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                kfree(fp);
 503        }
 504
 505The container_of() primitive is a macro that, given a pointer into a
 506struct, the type of the struct, and the pointed-to field within the
 507struct, returns a pointer to the beginning of the struct.
 508
 509The use of call_rcu() permits the caller of foo_update_a() to
 510immediately regain control, without needing to worry further about the
 511old version of the newly updated element.  It also clearly shows the
 512RCU distinction between updater, namely foo_update_a(), and reclaimer,
 513namely foo_reclaim().
 514
 515The summary of advice is the same as for the previous section, except
 516that we are now using call_rcu() rather than synchronize_rcu():
 517
 518o       Use call_rcu() -after- removing a data element from an
 519        RCU-protected data structure in order to register a callback
 520        function that will be invoked after the completion of all RCU
 521        read-side critical sections that might be referencing that
 522        data item.
 523
 524Again, see checklist.txt for additional rules governing the use of RCU.
 525
 526
 5275.  WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
 528
 529One of the nice things about RCU is that it has extremely simple "toy"
 530implementations that are a good first step towards understanding the
 531production-quality implementations in the Linux kernel.  This section
 532presents two such "toy" implementations of RCU, one that is implemented
 533in terms of familiar locking primitives, and another that more closely
 534resembles "classic" RCU.  Both are way too simple for real-world use,
 535lacking both functionality and performance.  However, they are useful
 536in getting a feel for how RCU works.  See kernel/rcupdate.c for a
 537production-quality implementation, and see:
 538
 539        http://www.rdrop.com/users/paulmck/RCU
 540
 541for papers describing the Linux kernel RCU implementation.  The OLS'01
 542and OLS'02 papers are a good introduction, and the dissertation provides
 543more details on the current implementation as of early 2004.
 544
 545
 5465A.  "TOY" IMPLEMENTATION #1: LOCKING
 547
 548This section presents a "toy" RCU implementation that is based on
 549familiar locking primitives.  Its overhead makes it a non-starter for
 550real-life use, as does its lack of scalability.  It is also unsuitable
 551for realtime use, since it allows scheduling latency to "bleed" from
 552one read-side critical section to another.
 553
 554However, it is probably the easiest implementation to relate to, so is
 555a good starting point.
 556
 557It is extremely simple:
 558
 559        static DEFINE_RWLOCK(rcu_gp_mutex);
 560
 561        void rcu_read_lock(void)
 562        {
 563                read_lock(&rcu_gp_mutex);
 564        }
 565
 566        void rcu_read_unlock(void)
 567        {
 568                read_unlock(&rcu_gp_mutex);
 569        }
 570
 571        void synchronize_rcu(void)
 572        {
 573                write_lock(&rcu_gp_mutex);
 574                write_unlock(&rcu_gp_mutex);
 575        }
 576
 577[You can ignore rcu_assign_pointer() and rcu_dereference() without
 578missing much.  But here they are anyway.  And whatever you do, don't
 579forget about them when submitting patches making use of RCU!]
 580
 581        #define rcu_assign_pointer(p, v)        ({ \
 582                                                        smp_wmb(); \
 583                                                        (p) = (v); \
 584                                                })
 585
 586        #define rcu_dereference(p)     ({ \
 587                                        typeof(p) _________p1 = p; \
 588                                        smp_read_barrier_depends(); \
 589                                        (_________p1); \
 590                                        })
 591
 592
 593The rcu_read_lock() and rcu_read_unlock() primitive read-acquire
 594and release a global reader-writer lock.  The synchronize_rcu()
 595primitive write-acquires this same lock, then immediately releases
 596it.  This means that once synchronize_rcu() exits, all RCU read-side
 597critical sections that were in progress before synchronize_rcu() was
 598called are guaranteed to have completed -- there is no way that
 599synchronize_rcu() would have been able to write-acquire the lock
 600otherwise.
 601
 602It is possible to nest rcu_read_lock(), since reader-writer locks may
 603be recursively acquired.  Note also that rcu_read_lock() is immune
 604from deadlock (an important property of RCU).  The reason for this is
 605that the only thing that can block rcu_read_lock() is a synchronize_rcu().
 606But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex,
 607so there can be no deadlock cycle.
 608
 609Quick Quiz #1:  Why is this argument naive?  How could a deadlock
 610                occur when using this algorithm in a real-world Linux
 611                kernel?  How could this deadlock be avoided?
 612
 613
 6145B.  "TOY" EXAMPLE #2: CLASSIC RCU
 615
 616This section presents a "toy" RCU implementation that is based on
 617"classic RCU".  It is also short on performance (but only for updates) and
 618on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT
 619kernels.  The definitions of rcu_dereference() and rcu_assign_pointer()
 620are the same as those shown in the preceding section, so they are omitted.
 621
 622        void rcu_read_lock(void) { }
 623
 624        void rcu_read_unlock(void) { }
 625
 626        void synchronize_rcu(void)
 627        {
 628                int cpu;
 629
 630                for_each_possible_cpu(cpu)
 631                        run_on(cpu);
 632        }
 633
 634Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing.
 635This is the great strength of classic RCU in a non-preemptive kernel:
 636read-side overhead is precisely zero, at least on non-Alpha CPUs.
 637And there is absolutely no way that rcu_read_lock() can possibly
 638participate in a deadlock cycle!
 639
 640The implementation of synchronize_rcu() simply schedules itself on each
 641CPU in turn.  The run_on() primitive can be implemented straightforwardly
 642in terms of the sched_setaffinity() primitive.  Of course, a somewhat less
 643"toy" implementation would restore the affinity upon completion rather
 644than just leaving all tasks running on the last CPU, but when I said
 645"toy", I meant -toy-!
 646
 647So how the heck is this supposed to work???
 648
 649Remember that it is illegal to block while in an RCU read-side critical
 650section.  Therefore, if a given CPU executes a context switch, we know
 651that it must have completed all preceding RCU read-side critical sections.
 652Once -all- CPUs have executed a context switch, then -all- preceding
 653RCU read-side critical sections will have completed.
 654
 655So, suppose that we remove a data item from its structure and then invoke
 656synchronize_rcu().  Once synchronize_rcu() returns, we are guaranteed
 657that there are no RCU read-side critical sections holding a reference
 658to that data item, so we can safely reclaim it.
 659
 660Quick Quiz #2:  Give an example where Classic RCU's read-side
 661                overhead is -negative-.
 662
 663Quick Quiz #3:  If it is illegal to block in an RCU read-side
 664                critical section, what the heck do you do in
 665                PREEMPT_RT, where normal spinlocks can block???
 666
 667
 6686.  ANALOGY WITH READER-WRITER LOCKING
 669
 670Although RCU can be used in many different ways, a very common use of
 671RCU is analogous to reader-writer locking.  The following unified
 672diff shows how closely related RCU and reader-writer locking can be.
 673
 674        @@ -13,15 +14,15 @@
 675                struct list_head *lp;
 676                struct el *p;
 677
 678        -       read_lock();
 679        -       list_for_each_entry(p, head, lp) {
 680        +       rcu_read_lock();
 681        +       list_for_each_entry_rcu(p, head, lp) {
 682                        if (p->key == key) {
 683                                *result = p->data;
 684        -                       read_unlock();
 685        +                       rcu_read_unlock();
 686                                return 1;
 687                        }
 688                }
 689        -       read_unlock();
 690        +       rcu_read_unlock();
 691                return 0;
 692         }
 693
 694        @@ -29,15 +30,16 @@
 695         {
 696                struct el *p;
 697
 698        -       write_lock(&listmutex);
 699        +       spin_lock(&listmutex);
 700                list_for_each_entry(p, head, lp) {
 701                        if (p->key == key) {
 702        -                       list_del(&p->list);
 703        -                       write_unlock(&listmutex);
 704        +                       list_del_rcu(&p->list);
 705        +                       spin_unlock(&listmutex);
 706        +                       synchronize_rcu();
 707                                kfree(p);
 708                                return 1;
 709                        }
 710                }
 711        -       write_unlock(&listmutex);
 712        +       spin_unlock(&listmutex);
 713                return 0;
 714         }
 715
 716Or, for those who prefer a side-by-side listing:
 717
 718 1 struct el {                          1 struct el {
 719 2   struct list_head list;             2   struct list_head list;
 720 3   long key;                          3   long key;
 721 4   spinlock_t mutex;                  4   spinlock_t mutex;
 722 5   int data;                          5   int data;
 723 6   /* Other data fields */            6   /* Other data fields */
 724 7 };                                   7 };
 725 8 spinlock_t listmutex;                8 spinlock_t listmutex;
 726 9 struct el head;                      9 struct el head;
 727
 728 1 int search(long key, int *result)    1 int search(long key, int *result)
 729 2 {                                    2 {
 730 3   struct list_head *lp;              3   struct list_head *lp;
 731 4   struct el *p;                      4   struct el *p;
 732 5                                      5
 733 6   read_lock();                       6   rcu_read_lock();
 734 7   list_for_each_entry(p, head, lp) { 7   list_for_each_entry_rcu(p, head, lp) {
 735 8     if (p->key == key) {             8     if (p->key == key) {
 736 9       *result = p->data;             9       *result = p->data;
 73710       read_unlock();                10       rcu_read_unlock();
 73811       return 1;                     11       return 1;
 73912     }                               12     }
 74013   }                                 13   }
 74114   read_unlock();                    14   rcu_read_unlock();
 74215   return 0;                         15   return 0;
 74316 }                                   16 }
 744
 745 1 int delete(long key)                 1 int delete(long key)
 746 2 {                                    2 {
 747 3   struct el *p;                      3   struct el *p;
 748 4                                      4
 749 5   write_lock(&listmutex);            5   spin_lock(&listmutex);
 750 6   list_for_each_entry(p, head, lp) { 6   list_for_each_entry(p, head, lp) {
 751 7     if (p->key == key) {             7     if (p->key == key) {
 752 8       list_del(&p->list);            8       list_del_rcu(&p->list);
 753 9       write_unlock(&listmutex);      9       spin_unlock(&listmutex);
 754                                       10       synchronize_rcu();
 75510       kfree(p);                     11       kfree(p);
 75611       return 1;                     12       return 1;
 75712     }                               13     }
 75813   }                                 14   }
 75914   write_unlock(&listmutex);         15   spin_unlock(&listmutex);
 76015   return 0;                         16   return 0;
 76116 }                                   17 }
 762
 763Either way, the differences are quite small.  Read-side locking moves
 764to rcu_read_lock() and rcu_read_unlock, update-side locking moves from
 765a reader-writer lock to a simple spinlock, and a synchronize_rcu()
 766precedes the kfree().
 767
 768However, there is one potential catch: the read-side and update-side
 769critical sections can now run concurrently.  In many cases, this will
 770not be a problem, but it is necessary to check carefully regardless.
 771For example, if multiple independent list updates must be seen as
 772a single atomic update, converting to RCU will require special care.
 773
 774Also, the presence of synchronize_rcu() means that the RCU version of
 775delete() can now block.  If this is a problem, there is a callback-based
 776mechanism that never blocks, namely call_rcu(), that can be used in
 777place of synchronize_rcu().
 778
 779
 7807.  FULL LIST OF RCU APIs
 781
 782The RCU APIs are documented in docbook-format header comments in the
 783Linux-kernel source code, but it helps to have a full list of the
 784APIs, since there does not appear to be a way to categorize them
 785in docbook.  Here is the list, by category.
 786
 787RCU list traversal:
 788
 789        list_for_each_entry_rcu
 790        hlist_for_each_entry_rcu
 791        hlist_nulls_for_each_entry_rcu
 792
 793        list_for_each_continue_rcu      (to be deprecated in favor of new
 794                                         list_for_each_entry_continue_rcu)
 795
 796RCU pointer/list update:
 797
 798        rcu_assign_pointer
 799        list_add_rcu
 800        list_add_tail_rcu
 801        list_del_rcu
 802        list_replace_rcu
 803        hlist_del_rcu
 804        hlist_add_after_rcu
 805        hlist_add_before_rcu
 806        hlist_add_head_rcu
 807        hlist_replace_rcu
 808        list_splice_init_rcu()
 809
 810RCU:    Critical sections       Grace period            Barrier
 811
 812        rcu_read_lock           synchronize_net         rcu_barrier
 813        rcu_read_unlock         synchronize_rcu
 814        rcu_dereference         synchronize_rcu_expedited
 815                                call_rcu
 816
 817
 818bh:     Critical sections       Grace period            Barrier
 819
 820        rcu_read_lock_bh        call_rcu_bh             rcu_barrier_bh
 821        rcu_read_unlock_bh      synchronize_rcu_bh
 822        rcu_dereference_bh      synchronize_rcu_bh_expedited
 823
 824
 825sched:  Critical sections       Grace period            Barrier
 826
 827        rcu_read_lock_sched     synchronize_sched       rcu_barrier_sched
 828        rcu_read_unlock_sched   call_rcu_sched
 829        [preempt_disable]       synchronize_sched_expedited
 830        [and friends]
 831        rcu_dereference_sched
 832
 833
 834SRCU:   Critical sections       Grace period            Barrier
 835
 836        srcu_read_lock          synchronize_srcu        srcu_barrier
 837        srcu_read_unlock        call_srcu
 838        srcu_read_lock_raw      synchronize_srcu_expedited
 839        srcu_read_unlock_raw
 840        srcu_dereference
 841
 842SRCU:   Initialization/cleanup
 843        init_srcu_struct
 844        cleanup_srcu_struct
 845
 846All:  lockdep-checked RCU-protected pointer access
 847
 848        rcu_dereference_check
 849        rcu_dereference_protected
 850        rcu_access_pointer
 851
 852See the comment headers in the source code (or the docbook generated
 853from them) for more information.
 854
 855However, given that there are no fewer than four families of RCU APIs
 856in the Linux kernel, how do you choose which one to use?  The following
 857list can be helpful:
 858
 859a.      Will readers need to block?  If so, you need SRCU.
 860
 861b.      Is it necessary to start a read-side critical section in a
 862        hardirq handler or exception handler, and then to complete
 863        this read-side critical section in the task that was
 864        interrupted?  If so, you need SRCU's srcu_read_lock_raw() and
 865        srcu_read_unlock_raw() primitives.
 866
 867c.      What about the -rt patchset?  If readers would need to block
 868        in an non-rt kernel, you need SRCU.  If readers would block
 869        in a -rt kernel, but not in a non-rt kernel, SRCU is not
 870        necessary.
 871
 872d.      Do you need to treat NMI handlers, hardirq handlers,
 873        and code segments with preemption disabled (whether
 874        via preempt_disable(), local_irq_save(), local_bh_disable(),
 875        or some other mechanism) as if they were explicit RCU readers?
 876        If so, RCU-sched is the only choice that will work for you.
 877
 878e.      Do you need RCU grace periods to complete even in the face
 879        of softirq monopolization of one or more of the CPUs?  For
 880        example, is your code subject to network-based denial-of-service
 881        attacks?  If so, you need RCU-bh.
 882
 883f.      Is your workload too update-intensive for normal use of
 884        RCU, but inappropriate for other synchronization mechanisms?
 885        If so, consider SLAB_DESTROY_BY_RCU.  But please be careful!
 886
 887g.      Do you need read-side critical sections that are respected
 888        even though they are in the middle of the idle loop, during
 889        user-mode execution, or on an offlined CPU?  If so, SRCU is the
 890        only choice that will work for you.
 891
 892h.      Otherwise, use RCU.
 893
 894Of course, this all assumes that you have determined that RCU is in fact
 895the right tool for your job.
 896
 897
 8988.  ANSWERS TO QUICK QUIZZES
 899
 900Quick Quiz #1:  Why is this argument naive?  How could a deadlock
 901                occur when using this algorithm in a real-world Linux
 902                kernel?  [Referring to the lock-based "toy" RCU
 903                algorithm.]
 904
 905Answer:         Consider the following sequence of events:
 906
 907                1.      CPU 0 acquires some unrelated lock, call it
 908                        "problematic_lock", disabling irq via
 909                        spin_lock_irqsave().
 910
 911                2.      CPU 1 enters synchronize_rcu(), write-acquiring
 912                        rcu_gp_mutex.
 913
 914                3.      CPU 0 enters rcu_read_lock(), but must wait
 915                        because CPU 1 holds rcu_gp_mutex.
 916
 917                4.      CPU 1 is interrupted, and the irq handler
 918                        attempts to acquire problematic_lock.
 919
 920                The system is now deadlocked.
 921
 922                One way to avoid this deadlock is to use an approach like
 923                that of CONFIG_PREEMPT_RT, where all normal spinlocks
 924                become blocking locks, and all irq handlers execute in
 925                the context of special tasks.  In this case, in step 4
 926                above, the irq handler would block, allowing CPU 1 to
 927                release rcu_gp_mutex, avoiding the deadlock.
 928
 929                Even in the absence of deadlock, this RCU implementation
 930                allows latency to "bleed" from readers to other
 931                readers through synchronize_rcu().  To see this,
 932                consider task A in an RCU read-side critical section
 933                (thus read-holding rcu_gp_mutex), task B blocked
 934                attempting to write-acquire rcu_gp_mutex, and
 935                task C blocked in rcu_read_lock() attempting to
 936                read_acquire rcu_gp_mutex.  Task A's RCU read-side
 937                latency is holding up task C, albeit indirectly via
 938                task B.
 939
 940                Realtime RCU implementations therefore use a counter-based
 941                approach where tasks in RCU read-side critical sections
 942                cannot be blocked by tasks executing synchronize_rcu().
 943
 944Quick Quiz #2:  Give an example where Classic RCU's read-side
 945                overhead is -negative-.
 946
 947Answer:         Imagine a single-CPU system with a non-CONFIG_PREEMPT
 948                kernel where a routing table is used by process-context
 949                code, but can be updated by irq-context code (for example,
 950                by an "ICMP REDIRECT" packet).  The usual way of handling
 951                this would be to have the process-context code disable
 952                interrupts while searching the routing table.  Use of
 953                RCU allows such interrupt-disabling to be dispensed with.
 954                Thus, without RCU, you pay the cost of disabling interrupts,
 955                and with RCU you don't.
 956
 957                One can argue that the overhead of RCU in this
 958                case is negative with respect to the single-CPU
 959                interrupt-disabling approach.  Others might argue that
 960                the overhead of RCU is merely zero, and that replacing
 961                the positive overhead of the interrupt-disabling scheme
 962                with the zero-overhead RCU scheme does not constitute
 963                negative overhead.
 964
 965                In real life, of course, things are more complex.  But
 966                even the theoretical possibility of negative overhead for
 967                a synchronization primitive is a bit unexpected.  ;-)
 968
 969Quick Quiz #3:  If it is illegal to block in an RCU read-side
 970                critical section, what the heck do you do in
 971                PREEMPT_RT, where normal spinlocks can block???
 972
 973Answer:         Just as PREEMPT_RT permits preemption of spinlock
 974                critical sections, it permits preemption of RCU
 975                read-side critical sections.  It also permits
 976                spinlocks blocking while in RCU read-side critical
 977                sections.
 978
 979                Why the apparent inconsistency?  Because it is it
 980                possible to use priority boosting to keep the RCU
 981                grace periods short if need be (for example, if running
 982                short of memory).  In contrast, if blocking waiting
 983                for (say) network reception, there is no way to know
 984                what should be boosted.  Especially given that the
 985                process we need to boost might well be a human being
 986                who just went out for a pizza or something.  And although
 987                a computer-operated cattle prod might arouse serious
 988                interest, it might also provoke serious objections.
 989                Besides, how does the computer know what pizza parlor
 990                the human being went to???
 991
 992
 993ACKNOWLEDGEMENTS
 994
 995My thanks to the people who helped make this human-readable, including
 996Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern.
 997
 998
 999For more information, see http://www.rdrop.com/users/paulmck/RCU.
1000
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.