Will sleep */ 
  </programlisting>

  <para>
   A short note on interface design: the UNIX system call motto is
   "Provide mechanism not policy".
  </para>
 </chapter>

 <chapter id="deadlock-recipes">
  <title>Recipes for Deadlock</title>

  <para>
   You cannot call any routines which may sleep, unless:
  </para>
  <itemizedlist>
   <listitem>
    <para>
     You are in user context.
    </para>
   </listitem>

   <listitem>
    <para>
     You do not own any spinlocks.
    </para>
   </listitem>

   <listitem>
    <para>
     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).
    </para>
   </listitem>
  </itemizedlist>

  <para>
   Note that some functions may sleep implicitly: common ones are
   the user space access functions (*_user) and memory allocation
   functions without <symbol>GFP_ATOMIC</symbol>.
  </para>

  <para>
   You should always compile your kernel
   <symbol>CONFIG_DEBUG_SPINLOCK_SLEEP</symbol> on, and it will warn
   you if you break these rules.  If you <emphasis>do</emphasis> break
   the rules, you will eventually lock up your box.
  </para>

  <para>
   Really.
  </para>
 </chapter>

 <chapter id="common-routines">
  <title>Common Routines</title>

  <sect1 id="routines-printk">
   <title>
    <function>printk()</function>
    <filename class="headerfile">include/linux/kernel.h</filename>
   </title>

   <para>
    <function>printk()</function> 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:
   </para>

   <programlisting>
printk(KERN_INFO "i = %u\n", i);
   </programlisting>