1 What is RCU? -- "Read, Copy, Update"
3 Please note that the "What is RCU?" LWN series is an excellent place
4 to start learning about RCU:
6 1. What is RCU, Fundamentally? http://lwn.net/Articles/262464/
7 2. What is RCU? Part 2: Usage http://lwn.net/Articles/263130/
8 3. RCU part 3: the RCU API http://lwn.net/Articles/264090/
9 4. The RCU API, 2010 Edition http://lwn.net/Articles/418853/
10 2010 Big API Table http://lwn.net/Articles/419086/
11 5. The RCU API, 2014 Edition http://lwn.net/Articles/609904/
12 2014 Big API Table http://lwn.net/Articles/609973/
17 RCU is a synchronization mechanism that was added to the Linux kernel
18 during the 2.5 development effort that is optimized for read-mostly
19 situations. Although RCU is actually quite simple once you understand it,
20 getting there can sometimes be a challenge. Part of the problem is that
21 most of the past descriptions of RCU have been written with the mistaken
22 assumption that there is "one true way" to describe RCU. Instead,
23 the experience has been that different people must take different paths
24 to arrive at an understanding of RCU. This document provides several
25 different paths, as follows:
28 2. WHAT IS RCU'S CORE API?
29 3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
30 4. WHAT IF MY UPDATING THREAD CANNOT BLOCK?
31 5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
32 6. ANALOGY WITH READER-WRITER LOCKING
33 7. FULL LIST OF RCU APIs
34 8. ANSWERS TO QUICK QUIZZES
36 People who prefer starting with a conceptual overview should focus on
37 Section 1, though most readers will profit by reading this section at
38 some point. People who prefer to start with an API that they can then
39 experiment with should focus on Section 2. People who prefer to start
40 with example uses should focus on Sections 3 and 4. People who need to
41 understand the RCU implementation should focus on Section 5, then dive
42 into the kernel source code. People who reason best by analogy should
43 focus on Section 6. Section 7 serves as an index to the docbook API
44 documentation, and Section 8 is the traditional answer key.
46 So, start with the section that makes the most sense to you and your
47 preferred method of learning. If you need to know everything about
48 everything, feel free to read the whole thing -- but if you are really
49 that type of person, you have perused the source code and will therefore
50 never need this document anyway. ;-)
55 The basic idea behind RCU is to split updates into "removal" and
56 "reclamation" phases. The removal phase removes references to data items
57 within a data structure (possibly by replacing them with references to
58 new versions of these data items), and can run concurrently with readers.
59 The reason that it is safe to run the removal phase concurrently with
60 readers is the semantics of modern CPUs guarantee that readers will see
61 either the old or the new version of the data structure rather than a
62 partially updated reference. The reclamation phase does the work of reclaiming
63 (e.g., freeing) the data items removed from the data structure during the
64 removal phase. Because reclaiming data items can disrupt any readers
65 concurrently referencing those data items, the reclamation phase must
66 not start until readers no longer hold references to those data items.
68 Splitting the update into removal and reclamation phases permits the
69 updater to perform the removal phase immediately, and to defer the
70 reclamation phase until all readers active during the removal phase have
71 completed, either by blocking until they finish or by registering a
72 callback that is invoked after they finish. Only readers that are active
73 during the removal phase need be considered, because any reader starting
74 after the removal phase will be unable to gain a reference to the removed
75 data items, and therefore cannot be disrupted by the reclamation phase.
77 So the typical RCU update sequence goes something like the following:
79 a. Remove pointers to a data structure, so that subsequent
80 readers cannot gain a reference to it.
82 b. Wait for all previous readers to complete their RCU read-side
85 c. At this point, there cannot be any readers who hold references
86 to the data structure, so it now may safely be reclaimed
89 Step (b) above is the key idea underlying RCU's deferred destruction.
90 The ability to wait until all readers are done allows RCU readers to
91 use much lighter-weight synchronization, in some cases, absolutely no
92 synchronization at all. In contrast, in more conventional lock-based
93 schemes, readers must use heavy-weight synchronization in order to
94 prevent an updater from deleting the data structure out from under them.
95 This is because lock-based updaters typically update data items in place,
96 and must therefore exclude readers. In contrast, RCU-based updaters
97 typically take advantage of the fact that writes to single aligned
98 pointers are atomic on modern CPUs, allowing atomic insertion, removal,
99 and replacement of data items in a linked structure without disrupting
100 readers. Concurrent RCU readers can then continue accessing the old
101 versions, and can dispense with the atomic operations, memory barriers,
102 and communications cache misses that are so expensive on present-day
103 SMP computer systems, even in absence of lock contention.
105 In the three-step procedure shown above, the updater is performing both
106 the removal and the reclamation step, but it is often helpful for an
107 entirely different thread to do the reclamation, as is in fact the case
108 in the Linux kernel's directory-entry cache (dcache). Even if the same
109 thread performs both the update step (step (a) above) and the reclamation
110 step (step (c) above), it is often helpful to think of them separately.
111 For example, RCU readers and updaters need not communicate at all,
112 but RCU provides implicit low-overhead communication between readers
113 and reclaimers, namely, in step (b) above.
115 So how the heck can a reclaimer tell when a reader is done, given
116 that readers are not doing any sort of synchronization operations???
117 Read on to learn about how RCU's API makes this easy.
120 2. WHAT IS RCU'S CORE API?
122 The core RCU API is quite small:
126 c. synchronize_rcu() / call_rcu()
127 d. rcu_assign_pointer()
130 There are many other members of the RCU API, but the rest can be
131 expressed in terms of these five, though most implementations instead
132 express synchronize_rcu() in terms of the call_rcu() callback API.
134 The five core RCU APIs are described below, the other 18 will be enumerated
135 later. See the kernel docbook documentation for more info, or look directly
136 at the function header comments.
140 void rcu_read_lock(void);
142 Used by a reader to inform the reclaimer that the reader is
143 entering an RCU read-side critical section. It is illegal
144 to block while in an RCU read-side critical section, though
145 kernels built with CONFIG_PREEMPT_RCU can preempt RCU
146 read-side critical sections. Any RCU-protected data structure
147 accessed during an RCU read-side critical section is guaranteed to
148 remain unreclaimed for the full duration of that critical section.
149 Reference counts may be used in conjunction with RCU to maintain
150 longer-term references to data structures.
154 void rcu_read_unlock(void);
156 Used by a reader to inform the reclaimer that the reader is
157 exiting an RCU read-side critical section. Note that RCU
158 read-side critical sections may be nested and/or overlapping.
162 void synchronize_rcu(void);
164 Marks the end of updater code and the beginning of reclaimer
165 code. It does this by blocking until all pre-existing RCU
166 read-side critical sections on all CPUs have completed.
167 Note that synchronize_rcu() will -not- necessarily wait for
168 any subsequent RCU read-side critical sections to complete.
169 For example, consider the following sequence of events:
172 ----------------- ------------------------- ---------------
174 2. enters synchronize_rcu()
177 5. exits synchronize_rcu()
180 To reiterate, synchronize_rcu() waits only for ongoing RCU
181 read-side critical sections to complete, not necessarily for
182 any that begin after synchronize_rcu() is invoked.
184 Of course, synchronize_rcu() does not necessarily return
185 -immediately- after the last pre-existing RCU read-side critical
186 section completes. For one thing, there might well be scheduling
187 delays. For another thing, many RCU implementations process
188 requests in batches in order to improve efficiencies, which can
189 further delay synchronize_rcu().
191 Since synchronize_rcu() is the API that must figure out when
192 readers are done, its implementation is key to RCU. For RCU
193 to be useful in all but the most read-intensive situations,
194 synchronize_rcu()'s overhead must also be quite small.
196 The call_rcu() API is a callback form of synchronize_rcu(),
197 and is described in more detail in a later section. Instead of
198 blocking, it registers a function and argument which are invoked
199 after all ongoing RCU read-side critical sections have completed.
200 This callback variant is particularly useful in situations where
201 it is illegal to block or where update-side performance is
202 critically important.
204 However, the call_rcu() API should not be used lightly, as use
205 of the synchronize_rcu() API generally results in simpler code.
206 In addition, the synchronize_rcu() API has the nice property
207 of automatically limiting update rate should grace periods
208 be delayed. This property results in system resilience in face
209 of denial-of-service attacks. Code using call_rcu() should limit
210 update rate in order to gain this same sort of resilience. See
211 checklist.txt for some approaches to limiting the update rate.
215 typeof(p) rcu_assign_pointer(p, typeof(p) v);
217 Yes, rcu_assign_pointer() -is- implemented as a macro, though it
218 would be cool to be able to declare a function in this manner.
219 (Compiler experts will no doubt disagree.)
221 The updater uses this function to assign a new value to an
222 RCU-protected pointer, in order to safely communicate the change
223 in value from the updater to the reader. This function returns
224 the new value, and also executes any memory-barrier instructions
225 required for a given CPU architecture.
227 Perhaps just as important, it serves to document (1) which
228 pointers are protected by RCU and (2) the point at which a
229 given structure becomes accessible to other CPUs. That said,
230 rcu_assign_pointer() is most frequently used indirectly, via
231 the _rcu list-manipulation primitives such as list_add_rcu().
235 typeof(p) rcu_dereference(p);
237 Like rcu_assign_pointer(), rcu_dereference() must be implemented
240 The reader uses rcu_dereference() to fetch an RCU-protected
241 pointer, which returns a value that may then be safely
242 dereferenced. Note that rcu_dereference() does not actually
243 dereference the pointer, instead, it protects the pointer for
244 later dereferencing. It also executes any needed memory-barrier
245 instructions for a given CPU architecture. Currently, only Alpha
246 needs memory barriers within rcu_dereference() -- on other CPUs,
247 it compiles to nothing, not even a compiler directive.
249 Common coding practice uses rcu_dereference() to copy an
250 RCU-protected pointer to a local variable, then dereferences
251 this local variable, for example as follows:
253 p = rcu_dereference(head.next);
256 However, in this case, one could just as easily combine these
259 return rcu_dereference(head.next)->data;
261 If you are going to be fetching multiple fields from the
262 RCU-protected structure, using the local variable is of
263 course preferred. Repeated rcu_dereference() calls look
264 ugly, do not guarantee that the same pointer will be returned
265 if an update happened while in the critical section, and incur
266 unnecessary overhead on Alpha CPUs.
268 Note that the value returned by rcu_dereference() is valid
269 only within the enclosing RCU read-side critical section.
270 For example, the following is -not- legal:
273 p = rcu_dereference(head.next);
275 x = p->address; /* BUG!!! */
277 y = p->data; /* BUG!!! */
280 Holding a reference from one RCU read-side critical section
281 to another is just as illegal as holding a reference from
282 one lock-based critical section to another! Similarly,
283 using a reference outside of the critical section in which
284 it was acquired is just as illegal as doing so with normal
287 As with rcu_assign_pointer(), an important function of
288 rcu_dereference() is to document which pointers are protected by
289 RCU, in particular, flagging a pointer that is subject to changing
290 at any time, including immediately after the rcu_dereference().
291 And, again like rcu_assign_pointer(), rcu_dereference() is
292 typically used indirectly, via the _rcu list-manipulation
293 primitives, such as list_for_each_entry_rcu().
295 The following diagram shows how each API communicates among the
296 reader, updater, and reclaimer.
301 +---------------------->| reader |---------+
305 | | | rcu_read_lock()
306 | | | rcu_read_unlock()
307 | rcu_dereference() | |
309 | updater |<---------------------+ |
312 +----------------------------------->| reclaimer |
315 synchronize_rcu() & call_rcu()
318 The RCU infrastructure observes the time sequence of rcu_read_lock(),
319 rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in
320 order to determine when (1) synchronize_rcu() invocations may return
321 to their callers and (2) call_rcu() callbacks may be invoked. Efficient
322 implementations of the RCU infrastructure make heavy use of batching in
323 order to amortize their overhead over many uses of the corresponding APIs.
325 There are at least three flavors of RCU usage in the Linux kernel. The diagram
326 above shows the most common one. On the updater side, the rcu_assign_pointer(),
327 sychronize_rcu() and call_rcu() primitives used are the same for all three
328 flavors. However for protection (on the reader side), the primitives used vary
329 depending on the flavor:
331 a. rcu_read_lock() / rcu_read_unlock()
334 b. rcu_read_lock_bh() / rcu_read_unlock_bh()
335 local_bh_disable() / local_bh_enable()
338 c. rcu_read_lock_sched() / rcu_read_unlock_sched()
339 preempt_disable() / preempt_enable()
340 local_irq_save() / local_irq_restore()
341 hardirq enter / hardirq exit
343 rcu_dereference_sched()
345 These three flavors are used as follows:
347 a. RCU applied to normal data structures.
349 b. RCU applied to networking data structures that may be subjected
350 to remote denial-of-service attacks.
352 c. RCU applied to scheduler and interrupt/NMI-handler tasks.
354 Again, most uses will be of (a). The (b) and (c) cases are important
355 for specialized uses, but are relatively uncommon.
358 3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
360 This section shows a simple use of the core RCU API to protect a
361 global pointer to a dynamically allocated structure. More-typical
362 uses of RCU may be found in listRCU.txt, arrayRCU.txt, and NMI-RCU.txt.
369 DEFINE_SPINLOCK(foo_mutex);
371 struct foo __rcu *gbl_foo;
374 * Create a new struct foo that is the same as the one currently
375 * pointed to by gbl_foo, except that field "a" is replaced
376 * with "new_a". Points gbl_foo to the new structure, and
377 * frees up the old structure after a grace period.
379 * Uses rcu_assign_pointer() to ensure that concurrent readers
380 * see the initialized version of the new structure.
382 * Uses synchronize_rcu() to ensure that any readers that might
383 * have references to the old structure complete before freeing
386 void foo_update_a(int new_a)
391 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
392 spin_lock(&foo_mutex);
393 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
396 rcu_assign_pointer(gbl_foo, new_fp);
397 spin_unlock(&foo_mutex);
403 * Return the value of field "a" of the current gbl_foo
404 * structure. Use rcu_read_lock() and rcu_read_unlock()
405 * to ensure that the structure does not get deleted out
406 * from under us, and use rcu_dereference() to ensure that
407 * we see the initialized version of the structure (important
408 * for DEC Alpha and for people reading the code).
415 retval = rcu_dereference(gbl_foo)->a;
422 o Use rcu_read_lock() and rcu_read_unlock() to guard RCU
423 read-side critical sections.
425 o Within an RCU read-side critical section, use rcu_dereference()
426 to dereference RCU-protected pointers.
428 o Use some solid scheme (such as locks or semaphores) to
429 keep concurrent updates from interfering with each other.
431 o Use rcu_assign_pointer() to update an RCU-protected pointer.
432 This primitive protects concurrent readers from the updater,
433 -not- concurrent updates from each other! You therefore still
434 need to use locking (or something similar) to keep concurrent
435 rcu_assign_pointer() primitives from interfering with each other.
437 o Use synchronize_rcu() -after- removing a data element from an
438 RCU-protected data structure, but -before- reclaiming/freeing
439 the data element, in order to wait for the completion of all
440 RCU read-side critical sections that might be referencing that
443 See checklist.txt for additional rules to follow when using RCU.
444 And again, more-typical uses of RCU may be found in listRCU.txt,
445 arrayRCU.txt, and NMI-RCU.txt.
448 4. WHAT IF MY UPDATING THREAD CANNOT BLOCK?
450 In the example above, foo_update_a() blocks until a grace period elapses.
451 This is quite simple, but in some cases one cannot afford to wait so
452 long -- there might be other high-priority work to be done.
454 In such cases, one uses call_rcu() rather than synchronize_rcu().
455 The call_rcu() API is as follows:
457 void call_rcu(struct rcu_head * head,
458 void (*func)(struct rcu_head *head));
460 This function invokes func(head) after a grace period has elapsed.
461 This invocation might happen from either softirq or process context,
462 so the function is not permitted to block. The foo struct needs to
463 have an rcu_head structure added, perhaps as follows:
472 The foo_update_a() function might then be written as follows:
475 * Create a new struct foo that is the same as the one currently
476 * pointed to by gbl_foo, except that field "a" is replaced
477 * with "new_a". Points gbl_foo to the new structure, and
478 * frees up the old structure after a grace period.
480 * Uses rcu_assign_pointer() to ensure that concurrent readers
481 * see the initialized version of the new structure.
483 * Uses call_rcu() to ensure that any readers that might have
484 * references to the old structure complete before freeing the
487 void foo_update_a(int new_a)
492 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
493 spin_lock(&foo_mutex);
494 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
497 rcu_assign_pointer(gbl_foo, new_fp);
498 spin_unlock(&foo_mutex);
499 call_rcu(&old_fp->rcu, foo_reclaim);
502 The foo_reclaim() function might appear as follows:
504 void foo_reclaim(struct rcu_head *rp)
506 struct foo *fp = container_of(rp, struct foo, rcu);
513 The container_of() primitive is a macro that, given a pointer into a
514 struct, the type of the struct, and the pointed-to field within the
515 struct, returns a pointer to the beginning of the struct.
517 The use of call_rcu() permits the caller of foo_update_a() to
518 immediately regain control, without needing to worry further about the
519 old version of the newly updated element. It also clearly shows the
520 RCU distinction between updater, namely foo_update_a(), and reclaimer,
521 namely foo_reclaim().
523 The summary of advice is the same as for the previous section, except
524 that we are now using call_rcu() rather than synchronize_rcu():
526 o Use call_rcu() -after- removing a data element from an
527 RCU-protected data structure in order to register a callback
528 function that will be invoked after the completion of all RCU
529 read-side critical sections that might be referencing that
532 If the callback for call_rcu() is not doing anything more than calling
533 kfree() on the structure, you can use kfree_rcu() instead of call_rcu()
534 to avoid having to write your own callback:
536 kfree_rcu(old_fp, rcu);
538 Again, see checklist.txt for additional rules governing the use of RCU.
541 5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
543 One of the nice things about RCU is that it has extremely simple "toy"
544 implementations that are a good first step towards understanding the
545 production-quality implementations in the Linux kernel. This section
546 presents two such "toy" implementations of RCU, one that is implemented
547 in terms of familiar locking primitives, and another that more closely
548 resembles "classic" RCU. Both are way too simple for real-world use,
549 lacking both functionality and performance. However, they are useful
550 in getting a feel for how RCU works. See kernel/rcupdate.c for a
551 production-quality implementation, and see:
553 http://www.rdrop.com/users/paulmck/RCU
555 for papers describing the Linux kernel RCU implementation. The OLS'01
556 and OLS'02 papers are a good introduction, and the dissertation provides
557 more details on the current implementation as of early 2004.
560 5A. "TOY" IMPLEMENTATION #1: LOCKING
562 This section presents a "toy" RCU implementation that is based on
563 familiar locking primitives. Its overhead makes it a non-starter for
564 real-life use, as does its lack of scalability. It is also unsuitable
565 for realtime use, since it allows scheduling latency to "bleed" from
566 one read-side critical section to another. It also assumes recursive
567 reader-writer locks: If you try this with non-recursive locks, and
568 you allow nested rcu_read_lock() calls, you can deadlock.
570 However, it is probably the easiest implementation to relate to, so is
571 a good starting point.
573 It is extremely simple:
575 static DEFINE_RWLOCK(rcu_gp_mutex);
577 void rcu_read_lock(void)
579 read_lock(&rcu_gp_mutex);
582 void rcu_read_unlock(void)
584 read_unlock(&rcu_gp_mutex);
587 void synchronize_rcu(void)
589 write_lock(&rcu_gp_mutex);
590 smp_mb__after_spinlock();
591 write_unlock(&rcu_gp_mutex);
594 [You can ignore rcu_assign_pointer() and rcu_dereference() without missing
595 much. But here are simplified versions anyway. And whatever you do,
596 don't forget about them when submitting patches making use of RCU!]
598 #define rcu_assign_pointer(p, v) \
600 smp_store_release(&(p), (v)); \
603 #define rcu_dereference(p) \
605 typeof(p) _________p1 = READ_ONCE(p); \
610 The rcu_read_lock() and rcu_read_unlock() primitive read-acquire
611 and release a global reader-writer lock. The synchronize_rcu()
612 primitive write-acquires this same lock, then releases it. This means
613 that once synchronize_rcu() exits, all RCU read-side critical sections
614 that were in progress before synchronize_rcu() was called are guaranteed
615 to have completed -- there is no way that synchronize_rcu() would have
616 been able to write-acquire the lock otherwise. The smp_mb__after_spinlock()
617 promotes synchronize_rcu() to a full memory barrier in compliance with
618 the "Memory-Barrier Guarantees" listed in:
620 Documentation/RCU/Design/Requirements/Requirements.html.
622 It is possible to nest rcu_read_lock(), since reader-writer locks may
623 be recursively acquired. Note also that rcu_read_lock() is immune
624 from deadlock (an important property of RCU). The reason for this is
625 that the only thing that can block rcu_read_lock() is a synchronize_rcu().
626 But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex,
627 so there can be no deadlock cycle.
629 Quick Quiz #1: Why is this argument naive? How could a deadlock
630 occur when using this algorithm in a real-world Linux
631 kernel? How could this deadlock be avoided?
634 5B. "TOY" EXAMPLE #2: CLASSIC RCU
636 This section presents a "toy" RCU implementation that is based on
637 "classic RCU". It is also short on performance (but only for updates) and
638 on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT
639 kernels. The definitions of rcu_dereference() and rcu_assign_pointer()
640 are the same as those shown in the preceding section, so they are omitted.
642 void rcu_read_lock(void) { }
644 void rcu_read_unlock(void) { }
646 void synchronize_rcu(void)
650 for_each_possible_cpu(cpu)
654 Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing.
655 This is the great strength of classic RCU in a non-preemptive kernel:
656 read-side overhead is precisely zero, at least on non-Alpha CPUs.
657 And there is absolutely no way that rcu_read_lock() can possibly
658 participate in a deadlock cycle!
660 The implementation of synchronize_rcu() simply schedules itself on each
661 CPU in turn. The run_on() primitive can be implemented straightforwardly
662 in terms of the sched_setaffinity() primitive. Of course, a somewhat less
663 "toy" implementation would restore the affinity upon completion rather
664 than just leaving all tasks running on the last CPU, but when I said
665 "toy", I meant -toy-!
667 So how the heck is this supposed to work???
669 Remember that it is illegal to block while in an RCU read-side critical
670 section. Therefore, if a given CPU executes a context switch, we know
671 that it must have completed all preceding RCU read-side critical sections.
672 Once -all- CPUs have executed a context switch, then -all- preceding
673 RCU read-side critical sections will have completed.
675 So, suppose that we remove a data item from its structure and then invoke
676 synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed
677 that there are no RCU read-side critical sections holding a reference
678 to that data item, so we can safely reclaim it.
680 Quick Quiz #2: Give an example where Classic RCU's read-side
681 overhead is -negative-.
683 Quick Quiz #3: If it is illegal to block in an RCU read-side
684 critical section, what the heck do you do in
685 PREEMPT_RT, where normal spinlocks can block???
688 6. ANALOGY WITH READER-WRITER LOCKING
690 Although RCU can be used in many different ways, a very common use of
691 RCU is analogous to reader-writer locking. The following unified
692 diff shows how closely related RCU and reader-writer locking can be.
694 @@ -5,5 +5,5 @@ struct el {
696 /* Other data fields */
699 +spinlock_t listmutex;
703 struct list_head *lp;
706 - read_lock(&listmutex);
707 - list_for_each_entry(p, head, lp) {
709 + list_for_each_entry_rcu(p, head, lp) {
712 - read_unlock(&listmutex);
717 - read_unlock(&listmutex);
726 - write_lock(&listmutex);
727 + spin_lock(&listmutex);
728 list_for_each_entry(p, head, lp) {
730 - list_del(&p->list);
731 - write_unlock(&listmutex);
732 + list_del_rcu(&p->list);
733 + spin_unlock(&listmutex);
739 - write_unlock(&listmutex);
740 + spin_unlock(&listmutex);
744 Or, for those who prefer a side-by-side listing:
746 1 struct el { 1 struct el {
747 2 struct list_head list; 2 struct list_head list;
748 3 long key; 3 long key;
749 4 spinlock_t mutex; 4 spinlock_t mutex;
750 5 int data; 5 int data;
751 6 /* Other data fields */ 6 /* Other data fields */
753 8 rwlock_t listmutex; 8 spinlock_t listmutex;
754 9 struct el head; 9 struct el head;
756 1 int search(long key, int *result) 1 int search(long key, int *result)
758 3 struct list_head *lp; 3 struct list_head *lp;
759 4 struct el *p; 4 struct el *p;
761 6 read_lock(&listmutex); 6 rcu_read_lock();
762 7 list_for_each_entry(p, head, lp) { 7 list_for_each_entry_rcu(p, head, lp) {
763 8 if (p->key == key) { 8 if (p->key == key) {
764 9 *result = p->data; 9 *result = p->data;
765 10 read_unlock(&listmutex); 10 rcu_read_unlock();
766 11 return 1; 11 return 1;
769 14 read_unlock(&listmutex); 14 rcu_read_unlock();
770 15 return 0; 15 return 0;
773 1 int delete(long key) 1 int delete(long key)
775 3 struct el *p; 3 struct el *p;
777 5 write_lock(&listmutex); 5 spin_lock(&listmutex);
778 6 list_for_each_entry(p, head, lp) { 6 list_for_each_entry(p, head, lp) {
779 7 if (p->key == key) { 7 if (p->key == key) {
780 8 list_del(&p->list); 8 list_del_rcu(&p->list);
781 9 write_unlock(&listmutex); 9 spin_unlock(&listmutex);
782 10 synchronize_rcu();
783 10 kfree(p); 11 kfree(p);
784 11 return 1; 12 return 1;
787 14 write_unlock(&listmutex); 15 spin_unlock(&listmutex);
788 15 return 0; 16 return 0;
791 Either way, the differences are quite small. Read-side locking moves
792 to rcu_read_lock() and rcu_read_unlock, update-side locking moves from
793 a reader-writer lock to a simple spinlock, and a synchronize_rcu()
794 precedes the kfree().
796 However, there is one potential catch: the read-side and update-side
797 critical sections can now run concurrently. In many cases, this will
798 not be a problem, but it is necessary to check carefully regardless.
799 For example, if multiple independent list updates must be seen as
800 a single atomic update, converting to RCU will require special care.
802 Also, the presence of synchronize_rcu() means that the RCU version of
803 delete() can now block. If this is a problem, there is a callback-based
804 mechanism that never blocks, namely call_rcu() or kfree_rcu(), that can
805 be used in place of synchronize_rcu().
808 7. FULL LIST OF RCU APIs
810 The RCU APIs are documented in docbook-format header comments in the
811 Linux-kernel source code, but it helps to have a full list of the
812 APIs, since there does not appear to be a way to categorize them
813 in docbook. Here is the list, by category.
820 list_for_each_entry_rcu
821 list_for_each_entry_continue_rcu
822 list_for_each_entry_from_rcu
826 hlist_for_each_entry_rcu
827 hlist_for_each_entry_rcu_bh
828 hlist_for_each_entry_from_rcu
829 hlist_for_each_entry_continue_rcu
830 hlist_for_each_entry_continue_rcu_bh
831 hlist_nulls_first_rcu
832 hlist_nulls_for_each_entry_rcu
834 hlist_bl_for_each_entry_rcu
836 RCU pointer/list update:
849 list_splice_init_rcu()
850 hlist_nulls_del_init_rcu
852 hlist_nulls_add_head_rcu
853 hlist_bl_add_head_rcu
854 hlist_bl_del_init_rcu
856 hlist_bl_set_first_rcu
858 RCU: Critical sections Grace period Barrier
860 rcu_read_lock synchronize_net rcu_barrier
861 rcu_read_unlock synchronize_rcu
862 rcu_dereference synchronize_rcu_expedited
863 rcu_read_lock_held call_rcu
864 rcu_dereference_check kfree_rcu
865 rcu_dereference_protected
867 bh: Critical sections Grace period Barrier
869 rcu_read_lock_bh call_rcu rcu_barrier
870 rcu_read_unlock_bh synchronize_rcu
871 [local_bh_disable] synchronize_rcu_expedited
874 rcu_dereference_bh_check
875 rcu_dereference_bh_protected
876 rcu_read_lock_bh_held
878 sched: Critical sections Grace period Barrier
880 rcu_read_lock_sched call_rcu rcu_barrier
881 rcu_read_unlock_sched synchronize_rcu
882 [preempt_disable] synchronize_rcu_expedited
884 rcu_read_lock_sched_notrace
885 rcu_read_unlock_sched_notrace
886 rcu_dereference_sched
887 rcu_dereference_sched_check
888 rcu_dereference_sched_protected
889 rcu_read_lock_sched_held
892 SRCU: Critical sections Grace period Barrier
894 srcu_read_lock call_srcu srcu_barrier
895 srcu_read_unlock synchronize_srcu
896 srcu_dereference synchronize_srcu_expedited
897 srcu_dereference_check
900 SRCU: Initialization/cleanup
906 All: lockdep-checked RCU-protected pointer access
914 See the comment headers in the source code (or the docbook generated
915 from them) for more information.
917 However, given that there are no fewer than four families of RCU APIs
918 in the Linux kernel, how do you choose which one to use? The following
921 a. Will readers need to block? If so, you need SRCU.
923 b. What about the -rt patchset? If readers would need to block
924 in an non-rt kernel, you need SRCU. If readers would block
925 in a -rt kernel, but not in a non-rt kernel, SRCU is not
926 necessary. (The -rt patchset turns spinlocks into sleeplocks,
927 hence this distinction.)
929 c. Do you need to treat NMI handlers, hardirq handlers,
930 and code segments with preemption disabled (whether
931 via preempt_disable(), local_irq_save(), local_bh_disable(),
932 or some other mechanism) as if they were explicit RCU readers?
933 If so, RCU-sched is the only choice that will work for you.
935 d. Do you need RCU grace periods to complete even in the face
936 of softirq monopolization of one or more of the CPUs? For
937 example, is your code subject to network-based denial-of-service
938 attacks? If so, you should disable softirq across your readers,
939 for example, by using rcu_read_lock_bh().
941 e. Is your workload too update-intensive for normal use of
942 RCU, but inappropriate for other synchronization mechanisms?
943 If so, consider SLAB_TYPESAFE_BY_RCU (which was originally
944 named SLAB_DESTROY_BY_RCU). But please be careful!
946 f. Do you need read-side critical sections that are respected
947 even though they are in the middle of the idle loop, during
948 user-mode execution, or on an offlined CPU? If so, SRCU is the
949 only choice that will work for you.
951 g. Otherwise, use RCU.
953 Of course, this all assumes that you have determined that RCU is in fact
954 the right tool for your job.
957 8. ANSWERS TO QUICK QUIZZES
959 Quick Quiz #1: Why is this argument naive? How could a deadlock
960 occur when using this algorithm in a real-world Linux
961 kernel? [Referring to the lock-based "toy" RCU
964 Answer: Consider the following sequence of events:
966 1. CPU 0 acquires some unrelated lock, call it
967 "problematic_lock", disabling irq via
970 2. CPU 1 enters synchronize_rcu(), write-acquiring
973 3. CPU 0 enters rcu_read_lock(), but must wait
974 because CPU 1 holds rcu_gp_mutex.
976 4. CPU 1 is interrupted, and the irq handler
977 attempts to acquire problematic_lock.
979 The system is now deadlocked.
981 One way to avoid this deadlock is to use an approach like
982 that of CONFIG_PREEMPT_RT, where all normal spinlocks
983 become blocking locks, and all irq handlers execute in
984 the context of special tasks. In this case, in step 4
985 above, the irq handler would block, allowing CPU 1 to
986 release rcu_gp_mutex, avoiding the deadlock.
988 Even in the absence of deadlock, this RCU implementation
989 allows latency to "bleed" from readers to other
990 readers through synchronize_rcu(). To see this,
991 consider task A in an RCU read-side critical section
992 (thus read-holding rcu_gp_mutex), task B blocked
993 attempting to write-acquire rcu_gp_mutex, and
994 task C blocked in rcu_read_lock() attempting to
995 read_acquire rcu_gp_mutex. Task A's RCU read-side
996 latency is holding up task C, albeit indirectly via
999 Realtime RCU implementations therefore use a counter-based
1000 approach where tasks in RCU read-side critical sections
1001 cannot be blocked by tasks executing synchronize_rcu().
1003 Quick Quiz #2: Give an example where Classic RCU's read-side
1004 overhead is -negative-.
1006 Answer: Imagine a single-CPU system with a non-CONFIG_PREEMPT
1007 kernel where a routing table is used by process-context
1008 code, but can be updated by irq-context code (for example,
1009 by an "ICMP REDIRECT" packet). The usual way of handling
1010 this would be to have the process-context code disable
1011 interrupts while searching the routing table. Use of
1012 RCU allows such interrupt-disabling to be dispensed with.
1013 Thus, without RCU, you pay the cost of disabling interrupts,
1014 and with RCU you don't.
1016 One can argue that the overhead of RCU in this
1017 case is negative with respect to the single-CPU
1018 interrupt-disabling approach. Others might argue that
1019 the overhead of RCU is merely zero, and that replacing
1020 the positive overhead of the interrupt-disabling scheme
1021 with the zero-overhead RCU scheme does not constitute
1024 In real life, of course, things are more complex. But
1025 even the theoretical possibility of negative overhead for
1026 a synchronization primitive is a bit unexpected. ;-)
1028 Quick Quiz #3: If it is illegal to block in an RCU read-side
1029 critical section, what the heck do you do in
1030 PREEMPT_RT, where normal spinlocks can block???
1032 Answer: Just as PREEMPT_RT permits preemption of spinlock
1033 critical sections, it permits preemption of RCU
1034 read-side critical sections. It also permits
1035 spinlocks blocking while in RCU read-side critical
1038 Why the apparent inconsistency? Because it is
1039 possible to use priority boosting to keep the RCU
1040 grace periods short if need be (for example, if running
1041 short of memory). In contrast, if blocking waiting
1042 for (say) network reception, there is no way to know
1043 what should be boosted. Especially given that the
1044 process we need to boost might well be a human being
1045 who just went out for a pizza or something. And although
1046 a computer-operated cattle prod might arouse serious
1047 interest, it might also provoke serious objections.
1048 Besides, how does the computer know what pizza parlor
1049 the human being went to???
1054 My thanks to the people who helped make this human-readable, including
1055 Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern.
1058 For more information, see http://www.rdrop.com/users/paulmck/RCU.