linux/Documentation/DocBook/kernel-locking.tmpl
<<
>>
Prefs
   1<?xml version="1.0" encoding="UTF-8"?>
   2<!DOCTYPE book PUBLIC "-//OASIS//DTD DocBook XML V4.1.2//EN"
   3        "http://www.oasis-open.org/docbook/xml/4.1.2/docbookx.dtd" []>
   4
   5<book id="LKLockingGuide">
   6 <bookinfo>
   7  <title>Unreliable Guide To Locking</title>
   8  
   9  <authorgroup>
  10   <author>
  11    <firstname>Rusty</firstname>
  12    <surname>Russell</surname>
  13    <affiliation>
  14     <address>
  15      <email>rusty@rustcorp.com.au</email>
  16     </address>
  17    </affiliation>
  18   </author>
  19  </authorgroup>
  20
  21  <copyright>
  22   <year>2003</year>
  23   <holder>Rusty Russell</holder>
  24  </copyright>
  25
  26  <legalnotice>
  27   <para>
  28     This documentation is free software; you can redistribute
  29     it and/or modify it under the terms of the GNU General Public
  30     License as published by the Free Software Foundation; either
  31     version 2 of the License, or (at your option) any later
  32     version.
  33   </para>
  34      
  35   <para>
  36     This program is distributed in the hope that it will be
  37     useful, but WITHOUT ANY WARRANTY; without even the implied
  38     warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  39     See the GNU General Public License for more details.
  40   </para>
  41      
  42   <para>
  43     You should have received a copy of the GNU General Public
  44     License along with this program; if not, write to the Free
  45     Software Foundation, Inc., 59 Temple Place, Suite 330, Boston,
  46     MA 02111-1307 USA
  47   </para>
  48      
  49   <para>
  50     For more details see the file COPYING in the source
  51     distribution of Linux.
  52   </para>
  53  </legalnotice>
  54 </bookinfo>
  55
  56 <toc></toc>
  57  <chapter id="intro">
  58   <title>Introduction</title>
  59   <para>
  60     Welcome, to Rusty's Remarkably Unreliable Guide to Kernel
  61     Locking issues.  This document describes the locking systems in
  62     the Linux Kernel in 2.6.
  63   </para>
  64   <para>
  65     With the wide availability of HyperThreading, and <firstterm
  66     linkend="gloss-preemption">preemption </firstterm> in the Linux
  67     Kernel, everyone hacking on the kernel needs to know the
  68     fundamentals of concurrency and locking for
  69     <firstterm linkend="gloss-smp"><acronym>SMP</acronym></firstterm>.
  70   </para>
  71  </chapter>
  72
  73   <chapter id="races">
  74    <title>The Problem With Concurrency</title>
  75    <para>
  76      (Skip this if you know what a Race Condition is).
  77    </para>
  78    <para>
  79      In a normal program, you can increment a counter like so:
  80    </para>
  81    <programlisting>
  82      very_important_count++;
  83    </programlisting>
  84
  85    <para>
  86      This is what they would expect to happen:
  87    </para>
  88
  89    <table>
  90     <title>Expected Results</title>
  91
  92     <tgroup cols="2" align="left">
  93
  94      <thead>
  95       <row>
  96        <entry>Instance 1</entry>
  97        <entry>Instance 2</entry>
  98       </row>
  99      </thead>
 100
 101      <tbody>
 102       <row>
 103        <entry>read very_important_count (5)</entry>
 104        <entry></entry>
 105       </row>
 106       <row>
 107        <entry>add 1 (6)</entry>
 108        <entry></entry>
 109       </row>
 110       <row>
 111        <entry>write very_important_count (6)</entry>
 112        <entry></entry>
 113       </row>
 114       <row>
 115        <entry></entry>
 116        <entry>read very_important_count (6)</entry>
 117       </row>
 118       <row>
 119        <entry></entry>
 120        <entry>add 1 (7)</entry>
 121       </row>
 122       <row>
 123        <entry></entry>
 124        <entry>write very_important_count (7)</entry>
 125       </row>
 126      </tbody>
 127
 128     </tgroup>
 129    </table>
 130
 131    <para>
 132     This is what might happen:
 133    </para>
 134
 135    <table>
 136     <title>Possible Results</title>
 137
 138     <tgroup cols="2" align="left">
 139      <thead>
 140       <row>
 141        <entry>Instance 1</entry>
 142        <entry>Instance 2</entry>
 143       </row>
 144      </thead>
 145
 146      <tbody>
 147       <row>
 148        <entry>read very_important_count (5)</entry>
 149        <entry></entry>
 150       </row>
 151       <row>
 152        <entry></entry>
 153        <entry>read very_important_count (5)</entry>
 154       </row>
 155       <row>
 156        <entry>add 1 (6)</entry>
 157        <entry></entry>
 158       </row>
 159       <row>
 160        <entry></entry>
 161        <entry>add 1 (6)</entry>
 162       </row>
 163       <row>
 164        <entry>write very_important_count (6)</entry>
 165        <entry></entry>
 166       </row>
 167       <row>
 168        <entry></entry>
 169        <entry>write very_important_count (6)</entry>
 170       </row>
 171      </tbody>
 172     </tgroup>
 173    </table>
 174
 175    <sect1 id="race-condition">
 176    <title>Race Conditions and Critical Regions</title>
 177    <para>
 178      This overlap, where the result depends on the
 179      relative timing of multiple tasks, is called a <firstterm>race condition</firstterm>.
 180      The piece of code containing the concurrency issue is called a
 181      <firstterm>critical region</firstterm>.  And especially since Linux starting running
 182      on SMP machines, they became one of the major issues in kernel
 183      design and implementation.
 184    </para>
 185    <para>
 186      Preemption can have the same effect, even if there is only one
 187      CPU: by preempting one task during the critical region, we have
 188      exactly the same race condition.  In this case the thread which
 189      preempts might run the critical region itself.
 190    </para>
 191    <para>
 192      The solution is to recognize when these simultaneous accesses
 193      occur, and use locks to make sure that only one instance can
 194      enter the critical region at any time.  There are many
 195      friendly primitives in the Linux kernel to help you do this.
 196      And then there are the unfriendly primitives, but I'll pretend
 197      they don't exist.
 198    </para>
 199    </sect1>
 200  </chapter>
 201
 202  <chapter id="locks">
 203   <title>Locking in the Linux Kernel</title>
 204
 205   <para>
 206     If I could give you one piece of advice: never sleep with anyone
 207     crazier than yourself.  But if I had to give you advice on
 208     locking: <emphasis>keep it simple</emphasis>.
 209   </para>
 210
 211   <para>
 212     Be reluctant to introduce new locks.
 213   </para>
 214
 215   <para>
 216     Strangely enough, this last one is the exact reverse of my advice when
 217     you <emphasis>have</emphasis> slept with someone crazier than yourself.
 218     And you should think about getting a big dog.
 219   </para>
 220
 221   <sect1 id="lock-intro">
 222   <title>Two Main Types of Kernel Locks: Spinlocks and Mutexes</title>
 223
 224   <para>
 225     There are two main types of kernel locks.  The fundamental type
 226     is the spinlock 
 227     (<filename class="headerfile">include/asm/spinlock.h</filename>),
 228     which is a very simple single-holder lock: if you can't get the 
 229     spinlock, you keep trying (spinning) until you can.  Spinlocks are 
 230     very small and fast, and can be used anywhere.
 231   </para>
 232   <para>
 233     The second type is a mutex
 234     (<filename class="headerfile">include/linux/mutex.h</filename>): it
 235     is like a spinlock, but you may block holding a mutex.
 236     If you can't lock a mutex, your task will suspend itself, and be woken
 237     up when the mutex is released.  This means the CPU can do something
 238     else while you are waiting.  There are many cases when you simply
 239     can't sleep (see <xref linkend="sleeping-things"/>), and so have to
 240     use a spinlock instead.
 241   </para>
 242   <para>
 243     Neither type of lock is recursive: see
 244     <xref linkend="deadlock"/>.
 245   </para>
 246   </sect1>
 247 
 248   <sect1 id="uniprocessor">
 249    <title>Locks and Uniprocessor Kernels</title>
 250
 251    <para>
 252      For kernels compiled without <symbol>CONFIG_SMP</symbol>, and
 253      without <symbol>CONFIG_PREEMPT</symbol> spinlocks do not exist at
 254      all.  This is an excellent design decision: when no-one else can
 255      run at the same time, there is no reason to have a lock.
 256    </para>
 257
 258    <para>
 259      If the kernel is compiled without <symbol>CONFIG_SMP</symbol>,
 260      but <symbol>CONFIG_PREEMPT</symbol> is set, then spinlocks
 261      simply disable preemption, which is sufficient to prevent any
 262      races.  For most purposes, we can think of preemption as
 263      equivalent to SMP, and not worry about it separately.
 264    </para>
 265
 266    <para>
 267      You should always test your locking code with <symbol>CONFIG_SMP</symbol>
 268      and <symbol>CONFIG_PREEMPT</symbol> enabled, even if you don't have an SMP test box, because it
 269      will still catch some kinds of locking bugs.
 270    </para>
 271
 272    <para>
 273      Mutexes still exist, because they are required for
 274      synchronization between <firstterm linkend="gloss-usercontext">user 
 275      contexts</firstterm>, as we will see below.
 276    </para>
 277   </sect1>
 278
 279    <sect1 id="usercontextlocking">
 280     <title>Locking Only In User Context</title>
 281
 282     <para>
 283       If you have a data structure which is only ever accessed from
 284       user context, then you can use a simple mutex
 285       (<filename>include/linux/mutex.h</filename>) to protect it.  This
 286       is the most trivial case: you initialize the mutex.  Then you can
 287       call <function>mutex_lock_interruptible()</function> to grab the mutex,
 288       and <function>mutex_unlock()</function> to release it.  There is also a 
 289       <function>mutex_lock()</function>, which should be avoided, because it 
 290       will not return if a signal is received.
 291     </para>
 292
 293     <para>
 294       Example: <filename>net/netfilter/nf_sockopt.c</filename> allows 
 295       registration of new <function>setsockopt()</function> and 
 296       <function>getsockopt()</function> calls, with
 297       <function>nf_register_sockopt()</function>.  Registration and 
 298       de-registration are only done on module load and unload (and boot 
 299       time, where there is no concurrency), and the list of registrations 
 300       is only consulted for an unknown <function>setsockopt()</function>
 301       or <function>getsockopt()</function> system call.  The 
 302       <varname>nf_sockopt_mutex</varname> is perfect to protect this,
 303       especially since the setsockopt and getsockopt calls may well
 304       sleep.
 305     </para>
 306   </sect1>
 307
 308   <sect1 id="lock-user-bh">
 309    <title>Locking Between User Context and Softirqs</title>
 310
 311    <para>
 312      If a <firstterm linkend="gloss-softirq">softirq</firstterm> shares
 313      data with user context, you have two problems.  Firstly, the current 
 314      user context can be interrupted by a softirq, and secondly, the
 315      critical region could be entered from another CPU.  This is where
 316      <function>spin_lock_bh()</function> 
 317      (<filename class="headerfile">include/linux/spinlock.h</filename>) is
 318      used.  It disables softirqs on that CPU, then grabs the lock.
 319      <function>spin_unlock_bh()</function> does the reverse.  (The
 320      '_bh' suffix is a historical reference to "Bottom Halves", the
 321      old name for software interrupts.  It should really be
 322      called spin_lock_softirq()' in a perfect world).
 323    </para>
 324
 325    <para>
 326      Note that you can also use <function>spin_lock_irq()</function>
 327      or <function>spin_lock_irqsave()</function> here, which stop
 328      hardware interrupts as well: see <xref linkend="hardirq-context"/>.
 329    </para>
 330
 331    <para>
 332      This works perfectly for <firstterm linkend="gloss-up"><acronym>UP
 333      </acronym></firstterm> as well: the spin lock vanishes, and this macro 
 334      simply becomes <function>local_bh_disable()</function>
 335      (<filename class="headerfile">include/linux/interrupt.h</filename>), which
 336      protects you from the softirq being run.
 337    </para>
 338   </sect1>
 339
 340   <sect1 id="lock-user-tasklet">
 341    <title>Locking Between User Context and Tasklets</title>
 342
 343    <para>
 344      This is exactly the same as above, because <firstterm
 345      linkend="gloss-tasklet">tasklets</firstterm> are actually run
 346      from a softirq.
 347    </para>
 348   </sect1>
 349
 350   <sect1 id="lock-user-timers">
 351    <title>Locking Between User Context and Timers</title>
 352
 353    <para>
 354      This, too, is exactly the same as above, because <firstterm
 355      linkend="gloss-timers">timers</firstterm> are actually run from
 356      a softirq.  From a locking point of view, tasklets and timers
 357      are identical.
 358    </para>
 359   </sect1>
 360
 361   <sect1 id="lock-tasklets">
 362    <title>Locking Between Tasklets/Timers</title>
 363
 364    <para>
 365      Sometimes a tasklet or timer might want to share data with
 366      another tasklet or timer.
 367    </para>
 368
 369    <sect2 id="lock-tasklets-same">
 370     <title>The Same Tasklet/Timer</title>
 371     <para>
 372       Since a tasklet is never run on two CPUs at once, you don't
 373       need to worry about your tasklet being reentrant (running
 374       twice at once), even on SMP.
 375     </para>
 376    </sect2>
 377
 378    <sect2 id="lock-tasklets-different">
 379     <title>Different Tasklets/Timers</title>
 380     <para>
 381       If another tasklet/timer wants
 382       to share data with your tasklet or timer , you will both need to use
 383       <function>spin_lock()</function> and
 384       <function>spin_unlock()</function> calls.  
 385       <function>spin_lock_bh()</function> is
 386       unnecessary here, as you are already in a tasklet, and
 387       none will be run on the same CPU.
 388     </para>
 389    </sect2>
 390   </sect1>
 391
 392   <sect1 id="lock-softirqs">
 393    <title>Locking Between Softirqs</title>
 394
 395    <para>
 396      Often a softirq might
 397      want to share data with itself or a tasklet/timer.
 398    </para>
 399
 400    <sect2 id="lock-softirqs-same">
 401     <title>The Same Softirq</title>
 402
 403     <para>
 404       The same softirq can run on the other CPUs: you can use a
 405       per-CPU array (see <xref linkend="per-cpu"/>) for better
 406       performance.  If you're going so far as to use a softirq,
 407       you probably care about scalable performance enough
 408       to justify the extra complexity.
 409     </para>
 410
 411     <para>
 412       You'll need to use <function>spin_lock()</function> and 
 413       <function>spin_unlock()</function> for shared data.
 414     </para>
 415    </sect2>
 416
 417    <sect2 id="lock-softirqs-different">
 418     <title>Different Softirqs</title>
 419
 420     <para>
 421       You'll need to use <function>spin_lock()</function> and
 422       <function>spin_unlock()</function> for shared data, whether it
 423       be a timer, tasklet, different softirq or the same or another
 424       softirq: any of them could be running on a different CPU.
 425     </para>
 426    </sect2>
 427   </sect1>
 428  </chapter>
 429
 430  <chapter id="hardirq-context">
 431   <title>Hard IRQ Context</title>
 432
 433   <para>
 434     Hardware interrupts usually communicate with a
 435     tasklet or softirq.  Frequently this involves putting work in a
 436     queue, which the softirq will take out.
 437   </para>
 438
 439   <sect1 id="hardirq-softirq">
 440    <title>Locking Between Hard IRQ and Softirqs/Tasklets</title>
 441
 442    <para>
 443      If a hardware irq handler shares data with a softirq, you have
 444      two concerns.  Firstly, the softirq processing can be
 445      interrupted by a hardware interrupt, and secondly, the
 446      critical region could be entered by a hardware interrupt on
 447      another CPU.  This is where <function>spin_lock_irq()</function> is 
 448      used.  It is defined to disable interrupts on that cpu, then grab 
 449      the lock. <function>spin_unlock_irq()</function> does the reverse.
 450    </para>
 451
 452    <para>
 453      The irq handler does not to use
 454      <function>spin_lock_irq()</function>, because the softirq cannot
 455      run while the irq handler is running: it can use
 456      <function>spin_lock()</function>, which is slightly faster.  The
 457      only exception would be if a different hardware irq handler uses
 458      the same lock: <function>spin_lock_irq()</function> will stop
 459      that from interrupting us.
 460    </para>
 461
 462    <para>
 463      This works perfectly for UP as well: the spin lock vanishes,
 464      and this macro simply becomes <function>local_irq_disable()</function>
 465      (<filename class="headerfile">include/asm/smp.h</filename>), which
 466      protects you from the softirq/tasklet/BH being run.
 467    </para>
 468
 469    <para>
 470      <function>spin_lock_irqsave()</function> 
 471      (<filename>include/linux/spinlock.h</filename>) is a variant
 472      which saves whether interrupts were on or off in a flags word,
 473      which is passed to <function>spin_unlock_irqrestore()</function>.  This
 474      means that the same code can be used inside an hard irq handler (where
 475      interrupts are already off) and in softirqs (where the irq
 476      disabling is required).
 477    </para>
 478
 479    <para>
 480      Note that softirqs (and hence tasklets and timers) are run on
 481      return from hardware interrupts, so
 482      <function>spin_lock_irq()</function> also stops these.  In that
 483      sense, <function>spin_lock_irqsave()</function> is the most
 484      general and powerful locking function.
 485    </para>
 486
 487   </sect1>
 488   <sect1 id="hardirq-hardirq">
 489    <title>Locking Between Two Hard IRQ Handlers</title>
 490    <para>
 491      It is rare to have to share data between two IRQ handlers, but
 492      if you do, <function>spin_lock_irqsave()</function> should be
 493      used: it is architecture-specific whether all interrupts are
 494      disabled inside irq handlers themselves.
 495    </para>
 496   </sect1>
 497
 498  </chapter>
 499
 500  <chapter id="cheatsheet">
 501   <title>Cheat Sheet For Locking</title>
 502   <para>
 503     Pete Zaitcev gives the following summary:
 504   </para>
 505   <itemizedlist>
 506      <listitem>
 507        <para>
 508          If you are in a process context (any syscall) and want to
 509        lock other process out, use a mutex.  You can take a mutex
 510        and sleep (<function>copy_from_user*(</function> or
 511        <function>kmalloc(x,GFP_KERNEL)</function>).
 512      </para>
 513      </listitem>
 514      <listitem>
 515        <para>
 516        Otherwise (== data can be touched in an interrupt), use
 517        <function>spin_lock_irqsave()</function> and
 518        <function>spin_unlock_irqrestore()</function>.
 519        </para>
 520      </listitem>
 521      <listitem>
 522        <para>
 523        Avoid holding spinlock for more than 5 lines of code and
 524        across any function call (except accessors like
 525        <function>readb</function>).
 526        </para>
 527      </listitem>
 528    </itemizedlist>
 529
 530   <sect1 id="minimum-lock-reqirements">
 531   <title>Table of Minimum Requirements</title>
 532
 533   <para> The following table lists the <emphasis>minimum</emphasis>
 534        locking requirements between various contexts.  In some cases,
 535        the same context can only be running on one CPU at a time, so
 536        no locking is required for that context (eg. a particular
 537        thread can only run on one CPU at a time, but if it needs
 538        shares data with another thread, locking is required).
 539   </para>
 540   <para>
 541        Remember the advice above: you can always use
 542        <function>spin_lock_irqsave()</function>, which is a superset
 543        of all other spinlock primitives.
 544   </para>
 545
 546   <table>
 547<title>Table of Locking Requirements</title>
 548<tgroup cols="11">
 549<tbody>
 550
 551<row>
 552<entry></entry>
 553<entry>IRQ Handler A</entry>
 554<entry>IRQ Handler B</entry>
 555<entry>Softirq A</entry>
 556<entry>Softirq B</entry>
 557<entry>Tasklet A</entry>
 558<entry>Tasklet B</entry>
 559<entry>Timer A</entry>
 560<entry>Timer B</entry>
 561<entry>User Context A</entry>
 562<entry>User Context B</entry>
 563</row>
 564
 565<row>
 566<entry>IRQ Handler A</entry>
 567<entry>None</entry>
 568</row>
 569
 570<row>
 571<entry>IRQ Handler B</entry>
 572<entry>SLIS</entry>
 573<entry>None</entry>
 574</row>
 575
 576<row>
 577<entry>Softirq A</entry>
 578<entry>SLI</entry>
 579<entry>SLI</entry>
 580<entry>SL</entry>
 581</row>
 582
 583<row>
 584<entry>Softirq B</entry>
 585<entry>SLI</entry>
 586<entry>SLI</entry>
 587<entry>SL</entry>
 588<entry>SL</entry>
 589</row>
 590
 591<row>
 592<entry>Tasklet A</entry>
 593<entry>SLI</entry>
 594<entry>SLI</entry>
 595<entry>SL</entry>
 596<entry>SL</entry>
 597<entry>None</entry>
 598</row>
 599
 600<row>
 601<entry>Tasklet B</entry>
 602<entry>SLI</entry>
 603<entry>SLI</entry>
 604<entry>SL</entry>
 605<entry>SL</entry>
 606<entry>SL</entry>
 607<entry>None</entry>
 608</row>
 609
 610<row>
 611<entry>Timer A</entry>
 612<entry>SLI</entry>
 613<entry>SLI</entry>
 614<entry>SL</entry>
 615<entry>SL</entry>
 616<entry>SL</entry>
 617<entry>SL</entry>
 618<entry>None</entry>
 619</row>
 620
 621<row>
 622<entry>Timer B</entry>
 623<entry>SLI</entry>
 624<entry>SLI</entry>
 625<entry>SL</entry>
 626<entry>SL</entry>
 627<entry>SL</entry>
 628<entry>SL</entry>
 629<entry>SL</entry>
 630<entry>None</entry>
 631</row>
 632
 633<row>
 634<entry>User Context A</entry>
 635<entry>SLI</entry>
 636<entry>SLI</entry>
 637<entry>SLBH</entry>
 638<entry>SLBH</entry>
 639<entry>SLBH</entry>
 640<entry>SLBH</entry>
 641<entry>SLBH</entry>
 642<entry>SLBH</entry>
 643<entry>None</entry>
 644</row>
 645
 646<row>
 647<entry>User Context B</entry>
 648<entry>SLI</entry>
 649<entry>SLI</entry>
 650<entry>SLBH</entry>
 651<entry>SLBH</entry>
 652<entry>SLBH</entry>
 653<entry>SLBH</entry>
 654<entry>SLBH</entry>
 655<entry>SLBH</entry>
 656<entry>MLI</entry>
 657<entry>None</entry>
 658</row>
 659
 660</tbody>
 661</tgroup>
 662</table>
 663
 664   <table>
 665<title>Legend for Locking Requirements Table</title>
 666<tgroup cols="2">
 667<tbody>
 668
 669<row>
 670<entry>SLIS</entry>
 671<entry>spin_lock_irqsave</entry>
 672</row>
 673<row>
 674<entry>SLI</entry>
 675<entry>spin_lock_irq</entry>
 676</row>
 677<row>
 678<entry>SL</entry>
 679<entry>spin_lock</entry>
 680</row>
 681<row>
 682<entry>SLBH</entry>
 683<entry>spin_lock_bh</entry>
 684</row>
 685<row>
 686<entry>MLI</entry>
 687<entry>mutex_lock_interruptible</entry>
 688</row>
 689
 690</tbody>
 691</tgroup>
 692</table>
 693
 694</sect1>
 695</chapter>
 696
 697<chapter id="trylock-functions">
 698 <title>The trylock Functions</title>
 699  <para>
 700   There are functions that try to acquire a lock only once and immediately
 701   return a value telling about success or failure to acquire the lock.
 702   They can be used if you need no access to the data protected with the lock
 703   when some other thread is holding the lock. You should acquire the lock
 704   later if you then need access to the data protected with the lock.
 705  </para>
 706
 707  <para>
 708    <function>spin_trylock()</function> does not spin but returns non-zero if
 709    it acquires the spinlock on the first try or 0 if not. This function can
 710    be used in all contexts like <function>spin_lock</function>: you must have
 711    disabled the contexts that might interrupt you and acquire the spin lock.
 712  </para>
 713
 714  <para>
 715    <function>mutex_trylock()</function> does not suspend your task
 716    but returns non-zero if it could lock the mutex on the first try
 717    or 0 if not. This function cannot be safely used in hardware or software
 718    interrupt contexts despite not sleeping.
 719  </para>
 720</chapter>
 721
 722  <chapter id="Examples">
 723   <title>Common Examples</title>
 724    <para>
 725Let's step through a simple example: a cache of number to name
 726mappings.  The cache keeps a count of how often each of the objects is
 727used, and when it gets full, throws out the least used one.
 728
 729    </para>
 730
 731   <sect1 id="examples-usercontext">
 732    <title>All In User Context</title>
 733    <para>
 734For our first example, we assume that all operations are in user
 735context (ie. from system calls), so we can sleep.  This means we can
 736use a mutex to protect the cache and all the objects within
 737it.  Here's the code:
 738    </para>
 739
 740    <programlisting>
 741#include &lt;linux/list.h&gt;
 742#include &lt;linux/slab.h&gt;
 743#include &lt;linux/string.h&gt;
 744#include &lt;linux/mutex.h&gt;
 745#include &lt;asm/errno.h&gt;
 746
 747struct object
 748{
 749        struct list_head list;
 750        int id;
 751        char name[32];
 752        int popularity;
 753};
 754
 755/* Protects the cache, cache_num, and the objects within it */
 756static DEFINE_MUTEX(cache_lock);
 757static LIST_HEAD(cache);
 758static unsigned int cache_num = 0;
 759#define MAX_CACHE_SIZE 10
 760
 761/* Must be holding cache_lock */
 762static struct object *__cache_find(int id)
 763{
 764        struct object *i;
 765
 766        list_for_each_entry(i, &amp;cache, list)
 767                if (i-&gt;id == id) {
 768                        i-&gt;popularity++;
 769                        return i;
 770                }
 771        return NULL;
 772}
 773
 774/* Must be holding cache_lock */
 775static void __cache_delete(struct object *obj)
 776{
 777        BUG_ON(!obj);
 778        list_del(&amp;obj-&gt;list);
 779        kfree(obj);
 780        cache_num--;
 781}
 782
 783/* Must be holding cache_lock */
 784static void __cache_add(struct object *obj)
 785{
 786        list_add(&amp;obj-&gt;list, &amp;cache);
 787        if (++cache_num > MAX_CACHE_SIZE) {
 788                struct object *i, *outcast = NULL;
 789                list_for_each_entry(i, &amp;cache, list) {
 790                        if (!outcast || i-&gt;popularity &lt; outcast-&gt;popularity)
 791                                outcast = i;
 792                }
 793                __cache_delete(outcast);
 794        }
 795}
 796
 797int cache_add(int id, const char *name)
 798{
 799        struct object *obj;
 800
 801        if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
 802                return -ENOMEM;
 803
 804        strlcpy(obj-&gt;name, name, sizeof(obj-&gt;name));
 805        obj-&gt;id = id;
 806        obj-&gt;popularity = 0;
 807
 808        mutex_lock(&amp;cache_lock);
 809        __cache_add(obj);
 810        mutex_unlock(&amp;cache_lock);
 811        return 0;
 812}
 813
 814void cache_delete(int id)
 815{
 816        mutex_lock(&amp;cache_lock);
 817        __cache_delete(__cache_find(id));
 818        mutex_unlock(&amp;cache_lock);
 819}
 820
 821int cache_find(int id, char *name)
 822{
 823        struct object *obj;
 824        int ret = -ENOENT;
 825
 826        mutex_lock(&amp;cache_lock);
 827        obj = __cache_find(id);
 828        if (obj) {
 829                ret = 0;
 830                strcpy(name, obj-&gt;name);
 831        }
 832        mutex_unlock(&amp;cache_lock);
 833        return ret;
 834}
 835</programlisting>
 836
 837    <para>
 838Note that we always make sure we have the cache_lock when we add,
 839delete, or look up the cache: both the cache infrastructure itself and
 840the contents of the objects are protected by the lock.  In this case
 841it's easy, since we copy the data for the user, and never let them
 842access the objects directly.
 843    </para>
 844    <para>
 845There is a slight (and common) optimization here: in
 846<function>cache_add</function> we set up the fields of the object
 847before grabbing the lock.  This is safe, as no-one else can access it
 848until we put it in cache.
 849    </para>
 850    </sect1>
 851
 852   <sect1 id="examples-interrupt">
 853    <title>Accessing From Interrupt Context</title>
 854    <para>
 855Now consider the case where <function>cache_find</function> can be
 856called from interrupt context: either a hardware interrupt or a
 857softirq.  An example would be a timer which deletes object from the
 858cache.
 859    </para>
 860    <para>
 861The change is shown below, in standard patch format: the
 862<symbol>-</symbol> are lines which are taken away, and the
 863<symbol>+</symbol> are lines which are added.
 864    </para>
 865<programlisting>
 866--- cache.c.usercontext 2003-12-09 13:58:54.000000000 +1100
 867+++ cache.c.interrupt   2003-12-09 14:07:49.000000000 +1100
 868@@ -12,7 +12,7 @@
 869         int popularity;
 870 };
 871
 872-static DEFINE_MUTEX(cache_lock);
 873+static DEFINE_SPINLOCK(cache_lock);
 874 static LIST_HEAD(cache);
 875 static unsigned int cache_num = 0;
 876 #define MAX_CACHE_SIZE 10
 877@@ -55,6 +55,7 @@
 878 int cache_add(int id, const char *name)
 879 {
 880         struct object *obj;
 881+        unsigned long flags;
 882
 883         if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
 884                 return -ENOMEM;
 885@@ -63,30 +64,33 @@
 886         obj-&gt;id = id;
 887         obj-&gt;popularity = 0;
 888
 889-        mutex_lock(&amp;cache_lock);
 890+        spin_lock_irqsave(&amp;cache_lock, flags);
 891         __cache_add(obj);
 892-        mutex_unlock(&amp;cache_lock);
 893+        spin_unlock_irqrestore(&amp;cache_lock, flags);
 894         return 0;
 895 }
 896
 897 void cache_delete(int id)
 898 {
 899-        mutex_lock(&amp;cache_lock);
 900+        unsigned long flags;
 901+
 902+        spin_lock_irqsave(&amp;cache_lock, flags);
 903         __cache_delete(__cache_find(id));
 904-        mutex_unlock(&amp;cache_lock);
 905+        spin_unlock_irqrestore(&amp;cache_lock, flags);
 906 }
 907
 908 int cache_find(int id, char *name)
 909 {
 910         struct object *obj;
 911         int ret = -ENOENT;
 912+        unsigned long flags;
 913
 914-        mutex_lock(&amp;cache_lock);
 915+        spin_lock_irqsave(&amp;cache_lock, flags);
 916         obj = __cache_find(id);
 917         if (obj) {
 918                 ret = 0;
 919                 strcpy(name, obj-&gt;name);
 920         }
 921-        mutex_unlock(&amp;cache_lock);
 922+        spin_unlock_irqrestore(&amp;cache_lock, flags);
 923         return ret;
 924 }
 925</programlisting>
 926
 927    <para>
 928Note that the <function>spin_lock_irqsave</function> will turn off
 929interrupts if they are on, otherwise does nothing (if we are already
 930in an interrupt handler), hence these functions are safe to call from
 931any context.
 932    </para>
 933    <para>
 934Unfortunately, <function>cache_add</function> calls
 935<function>kmalloc</function> with the <symbol>GFP_KERNEL</symbol>
 936flag, which is only legal in user context.  I have assumed that
 937<function>cache_add</function> is still only called in user context,
 938otherwise this should become a parameter to
 939<function>cache_add</function>.
 940    </para>
 941  </sect1>
 942   <sect1 id="examples-refcnt">
 943    <title>Exposing Objects Outside This File</title>
 944    <para>
 945If our objects contained more information, it might not be sufficient
 946to copy the information in and out: other parts of the code might want
 947to keep pointers to these objects, for example, rather than looking up
 948the id every time.  This produces two problems.
 949    </para>
 950    <para>
 951The first problem is that we use the <symbol>cache_lock</symbol> to
 952protect objects: we'd need to make this non-static so the rest of the
 953code can use it.  This makes locking trickier, as it is no longer all
 954in one place.
 955    </para>
 956    <para>
 957The second problem is the lifetime problem: if another structure keeps
 958a pointer to an object, it presumably expects that pointer to remain
 959valid.  Unfortunately, this is only guaranteed while you hold the
 960lock, otherwise someone might call <function>cache_delete</function>
 961and even worse, add another object, re-using the same address.
 962    </para>
 963    <para>
 964As there is only one lock, you can't hold it forever: no-one else would
 965get any work done.
 966    </para>
 967    <para>
 968The solution to this problem is to use a reference count: everyone who
 969has a pointer to the object increases it when they first get the
 970object, and drops the reference count when they're finished with it.
 971Whoever drops it to zero knows it is unused, and can actually delete it.
 972    </para>
 973    <para>
 974Here is the code:
 975    </para>
 976
 977<programlisting>
 978--- cache.c.interrupt   2003-12-09 14:25:43.000000000 +1100
 979+++ cache.c.refcnt      2003-12-09 14:33:05.000000000 +1100
 980@@ -7,6 +7,7 @@
 981 struct object
 982 {
 983         struct list_head list;
 984+        unsigned int refcnt;
 985         int id;
 986         char name[32];
 987         int popularity;
 988@@ -17,6 +18,35 @@
 989 static unsigned int cache_num = 0;
 990 #define MAX_CACHE_SIZE 10
 991
 992+static void __object_put(struct object *obj)
 993+{
 994+        if (--obj-&gt;refcnt == 0)
 995+                kfree(obj);
 996+}
 997+
 998+static void __object_get(struct object *obj)
 999+{
1000+        obj-&gt;refcnt++;
1001+}
1002+
1003+void object_put(struct object *obj)
1004+{
1005+        unsigned long flags;
1006+
1007+        spin_lock_irqsave(&amp;cache_lock, flags);
1008+        __object_put(obj);
1009+        spin_unlock_irqrestore(&amp;cache_lock, flags);
1010+}
1011+
1012+void object_get(struct object *obj)
1013+{
1014+        unsigned long flags;
1015+
1016+        spin_lock_irqsave(&amp;cache_lock, flags);
1017+        __object_get(obj);
1018+        spin_unlock_irqrestore(&amp;cache_lock, flags);
1019+}
1020+
1021 /* Must be holding cache_lock */
1022 static struct object *__cache_find(int id)
1023 {
1024@@ -35,6 +65,7 @@
1025 {
1026         BUG_ON(!obj);
1027         list_del(&amp;obj-&gt;list);
1028+        __object_put(obj);
1029         cache_num--;
1030 }
1031
1032@@ -63,6 +94,7 @@
1033         strlcpy(obj-&gt;name, name, sizeof(obj-&gt;name));
1034         obj-&gt;id = id;
1035         obj-&gt;popularity = 0;
1036+        obj-&gt;refcnt = 1; /* The cache holds a reference */
1037
1038         spin_lock_irqsave(&amp;cache_lock, flags);
1039         __cache_add(obj);
1040@@ -79,18 +111,15 @@
1041         spin_unlock_irqrestore(&amp;cache_lock, flags);
1042 }
1043
1044-int cache_find(int id, char *name)
1045+struct object *cache_find(int id)
1046 {
1047         struct object *obj;
1048-        int ret = -ENOENT;
1049         unsigned long flags;
1050
1051         spin_lock_irqsave(&amp;cache_lock, flags);
1052         obj = __cache_find(id);
1053-        if (obj) {
1054-                ret = 0;
1055-                strcpy(name, obj-&gt;name);
1056-        }
1057+        if (obj)
1058+                __object_get(obj);
1059         spin_unlock_irqrestore(&amp;cache_lock, flags);
1060-        return ret;
1061+        return obj;
1062 }
1063</programlisting>
1064
1065<para>
1066We encapsulate the reference counting in the standard 'get' and 'put'
1067functions.  Now we can return the object itself from
1068<function>cache_find</function> which has the advantage that the user
1069can now sleep holding the object (eg. to
1070<function>copy_to_user</function> to name to userspace).
1071</para>
1072<para>
1073The other point to note is that I said a reference should be held for
1074every pointer to the object: thus the reference count is 1 when first
1075inserted into the cache.  In some versions the framework does not hold
1076a reference count, but they are more complicated.
1077</para>
1078
1079   <sect2 id="examples-refcnt-atomic">
1080    <title>Using Atomic Operations For The Reference Count</title>
1081<para>
1082In practice, <type>atomic_t</type> would usually be used for
1083<structfield>refcnt</structfield>.  There are a number of atomic
1084operations defined in
1085
1086<filename class="headerfile">include/asm/atomic.h</filename>: these are
1087guaranteed to be seen atomically from all CPUs in the system, so no
1088lock is required.  In this case, it is simpler than using spinlocks,
1089although for anything non-trivial using spinlocks is clearer.  The
1090<function>atomic_inc</function> and
1091<function>atomic_dec_and_test</function> are used instead of the
1092standard increment and decrement operators, and the lock is no longer
1093used to protect the reference count itself.
1094</para>
1095
1096<programlisting>
1097--- cache.c.refcnt      2003-12-09 15:00:35.000000000 +1100
1098+++ cache.c.refcnt-atomic       2003-12-11 15:49:42.000000000 +1100
1099@@ -7,7 +7,7 @@
1100 struct object
1101 {
1102         struct list_head list;
1103-        unsigned int refcnt;
1104+        atomic_t refcnt;
1105         int id;
1106         char name[32];
1107         int popularity;
1108@@ -18,33 +18,15 @@
1109 static unsigned int cache_num = 0;
1110 #define MAX_CACHE_SIZE 10
1111
1112-static void __object_put(struct object *obj)
1113-{
1114-        if (--obj-&gt;refcnt == 0)
1115-                kfree(obj);
1116-}
1117-
1118-static void __object_get(struct object *obj)
1119-{
1120-        obj-&gt;refcnt++;
1121-}
1122-
1123 void object_put(struct object *obj)
1124 {
1125-        unsigned long flags;
1126-
1127-        spin_lock_irqsave(&amp;cache_lock, flags);
1128-        __object_put(obj);
1129-        spin_unlock_irqrestore(&amp;cache_lock, flags);
1130+        if (atomic_dec_and_test(&amp;obj-&gt;refcnt))
1131+                kfree(obj);
1132 }
1133
1134 void object_get(struct object *obj)
1135 {
1136-        unsigned long flags;
1137-
1138-        spin_lock_irqsave(&amp;cache_lock, flags);
1139-        __object_get(obj);
1140-        spin_unlock_irqrestore(&amp;cache_lock, flags);
1141+        atomic_inc(&amp;obj-&gt;refcnt);
1142 }
1143
1144 /* Must be holding cache_lock */
1145@@ -65,7 +47,7 @@
1146 {
1147         BUG_ON(!obj);
1148         list_del(&amp;obj-&gt;list);
1149-        __object_put(obj);
1150+        object_put(obj);
1151         cache_num--;
1152 }
1153
1154@@ -94,7 +76,7 @@
1155         strlcpy(obj-&gt;name, name, sizeof(obj-&gt;name));
1156         obj-&gt;id = id;
1157         obj-&gt;popularity = 0;
1158-        obj-&gt;refcnt = 1; /* The cache holds a reference */
1159+        atomic_set(&amp;obj-&gt;refcnt, 1); /* The cache holds a reference */
1160
1161         spin_lock_irqsave(&amp;cache_lock, flags);
1162         __cache_add(obj);
1163@@ -119,7 +101,7 @@
1164         spin_lock_irqsave(&amp;cache_lock, flags);
1165         obj = __cache_find(id);
1166         if (obj)
1167-                __object_get(obj);
1168+                object_get(obj);
1169         spin_unlock_irqrestore(&amp;cache_lock, flags);
1170         return obj;
1171 }
1172</programlisting>
1173</sect2>
1174</sect1>
1175
1176   <sect1 id="examples-lock-per-obj">
1177    <title>Protecting The Objects Themselves</title>
1178    <para>
1179In these examples, we assumed that the objects (except the reference
1180counts) never changed once they are created.  If we wanted to allow
1181the name to change, there are three possibilities:
1182    </para>
1183    <itemizedlist>
1184      <listitem>
1185        <para>
1186You can make <symbol>cache_lock</symbol> non-static, and tell people
1187to grab that lock before changing the name in any object.
1188        </para>
1189      </listitem>
1190      <listitem>
1191        <para>
1192You can provide a <function>cache_obj_rename</function> which grabs
1193this lock and changes the name for the caller, and tell everyone to
1194use that function.
1195        </para>
1196      </listitem>
1197      <listitem>
1198        <para>
1199You can make the <symbol>cache_lock</symbol> protect only the cache
1200itself, and use another lock to protect the name.
1201        </para>
1202      </listitem>
1203    </itemizedlist>
1204
1205      <para>
1206Theoretically, you can make the locks as fine-grained as one lock for
1207every field, for every object.  In practice, the most common variants
1208are:
1209</para>
1210    <itemizedlist>
1211      <listitem>
1212        <para>
1213One lock which protects the infrastructure (the <symbol>cache</symbol>
1214list in this example) and all the objects.  This is what we have done
1215so far.
1216        </para>
1217      </listitem>
1218      <listitem>
1219        <para>
1220One lock which protects the infrastructure (including the list
1221pointers inside the objects), and one lock inside the object which
1222protects the rest of that object.
1223        </para>
1224      </listitem>
1225      <listitem>
1226        <para>
1227Multiple locks to protect the infrastructure (eg. one lock per hash
1228chain), possibly with a separate per-object lock.
1229        </para>
1230      </listitem>
1231    </itemizedlist>
1232
1233<para>
1234Here is the "lock-per-object" implementation:
1235</para>
1236<programlisting>
1237--- cache.c.refcnt-atomic       2003-12-11 15:50:54.000000000 +1100
1238+++ cache.c.perobjectlock       2003-12-11 17:15:03.000000000 +1100
1239@@ -6,11 +6,17 @@
1240
1241 struct object
1242 {
1243+        /* These two protected by cache_lock. */
1244         struct list_head list;
1245+        int popularity;
1246+
1247         atomic_t refcnt;
1248+
1249+        /* Doesn't change once created. */
1250         int id;
1251+
1252+        spinlock_t lock; /* Protects the name */
1253         char name[32];
1254-        int popularity;
1255 };
1256
1257 static DEFINE_SPINLOCK(cache_lock);
1258@@ -77,6 +84,7 @@
1259         obj-&gt;id = id;
1260         obj-&gt;popularity = 0;
1261         atomic_set(&amp;obj-&gt;refcnt, 1); /* The cache holds a reference */
1262+        spin_lock_init(&amp;obj-&gt;lock);
1263
1264         spin_lock_irqsave(&amp;cache_lock, flags);
1265         __cache_add(obj);
1266</programlisting>
1267
1268<para>
1269Note that I decide that the <structfield>popularity</structfield>
1270count should be protected by the <symbol>cache_lock</symbol> rather
1271than the per-object lock: this is because it (like the
1272<structname>struct list_head</structname> inside the object) is
1273logically part of the infrastructure.  This way, I don't need to grab
1274the lock of every object in <function>__cache_add</function> when
1275seeking the least popular.
1276</para>
1277
1278<para>
1279I also decided that the <structfield>id</structfield> member is
1280unchangeable, so I don't need to grab each object lock in
1281<function>__cache_find()</function> to examine the
1282<structfield>id</structfield>: the object lock is only used by a
1283caller who wants to read or write the <structfield>name</structfield>
1284field.
1285</para>
1286
1287<para>
1288Note also that I added a comment describing what data was protected by
1289which locks.  This is extremely important, as it describes the runtime
1290behavior of the code, and can be hard to gain from just reading.  And
1291as Alan Cox says, <quote>Lock data, not code</quote>.
1292</para>
1293</sect1>
1294</chapter>
1295
1296   <chapter id="common-problems">
1297    <title>Common Problems</title>
1298    <sect1 id="deadlock">
1299    <title>Deadlock: Simple and Advanced</title>
1300
1301    <para>
1302      There is a coding bug where a piece of code tries to grab a
1303      spinlock twice: it will spin forever, waiting for the lock to
1304      be released (spinlocks, rwlocks and mutexes are not
1305      recursive in Linux).  This is trivial to diagnose: not a
1306      stay-up-five-nights-talk-to-fluffy-code-bunnies kind of
1307      problem.
1308    </para>
1309
1310    <para>
1311      For a slightly more complex case, imagine you have a region
1312      shared by a softirq and user context.  If you use a
1313      <function>spin_lock()</function> call to protect it, it is 
1314      possible that the user context will be interrupted by the softirq
1315      while it holds the lock, and the softirq will then spin
1316      forever trying to get the same lock.
1317    </para>
1318
1319    <para>
1320      Both of these are called deadlock, and as shown above, it can
1321      occur even with a single CPU (although not on UP compiles,
1322      since spinlocks vanish on kernel compiles with 
1323      <symbol>CONFIG_SMP</symbol>=n. You'll still get data corruption 
1324      in the second example).
1325    </para>
1326
1327    <para>
1328      This complete lockup is easy to diagnose: on SMP boxes the
1329      watchdog timer or compiling with <symbol>DEBUG_SPINLOCK</symbol> set
1330      (<filename>include/linux/spinlock.h</filename>) will show this up 
1331      immediately when it happens.
1332    </para>
1333
1334    <para>
1335      A more complex problem is the so-called 'deadly embrace',
1336      involving two or more locks.  Say you have a hash table: each
1337      entry in the table is a spinlock, and a chain of hashed
1338      objects.  Inside a softirq handler, you sometimes want to
1339      alter an object from one place in the hash to another: you
1340      grab the spinlock of the old hash chain and the spinlock of
1341      the new hash chain, and delete the object from the old one,
1342      and insert it in the new one.
1343    </para>
1344
1345    <para>
1346      There are two problems here.  First, if your code ever
1347      tries to move the object to the same chain, it will deadlock
1348      with itself as it tries to lock it twice.  Secondly, if the
1349      same softirq on another CPU is trying to move another object
1350      in the reverse direction, the following could happen:
1351    </para>
1352
1353    <table>
1354     <title>Consequences</title>
1355
1356     <tgroup cols="2" align="left">
1357
1358      <thead>
1359       <row>
1360        <entry>CPU 1</entry>
1361        <entry>CPU 2</entry>
1362       </row>
1363      </thead>
1364
1365      <tbody>
1366       <row>
1367        <entry>Grab lock A -&gt; OK</entry>
1368        <entry>Grab lock B -&gt; OK</entry>
1369       </row>
1370       <row>
1371        <entry>Grab lock B -&gt; spin</entry>
1372        <entry>Grab lock A -&gt; spin</entry>
1373       </row>
1374      </tbody>
1375     </tgroup>
1376    </table>
1377
1378    <para>
1379      The two CPUs will spin forever, waiting for the other to give up
1380      their lock.  It will look, smell, and feel like a crash.
1381    </para>
1382    </sect1>
1383
1384    <sect1 id="techs-deadlock-prevent">
1385     <title>Preventing Deadlock</title>
1386
1387     <para>
1388       Textbooks will tell you that if you always lock in the same
1389       order, you will never get this kind of deadlock.  Practice
1390       will tell you that this approach doesn't scale: when I
1391       create a new lock, I don't understand enough of the kernel
1392       to figure out where in the 5000 lock hierarchy it will fit.
1393     </para>
1394
1395     <para>
1396       The best locks are encapsulated: they never get exposed in
1397       headers, and are never held around calls to non-trivial
1398       functions outside the same file.  You can read through this
1399       code and see that it will never deadlock, because it never
1400       tries to grab another lock while it has that one.  People
1401       using your code don't even need to know you are using a
1402       lock.
1403     </para>
1404
1405     <para>
1406       A classic problem here is when you provide callbacks or
1407       hooks: if you call these with the lock held, you risk simple
1408       deadlock, or a deadly embrace (who knows what the callback
1409       will do?).  Remember, the other programmers are out to get
1410       you, so don't do this.
1411     </para>
1412
1413    <sect2 id="techs-deadlock-overprevent">
1414     <title>Overzealous Prevention Of Deadlocks</title>
1415
1416     <para>
1417       Deadlocks are problematic, but not as bad as data
1418       corruption.  Code which grabs a read lock, searches a list,
1419       fails to find what it wants, drops the read lock, grabs a
1420       write lock and inserts the object has a race condition.
1421     </para>
1422
1423     <para>
1424       If you don't see why, please stay the fuck away from my code.
1425     </para>
1426    </sect2>
1427    </sect1>
1428
1429   <sect1 id="racing-timers">
1430    <title>Racing Timers: A Kernel Pastime</title>
1431
1432    <para>
1433      Timers can produce their own special problems with races.
1434      Consider a collection of objects (list, hash, etc) where each
1435      object has a timer which is due to destroy it.
1436    </para>
1437
1438    <para>
1439      If you want to destroy the entire collection (say on module
1440      removal), you might do the following:
1441    </para>
1442
1443    <programlisting>
1444        /* THIS CODE BAD BAD BAD BAD: IF IT WAS ANY WORSE IT WOULD USE
1445           HUNGARIAN NOTATION */
1446        spin_lock_bh(&amp;list_lock);
1447
1448        while (list) {
1449                struct foo *next = list-&gt;next;
1450                del_timer(&amp;list-&gt;timer);
1451                kfree(list);
1452                list = next;
1453        }
1454
1455        spin_unlock_bh(&amp;list_lock);
1456    </programlisting>
1457
1458    <para>
1459      Sooner or later, this will crash on SMP, because a timer can
1460      have just gone off before the <function>spin_lock_bh()</function>,
1461      and it will only get the lock after we
1462      <function>spin_unlock_bh()</function>, and then try to free
1463      the element (which has already been freed!).
1464    </para>
1465
1466    <para>
1467      This can be avoided by checking the result of
1468      <function>del_timer()</function>: if it returns
1469      <returnvalue>1</returnvalue>, the timer has been deleted.
1470      If <returnvalue>0</returnvalue>, it means (in this
1471      case) that it is currently running, so we can do:
1472    </para>
1473
1474    <programlisting>
1475        retry:
1476                spin_lock_bh(&amp;list_lock);
1477
1478                while (list) {
1479                        struct foo *next = list-&gt;next;
1480                        if (!del_timer(&amp;list-&gt;timer)) {
1481                                /* Give timer a chance to delete this */
1482                                spin_unlock_bh(&amp;list_lock);
1483                                goto retry;
1484                        }
1485                        kfree(list);
1486                        list = next;
1487                }
1488
1489                spin_unlock_bh(&amp;list_lock);
1490    </programlisting>
1491
1492    <para>
1493      Another common problem is deleting timers which restart
1494      themselves (by calling <function>add_timer()</function> at the end
1495      of their timer function).  Because this is a fairly common case
1496      which is prone to races, you should use <function>del_timer_sync()</function>
1497      (<filename class="headerfile">include/linux/timer.h</filename>)
1498      to handle this case.  It returns the number of times the timer
1499      had to be deleted before we finally stopped it from adding itself back
1500      in.
1501    </para>
1502   </sect1>
1503
1504  </chapter>
1505
1506 <chapter id="Efficiency">
1507    <title>Locking Speed</title>
1508
1509    <para>
1510There are three main things to worry about when considering speed of
1511some code which does locking.  First is concurrency: how many things
1512are going to be waiting while someone else is holding a lock.  Second
1513is the time taken to actually acquire and release an uncontended lock.
1514Third is using fewer, or smarter locks.  I'm assuming that the lock is
1515used fairly often: otherwise, you wouldn't be concerned about
1516efficiency.
1517</para>
1518    <para>
1519Concurrency depends on how long the lock is usually held: you should
1520hold the lock for as long as needed, but no longer.  In the cache
1521example, we always create the object without the lock held, and then
1522grab the lock only when we are ready to insert it in the list.
1523</para>
1524    <para>
1525Acquisition times depend on how much damage the lock operations do to
1526the pipeline (pipeline stalls) and how likely it is that this CPU was
1527the last one to grab the lock (ie. is the lock cache-hot for this
1528CPU): on a machine with more CPUs, this likelihood drops fast.
1529Consider a 700MHz Intel Pentium III: an instruction takes about 0.7ns,
1530an atomic increment takes about 58ns, a lock which is cache-hot on
1531this CPU takes 160ns, and a cacheline transfer from another CPU takes
1532an additional 170 to 360ns.  (These figures from Paul McKenney's
1533<ulink url="http://www.linuxjournal.com/article.php?sid=6993"> Linux
1534Journal RCU article</ulink>).
1535</para>
1536    <para>
1537These two aims conflict: holding a lock for a short time might be done
1538by splitting locks into parts (such as in our final per-object-lock
1539example), but this increases the number of lock acquisitions, and the
1540results are often slower than having a single lock.  This is another
1541reason to advocate locking simplicity.
1542</para>
1543    <para>
1544The third concern is addressed below: there are some methods to reduce
1545the amount of locking which needs to be done.
1546</para>
1547
1548  <sect1 id="efficiency-rwlocks">
1549   <title>Read/Write Lock Variants</title>
1550
1551   <para>
1552      Both spinlocks and mutexes have read/write variants:
1553      <type>rwlock_t</type> and <structname>struct rw_semaphore</structname>.
1554      These divide users into two classes: the readers and the writers.  If
1555      you are only reading the data, you can get a read lock, but to write to
1556      the data you need the write lock.  Many people can hold a read lock,
1557      but a writer must be sole holder.
1558    </para>
1559
1560   <para>
1561      If your code divides neatly along reader/writer lines (as our
1562      cache code does), and the lock is held by readers for
1563      significant lengths of time, using these locks can help.  They
1564      are slightly slower than the normal locks though, so in practice
1565      <type>rwlock_t</type> is not usually worthwhile.
1566    </para>
1567   </sect1>
1568
1569   <sect1 id="efficiency-read-copy-update">
1570    <title>Avoiding Locks: Read Copy Update</title>
1571
1572    <para>
1573      There is a special method of read/write locking called Read Copy
1574      Update.  Using RCU, the readers can avoid taking a lock
1575      altogether: as we expect our cache to be read more often than
1576      updated (otherwise the cache is a waste of time), it is a
1577      candidate for this optimization.
1578    </para>
1579
1580    <para>
1581      How do we get rid of read locks?  Getting rid of read locks
1582      means that writers may be changing the list underneath the
1583      readers.  That is actually quite simple: we can read a linked
1584      list while an element is being added if the writer adds the
1585      element very carefully.  For example, adding
1586      <symbol>new</symbol> to a single linked list called
1587      <symbol>list</symbol>:
1588    </para>
1589
1590    <programlisting>
1591        new-&gt;next = list-&gt;next;
1592        wmb();
1593        list-&gt;next = new;
1594    </programlisting>
1595
1596    <para>
1597      The <function>wmb()</function> is a write memory barrier.  It
1598      ensures that the first operation (setting the new element's
1599      <symbol>next</symbol> pointer) is complete and will be seen by
1600      all CPUs, before the second operation is (putting the new
1601      element into the list).  This is important, since modern
1602      compilers and modern CPUs can both reorder instructions unless
1603      told otherwise: we want a reader to either not see the new
1604      element at all, or see the new element with the
1605      <symbol>next</symbol> pointer correctly pointing at the rest of
1606      the list.
1607    </para>
1608    <para>
1609      Fortunately, there is a function to do this for standard
1610      <structname>struct list_head</structname> lists:
1611      <function>list_add_rcu()</function>
1612      (<filename>include/linux/list.h</filename>).
1613    </para>
1614    <para>
1615      Removing an element from the list is even simpler: we replace
1616      the pointer to the old element with a pointer to its successor,
1617      and readers will either see it, or skip over it.
1618    </para>
1619    <programlisting>
1620        list-&gt;next = old-&gt;next;
1621    </programlisting>
1622    <para>
1623      There is <function>list_del_rcu()</function>
1624      (<filename>include/linux/list.h</filename>) which does this (the
1625      normal version poisons the old object, which we don't want).
1626    </para>
1627    <para>
1628      The reader must also be careful: some CPUs can look through the
1629      <symbol>next</symbol> pointer to start reading the contents of
1630      the next element early, but don't realize that the pre-fetched
1631      contents is wrong when the <symbol>next</symbol> pointer changes
1632      underneath them.  Once again, there is a
1633      <function>list_for_each_entry_rcu()</function>
1634      (<filename>include/linux/list.h</filename>) to help you.  Of
1635      course, writers can just use
1636      <function>list_for_each_entry()</function>, since there cannot
1637      be two simultaneous writers.
1638    </para>
1639    <para>
1640      Our final dilemma is this: when can we actually destroy the
1641      removed element?  Remember, a reader might be stepping through
1642      this element in the list right now: if we free this element and
1643      the <symbol>next</symbol> pointer changes, the reader will jump
1644      off into garbage and crash.  We need to wait until we know that
1645      all the readers who were traversing the list when we deleted the
1646      element are finished.  We use <function>call_rcu()</function> to
1647      register a callback which will actually destroy the object once
1648      the readers are finished.
1649    </para>
1650    <para>
1651      But how does Read Copy Update know when the readers are
1652      finished?  The method is this: firstly, the readers always
1653      traverse the list inside
1654      <function>rcu_read_lock()</function>/<function>rcu_read_unlock()</function>
1655      pairs: these simply disable preemption so the reader won't go to
1656      sleep while reading the list.
1657    </para>
1658    <para>
1659      RCU then waits until every other CPU has slept at least once:
1660      since readers cannot sleep, we know that any readers which were
1661      traversing the list during the deletion are finished, and the
1662      callback is triggered.  The real Read Copy Update code is a
1663      little more optimized than this, but this is the fundamental
1664      idea.
1665    </para>
1666
1667<programlisting>
1668--- cache.c.perobjectlock       2003-12-11 17:15:03.000000000 +1100
1669+++ cache.c.rcupdate    2003-12-11 17:55:14.000000000 +1100
1670@@ -1,15 +1,18 @@
1671 #include &lt;linux/list.h&gt;
1672 #include &lt;linux/slab.h&gt;
1673 #include &lt;linux/string.h&gt;
1674+#include &lt;linux/rcupdate.h&gt;
1675 #include &lt;linux/mutex.h&gt;
1676 #include &lt;asm/errno.h&gt;
1677
1678 struct object
1679 {
1680-        /* These two protected by cache_lock. */
1681+        /* This is protected by RCU */
1682         struct list_head list;
1683         int popularity;
1684
1685+        struct rcu_head rcu;
1686+
1687         atomic_t refcnt;
1688
1689         /* Doesn't change once created. */
1690@@ -40,7 +43,7 @@
1691 {
1692         struct object *i;
1693
1694-        list_for_each_entry(i, &amp;cache, list) {
1695+        list_for_each_entry_rcu(i, &amp;cache, list) {
1696                 if (i-&gt;id == id) {
1697                         i-&gt;popularity++;
1698                         return i;
1699@@ -49,19 +52,25 @@
1700         return NULL;
1701 }
1702
1703+/* Final discard done once we know no readers are looking. */
1704+static void cache_delete_rcu(void *arg)
1705+{
1706+        object_put(arg);
1707+}
1708+
1709 /* Must be holding cache_lock */
1710 static void __cache_delete(struct object *obj)
1711 {
1712         BUG_ON(!obj);
1713-        list_del(&amp;obj-&gt;list);
1714-        object_put(obj);
1715+        list_del_rcu(&amp;obj-&gt;list);
1716         cache_num--;
1717+        call_rcu(&amp;obj-&gt;rcu, cache_delete_rcu, obj);
1718 }
1719
1720 /* Must be holding cache_lock */
1721 static void __cache_add(struct object *obj)
1722 {
1723-        list_add(&amp;obj-&gt;list, &amp;cache);
1724+        list_add_rcu(&amp;obj-&gt;list, &amp;cache);
1725         if (++cache_num > MAX_CACHE_SIZE) {
1726                 struct object *i, *outcast = NULL;
1727                 list_for_each_entry(i, &amp;cache, list) {
1728@@ -85,6 +94,7 @@
1729         obj-&gt;popularity = 0;
1730         atomic_set(&amp;obj-&gt;refcnt, 1); /* The cache holds a reference */
1731         spin_lock_init(&amp;obj-&gt;lock);
1732+        INIT_RCU_HEAD(&amp;obj-&gt;rcu);
1733
1734         spin_lock_irqsave(&amp;cache_lock, flags);
1735         __cache_add(obj);
1736@@ -104,12 +114,11 @@
1737 struct object *cache_find(int id)
1738 {
1739         struct object *obj;
1740-        unsigned long flags;
1741
1742-        spin_lock_irqsave(&amp;cache_lock, flags);
1743+        rcu_read_lock();
1744         obj = __cache_find(id);
1745         if (obj)
1746                 object_get(obj);
1747-        spin_unlock_irqrestore(&amp;cache_lock, flags);
1748+        rcu_read_unlock();
1749         return obj;
1750 }
1751</programlisting>
1752
1753<para>
1754Note that the reader will alter the
1755<structfield>popularity</structfield> member in
1756<function>__cache_find()</function>, and now it doesn't hold a lock.
1757One solution would be to make it an <type>atomic_t</type>, but for
1758this usage, we don't really care about races: an approximate result is
1759good enough, so I didn't change it.
1760</para>
1761
1762<para>
1763The result is that <function>cache_find()</function> requires no
1764synchronization with any other functions, so is almost as fast on SMP
1765as it would be on UP.
1766</para>
1767
1768<para>
1769There is a furthur optimization possible here: remember our original
1770cache code, where there were no reference counts and the caller simply
1771held the lock whenever using the object?  This is still possible: if
1772you hold the lock, noone can delete the object, so you don't need to
1773get and put the reference count.
1774</para>
1775
1776<para>
1777Now, because the 'read lock' in RCU is simply disabling preemption, a
1778caller which always has preemption disabled between calling
1779<function>cache_find()</function> and
1780<function>object_put()</function> does not need to actually get and
1781put the reference count: we could expose
1782<function>__cache_find()</function> by making it non-static, and
1783such callers could simply call that.
1784</para>
1785<para>
1786The benefit here is that the reference count is not written to: the
1787object is not altered in any way, which is much faster on SMP
1788machines due to caching.
1789</para>
1790  </sect1>
1791
1792   <sect1 id="per-cpu">
1793    <title>Per-CPU Data</title>
1794
1795    <para>
1796      Another technique for avoiding locking which is used fairly
1797      widely is to duplicate information for each CPU.  For example,
1798      if you wanted to keep a count of a common condition, you could
1799      use a spin lock and a single counter.  Nice and simple.
1800    </para>
1801
1802    <para>
1803      If that was too slow (it's usually not, but if you've got a
1804      really big machine to test on and can show that it is), you
1805      could instead use a counter for each CPU, then none of them need
1806      an exclusive lock.  See <function>DEFINE_PER_CPU()</function>,
1807      <function>get_cpu_var()</function> and
1808      <function>put_cpu_var()</function>
1809      (<filename class="headerfile">include/linux/percpu.h</filename>).
1810    </para>
1811
1812    <para>
1813      Of particular use for simple per-cpu counters is the
1814      <type>local_t</type> type, and the
1815      <function>cpu_local_inc()</function> and related functions,
1816      which are more efficient than simple code on some architectures
1817      (<filename class="headerfile">include/asm/local.h</filename>).
1818    </para>
1819
1820    <para>
1821      Note that there is no simple, reliable way of getting an exact
1822      value of such a counter, without introducing more locks.  This
1823      is not a problem for some uses.
1824    </para>
1825   </sect1>
1826
1827   <sect1 id="mostly-hardirq">
1828    <title>Data Which Mostly Used By An IRQ Handler</title>
1829
1830    <para>
1831      If data is always accessed from within the same IRQ handler, you
1832      don't need a lock at all: the kernel already guarantees that the
1833      irq handler will not run simultaneously on multiple CPUs.
1834    </para>
1835    <para>
1836      Manfred Spraul points out that you can still do this, even if
1837      the data is very occasionally accessed in user context or
1838      softirqs/tasklets.  The irq handler doesn't use a lock, and
1839      all other accesses are done as so:
1840    </para>
1841
1842<programlisting>
1843        spin_lock(&amp;lock);
1844        disable_irq(irq);
1845        ...
1846        enable_irq(irq);
1847        spin_unlock(&amp;lock);
1848</programlisting>
1849    <para>
1850      The <function>disable_irq()</function> prevents the irq handler
1851      from running (and waits for it to finish if it's currently
1852      running on other CPUs).  The spinlock prevents any other
1853      accesses happening at the same time.  Naturally, this is slower
1854      than just a <function>spin_lock_irq()</function> call, so it
1855      only makes sense if this type of access happens extremely
1856      rarely.
1857    </para>
1858   </sect1>
1859  </chapter>
1860
1861 <chapter id="sleeping-things">
1862    <title>What Functions Are Safe To Call From Interrupts?</title>
1863
1864    <para>
1865      Many functions in the kernel sleep (ie. call schedule())
1866      directly or indirectly: you can never call them while holding a
1867      spinlock, or with preemption disabled.  This also means you need
1868      to be in user context: calling them from an interrupt is illegal.
1869    </para>
1870
1871   <sect1 id="sleeping">
1872    <title>Some Functions Which Sleep</title>
1873
1874    <para>
1875      The most common ones are listed below, but you usually have to
1876      read the code to find out if other calls are safe.  If everyone
1877      else who calls it can sleep, you probably need to be able to
1878      sleep, too.  In particular, registration and deregistration
1879      functions usually expect to be called from user context, and can
1880      sleep.
1881    </para>
1882
1883    <itemizedlist>
1884     <listitem>
1885      <para>
1886        Accesses to 
1887        <firstterm linkend="gloss-userspace">userspace</firstterm>:
1888      </para>
1889      <itemizedlist>
1890       <listitem>
1891        <para>
1892          <function>copy_from_user()</function>
1893        </para>
1894       </listitem>
1895       <listitem>
1896        <para>
1897          <function>copy_to_user()</function>
1898        </para>
1899       </listitem>
1900       <listitem>
1901        <para>
1902          <function>get_user()</function>
1903        </para>
1904       </listitem>
1905       <listitem>
1906        <para>
1907          <function>put_user()</function>
1908        </para>
1909       </listitem>
1910      </itemizedlist>
1911     </listitem>
1912
1913     <listitem>
1914      <para>
1915        <function>kmalloc(GFP_KERNEL)</function>
1916      </para>
1917     </listitem>
1918
1919     <listitem>
1920      <para>
1921      <function>mutex_lock_interruptible()</function> and
1922      <function>mutex_lock()</function>
1923      </para>
1924      <para>
1925       There is a <function>mutex_trylock()</function> which can be
1926       used inside interrupt context, as it will not sleep.
1927       <function>mutex_unlock()</function> will also never sleep.
1928      </para>
1929     </listitem>
1930    </itemizedlist>
1931   </sect1>
1932
1933   <sect1 id="dont-sleep">
1934    <title>Some Functions Which Don't Sleep</title>
1935
1936    <para>
1937     Some functions are safe to call from any context, or holding
1938     almost any lock.
1939    </para>
1940
1941    <itemizedlist>
1942     <listitem>
1943      <para>
1944        <function>printk()</function>
1945      </para>
1946     </listitem>
1947     <listitem>
1948      <para>
1949        <function>kfree()</function>
1950      </para>
1951     </listitem>
1952     <listitem>
1953      <para>
1954        <function>add_timer()</function> and <function>del_timer()</function>
1955      </para>
1956     </listitem>
1957    </itemizedlist>
1958   </sect1>
1959  </chapter>
1960
1961  <chapter id="references">
1962   <title>Further reading</title>
1963
1964   <itemizedlist>
1965    <listitem>
1966     <para>
1967       <filename>Documentation/spinlocks.txt</filename>: 
1968       Linus Torvalds' spinlocking tutorial in the kernel sources.
1969     </para>
1970    </listitem>
1971
1972    <listitem>
1973     <para>
1974       Unix Systems for Modern Architectures: Symmetric
1975       Multiprocessing and Caching for Kernel Programmers:
1976     </para>
1977
1978     <para>
1979       Curt Schimmel's very good introduction to kernel level
1980       locking (not written for Linux, but nearly everything
1981       applies).  The book is expensive, but really worth every
1982       penny to understand SMP locking. [ISBN: 0201633388]
1983     </para>
1984    </listitem>
1985   </itemizedlist>
1986  </chapter>
1987
1988  <chapter id="thanks">
1989    <title>Thanks</title>
1990
1991    <para>
1992      Thanks to Telsa Gwynne for DocBooking, neatening and adding
1993      style.
1994    </para>
1995
1996    <para>
1997      Thanks to Martin Pool, Philipp Rumpf, Stephen Rothwell, Paul
1998      Mackerras, Ruedi Aschwanden, Alan Cox, Manfred Spraul, Tim
1999      Waugh, Pete Zaitcev, James Morris, Robert Love, Paul McKenney,
2000      John Ashby for proofreading, correcting, flaming, commenting.
2001    </para>
2002
2003    <para>
2004      Thanks to the cabal for having no influence on this document.
2005    </para>
2006  </chapter>
2007
2008  <glossary id="glossary">
2009   <title>Glossary</title>
2010
2011   <glossentry id="gloss-preemption">
2012    <glossterm>preemption</glossterm>
2013     <glossdef>
2014      <para>
2015        Prior to 2.5, or when <symbol>CONFIG_PREEMPT</symbol> is
2016        unset, processes in user context inside the kernel would not
2017        preempt each other (ie. you had that CPU until you gave it up,
2018        except for interrupts).  With the addition of
2019        <symbol>CONFIG_PREEMPT</symbol> in 2.5.4, this changed: when
2020        in user context, higher priority tasks can "cut in": spinlocks
2021        were changed to disable preemption, even on UP.
2022     </para>
2023    </glossdef>
2024   </glossentry>
2025
2026   <glossentry id="gloss-bh">
2027    <glossterm>bh</glossterm>
2028     <glossdef>
2029      <para>
2030        Bottom Half: for historical reasons, functions with
2031        '_bh' in them often now refer to any software interrupt, e.g.
2032        <function>spin_lock_bh()</function> blocks any software interrupt 
2033        on the current CPU.  Bottom halves are deprecated, and will 
2034        eventually be replaced by tasklets.  Only one bottom half will be 
2035        running at any time.
2036     </para>
2037    </glossdef>
2038   </glossentry>
2039
2040   <glossentry id="gloss-hwinterrupt">
2041    <glossterm>Hardware Interrupt / Hardware IRQ</glossterm>
2042    <glossdef>
2043     <para>
2044       Hardware interrupt request.  <function>in_irq()</function> returns 
2045       <returnvalue>true</returnvalue> in a hardware interrupt handler.
2046     </para>
2047    </glossdef>
2048   </glossentry>
2049
2050   <glossentry id="gloss-interruptcontext">
2051    <glossterm>Interrupt Context</glossterm>
2052    <glossdef>
2053     <para>
2054       Not user context: processing a hardware irq or software irq.
2055       Indicated by the <function>in_interrupt()</function> macro 
2056       returning <returnvalue>true</returnvalue>.
2057     </para>
2058    </glossdef>
2059   </glossentry>
2060
2061   <glossentry id="gloss-smp">
2062    <glossterm><acronym>SMP</acronym></glossterm>
2063    <glossdef>
2064     <para>
2065       Symmetric Multi-Processor: kernels compiled for multiple-CPU
2066       machines.  (CONFIG_SMP=y).
2067     </para>
2068    </glossdef>
2069   </glossentry>
2070
2071   <glossentry id="gloss-softirq">
2072    <glossterm>Software Interrupt / softirq</glossterm>
2073    <glossdef>
2074     <para>
2075       Software interrupt handler.  <function>in_irq()</function> returns
2076       <returnvalue>false</returnvalue>; <function>in_softirq()</function>
2077       returns <returnvalue>true</returnvalue>.  Tasklets and softirqs
2078        both fall into the category of 'software interrupts'.
2079     </para>
2080     <para>
2081       Strictly speaking a softirq is one of up to 32 enumerated software
2082       interrupts which can run on multiple CPUs at once.
2083       Sometimes used to refer to tasklets as
2084       well (ie. all software interrupts).
2085     </para>
2086    </glossdef>
2087   </glossentry>
2088
2089   <glossentry id="gloss-tasklet">
2090    <glossterm>tasklet</glossterm>
2091    <glossdef>
2092     <para>
2093       A dynamically-registrable software interrupt,
2094       which is guaranteed to only run on one CPU at a time.
2095     </para>
2096    </glossdef>
2097   </glossentry>
2098
2099   <glossentry id="gloss-timers">
2100    <glossterm>timer</glossterm>
2101    <glossdef>
2102     <para>
2103       A dynamically-registrable software interrupt, which is run at
2104       (or close to) a given time.  When running, it is just like a
2105       tasklet (in fact, they are called from the TIMER_SOFTIRQ).
2106     </para>
2107    </glossdef>
2108   </glossentry>
2109
2110   <glossentry id="gloss-up">
2111    <glossterm><acronym>UP</acronym></glossterm>
2112    <glossdef>
2113     <para>
2114       Uni-Processor: Non-SMP.  (CONFIG_SMP=n).
2115     </para>
2116    </glossdef>
2117   </glossentry>
2118
2119   <glossentry id="gloss-usercontext">
2120    <glossterm>User Context</glossterm>
2121    <glossdef>
2122     <para>
2123       The kernel executing on behalf of a particular process (ie. a
2124       system call or trap) or kernel thread.  You can tell which
2125       process with the <symbol>current</symbol> macro.)  Not to
2126       be confused with userspace.  Can be interrupted by software or
2127       hardware interrupts.
2128     </para>
2129    </glossdef>
2130   </glossentry>
2131
2132   <glossentry id="gloss-userspace">
2133    <glossterm>Userspace</glossterm>
2134    <glossdef>
2135     <para>
2136       A process executing its own code outside the kernel.
2137     </para>
2138    </glossdef>
2139   </glossentry>      
2140
2141  </glossary>
2142</book>
2143
2144