3 What is RCU? -- "Read, Copy, Update"
4 ======================================
6 Please note that the "What is RCU?" LWN series is an excellent place
7 to start learning about RCU:
9 | 1. What is RCU, Fundamentally? http://lwn.net/Articles/262464/
10 | 2. What is RCU? Part 2: Usage http://lwn.net/Articles/263130/
11 | 3. RCU part 3: the RCU API http://lwn.net/Articles/264090/
12 | 4. The RCU API, 2010 Edition http://lwn.net/Articles/418853/
13 | 2010 Big API Table http://lwn.net/Articles/419086/
14 | 5. The RCU API, 2014 Edition http://lwn.net/Articles/609904/
15 | 2014 Big API Table http://lwn.net/Articles/609973/
20 RCU is a synchronization mechanism that was added to the Linux kernel
21 during the 2.5 development effort that is optimized for read-mostly
22 situations. Although RCU is actually quite simple once you understand it,
23 getting there can sometimes be a challenge. Part of the problem is that
24 most of the past descriptions of RCU have been written with the mistaken
25 assumption that there is "one true way" to describe RCU. Instead,
26 the experience has been that different people must take different paths
27 to arrive at an understanding of RCU. This document provides several
28 different paths, as follows:
30 :ref:`1. RCU OVERVIEW <1_whatisRCU>`
32 :ref:`2. WHAT IS RCU'S CORE API? <2_whatisRCU>`
34 :ref:`3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API? <3_whatisRCU>`
36 :ref:`4. WHAT IF MY UPDATING THREAD CANNOT BLOCK? <4_whatisRCU>`
38 :ref:`5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU? <5_whatisRCU>`
40 :ref:`6. ANALOGY WITH READER-WRITER LOCKING <6_whatisRCU>`
42 :ref:`7. ANALOGY WITH REFERENCE COUNTING <7_whatisRCU>`
44 :ref:`8. FULL LIST OF RCU APIs <8_whatisRCU>`
46 :ref:`9. ANSWERS TO QUICK QUIZZES <9_whatisRCU>`
48 People who prefer starting with a conceptual overview should focus on
49 Section 1, though most readers will profit by reading this section at
50 some point. People who prefer to start with an API that they can then
51 experiment with should focus on Section 2. People who prefer to start
52 with example uses should focus on Sections 3 and 4. People who need to
53 understand the RCU implementation should focus on Section 5, then dive
54 into the kernel source code. People who reason best by analogy should
55 focus on Section 6. Section 7 serves as an index to the docbook API
56 documentation, and Section 8 is the traditional answer key.
58 So, start with the section that makes the most sense to you and your
59 preferred method of learning. If you need to know everything about
60 everything, feel free to read the whole thing -- but if you are really
61 that type of person, you have perused the source code and will therefore
62 never need this document anyway. ;-)
69 The basic idea behind RCU is to split updates into "removal" and
70 "reclamation" phases. The removal phase removes references to data items
71 within a data structure (possibly by replacing them with references to
72 new versions of these data items), and can run concurrently with readers.
73 The reason that it is safe to run the removal phase concurrently with
74 readers is the semantics of modern CPUs guarantee that readers will see
75 either the old or the new version of the data structure rather than a
76 partially updated reference. The reclamation phase does the work of reclaiming
77 (e.g., freeing) the data items removed from the data structure during the
78 removal phase. Because reclaiming data items can disrupt any readers
79 concurrently referencing those data items, the reclamation phase must
80 not start until readers no longer hold references to those data items.
82 Splitting the update into removal and reclamation phases permits the
83 updater to perform the removal phase immediately, and to defer the
84 reclamation phase until all readers active during the removal phase have
85 completed, either by blocking until they finish or by registering a
86 callback that is invoked after they finish. Only readers that are active
87 during the removal phase need be considered, because any reader starting
88 after the removal phase will be unable to gain a reference to the removed
89 data items, and therefore cannot be disrupted by the reclamation phase.
91 So the typical RCU update sequence goes something like the following:
93 a. Remove pointers to a data structure, so that subsequent
94 readers cannot gain a reference to it.
96 b. Wait for all previous readers to complete their RCU read-side
99 c. At this point, there cannot be any readers who hold references
100 to the data structure, so it now may safely be reclaimed
103 Step (b) above is the key idea underlying RCU's deferred destruction.
104 The ability to wait until all readers are done allows RCU readers to
105 use much lighter-weight synchronization, in some cases, absolutely no
106 synchronization at all. In contrast, in more conventional lock-based
107 schemes, readers must use heavy-weight synchronization in order to
108 prevent an updater from deleting the data structure out from under them.
109 This is because lock-based updaters typically update data items in place,
110 and must therefore exclude readers. In contrast, RCU-based updaters
111 typically take advantage of the fact that writes to single aligned
112 pointers are atomic on modern CPUs, allowing atomic insertion, removal,
113 and replacement of data items in a linked structure without disrupting
114 readers. Concurrent RCU readers can then continue accessing the old
115 versions, and can dispense with the atomic operations, memory barriers,
116 and communications cache misses that are so expensive on present-day
117 SMP computer systems, even in absence of lock contention.
119 In the three-step procedure shown above, the updater is performing both
120 the removal and the reclamation step, but it is often helpful for an
121 entirely different thread to do the reclamation, as is in fact the case
122 in the Linux kernel's directory-entry cache (dcache). Even if the same
123 thread performs both the update step (step (a) above) and the reclamation
124 step (step (c) above), it is often helpful to think of them separately.
125 For example, RCU readers and updaters need not communicate at all,
126 but RCU provides implicit low-overhead communication between readers
127 and reclaimers, namely, in step (b) above.
129 So how the heck can a reclaimer tell when a reader is done, given
130 that readers are not doing any sort of synchronization operations???
131 Read on to learn about how RCU's API makes this easy.
135 2. WHAT IS RCU'S CORE API?
136 ---------------------------
138 The core RCU API is quite small:
142 c. synchronize_rcu() / call_rcu()
143 d. rcu_assign_pointer()
146 There are many other members of the RCU API, but the rest can be
147 expressed in terms of these five, though most implementations instead
148 express synchronize_rcu() in terms of the call_rcu() callback API.
150 The five core RCU APIs are described below, the other 18 will be enumerated
151 later. See the kernel docbook documentation for more info, or look directly
152 at the function header comments.
156 void rcu_read_lock(void);
158 Used by a reader to inform the reclaimer that the reader is
159 entering an RCU read-side critical section. It is illegal
160 to block while in an RCU read-side critical section, though
161 kernels built with CONFIG_PREEMPT_RCU can preempt RCU
162 read-side critical sections. Any RCU-protected data structure
163 accessed during an RCU read-side critical section is guaranteed to
164 remain unreclaimed for the full duration of that critical section.
165 Reference counts may be used in conjunction with RCU to maintain
166 longer-term references to data structures.
170 void rcu_read_unlock(void);
172 Used by a reader to inform the reclaimer that the reader is
173 exiting an RCU read-side critical section. Note that RCU
174 read-side critical sections may be nested and/or overlapping.
178 void synchronize_rcu(void);
180 Marks the end of updater code and the beginning of reclaimer
181 code. It does this by blocking until all pre-existing RCU
182 read-side critical sections on all CPUs have completed.
183 Note that synchronize_rcu() will **not** necessarily wait for
184 any subsequent RCU read-side critical sections to complete.
185 For example, consider the following sequence of events::
188 ----------------- ------------------------- ---------------
190 2. enters synchronize_rcu()
193 5. exits synchronize_rcu()
196 To reiterate, synchronize_rcu() waits only for ongoing RCU
197 read-side critical sections to complete, not necessarily for
198 any that begin after synchronize_rcu() is invoked.
200 Of course, synchronize_rcu() does not necessarily return
201 **immediately** after the last pre-existing RCU read-side critical
202 section completes. For one thing, there might well be scheduling
203 delays. For another thing, many RCU implementations process
204 requests in batches in order to improve efficiencies, which can
205 further delay synchronize_rcu().
207 Since synchronize_rcu() is the API that must figure out when
208 readers are done, its implementation is key to RCU. For RCU
209 to be useful in all but the most read-intensive situations,
210 synchronize_rcu()'s overhead must also be quite small.
212 The call_rcu() API is a callback form of synchronize_rcu(),
213 and is described in more detail in a later section. Instead of
214 blocking, it registers a function and argument which are invoked
215 after all ongoing RCU read-side critical sections have completed.
216 This callback variant is particularly useful in situations where
217 it is illegal to block or where update-side performance is
218 critically important.
220 However, the call_rcu() API should not be used lightly, as use
221 of the synchronize_rcu() API generally results in simpler code.
222 In addition, the synchronize_rcu() API has the nice property
223 of automatically limiting update rate should grace periods
224 be delayed. This property results in system resilience in face
225 of denial-of-service attacks. Code using call_rcu() should limit
226 update rate in order to gain this same sort of resilience. See
227 checklist.txt for some approaches to limiting the update rate.
231 void rcu_assign_pointer(p, typeof(p) v);
233 Yes, rcu_assign_pointer() **is** implemented as a macro, though it
234 would be cool to be able to declare a function in this manner.
235 (Compiler experts will no doubt disagree.)
237 The updater uses this function to assign a new value to an
238 RCU-protected pointer, in order to safely communicate the change
239 in value from the updater to the reader. This macro does not
240 evaluate to an rvalue, but it does execute any memory-barrier
241 instructions required for a given CPU architecture.
243 Perhaps just as important, it serves to document (1) which
244 pointers are protected by RCU and (2) the point at which a
245 given structure becomes accessible to other CPUs. That said,
246 rcu_assign_pointer() is most frequently used indirectly, via
247 the _rcu list-manipulation primitives such as list_add_rcu().
251 typeof(p) rcu_dereference(p);
253 Like rcu_assign_pointer(), rcu_dereference() must be implemented
256 The reader uses rcu_dereference() to fetch an RCU-protected
257 pointer, which returns a value that may then be safely
258 dereferenced. Note that rcu_dereference() does not actually
259 dereference the pointer, instead, it protects the pointer for
260 later dereferencing. It also executes any needed memory-barrier
261 instructions for a given CPU architecture. Currently, only Alpha
262 needs memory barriers within rcu_dereference() -- on other CPUs,
263 it compiles to nothing, not even a compiler directive.
265 Common coding practice uses rcu_dereference() to copy an
266 RCU-protected pointer to a local variable, then dereferences
267 this local variable, for example as follows::
269 p = rcu_dereference(head.next);
272 However, in this case, one could just as easily combine these
275 return rcu_dereference(head.next)->data;
277 If you are going to be fetching multiple fields from the
278 RCU-protected structure, using the local variable is of
279 course preferred. Repeated rcu_dereference() calls look
280 ugly, do not guarantee that the same pointer will be returned
281 if an update happened while in the critical section, and incur
282 unnecessary overhead on Alpha CPUs.
284 Note that the value returned by rcu_dereference() is valid
285 only within the enclosing RCU read-side critical section [1]_.
286 For example, the following is **not** legal::
289 p = rcu_dereference(head.next);
291 x = p->address; /* BUG!!! */
293 y = p->data; /* BUG!!! */
296 Holding a reference from one RCU read-side critical section
297 to another is just as illegal as holding a reference from
298 one lock-based critical section to another! Similarly,
299 using a reference outside of the critical section in which
300 it was acquired is just as illegal as doing so with normal
303 As with rcu_assign_pointer(), an important function of
304 rcu_dereference() is to document which pointers are protected by
305 RCU, in particular, flagging a pointer that is subject to changing
306 at any time, including immediately after the rcu_dereference().
307 And, again like rcu_assign_pointer(), rcu_dereference() is
308 typically used indirectly, via the _rcu list-manipulation
309 primitives, such as list_for_each_entry_rcu() [2]_.
311 .. [1] The variant rcu_dereference_protected() can be used outside
312 of an RCU read-side critical section as long as the usage is
313 protected by locks acquired by the update-side code. This variant
314 avoids the lockdep warning that would happen when using (for
315 example) rcu_dereference() without rcu_read_lock() protection.
316 Using rcu_dereference_protected() also has the advantage
317 of permitting compiler optimizations that rcu_dereference()
318 must prohibit. The rcu_dereference_protected() variant takes
319 a lockdep expression to indicate which locks must be acquired
320 by the caller. If the indicated protection is not provided,
321 a lockdep splat is emitted. See Documentation/RCU/Design/Requirements/Requirements.rst
322 and the API's code comments for more details and example usage.
324 .. [2] If the list_for_each_entry_rcu() instance might be used by
325 update-side code as well as by RCU readers, then an additional
326 lockdep expression can be added to its list of arguments.
327 For example, given an additional "lock_is_held(&mylock)" argument,
328 the RCU lockdep code would complain only if this instance was
329 invoked outside of an RCU read-side critical section and without
330 the protection of mylock.
332 The following diagram shows how each API communicates among the
333 reader, updater, and reclaimer.
339 +---------------------->| reader |---------+
343 | | | rcu_read_lock()
344 | | | rcu_read_unlock()
345 | rcu_dereference() | |
347 | updater |<----------------+ |
350 +----------------------------------->| reclaimer |
353 synchronize_rcu() & call_rcu()
356 The RCU infrastructure observes the time sequence of rcu_read_lock(),
357 rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in
358 order to determine when (1) synchronize_rcu() invocations may return
359 to their callers and (2) call_rcu() callbacks may be invoked. Efficient
360 implementations of the RCU infrastructure make heavy use of batching in
361 order to amortize their overhead over many uses of the corresponding APIs.
363 There are at least three flavors of RCU usage in the Linux kernel. The diagram
364 above shows the most common one. On the updater side, the rcu_assign_pointer(),
365 synchronize_rcu() and call_rcu() primitives used are the same for all three
366 flavors. However for protection (on the reader side), the primitives used vary
367 depending on the flavor:
369 a. rcu_read_lock() / rcu_read_unlock()
372 b. rcu_read_lock_bh() / rcu_read_unlock_bh()
373 local_bh_disable() / local_bh_enable()
376 c. rcu_read_lock_sched() / rcu_read_unlock_sched()
377 preempt_disable() / preempt_enable()
378 local_irq_save() / local_irq_restore()
379 hardirq enter / hardirq exit
381 rcu_dereference_sched()
383 These three flavors are used as follows:
385 a. RCU applied to normal data structures.
387 b. RCU applied to networking data structures that may be subjected
388 to remote denial-of-service attacks.
390 c. RCU applied to scheduler and interrupt/NMI-handler tasks.
392 Again, most uses will be of (a). The (b) and (c) cases are important
393 for specialized uses, but are relatively uncommon.
397 3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
398 -----------------------------------------------
400 This section shows a simple use of the core RCU API to protect a
401 global pointer to a dynamically allocated structure. More-typical
402 uses of RCU may be found in :ref:`listRCU.rst <list_rcu_doc>`,
403 :ref:`arrayRCU.rst <array_rcu_doc>`, and :ref:`NMI-RCU.rst <NMI_rcu_doc>`.
411 DEFINE_SPINLOCK(foo_mutex);
413 struct foo __rcu *gbl_foo;
416 * Create a new struct foo that is the same as the one currently
417 * pointed to by gbl_foo, except that field "a" is replaced
418 * with "new_a". Points gbl_foo to the new structure, and
419 * frees up the old structure after a grace period.
421 * Uses rcu_assign_pointer() to ensure that concurrent readers
422 * see the initialized version of the new structure.
424 * Uses synchronize_rcu() to ensure that any readers that might
425 * have references to the old structure complete before freeing
428 void foo_update_a(int new_a)
433 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
434 spin_lock(&foo_mutex);
435 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
438 rcu_assign_pointer(gbl_foo, new_fp);
439 spin_unlock(&foo_mutex);
445 * Return the value of field "a" of the current gbl_foo
446 * structure. Use rcu_read_lock() and rcu_read_unlock()
447 * to ensure that the structure does not get deleted out
448 * from under us, and use rcu_dereference() to ensure that
449 * we see the initialized version of the structure (important
450 * for DEC Alpha and for people reading the code).
457 retval = rcu_dereference(gbl_foo)->a;
464 - Use rcu_read_lock() and rcu_read_unlock() to guard RCU
465 read-side critical sections.
467 - Within an RCU read-side critical section, use rcu_dereference()
468 to dereference RCU-protected pointers.
470 - Use some solid scheme (such as locks or semaphores) to
471 keep concurrent updates from interfering with each other.
473 - Use rcu_assign_pointer() to update an RCU-protected pointer.
474 This primitive protects concurrent readers from the updater,
475 **not** concurrent updates from each other! You therefore still
476 need to use locking (or something similar) to keep concurrent
477 rcu_assign_pointer() primitives from interfering with each other.
479 - Use synchronize_rcu() **after** removing a data element from an
480 RCU-protected data structure, but **before** reclaiming/freeing
481 the data element, in order to wait for the completion of all
482 RCU read-side critical sections that might be referencing that
485 See checklist.txt for additional rules to follow when using RCU.
486 And again, more-typical uses of RCU may be found in :ref:`listRCU.rst
487 <list_rcu_doc>`, :ref:`arrayRCU.rst <array_rcu_doc>`, and :ref:`NMI-RCU.rst
492 4. WHAT IF MY UPDATING THREAD CANNOT BLOCK?
493 --------------------------------------------
495 In the example above, foo_update_a() blocks until a grace period elapses.
496 This is quite simple, but in some cases one cannot afford to wait so
497 long -- there might be other high-priority work to be done.
499 In such cases, one uses call_rcu() rather than synchronize_rcu().
500 The call_rcu() API is as follows::
502 void call_rcu(struct rcu_head *head, rcu_callback_t func);
504 This function invokes func(head) after a grace period has elapsed.
505 This invocation might happen from either softirq or process context,
506 so the function is not permitted to block. The foo struct needs to
507 have an rcu_head structure added, perhaps as follows::
516 The foo_update_a() function might then be written as follows::
519 * Create a new struct foo that is the same as the one currently
520 * pointed to by gbl_foo, except that field "a" is replaced
521 * with "new_a". Points gbl_foo to the new structure, and
522 * frees up the old structure after a grace period.
524 * Uses rcu_assign_pointer() to ensure that concurrent readers
525 * see the initialized version of the new structure.
527 * Uses call_rcu() to ensure that any readers that might have
528 * references to the old structure complete before freeing the
531 void foo_update_a(int new_a)
536 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
537 spin_lock(&foo_mutex);
538 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
541 rcu_assign_pointer(gbl_foo, new_fp);
542 spin_unlock(&foo_mutex);
543 call_rcu(&old_fp->rcu, foo_reclaim);
546 The foo_reclaim() function might appear as follows::
548 void foo_reclaim(struct rcu_head *rp)
550 struct foo *fp = container_of(rp, struct foo, rcu);
557 The container_of() primitive is a macro that, given a pointer into a
558 struct, the type of the struct, and the pointed-to field within the
559 struct, returns a pointer to the beginning of the struct.
561 The use of call_rcu() permits the caller of foo_update_a() to
562 immediately regain control, without needing to worry further about the
563 old version of the newly updated element. It also clearly shows the
564 RCU distinction between updater, namely foo_update_a(), and reclaimer,
565 namely foo_reclaim().
567 The summary of advice is the same as for the previous section, except
568 that we are now using call_rcu() rather than synchronize_rcu():
570 - Use call_rcu() **after** removing a data element from an
571 RCU-protected data structure in order to register a callback
572 function that will be invoked after the completion of all RCU
573 read-side critical sections that might be referencing that
576 If the callback for call_rcu() is not doing anything more than calling
577 kfree() on the structure, you can use kfree_rcu() instead of call_rcu()
578 to avoid having to write your own callback::
580 kfree_rcu(old_fp, rcu);
582 Again, see checklist.txt for additional rules governing the use of RCU.
586 5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
587 ------------------------------------------------
589 One of the nice things about RCU is that it has extremely simple "toy"
590 implementations that are a good first step towards understanding the
591 production-quality implementations in the Linux kernel. This section
592 presents two such "toy" implementations of RCU, one that is implemented
593 in terms of familiar locking primitives, and another that more closely
594 resembles "classic" RCU. Both are way too simple for real-world use,
595 lacking both functionality and performance. However, they are useful
596 in getting a feel for how RCU works. See kernel/rcu/update.c for a
597 production-quality implementation, and see:
599 http://www.rdrop.com/users/paulmck/RCU
601 for papers describing the Linux kernel RCU implementation. The OLS'01
602 and OLS'02 papers are a good introduction, and the dissertation provides
603 more details on the current implementation as of early 2004.
606 5A. "TOY" IMPLEMENTATION #1: LOCKING
607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
608 This section presents a "toy" RCU implementation that is based on
609 familiar locking primitives. Its overhead makes it a non-starter for
610 real-life use, as does its lack of scalability. It is also unsuitable
611 for realtime use, since it allows scheduling latency to "bleed" from
612 one read-side critical section to another. It also assumes recursive
613 reader-writer locks: If you try this with non-recursive locks, and
614 you allow nested rcu_read_lock() calls, you can deadlock.
616 However, it is probably the easiest implementation to relate to, so is
617 a good starting point.
619 It is extremely simple::
621 static DEFINE_RWLOCK(rcu_gp_mutex);
623 void rcu_read_lock(void)
625 read_lock(&rcu_gp_mutex);
628 void rcu_read_unlock(void)
630 read_unlock(&rcu_gp_mutex);
633 void synchronize_rcu(void)
635 write_lock(&rcu_gp_mutex);
636 smp_mb__after_spinlock();
637 write_unlock(&rcu_gp_mutex);
640 [You can ignore rcu_assign_pointer() and rcu_dereference() without missing
641 much. But here are simplified versions anyway. And whatever you do,
642 don't forget about them when submitting patches making use of RCU!]::
644 #define rcu_assign_pointer(p, v) \
646 smp_store_release(&(p), (v)); \
649 #define rcu_dereference(p) \
651 typeof(p) _________p1 = READ_ONCE(p); \
656 The rcu_read_lock() and rcu_read_unlock() primitive read-acquire
657 and release a global reader-writer lock. The synchronize_rcu()
658 primitive write-acquires this same lock, then releases it. This means
659 that once synchronize_rcu() exits, all RCU read-side critical sections
660 that were in progress before synchronize_rcu() was called are guaranteed
661 to have completed -- there is no way that synchronize_rcu() would have
662 been able to write-acquire the lock otherwise. The smp_mb__after_spinlock()
663 promotes synchronize_rcu() to a full memory barrier in compliance with
664 the "Memory-Barrier Guarantees" listed in:
666 Documentation/RCU/Design/Requirements/Requirements.rst
668 It is possible to nest rcu_read_lock(), since reader-writer locks may
669 be recursively acquired. Note also that rcu_read_lock() is immune
670 from deadlock (an important property of RCU). The reason for this is
671 that the only thing that can block rcu_read_lock() is a synchronize_rcu().
672 But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex,
673 so there can be no deadlock cycle.
678 Why is this argument naive? How could a deadlock
679 occur when using this algorithm in a real-world Linux
680 kernel? How could this deadlock be avoided?
682 :ref:`Answers to Quick Quiz <9_whatisRCU>`
684 5B. "TOY" EXAMPLE #2: CLASSIC RCU
685 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
686 This section presents a "toy" RCU implementation that is based on
687 "classic RCU". It is also short on performance (but only for updates) and
688 on features such as hotplug CPU and the ability to run in CONFIG_PREEMPTION
689 kernels. The definitions of rcu_dereference() and rcu_assign_pointer()
690 are the same as those shown in the preceding section, so they are omitted.
693 void rcu_read_lock(void) { }
695 void rcu_read_unlock(void) { }
697 void synchronize_rcu(void)
701 for_each_possible_cpu(cpu)
705 Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing.
706 This is the great strength of classic RCU in a non-preemptive kernel:
707 read-side overhead is precisely zero, at least on non-Alpha CPUs.
708 And there is absolutely no way that rcu_read_lock() can possibly
709 participate in a deadlock cycle!
711 The implementation of synchronize_rcu() simply schedules itself on each
712 CPU in turn. The run_on() primitive can be implemented straightforwardly
713 in terms of the sched_setaffinity() primitive. Of course, a somewhat less
714 "toy" implementation would restore the affinity upon completion rather
715 than just leaving all tasks running on the last CPU, but when I said
716 "toy", I meant **toy**!
718 So how the heck is this supposed to work???
720 Remember that it is illegal to block while in an RCU read-side critical
721 section. Therefore, if a given CPU executes a context switch, we know
722 that it must have completed all preceding RCU read-side critical sections.
723 Once **all** CPUs have executed a context switch, then **all** preceding
724 RCU read-side critical sections will have completed.
726 So, suppose that we remove a data item from its structure and then invoke
727 synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed
728 that there are no RCU read-side critical sections holding a reference
729 to that data item, so we can safely reclaim it.
734 Give an example where Classic RCU's read-side
735 overhead is **negative**.
737 :ref:`Answers to Quick Quiz <9_whatisRCU>`
742 If it is illegal to block in an RCU read-side
743 critical section, what the heck do you do in
744 CONFIG_PREEMPT_RT, where normal spinlocks can block???
746 :ref:`Answers to Quick Quiz <9_whatisRCU>`
750 6. ANALOGY WITH READER-WRITER LOCKING
751 --------------------------------------
753 Although RCU can be used in many different ways, a very common use of
754 RCU is analogous to reader-writer locking. The following unified
755 diff shows how closely related RCU and reader-writer locking can be.
758 @@ -5,5 +5,5 @@ struct el {
760 /* Other data fields */
763 +spinlock_t listmutex;
767 struct list_head *lp;
770 - read_lock(&listmutex);
771 - list_for_each_entry(p, head, lp) {
773 + list_for_each_entry_rcu(p, head, lp) {
776 - read_unlock(&listmutex);
781 - read_unlock(&listmutex);
790 - write_lock(&listmutex);
791 + spin_lock(&listmutex);
792 list_for_each_entry(p, head, lp) {
794 - list_del(&p->list);
795 - write_unlock(&listmutex);
796 + list_del_rcu(&p->list);
797 + spin_unlock(&listmutex);
803 - write_unlock(&listmutex);
804 + spin_unlock(&listmutex);
808 Or, for those who prefer a side-by-side listing::
810 1 struct el { 1 struct el {
811 2 struct list_head list; 2 struct list_head list;
812 3 long key; 3 long key;
813 4 spinlock_t mutex; 4 spinlock_t mutex;
814 5 int data; 5 int data;
815 6 /* Other data fields */ 6 /* Other data fields */
817 8 rwlock_t listmutex; 8 spinlock_t listmutex;
818 9 struct el head; 9 struct el head;
822 1 int search(long key, int *result) 1 int search(long key, int *result)
824 3 struct list_head *lp; 3 struct list_head *lp;
825 4 struct el *p; 4 struct el *p;
827 6 read_lock(&listmutex); 6 rcu_read_lock();
828 7 list_for_each_entry(p, head, lp) { 7 list_for_each_entry_rcu(p, head, lp) {
829 8 if (p->key == key) { 8 if (p->key == key) {
830 9 *result = p->data; 9 *result = p->data;
831 10 read_unlock(&listmutex); 10 rcu_read_unlock();
832 11 return 1; 11 return 1;
835 14 read_unlock(&listmutex); 14 rcu_read_unlock();
836 15 return 0; 15 return 0;
841 1 int delete(long key) 1 int delete(long key)
843 3 struct el *p; 3 struct el *p;
845 5 write_lock(&listmutex); 5 spin_lock(&listmutex);
846 6 list_for_each_entry(p, head, lp) { 6 list_for_each_entry(p, head, lp) {
847 7 if (p->key == key) { 7 if (p->key == key) {
848 8 list_del(&p->list); 8 list_del_rcu(&p->list);
849 9 write_unlock(&listmutex); 9 spin_unlock(&listmutex);
850 10 synchronize_rcu();
851 10 kfree(p); 11 kfree(p);
852 11 return 1; 12 return 1;
855 14 write_unlock(&listmutex); 15 spin_unlock(&listmutex);
856 15 return 0; 16 return 0;
859 Either way, the differences are quite small. Read-side locking moves
860 to rcu_read_lock() and rcu_read_unlock, update-side locking moves from
861 a reader-writer lock to a simple spinlock, and a synchronize_rcu()
862 precedes the kfree().
864 However, there is one potential catch: the read-side and update-side
865 critical sections can now run concurrently. In many cases, this will
866 not be a problem, but it is necessary to check carefully regardless.
867 For example, if multiple independent list updates must be seen as
868 a single atomic update, converting to RCU will require special care.
870 Also, the presence of synchronize_rcu() means that the RCU version of
871 delete() can now block. If this is a problem, there is a callback-based
872 mechanism that never blocks, namely call_rcu() or kfree_rcu(), that can
873 be used in place of synchronize_rcu().
877 7. ANALOGY WITH REFERENCE COUNTING
878 -----------------------------------
880 The reader-writer analogy (illustrated by the previous section) is not
881 always the best way to think about using RCU. Another helpful analogy
882 considers RCU an effective reference count on everything which is
885 A reference count typically does not prevent the referenced object's
886 values from changing, but does prevent changes to type -- particularly the
887 gross change of type that happens when that object's memory is freed and
888 re-allocated for some other purpose. Once a type-safe reference to the
889 object is obtained, some other mechanism is needed to ensure consistent
890 access to the data in the object. This could involve taking a spinlock,
891 but with RCU the typical approach is to perform reads with SMP-aware
892 operations such as smp_load_acquire(), to perform updates with atomic
893 read-modify-write operations, and to provide the necessary ordering.
894 RCU provides a number of support functions that embed the required
895 operations and ordering, such as the list_for_each_entry_rcu() macro
896 used in the previous section.
898 A more focused view of the reference counting behavior is that,
899 between rcu_read_lock() and rcu_read_unlock(), any reference taken with
900 rcu_dereference() on a pointer marked as ``__rcu`` can be treated as
901 though a reference-count on that object has been temporarily increased.
902 This prevents the object from changing type. Exactly what this means
903 will depend on normal expectations of objects of that type, but it
904 typically includes that spinlocks can still be safely locked, normal
905 reference counters can be safely manipulated, and ``__rcu`` pointers
906 can be safely dereferenced.
908 Some operations that one might expect to see on an object for
909 which an RCU reference is held include:
911 - Copying out data that is guaranteed to be stable by the object's type.
912 - Using kref_get_unless_zero() or similar to get a longer-term
913 reference. This may fail of course.
914 - Acquiring a spinlock in the object, and checking if the object still
915 is the expected object and if so, manipulating it freely.
917 The understanding that RCU provides a reference that only prevents a
918 change of type is particularly visible with objects allocated from a
919 slab cache marked ``SLAB_TYPESAFE_BY_RCU``. RCU operations may yield a
920 reference to an object from such a cache that has been concurrently
921 freed and the memory reallocated to a completely different object,
922 though of the same type. In this case RCU doesn't even protect the
923 identity of the object from changing, only its type. So the object
924 found may not be the one expected, but it will be one where it is safe
925 to take a reference or spinlock and then confirm that the identity
926 matches the expectations.
928 With traditional reference counting -- such as that implemented by the
929 kref library in Linux -- there is typically code that runs when the last
930 reference to an object is dropped. With kref, this is the function
931 passed to kref_put(). When RCU is being used, such finalization code
932 must not be run until all ``__rcu`` pointers referencing the object have
933 been updated, and then a grace period has passed. Every remaining
934 globally visible pointer to the object must be considered to be a
935 potential counted reference, and the finalization code is typically run
936 using call_rcu() only after all those pointers have been changed.
938 To see how to choose between these two analogies -- of RCU as a
939 reader-writer lock and RCU as a reference counting system -- it is useful
940 to reflect on the scale of the thing being protected. The reader-writer
941 lock analogy looks at larger multi-part objects such as a linked list
942 and shows how RCU can facilitate concurrency while elements are added
943 to, and removed from, the list. The reference-count analogy looks at
944 the individual objects and looks at how they can be accessed safely
945 within whatever whole they are a part of.
949 8. FULL LIST OF RCU APIs
950 -------------------------
952 The RCU APIs are documented in docbook-format header comments in the
953 Linux-kernel source code, but it helps to have a full list of the
954 APIs, since there does not appear to be a way to categorize them
955 in docbook. Here is the list, by category.
963 list_for_each_entry_rcu
964 list_for_each_entry_continue_rcu
965 list_for_each_entry_from_rcu
966 list_first_or_null_rcu
967 list_next_or_null_rcu
971 hlist_for_each_entry_rcu
972 hlist_for_each_entry_rcu_bh
973 hlist_for_each_entry_from_rcu
974 hlist_for_each_entry_continue_rcu
975 hlist_for_each_entry_continue_rcu_bh
976 hlist_nulls_first_rcu
977 hlist_nulls_for_each_entry_rcu
979 hlist_bl_for_each_entry_rcu
981 RCU pointer/list update::
996 list_splice_tail_init_rcu
997 hlist_nulls_del_init_rcu
999 hlist_nulls_add_head_rcu
1000 hlist_bl_add_head_rcu
1001 hlist_bl_del_init_rcu
1003 hlist_bl_set_first_rcu
1007 Critical sections Grace period Barrier
1009 rcu_read_lock synchronize_net rcu_barrier
1010 rcu_read_unlock synchronize_rcu
1011 rcu_dereference synchronize_rcu_expedited
1012 rcu_read_lock_held call_rcu
1013 rcu_dereference_check kfree_rcu
1014 rcu_dereference_protected
1018 Critical sections Grace period Barrier
1020 rcu_read_lock_bh call_rcu rcu_barrier
1021 rcu_read_unlock_bh synchronize_rcu
1022 [local_bh_disable] synchronize_rcu_expedited
1025 rcu_dereference_bh_check
1026 rcu_dereference_bh_protected
1027 rcu_read_lock_bh_held
1031 Critical sections Grace period Barrier
1033 rcu_read_lock_sched call_rcu rcu_barrier
1034 rcu_read_unlock_sched synchronize_rcu
1035 [preempt_disable] synchronize_rcu_expedited
1037 rcu_read_lock_sched_notrace
1038 rcu_read_unlock_sched_notrace
1039 rcu_dereference_sched
1040 rcu_dereference_sched_check
1041 rcu_dereference_sched_protected
1042 rcu_read_lock_sched_held
1047 Critical sections Grace period Barrier
1049 srcu_read_lock call_srcu srcu_barrier
1050 srcu_read_unlock synchronize_srcu
1051 srcu_dereference synchronize_srcu_expedited
1052 srcu_dereference_check
1055 SRCU: Initialization/cleanup::
1062 All: lockdep-checked RCU-protected pointer access::
1070 See the comment headers in the source code (or the docbook generated
1071 from them) for more information.
1073 However, given that there are no fewer than four families of RCU APIs
1074 in the Linux kernel, how do you choose which one to use? The following
1075 list can be helpful:
1077 a. Will readers need to block? If so, you need SRCU.
1079 b. What about the -rt patchset? If readers would need to block
1080 in an non-rt kernel, you need SRCU. If readers would block
1081 in a -rt kernel, but not in a non-rt kernel, SRCU is not
1082 necessary. (The -rt patchset turns spinlocks into sleeplocks,
1083 hence this distinction.)
1085 c. Do you need to treat NMI handlers, hardirq handlers,
1086 and code segments with preemption disabled (whether
1087 via preempt_disable(), local_irq_save(), local_bh_disable(),
1088 or some other mechanism) as if they were explicit RCU readers?
1089 If so, RCU-sched is the only choice that will work for you.
1091 d. Do you need RCU grace periods to complete even in the face
1092 of softirq monopolization of one or more of the CPUs? For
1093 example, is your code subject to network-based denial-of-service
1094 attacks? If so, you should disable softirq across your readers,
1095 for example, by using rcu_read_lock_bh().
1097 e. Is your workload too update-intensive for normal use of
1098 RCU, but inappropriate for other synchronization mechanisms?
1099 If so, consider SLAB_TYPESAFE_BY_RCU (which was originally
1100 named SLAB_DESTROY_BY_RCU). But please be careful!
1102 f. Do you need read-side critical sections that are respected
1103 even though they are in the middle of the idle loop, during
1104 user-mode execution, or on an offlined CPU? If so, SRCU is the
1105 only choice that will work for you.
1107 g. Otherwise, use RCU.
1109 Of course, this all assumes that you have determined that RCU is in fact
1110 the right tool for your job.
1114 9. ANSWERS TO QUICK QUIZZES
1115 ----------------------------
1118 Why is this argument naive? How could a deadlock
1119 occur when using this algorithm in a real-world Linux
1120 kernel? [Referring to the lock-based "toy" RCU
1124 Consider the following sequence of events:
1126 1. CPU 0 acquires some unrelated lock, call it
1127 "problematic_lock", disabling irq via
1128 spin_lock_irqsave().
1130 2. CPU 1 enters synchronize_rcu(), write-acquiring
1133 3. CPU 0 enters rcu_read_lock(), but must wait
1134 because CPU 1 holds rcu_gp_mutex.
1136 4. CPU 1 is interrupted, and the irq handler
1137 attempts to acquire problematic_lock.
1139 The system is now deadlocked.
1141 One way to avoid this deadlock is to use an approach like
1142 that of CONFIG_PREEMPT_RT, where all normal spinlocks
1143 become blocking locks, and all irq handlers execute in
1144 the context of special tasks. In this case, in step 4
1145 above, the irq handler would block, allowing CPU 1 to
1146 release rcu_gp_mutex, avoiding the deadlock.
1148 Even in the absence of deadlock, this RCU implementation
1149 allows latency to "bleed" from readers to other
1150 readers through synchronize_rcu(). To see this,
1151 consider task A in an RCU read-side critical section
1152 (thus read-holding rcu_gp_mutex), task B blocked
1153 attempting to write-acquire rcu_gp_mutex, and
1154 task C blocked in rcu_read_lock() attempting to
1155 read_acquire rcu_gp_mutex. Task A's RCU read-side
1156 latency is holding up task C, albeit indirectly via
1159 Realtime RCU implementations therefore use a counter-based
1160 approach where tasks in RCU read-side critical sections
1161 cannot be blocked by tasks executing synchronize_rcu().
1163 :ref:`Back to Quick Quiz #1 <quiz_1>`
1166 Give an example where Classic RCU's read-side
1167 overhead is **negative**.
1170 Imagine a single-CPU system with a non-CONFIG_PREEMPTION
1171 kernel where a routing table is used by process-context
1172 code, but can be updated by irq-context code (for example,
1173 by an "ICMP REDIRECT" packet). The usual way of handling
1174 this would be to have the process-context code disable
1175 interrupts while searching the routing table. Use of
1176 RCU allows such interrupt-disabling to be dispensed with.
1177 Thus, without RCU, you pay the cost of disabling interrupts,
1178 and with RCU you don't.
1180 One can argue that the overhead of RCU in this
1181 case is negative with respect to the single-CPU
1182 interrupt-disabling approach. Others might argue that
1183 the overhead of RCU is merely zero, and that replacing
1184 the positive overhead of the interrupt-disabling scheme
1185 with the zero-overhead RCU scheme does not constitute
1188 In real life, of course, things are more complex. But
1189 even the theoretical possibility of negative overhead for
1190 a synchronization primitive is a bit unexpected. ;-)
1192 :ref:`Back to Quick Quiz #2 <quiz_2>`
1195 If it is illegal to block in an RCU read-side
1196 critical section, what the heck do you do in
1197 CONFIG_PREEMPT_RT, where normal spinlocks can block???
1200 Just as CONFIG_PREEMPT_RT permits preemption of spinlock
1201 critical sections, it permits preemption of RCU
1202 read-side critical sections. It also permits
1203 spinlocks blocking while in RCU read-side critical
1206 Why the apparent inconsistency? Because it is
1207 possible to use priority boosting to keep the RCU
1208 grace periods short if need be (for example, if running
1209 short of memory). In contrast, if blocking waiting
1210 for (say) network reception, there is no way to know
1211 what should be boosted. Especially given that the
1212 process we need to boost might well be a human being
1213 who just went out for a pizza or something. And although
1214 a computer-operated cattle prod might arouse serious
1215 interest, it might also provoke serious objections.
1216 Besides, how does the computer know what pizza parlor
1217 the human being went to???
1219 :ref:`Back to Quick Quiz #3 <quiz_3>`
1223 My thanks to the people who helped make this human-readable, including
1224 Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern.
1227 For more information, see http://www.rdrop.com/users/paulmck/RCU.