Linux Audio

Check our new training course

Loading...
v5.14.15
   1.. _kernel_hacking_lock:
   2
   3===========================
   4Unreliable Guide To Locking
   5===========================
   6
   7:Author: Rusty Russell
   8
   9Introduction
  10============
  11
  12Welcome, to Rusty's Remarkably Unreliable Guide to Kernel Locking
  13issues. This document describes the locking systems in the Linux Kernel
  14in 2.6.
  15
  16With the wide availability of HyperThreading, and preemption in the
  17Linux Kernel, everyone hacking on the kernel needs to know the
  18fundamentals of concurrency and locking for SMP.
  19
  20The Problem With Concurrency
  21============================
  22
  23(Skip this if you know what a Race Condition is).
  24
  25In a normal program, you can increment a counter like so:
  26
  27::
  28
  29          very_important_count++;
  30
  31
  32This is what they would expect to happen:
  33
  34
  35.. table:: Expected Results
  36
  37  +------------------------------------+------------------------------------+
  38  | Instance 1                         | Instance 2                         |
  39  +====================================+====================================+
  40  | read very_important_count (5)      |                                    |
  41  +------------------------------------+------------------------------------+
  42  | add 1 (6)                          |                                    |
  43  +------------------------------------+------------------------------------+
  44  | write very_important_count (6)     |                                    |
  45  +------------------------------------+------------------------------------+
  46  |                                    | read very_important_count (6)      |
  47  +------------------------------------+------------------------------------+
  48  |                                    | add 1 (7)                          |
  49  +------------------------------------+------------------------------------+
  50  |                                    | write very_important_count (7)     |
  51  +------------------------------------+------------------------------------+
  52
  53This is what might happen:
  54
  55.. table:: Possible Results
  56
  57  +------------------------------------+------------------------------------+
  58  | Instance 1                         | Instance 2                         |
  59  +====================================+====================================+
  60  | read very_important_count (5)      |                                    |
  61  +------------------------------------+------------------------------------+
  62  |                                    | read very_important_count (5)      |
  63  +------------------------------------+------------------------------------+
  64  | add 1 (6)                          |                                    |
  65  +------------------------------------+------------------------------------+
  66  |                                    | add 1 (6)                          |
  67  +------------------------------------+------------------------------------+
  68  | write very_important_count (6)     |                                    |
  69  +------------------------------------+------------------------------------+
  70  |                                    | write very_important_count (6)     |
  71  +------------------------------------+------------------------------------+
  72
  73
  74Race Conditions and Critical Regions
  75------------------------------------
  76
  77This overlap, where the result depends on the relative timing of
  78multiple tasks, is called a race condition. The piece of code containing
  79the concurrency issue is called a critical region. And especially since
  80Linux starting running on SMP machines, they became one of the major
  81issues in kernel design and implementation.
  82
  83Preemption can have the same effect, even if there is only one CPU: by
  84preempting one task during the critical region, we have exactly the same
  85race condition. In this case the thread which preempts might run the
  86critical region itself.
  87
  88The solution is to recognize when these simultaneous accesses occur, and
  89use locks to make sure that only one instance can enter the critical
  90region at any time. There are many friendly primitives in the Linux
  91kernel to help you do this. And then there are the unfriendly
  92primitives, but I'll pretend they don't exist.
  93
  94Locking in the Linux Kernel
  95===========================
  96
  97If I could give you one piece of advice: never sleep with anyone crazier
  98than yourself. But if I had to give you advice on locking: **keep it
  99simple**.
 100
 101Be reluctant to introduce new locks.
 102
 103Strangely enough, this last one is the exact reverse of my advice when
 104you **have** slept with someone crazier than yourself. And you should
 105think about getting a big dog.
 106
 107Two Main Types of Kernel Locks: Spinlocks and Mutexes
 108-----------------------------------------------------
 109
 110There are two main types of kernel locks. The fundamental type is the
 111spinlock (``include/asm/spinlock.h``), which is a very simple
 112single-holder lock: if you can't get the spinlock, you keep trying
 113(spinning) until you can. Spinlocks are very small and fast, and can be
 114used anywhere.
 115
 116The second type is a mutex (``include/linux/mutex.h``): it is like a
 117spinlock, but you may block holding a mutex. If you can't lock a mutex,
 118your task will suspend itself, and be woken up when the mutex is
 119released. This means the CPU can do something else while you are
 120waiting. There are many cases when you simply can't sleep (see
 121`What Functions Are Safe To Call From Interrupts?`_),
 122and so have to use a spinlock instead.
 123
 124Neither type of lock is recursive: see
 125`Deadlock: Simple and Advanced`_.
 126
 127Locks and Uniprocessor Kernels
 128------------------------------
 129
 130For kernels compiled without ``CONFIG_SMP``, and without
 131``CONFIG_PREEMPT`` spinlocks do not exist at all. This is an excellent
 132design decision: when no-one else can run at the same time, there is no
 133reason to have a lock.
 134
 135If the kernel is compiled without ``CONFIG_SMP``, but ``CONFIG_PREEMPT``
 136is set, then spinlocks simply disable preemption, which is sufficient to
 137prevent any races. For most purposes, we can think of preemption as
 138equivalent to SMP, and not worry about it separately.
 139
 140You should always test your locking code with ``CONFIG_SMP`` and
 141``CONFIG_PREEMPT`` enabled, even if you don't have an SMP test box,
 142because it will still catch some kinds of locking bugs.
 143
 144Mutexes still exist, because they are required for synchronization
 145between user contexts, as we will see below.
 146
 147Locking Only In User Context
 148----------------------------
 149
 150If you have a data structure which is only ever accessed from user
 151context, then you can use a simple mutex (``include/linux/mutex.h``) to
 152protect it. This is the most trivial case: you initialize the mutex.
 153Then you can call mutex_lock_interruptible() to grab the
 154mutex, and mutex_unlock() to release it. There is also a
 155mutex_lock(), which should be avoided, because it will
 156not return if a signal is received.
 157
 158Example: ``net/netfilter/nf_sockopt.c`` allows registration of new
 159setsockopt() and getsockopt() calls, with
 160nf_register_sockopt(). Registration and de-registration
 161are only done on module load and unload (and boot time, where there is
 162no concurrency), and the list of registrations is only consulted for an
 163unknown setsockopt() or getsockopt() system
 164call. The ``nf_sockopt_mutex`` is perfect to protect this, especially
 165since the setsockopt and getsockopt calls may well sleep.
 166
 167Locking Between User Context and Softirqs
 168-----------------------------------------
 169
 170If a softirq shares data with user context, you have two problems.
 171Firstly, the current user context can be interrupted by a softirq, and
 172secondly, the critical region could be entered from another CPU. This is
 173where spin_lock_bh() (``include/linux/spinlock.h``) is
 174used. It disables softirqs on that CPU, then grabs the lock.
 175spin_unlock_bh() does the reverse. (The '_bh' suffix is
 176a historical reference to "Bottom Halves", the old name for software
 177interrupts. It should really be called spin_lock_softirq()' in a
 178perfect world).
 179
 180Note that you can also use spin_lock_irq() or
 181spin_lock_irqsave() here, which stop hardware interrupts
 182as well: see `Hard IRQ Context`_.
 183
 184This works perfectly for UP as well: the spin lock vanishes, and this
 185macro simply becomes local_bh_disable()
 186(``include/linux/interrupt.h``), which protects you from the softirq
 187being run.
 188
 189Locking Between User Context and Tasklets
 190-----------------------------------------
 191
 192This is exactly the same as above, because tasklets are actually run
 193from a softirq.
 194
 195Locking Between User Context and Timers
 196---------------------------------------
 197
 198This, too, is exactly the same as above, because timers are actually run
 199from a softirq. From a locking point of view, tasklets and timers are
 200identical.
 201
 202Locking Between Tasklets/Timers
 203-------------------------------
 204
 205Sometimes a tasklet or timer might want to share data with another
 206tasklet or timer.
 207
 208The Same Tasklet/Timer
 209~~~~~~~~~~~~~~~~~~~~~~
 210
 211Since a tasklet is never run on two CPUs at once, you don't need to
 212worry about your tasklet being reentrant (running twice at once), even
 213on SMP.
 214
 215Different Tasklets/Timers
 216~~~~~~~~~~~~~~~~~~~~~~~~~
 217
 218If another tasklet/timer wants to share data with your tasklet or timer
 219, you will both need to use spin_lock() and
 220spin_unlock() calls. spin_lock_bh() is
 221unnecessary here, as you are already in a tasklet, and none will be run
 222on the same CPU.
 223
 224Locking Between Softirqs
 225------------------------
 226
 227Often a softirq might want to share data with itself or a tasklet/timer.
 228
 229The Same Softirq
 230~~~~~~~~~~~~~~~~
 231
 232The same softirq can run on the other CPUs: you can use a per-CPU array
 233(see `Per-CPU Data`_) for better performance. If you're
 234going so far as to use a softirq, you probably care about scalable
 235performance enough to justify the extra complexity.
 236
 237You'll need to use spin_lock() and
 238spin_unlock() for shared data.
 239
 240Different Softirqs
 241~~~~~~~~~~~~~~~~~~
 242
 243You'll need to use spin_lock() and
 244spin_unlock() for shared data, whether it be a timer,
 245tasklet, different softirq or the same or another softirq: any of them
 246could be running on a different CPU.
 247
 248Hard IRQ Context
 249================
 250
 251Hardware interrupts usually communicate with a tasklet or softirq.
 252Frequently this involves putting work in a queue, which the softirq will
 253take out.
 254
 255Locking Between Hard IRQ and Softirqs/Tasklets
 256----------------------------------------------
 257
 258If a hardware irq handler shares data with a softirq, you have two
 259concerns. Firstly, the softirq processing can be interrupted by a
 260hardware interrupt, and secondly, the critical region could be entered
 261by a hardware interrupt on another CPU. This is where
 262spin_lock_irq() is used. It is defined to disable
 263interrupts on that cpu, then grab the lock.
 264spin_unlock_irq() does the reverse.
 265
 266The irq handler does not need to use spin_lock_irq(), because
 267the softirq cannot run while the irq handler is running: it can use
 268spin_lock(), which is slightly faster. The only exception
 269would be if a different hardware irq handler uses the same lock:
 270spin_lock_irq() will stop that from interrupting us.
 271
 272This works perfectly for UP as well: the spin lock vanishes, and this
 273macro simply becomes local_irq_disable()
 274(``include/asm/smp.h``), which protects you from the softirq/tasklet/BH
 275being run.
 276
 277spin_lock_irqsave() (``include/linux/spinlock.h``) is a
 278variant which saves whether interrupts were on or off in a flags word,
 279which is passed to spin_unlock_irqrestore(). This means
 280that the same code can be used inside an hard irq handler (where
 281interrupts are already off) and in softirqs (where the irq disabling is
 282required).
 283
 284Note that softirqs (and hence tasklets and timers) are run on return
 285from hardware interrupts, so spin_lock_irq() also stops
 286these. In that sense, spin_lock_irqsave() is the most
 287general and powerful locking function.
 288
 289Locking Between Two Hard IRQ Handlers
 290-------------------------------------
 291
 292It is rare to have to share data between two IRQ handlers, but if you
 293do, spin_lock_irqsave() should be used: it is
 294architecture-specific whether all interrupts are disabled inside irq
 295handlers themselves.
 296
 297Cheat Sheet For Locking
 298=======================
 299
 300Pete Zaitcev gives the following summary:
 301
 302-  If you are in a process context (any syscall) and want to lock other
 303   process out, use a mutex. You can take a mutex and sleep
 304   (``copy_from_user*(`` or ``kmalloc(x,GFP_KERNEL)``).
 305
 306-  Otherwise (== data can be touched in an interrupt), use
 307   spin_lock_irqsave() and
 308   spin_unlock_irqrestore().
 309
 310-  Avoid holding spinlock for more than 5 lines of code and across any
 311   function call (except accessors like readb()).
 312
 313Table of Minimum Requirements
 314-----------------------------
 315
 316The following table lists the **minimum** locking requirements between
 317various contexts. In some cases, the same context can only be running on
 318one CPU at a time, so no locking is required for that context (eg. a
 319particular thread can only run on one CPU at a time, but if it needs
 320shares data with another thread, locking is required).
 321
 322Remember the advice above: you can always use
 323spin_lock_irqsave(), which is a superset of all other
 324spinlock primitives.
 325
 326============== ============= ============= ========= ========= ========= ========= ======= ======= ============== ==============
 327.              IRQ Handler A IRQ Handler B Softirq A Softirq B Tasklet A Tasklet B Timer A Timer B User Context A User Context B
 328============== ============= ============= ========= ========= ========= ========= ======= ======= ============== ==============
 329IRQ Handler A  None
 330IRQ Handler B  SLIS          None
 331Softirq A      SLI           SLI           SL
 332Softirq B      SLI           SLI           SL        SL
 333Tasklet A      SLI           SLI           SL        SL        None
 334Tasklet B      SLI           SLI           SL        SL        SL        None
 335Timer A        SLI           SLI           SL        SL        SL        SL        None
 336Timer B        SLI           SLI           SL        SL        SL        SL        SL      None
 337User Context A SLI           SLI           SLBH      SLBH      SLBH      SLBH      SLBH    SLBH    None
 338User Context B SLI           SLI           SLBH      SLBH      SLBH      SLBH      SLBH    SLBH    MLI            None
 339============== ============= ============= ========= ========= ========= ========= ======= ======= ============== ==============
 340
 341Table: Table of Locking Requirements
 342
 343+--------+----------------------------+
 344| SLIS   | spin_lock_irqsave          |
 345+--------+----------------------------+
 346| SLI    | spin_lock_irq              |
 347+--------+----------------------------+
 348| SL     | spin_lock                  |
 349+--------+----------------------------+
 350| SLBH   | spin_lock_bh               |
 351+--------+----------------------------+
 352| MLI    | mutex_lock_interruptible   |
 353+--------+----------------------------+
 354
 355Table: Legend for Locking Requirements Table
 356
 357The trylock Functions
 358=====================
 359
 360There are functions that try to acquire a lock only once and immediately
 361return a value telling about success or failure to acquire the lock.
 362They can be used if you need no access to the data protected with the
 363lock when some other thread is holding the lock. You should acquire the
 364lock later if you then need access to the data protected with the lock.
 365
 366spin_trylock() does not spin but returns non-zero if it
 367acquires the spinlock on the first try or 0 if not. This function can be
 368used in all contexts like spin_lock(): you must have
 369disabled the contexts that might interrupt you and acquire the spin
 370lock.
 371
 372mutex_trylock() does not suspend your task but returns
 373non-zero if it could lock the mutex on the first try or 0 if not. This
 374function cannot be safely used in hardware or software interrupt
 375contexts despite not sleeping.
 376
 377Common Examples
 378===============
 379
 380Let's step through a simple example: a cache of number to name mappings.
 381The cache keeps a count of how often each of the objects is used, and
 382when it gets full, throws out the least used one.
 383
 384All In User Context
 385-------------------
 386
 387For our first example, we assume that all operations are in user context
 388(ie. from system calls), so we can sleep. This means we can use a mutex
 389to protect the cache and all the objects within it. Here's the code::
 390
 391    #include <linux/list.h>
 392    #include <linux/slab.h>
 393    #include <linux/string.h>
 394    #include <linux/mutex.h>
 395    #include <asm/errno.h>
 396
 397    struct object
 398    {
 399            struct list_head list;
 400            int id;
 401            char name[32];
 402            int popularity;
 403    };
 404
 405    /* Protects the cache, cache_num, and the objects within it */
 406    static DEFINE_MUTEX(cache_lock);
 407    static LIST_HEAD(cache);
 408    static unsigned int cache_num = 0;
 409    #define MAX_CACHE_SIZE 10
 410
 411    /* Must be holding cache_lock */
 412    static struct object *__cache_find(int id)
 413    {
 414            struct object *i;
 415
 416            list_for_each_entry(i, &cache, list)
 417                    if (i->id == id) {
 418                            i->popularity++;
 419                            return i;
 420                    }
 421            return NULL;
 422    }
 423
 424    /* Must be holding cache_lock */
 425    static void __cache_delete(struct object *obj)
 426    {
 427            BUG_ON(!obj);
 428            list_del(&obj->list);
 429            kfree(obj);
 430            cache_num--;
 431    }
 432
 433    /* Must be holding cache_lock */
 434    static void __cache_add(struct object *obj)
 435    {
 436            list_add(&obj->list, &cache);
 437            if (++cache_num > MAX_CACHE_SIZE) {
 438                    struct object *i, *outcast = NULL;
 439                    list_for_each_entry(i, &cache, list) {
 440                            if (!outcast || i->popularity < outcast->popularity)
 441                                    outcast = i;
 442                    }
 443                    __cache_delete(outcast);
 444            }
 445    }
 446
 447    int cache_add(int id, const char *name)
 448    {
 449            struct object *obj;
 450
 451            if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
 452                    return -ENOMEM;
 453
 454            strscpy(obj->name, name, sizeof(obj->name));
 455            obj->id = id;
 456            obj->popularity = 0;
 457
 458            mutex_lock(&cache_lock);
 459            __cache_add(obj);
 460            mutex_unlock(&cache_lock);
 461            return 0;
 462    }
 463
 464    void cache_delete(int id)
 465    {
 466            mutex_lock(&cache_lock);
 467            __cache_delete(__cache_find(id));
 468            mutex_unlock(&cache_lock);
 469    }
 470
 471    int cache_find(int id, char *name)
 472    {
 473            struct object *obj;
 474            int ret = -ENOENT;
 475
 476            mutex_lock(&cache_lock);
 477            obj = __cache_find(id);
 478            if (obj) {
 479                    ret = 0;
 480                    strcpy(name, obj->name);
 481            }
 482            mutex_unlock(&cache_lock);
 483            return ret;
 484    }
 485
 486Note that we always make sure we have the cache_lock when we add,
 487delete, or look up the cache: both the cache infrastructure itself and
 488the contents of the objects are protected by the lock. In this case it's
 489easy, since we copy the data for the user, and never let them access the
 490objects directly.
 491
 492There is a slight (and common) optimization here: in
 493cache_add() we set up the fields of the object before
 494grabbing the lock. This is safe, as no-one else can access it until we
 495put it in cache.
 496
 497Accessing From Interrupt Context
 498--------------------------------
 499
 500Now consider the case where cache_find() can be called
 501from interrupt context: either a hardware interrupt or a softirq. An
 502example would be a timer which deletes object from the cache.
 503
 504The change is shown below, in standard patch format: the ``-`` are lines
 505which are taken away, and the ``+`` are lines which are added.
 506
 507::
 508
 509    --- cache.c.usercontext 2003-12-09 13:58:54.000000000 +1100
 510    +++ cache.c.interrupt   2003-12-09 14:07:49.000000000 +1100
 511    @@ -12,7 +12,7 @@
 512             int popularity;
 513     };
 514
 515    -static DEFINE_MUTEX(cache_lock);
 516    +static DEFINE_SPINLOCK(cache_lock);
 517     static LIST_HEAD(cache);
 518     static unsigned int cache_num = 0;
 519     #define MAX_CACHE_SIZE 10
 520    @@ -55,6 +55,7 @@
 521     int cache_add(int id, const char *name)
 522     {
 523             struct object *obj;
 524    +        unsigned long flags;
 525
 526             if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
 527                     return -ENOMEM;
 528    @@ -63,30 +64,33 @@
 529             obj->id = id;
 530             obj->popularity = 0;
 531
 532    -        mutex_lock(&cache_lock);
 533    +        spin_lock_irqsave(&cache_lock, flags);
 534             __cache_add(obj);
 535    -        mutex_unlock(&cache_lock);
 536    +        spin_unlock_irqrestore(&cache_lock, flags);
 537             return 0;
 538     }
 539
 540     void cache_delete(int id)
 541     {
 542    -        mutex_lock(&cache_lock);
 543    +        unsigned long flags;
 544    +
 545    +        spin_lock_irqsave(&cache_lock, flags);
 546             __cache_delete(__cache_find(id));
 547    -        mutex_unlock(&cache_lock);
 548    +        spin_unlock_irqrestore(&cache_lock, flags);
 549     }
 550
 551     int cache_find(int id, char *name)
 552     {
 553             struct object *obj;
 554             int ret = -ENOENT;
 555    +        unsigned long flags;
 556
 557    -        mutex_lock(&cache_lock);
 558    +        spin_lock_irqsave(&cache_lock, flags);
 559             obj = __cache_find(id);
 560             if (obj) {
 561                     ret = 0;
 562                     strcpy(name, obj->name);
 563             }
 564    -        mutex_unlock(&cache_lock);
 565    +        spin_unlock_irqrestore(&cache_lock, flags);
 566             return ret;
 567     }
 568
 569Note that the spin_lock_irqsave() will turn off
 570interrupts if they are on, otherwise does nothing (if we are already in
 571an interrupt handler), hence these functions are safe to call from any
 572context.
 573
 574Unfortunately, cache_add() calls kmalloc()
 575with the ``GFP_KERNEL`` flag, which is only legal in user context. I
 576have assumed that cache_add() is still only called in
 577user context, otherwise this should become a parameter to
 578cache_add().
 579
 580Exposing Objects Outside This File
 581----------------------------------
 582
 583If our objects contained more information, it might not be sufficient to
 584copy the information in and out: other parts of the code might want to
 585keep pointers to these objects, for example, rather than looking up the
 586id every time. This produces two problems.
 587
 588The first problem is that we use the ``cache_lock`` to protect objects:
 589we'd need to make this non-static so the rest of the code can use it.
 590This makes locking trickier, as it is no longer all in one place.
 591
 592The second problem is the lifetime problem: if another structure keeps a
 593pointer to an object, it presumably expects that pointer to remain
 594valid. Unfortunately, this is only guaranteed while you hold the lock,
 595otherwise someone might call cache_delete() and even
 596worse, add another object, re-using the same address.
 597
 598As there is only one lock, you can't hold it forever: no-one else would
 599get any work done.
 600
 601The solution to this problem is to use a reference count: everyone who
 602has a pointer to the object increases it when they first get the object,
 603and drops the reference count when they're finished with it. Whoever
 604drops it to zero knows it is unused, and can actually delete it.
 605
 606Here is the code::
 607
 608    --- cache.c.interrupt   2003-12-09 14:25:43.000000000 +1100
 609    +++ cache.c.refcnt  2003-12-09 14:33:05.000000000 +1100
 610    @@ -7,6 +7,7 @@
 611     struct object
 612     {
 613             struct list_head list;
 614    +        unsigned int refcnt;
 615             int id;
 616             char name[32];
 617             int popularity;
 618    @@ -17,6 +18,35 @@
 619     static unsigned int cache_num = 0;
 620     #define MAX_CACHE_SIZE 10
 621
 622    +static void __object_put(struct object *obj)
 623    +{
 624    +        if (--obj->refcnt == 0)
 625    +                kfree(obj);
 626    +}
 627    +
 628    +static void __object_get(struct object *obj)
 629    +{
 630    +        obj->refcnt++;
 631    +}
 632    +
 633    +void object_put(struct object *obj)
 634    +{
 635    +        unsigned long flags;
 636    +
 637    +        spin_lock_irqsave(&cache_lock, flags);
 638    +        __object_put(obj);
 639    +        spin_unlock_irqrestore(&cache_lock, flags);
 640    +}
 641    +
 642    +void object_get(struct object *obj)
 643    +{
 644    +        unsigned long flags;
 645    +
 646    +        spin_lock_irqsave(&cache_lock, flags);
 647    +        __object_get(obj);
 648    +        spin_unlock_irqrestore(&cache_lock, flags);
 649    +}
 650    +
 651     /* Must be holding cache_lock */
 652     static struct object *__cache_find(int id)
 653     {
 654    @@ -35,6 +65,7 @@
 655     {
 656             BUG_ON(!obj);
 657             list_del(&obj->list);
 658    +        __object_put(obj);
 659             cache_num--;
 660     }
 661
 662    @@ -63,6 +94,7 @@
 663             strscpy(obj->name, name, sizeof(obj->name));
 664             obj->id = id;
 665             obj->popularity = 0;
 666    +        obj->refcnt = 1; /* The cache holds a reference */
 667
 668             spin_lock_irqsave(&cache_lock, flags);
 669             __cache_add(obj);
 670    @@ -79,18 +111,15 @@
 671             spin_unlock_irqrestore(&cache_lock, flags);
 672     }
 673
 674    -int cache_find(int id, char *name)
 675    +struct object *cache_find(int id)
 676     {
 677             struct object *obj;
 678    -        int ret = -ENOENT;
 679             unsigned long flags;
 680
 681             spin_lock_irqsave(&cache_lock, flags);
 682             obj = __cache_find(id);
 683    -        if (obj) {
 684    -                ret = 0;
 685    -                strcpy(name, obj->name);
 686    -        }
 687    +        if (obj)
 688    +                __object_get(obj);
 689             spin_unlock_irqrestore(&cache_lock, flags);
 690    -        return ret;
 691    +        return obj;
 692     }
 693
 694We encapsulate the reference counting in the standard 'get' and 'put'
 695functions. Now we can return the object itself from
 696cache_find() which has the advantage that the user can
 697now sleep holding the object (eg. to copy_to_user() to
 698name to userspace).
 699
 700The other point to note is that I said a reference should be held for
 701every pointer to the object: thus the reference count is 1 when first
 702inserted into the cache. In some versions the framework does not hold a
 703reference count, but they are more complicated.
 704
 705Using Atomic Operations For The Reference Count
 706~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 707
 708In practice, :c:type:`atomic_t` would usually be used for refcnt. There are a
 709number of atomic operations defined in ``include/asm/atomic.h``: these
 710are guaranteed to be seen atomically from all CPUs in the system, so no
 711lock is required. In this case, it is simpler than using spinlocks,
 712although for anything non-trivial using spinlocks is clearer. The
 713atomic_inc() and atomic_dec_and_test()
 714are used instead of the standard increment and decrement operators, and
 715the lock is no longer used to protect the reference count itself.
 716
 717::
 718
 719    --- cache.c.refcnt  2003-12-09 15:00:35.000000000 +1100
 720    +++ cache.c.refcnt-atomic   2003-12-11 15:49:42.000000000 +1100
 721    @@ -7,7 +7,7 @@
 722     struct object
 723     {
 724             struct list_head list;
 725    -        unsigned int refcnt;
 726    +        atomic_t refcnt;
 727             int id;
 728             char name[32];
 729             int popularity;
 730    @@ -18,33 +18,15 @@
 731     static unsigned int cache_num = 0;
 732     #define MAX_CACHE_SIZE 10
 733
 734    -static void __object_put(struct object *obj)
 735    -{
 736    -        if (--obj->refcnt == 0)
 737    -                kfree(obj);
 738    -}
 739    -
 740    -static void __object_get(struct object *obj)
 741    -{
 742    -        obj->refcnt++;
 743    -}
 744    -
 745     void object_put(struct object *obj)
 746     {
 747    -        unsigned long flags;
 748    -
 749    -        spin_lock_irqsave(&cache_lock, flags);
 750    -        __object_put(obj);
 751    -        spin_unlock_irqrestore(&cache_lock, flags);
 752    +        if (atomic_dec_and_test(&obj->refcnt))
 753    +                kfree(obj);
 754     }
 755
 756     void object_get(struct object *obj)
 757     {
 758    -        unsigned long flags;
 759    -
 760    -        spin_lock_irqsave(&cache_lock, flags);
 761    -        __object_get(obj);
 762    -        spin_unlock_irqrestore(&cache_lock, flags);
 763    +        atomic_inc(&obj->refcnt);
 764     }
 765
 766     /* Must be holding cache_lock */
 767    @@ -65,7 +47,7 @@
 768     {
 769             BUG_ON(!obj);
 770             list_del(&obj->list);
 771    -        __object_put(obj);
 772    +        object_put(obj);
 773             cache_num--;
 774     }
 775
 776    @@ -94,7 +76,7 @@
 777             strscpy(obj->name, name, sizeof(obj->name));
 778             obj->id = id;
 779             obj->popularity = 0;
 780    -        obj->refcnt = 1; /* The cache holds a reference */
 781    +        atomic_set(&obj->refcnt, 1); /* The cache holds a reference */
 782
 783             spin_lock_irqsave(&cache_lock, flags);
 784             __cache_add(obj);
 785    @@ -119,7 +101,7 @@
 786             spin_lock_irqsave(&cache_lock, flags);
 787             obj = __cache_find(id);
 788             if (obj)
 789    -                __object_get(obj);
 790    +                object_get(obj);
 791             spin_unlock_irqrestore(&cache_lock, flags);
 792             return obj;
 793     }
 794
 795Protecting The Objects Themselves
 796---------------------------------
 797
 798In these examples, we assumed that the objects (except the reference
 799counts) never changed once they are created. If we wanted to allow the
 800name to change, there are three possibilities:
 801
 802-  You can make ``cache_lock`` non-static, and tell people to grab that
 803   lock before changing the name in any object.
 804
 805-  You can provide a cache_obj_rename() which grabs this
 806   lock and changes the name for the caller, and tell everyone to use
 807   that function.
 808
 809-  You can make the ``cache_lock`` protect only the cache itself, and
 810   use another lock to protect the name.
 811
 812Theoretically, you can make the locks as fine-grained as one lock for
 813every field, for every object. In practice, the most common variants
 814are:
 815
 816-  One lock which protects the infrastructure (the ``cache`` list in
 817   this example) and all the objects. This is what we have done so far.
 818
 819-  One lock which protects the infrastructure (including the list
 820   pointers inside the objects), and one lock inside the object which
 821   protects the rest of that object.
 822
 823-  Multiple locks to protect the infrastructure (eg. one lock per hash
 824   chain), possibly with a separate per-object lock.
 825
 826Here is the "lock-per-object" implementation:
 827
 828::
 829
 830    --- cache.c.refcnt-atomic   2003-12-11 15:50:54.000000000 +1100
 831    +++ cache.c.perobjectlock   2003-12-11 17:15:03.000000000 +1100
 832    @@ -6,11 +6,17 @@
 833
 834     struct object
 835     {
 836    +        /* These two protected by cache_lock. */
 837             struct list_head list;
 838    +        int popularity;
 839    +
 840             atomic_t refcnt;
 841    +
 842    +        /* Doesn't change once created. */
 843             int id;
 844    +
 845    +        spinlock_t lock; /* Protects the name */
 846             char name[32];
 847    -        int popularity;
 848     };
 849
 850     static DEFINE_SPINLOCK(cache_lock);
 851    @@ -77,6 +84,7 @@
 852             obj->id = id;
 853             obj->popularity = 0;
 854             atomic_set(&obj->refcnt, 1); /* The cache holds a reference */
 855    +        spin_lock_init(&obj->lock);
 856
 857             spin_lock_irqsave(&cache_lock, flags);
 858             __cache_add(obj);
 859
 860Note that I decide that the popularity count should be protected by the
 861``cache_lock`` rather than the per-object lock: this is because it (like
 862the :c:type:`struct list_head <list_head>` inside the object)
 863is logically part of the infrastructure. This way, I don't need to grab
 864the lock of every object in __cache_add() when seeking
 865the least popular.
 866
 867I also decided that the id member is unchangeable, so I don't need to
 868grab each object lock in __cache_find() to examine the
 869id: the object lock is only used by a caller who wants to read or write
 870the name field.
 871
 872Note also that I added a comment describing what data was protected by
 873which locks. This is extremely important, as it describes the runtime
 874behavior of the code, and can be hard to gain from just reading. And as
 875Alan Cox says, “Lock data, not code”.
 876
 877Common Problems
 878===============
 879
 880Deadlock: Simple and Advanced
 881-----------------------------
 882
 883There is a coding bug where a piece of code tries to grab a spinlock
 884twice: it will spin forever, waiting for the lock to be released
 885(spinlocks, rwlocks and mutexes are not recursive in Linux). This is
 886trivial to diagnose: not a
 887stay-up-five-nights-talk-to-fluffy-code-bunnies kind of problem.
 888
 889For a slightly more complex case, imagine you have a region shared by a
 890softirq and user context. If you use a spin_lock() call
 891to protect it, it is possible that the user context will be interrupted
 892by the softirq while it holds the lock, and the softirq will then spin
 893forever trying to get the same lock.
 894
 895Both of these are called deadlock, and as shown above, it can occur even
 896with a single CPU (although not on UP compiles, since spinlocks vanish
 897on kernel compiles with ``CONFIG_SMP``\ =n. You'll still get data
 898corruption in the second example).
 899
 900This complete lockup is easy to diagnose: on SMP boxes the watchdog
 901timer or compiling with ``DEBUG_SPINLOCK`` set
 902(``include/linux/spinlock.h``) will show this up immediately when it
 903happens.
 904
 905A more complex problem is the so-called 'deadly embrace', involving two
 906or more locks. Say you have a hash table: each entry in the table is a
 907spinlock, and a chain of hashed objects. Inside a softirq handler, you
 908sometimes want to alter an object from one place in the hash to another:
 909you grab the spinlock of the old hash chain and the spinlock of the new
 910hash chain, and delete the object from the old one, and insert it in the
 911new one.
 912
 913There are two problems here. First, if your code ever tries to move the
 914object to the same chain, it will deadlock with itself as it tries to
 915lock it twice. Secondly, if the same softirq on another CPU is trying to
 916move another object in the reverse direction, the following could
 917happen:
 918
 919+-----------------------+-----------------------+
 920| CPU 1                 | CPU 2                 |
 921+=======================+=======================+
 922| Grab lock A -> OK     | Grab lock B -> OK     |
 923+-----------------------+-----------------------+
 924| Grab lock B -> spin   | Grab lock A -> spin   |
 925+-----------------------+-----------------------+
 926
 927Table: Consequences
 928
 929The two CPUs will spin forever, waiting for the other to give up their
 930lock. It will look, smell, and feel like a crash.
 931
 932Preventing Deadlock
 933-------------------
 934
 935Textbooks will tell you that if you always lock in the same order, you
 936will never get this kind of deadlock. Practice will tell you that this
 937approach doesn't scale: when I create a new lock, I don't understand
 938enough of the kernel to figure out where in the 5000 lock hierarchy it
 939will fit.
 940
 941The best locks are encapsulated: they never get exposed in headers, and
 942are never held around calls to non-trivial functions outside the same
 943file. You can read through this code and see that it will never
 944deadlock, because it never tries to grab another lock while it has that
 945one. People using your code don't even need to know you are using a
 946lock.
 947
 948A classic problem here is when you provide callbacks or hooks: if you
 949call these with the lock held, you risk simple deadlock, or a deadly
 950embrace (who knows what the callback will do?). Remember, the other
 951programmers are out to get you, so don't do this.
 952
 953Overzealous Prevention Of Deadlocks
 954~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 955
 956Deadlocks are problematic, but not as bad as data corruption. Code which
 957grabs a read lock, searches a list, fails to find what it wants, drops
 958the read lock, grabs a write lock and inserts the object has a race
 959condition.
 960
 961If you don't see why, please stay away from my code.
 962
 963Racing Timers: A Kernel Pastime
 964-------------------------------
 965
 966Timers can produce their own special problems with races. Consider a
 967collection of objects (list, hash, etc) where each object has a timer
 968which is due to destroy it.
 969
 970If you want to destroy the entire collection (say on module removal),
 971you might do the following::
 972
 973            /* THIS CODE BAD BAD BAD BAD: IF IT WAS ANY WORSE IT WOULD USE
 974               HUNGARIAN NOTATION */
 975            spin_lock_bh(&list_lock);
 976
 977            while (list) {
 978                    struct foo *next = list->next;
 979                    del_timer(&list->timer);
 980                    kfree(list);
 981                    list = next;
 982            }
 983
 984            spin_unlock_bh(&list_lock);
 985
 986
 987Sooner or later, this will crash on SMP, because a timer can have just
 988gone off before the spin_lock_bh(), and it will only get
 989the lock after we spin_unlock_bh(), and then try to free
 990the element (which has already been freed!).
 991
 992This can be avoided by checking the result of
 993del_timer(): if it returns 1, the timer has been deleted.
 994If 0, it means (in this case) that it is currently running, so we can
 995do::
 996
 997            retry:
 998                    spin_lock_bh(&list_lock);
 999
1000                    while (list) {
1001                            struct foo *next = list->next;
1002                            if (!del_timer(&list->timer)) {
1003                                    /* Give timer a chance to delete this */
1004                                    spin_unlock_bh(&list_lock);
1005                                    goto retry;
1006                            }
1007                            kfree(list);
1008                            list = next;
1009                    }
1010
1011                    spin_unlock_bh(&list_lock);
1012
1013
1014Another common problem is deleting timers which restart themselves (by
1015calling add_timer() at the end of their timer function).
1016Because this is a fairly common case which is prone to races, you should
1017use del_timer_sync() (``include/linux/timer.h``) to
1018handle this case. It returns the number of times the timer had to be
1019deleted before we finally stopped it from adding itself back in.
1020
1021Locking Speed
1022=============
1023
1024There are three main things to worry about when considering speed of
1025some code which does locking. First is concurrency: how many things are
1026going to be waiting while someone else is holding a lock. Second is the
1027time taken to actually acquire and release an uncontended lock. Third is
1028using fewer, or smarter locks. I'm assuming that the lock is used fairly
1029often: otherwise, you wouldn't be concerned about efficiency.
1030
1031Concurrency depends on how long the lock is usually held: you should
1032hold the lock for as long as needed, but no longer. In the cache
1033example, we always create the object without the lock held, and then
1034grab the lock only when we are ready to insert it in the list.
1035
1036Acquisition times depend on how much damage the lock operations do to
1037the pipeline (pipeline stalls) and how likely it is that this CPU was
1038the last one to grab the lock (ie. is the lock cache-hot for this CPU):
1039on a machine with more CPUs, this likelihood drops fast. Consider a
1040700MHz Intel Pentium III: an instruction takes about 0.7ns, an atomic
1041increment takes about 58ns, a lock which is cache-hot on this CPU takes
1042160ns, and a cacheline transfer from another CPU takes an additional 170
1043to 360ns. (These figures from Paul McKenney's `Linux Journal RCU
1044article <http://www.linuxjournal.com/article.php?sid=6993>`__).
1045
1046These two aims conflict: holding a lock for a short time might be done
1047by splitting locks into parts (such as in our final per-object-lock
1048example), but this increases the number of lock acquisitions, and the
1049results are often slower than having a single lock. This is another
1050reason to advocate locking simplicity.
1051
1052The third concern is addressed below: there are some methods to reduce
1053the amount of locking which needs to be done.
1054
1055Read/Write Lock Variants
1056------------------------
1057
1058Both spinlocks and mutexes have read/write variants: ``rwlock_t`` and
1059:c:type:`struct rw_semaphore <rw_semaphore>`. These divide
1060users into two classes: the readers and the writers. If you are only
1061reading the data, you can get a read lock, but to write to the data you
1062need the write lock. Many people can hold a read lock, but a writer must
1063be sole holder.
1064
1065If your code divides neatly along reader/writer lines (as our cache code
1066does), and the lock is held by readers for significant lengths of time,
1067using these locks can help. They are slightly slower than the normal
1068locks though, so in practice ``rwlock_t`` is not usually worthwhile.
1069
1070Avoiding Locks: Read Copy Update
1071--------------------------------
1072
1073There is a special method of read/write locking called Read Copy Update.
1074Using RCU, the readers can avoid taking a lock altogether: as we expect
1075our cache to be read more often than updated (otherwise the cache is a
1076waste of time), it is a candidate for this optimization.
1077
1078How do we get rid of read locks? Getting rid of read locks means that
1079writers may be changing the list underneath the readers. That is
1080actually quite simple: we can read a linked list while an element is
1081being added if the writer adds the element very carefully. For example,
1082adding ``new`` to a single linked list called ``list``::
1083
1084            new->next = list->next;
1085            wmb();
1086            list->next = new;
1087
1088
1089The wmb() is a write memory barrier. It ensures that the
1090first operation (setting the new element's ``next`` pointer) is complete
1091and will be seen by all CPUs, before the second operation is (putting
1092the new element into the list). This is important, since modern
1093compilers and modern CPUs can both reorder instructions unless told
1094otherwise: we want a reader to either not see the new element at all, or
1095see the new element with the ``next`` pointer correctly pointing at the
1096rest of the list.
1097
1098Fortunately, there is a function to do this for standard
1099:c:type:`struct list_head <list_head>` lists:
1100list_add_rcu() (``include/linux/list.h``).
1101
1102Removing an element from the list is even simpler: we replace the
1103pointer to the old element with a pointer to its successor, and readers
1104will either see it, or skip over it.
1105
1106::
1107
1108            list->next = old->next;
1109
1110
1111There is list_del_rcu() (``include/linux/list.h``) which
1112does this (the normal version poisons the old object, which we don't
1113want).
1114
1115The reader must also be careful: some CPUs can look through the ``next``
1116pointer to start reading the contents of the next element early, but
1117don't realize that the pre-fetched contents is wrong when the ``next``
1118pointer changes underneath them. Once again, there is a
1119list_for_each_entry_rcu() (``include/linux/list.h``)
1120to help you. Of course, writers can just use
1121list_for_each_entry(), since there cannot be two
1122simultaneous writers.
1123
1124Our final dilemma is this: when can we actually destroy the removed
1125element? Remember, a reader might be stepping through this element in
1126the list right now: if we free this element and the ``next`` pointer
1127changes, the reader will jump off into garbage and crash. We need to
1128wait until we know that all the readers who were traversing the list
1129when we deleted the element are finished. We use
1130call_rcu() to register a callback which will actually
1131destroy the object once all pre-existing readers are finished.
1132Alternatively, synchronize_rcu() may be used to block
1133until all pre-existing are finished.
1134
1135But how does Read Copy Update know when the readers are finished? The
1136method is this: firstly, the readers always traverse the list inside
1137rcu_read_lock()/rcu_read_unlock() pairs:
1138these simply disable preemption so the reader won't go to sleep while
1139reading the list.
1140
1141RCU then waits until every other CPU has slept at least once: since
1142readers cannot sleep, we know that any readers which were traversing the
1143list during the deletion are finished, and the callback is triggered.
1144The real Read Copy Update code is a little more optimized than this, but
1145this is the fundamental idea.
1146
1147::
1148
1149    --- cache.c.perobjectlock   2003-12-11 17:15:03.000000000 +1100
1150    +++ cache.c.rcupdate    2003-12-11 17:55:14.000000000 +1100
1151    @@ -1,15 +1,18 @@
1152     #include <linux/list.h>
1153     #include <linux/slab.h>
1154     #include <linux/string.h>
1155    +#include <linux/rcupdate.h>
1156     #include <linux/mutex.h>
1157     #include <asm/errno.h>
1158
1159     struct object
1160     {
1161    -        /* These two protected by cache_lock. */
1162    +        /* This is protected by RCU */
1163             struct list_head list;
1164             int popularity;
1165
1166    +        struct rcu_head rcu;
1167    +
1168             atomic_t refcnt;
1169
1170             /* Doesn't change once created. */
1171    @@ -40,7 +43,7 @@
1172     {
1173             struct object *i;
1174
1175    -        list_for_each_entry(i, &cache, list) {
1176    +        list_for_each_entry_rcu(i, &cache, list) {
1177                     if (i->id == id) {
1178                             i->popularity++;
1179                             return i;
1180    @@ -49,19 +52,25 @@
1181             return NULL;
1182     }
1183
1184    +/* Final discard done once we know no readers are looking. */
1185    +static void cache_delete_rcu(void *arg)
1186    +{
1187    +        object_put(arg);
1188    +}
1189    +
1190     /* Must be holding cache_lock */
1191     static void __cache_delete(struct object *obj)
1192     {
1193             BUG_ON(!obj);
1194    -        list_del(&obj->list);
1195    -        object_put(obj);
1196    +        list_del_rcu(&obj->list);
1197             cache_num--;
1198    +        call_rcu(&obj->rcu, cache_delete_rcu);
1199     }
1200
1201     /* Must be holding cache_lock */
1202     static void __cache_add(struct object *obj)
1203     {
1204    -        list_add(&obj->list, &cache);
1205    +        list_add_rcu(&obj->list, &cache);
1206             if (++cache_num > MAX_CACHE_SIZE) {
1207                     struct object *i, *outcast = NULL;
1208                     list_for_each_entry(i, &cache, list) {
1209    @@ -104,12 +114,11 @@
1210     struct object *cache_find(int id)
1211     {
1212             struct object *obj;
1213    -        unsigned long flags;
1214
1215    -        spin_lock_irqsave(&cache_lock, flags);
1216    +        rcu_read_lock();
1217             obj = __cache_find(id);
1218             if (obj)
1219                     object_get(obj);
1220    -        spin_unlock_irqrestore(&cache_lock, flags);
1221    +        rcu_read_unlock();
1222             return obj;
1223     }
1224
1225Note that the reader will alter the popularity member in
1226__cache_find(), and now it doesn't hold a lock. One
1227solution would be to make it an ``atomic_t``, but for this usage, we
1228don't really care about races: an approximate result is good enough, so
1229I didn't change it.
1230
1231The result is that cache_find() requires no
1232synchronization with any other functions, so is almost as fast on SMP as
1233it would be on UP.
1234
1235There is a further optimization possible here: remember our original
1236cache code, where there were no reference counts and the caller simply
1237held the lock whenever using the object? This is still possible: if you
1238hold the lock, no one can delete the object, so you don't need to get
1239and put the reference count.
1240
1241Now, because the 'read lock' in RCU is simply disabling preemption, a
1242caller which always has preemption disabled between calling
1243cache_find() and object_put() does not
1244need to actually get and put the reference count: we could expose
1245__cache_find() by making it non-static, and such
1246callers could simply call that.
1247
1248The benefit here is that the reference count is not written to: the
1249object is not altered in any way, which is much faster on SMP machines
1250due to caching.
1251
1252Per-CPU Data
1253------------
1254
1255Another technique for avoiding locking which is used fairly widely is to
1256duplicate information for each CPU. For example, if you wanted to keep a
1257count of a common condition, you could use a spin lock and a single
1258counter. Nice and simple.
1259
1260If that was too slow (it's usually not, but if you've got a really big
1261machine to test on and can show that it is), you could instead use a
1262counter for each CPU, then none of them need an exclusive lock. See
1263DEFINE_PER_CPU(), get_cpu_var() and
1264put_cpu_var() (``include/linux/percpu.h``).
1265
1266Of particular use for simple per-cpu counters is the ``local_t`` type,
1267and the cpu_local_inc() and related functions, which are
1268more efficient than simple code on some architectures
1269(``include/asm/local.h``).
1270
1271Note that there is no simple, reliable way of getting an exact value of
1272such a counter, without introducing more locks. This is not a problem
1273for some uses.
1274
1275Data Which Mostly Used By An IRQ Handler
1276----------------------------------------
1277
1278If data is always accessed from within the same IRQ handler, you don't
1279need a lock at all: the kernel already guarantees that the irq handler
1280will not run simultaneously on multiple CPUs.
1281
1282Manfred Spraul points out that you can still do this, even if the data
1283is very occasionally accessed in user context or softirqs/tasklets. The
1284irq handler doesn't use a lock, and all other accesses are done as so::
1285
1286        spin_lock(&lock);
1287        disable_irq(irq);
1288        ...
1289        enable_irq(irq);
1290        spin_unlock(&lock);
1291
1292The disable_irq() prevents the irq handler from running
1293(and waits for it to finish if it's currently running on other CPUs).
1294The spinlock prevents any other accesses happening at the same time.
1295Naturally, this is slower than just a spin_lock_irq()
1296call, so it only makes sense if this type of access happens extremely
1297rarely.
1298
1299What Functions Are Safe To Call From Interrupts?
1300================================================
1301
1302Many functions in the kernel sleep (ie. call schedule()) directly or
1303indirectly: you can never call them while holding a spinlock, or with
1304preemption disabled. This also means you need to be in user context:
1305calling them from an interrupt is illegal.
1306
1307Some Functions Which Sleep
1308--------------------------
1309
1310The most common ones are listed below, but you usually have to read the
1311code to find out if other calls are safe. If everyone else who calls it
1312can sleep, you probably need to be able to sleep, too. In particular,
1313registration and deregistration functions usually expect to be called
1314from user context, and can sleep.
1315
1316-  Accesses to userspace:
1317
1318   -  copy_from_user()
1319
1320   -  copy_to_user()
1321
1322   -  get_user()
1323
1324   -  put_user()
1325
1326-  kmalloc(GP_KERNEL) <kmalloc>`
1327
1328-  mutex_lock_interruptible() and
1329   mutex_lock()
1330
1331   There is a mutex_trylock() which does not sleep.
1332   Still, it must not be used inside interrupt context since its
1333   implementation is not safe for that. mutex_unlock()
1334   will also never sleep. It cannot be used in interrupt context either
1335   since a mutex must be released by the same task that acquired it.
1336
1337Some Functions Which Don't Sleep
1338--------------------------------
1339
1340Some functions are safe to call from any context, or holding almost any
1341lock.
1342
1343-  printk()
1344
1345-  kfree()
1346
1347-  add_timer() and del_timer()
1348
1349Mutex API reference
1350===================
1351
1352.. kernel-doc:: include/linux/mutex.h
1353   :internal:
1354
1355.. kernel-doc:: kernel/locking/mutex.c
1356   :export:
1357
1358Futex API reference
1359===================
1360
1361.. kernel-doc:: kernel/futex.c
1362   :internal:
1363
1364Further reading
1365===============
1366
1367-  ``Documentation/locking/spinlocks.rst``: Linus Torvalds' spinlocking
1368   tutorial in the kernel sources.
1369
1370-  Unix Systems for Modern Architectures: Symmetric Multiprocessing and
1371   Caching for Kernel Programmers:
1372
1373   Curt Schimmel's very good introduction to kernel level locking (not
1374   written for Linux, but nearly everything applies). The book is
1375   expensive, but really worth every penny to understand SMP locking.
1376   [ISBN: 0201633388]
1377
1378Thanks
1379======
1380
1381Thanks to Telsa Gwynne for DocBooking, neatening and adding style.
1382
1383Thanks to Martin Pool, Philipp Rumpf, Stephen Rothwell, Paul Mackerras,
1384Ruedi Aschwanden, Alan Cox, Manfred Spraul, Tim Waugh, Pete Zaitcev,
1385James Morris, Robert Love, Paul McKenney, John Ashby for proofreading,
1386correcting, flaming, commenting.
1387
1388Thanks to the cabal for having no influence on this document.
1389
1390Glossary
1391========
1392
1393preemption
1394  Prior to 2.5, or when ``CONFIG_PREEMPT`` is unset, processes in user
1395  context inside the kernel would not preempt each other (ie. you had that
1396  CPU until you gave it up, except for interrupts). With the addition of
1397  ``CONFIG_PREEMPT`` in 2.5.4, this changed: when in user context, higher
1398  priority tasks can "cut in": spinlocks were changed to disable
1399  preemption, even on UP.
1400
1401bh
1402  Bottom Half: for historical reasons, functions with '_bh' in them often
1403  now refer to any software interrupt, e.g. spin_lock_bh()
1404  blocks any software interrupt on the current CPU. Bottom halves are
1405  deprecated, and will eventually be replaced by tasklets. Only one bottom
1406  half will be running at any time.
1407
1408Hardware Interrupt / Hardware IRQ
1409  Hardware interrupt request. in_irq() returns true in a
1410  hardware interrupt handler.
1411
1412Interrupt Context
1413  Not user context: processing a hardware irq or software irq. Indicated
1414  by the in_interrupt() macro returning true.
1415
1416SMP
1417  Symmetric Multi-Processor: kernels compiled for multiple-CPU machines.
1418  (``CONFIG_SMP=y``).
1419
1420Software Interrupt / softirq
1421  Software interrupt handler. in_irq() returns false;
1422  in_softirq() returns true. Tasklets and softirqs both
1423  fall into the category of 'software interrupts'.
1424
1425  Strictly speaking a softirq is one of up to 32 enumerated software
1426  interrupts which can run on multiple CPUs at once. Sometimes used to
1427  refer to tasklets as well (ie. all software interrupts).
1428
1429tasklet
1430  A dynamically-registrable software interrupt, which is guaranteed to
1431  only run on one CPU at a time.
1432
1433timer
1434  A dynamically-registrable software interrupt, which is run at (or close
1435  to) a given time. When running, it is just like a tasklet (in fact, they
1436  are called from the ``TIMER_SOFTIRQ``).
1437
1438UP
1439  Uni-Processor: Non-SMP. (``CONFIG_SMP=n``).
1440
1441User Context
1442  The kernel executing on behalf of a particular process (ie. a system
1443  call or trap) or kernel thread. You can tell which process with the
1444  ``current`` macro.) Not to be confused with userspace. Can be
1445  interrupted by software or hardware interrupts.
1446
1447Userspace
1448  A process executing its own code outside the kernel.
v4.17
 
 
   1===========================
   2Unreliable Guide To Locking
   3===========================
   4
   5:Author: Rusty Russell
   6
   7Introduction
   8============
   9
  10Welcome, to Rusty's Remarkably Unreliable Guide to Kernel Locking
  11issues. This document describes the locking systems in the Linux Kernel
  12in 2.6.
  13
  14With the wide availability of HyperThreading, and preemption in the
  15Linux Kernel, everyone hacking on the kernel needs to know the
  16fundamentals of concurrency and locking for SMP.
  17
  18The Problem With Concurrency
  19============================
  20
  21(Skip this if you know what a Race Condition is).
  22
  23In a normal program, you can increment a counter like so:
  24
  25::
  26
  27          very_important_count++;
  28
  29
  30This is what they would expect to happen:
  31
  32
  33.. table:: Expected Results
  34
  35  +------------------------------------+------------------------------------+
  36  | Instance 1                         | Instance 2                         |
  37  +====================================+====================================+
  38  | read very_important_count (5)      |                                    |
  39  +------------------------------------+------------------------------------+
  40  | add 1 (6)                          |                                    |
  41  +------------------------------------+------------------------------------+
  42  | write very_important_count (6)     |                                    |
  43  +------------------------------------+------------------------------------+
  44  |                                    | read very_important_count (6)      |
  45  +------------------------------------+------------------------------------+
  46  |                                    | add 1 (7)                          |
  47  +------------------------------------+------------------------------------+
  48  |                                    | write very_important_count (7)     |
  49  +------------------------------------+------------------------------------+
  50
  51This is what might happen:
  52
  53.. table:: Possible Results
  54
  55  +------------------------------------+------------------------------------+
  56  | Instance 1                         | Instance 2                         |
  57  +====================================+====================================+
  58  | read very_important_count (5)      |                                    |
  59  +------------------------------------+------------------------------------+
  60  |                                    | read very_important_count (5)      |
  61  +------------------------------------+------------------------------------+
  62  | add 1 (6)                          |                                    |
  63  +------------------------------------+------------------------------------+
  64  |                                    | add 1 (6)                          |
  65  +------------------------------------+------------------------------------+
  66  | write very_important_count (6)     |                                    |
  67  +------------------------------------+------------------------------------+
  68  |                                    | write very_important_count (6)     |
  69  +------------------------------------+------------------------------------+
  70
  71
  72Race Conditions and Critical Regions
  73------------------------------------
  74
  75This overlap, where the result depends on the relative timing of
  76multiple tasks, is called a race condition. The piece of code containing
  77the concurrency issue is called a critical region. And especially since
  78Linux starting running on SMP machines, they became one of the major
  79issues in kernel design and implementation.
  80
  81Preemption can have the same effect, even if there is only one CPU: by
  82preempting one task during the critical region, we have exactly the same
  83race condition. In this case the thread which preempts might run the
  84critical region itself.
  85
  86The solution is to recognize when these simultaneous accesses occur, and
  87use locks to make sure that only one instance can enter the critical
  88region at any time. There are many friendly primitives in the Linux
  89kernel to help you do this. And then there are the unfriendly
  90primitives, but I'll pretend they don't exist.
  91
  92Locking in the Linux Kernel
  93===========================
  94
  95If I could give you one piece of advice: never sleep with anyone crazier
  96than yourself. But if I had to give you advice on locking: **keep it
  97simple**.
  98
  99Be reluctant to introduce new locks.
 100
 101Strangely enough, this last one is the exact reverse of my advice when
 102you **have** slept with someone crazier than yourself. And you should
 103think about getting a big dog.
 104
 105Two Main Types of Kernel Locks: Spinlocks and Mutexes
 106-----------------------------------------------------
 107
 108There are two main types of kernel locks. The fundamental type is the
 109spinlock (``include/asm/spinlock.h``), which is a very simple
 110single-holder lock: if you can't get the spinlock, you keep trying
 111(spinning) until you can. Spinlocks are very small and fast, and can be
 112used anywhere.
 113
 114The second type is a mutex (``include/linux/mutex.h``): it is like a
 115spinlock, but you may block holding a mutex. If you can't lock a mutex,
 116your task will suspend itself, and be woken up when the mutex is
 117released. This means the CPU can do something else while you are
 118waiting. There are many cases when you simply can't sleep (see
 119`What Functions Are Safe To Call From Interrupts? <#sleeping-things>`__),
 120and so have to use a spinlock instead.
 121
 122Neither type of lock is recursive: see
 123`Deadlock: Simple and Advanced <#deadlock>`__.
 124
 125Locks and Uniprocessor Kernels
 126------------------------------
 127
 128For kernels compiled without ``CONFIG_SMP``, and without
 129``CONFIG_PREEMPT`` spinlocks do not exist at all. This is an excellent
 130design decision: when no-one else can run at the same time, there is no
 131reason to have a lock.
 132
 133If the kernel is compiled without ``CONFIG_SMP``, but ``CONFIG_PREEMPT``
 134is set, then spinlocks simply disable preemption, which is sufficient to
 135prevent any races. For most purposes, we can think of preemption as
 136equivalent to SMP, and not worry about it separately.
 137
 138You should always test your locking code with ``CONFIG_SMP`` and
 139``CONFIG_PREEMPT`` enabled, even if you don't have an SMP test box,
 140because it will still catch some kinds of locking bugs.
 141
 142Mutexes still exist, because they are required for synchronization
 143between user contexts, as we will see below.
 144
 145Locking Only In User Context
 146----------------------------
 147
 148If you have a data structure which is only ever accessed from user
 149context, then you can use a simple mutex (``include/linux/mutex.h``) to
 150protect it. This is the most trivial case: you initialize the mutex.
 151Then you can call :c:func:`mutex_lock_interruptible()` to grab the
 152mutex, and :c:func:`mutex_unlock()` to release it. There is also a
 153:c:func:`mutex_lock()`, which should be avoided, because it will
 154not return if a signal is received.
 155
 156Example: ``net/netfilter/nf_sockopt.c`` allows registration of new
 157:c:func:`setsockopt()` and :c:func:`getsockopt()` calls, with
 158:c:func:`nf_register_sockopt()`. Registration and de-registration
 159are only done on module load and unload (and boot time, where there is
 160no concurrency), and the list of registrations is only consulted for an
 161unknown :c:func:`setsockopt()` or :c:func:`getsockopt()` system
 162call. The ``nf_sockopt_mutex`` is perfect to protect this, especially
 163since the setsockopt and getsockopt calls may well sleep.
 164
 165Locking Between User Context and Softirqs
 166-----------------------------------------
 167
 168If a softirq shares data with user context, you have two problems.
 169Firstly, the current user context can be interrupted by a softirq, and
 170secondly, the critical region could be entered from another CPU. This is
 171where :c:func:`spin_lock_bh()` (``include/linux/spinlock.h``) is
 172used. It disables softirqs on that CPU, then grabs the lock.
 173:c:func:`spin_unlock_bh()` does the reverse. (The '_bh' suffix is
 174a historical reference to "Bottom Halves", the old name for software
 175interrupts. It should really be called spin_lock_softirq()' in a
 176perfect world).
 177
 178Note that you can also use :c:func:`spin_lock_irq()` or
 179:c:func:`spin_lock_irqsave()` here, which stop hardware interrupts
 180as well: see `Hard IRQ Context <#hardirq-context>`__.
 181
 182This works perfectly for UP as well: the spin lock vanishes, and this
 183macro simply becomes :c:func:`local_bh_disable()`
 184(``include/linux/interrupt.h``), which protects you from the softirq
 185being run.
 186
 187Locking Between User Context and Tasklets
 188-----------------------------------------
 189
 190This is exactly the same as above, because tasklets are actually run
 191from a softirq.
 192
 193Locking Between User Context and Timers
 194---------------------------------------
 195
 196This, too, is exactly the same as above, because timers are actually run
 197from a softirq. From a locking point of view, tasklets and timers are
 198identical.
 199
 200Locking Between Tasklets/Timers
 201-------------------------------
 202
 203Sometimes a tasklet or timer might want to share data with another
 204tasklet or timer.
 205
 206The Same Tasklet/Timer
 207~~~~~~~~~~~~~~~~~~~~~~
 208
 209Since a tasklet is never run on two CPUs at once, you don't need to
 210worry about your tasklet being reentrant (running twice at once), even
 211on SMP.
 212
 213Different Tasklets/Timers
 214~~~~~~~~~~~~~~~~~~~~~~~~~
 215
 216If another tasklet/timer wants to share data with your tasklet or timer
 217, you will both need to use :c:func:`spin_lock()` and
 218:c:func:`spin_unlock()` calls. :c:func:`spin_lock_bh()` is
 219unnecessary here, as you are already in a tasklet, and none will be run
 220on the same CPU.
 221
 222Locking Between Softirqs
 223------------------------
 224
 225Often a softirq might want to share data with itself or a tasklet/timer.
 226
 227The Same Softirq
 228~~~~~~~~~~~~~~~~
 229
 230The same softirq can run on the other CPUs: you can use a per-CPU array
 231(see `Per-CPU Data <#per-cpu>`__) for better performance. If you're
 232going so far as to use a softirq, you probably care about scalable
 233performance enough to justify the extra complexity.
 234
 235You'll need to use :c:func:`spin_lock()` and
 236:c:func:`spin_unlock()` for shared data.
 237
 238Different Softirqs
 239~~~~~~~~~~~~~~~~~~
 240
 241You'll need to use :c:func:`spin_lock()` and
 242:c:func:`spin_unlock()` for shared data, whether it be a timer,
 243tasklet, different softirq or the same or another softirq: any of them
 244could be running on a different CPU.
 245
 246Hard IRQ Context
 247================
 248
 249Hardware interrupts usually communicate with a tasklet or softirq.
 250Frequently this involves putting work in a queue, which the softirq will
 251take out.
 252
 253Locking Between Hard IRQ and Softirqs/Tasklets
 254----------------------------------------------
 255
 256If a hardware irq handler shares data with a softirq, you have two
 257concerns. Firstly, the softirq processing can be interrupted by a
 258hardware interrupt, and secondly, the critical region could be entered
 259by a hardware interrupt on another CPU. This is where
 260:c:func:`spin_lock_irq()` is used. It is defined to disable
 261interrupts on that cpu, then grab the lock.
 262:c:func:`spin_unlock_irq()` does the reverse.
 263
 264The irq handler does not to use :c:func:`spin_lock_irq()`, because
 265the softirq cannot run while the irq handler is running: it can use
 266:c:func:`spin_lock()`, which is slightly faster. The only exception
 267would be if a different hardware irq handler uses the same lock:
 268:c:func:`spin_lock_irq()` will stop that from interrupting us.
 269
 270This works perfectly for UP as well: the spin lock vanishes, and this
 271macro simply becomes :c:func:`local_irq_disable()`
 272(``include/asm/smp.h``), which protects you from the softirq/tasklet/BH
 273being run.
 274
 275:c:func:`spin_lock_irqsave()` (``include/linux/spinlock.h``) is a
 276variant which saves whether interrupts were on or off in a flags word,
 277which is passed to :c:func:`spin_unlock_irqrestore()`. This means
 278that the same code can be used inside an hard irq handler (where
 279interrupts are already off) and in softirqs (where the irq disabling is
 280required).
 281
 282Note that softirqs (and hence tasklets and timers) are run on return
 283from hardware interrupts, so :c:func:`spin_lock_irq()` also stops
 284these. In that sense, :c:func:`spin_lock_irqsave()` is the most
 285general and powerful locking function.
 286
 287Locking Between Two Hard IRQ Handlers
 288-------------------------------------
 289
 290It is rare to have to share data between two IRQ handlers, but if you
 291do, :c:func:`spin_lock_irqsave()` should be used: it is
 292architecture-specific whether all interrupts are disabled inside irq
 293handlers themselves.
 294
 295Cheat Sheet For Locking
 296=======================
 297
 298Pete Zaitcev gives the following summary:
 299
 300-  If you are in a process context (any syscall) and want to lock other
 301   process out, use a mutex. You can take a mutex and sleep
 302   (``copy_from_user*(`` or ``kmalloc(x,GFP_KERNEL)``).
 303
 304-  Otherwise (== data can be touched in an interrupt), use
 305   :c:func:`spin_lock_irqsave()` and
 306   :c:func:`spin_unlock_irqrestore()`.
 307
 308-  Avoid holding spinlock for more than 5 lines of code and across any
 309   function call (except accessors like :c:func:`readb()`).
 310
 311Table of Minimum Requirements
 312-----------------------------
 313
 314The following table lists the **minimum** locking requirements between
 315various contexts. In some cases, the same context can only be running on
 316one CPU at a time, so no locking is required for that context (eg. a
 317particular thread can only run on one CPU at a time, but if it needs
 318shares data with another thread, locking is required).
 319
 320Remember the advice above: you can always use
 321:c:func:`spin_lock_irqsave()`, which is a superset of all other
 322spinlock primitives.
 323
 324============== ============= ============= ========= ========= ========= ========= ======= ======= ============== ==============
 325.              IRQ Handler A IRQ Handler B Softirq A Softirq B Tasklet A Tasklet B Timer A Timer B User Context A User Context B
 326============== ============= ============= ========= ========= ========= ========= ======= ======= ============== ==============
 327IRQ Handler A  None
 328IRQ Handler B  SLIS          None
 329Softirq A      SLI           SLI           SL
 330Softirq B      SLI           SLI           SL        SL
 331Tasklet A      SLI           SLI           SL        SL        None
 332Tasklet B      SLI           SLI           SL        SL        SL        None
 333Timer A        SLI           SLI           SL        SL        SL        SL        None
 334Timer B        SLI           SLI           SL        SL        SL        SL        SL      None
 335User Context A SLI           SLI           SLBH      SLBH      SLBH      SLBH      SLBH    SLBH    None
 336User Context B SLI           SLI           SLBH      SLBH      SLBH      SLBH      SLBH    SLBH    MLI            None
 337============== ============= ============= ========= ========= ========= ========= ======= ======= ============== ==============
 338
 339Table: Table of Locking Requirements
 340
 341+--------+----------------------------+
 342| SLIS   | spin_lock_irqsave          |
 343+--------+----------------------------+
 344| SLI    | spin_lock_irq              |
 345+--------+----------------------------+
 346| SL     | spin_lock                  |
 347+--------+----------------------------+
 348| SLBH   | spin_lock_bh               |
 349+--------+----------------------------+
 350| MLI    | mutex_lock_interruptible   |
 351+--------+----------------------------+
 352
 353Table: Legend for Locking Requirements Table
 354
 355The trylock Functions
 356=====================
 357
 358There are functions that try to acquire a lock only once and immediately
 359return a value telling about success or failure to acquire the lock.
 360They can be used if you need no access to the data protected with the
 361lock when some other thread is holding the lock. You should acquire the
 362lock later if you then need access to the data protected with the lock.
 363
 364:c:func:`spin_trylock()` does not spin but returns non-zero if it
 365acquires the spinlock on the first try or 0 if not. This function can be
 366used in all contexts like :c:func:`spin_lock()`: you must have
 367disabled the contexts that might interrupt you and acquire the spin
 368lock.
 369
 370:c:func:`mutex_trylock()` does not suspend your task but returns
 371non-zero if it could lock the mutex on the first try or 0 if not. This
 372function cannot be safely used in hardware or software interrupt
 373contexts despite not sleeping.
 374
 375Common Examples
 376===============
 377
 378Let's step through a simple example: a cache of number to name mappings.
 379The cache keeps a count of how often each of the objects is used, and
 380when it gets full, throws out the least used one.
 381
 382All In User Context
 383-------------------
 384
 385For our first example, we assume that all operations are in user context
 386(ie. from system calls), so we can sleep. This means we can use a mutex
 387to protect the cache and all the objects within it. Here's the code::
 388
 389    #include <linux/list.h>
 390    #include <linux/slab.h>
 391    #include <linux/string.h>
 392    #include <linux/mutex.h>
 393    #include <asm/errno.h>
 394
 395    struct object
 396    {
 397            struct list_head list;
 398            int id;
 399            char name[32];
 400            int popularity;
 401    };
 402
 403    /* Protects the cache, cache_num, and the objects within it */
 404    static DEFINE_MUTEX(cache_lock);
 405    static LIST_HEAD(cache);
 406    static unsigned int cache_num = 0;
 407    #define MAX_CACHE_SIZE 10
 408
 409    /* Must be holding cache_lock */
 410    static struct object *__cache_find(int id)
 411    {
 412            struct object *i;
 413
 414            list_for_each_entry(i, &cache, list)
 415                    if (i->id == id) {
 416                            i->popularity++;
 417                            return i;
 418                    }
 419            return NULL;
 420    }
 421
 422    /* Must be holding cache_lock */
 423    static void __cache_delete(struct object *obj)
 424    {
 425            BUG_ON(!obj);
 426            list_del(&obj->list);
 427            kfree(obj);
 428            cache_num--;
 429    }
 430
 431    /* Must be holding cache_lock */
 432    static void __cache_add(struct object *obj)
 433    {
 434            list_add(&obj->list, &cache);
 435            if (++cache_num > MAX_CACHE_SIZE) {
 436                    struct object *i, *outcast = NULL;
 437                    list_for_each_entry(i, &cache, list) {
 438                            if (!outcast || i->popularity < outcast->popularity)
 439                                    outcast = i;
 440                    }
 441                    __cache_delete(outcast);
 442            }
 443    }
 444
 445    int cache_add(int id, const char *name)
 446    {
 447            struct object *obj;
 448
 449            if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
 450                    return -ENOMEM;
 451
 452            strlcpy(obj->name, name, sizeof(obj->name));
 453            obj->id = id;
 454            obj->popularity = 0;
 455
 456            mutex_lock(&cache_lock);
 457            __cache_add(obj);
 458            mutex_unlock(&cache_lock);
 459            return 0;
 460    }
 461
 462    void cache_delete(int id)
 463    {
 464            mutex_lock(&cache_lock);
 465            __cache_delete(__cache_find(id));
 466            mutex_unlock(&cache_lock);
 467    }
 468
 469    int cache_find(int id, char *name)
 470    {
 471            struct object *obj;
 472            int ret = -ENOENT;
 473
 474            mutex_lock(&cache_lock);
 475            obj = __cache_find(id);
 476            if (obj) {
 477                    ret = 0;
 478                    strcpy(name, obj->name);
 479            }
 480            mutex_unlock(&cache_lock);
 481            return ret;
 482    }
 483
 484Note that we always make sure we have the cache_lock when we add,
 485delete, or look up the cache: both the cache infrastructure itself and
 486the contents of the objects are protected by the lock. In this case it's
 487easy, since we copy the data for the user, and never let them access the
 488objects directly.
 489
 490There is a slight (and common) optimization here: in
 491:c:func:`cache_add()` we set up the fields of the object before
 492grabbing the lock. This is safe, as no-one else can access it until we
 493put it in cache.
 494
 495Accessing From Interrupt Context
 496--------------------------------
 497
 498Now consider the case where :c:func:`cache_find()` can be called
 499from interrupt context: either a hardware interrupt or a softirq. An
 500example would be a timer which deletes object from the cache.
 501
 502The change is shown below, in standard patch format: the ``-`` are lines
 503which are taken away, and the ``+`` are lines which are added.
 504
 505::
 506
 507    --- cache.c.usercontext 2003-12-09 13:58:54.000000000 +1100
 508    +++ cache.c.interrupt   2003-12-09 14:07:49.000000000 +1100
 509    @@ -12,7 +12,7 @@
 510             int popularity;
 511     };
 512
 513    -static DEFINE_MUTEX(cache_lock);
 514    +static DEFINE_SPINLOCK(cache_lock);
 515     static LIST_HEAD(cache);
 516     static unsigned int cache_num = 0;
 517     #define MAX_CACHE_SIZE 10
 518    @@ -55,6 +55,7 @@
 519     int cache_add(int id, const char *name)
 520     {
 521             struct object *obj;
 522    +        unsigned long flags;
 523
 524             if ((obj = kmalloc(sizeof(*obj), GFP_KERNEL)) == NULL)
 525                     return -ENOMEM;
 526    @@ -63,30 +64,33 @@
 527             obj->id = id;
 528             obj->popularity = 0;
 529
 530    -        mutex_lock(&cache_lock);
 531    +        spin_lock_irqsave(&cache_lock, flags);
 532             __cache_add(obj);
 533    -        mutex_unlock(&cache_lock);
 534    +        spin_unlock_irqrestore(&cache_lock, flags);
 535             return 0;
 536     }
 537
 538     void cache_delete(int id)
 539     {
 540    -        mutex_lock(&cache_lock);
 541    +        unsigned long flags;
 542    +
 543    +        spin_lock_irqsave(&cache_lock, flags);
 544             __cache_delete(__cache_find(id));
 545    -        mutex_unlock(&cache_lock);
 546    +        spin_unlock_irqrestore(&cache_lock, flags);
 547     }
 548
 549     int cache_find(int id, char *name)
 550     {
 551             struct object *obj;
 552             int ret = -ENOENT;
 553    +        unsigned long flags;
 554
 555    -        mutex_lock(&cache_lock);
 556    +        spin_lock_irqsave(&cache_lock, flags);
 557             obj = __cache_find(id);
 558             if (obj) {
 559                     ret = 0;
 560                     strcpy(name, obj->name);
 561             }
 562    -        mutex_unlock(&cache_lock);
 563    +        spin_unlock_irqrestore(&cache_lock, flags);
 564             return ret;
 565     }
 566
 567Note that the :c:func:`spin_lock_irqsave()` will turn off
 568interrupts if they are on, otherwise does nothing (if we are already in
 569an interrupt handler), hence these functions are safe to call from any
 570context.
 571
 572Unfortunately, :c:func:`cache_add()` calls :c:func:`kmalloc()`
 573with the ``GFP_KERNEL`` flag, which is only legal in user context. I
 574have assumed that :c:func:`cache_add()` is still only called in
 575user context, otherwise this should become a parameter to
 576:c:func:`cache_add()`.
 577
 578Exposing Objects Outside This File
 579----------------------------------
 580
 581If our objects contained more information, it might not be sufficient to
 582copy the information in and out: other parts of the code might want to
 583keep pointers to these objects, for example, rather than looking up the
 584id every time. This produces two problems.
 585
 586The first problem is that we use the ``cache_lock`` to protect objects:
 587we'd need to make this non-static so the rest of the code can use it.
 588This makes locking trickier, as it is no longer all in one place.
 589
 590The second problem is the lifetime problem: if another structure keeps a
 591pointer to an object, it presumably expects that pointer to remain
 592valid. Unfortunately, this is only guaranteed while you hold the lock,
 593otherwise someone might call :c:func:`cache_delete()` and even
 594worse, add another object, re-using the same address.
 595
 596As there is only one lock, you can't hold it forever: no-one else would
 597get any work done.
 598
 599The solution to this problem is to use a reference count: everyone who
 600has a pointer to the object increases it when they first get the object,
 601and drops the reference count when they're finished with it. Whoever
 602drops it to zero knows it is unused, and can actually delete it.
 603
 604Here is the code::
 605
 606    --- cache.c.interrupt   2003-12-09 14:25:43.000000000 +1100
 607    +++ cache.c.refcnt  2003-12-09 14:33:05.000000000 +1100
 608    @@ -7,6 +7,7 @@
 609     struct object
 610     {
 611             struct list_head list;
 612    +        unsigned int refcnt;
 613             int id;
 614             char name[32];
 615             int popularity;
 616    @@ -17,6 +18,35 @@
 617     static unsigned int cache_num = 0;
 618     #define MAX_CACHE_SIZE 10
 619
 620    +static void __object_put(struct object *obj)
 621    +{
 622    +        if (--obj->refcnt == 0)
 623    +                kfree(obj);
 624    +}
 625    +
 626    +static void __object_get(struct object *obj)
 627    +{
 628    +        obj->refcnt++;
 629    +}
 630    +
 631    +void object_put(struct object *obj)
 632    +{
 633    +        unsigned long flags;
 634    +
 635    +        spin_lock_irqsave(&cache_lock, flags);
 636    +        __object_put(obj);
 637    +        spin_unlock_irqrestore(&cache_lock, flags);
 638    +}
 639    +
 640    +void object_get(struct object *obj)
 641    +{
 642    +        unsigned long flags;
 643    +
 644    +        spin_lock_irqsave(&cache_lock, flags);
 645    +        __object_get(obj);
 646    +        spin_unlock_irqrestore(&cache_lock, flags);
 647    +}
 648    +
 649     /* Must be holding cache_lock */
 650     static struct object *__cache_find(int id)
 651     {
 652    @@ -35,6 +65,7 @@
 653     {
 654             BUG_ON(!obj);
 655             list_del(&obj->list);
 656    +        __object_put(obj);
 657             cache_num--;
 658     }
 659
 660    @@ -63,6 +94,7 @@
 661             strlcpy(obj->name, name, sizeof(obj->name));
 662             obj->id = id;
 663             obj->popularity = 0;
 664    +        obj->refcnt = 1; /* The cache holds a reference */
 665
 666             spin_lock_irqsave(&cache_lock, flags);
 667             __cache_add(obj);
 668    @@ -79,18 +111,15 @@
 669             spin_unlock_irqrestore(&cache_lock, flags);
 670     }
 671
 672    -int cache_find(int id, char *name)
 673    +struct object *cache_find(int id)
 674     {
 675             struct object *obj;
 676    -        int ret = -ENOENT;
 677             unsigned long flags;
 678
 679             spin_lock_irqsave(&cache_lock, flags);
 680             obj = __cache_find(id);
 681    -        if (obj) {
 682    -                ret = 0;
 683    -                strcpy(name, obj->name);
 684    -        }
 685    +        if (obj)
 686    +                __object_get(obj);
 687             spin_unlock_irqrestore(&cache_lock, flags);
 688    -        return ret;
 689    +        return obj;
 690     }
 691
 692We encapsulate the reference counting in the standard 'get' and 'put'
 693functions. Now we can return the object itself from
 694:c:func:`cache_find()` which has the advantage that the user can
 695now sleep holding the object (eg. to :c:func:`copy_to_user()` to
 696name to userspace).
 697
 698The other point to note is that I said a reference should be held for
 699every pointer to the object: thus the reference count is 1 when first
 700inserted into the cache. In some versions the framework does not hold a
 701reference count, but they are more complicated.
 702
 703Using Atomic Operations For The Reference Count
 704~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 705
 706In practice, :c:type:`atomic_t` would usually be used for refcnt. There are a
 707number of atomic operations defined in ``include/asm/atomic.h``: these
 708are guaranteed to be seen atomically from all CPUs in the system, so no
 709lock is required. In this case, it is simpler than using spinlocks,
 710although for anything non-trivial using spinlocks is clearer. The
 711:c:func:`atomic_inc()` and :c:func:`atomic_dec_and_test()`
 712are used instead of the standard increment and decrement operators, and
 713the lock is no longer used to protect the reference count itself.
 714
 715::
 716
 717    --- cache.c.refcnt  2003-12-09 15:00:35.000000000 +1100
 718    +++ cache.c.refcnt-atomic   2003-12-11 15:49:42.000000000 +1100
 719    @@ -7,7 +7,7 @@
 720     struct object
 721     {
 722             struct list_head list;
 723    -        unsigned int refcnt;
 724    +        atomic_t refcnt;
 725             int id;
 726             char name[32];
 727             int popularity;
 728    @@ -18,33 +18,15 @@
 729     static unsigned int cache_num = 0;
 730     #define MAX_CACHE_SIZE 10
 731
 732    -static void __object_put(struct object *obj)
 733    -{
 734    -        if (--obj->refcnt == 0)
 735    -                kfree(obj);
 736    -}
 737    -
 738    -static void __object_get(struct object *obj)
 739    -{
 740    -        obj->refcnt++;
 741    -}
 742    -
 743     void object_put(struct object *obj)
 744     {
 745    -        unsigned long flags;
 746    -
 747    -        spin_lock_irqsave(&cache_lock, flags);
 748    -        __object_put(obj);
 749    -        spin_unlock_irqrestore(&cache_lock, flags);
 750    +        if (atomic_dec_and_test(&obj->refcnt))
 751    +                kfree(obj);
 752     }
 753
 754     void object_get(struct object *obj)
 755     {
 756    -        unsigned long flags;
 757    -
 758    -        spin_lock_irqsave(&cache_lock, flags);
 759    -        __object_get(obj);
 760    -        spin_unlock_irqrestore(&cache_lock, flags);
 761    +        atomic_inc(&obj->refcnt);
 762     }
 763
 764     /* Must be holding cache_lock */
 765    @@ -65,7 +47,7 @@
 766     {
 767             BUG_ON(!obj);
 768             list_del(&obj->list);
 769    -        __object_put(obj);
 770    +        object_put(obj);
 771             cache_num--;
 772     }
 773
 774    @@ -94,7 +76,7 @@
 775             strlcpy(obj->name, name, sizeof(obj->name));
 776             obj->id = id;
 777             obj->popularity = 0;
 778    -        obj->refcnt = 1; /* The cache holds a reference */
 779    +        atomic_set(&obj->refcnt, 1); /* The cache holds a reference */
 780
 781             spin_lock_irqsave(&cache_lock, flags);
 782             __cache_add(obj);
 783    @@ -119,7 +101,7 @@
 784             spin_lock_irqsave(&cache_lock, flags);
 785             obj = __cache_find(id);
 786             if (obj)
 787    -                __object_get(obj);
 788    +                object_get(obj);
 789             spin_unlock_irqrestore(&cache_lock, flags);
 790             return obj;
 791     }
 792
 793Protecting The Objects Themselves
 794---------------------------------
 795
 796In these examples, we assumed that the objects (except the reference
 797counts) never changed once they are created. If we wanted to allow the
 798name to change, there are three possibilities:
 799
 800-  You can make ``cache_lock`` non-static, and tell people to grab that
 801   lock before changing the name in any object.
 802
 803-  You can provide a :c:func:`cache_obj_rename()` which grabs this
 804   lock and changes the name for the caller, and tell everyone to use
 805   that function.
 806
 807-  You can make the ``cache_lock`` protect only the cache itself, and
 808   use another lock to protect the name.
 809
 810Theoretically, you can make the locks as fine-grained as one lock for
 811every field, for every object. In practice, the most common variants
 812are:
 813
 814-  One lock which protects the infrastructure (the ``cache`` list in
 815   this example) and all the objects. This is what we have done so far.
 816
 817-  One lock which protects the infrastructure (including the list
 818   pointers inside the objects), and one lock inside the object which
 819   protects the rest of that object.
 820
 821-  Multiple locks to protect the infrastructure (eg. one lock per hash
 822   chain), possibly with a separate per-object lock.
 823
 824Here is the "lock-per-object" implementation:
 825
 826::
 827
 828    --- cache.c.refcnt-atomic   2003-12-11 15:50:54.000000000 +1100
 829    +++ cache.c.perobjectlock   2003-12-11 17:15:03.000000000 +1100
 830    @@ -6,11 +6,17 @@
 831
 832     struct object
 833     {
 834    +        /* These two protected by cache_lock. */
 835             struct list_head list;
 836    +        int popularity;
 837    +
 838             atomic_t refcnt;
 839    +
 840    +        /* Doesn't change once created. */
 841             int id;
 842    +
 843    +        spinlock_t lock; /* Protects the name */
 844             char name[32];
 845    -        int popularity;
 846     };
 847
 848     static DEFINE_SPINLOCK(cache_lock);
 849    @@ -77,6 +84,7 @@
 850             obj->id = id;
 851             obj->popularity = 0;
 852             atomic_set(&obj->refcnt, 1); /* The cache holds a reference */
 853    +        spin_lock_init(&obj->lock);
 854
 855             spin_lock_irqsave(&cache_lock, flags);
 856             __cache_add(obj);
 857
 858Note that I decide that the popularity count should be protected by the
 859``cache_lock`` rather than the per-object lock: this is because it (like
 860the :c:type:`struct list_head <list_head>` inside the object)
 861is logically part of the infrastructure. This way, I don't need to grab
 862the lock of every object in :c:func:`__cache_add()` when seeking
 863the least popular.
 864
 865I also decided that the id member is unchangeable, so I don't need to
 866grab each object lock in :c:func:`__cache_find()` to examine the
 867id: the object lock is only used by a caller who wants to read or write
 868the name field.
 869
 870Note also that I added a comment describing what data was protected by
 871which locks. This is extremely important, as it describes the runtime
 872behavior of the code, and can be hard to gain from just reading. And as
 873Alan Cox says, “Lock data, not code”.
 874
 875Common Problems
 876===============
 877
 878Deadlock: Simple and Advanced
 879-----------------------------
 880
 881There is a coding bug where a piece of code tries to grab a spinlock
 882twice: it will spin forever, waiting for the lock to be released
 883(spinlocks, rwlocks and mutexes are not recursive in Linux). This is
 884trivial to diagnose: not a
 885stay-up-five-nights-talk-to-fluffy-code-bunnies kind of problem.
 886
 887For a slightly more complex case, imagine you have a region shared by a
 888softirq and user context. If you use a :c:func:`spin_lock()` call
 889to protect it, it is possible that the user context will be interrupted
 890by the softirq while it holds the lock, and the softirq will then spin
 891forever trying to get the same lock.
 892
 893Both of these are called deadlock, and as shown above, it can occur even
 894with a single CPU (although not on UP compiles, since spinlocks vanish
 895on kernel compiles with ``CONFIG_SMP``\ =n. You'll still get data
 896corruption in the second example).
 897
 898This complete lockup is easy to diagnose: on SMP boxes the watchdog
 899timer or compiling with ``DEBUG_SPINLOCK`` set
 900(``include/linux/spinlock.h``) will show this up immediately when it
 901happens.
 902
 903A more complex problem is the so-called 'deadly embrace', involving two
 904or more locks. Say you have a hash table: each entry in the table is a
 905spinlock, and a chain of hashed objects. Inside a softirq handler, you
 906sometimes want to alter an object from one place in the hash to another:
 907you grab the spinlock of the old hash chain and the spinlock of the new
 908hash chain, and delete the object from the old one, and insert it in the
 909new one.
 910
 911There are two problems here. First, if your code ever tries to move the
 912object to the same chain, it will deadlock with itself as it tries to
 913lock it twice. Secondly, if the same softirq on another CPU is trying to
 914move another object in the reverse direction, the following could
 915happen:
 916
 917+-----------------------+-----------------------+
 918| CPU 1                 | CPU 2                 |
 919+=======================+=======================+
 920| Grab lock A -> OK     | Grab lock B -> OK     |
 921+-----------------------+-----------------------+
 922| Grab lock B -> spin   | Grab lock A -> spin   |
 923+-----------------------+-----------------------+
 924
 925Table: Consequences
 926
 927The two CPUs will spin forever, waiting for the other to give up their
 928lock. It will look, smell, and feel like a crash.
 929
 930Preventing Deadlock
 931-------------------
 932
 933Textbooks will tell you that if you always lock in the same order, you
 934will never get this kind of deadlock. Practice will tell you that this
 935approach doesn't scale: when I create a new lock, I don't understand
 936enough of the kernel to figure out where in the 5000 lock hierarchy it
 937will fit.
 938
 939The best locks are encapsulated: they never get exposed in headers, and
 940are never held around calls to non-trivial functions outside the same
 941file. You can read through this code and see that it will never
 942deadlock, because it never tries to grab another lock while it has that
 943one. People using your code don't even need to know you are using a
 944lock.
 945
 946A classic problem here is when you provide callbacks or hooks: if you
 947call these with the lock held, you risk simple deadlock, or a deadly
 948embrace (who knows what the callback will do?). Remember, the other
 949programmers are out to get you, so don't do this.
 950
 951Overzealous Prevention Of Deadlocks
 952~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 953
 954Deadlocks are problematic, but not as bad as data corruption. Code which
 955grabs a read lock, searches a list, fails to find what it wants, drops
 956the read lock, grabs a write lock and inserts the object has a race
 957condition.
 958
 959If you don't see why, please stay the fuck away from my code.
 960
 961Racing Timers: A Kernel Pastime
 962-------------------------------
 963
 964Timers can produce their own special problems with races. Consider a
 965collection of objects (list, hash, etc) where each object has a timer
 966which is due to destroy it.
 967
 968If you want to destroy the entire collection (say on module removal),
 969you might do the following::
 970
 971            /* THIS CODE BAD BAD BAD BAD: IF IT WAS ANY WORSE IT WOULD USE
 972               HUNGARIAN NOTATION */
 973            spin_lock_bh(&list_lock);
 974
 975            while (list) {
 976                    struct foo *next = list->next;
 977                    del_timer(&list->timer);
 978                    kfree(list);
 979                    list = next;
 980            }
 981
 982            spin_unlock_bh(&list_lock);
 983
 984
 985Sooner or later, this will crash on SMP, because a timer can have just
 986gone off before the :c:func:`spin_lock_bh()`, and it will only get
 987the lock after we :c:func:`spin_unlock_bh()`, and then try to free
 988the element (which has already been freed!).
 989
 990This can be avoided by checking the result of
 991:c:func:`del_timer()`: if it returns 1, the timer has been deleted.
 992If 0, it means (in this case) that it is currently running, so we can
 993do::
 994
 995            retry:
 996                    spin_lock_bh(&list_lock);
 997
 998                    while (list) {
 999                            struct foo *next = list->next;
1000                            if (!del_timer(&list->timer)) {
1001                                    /* Give timer a chance to delete this */
1002                                    spin_unlock_bh(&list_lock);
1003                                    goto retry;
1004                            }
1005                            kfree(list);
1006                            list = next;
1007                    }
1008
1009                    spin_unlock_bh(&list_lock);
1010
1011
1012Another common problem is deleting timers which restart themselves (by
1013calling :c:func:`add_timer()` at the end of their timer function).
1014Because this is a fairly common case which is prone to races, you should
1015use :c:func:`del_timer_sync()` (``include/linux/timer.h``) to
1016handle this case. It returns the number of times the timer had to be
1017deleted before we finally stopped it from adding itself back in.
1018
1019Locking Speed
1020=============
1021
1022There are three main things to worry about when considering speed of
1023some code which does locking. First is concurrency: how many things are
1024going to be waiting while someone else is holding a lock. Second is the
1025time taken to actually acquire and release an uncontended lock. Third is
1026using fewer, or smarter locks. I'm assuming that the lock is used fairly
1027often: otherwise, you wouldn't be concerned about efficiency.
1028
1029Concurrency depends on how long the lock is usually held: you should
1030hold the lock for as long as needed, but no longer. In the cache
1031example, we always create the object without the lock held, and then
1032grab the lock only when we are ready to insert it in the list.
1033
1034Acquisition times depend on how much damage the lock operations do to
1035the pipeline (pipeline stalls) and how likely it is that this CPU was
1036the last one to grab the lock (ie. is the lock cache-hot for this CPU):
1037on a machine with more CPUs, this likelihood drops fast. Consider a
1038700MHz Intel Pentium III: an instruction takes about 0.7ns, an atomic
1039increment takes about 58ns, a lock which is cache-hot on this CPU takes
1040160ns, and a cacheline transfer from another CPU takes an additional 170
1041to 360ns. (These figures from Paul McKenney's `Linux Journal RCU
1042article <http://www.linuxjournal.com/article.php?sid=6993>`__).
1043
1044These two aims conflict: holding a lock for a short time might be done
1045by splitting locks into parts (such as in our final per-object-lock
1046example), but this increases the number of lock acquisitions, and the
1047results are often slower than having a single lock. This is another
1048reason to advocate locking simplicity.
1049
1050The third concern is addressed below: there are some methods to reduce
1051the amount of locking which needs to be done.
1052
1053Read/Write Lock Variants
1054------------------------
1055
1056Both spinlocks and mutexes have read/write variants: ``rwlock_t`` and
1057:c:type:`struct rw_semaphore <rw_semaphore>`. These divide
1058users into two classes: the readers and the writers. If you are only
1059reading the data, you can get a read lock, but to write to the data you
1060need the write lock. Many people can hold a read lock, but a writer must
1061be sole holder.
1062
1063If your code divides neatly along reader/writer lines (as our cache code
1064does), and the lock is held by readers for significant lengths of time,
1065using these locks can help. They are slightly slower than the normal
1066locks though, so in practice ``rwlock_t`` is not usually worthwhile.
1067
1068Avoiding Locks: Read Copy Update
1069--------------------------------
1070
1071There is a special method of read/write locking called Read Copy Update.
1072Using RCU, the readers can avoid taking a lock altogether: as we expect
1073our cache to be read more often than updated (otherwise the cache is a
1074waste of time), it is a candidate for this optimization.
1075
1076How do we get rid of read locks? Getting rid of read locks means that
1077writers may be changing the list underneath the readers. That is
1078actually quite simple: we can read a linked list while an element is
1079being added if the writer adds the element very carefully. For example,
1080adding ``new`` to a single linked list called ``list``::
1081
1082            new->next = list->next;
1083            wmb();
1084            list->next = new;
1085
1086
1087The :c:func:`wmb()` is a write memory barrier. It ensures that the
1088first operation (setting the new element's ``next`` pointer) is complete
1089and will be seen by all CPUs, before the second operation is (putting
1090the new element into the list). This is important, since modern
1091compilers and modern CPUs can both reorder instructions unless told
1092otherwise: we want a reader to either not see the new element at all, or
1093see the new element with the ``next`` pointer correctly pointing at the
1094rest of the list.
1095
1096Fortunately, there is a function to do this for standard
1097:c:type:`struct list_head <list_head>` lists:
1098:c:func:`list_add_rcu()` (``include/linux/list.h``).
1099
1100Removing an element from the list is even simpler: we replace the
1101pointer to the old element with a pointer to its successor, and readers
1102will either see it, or skip over it.
1103
1104::
1105
1106            list->next = old->next;
1107
1108
1109There is :c:func:`list_del_rcu()` (``include/linux/list.h``) which
1110does this (the normal version poisons the old object, which we don't
1111want).
1112
1113The reader must also be careful: some CPUs can look through the ``next``
1114pointer to start reading the contents of the next element early, but
1115don't realize that the pre-fetched contents is wrong when the ``next``
1116pointer changes underneath them. Once again, there is a
1117:c:func:`list_for_each_entry_rcu()` (``include/linux/list.h``)
1118to help you. Of course, writers can just use
1119:c:func:`list_for_each_entry()`, since there cannot be two
1120simultaneous writers.
1121
1122Our final dilemma is this: when can we actually destroy the removed
1123element? Remember, a reader might be stepping through this element in
1124the list right now: if we free this element and the ``next`` pointer
1125changes, the reader will jump off into garbage and crash. We need to
1126wait until we know that all the readers who were traversing the list
1127when we deleted the element are finished. We use
1128:c:func:`call_rcu()` to register a callback which will actually
1129destroy the object once all pre-existing readers are finished.
1130Alternatively, :c:func:`synchronize_rcu()` may be used to block
1131until all pre-existing are finished.
1132
1133But how does Read Copy Update know when the readers are finished? The
1134method is this: firstly, the readers always traverse the list inside
1135:c:func:`rcu_read_lock()`/:c:func:`rcu_read_unlock()` pairs:
1136these simply disable preemption so the reader won't go to sleep while
1137reading the list.
1138
1139RCU then waits until every other CPU has slept at least once: since
1140readers cannot sleep, we know that any readers which were traversing the
1141list during the deletion are finished, and the callback is triggered.
1142The real Read Copy Update code is a little more optimized than this, but
1143this is the fundamental idea.
1144
1145::
1146
1147    --- cache.c.perobjectlock   2003-12-11 17:15:03.000000000 +1100
1148    +++ cache.c.rcupdate    2003-12-11 17:55:14.000000000 +1100
1149    @@ -1,15 +1,18 @@
1150     #include <linux/list.h>
1151     #include <linux/slab.h>
1152     #include <linux/string.h>
1153    +#include <linux/rcupdate.h>
1154     #include <linux/mutex.h>
1155     #include <asm/errno.h>
1156
1157     struct object
1158     {
1159    -        /* These two protected by cache_lock. */
1160    +        /* This is protected by RCU */
1161             struct list_head list;
1162             int popularity;
1163
1164    +        struct rcu_head rcu;
1165    +
1166             atomic_t refcnt;
1167
1168             /* Doesn't change once created. */
1169    @@ -40,7 +43,7 @@
1170     {
1171             struct object *i;
1172
1173    -        list_for_each_entry(i, &cache, list) {
1174    +        list_for_each_entry_rcu(i, &cache, list) {
1175                     if (i->id == id) {
1176                             i->popularity++;
1177                             return i;
1178    @@ -49,19 +52,25 @@
1179             return NULL;
1180     }
1181
1182    +/* Final discard done once we know no readers are looking. */
1183    +static void cache_delete_rcu(void *arg)
1184    +{
1185    +        object_put(arg);
1186    +}
1187    +
1188     /* Must be holding cache_lock */
1189     static void __cache_delete(struct object *obj)
1190     {
1191             BUG_ON(!obj);
1192    -        list_del(&obj->list);
1193    -        object_put(obj);
1194    +        list_del_rcu(&obj->list);
1195             cache_num--;
1196    +        call_rcu(&obj->rcu, cache_delete_rcu);
1197     }
1198
1199     /* Must be holding cache_lock */
1200     static void __cache_add(struct object *obj)
1201     {
1202    -        list_add(&obj->list, &cache);
1203    +        list_add_rcu(&obj->list, &cache);
1204             if (++cache_num > MAX_CACHE_SIZE) {
1205                     struct object *i, *outcast = NULL;
1206                     list_for_each_entry(i, &cache, list) {
1207    @@ -104,12 +114,11 @@
1208     struct object *cache_find(int id)
1209     {
1210             struct object *obj;
1211    -        unsigned long flags;
1212
1213    -        spin_lock_irqsave(&cache_lock, flags);
1214    +        rcu_read_lock();
1215             obj = __cache_find(id);
1216             if (obj)
1217                     object_get(obj);
1218    -        spin_unlock_irqrestore(&cache_lock, flags);
1219    +        rcu_read_unlock();
1220             return obj;
1221     }
1222
1223Note that the reader will alter the popularity member in
1224:c:func:`__cache_find()`, and now it doesn't hold a lock. One
1225solution would be to make it an ``atomic_t``, but for this usage, we
1226don't really care about races: an approximate result is good enough, so
1227I didn't change it.
1228
1229The result is that :c:func:`cache_find()` requires no
1230synchronization with any other functions, so is almost as fast on SMP as
1231it would be on UP.
1232
1233There is a further optimization possible here: remember our original
1234cache code, where there were no reference counts and the caller simply
1235held the lock whenever using the object? This is still possible: if you
1236hold the lock, no one can delete the object, so you don't need to get
1237and put the reference count.
1238
1239Now, because the 'read lock' in RCU is simply disabling preemption, a
1240caller which always has preemption disabled between calling
1241:c:func:`cache_find()` and :c:func:`object_put()` does not
1242need to actually get and put the reference count: we could expose
1243:c:func:`__cache_find()` by making it non-static, and such
1244callers could simply call that.
1245
1246The benefit here is that the reference count is not written to: the
1247object is not altered in any way, which is much faster on SMP machines
1248due to caching.
1249
1250Per-CPU Data
1251------------
1252
1253Another technique for avoiding locking which is used fairly widely is to
1254duplicate information for each CPU. For example, if you wanted to keep a
1255count of a common condition, you could use a spin lock and a single
1256counter. Nice and simple.
1257
1258If that was too slow (it's usually not, but if you've got a really big
1259machine to test on and can show that it is), you could instead use a
1260counter for each CPU, then none of them need an exclusive lock. See
1261:c:func:`DEFINE_PER_CPU()`, :c:func:`get_cpu_var()` and
1262:c:func:`put_cpu_var()` (``include/linux/percpu.h``).
1263
1264Of particular use for simple per-cpu counters is the ``local_t`` type,
1265and the :c:func:`cpu_local_inc()` and related functions, which are
1266more efficient than simple code on some architectures
1267(``include/asm/local.h``).
1268
1269Note that there is no simple, reliable way of getting an exact value of
1270such a counter, without introducing more locks. This is not a problem
1271for some uses.
1272
1273Data Which Mostly Used By An IRQ Handler
1274----------------------------------------
1275
1276If data is always accessed from within the same IRQ handler, you don't
1277need a lock at all: the kernel already guarantees that the irq handler
1278will not run simultaneously on multiple CPUs.
1279
1280Manfred Spraul points out that you can still do this, even if the data
1281is very occasionally accessed in user context or softirqs/tasklets. The
1282irq handler doesn't use a lock, and all other accesses are done as so::
1283
1284        spin_lock(&lock);
1285        disable_irq(irq);
1286        ...
1287        enable_irq(irq);
1288        spin_unlock(&lock);
1289
1290The :c:func:`disable_irq()` prevents the irq handler from running
1291(and waits for it to finish if it's currently running on other CPUs).
1292The spinlock prevents any other accesses happening at the same time.
1293Naturally, this is slower than just a :c:func:`spin_lock_irq()`
1294call, so it only makes sense if this type of access happens extremely
1295rarely.
1296
1297What Functions Are Safe To Call From Interrupts?
1298================================================
1299
1300Many functions in the kernel sleep (ie. call schedule()) directly or
1301indirectly: you can never call them while holding a spinlock, or with
1302preemption disabled. This also means you need to be in user context:
1303calling them from an interrupt is illegal.
1304
1305Some Functions Which Sleep
1306--------------------------
1307
1308The most common ones are listed below, but you usually have to read the
1309code to find out if other calls are safe. If everyone else who calls it
1310can sleep, you probably need to be able to sleep, too. In particular,
1311registration and deregistration functions usually expect to be called
1312from user context, and can sleep.
1313
1314-  Accesses to userspace:
1315
1316   -  :c:func:`copy_from_user()`
1317
1318   -  :c:func:`copy_to_user()`
1319
1320   -  :c:func:`get_user()`
1321
1322   -  :c:func:`put_user()`
1323
1324-  :c:func:`kmalloc(GFP_KERNEL) <kmalloc>`
1325
1326-  :c:func:`mutex_lock_interruptible()` and
1327   :c:func:`mutex_lock()`
1328
1329   There is a :c:func:`mutex_trylock()` which does not sleep.
1330   Still, it must not be used inside interrupt context since its
1331   implementation is not safe for that. :c:func:`mutex_unlock()`
1332   will also never sleep. It cannot be used in interrupt context either
1333   since a mutex must be released by the same task that acquired it.
1334
1335Some Functions Which Don't Sleep
1336--------------------------------
1337
1338Some functions are safe to call from any context, or holding almost any
1339lock.
1340
1341-  :c:func:`printk()`
1342
1343-  :c:func:`kfree()`
1344
1345-  :c:func:`add_timer()` and :c:func:`del_timer()`
1346
1347Mutex API reference
1348===================
1349
1350.. kernel-doc:: include/linux/mutex.h
1351   :internal:
1352
1353.. kernel-doc:: kernel/locking/mutex.c
1354   :export:
1355
1356Futex API reference
1357===================
1358
1359.. kernel-doc:: kernel/futex.c
1360   :internal:
1361
1362Further reading
1363===============
1364
1365-  ``Documentation/locking/spinlocks.txt``: Linus Torvalds' spinlocking
1366   tutorial in the kernel sources.
1367
1368-  Unix Systems for Modern Architectures: Symmetric Multiprocessing and
1369   Caching for Kernel Programmers:
1370
1371   Curt Schimmel's very good introduction to kernel level locking (not
1372   written for Linux, but nearly everything applies). The book is
1373   expensive, but really worth every penny to understand SMP locking.
1374   [ISBN: 0201633388]
1375
1376Thanks
1377======
1378
1379Thanks to Telsa Gwynne for DocBooking, neatening and adding style.
1380
1381Thanks to Martin Pool, Philipp Rumpf, Stephen Rothwell, Paul Mackerras,
1382Ruedi Aschwanden, Alan Cox, Manfred Spraul, Tim Waugh, Pete Zaitcev,
1383James Morris, Robert Love, Paul McKenney, John Ashby for proofreading,
1384correcting, flaming, commenting.
1385
1386Thanks to the cabal for having no influence on this document.
1387
1388Glossary
1389========
1390
1391preemption
1392  Prior to 2.5, or when ``CONFIG_PREEMPT`` is unset, processes in user
1393  context inside the kernel would not preempt each other (ie. you had that
1394  CPU until you gave it up, except for interrupts). With the addition of
1395  ``CONFIG_PREEMPT`` in 2.5.4, this changed: when in user context, higher
1396  priority tasks can "cut in": spinlocks were changed to disable
1397  preemption, even on UP.
1398
1399bh
1400  Bottom Half: for historical reasons, functions with '_bh' in them often
1401  now refer to any software interrupt, e.g. :c:func:`spin_lock_bh()`
1402  blocks any software interrupt on the current CPU. Bottom halves are
1403  deprecated, and will eventually be replaced by tasklets. Only one bottom
1404  half will be running at any time.
1405
1406Hardware Interrupt / Hardware IRQ
1407  Hardware interrupt request. :c:func:`in_irq()` returns true in a
1408  hardware interrupt handler.
1409
1410Interrupt Context
1411  Not user context: processing a hardware irq or software irq. Indicated
1412  by the :c:func:`in_interrupt()` macro returning true.
1413
1414SMP
1415  Symmetric Multi-Processor: kernels compiled for multiple-CPU machines.
1416  (``CONFIG_SMP=y``).
1417
1418Software Interrupt / softirq
1419  Software interrupt handler. :c:func:`in_irq()` returns false;
1420  :c:func:`in_softirq()` returns true. Tasklets and softirqs both
1421  fall into the category of 'software interrupts'.
1422
1423  Strictly speaking a softirq is one of up to 32 enumerated software
1424  interrupts which can run on multiple CPUs at once. Sometimes used to
1425  refer to tasklets as well (ie. all software interrupts).
1426
1427tasklet
1428  A dynamically-registrable software interrupt, which is guaranteed to
1429  only run on one CPU at a time.
1430
1431timer
1432  A dynamically-registrable software interrupt, which is run at (or close
1433  to) a given time. When running, it is just like a tasklet (in fact, they
1434  are called from the ``TIMER_SOFTIRQ``).
1435
1436UP
1437  Uni-Processor: Non-SMP. (``CONFIG_SMP=n``).
1438
1439User Context
1440  The kernel executing on behalf of a particular process (ie. a system
1441  call or trap) or kernel thread. You can tell which process with the
1442  ``current`` macro.) Not to be confused with userspace. Can be
1443  interrupted by software or hardware interrupts.
1444
1445Userspace
1446  A process executing its own code outside the kernel.