linux/Documentation/RCU/whatisRCU.txt
<<
>>
Prefs
   1Please note that the "What is RCU?" LWN series is an excellent place
   2to start learning about RCU:
   3
   41.      What is RCU, Fundamentally?  http://lwn.net/Articles/262464/
   52.      What is RCU? Part 2: Usage   http://lwn.net/Articles/263130/
   63.      RCU part 3: the RCU API      http://lwn.net/Articles/264090/
   74.      The RCU API, 2010 Edition    http://lwn.net/Articles/418853/
   8
   9
  10What is RCU?
  11
  12RCU is a synchronization mechanism that was added to the Linux kernel
  13during the 2.5 development effort that is optimized for read-mostly
  14situations.  Although RCU is actually quite simple once you understand it,
  15getting there can sometimes be a challenge.  Part of the problem is that
  16most of the past descriptions of RCU have been written with the mistaken
  17assumption that there is "one true way" to describe RCU.  Instead,
  18the experience has been that different people must take different paths
  19to arrive at an understanding of RCU.  This document provides several
  20different paths, as follows:
  21
  221.      RCU OVERVIEW
  232.      WHAT IS RCU'S CORE API?
  243.      WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
  254.      WHAT IF MY UPDATING THREAD CANNOT BLOCK?
  265.      WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
  276.      ANALOGY WITH READER-WRITER LOCKING
  287.      FULL LIST OF RCU APIs
  298.      ANSWERS TO QUICK QUIZZES
  30
  31People who prefer starting with a conceptual overview should focus on
  32Section 1, though most readers will profit by reading this section at
  33some point.  People who prefer to start with an API that they can then
  34experiment with should focus on Section 2.  People who prefer to start
  35with example uses should focus on Sections 3 and 4.  People who need to
  36understand the RCU implementation should focus on Section 5, then dive
  37into the kernel source code.  People who reason best by analogy should
  38focus on Section 6.  Section 7 serves as an index to the docbook API
  39documentation, and Section 8 is the traditional answer key.
  40
  41So, start with the section that makes the most sense to you and your
  42preferred method of learning.  If you need to know everything about
  43everything, feel free to read the whole thing -- but if you are really
  44that type of person, you have perused the source code and will therefore
  45never need this document anyway.  ;-)
  46
  47
  481.  RCU OVERVIEW
  49
  50The basic idea behind RCU is to split updates into "removal" and
  51"reclamation" phases.  The removal phase removes references to data items
  52within a data structure (possibly by replacing them with references to
  53new versions of these data items), and can run concurrently with readers.
  54The reason that it is safe to run the removal phase concurrently with
  55readers is the semantics of modern CPUs guarantee that readers will see
  56either the old or the new version of the data structure rather than a
  57partially updated reference.  The reclamation phase does the work of reclaiming
  58(e.g., freeing) the data items removed from the data structure during the
  59removal phase.  Because reclaiming data items can disrupt any readers
  60concurrently referencing those data items, the reclamation phase must
  61not start until readers no longer hold references to those data items.
  62
  63Splitting the update into removal and reclamation phases permits the
  64updater to perform the removal phase immediately, and to defer the
  65reclamation phase until all readers active during the removal phase have
  66completed, either by blocking until they finish or by registering a
  67callback that is invoked after they finish.  Only readers that are active
  68during the removal phase need be considered, because any reader starting
  69after the removal phase will be unable to gain a reference to the removed
  70data items, and therefore cannot be disrupted by the reclamation phase.
  71
  72So the typical RCU update sequence goes something like the following:
  73
  74a.      Remove pointers to a data structure, so that subsequent
  75        readers cannot gain a reference to it.
  76
  77b.      Wait for all previous readers to complete their RCU read-side
  78        critical sections.
  79
  80c.      At this point, there cannot be any readers who hold references
  81        to the data structure, so it now may safely be reclaimed
  82        (e.g., kfree()d).
  83
  84Step (b) above is the key idea underlying RCU's deferred destruction.
  85The ability to wait until all readers are done allows RCU readers to
  86use much lighter-weight synchronization, in some cases, absolutely no
  87synchronization at all.  In contrast, in more conventional lock-based
  88schemes, readers must use heavy-weight synchronization in order to
  89prevent an updater from deleting the data structure out from under them.
  90This is because lock-based updaters typically update data items in place,
  91and must therefore exclude readers.  In contrast, RCU-based updaters
  92typically take advantage of the fact that writes to single aligned
  93pointers are atomic on modern CPUs, allowing atomic insertion, removal,
  94and replacement of data items in a linked structure without disrupting
  95readers.  Concurrent RCU readers can then continue accessing the old
  96versions, and can dispense with the atomic operations, memory barriers,
  97and communications cache misses that are so expensive on present-day
  98SMP computer systems, even in absence of lock contention.
  99
 100In the three-step procedure shown above, the updater is performing both
 101the removal and the reclamation step, but it is often helpful for an
 102entirely different thread to do the reclamation, as is in fact the case
 103in the Linux kernel's directory-entry cache (dcache).  Even if the same
 104thread performs both the update step (step (a) above) and the reclamation
 105step (step (c) above), it is often helpful to think of them separately.
 106For example, RCU readers and updaters need not communicate at all,
 107but RCU provides implicit low-overhead communication between readers
 108and reclaimers, namely, in step (b) above.
 109
 110So how the heck can a reclaimer tell when a reader is done, given
 111that readers are not doing any sort of synchronization operations???
 112Read on to learn about how RCU's API makes this easy.
 113
 114
 1152.  WHAT IS RCU'S CORE API?
 116
 117The core RCU API is quite small:
 118
 119a.      rcu_read_lock()
 120b.      rcu_read_unlock()
 121c.      synchronize_rcu() / call_rcu()
 122d.      rcu_assign_pointer()
 123e.      rcu_dereference()
 124
 125There are many other members of the RCU API, but the rest can be
 126expressed in terms of these five, though most implementations instead
 127express synchronize_rcu() in terms of the call_rcu() callback API.
 128
 129The five core RCU APIs are described below, the other 18 will be enumerated
 130later.  See the kernel docbook documentation for more info, or look directly
 131at the function header comments.
 132
 133rcu_read_lock()
 134
 135        void rcu_read_lock(void);
 136
 137        Used by a reader to inform the reclaimer that the reader is
 138        entering an RCU read-side critical section.  It is illegal
 139        to block while in an RCU read-side critical section, though
 140        kernels built with CONFIG_TREE_PREEMPT_RCU can preempt RCU
 141        read-side critical sections.  Any RCU-protected data structure
 142        accessed during an RCU read-side critical section is guaranteed to
 143        remain unreclaimed for the full duration of that critical section.
 144        Reference counts may be used in conjunction with RCU to maintain
 145        longer-term references to data structures.
 146
 147rcu_read_unlock()
 148
 149        void rcu_read_unlock(void);
 150
 151        Used by a reader to inform the reclaimer that the reader is
 152        exiting an RCU read-side critical section.  Note that RCU
 153        read-side critical sections may be nested and/or overlapping.
 154
 155synchronize_rcu()
 156
 157        void synchronize_rcu(void);
 158
 159        Marks the end of updater code and the beginning of reclaimer
 160        code.  It does this by blocking until all pre-existing RCU
 161        read-side critical sections on all CPUs have completed.
 162        Note that synchronize_rcu() will -not- necessarily wait for
 163        any subsequent RCU read-side critical sections to complete.
 164        For example, consider the following sequence of events:
 165
 166                 CPU 0                  CPU 1                 CPU 2
 167             ----------------- ------------------------- ---------------
 168         1.  rcu_read_lock()
 169         2.                    enters synchronize_rcu()
 170         3.                                               rcu_read_lock()
 171         4.  rcu_read_unlock()
 172         5.                     exits synchronize_rcu()
 173         6.                                              rcu_read_unlock()
 174
 175        To reiterate, synchronize_rcu() waits only for ongoing RCU
 176        read-side critical sections to complete, not necessarily for
 177        any that begin after synchronize_rcu() is invoked.
 178
 179        Of course, synchronize_rcu() does not necessarily return
 180        -immediately- after the last pre-existing RCU read-side critical
 181        section completes.  For one thing, there might well be scheduling
 182        delays.  For another thing, many RCU implementations process
 183        requests in batches in order to improve efficiencies, which can
 184        further delay synchronize_rcu().
 185
 186        Since synchronize_rcu() is the API that must figure out when
 187        readers are done, its implementation is key to RCU.  For RCU
 188        to be useful in all but the most read-intensive situations,
 189        synchronize_rcu()'s overhead must also be quite small.
 190
 191        The call_rcu() API is a callback form of synchronize_rcu(),
 192        and is described in more detail in a later section.  Instead of
 193        blocking, it registers a function and argument which are invoked
 194        after all ongoing RCU read-side critical sections have completed.
 195        This callback variant is particularly useful in situations where
 196        it is illegal to block or where update-side performance is
 197        critically important.
 198
 199        However, the call_rcu() API should not be used lightly, as use
 200        of the synchronize_rcu() API generally results in simpler code.
 201        In addition, the synchronize_rcu() API has the nice property
 202        of automatically limiting update rate should grace periods
 203        be delayed.  This property results in system resilience in face
 204        of denial-of-service attacks.  Code using call_rcu() should limit
 205        update rate in order to gain this same sort of resilience.  See
 206        checklist.txt for some approaches to limiting the update rate.
 207
 208rcu_assign_pointer()
 209
 210        typeof(p) rcu_assign_pointer(p, typeof(p) v);
 211
 212        Yes, rcu_assign_pointer() -is- implemented as a macro, though it
 213        would be cool to be able to declare a function in this manner.
 214        (Compiler experts will no doubt disagree.)
 215
 216        The updater uses this function to assign a new value to an
 217        RCU-protected pointer, in order to safely communicate the change
 218        in value from the updater to the reader.  This function returns
 219        the new value, and also executes any memory-barrier instructions
 220        required for a given CPU architecture.
 221
 222        Perhaps just as important, it serves to document (1) which
 223        pointers are protected by RCU and (2) the point at which a
 224        given structure becomes accessible to other CPUs.  That said,
 225        rcu_assign_pointer() is most frequently used indirectly, via
 226        the _rcu list-manipulation primitives such as list_add_rcu().
 227
 228rcu_dereference()
 229
 230        typeof(p) rcu_dereference(p);
 231
 232        Like rcu_assign_pointer(), rcu_dereference() must be implemented
 233        as a macro.
 234
 235        The reader uses rcu_dereference() to fetch an RCU-protected
 236        pointer, which returns a value that may then be safely
 237        dereferenced.  Note that rcu_deference() does not actually
 238        dereference the pointer, instead, it protects the pointer for
 239        later dereferencing.  It also executes any needed memory-barrier
 240        instructions for a given CPU architecture.  Currently, only Alpha
 241        needs memory barriers within rcu_dereference() -- on other CPUs,
 242        it compiles to nothing, not even a compiler directive.
 243
 244        Common coding practice uses rcu_dereference() to copy an
 245        RCU-protected pointer to a local variable, then dereferences
 246        this local variable, for example as follows:
 247
 248                p = rcu_dereference(head.next);
 249                return p->data;
 250
 251        However, in this case, one could just as easily combine these
 252        into one statement:
 253
 254                return rcu_dereference(head.next)->data;
 255
 256        If you are going to be fetching multiple fields from the
 257        RCU-protected structure, using the local variable is of
 258        course preferred.  Repeated rcu_dereference() calls look
 259        ugly and incur unnecessary overhead on Alpha CPUs.
 260
 261        Note that the value returned by rcu_dereference() is valid
 262        only within the enclosing RCU read-side critical section.
 263        For example, the following is -not- legal:
 264
 265                rcu_read_lock();
 266                p = rcu_dereference(head.next);
 267                rcu_read_unlock();
 268                x = p->address;
 269                rcu_read_lock();
 270                y = p->data;
 271                rcu_read_unlock();
 272
 273        Holding a reference from one RCU read-side critical section
 274        to another is just as illegal as holding a reference from
 275        one lock-based critical section to another!  Similarly,
 276        using a reference outside of the critical section in which
 277        it was acquired is just as illegal as doing so with normal
 278        locking.
 279
 280        As with rcu_assign_pointer(), an important function of
 281        rcu_dereference() is to document which pointers are protected by
 282        RCU, in particular, flagging a pointer that is subject to changing
 283        at any time, including immediately after the rcu_dereference().
 284        And, again like rcu_assign_pointer(), rcu_dereference() is
 285        typically used indirectly, via the _rcu list-manipulation
 286        primitives, such as list_for_each_entry_rcu().
 287
 288The following diagram shows how each API communicates among the
 289reader, updater, and reclaimer.
 290
 291
 292            rcu_assign_pointer()
 293                                    +--------+
 294            +---------------------->| reader |---------+
 295            |                       +--------+         |
 296            |                           |              |
 297            |                           |              | Protect:
 298            |                           |              | rcu_read_lock()
 299            |                           |              | rcu_read_unlock()
 300            |        rcu_dereference()  |              |
 301       +---------+                      |              |
 302       | updater |<---------------------+              |
 303       +---------+                                     V
 304            |                                    +-----------+
 305            +----------------------------------->| reclaimer |
 306                                                 +-----------+
 307              Defer:
 308              synchronize_rcu() & call_rcu()
 309
 310
 311The RCU infrastructure observes the time sequence of rcu_read_lock(),
 312rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in
 313order to determine when (1) synchronize_rcu() invocations may return
 314to their callers and (2) call_rcu() callbacks may be invoked.  Efficient
 315implementations of the RCU infrastructure make heavy use of batching in
 316order to amortize their overhead over many uses of the corresponding APIs.
 317
 318There are no fewer than three RCU mechanisms in the Linux kernel; the
 319diagram above shows the first one, which is by far the most commonly used.
 320The rcu_dereference() and rcu_assign_pointer() primitives are used for
 321all three mechanisms, but different defer and protect primitives are
 322used as follows:
 323
 324        Defer                   Protect
 325
 326a.      synchronize_rcu()       rcu_read_lock() / rcu_read_unlock()
 327        call_rcu()              rcu_dereference()
 328
 329b.      call_rcu_bh()           rcu_read_lock_bh() / rcu_read_unlock_bh()
 330                                rcu_dereference_bh()
 331
 332c.      synchronize_sched()     rcu_read_lock_sched() / rcu_read_unlock_sched()
 333                                preempt_disable() / preempt_enable()
 334                                local_irq_save() / local_irq_restore()
 335                                hardirq enter / hardirq exit
 336                                NMI enter / NMI exit
 337                                rcu_dereference_sched()
 338
 339These three mechanisms are used as follows:
 340
 341a.      RCU applied to normal data structures.
 342
 343b.      RCU applied to networking data structures that may be subjected
 344        to remote denial-of-service attacks.
 345
 346c.      RCU applied to scheduler and interrupt/NMI-handler tasks.
 347
 348Again, most uses will be of (a).  The (b) and (c) cases are important
 349for specialized uses, but are relatively uncommon.
 350
 351
 3523.  WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
 353
 354This section shows a simple use of the core RCU API to protect a
 355global pointer to a dynamically allocated structure.  More-typical
 356uses of RCU may be found in listRCU.txt, arrayRCU.txt, and NMI-RCU.txt.
 357
 358        struct foo {
 359                int a;
 360                char b;
 361                long c;
 362        };
 363        DEFINE_SPINLOCK(foo_mutex);
 364
 365        struct foo *gbl_foo;
 366
 367        /*
 368         * Create a new struct foo that is the same as the one currently
 369         * pointed to by gbl_foo, except that field "a" is replaced
 370         * with "new_a".  Points gbl_foo to the new structure, and
 371         * frees up the old structure after a grace period.
 372         *
 373         * Uses rcu_assign_pointer() to ensure that concurrent readers
 374         * see the initialized version of the new structure.
 375         *
 376         * Uses synchronize_rcu() to ensure that any readers that might
 377         * have references to the old structure complete before freeing
 378         * the old structure.
 379         */
 380        void foo_update_a(int new_a)
 381        {
 382                struct foo *new_fp;
 383                struct foo *old_fp;
 384
 385                new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
 386                spin_lock(&foo_mutex);
 387                old_fp = gbl_foo;
 388                *new_fp = *old_fp;
 389                new_fp->a = new_a;
 390                rcu_assign_pointer(gbl_foo, new_fp);
 391                spin_unlock(&foo_mutex);
 392                synchronize_rcu();
 393                kfree(old_fp);
 394        }
 395
 396        /*
 397         * Return the value of field "a" of the current gbl_foo
 398         * structure.  Use rcu_read_lock() and rcu_read_unlock()
 399         * to ensure that the structure does not get deleted out
 400         * from under us, and use rcu_dereference() to ensure that
 401         * we see the initialized version of the structure (important
 402         * for DEC Alpha and for people reading the code).
 403         */
 404        int foo_get_a(void)
 405        {
 406                int retval;
 407
 408                rcu_read_lock();
 409                retval = rcu_dereference(gbl_foo)->a;
 410                rcu_read_unlock();
 411                return retval;
 412        }
 413
 414So, to sum up:
 415
 416o       Use rcu_read_lock() and rcu_read_unlock() to guard RCU
 417        read-side critical sections.
 418
 419o       Within an RCU read-side critical section, use rcu_dereference()
 420        to dereference RCU-protected pointers.
 421
 422o       Use some solid scheme (such as locks or semaphores) to
 423        keep concurrent updates from interfering with each other.
 424
 425o       Use rcu_assign_pointer() to update an RCU-protected pointer.
 426        This primitive protects concurrent readers from the updater,
 427        -not- concurrent updates from each other!  You therefore still
 428        need to use locking (or something similar) to keep concurrent
 429        rcu_assign_pointer() primitives from interfering with each other.
 430
 431o       Use synchronize_rcu() -after- removing a data element from an
 432        RCU-protected data structure, but -before- reclaiming/freeing
 433        the data element, in order to wait for the completion of all
 434        RCU read-side critical sections that might be referencing that
 435        data item.
 436
 437See checklist.txt for additional rules to follow when using RCU.
 438And again, more-typical uses of RCU may be found in listRCU.txt,
 439arrayRCU.txt, and NMI-RCU.txt.
 440
 441
 4424.  WHAT IF MY UPDATING THREAD CANNOT BLOCK?
 443
 444In the example above, foo_update_a() blocks until a grace period elapses.
 445This is quite simple, but in some cases one cannot afford to wait so
 446long -- there might be other high-priority work to be done.
 447
 448In such cases, one uses call_rcu() rather than synchronize_rcu().
 449The call_rcu() API is as follows:
 450
 451        void call_rcu(struct rcu_head * head,
 452                      void (*func)(struct rcu_head *head));
 453
 454This function invokes func(head) after a grace period has elapsed.
 455This invocation might happen from either softirq or process context,
 456so the function is not permitted to block.  The foo struct needs to
 457have an rcu_head structure added, perhaps as follows:
 458
 459        struct foo {
 460                int a;
 461                char b;
 462                long c;
 463                struct rcu_head rcu;
 464        };
 465
 466The foo_update_a() function might then be written as follows:
 467
 468        /*
 469         * Create a new struct foo that is the same as the one currently
 470         * pointed to by gbl_foo, except that field "a" is replaced
 471         * with "new_a".  Points gbl_foo to the new structure, and
 472         * frees up the old structure after a grace period.
 473         *
 474         * Uses rcu_assign_pointer() to ensure that concurrent readers
 475         * see the initialized version of the new structure.
 476         *
 477         * Uses call_rcu() to ensure that any readers that might have
 478         * references to the old structure complete before freeing the
 479         * old structure.
 480         */
 481        void foo_update_a(int new_a)
 482        {
 483                struct foo *new_fp;
 484                struct foo *old_fp;
 485
 486                new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
 487                spin_lock(&foo_mutex);
 488                old_fp = gbl_foo;
 489                *new_fp = *old_fp;
 490                new_fp->a = new_a;
 491                rcu_assign_pointer(gbl_foo, new_fp);
 492                spin_unlock(&foo_mutex);
 493                call_rcu(&old_fp->rcu, foo_reclaim);
 494        }
 495
 496The foo_reclaim() function might appear as follows:
 497
 498        void foo_reclaim(struct rcu_head *rp)
 499        {
 500                struct foo *fp = container_of(rp, struct foo, rcu);
 501
 502                foo_cleanup(fp->a);
 503
 504                kfree(fp);
 505        }
 506
 507The container_of() primitive is a macro that, given a pointer into a
 508struct, the type of the struct, and the pointed-to field within the
 509struct, returns a pointer to the beginning of the struct.
 510
 511The use of call_rcu() permits the caller of foo_update_a() to
 512immediately regain control, without needing to worry further about the
 513old version of the newly updated element.  It also clearly shows the
 514RCU distinction between updater, namely foo_update_a(), and reclaimer,
 515namely foo_reclaim().
 516
 517The summary of advice is the same as for the previous section, except
 518that we are now using call_rcu() rather than synchronize_rcu():
 519
 520o       Use call_rcu() -after- removing a data element from an
 521        RCU-protected data structure in order to register a callback
 522        function that will be invoked after the completion of all RCU
 523        read-side critical sections that might be referencing that
 524        data item.
 525
 526If the callback for call_rcu() is not doing anything more than calling
 527kfree() on the structure, you can use kfree_rcu() instead of call_rcu()
 528to avoid having to write your own callback:
 529
 530        kfree_rcu(old_fp, rcu);
 531
 532Again, see checklist.txt for additional rules governing the use of RCU.
 533
 534
 5355.  WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
 536
 537One of the nice things about RCU is that it has extremely simple "toy"
 538implementations that are a good first step towards understanding the
 539production-quality implementations in the Linux kernel.  This section
 540presents two such "toy" implementations of RCU, one that is implemented
 541in terms of familiar locking primitives, and another that more closely
 542resembles "classic" RCU.  Both are way too simple for real-world use,
 543lacking both functionality and performance.  However, they are useful
 544in getting a feel for how RCU works.  See kernel/rcupdate.c for a
 545production-quality implementation, and see:
 546
 547        http://www.rdrop.com/users/paulmck/RCU
 548
 549for papers describing the Linux kernel RCU implementation.  The OLS'01
 550and OLS'02 papers are a good introduction, and the dissertation provides
 551more details on the current implementation as of early 2004.
 552
 553
 5545A.  "TOY" IMPLEMENTATION #1: LOCKING
 555
 556This section presents a "toy" RCU implementation that is based on
 557familiar locking primitives.  Its overhead makes it a non-starter for
 558real-life use, as does its lack of scalability.  It is also unsuitable
 559for realtime use, since it allows scheduling latency to "bleed" from
 560one read-side critical section to another.
 561
 562However, it is probably the easiest implementation to relate to, so is
 563a good starting point.
 564
 565It is extremely simple:
 566
 567        static DEFINE_RWLOCK(rcu_gp_mutex);
 568
 569        void rcu_read_lock(void)
 570        {
 571                read_lock(&rcu_gp_mutex);
 572        }
 573
 574        void rcu_read_unlock(void)
 575        {
 576                read_unlock(&rcu_gp_mutex);
 577        }
 578
 579        void synchronize_rcu(void)
 580        {
 581                write_lock(&rcu_gp_mutex);
 582                write_unlock(&rcu_gp_mutex);
 583        }
 584
 585[You can ignore rcu_assign_pointer() and rcu_dereference() without
 586missing much.  But here they are anyway.  And whatever you do, don't
 587forget about them when submitting patches making use of RCU!]
 588
 589        #define rcu_assign_pointer(p, v)        ({ \
 590                                                        smp_wmb(); \
 591                                                        (p) = (v); \
 592                                                })
 593
 594        #define rcu_dereference(p)     ({ \
 595                                        typeof(p) _________p1 = p; \
 596                                        smp_read_barrier_depends(); \
 597                                        (_________p1); \
 598                                        })
 599
 600
 601The rcu_read_lock() and rcu_read_unlock() primitive read-acquire
 602and release a global reader-writer lock.  The synchronize_rcu()
 603primitive write-acquires this same lock, then immediately releases
 604it.  This means that once synchronize_rcu() exits, all RCU read-side
 605critical sections that were in progress before synchronize_rcu() was
 606called are guaranteed to have completed -- there is no way that
 607synchronize_rcu() would have been able to write-acquire the lock
 608otherwise.
 609
 610It is possible to nest rcu_read_lock(), since reader-writer locks may
 611be recursively acquired.  Note also that rcu_read_lock() is immune
 612from deadlock (an important property of RCU).  The reason for this is
 613that the only thing that can block rcu_read_lock() is a synchronize_rcu().
 614But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex,
 615so there can be no deadlock cycle.
 616
 617Quick Quiz #1:  Why is this argument naive?  How could a deadlock
 618                occur when using this algorithm in a real-world Linux
 619                kernel?  How could this deadlock be avoided?
 620
 621
 6225B.  "TOY" EXAMPLE #2: CLASSIC RCU
 623
 624This section presents a "toy" RCU implementation that is based on
 625"classic RCU".  It is also short on performance (but only for updates) and
 626on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT
 627kernels.  The definitions of rcu_dereference() and rcu_assign_pointer()
 628are the same as those shown in the preceding section, so they are omitted.
 629
 630        void rcu_read_lock(void) { }
 631
 632        void rcu_read_unlock(void) { }
 633
 634        void synchronize_rcu(void)
 635        {
 636                int cpu;
 637
 638                for_each_possible_cpu(cpu)
 639                        run_on(cpu);
 640        }
 641
 642Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing.
 643This is the great strength of classic RCU in a non-preemptive kernel:
 644read-side overhead is precisely zero, at least on non-Alpha CPUs.
 645And there is absolutely no way that rcu_read_lock() can possibly
 646participate in a deadlock cycle!
 647
 648The implementation of synchronize_rcu() simply schedules itself on each
 649CPU in turn.  The run_on() primitive can be implemented straightforwardly
 650in terms of the sched_setaffinity() primitive.  Of course, a somewhat less
 651"toy" implementation would restore the affinity upon completion rather
 652than just leaving all tasks running on the last CPU, but when I said
 653"toy", I meant -toy-!
 654
 655So how the heck is this supposed to work???
 656
 657Remember that it is illegal to block while in an RCU read-side critical
 658section.  Therefore, if a given CPU executes a context switch, we know
 659that it must have completed all preceding RCU read-side critical sections.
 660Once -all- CPUs have executed a context switch, then -all- preceding
 661RCU read-side critical sections will have completed.
 662
 663So, suppose that we remove a data item from its structure and then invoke
 664synchronize_rcu().  Once synchronize_rcu() returns, we are guaranteed
 665that there are no RCU read-side critical sections holding a reference
 666to that data item, so we can safely reclaim it.
 667
 668Quick Quiz #2:  Give an example where Classic RCU's read-side
 669                overhead is -negative-.
 670
 671Quick Quiz #3:  If it is illegal to block in an RCU read-side
 672                critical section, what the heck do you do in
 673                PREEMPT_RT, where normal spinlocks can block???
 674
 675
 6766.  ANALOGY WITH READER-WRITER LOCKING
 677
 678Although RCU can be used in many different ways, a very common use of
 679RCU is analogous to reader-writer locking.  The following unified
 680diff shows how closely related RCU and reader-writer locking can be.
 681
 682        @@ -13,15 +14,15 @@
 683                struct list_head *lp;
 684                struct el *p;
 685
 686        -       read_lock();
 687        -       list_for_each_entry(p, head, lp) {
 688        +       rcu_read_lock();
 689        +       list_for_each_entry_rcu(p, head, lp) {
 690                        if (p->key == key) {
 691                                *result = p->data;
 692        -                       read_unlock();
 693        +                       rcu_read_unlock();
 694                                return 1;
 695                        }
 696                }
 697        -       read_unlock();
 698        +       rcu_read_unlock();
 699                return 0;
 700         }
 701
 702        @@ -29,15 +30,16 @@
 703         {
 704                struct el *p;
 705
 706        -       write_lock(&listmutex);
 707        +       spin_lock(&listmutex);
 708                list_for_each_entry(p, head, lp) {
 709                        if (p->key == key) {
 710        -                       list_del(&p->list);
 711        -                       write_unlock(&listmutex);
 712        +                       list_del_rcu(&p->list);
 713        +                       spin_unlock(&listmutex);
 714        +                       synchronize_rcu();
 715                                kfree(p);
 716                                return 1;
 717                        }
 718                }
 719        -       write_unlock(&listmutex);
 720        +       spin_unlock(&listmutex);
 721                return 0;
 722         }
 723
 724Or, for those who prefer a side-by-side listing:
 725
 726 1 struct el {                          1 struct el {
 727 2   struct list_head list;             2   struct list_head list;
 728 3   long key;                          3   long key;
 729 4   spinlock_t mutex;                  4   spinlock_t mutex;
 730 5   int data;                          5   int data;
 731 6   /* Other data fields */            6   /* Other data fields */
 732 7 };                                   7 };
 733 8 spinlock_t listmutex;                8 spinlock_t listmutex;
 734 9 struct el head;                      9 struct el head;
 735
 736 1 int search(long key, int *result)    1 int search(long key, int *result)
 737 2 {                                    2 {
 738 3   struct list_head *lp;              3   struct list_head *lp;
 739 4   struct el *p;                      4   struct el *p;
 740 5                                      5
 741 6   read_lock();                       6   rcu_read_lock();
 742 7   list_for_each_entry(p, head, lp) { 7   list_for_each_entry_rcu(p, head, lp) {
 743 8     if (p->key == key) {             8     if (p->key == key) {
 744 9       *result = p->data;             9       *result = p->data;
 74510       read_unlock();                10       rcu_read_unlock();
 74611       return 1;                     11       return 1;
 74712     }                               12     }
 74813   }                                 13   }
 74914   read_unlock();                    14   rcu_read_unlock();
 75015   return 0;                         15   return 0;
 75116 }                                   16 }
 752
 753 1 int delete(long key)                 1 int delete(long key)
 754 2 {                                    2 {
 755 3   struct el *p;                      3   struct el *p;
 756 4                                      4
 757 5   write_lock(&listmutex);            5   spin_lock(&listmutex);
 758 6   list_for_each_entry(p, head, lp) { 6   list_for_each_entry(p, head, lp) {
 759 7     if (p->key == key) {             7     if (p->key == key) {
 760 8       list_del(&p->list);            8       list_del_rcu(&p->list);
 761 9       write_unlock(&listmutex);      9       spin_unlock(&listmutex);
 762                                       10       synchronize_rcu();
 76310       kfree(p);                     11       kfree(p);
 76411       return 1;                     12       return 1;
 76512     }                               13     }
 76613   }                                 14   }
 76714   write_unlock(&listmutex);         15   spin_unlock(&listmutex);
 76815   return 0;                         16   return 0;
 76916 }                                   17 }
 770
 771Either way, the differences are quite small.  Read-side locking moves
 772to rcu_read_lock() and rcu_read_unlock, update-side locking moves from
 773a reader-writer lock to a simple spinlock, and a synchronize_rcu()
 774precedes the kfree().
 775
 776However, there is one potential catch: the read-side and update-side
 777critical sections can now run concurrently.  In many cases, this will
 778not be a problem, but it is necessary to check carefully regardless.
 779For example, if multiple independent list updates must be seen as
 780a single atomic update, converting to RCU will require special care.
 781
 782Also, the presence of synchronize_rcu() means that the RCU version of
 783delete() can now block.  If this is a problem, there is a callback-based
 784mechanism that never blocks, namely call_rcu() or kfree_rcu(), that can
 785be used in place of synchronize_rcu().
 786
 787
 7887.  FULL LIST OF RCU APIs
 789
 790The RCU APIs are documented in docbook-format header comments in the
 791Linux-kernel source code, but it helps to have a full list of the
 792APIs, since there does not appear to be a way to categorize them
 793in docbook.  Here is the list, by category.
 794
 795RCU list traversal:
 796
 797        list_for_each_entry_rcu
 798        hlist_for_each_entry_rcu
 799        hlist_nulls_for_each_entry_rcu
 800        list_for_each_entry_continue_rcu
 801
 802RCU pointer/list update:
 803
 804        rcu_assign_pointer
 805        list_add_rcu
 806        list_add_tail_rcu
 807        list_del_rcu
 808        list_replace_rcu
 809        hlist_del_rcu
 810        hlist_add_after_rcu
 811        hlist_add_before_rcu
 812        hlist_add_head_rcu
 813        hlist_replace_rcu
 814        list_splice_init_rcu()
 815
 816RCU:    Critical sections       Grace period            Barrier
 817
 818        rcu_read_lock           synchronize_net         rcu_barrier
 819        rcu_read_unlock         synchronize_rcu
 820        rcu_dereference         synchronize_rcu_expedited
 821                                call_rcu
 822                                kfree_rcu
 823
 824
 825bh:     Critical sections       Grace period            Barrier
 826
 827        rcu_read_lock_bh        call_rcu_bh             rcu_barrier_bh
 828        rcu_read_unlock_bh      synchronize_rcu_bh
 829        rcu_dereference_bh      synchronize_rcu_bh_expedited
 830
 831
 832sched:  Critical sections       Grace period            Barrier
 833
 834        rcu_read_lock_sched     synchronize_sched       rcu_barrier_sched
 835        rcu_read_unlock_sched   call_rcu_sched
 836        [preempt_disable]       synchronize_sched_expedited
 837        [and friends]
 838        rcu_dereference_sched
 839
 840
 841SRCU:   Critical sections       Grace period            Barrier
 842
 843        srcu_read_lock          synchronize_srcu        srcu_barrier
 844        srcu_read_unlock        call_srcu
 845        srcu_read_lock_raw      synchronize_srcu_expedited
 846        srcu_read_unlock_raw
 847        srcu_dereference
 848
 849SRCU:   Initialization/cleanup
 850        init_srcu_struct
 851        cleanup_srcu_struct
 852
 853All:  lockdep-checked RCU-protected pointer access
 854
 855        rcu_dereference_check
 856        rcu_dereference_protected
 857        rcu_access_pointer
 858
 859See the comment headers in the source code (or the docbook generated
 860from them) for more information.
 861
 862However, given that there are no fewer than four families of RCU APIs
 863in the Linux kernel, how do you choose which one to use?  The following
 864list can be helpful:
 865
 866a.      Will readers need to block?  If so, you need SRCU.
 867
 868b.      Is it necessary to start a read-side critical section in a
 869        hardirq handler or exception handler, and then to complete
 870        this read-side critical section in the task that was
 871        interrupted?  If so, you need SRCU's srcu_read_lock_raw() and
 872        srcu_read_unlock_raw() primitives.
 873
 874c.      What about the -rt patchset?  If readers would need to block
 875        in an non-rt kernel, you need SRCU.  If readers would block
 876        in a -rt kernel, but not in a non-rt kernel, SRCU is not
 877        necessary.
 878
 879d.      Do you need to treat NMI handlers, hardirq handlers,
 880        and code segments with preemption disabled (whether
 881        via preempt_disable(), local_irq_save(), local_bh_disable(),
 882        or some other mechanism) as if they were explicit RCU readers?
 883        If so, RCU-sched is the only choice that will work for you.
 884
 885e.      Do you need RCU grace periods to complete even in the face
 886        of softirq monopolization of one or more of the CPUs?  For
 887        example, is your code subject to network-based denial-of-service
 888        attacks?  If so, you need RCU-bh.
 889
 890f.      Is your workload too update-intensive for normal use of
 891        RCU, but inappropriate for other synchronization mechanisms?
 892        If so, consider SLAB_DESTROY_BY_RCU.  But please be careful!
 893
 894g.      Do you need read-side critical sections that are respected
 895        even though they are in the middle of the idle loop, during
 896        user-mode execution, or on an offlined CPU?  If so, SRCU is the
 897        only choice that will work for you.
 898
 899h.      Otherwise, use RCU.
 900
 901Of course, this all assumes that you have determined that RCU is in fact
 902the right tool for your job.
 903
 904
 9058.  ANSWERS TO QUICK QUIZZES
 906
 907Quick Quiz #1:  Why is this argument naive?  How could a deadlock
 908                occur when using this algorithm in a real-world Linux
 909                kernel?  [Referring to the lock-based "toy" RCU
 910                algorithm.]
 911
 912Answer:         Consider the following sequence of events:
 913
 914                1.      CPU 0 acquires some unrelated lock, call it
 915                        "problematic_lock", disabling irq via
 916                        spin_lock_irqsave().
 917
 918                2.      CPU 1 enters synchronize_rcu(), write-acquiring
 919                        rcu_gp_mutex.
 920
 921                3.      CPU 0 enters rcu_read_lock(), but must wait
 922                        because CPU 1 holds rcu_gp_mutex.
 923
 924                4.      CPU 1 is interrupted, and the irq handler
 925                        attempts to acquire problematic_lock.
 926
 927                The system is now deadlocked.
 928
 929                One way to avoid this deadlock is to use an approach like
 930                that of CONFIG_PREEMPT_RT, where all normal spinlocks
 931                become blocking locks, and all irq handlers execute in
 932                the context of special tasks.  In this case, in step 4
 933                above, the irq handler would block, allowing CPU 1 to
 934                release rcu_gp_mutex, avoiding the deadlock.
 935
 936                Even in the absence of deadlock, this RCU implementation
 937                allows latency to "bleed" from readers to other
 938                readers through synchronize_rcu().  To see this,
 939                consider task A in an RCU read-side critical section
 940                (thus read-holding rcu_gp_mutex), task B blocked
 941                attempting to write-acquire rcu_gp_mutex, and
 942                task C blocked in rcu_read_lock() attempting to
 943                read_acquire rcu_gp_mutex.  Task A's RCU read-side
 944                latency is holding up task C, albeit indirectly via
 945                task B.
 946
 947                Realtime RCU implementations therefore use a counter-based
 948                approach where tasks in RCU read-side critical sections
 949                cannot be blocked by tasks executing synchronize_rcu().
 950
 951Quick Quiz #2:  Give an example where Classic RCU's read-side
 952                overhead is -negative-.
 953
 954Answer:         Imagine a single-CPU system with a non-CONFIG_PREEMPT
 955                kernel where a routing table is used by process-context
 956                code, but can be updated by irq-context code (for example,
 957                by an "ICMP REDIRECT" packet).  The usual way of handling
 958                this would be to have the process-context code disable
 959                interrupts while searching the routing table.  Use of
 960                RCU allows such interrupt-disabling to be dispensed with.
 961                Thus, without RCU, you pay the cost of disabling interrupts,
 962                and with RCU you don't.
 963
 964                One can argue that the overhead of RCU in this
 965                case is negative with respect to the single-CPU
 966                interrupt-disabling approach.  Others might argue that
 967                the overhead of RCU is merely zero, and that replacing
 968                the positive overhead of the interrupt-disabling scheme
 969                with the zero-overhead RCU scheme does not constitute
 970                negative overhead.
 971
 972                In real life, of course, things are more complex.  But
 973                even the theoretical possibility of negative overhead for
 974                a synchronization primitive is a bit unexpected.  ;-)
 975
 976Quick Quiz #3:  If it is illegal to block in an RCU read-side
 977                critical section, what the heck do you do in
 978                PREEMPT_RT, where normal spinlocks can block???
 979
 980Answer:         Just as PREEMPT_RT permits preemption of spinlock
 981                critical sections, it permits preemption of RCU
 982                read-side critical sections.  It also permits
 983                spinlocks blocking while in RCU read-side critical
 984                sections.
 985
 986                Why the apparent inconsistency?  Because it is it
 987                possible to use priority boosting to keep the RCU
 988                grace periods short if need be (for example, if running
 989                short of memory).  In contrast, if blocking waiting
 990                for (say) network reception, there is no way to know
 991                what should be boosted.  Especially given that the
 992                process we need to boost might well be a human being
 993                who just went out for a pizza or something.  And although
 994                a computer-operated cattle prod might arouse serious
 995                interest, it might also provoke serious objections.
 996                Besides, how does the computer know what pizza parlor
 997                the human being went to???
 998
 999
1000ACKNOWLEDGEMENTS
1001
1002My thanks to the people who helped make this human-readable, including
1003Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern.
1004
1005
1006For more information, see http://www.rdrop.com/users/paulmck/RCU.
1007
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.