1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
2 "http://www.w3.org/TR/html4/loose.dtd">
4 <head><title>A Tour Through RCU's Requirements [LWN.net]</title>
5 <meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=utf-8">
7 <h1>A Tour Through RCU's Requirements</h1>
9 <p>Copyright IBM Corporation, 2015</p>
10 <p>Author: Paul E. McKenney</p>
11 <p><i>The initial version of this document appeared in the
12 <a href="https://lwn.net/">LWN</a> articles
13 <a href="https://lwn.net/Articles/652156/">here</a>,
14 <a href="https://lwn.net/Articles/652677/">here</a>, and
15 <a href="https://lwn.net/Articles/653326/">here</a>.</i></p>
20 Read-copy update (RCU) is a synchronization mechanism that is often
21 used as a replacement for reader-writer locking.
22 RCU is unusual in that updaters do not block readers,
23 which means that RCU's read-side primitives can be exceedingly fast
25 In addition, updaters can make useful forward progress concurrently
27 However, all this concurrency between RCU readers and updaters does raise
28 the question of exactly what RCU readers are doing, which in turn
29 raises the question of exactly what RCU's requirements are.
32 This document therefore summarizes RCU's requirements, and can be thought
33 of as an informal, high-level specification for RCU.
34 It is important to understand that RCU's specification is primarily
36 in fact, I learned about many of these requirements the hard way.
37 This situation might cause some consternation, however, not only
38 has this learning process been a lot of fun, but it has also been
39 a great privilege to work with so many people willing to apply
40 technologies in interesting new ways.
43 All that aside, here are the categories of currently known RCU requirements:
47 <li> <a href="#Fundamental Requirements">
48 Fundamental Requirements</a>
49 <li> <a href="#Fundamental Non-Requirements">Fundamental Non-Requirements</a>
50 <li> <a href="#Parallelism Facts of Life">
51 Parallelism Facts of Life</a>
52 <li> <a href="#Quality-of-Implementation Requirements">
53 Quality-of-Implementation Requirements</a>
54 <li> <a href="#Linux Kernel Complications">
55 Linux Kernel Complications</a>
56 <li> <a href="#Software-Engineering Requirements">
57 Software-Engineering Requirements</a>
58 <li> <a href="#Other RCU Flavors">
60 <li> <a href="#Possible Future Changes">
61 Possible Future Changes</a>
65 This is followed by a <a href="#Summary">summary</a>,
66 however, the answers to each quick quiz immediately follows the quiz.
67 Select the big white space with your mouse to see the answer.
69 <h2><a name="Fundamental Requirements">Fundamental Requirements</a></h2>
72 RCU's fundamental requirements are the closest thing RCU has to hard
73 mathematical requirements.
77 <li> <a href="#Grace-Period Guarantee">
78 Grace-Period Guarantee</a>
79 <li> <a href="#Publish-Subscribe Guarantee">
80 Publish-Subscribe Guarantee</a>
81 <li> <a href="#Memory-Barrier Guarantees">
82 Memory-Barrier Guarantees</a>
83 <li> <a href="#RCU Primitives Guaranteed to Execute Unconditionally">
84 RCU Primitives Guaranteed to Execute Unconditionally</a>
85 <li> <a href="#Guaranteed Read-to-Write Upgrade">
86 Guaranteed Read-to-Write Upgrade</a>
89 <h3><a name="Grace-Period Guarantee">Grace-Period Guarantee</a></h3>
92 RCU's grace-period guarantee is unusual in being premeditated:
93 Jack Slingwine and I had this guarantee firmly in mind when we started
94 work on RCU (then called “rclock”) in the early 1990s.
95 That said, the past two decades of experience with RCU have produced
96 a much more detailed understanding of this guarantee.
99 RCU's grace-period guarantee allows updaters to wait for the completion
100 of all pre-existing RCU read-side critical sections.
101 An RCU read-side critical section
102 begins with the marker <tt>rcu_read_lock()</tt> and ends with
103 the marker <tt>rcu_read_unlock()</tt>.
104 These markers may be nested, and RCU treats a nested set as one
105 big RCU read-side critical section.
106 Production-quality implementations of <tt>rcu_read_lock()</tt> and
107 <tt>rcu_read_unlock()</tt> are extremely lightweight, and in
108 fact have exactly zero overhead in Linux kernels built for production
109 use with <tt>CONFIG_PREEMPT=n</tt>.
112 This guarantee allows ordering to be enforced with extremely low
113 overhead to readers, for example:
127 11 void thread1(void)
130 14 synchronize_rcu();
137 Because the <tt>synchronize_rcu()</tt> on line 14 waits for
138 all pre-existing readers, any instance of <tt>thread0()</tt> that
139 loads a value of zero from <tt>x</tt> must complete before
140 <tt>thread1()</tt> stores to <tt>y</tt>, so that instance must
141 also load a value of zero from <tt>y</tt>.
142 Similarly, any instance of <tt>thread0()</tt> that loads a value of
143 one from <tt>y</tt> must have started after the
144 <tt>synchronize_rcu()</tt> started, and must therefore also load
145 a value of one from <tt>x</tt>.
146 Therefore, the outcome:
149 (r1 == 0 && r2 == 1)
155 <tr><th> </th></tr>
156 <tr><th align="left">Quick Quiz:</th></tr>
159 You said that updaters can make useful forward progress concurrently
160 with readers, but pre-existing readers will block
161 <tt>synchronize_rcu()</tt>!!!
162 Just who are you trying to fool???
164 <tr><th align="left">Answer:</th></tr>
165 <tr><td bgcolor="#ffffff"><font color="ffffff">
166 First, if updaters do not wish to be blocked by readers, they can use
167 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt>, which will
169 Second, even when using <tt>synchronize_rcu()</tt>, the other
170 update-side code does run concurrently with readers, whether
173 <tr><td> </td></tr>
177 This scenario resembles one of the first uses of RCU in
178 <a href="https://en.wikipedia.org/wiki/DYNIX">DYNIX/ptx</a>,
179 which managed a distributed lock manager's transition into
180 a state suitable for handling recovery from node failure,
181 more or less as follows:
185 1 #define STATE_NORMAL 0
186 2 #define STATE_WANT_RECOVERY 1
187 3 #define STATE_RECOVERING 2
188 4 #define STATE_WANT_NORMAL 3
190 6 int state = STATE_NORMAL;
192 8 void do_something_dlm(void)
197 13 state_snap = READ_ONCE(state);
198 14 if (state_snap == STATE_NORMAL)
201 17 do_something_carefully();
202 18 rcu_read_unlock();
205 21 void start_recovery(void)
207 23 WRITE_ONCE(state, STATE_WANT_RECOVERY);
208 24 synchronize_rcu();
209 25 WRITE_ONCE(state, STATE_RECOVERING);
211 27 WRITE_ONCE(state, STATE_WANT_NORMAL);
212 28 synchronize_rcu();
213 29 WRITE_ONCE(state, STATE_NORMAL);
219 The RCU read-side critical section in <tt>do_something_dlm()</tt>
220 works with the <tt>synchronize_rcu()</tt> in <tt>start_recovery()</tt>
221 to guarantee that <tt>do_something()</tt> never runs concurrently
222 with <tt>recovery()</tt>, but with little or no synchronization
223 overhead in <tt>do_something_dlm()</tt>.
226 <tr><th> </th></tr>
227 <tr><th align="left">Quick Quiz:</th></tr>
229 Why is the <tt>synchronize_rcu()</tt> on line 28 needed?
231 <tr><th align="left">Answer:</th></tr>
232 <tr><td bgcolor="#ffffff"><font color="ffffff">
233 Without that extra grace period, memory reordering could result in
234 <tt>do_something_dlm()</tt> executing <tt>do_something()</tt>
235 concurrently with the last bits of <tt>recovery()</tt>.
237 <tr><td> </td></tr>
241 In order to avoid fatal problems such as deadlocks,
242 an RCU read-side critical section must not contain calls to
243 <tt>synchronize_rcu()</tt>.
244 Similarly, an RCU read-side critical section must not
245 contain anything that waits, directly or indirectly, on completion of
246 an invocation of <tt>synchronize_rcu()</tt>.
249 Although RCU's grace-period guarantee is useful in and of itself, with
250 <a href="https://lwn.net/Articles/573497/">quite a few use cases</a>,
251 it would be good to be able to use RCU to coordinate read-side
252 access to linked data structures.
253 For this, the grace-period guarantee is not sufficient, as can
254 be seen in function <tt>add_gp_buggy()</tt> below.
255 We will look at the reader's code later, but in the meantime, just think of
256 the reader as locklessly picking up the <tt>gp</tt> pointer,
257 and, if the value loaded is non-<tt>NULL</tt>, locklessly accessing the
258 <tt>->a</tt> and <tt>->b</tt> fields.
262 1 bool add_gp_buggy(int a, int b)
264 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
267 6 spin_lock(&gp_lock);
268 7 if (rcu_access_pointer(gp)) {
269 8 spin_unlock(&gp_lock);
274 13 gp = p; /* ORDERING BUG */
275 14 spin_unlock(&gp_lock);
282 The problem is that both the compiler and weakly ordered CPUs are within
283 their rights to reorder this code as follows:
287 1 bool add_gp_buggy_optimized(int a, int b)
289 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
292 6 spin_lock(&gp_lock);
293 7 if (rcu_access_pointer(gp)) {
294 8 spin_unlock(&gp_lock);
297 <b>11 gp = p; /* ORDERING BUG */
300 14 spin_unlock(&gp_lock);
307 If an RCU reader fetches <tt>gp</tt> just after
308 <tt>add_gp_buggy_optimized</tt> executes line 11,
309 it will see garbage in the <tt>->a</tt> and <tt>->b</tt>
311 And this is but one of many ways in which compiler and hardware optimizations
313 Therefore, we clearly need some way to prevent the compiler and the CPU from
314 reordering in this manner, which brings us to the publish-subscribe
315 guarantee discussed in the next section.
317 <h3><a name="Publish-Subscribe Guarantee">Publish/Subscribe Guarantee</a></h3>
320 RCU's publish-subscribe guarantee allows data to be inserted
321 into a linked data structure without disrupting RCU readers.
322 The updater uses <tt>rcu_assign_pointer()</tt> to insert the
323 new data, and readers use <tt>rcu_dereference()</tt> to
324 access data, whether new or old.
325 The following shows an example of insertion:
329 1 bool add_gp(int a, int b)
331 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
334 6 spin_lock(&gp_lock);
335 7 if (rcu_access_pointer(gp)) {
336 8 spin_unlock(&gp_lock);
341 13 rcu_assign_pointer(gp, p);
342 14 spin_unlock(&gp_lock);
349 The <tt>rcu_assign_pointer()</tt> on line 13 is conceptually
350 equivalent to a simple assignment statement, but also guarantees
351 that its assignment will
352 happen after the two assignments in lines 11 and 12,
353 similar to the C11 <tt>memory_order_release</tt> store operation.
354 It also prevents any number of “interesting” compiler
355 optimizations, for example, the use of <tt>gp</tt> as a scratch
356 location immediately preceding the assignment.
359 <tr><th> </th></tr>
360 <tr><th align="left">Quick Quiz:</th></tr>
362 But <tt>rcu_assign_pointer()</tt> does nothing to prevent the
363 two assignments to <tt>p->a</tt> and <tt>p->b</tt>
364 from being reordered.
365 Can't that also cause problems?
367 <tr><th align="left">Answer:</th></tr>
368 <tr><td bgcolor="#ffffff"><font color="ffffff">
370 The readers cannot see either of these two fields until
371 the assignment to <tt>gp</tt>, by which time both fields are
373 So reordering the assignments
374 to <tt>p->a</tt> and <tt>p->b</tt> cannot possibly
377 <tr><td> </td></tr>
381 It is tempting to assume that the reader need not do anything special
382 to control its accesses to the RCU-protected data,
383 as shown in <tt>do_something_gp_buggy()</tt> below:
387 1 bool do_something_gp_buggy(void)
390 4 p = gp; /* OPTIMIZATIONS GALORE!!! */
392 6 do_something(p->a, p->b);
396 10 rcu_read_unlock();
403 However, this temptation must be resisted because there are a
404 surprisingly large number of ways that the compiler
406 <a href="https://h71000.www7.hp.com/wizard/wiz_2637.html">DEC Alpha CPUs</a>)
407 can trip this code up.
408 For but one example, if the compiler were short of registers, it
409 might choose to refetch from <tt>gp</tt> rather than keeping
410 a separate copy in <tt>p</tt> as follows:
414 1 bool do_something_gp_buggy_optimized(void)
417 4 if (gp) { /* OPTIMIZATIONS GALORE!!! */
418 <b> 5 do_something(gp->a, gp->b);</b>
429 If this function ran concurrently with a series of updates that
430 replaced the current structure with a new one,
431 the fetches of <tt>gp->a</tt>
432 and <tt>gp->b</tt> might well come from two different structures,
433 which could cause serious confusion.
434 To prevent this (and much else besides), <tt>do_something_gp()</tt> uses
435 <tt>rcu_dereference()</tt> to fetch from <tt>gp</tt>:
439 1 bool do_something_gp(void)
442 4 p = rcu_dereference(gp);
444 6 do_something(p->a, p->b);
448 10 rcu_read_unlock();
455 The <tt>rcu_dereference()</tt> uses volatile casts and (for DEC Alpha)
456 memory barriers in the Linux kernel.
458 <a href="http://www.rdrop.com/users/paulmck/RCU/consume.2015.07.13a.pdf">high-quality implementation of C11 <tt>memory_order_consume</tt> [PDF]</a>
459 ever appear, then <tt>rcu_dereference()</tt> could be implemented
460 as a <tt>memory_order_consume</tt> load.
461 Regardless of the exact implementation, a pointer fetched by
462 <tt>rcu_dereference()</tt> may not be used outside of the
463 outermost RCU read-side critical section containing that
464 <tt>rcu_dereference()</tt>, unless protection of
465 the corresponding data element has been passed from RCU to some
466 other synchronization mechanism, most commonly locking or
467 <a href="https://www.kernel.org/doc/Documentation/RCU/rcuref.txt">reference counting</a>.
470 In short, updaters use <tt>rcu_assign_pointer()</tt> and readers
471 use <tt>rcu_dereference()</tt>, and these two RCU API elements
472 work together to ensure that readers have a consistent view of
473 newly added data elements.
476 Of course, it is also necessary to remove elements from RCU-protected
477 data structures, for example, using the following process:
480 <li> Remove the data element from the enclosing structure.
481 <li> Wait for all pre-existing RCU read-side critical sections
482 to complete (because only pre-existing readers can possibly have
483 a reference to the newly removed data element).
484 <li> At this point, only the updater has a reference to the
485 newly removed data element, so it can safely reclaim
486 the data element, for example, by passing it to <tt>kfree()</tt>.
489 This process is implemented by <tt>remove_gp_synchronous()</tt>:
493 1 bool remove_gp_synchronous(void)
497 5 spin_lock(&gp_lock);
498 6 p = rcu_access_pointer(gp);
500 8 spin_unlock(&gp_lock);
503 11 rcu_assign_pointer(gp, NULL);
504 12 spin_unlock(&gp_lock);
505 13 synchronize_rcu();
513 This function is straightforward, with line 13 waiting for a grace
514 period before line 14 frees the old data element.
515 This waiting ensures that readers will reach line 7 of
516 <tt>do_something_gp()</tt> before the data element referenced by
518 The <tt>rcu_access_pointer()</tt> on line 6 is similar to
519 <tt>rcu_dereference()</tt>, except that:
522 <li> The value returned by <tt>rcu_access_pointer()</tt>
523 cannot be dereferenced.
524 If you want to access the value pointed to as well as
525 the pointer itself, use <tt>rcu_dereference()</tt>
526 instead of <tt>rcu_access_pointer()</tt>.
527 <li> The call to <tt>rcu_access_pointer()</tt> need not be
529 In contrast, <tt>rcu_dereference()</tt> must either be
530 within an RCU read-side critical section or in a code
531 segment where the pointer cannot change, for example, in
532 code protected by the corresponding update-side lock.
536 <tr><th> </th></tr>
537 <tr><th align="left">Quick Quiz:</th></tr>
539 Without the <tt>rcu_dereference()</tt> or the
540 <tt>rcu_access_pointer()</tt>, what destructive optimizations
541 might the compiler make use of?
543 <tr><th align="left">Answer:</th></tr>
544 <tr><td bgcolor="#ffffff"><font color="ffffff">
545 Let's start with what happens to <tt>do_something_gp()</tt>
546 if it fails to use <tt>rcu_dereference()</tt>.
547 It could reuse a value formerly fetched from this same pointer.
548 It could also fetch the pointer from <tt>gp</tt> in a byte-at-a-time
549 manner, resulting in <i>load tearing</i>, in turn resulting a bytewise
550 mash-up of two distinct pointer values.
551 It might even use value-speculation optimizations, where it makes
552 a wrong guess, but by the time it gets around to checking the
553 value, an update has changed the pointer to match the wrong guess.
554 Too bad about any dereferences that returned pre-initialization garbage
558 <p><font color="ffffff">
559 For <tt>remove_gp_synchronous()</tt>, as long as all modifications
560 to <tt>gp</tt> are carried out while holding <tt>gp_lock</tt>,
561 the above optimizations are harmless.
562 However, <tt>sparse</tt> will complain if you
563 define <tt>gp</tt> with <tt>__rcu</tt> and then
564 access it without using
565 either <tt>rcu_access_pointer()</tt> or <tt>rcu_dereference()</tt>.
567 <tr><td> </td></tr>
571 In short, RCU's publish-subscribe guarantee is provided by the combination
572 of <tt>rcu_assign_pointer()</tt> and <tt>rcu_dereference()</tt>.
573 This guarantee allows data elements to be safely added to RCU-protected
574 linked data structures without disrupting RCU readers.
575 This guarantee can be used in combination with the grace-period
576 guarantee to also allow data elements to be removed from RCU-protected
577 linked data structures, again without disrupting RCU readers.
580 This guarantee was only partially premeditated.
581 DYNIX/ptx used an explicit memory barrier for publication, but had nothing
582 resembling <tt>rcu_dereference()</tt> for subscription, nor did it
583 have anything resembling the <tt>smp_read_barrier_depends()</tt>
584 that was later subsumed into <tt>rcu_dereference()</tt>.
585 The need for these operations made itself known quite suddenly at a
586 late-1990s meeting with the DEC Alpha architects, back in the days when
587 DEC was still a free-standing company.
588 It took the Alpha architects a good hour to convince me that any sort
589 of barrier would ever be needed, and it then took me a good <i>two</i> hours
590 to convince them that their documentation did not make this point clear.
591 More recent work with the C and C++ standards committees have provided
592 much education on tricks and traps from the compiler.
593 In short, compilers were much less tricky in the early 1990s, but in
594 2015, don't even think about omitting <tt>rcu_dereference()</tt>!
596 <h3><a name="Memory-Barrier Guarantees">Memory-Barrier Guarantees</a></h3>
599 The previous section's simple linked-data-structure scenario clearly
600 demonstrates the need for RCU's stringent memory-ordering guarantees on
601 systems with more than one CPU:
604 <li> Each CPU that has an RCU read-side critical section that
605 begins before <tt>synchronize_rcu()</tt> starts is
606 guaranteed to execute a full memory barrier between the time
607 that the RCU read-side critical section ends and the time that
608 <tt>synchronize_rcu()</tt> returns.
609 Without this guarantee, a pre-existing RCU read-side critical section
610 might hold a reference to the newly removed <tt>struct foo</tt>
611 after the <tt>kfree()</tt> on line 14 of
612 <tt>remove_gp_synchronous()</tt>.
613 <li> Each CPU that has an RCU read-side critical section that ends
614 after <tt>synchronize_rcu()</tt> returns is guaranteed
615 to execute a full memory barrier between the time that
616 <tt>synchronize_rcu()</tt> begins and the time that the RCU
617 read-side critical section begins.
618 Without this guarantee, a later RCU read-side critical section
619 running after the <tt>kfree()</tt> on line 14 of
620 <tt>remove_gp_synchronous()</tt> might
621 later run <tt>do_something_gp()</tt> and find the
622 newly deleted <tt>struct foo</tt>.
623 <li> If the task invoking <tt>synchronize_rcu()</tt> remains
624 on a given CPU, then that CPU is guaranteed to execute a full
625 memory barrier sometime during the execution of
626 <tt>synchronize_rcu()</tt>.
627 This guarantee ensures that the <tt>kfree()</tt> on
628 line 14 of <tt>remove_gp_synchronous()</tt> really does
629 execute after the removal on line 11.
630 <li> If the task invoking <tt>synchronize_rcu()</tt> migrates
631 among a group of CPUs during that invocation, then each of the
632 CPUs in that group is guaranteed to execute a full memory barrier
633 sometime during the execution of <tt>synchronize_rcu()</tt>.
634 This guarantee also ensures that the <tt>kfree()</tt> on
635 line 14 of <tt>remove_gp_synchronous()</tt> really does
636 execute after the removal on
637 line 11, but also in the case where the thread executing the
638 <tt>synchronize_rcu()</tt> migrates in the meantime.
642 <tr><th> </th></tr>
643 <tr><th align="left">Quick Quiz:</th></tr>
645 Given that multiple CPUs can start RCU read-side critical sections
646 at any time without any ordering whatsoever, how can RCU possibly
647 tell whether or not a given RCU read-side critical section starts
648 before a given instance of <tt>synchronize_rcu()</tt>?
650 <tr><th align="left">Answer:</th></tr>
651 <tr><td bgcolor="#ffffff"><font color="ffffff">
652 If RCU cannot tell whether or not a given
653 RCU read-side critical section starts before a
654 given instance of <tt>synchronize_rcu()</tt>,
655 then it must assume that the RCU read-side critical section
657 In other words, a given instance of <tt>synchronize_rcu()</tt>
658 can avoid waiting on a given RCU read-side critical section only
659 if it can prove that <tt>synchronize_rcu()</tt> started first.
662 <p><font color="ffffff">
663 A related question is “When <tt>rcu_read_lock()</tt>
664 doesn't generate any code, why does it matter how it relates
665 to a grace period?”
666 The answer is that it is not the relationship of
667 <tt>rcu_read_lock()</tt> itself that is important, but rather
668 the relationship of the code within the enclosed RCU read-side
669 critical section to the code preceding and following the
671 If we take this viewpoint, then a given RCU read-side critical
672 section begins before a given grace period when some access
673 preceding the grace period observes the effect of some access
674 within the critical section, in which case none of the accesses
675 within the critical section may observe the effects of any
676 access following the grace period.
679 <p><font color="ffffff">
680 As of late 2016, mathematical models of RCU take this
681 viewpoint, for example, see slides 62 and 63
683 <a href="http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.2016.10.04c.LCE.pdf">2016 LinuxCon EU</a>
686 <tr><td> </td></tr>
690 <tr><th> </th></tr>
691 <tr><th align="left">Quick Quiz:</th></tr>
693 The first and second guarantees require unbelievably strict ordering!
694 Are all these memory barriers <i> really</i> required?
696 <tr><th align="left">Answer:</th></tr>
697 <tr><td bgcolor="#ffffff"><font color="ffffff">
698 Yes, they really are required.
699 To see why the first guarantee is required, consider the following
704 <li> <font color="ffffff">
705 CPU 1: <tt>rcu_read_lock()</tt>
707 <li> <font color="ffffff">
708 CPU 1: <tt>q = rcu_dereference(gp);
709 /* Very likely to return p. */</tt>
711 <li> <font color="ffffff">
712 CPU 0: <tt>list_del_rcu(p);</tt>
714 <li> <font color="ffffff">
715 CPU 0: <tt>synchronize_rcu()</tt> starts.
717 <li> <font color="ffffff">
718 CPU 1: <tt>do_something_with(q->a);
719 /* No smp_mb(), so might happen after kfree(). */</tt>
721 <li> <font color="ffffff">
722 CPU 1: <tt>rcu_read_unlock()</tt>
724 <li> <font color="ffffff">
725 CPU 0: <tt>synchronize_rcu()</tt> returns.
727 <li> <font color="ffffff">
728 CPU 0: <tt>kfree(p);</tt>
732 <p><font color="ffffff">
733 Therefore, there absolutely must be a full memory barrier between the
734 end of the RCU read-side critical section and the end of the
738 <p><font color="ffffff">
739 The sequence of events demonstrating the necessity of the second rule
744 <li> <font color="ffffff">CPU 0: <tt>list_del_rcu(p);</tt>
746 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> starts.
748 <li> <font color="ffffff">CPU 1: <tt>rcu_read_lock()</tt>
750 <li> <font color="ffffff">CPU 1: <tt>q = rcu_dereference(gp);
751 /* Might return p if no memory barrier. */</tt>
753 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> returns.
755 <li> <font color="ffffff">CPU 0: <tt>kfree(p);</tt>
757 <li> <font color="ffffff">
758 CPU 1: <tt>do_something_with(q->a); /* Boom!!! */</tt>
760 <li> <font color="ffffff">CPU 1: <tt>rcu_read_unlock()</tt>
764 <p><font color="ffffff">
765 And similarly, without a memory barrier between the beginning of the
766 grace period and the beginning of the RCU read-side critical section,
767 CPU 1 might end up accessing the freelist.
770 <p><font color="ffffff">
771 The “as if” rule of course applies, so that any
772 implementation that acts as if the appropriate memory barriers
773 were in place is a correct implementation.
774 That said, it is much easier to fool yourself into believing
775 that you have adhered to the as-if rule than it is to actually
778 <tr><td> </td></tr>
782 <tr><th> </th></tr>
783 <tr><th align="left">Quick Quiz:</th></tr>
785 You claim that <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
786 generate absolutely no code in some kernel builds.
787 This means that the compiler might arbitrarily rearrange consecutive
788 RCU read-side critical sections.
789 Given such rearrangement, if a given RCU read-side critical section
790 is done, how can you be sure that all prior RCU read-side critical
792 Won't the compiler rearrangements make that impossible to determine?
794 <tr><th align="left">Answer:</th></tr>
795 <tr><td bgcolor="#ffffff"><font color="ffffff">
796 In cases where <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
797 generate absolutely no code, RCU infers quiescent states only at
798 special locations, for example, within the scheduler.
799 Because calls to <tt>schedule()</tt> had better prevent calling-code
800 accesses to shared variables from being rearranged across the call to
801 <tt>schedule()</tt>, if RCU detects the end of a given RCU read-side
802 critical section, it will necessarily detect the end of all prior
803 RCU read-side critical sections, no matter how aggressively the
804 compiler scrambles the code.
807 <p><font color="ffffff">
808 Again, this all assumes that the compiler cannot scramble code across
809 calls to the scheduler, out of interrupt handlers, into the idle loop,
810 into user-mode code, and so on.
811 But if your kernel build allows that sort of scrambling, you have broken
812 far more than just RCU!
814 <tr><td> </td></tr>
818 Note that these memory-barrier requirements do not replace the fundamental
819 RCU requirement that a grace period wait for all pre-existing readers.
820 On the contrary, the memory barriers called out in this section must operate in
821 such a way as to <i>enforce</i> this fundamental requirement.
822 Of course, different implementations enforce this requirement in different
823 ways, but enforce it they must.
825 <h3><a name="RCU Primitives Guaranteed to Execute Unconditionally">RCU Primitives Guaranteed to Execute Unconditionally</a></h3>
828 The common-case RCU primitives are unconditional.
829 They are invoked, they do their job, and they return, with no possibility
830 of error, and no need to retry.
831 This is a key RCU design philosophy.
834 However, this philosophy is pragmatic rather than pigheaded.
835 If someone comes up with a good justification for a particular conditional
836 RCU primitive, it might well be implemented and added.
837 After all, this guarantee was reverse-engineered, not premeditated.
838 The unconditional nature of the RCU primitives was initially an
839 accident of implementation, and later experience with synchronization
840 primitives with conditional primitives caused me to elevate this
841 accident to a guarantee.
842 Therefore, the justification for adding a conditional primitive to
843 RCU would need to be based on detailed and compelling use cases.
845 <h3><a name="Guaranteed Read-to-Write Upgrade">Guaranteed Read-to-Write Upgrade</a></h3>
848 As far as RCU is concerned, it is always possible to carry out an
849 update within an RCU read-side critical section.
850 For example, that RCU read-side critical section might search for
851 a given data element, and then might acquire the update-side
852 spinlock in order to update that element, all while remaining
853 in that RCU read-side critical section.
854 Of course, it is necessary to exit the RCU read-side critical section
855 before invoking <tt>synchronize_rcu()</tt>, however, this
856 inconvenience can be avoided through use of the
857 <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt> API members
858 described later in this document.
861 <tr><th> </th></tr>
862 <tr><th align="left">Quick Quiz:</th></tr>
864 But how does the upgrade-to-write operation exclude other readers?
866 <tr><th align="left">Answer:</th></tr>
867 <tr><td bgcolor="#ffffff"><font color="ffffff">
868 It doesn't, just like normal RCU updates, which also do not exclude
871 <tr><td> </td></tr>
875 This guarantee allows lookup code to be shared between read-side
876 and update-side code, and was premeditated, appearing in the earliest
877 DYNIX/ptx RCU documentation.
879 <h2><a name="Fundamental Non-Requirements">Fundamental Non-Requirements</a></h2>
882 RCU provides extremely lightweight readers, and its read-side guarantees,
883 though quite useful, are correspondingly lightweight.
884 It is therefore all too easy to assume that RCU is guaranteeing more
886 Of course, the list of things that RCU does not guarantee is infinitely
887 long, however, the following sections list a few non-guarantees that
888 have caused confusion.
889 Except where otherwise noted, these non-guarantees were premeditated.
892 <li> <a href="#Readers Impose Minimal Ordering">
893 Readers Impose Minimal Ordering</a>
894 <li> <a href="#Readers Do Not Exclude Updaters">
895 Readers Do Not Exclude Updaters</a>
896 <li> <a href="#Updaters Only Wait For Old Readers">
897 Updaters Only Wait For Old Readers</a>
898 <li> <a href="#Grace Periods Don't Partition Read-Side Critical Sections">
899 Grace Periods Don't Partition Read-Side Critical Sections</a>
900 <li> <a href="#Read-Side Critical Sections Don't Partition Grace Periods">
901 Read-Side Critical Sections Don't Partition Grace Periods</a>
902 <li> <a href="#Disabling Preemption Does Not Block Grace Periods">
903 Disabling Preemption Does Not Block Grace Periods</a>
906 <h3><a name="Readers Impose Minimal Ordering">Readers Impose Minimal Ordering</a></h3>
909 Reader-side markers such as <tt>rcu_read_lock()</tt> and
910 <tt>rcu_read_unlock()</tt> provide absolutely no ordering guarantees
911 except through their interaction with the grace-period APIs such as
912 <tt>synchronize_rcu()</tt>.
913 To see this, consider the following pair of threads:
927 11 void thread1(void)
930 14 r1 = READ_ONCE(y);
931 15 rcu_read_unlock();
933 17 r2 = READ_ONCE(x);
934 18 rcu_read_unlock();
940 After <tt>thread0()</tt> and <tt>thread1()</tt> execute
941 concurrently, it is quite possible to have
945 (r1 == 1 && r2 == 0)
949 (that is, <tt>y</tt> appears to have been assigned before <tt>x</tt>),
950 which would not be possible if <tt>rcu_read_lock()</tt> and
951 <tt>rcu_read_unlock()</tt> had much in the way of ordering
953 But they do not, so the CPU is within its rights
954 to do significant reordering.
955 This is by design: Any significant ordering constraints would slow down
956 these fast-path APIs.
959 <tr><th> </th></tr>
960 <tr><th align="left">Quick Quiz:</th></tr>
962 Can't the compiler also reorder this code?
964 <tr><th align="left">Answer:</th></tr>
965 <tr><td bgcolor="#ffffff"><font color="ffffff">
966 No, the volatile casts in <tt>READ_ONCE()</tt> and
967 <tt>WRITE_ONCE()</tt> prevent the compiler from reordering in
968 this particular case.
970 <tr><td> </td></tr>
973 <h3><a name="Readers Do Not Exclude Updaters">Readers Do Not Exclude Updaters</a></h3>
976 Neither <tt>rcu_read_lock()</tt> nor <tt>rcu_read_unlock()</tt>
978 All they do is to prevent grace periods from ending.
979 The following example illustrates this:
988 6 do_something_with_nonzero_x();
990 8 WARN_ON(!r2); /* BUG!!! */
992 10 rcu_read_unlock();
995 13 void thread1(void)
997 15 spin_lock(&my_lock);
1000 18 spin_unlock(&my_lock);
1006 If the <tt>thread0()</tt> function's <tt>rcu_read_lock()</tt>
1007 excluded the <tt>thread1()</tt> function's update,
1008 the <tt>WARN_ON()</tt> could never fire.
1009 But the fact is that <tt>rcu_read_lock()</tt> does not exclude
1010 much of anything aside from subsequent grace periods, of which
1011 <tt>thread1()</tt> has none, so the
1012 <tt>WARN_ON()</tt> can and does fire.
1014 <h3><a name="Updaters Only Wait For Old Readers">Updaters Only Wait For Old Readers</a></h3>
1017 It might be tempting to assume that after <tt>synchronize_rcu()</tt>
1018 completes, there are no readers executing.
1019 This temptation must be avoided because
1020 new readers can start immediately after <tt>synchronize_rcu()</tt>
1021 starts, and <tt>synchronize_rcu()</tt> is under no
1022 obligation to wait for these new readers.
1025 <tr><th> </th></tr>
1026 <tr><th align="left">Quick Quiz:</th></tr>
1028 Suppose that synchronize_rcu() did wait until <i>all</i>
1029 readers had completed instead of waiting only on
1030 pre-existing readers.
1031 For how long would the updater be able to rely on there
1034 <tr><th align="left">Answer:</th></tr>
1035 <tr><td bgcolor="#ffffff"><font color="ffffff">
1037 Even if <tt>synchronize_rcu()</tt> were to wait until
1038 all readers had completed, a new reader might start immediately after
1039 <tt>synchronize_rcu()</tt> completed.
1040 Therefore, the code following
1041 <tt>synchronize_rcu()</tt> can <i>never</i> rely on there being
1044 <tr><td> </td></tr>
1047 <h3><a name="Grace Periods Don't Partition Read-Side Critical Sections">
1048 Grace Periods Don't Partition Read-Side Critical Sections</a></h3>
1051 It is tempting to assume that if any part of one RCU read-side critical
1052 section precedes a given grace period, and if any part of another RCU
1053 read-side critical section follows that same grace period, then all of
1054 the first RCU read-side critical section must precede all of the second.
1055 However, this just isn't the case: A single grace period does not
1056 partition the set of RCU read-side critical sections.
1057 An example of this situation can be illustrated as follows, where
1058 <tt>x</tt>, <tt>y</tt>, and <tt>z</tt> are initially all zero:
1062 1 void thread0(void)
1067 6 rcu_read_unlock();
1070 9 void thread1(void)
1072 11 r1 = READ_ONCE(a);
1073 12 synchronize_rcu();
1074 13 WRITE_ONCE(c, 1);
1077 16 void thread2(void)
1080 19 r2 = READ_ONCE(b);
1081 20 r3 = READ_ONCE(c);
1082 21 rcu_read_unlock();
1088 It turns out that the outcome:
1092 (r1 == 1 && r2 == 0 && r3 == 1)
1096 is entirely possible.
1097 The following figure show how this can happen, with each circled
1098 <tt>QS</tt> indicating the point at which RCU recorded a
1099 <i>quiescent state</i> for each thread, that is, a state in which
1100 RCU knows that the thread cannot be in the midst of an RCU read-side
1101 critical section that started before the current grace period:
1103 <p><img src="GPpartitionReaders1.svg" alt="GPpartitionReaders1.svg" width="60%"></p>
1106 If it is necessary to partition RCU read-side critical sections in this
1107 manner, it is necessary to use two grace periods, where the first
1108 grace period is known to end before the second grace period starts:
1112 1 void thread0(void)
1117 6 rcu_read_unlock();
1120 9 void thread1(void)
1122 11 r1 = READ_ONCE(a);
1123 12 synchronize_rcu();
1124 13 WRITE_ONCE(c, 1);
1127 16 void thread2(void)
1129 18 r2 = READ_ONCE(c);
1130 19 synchronize_rcu();
1131 20 WRITE_ONCE(d, 1);
1134 23 void thread3(void)
1137 26 r3 = READ_ONCE(b);
1138 27 r4 = READ_ONCE(d);
1139 28 rcu_read_unlock();
1145 Here, if <tt>(r1 == 1)</tt>, then
1146 <tt>thread0()</tt>'s write to <tt>b</tt> must happen
1147 before the end of <tt>thread1()</tt>'s grace period.
1148 If in addition <tt>(r4 == 1)</tt>, then
1149 <tt>thread3()</tt>'s read from <tt>b</tt> must happen
1150 after the beginning of <tt>thread2()</tt>'s grace period.
1151 If it is also the case that <tt>(r2 == 1)</tt>, then the
1152 end of <tt>thread1()</tt>'s grace period must precede the
1153 beginning of <tt>thread2()</tt>'s grace period.
1154 This mean that the two RCU read-side critical sections cannot overlap,
1155 guaranteeing that <tt>(r3 == 1)</tt>.
1156 As a result, the outcome:
1160 (r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
1167 This non-requirement was also non-premeditated, but became apparent
1168 when studying RCU's interaction with memory ordering.
1170 <h3><a name="Read-Side Critical Sections Don't Partition Grace Periods">
1171 Read-Side Critical Sections Don't Partition Grace Periods</a></h3>
1174 It is also tempting to assume that if an RCU read-side critical section
1175 happens between a pair of grace periods, then those grace periods cannot
1177 However, this temptation leads nowhere good, as can be illustrated by
1178 the following, with all variables initially zero:
1182 1 void thread0(void)
1187 6 rcu_read_unlock();
1190 9 void thread1(void)
1192 11 r1 = READ_ONCE(a);
1193 12 synchronize_rcu();
1194 13 WRITE_ONCE(c, 1);
1197 16 void thread2(void)
1200 19 WRITE_ONCE(d, 1);
1201 20 r2 = READ_ONCE(c);
1202 21 rcu_read_unlock();
1205 24 void thread3(void)
1207 26 r3 = READ_ONCE(d);
1208 27 synchronize_rcu();
1209 28 WRITE_ONCE(e, 1);
1212 31 void thread4(void)
1215 34 r4 = READ_ONCE(b);
1216 35 r5 = READ_ONCE(e);
1217 36 rcu_read_unlock();
1223 In this case, the outcome:
1227 (r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
1231 is entirely possible, as illustrated below:
1233 <p><img src="ReadersPartitionGP1.svg" alt="ReadersPartitionGP1.svg" width="100%"></p>
1236 Again, an RCU read-side critical section can overlap almost all of a
1237 given grace period, just so long as it does not overlap the entire
1239 As a result, an RCU read-side critical section cannot partition a pair
1240 of RCU grace periods.
1243 <tr><th> </th></tr>
1244 <tr><th align="left">Quick Quiz:</th></tr>
1246 How long a sequence of grace periods, each separated by an RCU
1247 read-side critical section, would be required to partition the RCU
1248 read-side critical sections at the beginning and end of the chain?
1250 <tr><th align="left">Answer:</th></tr>
1251 <tr><td bgcolor="#ffffff"><font color="ffffff">
1252 In theory, an infinite number.
1253 In practice, an unknown number that is sensitive to both implementation
1254 details and timing considerations.
1255 Therefore, even in practice, RCU users must abide by the
1256 theoretical rather than the practical answer.
1258 <tr><td> </td></tr>
1261 <h3><a name="Disabling Preemption Does Not Block Grace Periods">
1262 Disabling Preemption Does Not Block Grace Periods</a></h3>
1265 There was a time when disabling preemption on any given CPU would block
1266 subsequent grace periods.
1267 However, this was an accident of implementation and is not a requirement.
1268 And in the current Linux-kernel implementation, disabling preemption
1269 on a given CPU in fact does not block grace periods, as Oleg Nesterov
1270 <a href="https://lkml.kernel.org/g/20150614193825.GA19582@redhat.com">demonstrated</a>.
1273 If you need a preempt-disable region to block grace periods, you need to add
1274 <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>, for example
1279 1 preempt_disable();
1282 4 rcu_read_unlock();
1285 7 /* Spinlocks implicitly disable preemption. */
1286 8 spin_lock(&mylock);
1289 11 rcu_read_unlock();
1290 12 spin_unlock(&mylock);
1295 In theory, you could enter the RCU read-side critical section first,
1296 but it is more efficient to keep the entire RCU read-side critical
1297 section contained in the preempt-disable region as shown above.
1298 Of course, RCU read-side critical sections that extend outside of
1299 preempt-disable regions will work correctly, but such critical sections
1300 can be preempted, which forces <tt>rcu_read_unlock()</tt> to do
1302 And no, this is <i>not</i> an invitation to enclose all of your RCU
1303 read-side critical sections within preempt-disable regions, because
1304 doing so would degrade real-time response.
1307 This non-requirement appeared with preemptible RCU.
1308 If you need a grace period that waits on non-preemptible code regions, use
1309 <a href="#Sched Flavor">RCU-sched</a>.
1311 <h2><a name="Parallelism Facts of Life">Parallelism Facts of Life</a></h2>
1314 These parallelism facts of life are by no means specific to RCU, but
1315 the RCU implementation must abide by them.
1316 They therefore bear repeating:
1319 <li> Any CPU or task may be delayed at any time,
1320 and any attempts to avoid these delays by disabling
1321 preemption, interrupts, or whatever are completely futile.
1322 This is most obvious in preemptible user-level
1323 environments and in virtualized environments (where
1324 a given guest OS's VCPUs can be preempted at any time by
1325 the underlying hypervisor), but can also happen in bare-metal
1326 environments due to ECC errors, NMIs, and other hardware
1328 Although a delay of more than about 20 seconds can result
1329 in splats, the RCU implementation is obligated to use
1330 algorithms that can tolerate extremely long delays, but where
1331 “extremely long” is not long enough to allow
1332 wrap-around when incrementing a 64-bit counter.
1333 <li> Both the compiler and the CPU can reorder memory accesses.
1334 Where it matters, RCU must use compiler directives and
1335 memory-barrier instructions to preserve ordering.
1336 <li> Conflicting writes to memory locations in any given cache line
1337 will result in expensive cache misses.
1338 Greater numbers of concurrent writes and more-frequent
1339 concurrent writes will result in more dramatic slowdowns.
1340 RCU is therefore obligated to use algorithms that have
1341 sufficient locality to avoid significant performance and
1342 scalability problems.
1343 <li> As a rough rule of thumb, only one CPU's worth of processing
1344 may be carried out under the protection of any given exclusive
1346 RCU must therefore use scalable locking designs.
1347 <li> Counters are finite, especially on 32-bit systems.
1348 RCU's use of counters must therefore tolerate counter wrap,
1349 or be designed such that counter wrap would take way more
1350 time than a single system is likely to run.
1351 An uptime of ten years is quite possible, a runtime
1352 of a century much less so.
1353 As an example of the latter, RCU's dyntick-idle nesting counter
1354 allows 54 bits for interrupt nesting level (this counter
1355 is 64 bits even on a 32-bit system).
1356 Overflowing this counter requires 2<sup>54</sup>
1357 half-interrupts on a given CPU without that CPU ever going idle.
1358 If a half-interrupt happened every microsecond, it would take
1359 570 years of runtime to overflow this counter, which is currently
1360 believed to be an acceptably long time.
1361 <li> Linux systems can have thousands of CPUs running a single
1362 Linux kernel in a single shared-memory environment.
1363 RCU must therefore pay close attention to high-end scalability.
1367 This last parallelism fact of life means that RCU must pay special
1368 attention to the preceding facts of life.
1369 The idea that Linux might scale to systems with thousands of CPUs would
1370 have been met with some skepticism in the 1990s, but these requirements
1371 would have otherwise have been unsurprising, even in the early 1990s.
1373 <h2><a name="Quality-of-Implementation Requirements">Quality-of-Implementation Requirements</a></h2>
1376 These sections list quality-of-implementation requirements.
1377 Although an RCU implementation that ignores these requirements could
1378 still be used, it would likely be subject to limitations that would
1379 make it inappropriate for industrial-strength production use.
1380 Classes of quality-of-implementation requirements are as follows:
1383 <li> <a href="#Specialization">Specialization</a>
1384 <li> <a href="#Performance and Scalability">Performance and Scalability</a>
1385 <li> <a href="#Composability">Composability</a>
1386 <li> <a href="#Corner Cases">Corner Cases</a>
1390 These classes is covered in the following sections.
1392 <h3><a name="Specialization">Specialization</a></h3>
1395 RCU is and always has been intended primarily for read-mostly situations,
1396 which means that RCU's read-side primitives are optimized, often at the
1397 expense of its update-side primitives.
1398 Experience thus far is captured by the following list of situations:
1401 <li> Read-mostly data, where stale and inconsistent data is not
1402 a problem: RCU works great!
1403 <li> Read-mostly data, where data must be consistent:
1405 <li> Read-write data, where data must be consistent:
1406 RCU <i>might</i> work OK.
1408 <li> Write-mostly data, where data must be consistent:
1409 RCU is very unlikely to be the right tool for the job,
1410 with the following exceptions, where RCU can provide:
1412 <li> Existence guarantees for update-friendly mechanisms.
1413 <li> Wait-free read-side primitives for real-time use.
1418 This focus on read-mostly situations means that RCU must interoperate
1419 with other synchronization primitives.
1420 For example, the <tt>add_gp()</tt> and <tt>remove_gp_synchronous()</tt>
1421 examples discussed earlier use RCU to protect readers and locking to
1422 coordinate updaters.
1423 However, the need extends much farther, requiring that a variety of
1424 synchronization primitives be legal within RCU read-side critical sections,
1425 including spinlocks, sequence locks, atomic operations, reference
1426 counters, and memory barriers.
1429 <tr><th> </th></tr>
1430 <tr><th align="left">Quick Quiz:</th></tr>
1432 What about sleeping locks?
1434 <tr><th align="left">Answer:</th></tr>
1435 <tr><td bgcolor="#ffffff"><font color="ffffff">
1436 These are forbidden within Linux-kernel RCU read-side critical
1437 sections because it is not legal to place a quiescent state
1438 (in this case, voluntary context switch) within an RCU read-side
1440 However, sleeping locks may be used within userspace RCU read-side
1441 critical sections, and also within Linux-kernel sleepable RCU
1442 <a href="#Sleepable RCU"><font color="ffffff">(SRCU)</font></a>
1443 read-side critical sections.
1444 In addition, the -rt patchset turns spinlocks into a
1445 sleeping locks so that the corresponding critical sections
1446 can be preempted, which also means that these sleeplockified
1447 spinlocks (but not other sleeping locks!) may be acquire within
1448 -rt-Linux-kernel RCU read-side critical sections.
1451 <p><font color="ffffff">
1452 Note that it <i>is</i> legal for a normal RCU read-side
1453 critical section to conditionally acquire a sleeping locks
1454 (as in <tt>mutex_trylock()</tt>), but only as long as it does
1455 not loop indefinitely attempting to conditionally acquire that
1457 The key point is that things like <tt>mutex_trylock()</tt>
1458 either return with the mutex held, or return an error indication if
1459 the mutex was not immediately available.
1460 Either way, <tt>mutex_trylock()</tt> returns immediately without
1463 <tr><td> </td></tr>
1467 It often comes as a surprise that many algorithms do not require a
1468 consistent view of data, but many can function in that mode,
1469 with network routing being the poster child.
1470 Internet routing algorithms take significant time to propagate
1471 updates, so that by the time an update arrives at a given system,
1472 that system has been sending network traffic the wrong way for
1473 a considerable length of time.
1474 Having a few threads continue to send traffic the wrong way for a
1475 few more milliseconds is clearly not a problem: In the worst case,
1476 TCP retransmissions will eventually get the data where it needs to go.
1477 In general, when tracking the state of the universe outside of the
1478 computer, some level of inconsistency must be tolerated due to
1479 speed-of-light delays if nothing else.
1482 Furthermore, uncertainty about external state is inherent in many cases.
1483 For example, a pair of veterinarians might use heartbeat to determine
1484 whether or not a given cat was alive.
1485 But how long should they wait after the last heartbeat to decide that
1486 the cat is in fact dead?
1487 Waiting less than 400 milliseconds makes no sense because this would
1488 mean that a relaxed cat would be considered to cycle between death
1489 and life more than 100 times per minute.
1490 Moreover, just as with human beings, a cat's heart might stop for
1491 some period of time, so the exact wait period is a judgment call.
1492 One of our pair of veterinarians might wait 30 seconds before pronouncing
1493 the cat dead, while the other might insist on waiting a full minute.
1494 The two veterinarians would then disagree on the state of the cat during
1495 the final 30 seconds of the minute following the last heartbeat.
1498 Interestingly enough, this same situation applies to hardware.
1499 When push comes to shove, how do we tell whether or not some
1500 external server has failed?
1501 We send messages to it periodically, and declare it failed if we
1502 don't receive a response within a given period of time.
1503 Policy decisions can usually tolerate short
1504 periods of inconsistency.
1505 The policy was decided some time ago, and is only now being put into
1506 effect, so a few milliseconds of delay is normally inconsequential.
1509 However, there are algorithms that absolutely must see consistent data.
1510 For example, the translation between a user-level SystemV semaphore
1511 ID to the corresponding in-kernel data structure is protected by RCU,
1512 but it is absolutely forbidden to update a semaphore that has just been
1514 In the Linux kernel, this need for consistency is accommodated by acquiring
1515 spinlocks located in the in-kernel data structure from within
1516 the RCU read-side critical section, and this is indicated by the
1517 green box in the figure above.
1518 Many other techniques may be used, and are in fact used within the
1522 In short, RCU is not required to maintain consistency, and other
1523 mechanisms may be used in concert with RCU when consistency is required.
1524 RCU's specialization allows it to do its job extremely well, and its
1525 ability to interoperate with other synchronization mechanisms allows
1526 the right mix of synchronization tools to be used for a given job.
1528 <h3><a name="Performance and Scalability">Performance and Scalability</a></h3>
1531 Energy efficiency is a critical component of performance today,
1532 and Linux-kernel RCU implementations must therefore avoid unnecessarily
1533 awakening idle CPUs.
1534 I cannot claim that this requirement was premeditated.
1535 In fact, I learned of it during a telephone conversation in which I
1536 was given “frank and open” feedback on the importance
1537 of energy efficiency in battery-powered systems and on specific
1538 energy-efficiency shortcomings of the Linux-kernel RCU implementation.
1539 In my experience, the battery-powered embedded community will consider
1540 any unnecessary wakeups to be extremely unfriendly acts.
1541 So much so that mere Linux-kernel-mailing-list posts are
1542 insufficient to vent their ire.
1545 Memory consumption is not particularly important for in most
1546 situations, and has become decreasingly
1547 so as memory sizes have expanded and memory
1548 costs have plummeted.
1549 However, as I learned from Matt Mackall's
1550 <a href="http://elinux.org/Linux_Tiny-FAQ">bloatwatch</a>
1551 efforts, memory footprint is critically important on single-CPU systems with
1552 non-preemptible (<tt>CONFIG_PREEMPT=n</tt>) kernels, and thus
1553 <a href="https://lkml.kernel.org/g/20090113221724.GA15307@linux.vnet.ibm.com">tiny RCU</a>
1555 Josh Triplett has since taken over the small-memory banner with his
1556 <a href="https://tiny.wiki.kernel.org/">Linux kernel tinification</a>
1557 project, which resulted in
1558 <a href="#Sleepable RCU">SRCU</a>
1559 becoming optional for those kernels not needing it.
1562 The remaining performance requirements are, for the most part,
1564 For example, in keeping with RCU's read-side specialization,
1565 <tt>rcu_dereference()</tt> should have negligible overhead (for
1566 example, suppression of a few minor compiler optimizations).
1567 Similarly, in non-preemptible environments, <tt>rcu_read_lock()</tt> and
1568 <tt>rcu_read_unlock()</tt> should have exactly zero overhead.
1571 In preemptible environments, in the case where the RCU read-side
1572 critical section was not preempted (as will be the case for the
1573 highest-priority real-time process), <tt>rcu_read_lock()</tt> and
1574 <tt>rcu_read_unlock()</tt> should have minimal overhead.
1575 In particular, they should not contain atomic read-modify-write
1576 operations, memory-barrier instructions, preemption disabling,
1577 interrupt disabling, or backwards branches.
1578 However, in the case where the RCU read-side critical section was preempted,
1579 <tt>rcu_read_unlock()</tt> may acquire spinlocks and disable interrupts.
1580 This is why it is better to nest an RCU read-side critical section
1581 within a preempt-disable region than vice versa, at least in cases
1582 where that critical section is short enough to avoid unduly degrading
1583 real-time latencies.
1586 The <tt>synchronize_rcu()</tt> grace-period-wait primitive is
1587 optimized for throughput.
1588 It may therefore incur several milliseconds of latency in addition to
1589 the duration of the longest RCU read-side critical section.
1590 On the other hand, multiple concurrent invocations of
1591 <tt>synchronize_rcu()</tt> are required to use batching optimizations
1592 so that they can be satisfied by a single underlying grace-period-wait
1594 For example, in the Linux kernel, it is not unusual for a single
1595 grace-period-wait operation to serve more than
1596 <a href="https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response">1,000 separate invocations</a>
1597 of <tt>synchronize_rcu()</tt>, thus amortizing the per-invocation
1598 overhead down to nearly zero.
1599 However, the grace-period optimization is also required to avoid
1600 measurable degradation of real-time scheduling and interrupt latencies.
1603 In some cases, the multi-millisecond <tt>synchronize_rcu()</tt>
1604 latencies are unacceptable.
1605 In these cases, <tt>synchronize_rcu_expedited()</tt> may be used
1606 instead, reducing the grace-period latency down to a few tens of
1607 microseconds on small systems, at least in cases where the RCU read-side
1608 critical sections are short.
1609 There are currently no special latency requirements for
1610 <tt>synchronize_rcu_expedited()</tt> on large systems, but,
1611 consistent with the empirical nature of the RCU specification,
1612 that is subject to change.
1613 However, there most definitely are scalability requirements:
1614 A storm of <tt>synchronize_rcu_expedited()</tt> invocations on 4096
1615 CPUs should at least make reasonable forward progress.
1616 In return for its shorter latencies, <tt>synchronize_rcu_expedited()</tt>
1617 is permitted to impose modest degradation of real-time latency
1618 on non-idle online CPUs.
1619 Here, “modest” means roughly the same latency
1620 degradation as a scheduling-clock interrupt.
1623 There are a number of situations where even
1624 <tt>synchronize_rcu_expedited()</tt>'s reduced grace-period
1625 latency is unacceptable.
1626 In these situations, the asynchronous <tt>call_rcu()</tt> can be
1627 used in place of <tt>synchronize_rcu()</tt> as follows:
1634 4 struct rcu_head rh;
1637 7 static void remove_gp_cb(struct rcu_head *rhp)
1639 9 struct foo *p = container_of(rhp, struct foo, rh);
1644 14 bool remove_gp_asynchronous(void)
1648 18 spin_lock(&gp_lock);
1649 19 p = rcu_dereference(gp);
1651 21 spin_unlock(&gp_lock);
1654 24 rcu_assign_pointer(gp, NULL);
1655 25 call_rcu(&p->rh, remove_gp_cb);
1656 26 spin_unlock(&gp_lock);
1663 A definition of <tt>struct foo</tt> is finally needed, and appears
1665 The function <tt>remove_gp_cb()</tt> is passed to <tt>call_rcu()</tt>
1666 on line 25, and will be invoked after the end of a subsequent
1668 This gets the same effect as <tt>remove_gp_synchronous()</tt>,
1669 but without forcing the updater to wait for a grace period to elapse.
1670 The <tt>call_rcu()</tt> function may be used in a number of
1671 situations where neither <tt>synchronize_rcu()</tt> nor
1672 <tt>synchronize_rcu_expedited()</tt> would be legal,
1673 including within preempt-disable code, <tt>local_bh_disable()</tt> code,
1674 interrupt-disable code, and interrupt handlers.
1675 However, even <tt>call_rcu()</tt> is illegal within NMI handlers
1676 and from idle and offline CPUs.
1677 The callback function (<tt>remove_gp_cb()</tt> in this case) will be
1678 executed within softirq (software interrupt) environment within the
1680 either within a real softirq handler or under the protection
1681 of <tt>local_bh_disable()</tt>.
1682 In both the Linux kernel and in userspace, it is bad practice to
1683 write an RCU callback function that takes too long.
1684 Long-running operations should be relegated to separate threads or
1685 (in the Linux kernel) workqueues.
1688 <tr><th> </th></tr>
1689 <tr><th align="left">Quick Quiz:</th></tr>
1691 Why does line 19 use <tt>rcu_access_pointer()</tt>?
1692 After all, <tt>call_rcu()</tt> on line 25 stores into the
1693 structure, which would interact badly with concurrent insertions.
1694 Doesn't this mean that <tt>rcu_dereference()</tt> is required?
1696 <tr><th align="left">Answer:</th></tr>
1697 <tr><td bgcolor="#ffffff"><font color="ffffff">
1698 Presumably the <tt>->gp_lock</tt> acquired on line 18 excludes
1699 any changes, including any insertions that <tt>rcu_dereference()</tt>
1700 would protect against.
1701 Therefore, any insertions will be delayed until after
1702 <tt>->gp_lock</tt>
1703 is released on line 25, which in turn means that
1704 <tt>rcu_access_pointer()</tt> suffices.
1706 <tr><td> </td></tr>
1710 However, all that <tt>remove_gp_cb()</tt> is doing is
1711 invoking <tt>kfree()</tt> on the data element.
1712 This is a common idiom, and is supported by <tt>kfree_rcu()</tt>,
1713 which allows “fire and forget” operation as shown below:
1720 4 struct rcu_head rh;
1723 7 bool remove_gp_faf(void)
1727 11 spin_lock(&gp_lock);
1728 12 p = rcu_dereference(gp);
1730 14 spin_unlock(&gp_lock);
1733 17 rcu_assign_pointer(gp, NULL);
1734 18 kfree_rcu(p, rh);
1735 19 spin_unlock(&gp_lock);
1742 Note that <tt>remove_gp_faf()</tt> simply invokes
1743 <tt>kfree_rcu()</tt> and proceeds, without any need to pay any
1744 further attention to the subsequent grace period and <tt>kfree()</tt>.
1745 It is permissible to invoke <tt>kfree_rcu()</tt> from the same
1746 environments as for <tt>call_rcu()</tt>.
1747 Interestingly enough, DYNIX/ptx had the equivalents of
1748 <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>, but not
1749 <tt>synchronize_rcu()</tt>.
1750 This was due to the fact that RCU was not heavily used within DYNIX/ptx,
1751 so the very few places that needed something like
1752 <tt>synchronize_rcu()</tt> simply open-coded it.
1755 <tr><th> </th></tr>
1756 <tr><th align="left">Quick Quiz:</th></tr>
1758 Earlier it was claimed that <tt>call_rcu()</tt> and
1759 <tt>kfree_rcu()</tt> allowed updaters to avoid being blocked
1761 But how can that be correct, given that the invocation of the callback
1762 and the freeing of the memory (respectively) must still wait for
1763 a grace period to elapse?
1765 <tr><th align="left">Answer:</th></tr>
1766 <tr><td bgcolor="#ffffff"><font color="ffffff">
1767 We could define things this way, but keep in mind that this sort of
1768 definition would say that updates in garbage-collected languages
1769 cannot complete until the next time the garbage collector runs,
1770 which does not seem at all reasonable.
1771 The key point is that in most cases, an updater using either
1772 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> can proceed to the
1773 next update as soon as it has invoked <tt>call_rcu()</tt> or
1774 <tt>kfree_rcu()</tt>, without having to wait for a subsequent
1777 <tr><td> </td></tr>
1781 But what if the updater must wait for the completion of code to be
1782 executed after the end of the grace period, but has other tasks
1783 that can be carried out in the meantime?
1784 The polling-style <tt>get_state_synchronize_rcu()</tt> and
1785 <tt>cond_synchronize_rcu()</tt> functions may be used for this
1786 purpose, as shown below:
1790 1 bool remove_gp_poll(void)
1795 6 spin_lock(&gp_lock);
1796 7 p = rcu_access_pointer(gp);
1798 9 spin_unlock(&gp_lock);
1801 12 rcu_assign_pointer(gp, NULL);
1802 13 spin_unlock(&gp_lock);
1803 14 s = get_state_synchronize_rcu();
1804 15 do_something_while_waiting();
1805 16 cond_synchronize_rcu(s);
1813 On line 14, <tt>get_state_synchronize_rcu()</tt> obtains a
1814 “cookie” from RCU,
1815 then line 15 carries out other tasks,
1816 and finally, line 16 returns immediately if a grace period has
1817 elapsed in the meantime, but otherwise waits as required.
1818 The need for <tt>get_state_synchronize_rcu</tt> and
1819 <tt>cond_synchronize_rcu()</tt> has appeared quite recently,
1820 so it is too early to tell whether they will stand the test of time.
1823 RCU thus provides a range of tools to allow updaters to strike the
1824 required tradeoff between latency, flexibility and CPU overhead.
1826 <h3><a name="Composability">Composability</a></h3>
1829 Composability has received much attention in recent years, perhaps in part
1830 due to the collision of multicore hardware with object-oriented techniques
1831 designed in single-threaded environments for single-threaded use.
1832 And in theory, RCU read-side critical sections may be composed, and in
1833 fact may be nested arbitrarily deeply.
1834 In practice, as with all real-world implementations of composable
1835 constructs, there are limitations.
1838 Implementations of RCU for which <tt>rcu_read_lock()</tt>
1839 and <tt>rcu_read_unlock()</tt> generate no code, such as
1840 Linux-kernel RCU when <tt>CONFIG_PREEMPT=n</tt>, can be
1841 nested arbitrarily deeply.
1842 After all, there is no overhead.
1843 Except that if all these instances of <tt>rcu_read_lock()</tt>
1844 and <tt>rcu_read_unlock()</tt> are visible to the compiler,
1845 compilation will eventually fail due to exhausting memory,
1846 mass storage, or user patience, whichever comes first.
1847 If the nesting is not visible to the compiler, as is the case with
1848 mutually recursive functions each in its own translation unit,
1849 stack overflow will result.
1850 If the nesting takes the form of loops, perhaps in the guise of tail
1851 recursion, either the control variable
1852 will overflow or (in the Linux kernel) you will get an RCU CPU stall warning.
1853 Nevertheless, this class of RCU implementations is one
1854 of the most composable constructs in existence.
1857 RCU implementations that explicitly track nesting depth
1858 are limited by the nesting-depth counter.
1859 For example, the Linux kernel's preemptible RCU limits nesting to
1861 This should suffice for almost all practical purposes.
1862 That said, a consecutive pair of RCU read-side critical sections
1863 between which there is an operation that waits for a grace period
1864 cannot be enclosed in another RCU read-side critical section.
1865 This is because it is not legal to wait for a grace period within
1866 an RCU read-side critical section: To do so would result either
1868 in RCU implicitly splitting the enclosing RCU read-side critical
1869 section, neither of which is conducive to a long-lived and prosperous
1873 It is worth noting that RCU is not alone in limiting composability.
1874 For example, many transactional-memory implementations prohibit
1875 composing a pair of transactions separated by an irrevocable
1876 operation (for example, a network receive operation).
1877 For another example, lock-based critical sections can be composed
1878 surprisingly freely, but only if deadlock is avoided.
1881 In short, although RCU read-side critical sections are highly composable,
1882 care is required in some situations, just as is the case for any other
1883 composable synchronization mechanism.
1885 <h3><a name="Corner Cases">Corner Cases</a></h3>
1888 A given RCU workload might have an endless and intense stream of
1889 RCU read-side critical sections, perhaps even so intense that there
1890 was never a point in time during which there was not at least one
1891 RCU read-side critical section in flight.
1892 RCU cannot allow this situation to block grace periods: As long as
1893 all the RCU read-side critical sections are finite, grace periods
1894 must also be finite.
1897 That said, preemptible RCU implementations could potentially result
1898 in RCU read-side critical sections being preempted for long durations,
1899 which has the effect of creating a long-duration RCU read-side
1901 This situation can arise only in heavily loaded systems, but systems using
1902 real-time priorities are of course more vulnerable.
1903 Therefore, RCU priority boosting is provided to help deal with this
1905 That said, the exact requirements on RCU priority boosting will likely
1906 evolve as more experience accumulates.
1909 Other workloads might have very high update rates.
1910 Although one can argue that such workloads should instead use
1911 something other than RCU, the fact remains that RCU must
1912 handle such workloads gracefully.
1913 This requirement is another factor driving batching of grace periods,
1914 but it is also the driving force behind the checks for large numbers
1915 of queued RCU callbacks in the <tt>call_rcu()</tt> code path.
1916 Finally, high update rates should not delay RCU read-side critical
1917 sections, although some small read-side delays can occur when using
1918 <tt>synchronize_rcu_expedited()</tt>, courtesy of this function's use
1919 of <tt>smp_call_function_single()</tt>.
1922 Although all three of these corner cases were understood in the early
1923 1990s, a simple user-level test consisting of <tt>close(open(path))</tt>
1925 in the early 2000s suddenly provided a much deeper appreciation of the
1926 high-update-rate corner case.
1927 This test also motivated addition of some RCU code to react to high update
1928 rates, for example, if a given CPU finds itself with more than 10,000
1929 RCU callbacks queued, it will cause RCU to take evasive action by
1930 more aggressively starting grace periods and more aggressively forcing
1931 completion of grace-period processing.
1932 This evasive action causes the grace period to complete more quickly,
1933 but at the cost of restricting RCU's batching optimizations, thus
1934 increasing the CPU overhead incurred by that grace period.
1936 <h2><a name="Software-Engineering Requirements">
1937 Software-Engineering Requirements</a></h2>
1940 Between Murphy's Law and “To err is human”, it is necessary to
1941 guard against mishaps and misuse:
1944 <li> It is all too easy to forget to use <tt>rcu_read_lock()</tt>
1945 everywhere that it is needed, so kernels built with
1946 <tt>CONFIG_PROVE_RCU=y</tt> will splat if
1947 <tt>rcu_dereference()</tt> is used outside of an
1948 RCU read-side critical section.
1949 Update-side code can use <tt>rcu_dereference_protected()</tt>,
1951 <a href="https://lwn.net/Articles/371986/">lockdep expression</a>
1952 to indicate what is providing the protection.
1953 If the indicated protection is not provided, a lockdep splat
1957 Code shared between readers and updaters can use
1958 <tt>rcu_dereference_check()</tt>, which also takes a
1959 lockdep expression, and emits a lockdep splat if neither
1960 <tt>rcu_read_lock()</tt> nor the indicated protection
1962 In addition, <tt>rcu_dereference_raw()</tt> is used in those
1963 (hopefully rare) cases where the required protection cannot
1964 be easily described.
1965 Finally, <tt>rcu_read_lock_held()</tt> is provided to
1966 allow a function to verify that it has been invoked within
1967 an RCU read-side critical section.
1968 I was made aware of this set of requirements shortly after Thomas
1969 Gleixner audited a number of RCU uses.
1970 <li> A given function might wish to check for RCU-related preconditions
1971 upon entry, before using any other RCU API.
1972 The <tt>rcu_lockdep_assert()</tt> does this job,
1973 asserting the expression in kernels having lockdep enabled
1974 and doing nothing otherwise.
1975 <li> It is also easy to forget to use <tt>rcu_assign_pointer()</tt>
1976 and <tt>rcu_dereference()</tt>, perhaps (incorrectly)
1977 substituting a simple assignment.
1978 To catch this sort of error, a given RCU-protected pointer may be
1979 tagged with <tt>__rcu</tt>, after which sparse
1980 will complain about simple-assignment accesses to that pointer.
1981 Arnd Bergmann made me aware of this requirement, and also
1983 <a href="https://lwn.net/Articles/376011/">patch series</a>.
1984 <li> Kernels built with <tt>CONFIG_DEBUG_OBJECTS_RCU_HEAD=y</tt>
1985 will splat if a data element is passed to <tt>call_rcu()</tt>
1986 twice in a row, without a grace period in between.
1987 (This error is similar to a double free.)
1988 The corresponding <tt>rcu_head</tt> structures that are
1989 dynamically allocated are automatically tracked, but
1990 <tt>rcu_head</tt> structures allocated on the stack
1991 must be initialized with <tt>init_rcu_head_on_stack()</tt>
1992 and cleaned up with <tt>destroy_rcu_head_on_stack()</tt>.
1993 Similarly, statically allocated non-stack <tt>rcu_head</tt>
1994 structures must be initialized with <tt>init_rcu_head()</tt>
1995 and cleaned up with <tt>destroy_rcu_head()</tt>.
1996 Mathieu Desnoyers made me aware of this requirement, and also
1998 <a href="https://lkml.kernel.org/g/20100319013024.GA28456@Krystal">patch</a>.
1999 <li> An infinite loop in an RCU read-side critical section will
2000 eventually trigger an RCU CPU stall warning splat, with
2001 the duration of “eventually” being controlled by the
2002 <tt>RCU_CPU_STALL_TIMEOUT</tt> <tt>Kconfig</tt> option, or,
2003 alternatively, by the
2004 <tt>rcupdate.rcu_cpu_stall_timeout</tt> boot/sysfs
2006 However, RCU is not obligated to produce this splat
2007 unless there is a grace period waiting on that particular
2008 RCU read-side critical section.
2010 Some extreme workloads might intentionally delay
2011 RCU grace periods, and systems running those workloads can
2012 be booted with <tt>rcupdate.rcu_cpu_stall_suppress</tt>
2013 to suppress the splats.
2014 This kernel parameter may also be set via <tt>sysfs</tt>.
2015 Furthermore, RCU CPU stall warnings are counter-productive
2016 during sysrq dumps and during panics.
2017 RCU therefore supplies the <tt>rcu_sysrq_start()</tt> and
2018 <tt>rcu_sysrq_end()</tt> API members to be called before
2019 and after long sysrq dumps.
2020 RCU also supplies the <tt>rcu_panic()</tt> notifier that is
2021 automatically invoked at the beginning of a panic to suppress
2022 further RCU CPU stall warnings.
2025 This requirement made itself known in the early 1990s, pretty
2026 much the first time that it was necessary to debug a CPU stall.
2027 That said, the initial implementation in DYNIX/ptx was quite
2028 generic in comparison with that of Linux.
2029 <li> Although it would be very good to detect pointers leaking out
2030 of RCU read-side critical sections, there is currently no
2031 good way of doing this.
2032 One complication is the need to distinguish between pointers
2033 leaking and pointers that have been handed off from RCU to
2034 some other synchronization mechanism, for example, reference
2036 <li> In kernels built with <tt>CONFIG_RCU_TRACE=y</tt>, RCU-related
2037 information is provided via event tracing.
2038 <li> Open-coded use of <tt>rcu_assign_pointer()</tt> and
2039 <tt>rcu_dereference()</tt> to create typical linked
2040 data structures can be surprisingly error-prone.
2041 Therefore, RCU-protected
2042 <a href="https://lwn.net/Articles/609973/#RCU List APIs">linked lists</a>
2043 and, more recently, RCU-protected
2044 <a href="https://lwn.net/Articles/612100/">hash tables</a>
2046 Many other special-purpose RCU-protected data structures are
2047 available in the Linux kernel and the userspace RCU library.
2048 <li> Some linked structures are created at compile time, but still
2049 require <tt>__rcu</tt> checking.
2050 The <tt>RCU_POINTER_INITIALIZER()</tt> macro serves this
2052 <li> It is not necessary to use <tt>rcu_assign_pointer()</tt>
2053 when creating linked structures that are to be published via
2054 a single external pointer.
2055 The <tt>RCU_INIT_POINTER()</tt> macro is provided for
2056 this task and also for assigning <tt>NULL</tt> pointers
2061 This not a hard-and-fast list: RCU's diagnostic capabilities will
2062 continue to be guided by the number and type of usage bugs found
2063 in real-world RCU usage.
2065 <h2><a name="Linux Kernel Complications">Linux Kernel Complications</a></h2>
2068 The Linux kernel provides an interesting environment for all kinds of
2069 software, including RCU.
2070 Some of the relevant points of interest are as follows:
2073 <li> <a href="#Configuration">Configuration</a>.
2074 <li> <a href="#Firmware Interface">Firmware Interface</a>.
2075 <li> <a href="#Early Boot">Early Boot</a>.
2076 <li> <a href="#Interrupts and NMIs">
2077 Interrupts and non-maskable interrupts (NMIs)</a>.
2078 <li> <a href="#Loadable Modules">Loadable Modules</a>.
2079 <li> <a href="#Hotplug CPU">Hotplug CPU</a>.
2080 <li> <a href="#Scheduler and RCU">Scheduler and RCU</a>.
2081 <li> <a href="#Tracing and RCU">Tracing and RCU</a>.
2082 <li> <a href="#Energy Efficiency">Energy Efficiency</a>.
2083 <li> <a href="#Memory Efficiency">Memory Efficiency</a>.
2084 <li> <a href="#Performance, Scalability, Response Time, and Reliability">
2085 Performance, Scalability, Response Time, and Reliability</a>.
2089 This list is probably incomplete, but it does give a feel for the
2090 most notable Linux-kernel complications.
2091 Each of the following sections covers one of the above topics.
2093 <h3><a name="Configuration">Configuration</a></h3>
2096 RCU's goal is automatic configuration, so that almost nobody
2097 needs to worry about RCU's <tt>Kconfig</tt> options.
2098 And for almost all users, RCU does in fact work well
2099 “out of the box.”
2102 However, there are specialized use cases that are handled by
2103 kernel boot parameters and <tt>Kconfig</tt> options.
2104 Unfortunately, the <tt>Kconfig</tt> system will explicitly ask users
2105 about new <tt>Kconfig</tt> options, which requires almost all of them
2106 be hidden behind a <tt>CONFIG_RCU_EXPERT</tt> <tt>Kconfig</tt> option.
2109 This all should be quite obvious, but the fact remains that
2110 Linus Torvalds recently had to
2111 <a href="https://lkml.kernel.org/g/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com">remind</a>
2112 me of this requirement.
2114 <h3><a name="Firmware Interface">Firmware Interface</a></h3>
2117 In many cases, kernel obtains information about the system from the
2118 firmware, and sometimes things are lost in translation.
2119 Or the translation is accurate, but the original message is bogus.
2122 For example, some systems' firmware overreports the number of CPUs,
2123 sometimes by a large factor.
2124 If RCU naively believed the firmware, as it used to do,
2125 it would create too many per-CPU kthreads.
2126 Although the resulting system will still run correctly, the extra
2127 kthreads needlessly consume memory and can cause confusion
2128 when they show up in <tt>ps</tt> listings.
2131 RCU must therefore wait for a given CPU to actually come online before
2132 it can allow itself to believe that the CPU actually exists.
2133 The resulting “ghost CPUs” (which are never going to
2134 come online) cause a number of
2135 <a href="https://paulmck.livejournal.com/37494.html">interesting complications</a>.
2137 <h3><a name="Early Boot">Early Boot</a></h3>
2140 The Linux kernel's boot sequence is an interesting process,
2141 and RCU is used early, even before <tt>rcu_init()</tt>
2143 In fact, a number of RCU's primitives can be used as soon as the
2144 initial task's <tt>task_struct</tt> is available and the
2145 boot CPU's per-CPU variables are set up.
2146 The read-side primitives (<tt>rcu_read_lock()</tt>,
2147 <tt>rcu_read_unlock()</tt>, <tt>rcu_dereference()</tt>,
2148 and <tt>rcu_access_pointer()</tt>) will operate normally very early on,
2149 as will <tt>rcu_assign_pointer()</tt>.
2152 Although <tt>call_rcu()</tt> may be invoked at any
2153 time during boot, callbacks are not guaranteed to be invoked until after
2154 all of RCU's kthreads have been spawned, which occurs at
2155 <tt>early_initcall()</tt> time.
2156 This delay in callback invocation is due to the fact that RCU does not
2157 invoke callbacks until it is fully initialized, and this full initialization
2158 cannot occur until after the scheduler has initialized itself to the
2159 point where RCU can spawn and run its kthreads.
2160 In theory, it would be possible to invoke callbacks earlier,
2161 however, this is not a panacea because there would be severe restrictions
2162 on what operations those callbacks could invoke.
2165 Perhaps surprisingly, <tt>synchronize_rcu()</tt>,
2166 <a href="#Bottom-Half Flavor"><tt>synchronize_rcu_bh()</tt></a>
2167 (<a href="#Bottom-Half Flavor">discussed below</a>),
2168 <a href="#Sched Flavor"><tt>synchronize_sched()</tt></a>,
2169 <tt>synchronize_rcu_expedited()</tt>,
2170 <tt>synchronize_rcu_bh_expedited()</tt>, and
2171 <tt>synchronize_sched_expedited()</tt>
2172 will all operate normally
2173 during very early boot, the reason being that there is only one CPU
2174 and preemption is disabled.
2175 This means that the call <tt>synchronize_rcu()</tt> (or friends)
2176 itself is a quiescent
2177 state and thus a grace period, so the early-boot implementation can
2181 However, once the scheduler has spawned its first kthread, this early
2182 boot trick fails for <tt>synchronize_rcu()</tt> (as well as for
2183 <tt>synchronize_rcu_expedited()</tt>) in <tt>CONFIG_PREEMPT=y</tt>
2185 The reason is that an RCU read-side critical section might be preempted,
2186 which means that a subsequent <tt>synchronize_rcu()</tt> really does have
2187 to wait for something, as opposed to simply returning immediately.
2188 Unfortunately, <tt>synchronize_rcu()</tt> can't do this until all of
2189 its kthreads are spawned, which doesn't happen until some time during
2190 <tt>early_initcalls()</tt> time.
2191 But this is no excuse: RCU is nevertheless required to correctly handle
2192 synchronous grace periods during this time period.
2193 Once all of its kthreads are up and running, RCU starts running
2197 <tr><th> </th></tr>
2198 <tr><th align="left">Quick Quiz:</th></tr>
2200 How can RCU possibly handle grace periods before all of its
2201 kthreads have been spawned???
2203 <tr><th align="left">Answer:</th></tr>
2204 <tr><td bgcolor="#ffffff"><font color="ffffff">
2208 <p><font color="ffffff">
2209 During the “dead zone” between the time that the
2210 scheduler spawns the first task and the time that all of RCU's
2211 kthreads have been spawned, all synchronous grace periods are
2212 handled by the expedited grace-period mechanism.
2213 At runtime, this expedited mechanism relies on workqueues, but
2214 during the dead zone the requesting task itself drives the
2215 desired expedited grace period.
2216 Because dead-zone execution takes place within task context,
2218 Once the dead zone ends, expedited grace periods go back to
2219 using workqueues, as is required to avoid problems that would
2220 otherwise occur when a user task received a POSIX signal while
2221 driving an expedited grace period.
2224 <p><font color="ffffff">
2225 And yes, this does mean that it is unhelpful to send POSIX
2226 signals to random tasks between the time that the scheduler
2227 spawns its first kthread and the time that RCU's kthreads
2228 have all been spawned.
2229 If there ever turns out to be a good reason for sending POSIX
2230 signals during that time, appropriate adjustments will be made.
2231 (If it turns out that POSIX signals are sent during this time for
2232 no good reason, other adjustments will be made, appropriate
2235 <tr><td> </td></tr>
2239 I learned of these boot-time requirements as a result of a series of
2242 <h3><a name="Interrupts and NMIs">Interrupts and NMIs</a></h3>
2245 The Linux kernel has interrupts, and RCU read-side critical sections are
2246 legal within interrupt handlers and within interrupt-disabled regions
2247 of code, as are invocations of <tt>call_rcu()</tt>.
2250 Some Linux-kernel architectures can enter an interrupt handler from
2251 non-idle process context, and then just never leave it, instead stealthily
2252 transitioning back to process context.
2253 This trick is sometimes used to invoke system calls from inside the kernel.
2254 These “half-interrupts” mean that RCU has to be very careful
2255 about how it counts interrupt nesting levels.
2256 I learned of this requirement the hard way during a rewrite
2257 of RCU's dyntick-idle code.
2260 The Linux kernel has non-maskable interrupts (NMIs), and
2261 RCU read-side critical sections are legal within NMI handlers.
2262 Thankfully, RCU update-side primitives, including
2263 <tt>call_rcu()</tt>, are prohibited within NMI handlers.
2266 The name notwithstanding, some Linux-kernel architectures
2267 can have nested NMIs, which RCU must handle correctly.
2269 <a href="https://lkml.kernel.org/g/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com">surprised me</a>
2270 with this requirement;
2271 he also kindly surprised me with
2272 <a href="https://lkml.kernel.org/g/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com">an algorithm</a>
2273 that meets this requirement.
2275 <h3><a name="Loadable Modules">Loadable Modules</a></h3>
2278 The Linux kernel has loadable modules, and these modules can
2280 After a given module has been unloaded, any attempt to call
2281 one of its functions results in a segmentation fault.
2282 The module-unload functions must therefore cancel any
2283 delayed calls to loadable-module functions, for example,
2284 any outstanding <tt>mod_timer()</tt> must be dealt with
2285 via <tt>del_timer_sync()</tt> or similar.
2288 Unfortunately, there is no way to cancel an RCU callback;
2289 once you invoke <tt>call_rcu()</tt>, the callback function is
2290 going to eventually be invoked, unless the system goes down first.
2291 Because it is normally considered socially irresponsible to crash the system
2292 in response to a module unload request, we need some other way
2293 to deal with in-flight RCU callbacks.
2296 RCU therefore provides
2297 <tt><a href="https://lwn.net/Articles/217484/">rcu_barrier()</a></tt>,
2298 which waits until all in-flight RCU callbacks have been invoked.
2299 If a module uses <tt>call_rcu()</tt>, its exit function should therefore
2300 prevent any future invocation of <tt>call_rcu()</tt>, then invoke
2301 <tt>rcu_barrier()</tt>.
2302 In theory, the underlying module-unload code could invoke
2303 <tt>rcu_barrier()</tt> unconditionally, but in practice this would
2304 incur unacceptable latencies.
2307 Nikita Danilov noted this requirement for an analogous filesystem-unmount
2308 situation, and Dipankar Sarma incorporated <tt>rcu_barrier()</tt> into RCU.
2309 The need for <tt>rcu_barrier()</tt> for module unloading became
2313 <b>Important note</b>: The <tt>rcu_barrier()</tt> function is not,
2314 repeat, <i>not</i>, obligated to wait for a grace period.
2315 It is instead only required to wait for RCU callbacks that have
2316 already been posted.
2317 Therefore, if there are no RCU callbacks posted anywhere in the system,
2318 <tt>rcu_barrier()</tt> is within its rights to return immediately.
2319 Even if there are callbacks posted, <tt>rcu_barrier()</tt> does not
2320 necessarily need to wait for a grace period.
2323 <tr><th> </th></tr>
2324 <tr><th align="left">Quick Quiz:</th></tr>
2327 Each RCU callbacks must wait for a grace period to complete,
2328 and <tt>rcu_barrier()</tt> must wait for each pre-existing
2329 callback to be invoked.
2330 Doesn't <tt>rcu_barrier()</tt> therefore need to wait for
2331 a full grace period if there is even one callback posted anywhere
2334 <tr><th align="left">Answer:</th></tr>
2335 <tr><td bgcolor="#ffffff"><font color="ffffff">
2339 <p><font color="ffffff">
2340 Yes, each RCU callbacks must wait for a grace period to complete,
2341 but it might well be partly (or even completely) finished waiting
2342 by the time <tt>rcu_barrier()</tt> is invoked.
2343 In that case, <tt>rcu_barrier()</tt> need only wait for the
2344 remaining portion of the grace period to elapse.
2345 So even if there are quite a few callbacks posted,
2346 <tt>rcu_barrier()</tt> might well return quite quickly.
2349 <p><font color="ffffff">
2350 So if you need to wait for a grace period as well as for all
2351 pre-existing callbacks, you will need to invoke both
2352 <tt>synchronize_rcu()</tt> and <tt>rcu_barrier()</tt>.
2353 If latency is a concern, you can always use workqueues
2354 to invoke them concurrently.
2356 <tr><td> </td></tr>
2359 <h3><a name="Hotplug CPU">Hotplug CPU</a></h3>
2362 The Linux kernel supports CPU hotplug, which means that CPUs
2364 It is of course illegal to use any RCU API member from an offline CPU,
2365 with the exception of <a href="#Sleepable RCU">SRCU</a> read-side
2367 This requirement was present from day one in DYNIX/ptx, but
2368 on the other hand, the Linux kernel's CPU-hotplug implementation
2369 is “interesting.”
2372 The Linux-kernel CPU-hotplug implementation has notifiers that
2373 are used to allow the various kernel subsystems (including RCU)
2374 to respond appropriately to a given CPU-hotplug operation.
2375 Most RCU operations may be invoked from CPU-hotplug notifiers,
2376 including even synchronous grace-period operations such as
2377 <tt>synchronize_rcu()</tt> and <tt>synchronize_rcu_expedited()</tt>.
2380 However, all-callback-wait operations such as
2381 <tt>rcu_barrier()</tt> are also not supported, due to the
2382 fact that there are phases of CPU-hotplug operations where
2383 the outgoing CPU's callbacks will not be invoked until after
2384 the CPU-hotplug operation ends, which could also result in deadlock.
2385 Furthermore, <tt>rcu_barrier()</tt> blocks CPU-hotplug operations
2386 during its execution, which results in another type of deadlock
2387 when invoked from a CPU-hotplug notifier.
2389 <h3><a name="Scheduler and RCU">Scheduler and RCU</a></h3>
2392 RCU depends on the scheduler, and the scheduler uses RCU to
2393 protect some of its data structures.
2394 This means the scheduler is forbidden from acquiring
2395 the runqueue locks and the priority-inheritance locks
2396 in the middle of an outermost RCU read-side critical section unless either
2397 (1) it releases them before exiting that same
2398 RCU read-side critical section, or
2399 (2) interrupts are disabled across
2400 that entire RCU read-side critical section.
2401 This same prohibition also applies (recursively!) to any lock that is acquired
2402 while holding any lock to which this prohibition applies.
2403 Adhering to this rule prevents preemptible RCU from invoking
2404 <tt>rcu_read_unlock_special()</tt> while either runqueue or
2405 priority-inheritance locks are held, thus avoiding deadlock.
2408 Prior to v4.4, it was only necessary to disable preemption across
2409 RCU read-side critical sections that acquired scheduler locks.
2410 In v4.4, expedited grace periods started using IPIs, and these
2411 IPIs could force a <tt>rcu_read_unlock()</tt> to take the slowpath.
2412 Therefore, this expedited-grace-period change required disabling of
2413 interrupts, not just preemption.
2416 For RCU's part, the preemptible-RCU <tt>rcu_read_unlock()</tt>
2417 implementation must be written carefully to avoid similar deadlocks.
2418 In particular, <tt>rcu_read_unlock()</tt> must tolerate an
2419 interrupt where the interrupt handler invokes both
2420 <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
2421 This possibility requires <tt>rcu_read_unlock()</tt> to use
2422 negative nesting levels to avoid destructive recursion via
2423 interrupt handler's use of RCU.
2426 This pair of mutual scheduler-RCU requirements came as a
2427 <a href="https://lwn.net/Articles/453002/">complete surprise</a>.
2430 As noted above, RCU makes use of kthreads, and it is necessary to
2431 avoid excessive CPU-time accumulation by these kthreads.
2432 This requirement was no surprise, but RCU's violation of it
2433 when running context-switch-heavy workloads when built with
2434 <tt>CONFIG_NO_HZ_FULL=y</tt>
2435 <a href="http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf">did come as a surprise [PDF]</a>.
2436 RCU has made good progress towards meeting this requirement, even
2437 for context-switch-have <tt>CONFIG_NO_HZ_FULL=y</tt> workloads,
2438 but there is room for further improvement.
2440 <h3><a name="Tracing and RCU">Tracing and RCU</a></h3>
2443 It is possible to use tracing on RCU code, but tracing itself
2445 For this reason, <tt>rcu_dereference_raw_notrace()</tt>
2446 is provided for use by tracing, which avoids the destructive
2447 recursion that could otherwise ensue.
2448 This API is also used by virtualization in some architectures,
2449 where RCU readers execute in environments in which tracing
2451 The tracing folks both located the requirement and provided the
2452 needed fix, so this surprise requirement was relatively painless.
2454 <h3><a name="Energy Efficiency">Energy Efficiency</a></h3>
2457 Interrupting idle CPUs is considered socially unacceptable,
2458 especially by people with battery-powered embedded systems.
2459 RCU therefore conserves energy by detecting which CPUs are
2460 idle, including tracking CPUs that have been interrupted from idle.
2461 This is a large part of the energy-efficiency requirement,
2462 so I learned of this via an irate phone call.
2465 Because RCU avoids interrupting idle CPUs, it is illegal to
2466 execute an RCU read-side critical section on an idle CPU.
2467 (Kernels built with <tt>CONFIG_PROVE_RCU=y</tt> will splat
2469 The <tt>RCU_NONIDLE()</tt> macro and <tt>_rcuidle</tt>
2470 event tracing is provided to work around this restriction.
2471 In addition, <tt>rcu_is_watching()</tt> may be used to
2472 test whether or not it is currently legal to run RCU read-side
2473 critical sections on this CPU.
2474 I learned of the need for diagnostics on the one hand
2475 and <tt>RCU_NONIDLE()</tt> on the other while inspecting
2477 Steven Rostedt supplied <tt>_rcuidle</tt> event tracing,
2478 which is used quite heavily in the idle loop.
2479 However, there are some restrictions on the code placed within
2480 <tt>RCU_NONIDLE()</tt>:
2483 <li> Blocking is prohibited.
2484 In practice, this is not a serious restriction given that idle
2485 tasks are prohibited from blocking to begin with.
2486 <li> Although nesting <tt>RCU_NONIDLE()</tt> is permitted, they cannot
2487 nest indefinitely deeply.
2488 However, given that they can be nested on the order of a million
2489 deep, even on 32-bit systems, this should not be a serious
2491 This nesting limit would probably be reached long after the
2492 compiler OOMed or the stack overflowed.
2493 <li> Any code path that enters <tt>RCU_NONIDLE()</tt> must sequence
2494 out of that same <tt>RCU_NONIDLE()</tt>.
2495 For example, the following is grossly illegal:
2501 3 goto bad_idea; /* BUG!!! */
2502 4 do_something_else();});
2508 It is just as illegal to transfer control into the middle of
2509 <tt>RCU_NONIDLE()</tt>'s argument.
2510 Yes, in theory, you could transfer in as long as you also
2511 transferred out, but in practice you could also expect to get sharply
2512 worded review comments.
2516 It is similarly socially unacceptable to interrupt an
2517 <tt>nohz_full</tt> CPU running in userspace.
2518 RCU must therefore track <tt>nohz_full</tt> userspace
2520 RCU must therefore be able to sample state at two points in
2521 time, and be able to determine whether or not some other CPU spent
2522 any time idle and/or executing in userspace.
2525 These energy-efficiency requirements have proven quite difficult to
2526 understand and to meet, for example, there have been more than five
2527 clean-sheet rewrites of RCU's energy-efficiency code, the last of
2528 which was finally able to demonstrate
2529 <a href="http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf">real energy savings running on real hardware [PDF]</a>.
2531 I learned of many of these requirements via angry phone calls:
2532 Flaming me on the Linux-kernel mailing list was apparently not
2533 sufficient to fully vent their ire at RCU's energy-efficiency bugs!
2535 <h3><a name="Memory Efficiency">Memory Efficiency</a></h3>
2538 Although small-memory non-realtime systems can simply use Tiny RCU,
2539 code size is only one aspect of memory efficiency.
2540 Another aspect is the size of the <tt>rcu_head</tt> structure
2541 used by <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>.
2542 Although this structure contains nothing more than a pair of pointers,
2543 it does appear in many RCU-protected data structures, including
2544 some that are size critical.
2545 The <tt>page</tt> structure is a case in point, as evidenced by
2546 the many occurrences of the <tt>union</tt> keyword within that structure.
2549 This need for memory efficiency is one reason that RCU uses hand-crafted
2550 singly linked lists to track the <tt>rcu_head</tt> structures that
2551 are waiting for a grace period to elapse.
2552 It is also the reason why <tt>rcu_head</tt> structures do not contain
2553 debug information, such as fields tracking the file and line of the
2554 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> that posted them.
2555 Although this information might appear in debug-only kernel builds at some
2556 point, in the meantime, the <tt>->func</tt> field will often provide
2557 the needed debug information.
2560 However, in some cases, the need for memory efficiency leads to even
2561 more extreme measures.
2562 Returning to the <tt>page</tt> structure, the <tt>rcu_head</tt> field
2563 shares storage with a great many other structures that are used at
2564 various points in the corresponding page's lifetime.
2565 In order to correctly resolve certain
2566 <a href="https://lkml.kernel.org/g/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com">race conditions</a>,
2567 the Linux kernel's memory-management subsystem needs a particular bit
2568 to remain zero during all phases of grace-period processing,
2569 and that bit happens to map to the bottom bit of the
2570 <tt>rcu_head</tt> structure's <tt>->next</tt> field.
2571 RCU makes this guarantee as long as <tt>call_rcu()</tt>
2572 is used to post the callback, as opposed to <tt>kfree_rcu()</tt>
2573 or some future “lazy”
2574 variant of <tt>call_rcu()</tt> that might one day be created for
2575 energy-efficiency purposes.
2578 That said, there are limits.
2579 RCU requires that the <tt>rcu_head</tt> structure be aligned to a
2580 two-byte boundary, and passing a misaligned <tt>rcu_head</tt>
2581 structure to one of the <tt>call_rcu()</tt> family of functions
2582 will result in a splat.
2583 It is therefore necessary to exercise caution when packing
2584 structures containing fields of type <tt>rcu_head</tt>.
2585 Why not a four-byte or even eight-byte alignment requirement?
2586 Because the m68k architecture provides only two-byte alignment,
2587 and thus acts as alignment's least common denominator.
2590 The reason for reserving the bottom bit of pointers to
2591 <tt>rcu_head</tt> structures is to leave the door open to
2592 “lazy” callbacks whose invocations can safely be deferred.
2593 Deferring invocation could potentially have energy-efficiency
2594 benefits, but only if the rate of non-lazy callbacks decreases
2595 significantly for some important workload.
2596 In the meantime, reserving the bottom bit keeps this option open
2597 in case it one day becomes useful.
2599 <h3><a name="Performance, Scalability, Response Time, and Reliability">
2600 Performance, Scalability, Response Time, and Reliability</a></h3>
2604 <a href="#Performance and Scalability">earlier discussion</a>,
2605 RCU is used heavily by hot code paths in performance-critical
2606 portions of the Linux kernel's networking, security, virtualization,
2607 and scheduling code paths.
2608 RCU must therefore use efficient implementations, especially in its
2609 read-side primitives.
2610 To that end, it would be good if preemptible RCU's implementation
2611 of <tt>rcu_read_lock()</tt> could be inlined, however, doing
2612 this requires resolving <tt>#include</tt> issues with the
2613 <tt>task_struct</tt> structure.
2616 The Linux kernel supports hardware configurations with up to
2617 4096 CPUs, which means that RCU must be extremely scalable.
2618 Algorithms that involve frequent acquisitions of global locks or
2619 frequent atomic operations on global variables simply cannot be
2620 tolerated within the RCU implementation.
2621 RCU therefore makes heavy use of a combining tree based on the
2622 <tt>rcu_node</tt> structure.
2623 RCU is required to tolerate all CPUs continuously invoking any
2624 combination of RCU's runtime primitives with minimal per-operation
2626 In fact, in many cases, increasing load must <i>decrease</i> the
2627 per-operation overhead, witness the batching optimizations for
2628 <tt>synchronize_rcu()</tt>, <tt>call_rcu()</tt>,
2629 <tt>synchronize_rcu_expedited()</tt>, and <tt>rcu_barrier()</tt>.
2630 As a general rule, RCU must cheerfully accept whatever the
2631 rest of the Linux kernel decides to throw at it.
2634 The Linux kernel is used for real-time workloads, especially
2635 in conjunction with the
2636 <a href="https://rt.wiki.kernel.org/index.php/Main_Page">-rt patchset</a>.
2637 The real-time-latency response requirements are such that the
2638 traditional approach of disabling preemption across RCU
2639 read-side critical sections is inappropriate.
2640 Kernels built with <tt>CONFIG_PREEMPT=y</tt> therefore
2641 use an RCU implementation that allows RCU read-side critical
2642 sections to be preempted.
2643 This requirement made its presence known after users made it
2644 clear that an earlier
2645 <a href="https://lwn.net/Articles/107930/">real-time patch</a>
2646 did not meet their needs, in conjunction with some
2647 <a href="https://lkml.kernel.org/g/20050318002026.GA2693@us.ibm.com">RCU issues</a>
2648 encountered by a very early version of the -rt patchset.
2651 In addition, RCU must make do with a sub-100-microsecond real-time latency
2653 In fact, on smaller systems with the -rt patchset, the Linux kernel
2654 provides sub-20-microsecond real-time latencies for the whole kernel,
2656 RCU's scalability and latency must therefore be sufficient for
2657 these sorts of configurations.
2658 To my surprise, the sub-100-microsecond real-time latency budget
2659 <a href="http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf">
2660 applies to even the largest systems [PDF]</a>,
2661 up to and including systems with 4096 CPUs.
2662 This real-time requirement motivated the grace-period kthread, which
2663 also simplified handling of a number of race conditions.
2666 RCU must avoid degrading real-time response for CPU-bound threads, whether
2667 executing in usermode (which is one use case for
2668 <tt>CONFIG_NO_HZ_FULL=y</tt>) or in the kernel.
2669 That said, CPU-bound loops in the kernel must execute
2670 <tt>cond_resched_rcu_qs()</tt> at least once per few tens of milliseconds
2671 in order to avoid receiving an IPI from RCU.
2674 Finally, RCU's status as a synchronization primitive means that
2675 any RCU failure can result in arbitrary memory corruption that can be
2676 extremely difficult to debug.
2677 This means that RCU must be extremely reliable, which in
2678 practice also means that RCU must have an aggressive stress-test
2680 This stress-test suite is called <tt>rcutorture</tt>.
2683 Although the need for <tt>rcutorture</tt> was no surprise,
2684 the current immense popularity of the Linux kernel is posing
2685 interesting—and perhaps unprecedented—validation
2687 To see this, keep in mind that there are well over one billion
2688 instances of the Linux kernel running today, given Android
2689 smartphones, Linux-powered televisions, and servers.
2690 This number can be expected to increase sharply with the advent of
2691 the celebrated Internet of Things.
2694 Suppose that RCU contains a race condition that manifests on average
2695 once per million years of runtime.
2696 This bug will be occurring about three times per <i>day</i> across
2698 RCU could simply hide behind hardware error rates, given that no one
2699 should really expect their smartphone to last for a million years.
2700 However, anyone taking too much comfort from this thought should
2701 consider the fact that in most jurisdictions, a successful multi-year
2702 test of a given mechanism, which might include a Linux kernel,
2703 suffices for a number of types of safety-critical certifications.
2704 In fact, rumor has it that the Linux kernel is already being used
2705 in production for safety-critical applications.
2706 I don't know about you, but I would feel quite bad if a bug in RCU
2708 Which might explain my recent focus on validation and verification.
2710 <h2><a name="Other RCU Flavors">Other RCU Flavors</a></h2>
2713 One of the more surprising things about RCU is that there are now
2714 no fewer than five <i>flavors</i>, or API families.
2715 In addition, the primary flavor that has been the sole focus up to
2716 this point has two different implementations, non-preemptible and
2718 The other four flavors are listed below, with requirements for each
2719 described in a separate section.
2722 <li> <a href="#Bottom-Half Flavor">Bottom-Half Flavor</a>
2723 <li> <a href="#Sched Flavor">Sched Flavor</a>
2724 <li> <a href="#Sleepable RCU">Sleepable RCU</a>
2725 <li> <a href="#Tasks RCU">Tasks RCU</a>
2726 <li> <a href="#Waiting for Multiple Grace Periods">
2727 Waiting for Multiple Grace Periods</a>
2730 <h3><a name="Bottom-Half Flavor">Bottom-Half Flavor</a></h3>
2733 The softirq-disable (AKA “bottom-half”,
2734 hence the “_bh” abbreviations)
2735 flavor of RCU, or <i>RCU-bh</i>, was developed by
2736 Dipankar Sarma to provide a flavor of RCU that could withstand the
2737 network-based denial-of-service attacks researched by Robert
2739 These attacks placed so much networking load on the system
2740 that some of the CPUs never exited softirq execution,
2741 which in turn prevented those CPUs from ever executing a context switch,
2742 which, in the RCU implementation of that time, prevented grace periods
2744 The result was an out-of-memory condition and a system hang.
2747 The solution was the creation of RCU-bh, which does
2748 <tt>local_bh_disable()</tt>
2749 across its read-side critical sections, and which uses the transition
2750 from one type of softirq processing to another as a quiescent state
2751 in addition to context switch, idle, user mode, and offline.
2752 This means that RCU-bh grace periods can complete even when some of
2753 the CPUs execute in softirq indefinitely, thus allowing algorithms
2754 based on RCU-bh to withstand network-based denial-of-service attacks.
2758 <tt>rcu_read_lock_bh()</tt> and <tt>rcu_read_unlock_bh()</tt>
2759 disable and re-enable softirq handlers, any attempt to start a softirq
2761 RCU-bh read-side critical section will be deferred.
2762 In this case, <tt>rcu_read_unlock_bh()</tt>
2763 will invoke softirq processing, which can take considerable time.
2764 One can of course argue that this softirq overhead should be associated
2765 with the code following the RCU-bh read-side critical section rather
2766 than <tt>rcu_read_unlock_bh()</tt>, but the fact
2767 is that most profiling tools cannot be expected to make this sort
2768 of fine distinction.
2769 For example, suppose that a three-millisecond-long RCU-bh read-side
2770 critical section executes during a time of heavy networking load.
2771 There will very likely be an attempt to invoke at least one softirq
2772 handler during that three milliseconds, but any such invocation will
2773 be delayed until the time of the <tt>rcu_read_unlock_bh()</tt>.
2774 This can of course make it appear at first glance as if
2775 <tt>rcu_read_unlock_bh()</tt> was executing very slowly.
2779 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-bh API</a>
2781 <tt>rcu_read_lock_bh()</tt>,
2782 <tt>rcu_read_unlock_bh()</tt>,
2783 <tt>rcu_dereference_bh()</tt>,
2784 <tt>rcu_dereference_bh_check()</tt>,
2785 <tt>synchronize_rcu_bh()</tt>,
2786 <tt>synchronize_rcu_bh_expedited()</tt>,
2787 <tt>call_rcu_bh()</tt>,
2788 <tt>rcu_barrier_bh()</tt>, and
2789 <tt>rcu_read_lock_bh_held()</tt>.
2791 <h3><a name="Sched Flavor">Sched Flavor</a></h3>
2794 Before preemptible RCU, waiting for an RCU grace period had the
2795 side effect of also waiting for all pre-existing interrupt
2797 However, there are legitimate preemptible-RCU implementations that
2798 do not have this property, given that any point in the code outside
2799 of an RCU read-side critical section can be a quiescent state.
2800 Therefore, <i>RCU-sched</i> was created, which follows “classic”
2801 RCU in that an RCU-sched grace period waits for for pre-existing
2802 interrupt and NMI handlers.
2803 In kernels built with <tt>CONFIG_PREEMPT=n</tt>, the RCU and RCU-sched
2804 APIs have identical implementations, while kernels built with
2805 <tt>CONFIG_PREEMPT=y</tt> provide a separate implementation for each.
2808 Note well that in <tt>CONFIG_PREEMPT=y</tt> kernels,
2809 <tt>rcu_read_lock_sched()</tt> and <tt>rcu_read_unlock_sched()</tt>
2810 disable and re-enable preemption, respectively.
2811 This means that if there was a preemption attempt during the
2812 RCU-sched read-side critical section, <tt>rcu_read_unlock_sched()</tt>
2813 will enter the scheduler, with all the latency and overhead entailed.
2814 Just as with <tt>rcu_read_unlock_bh()</tt>, this can make it look
2815 as if <tt>rcu_read_unlock_sched()</tt> was executing very slowly.
2816 However, the highest-priority task won't be preempted, so that task
2817 will enjoy low-overhead <tt>rcu_read_unlock_sched()</tt> invocations.
2821 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-sched API</a>
2823 <tt>rcu_read_lock_sched()</tt>,
2824 <tt>rcu_read_unlock_sched()</tt>,
2825 <tt>rcu_read_lock_sched_notrace()</tt>,
2826 <tt>rcu_read_unlock_sched_notrace()</tt>,
2827 <tt>rcu_dereference_sched()</tt>,
2828 <tt>rcu_dereference_sched_check()</tt>,
2829 <tt>synchronize_sched()</tt>,
2830 <tt>synchronize_rcu_sched_expedited()</tt>,
2831 <tt>call_rcu_sched()</tt>,
2832 <tt>rcu_barrier_sched()</tt>, and
2833 <tt>rcu_read_lock_sched_held()</tt>.
2834 However, anything that disables preemption also marks an RCU-sched
2835 read-side critical section, including
2836 <tt>preempt_disable()</tt> and <tt>preempt_enable()</tt>,
2837 <tt>local_irq_save()</tt> and <tt>local_irq_restore()</tt>,
2840 <h3><a name="Sleepable RCU">Sleepable RCU</a></h3>
2843 For well over a decade, someone saying “I need to block within
2844 an RCU read-side critical section” was a reliable indication
2845 that this someone did not understand RCU.
2846 After all, if you are always blocking in an RCU read-side critical
2847 section, you can probably afford to use a higher-overhead synchronization
2849 However, that changed with the advent of the Linux kernel's notifiers,
2850 whose RCU read-side critical
2851 sections almost never sleep, but sometimes need to.
2852 This resulted in the introduction of
2853 <a href="https://lwn.net/Articles/202847/">sleepable RCU</a>,
2857 SRCU allows different domains to be defined, with each such domain
2858 defined by an instance of an <tt>srcu_struct</tt> structure.
2859 A pointer to this structure must be passed in to each SRCU function,
2860 for example, <tt>synchronize_srcu(&ss)</tt>, where
2861 <tt>ss</tt> is the <tt>srcu_struct</tt> structure.
2862 The key benefit of these domains is that a slow SRCU reader in one
2863 domain does not delay an SRCU grace period in some other domain.
2864 That said, one consequence of these domains is that read-side code
2865 must pass a “cookie” from <tt>srcu_read_lock()</tt>
2866 to <tt>srcu_read_unlock()</tt>, for example, as follows:
2872 3 idx = srcu_read_lock(&ss);
2874 5 srcu_read_unlock(&ss, idx);
2879 As noted above, it is legal to block within SRCU read-side critical sections,
2880 however, with great power comes great responsibility.
2881 If you block forever in one of a given domain's SRCU read-side critical
2882 sections, then that domain's grace periods will also be blocked forever.
2883 Of course, one good way to block forever is to deadlock, which can
2884 happen if any operation in a given domain's SRCU read-side critical
2885 section can block waiting, either directly or indirectly, for that domain's
2886 grace period to elapse.
2887 For example, this results in a self-deadlock:
2893 3 idx = srcu_read_lock(&ss);
2895 5 synchronize_srcu(&ss);
2896 6 srcu_read_unlock(&ss, idx);
2901 However, if line 5 acquired a mutex that was held across
2902 a <tt>synchronize_srcu()</tt> for domain <tt>ss</tt>,
2903 deadlock would still be possible.
2904 Furthermore, if line 5 acquired a mutex that was held across
2905 a <tt>synchronize_srcu()</tt> for some other domain <tt>ss1</tt>,
2906 and if an <tt>ss1</tt>-domain SRCU read-side critical section
2907 acquired another mutex that was held across as <tt>ss</tt>-domain
2908 <tt>synchronize_srcu()</tt>,
2909 deadlock would again be possible.
2910 Such a deadlock cycle could extend across an arbitrarily large number
2911 of different SRCU domains.
2912 Again, with great power comes great responsibility.
2915 Unlike the other RCU flavors, SRCU read-side critical sections can
2916 run on idle and even offline CPUs.
2917 This ability requires that <tt>srcu_read_lock()</tt> and
2918 <tt>srcu_read_unlock()</tt> contain memory barriers, which means
2919 that SRCU readers will run a bit slower than would RCU readers.
2920 It also motivates the <tt>smp_mb__after_srcu_read_unlock()</tt>
2921 API, which, in combination with <tt>srcu_read_unlock()</tt>,
2922 guarantees a full memory barrier.
2925 Also unlike other RCU flavors, SRCU's callbacks-wait function
2926 <tt>srcu_barrier()</tt> may be invoked from CPU-hotplug notifiers,
2927 though this is not necessarily a good idea.
2928 The reason that this is possible is that SRCU is insensitive
2929 to whether or not a CPU is online, which means that <tt>srcu_barrier()</tt>
2930 need not exclude CPU-hotplug operations.
2933 SRCU also differs from other RCU flavors in that SRCU's expedited and
2934 non-expedited grace periods are implemented by the same mechanism.
2935 This means that in the current SRCU implementation, expediting a
2936 future grace period has the side effect of expediting all prior
2937 grace periods that have not yet completed.
2938 (But please note that this is a property of the current implementation,
2939 not necessarily of future implementations.)
2940 In addition, if SRCU has been idle for longer than the interval
2941 specified by the <tt>srcutree.exp_holdoff</tt> kernel boot parameter
2942 (25 microseconds by default),
2943 and if a <tt>synchronize_srcu()</tt> invocation ends this idle period,
2944 that invocation will be automatically expedited.
2947 As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating
2948 a locking bottleneck present in prior kernel versions.
2949 Although this will allow users to put much heavier stress on
2950 <tt>call_srcu()</tt>, it is important to note that SRCU does not
2951 yet take any special steps to deal with callback flooding.
2952 So if you are posting (say) 10,000 SRCU callbacks per second per CPU,
2953 you are probably totally OK, but if you intend to post (say) 1,000,000
2954 SRCU callbacks per second per CPU, please run some tests first.
2955 SRCU just might need a few adjustment to deal with that sort of load.
2956 Of course, your mileage may vary based on the speed of your CPUs and
2957 the size of your memory.
2961 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">SRCU API</a>
2963 <tt>srcu_read_lock()</tt>,
2964 <tt>srcu_read_unlock()</tt>,
2965 <tt>srcu_dereference()</tt>,
2966 <tt>srcu_dereference_check()</tt>,
2967 <tt>synchronize_srcu()</tt>,
2968 <tt>synchronize_srcu_expedited()</tt>,
2969 <tt>call_srcu()</tt>,
2970 <tt>srcu_barrier()</tt>, and
2971 <tt>srcu_read_lock_held()</tt>.
2973 <tt>DEFINE_SRCU()</tt>,
2974 <tt>DEFINE_STATIC_SRCU()</tt>, and
2975 <tt>init_srcu_struct()</tt>
2976 APIs for defining and initializing <tt>srcu_struct</tt> structures.
2978 <h3><a name="Tasks RCU">Tasks RCU</a></h3>
2981 Some forms of tracing use “trampolines” to handle the
2982 binary rewriting required to install different types of probes.
2983 It would be good to be able to free old trampolines, which sounds
2984 like a job for some form of RCU.
2985 However, because it is necessary to be able to install a trace
2986 anywhere in the code, it is not possible to use read-side markers
2987 such as <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
2988 In addition, it does not work to have these markers in the trampoline
2989 itself, because there would need to be instructions following
2990 <tt>rcu_read_unlock()</tt>.
2991 Although <tt>synchronize_rcu()</tt> would guarantee that execution
2992 reached the <tt>rcu_read_unlock()</tt>, it would not be able to
2993 guarantee that execution had completely left the trampoline.
2996 The solution, in the form of
2997 <a href="https://lwn.net/Articles/607117/"><i>Tasks RCU</i></a>,
2999 read-side critical sections that are delimited by voluntary context
3000 switches, that is, calls to <tt>schedule()</tt>,
3001 <tt>cond_resched_rcu_qs()</tt>, and
3002 <tt>synchronize_rcu_tasks()</tt>.
3003 In addition, transitions to and from userspace execution also delimit
3004 tasks-RCU read-side critical sections.
3007 The tasks-RCU API is quite compact, consisting only of
3008 <tt>call_rcu_tasks()</tt>,
3009 <tt>synchronize_rcu_tasks()</tt>, and
3010 <tt>rcu_barrier_tasks()</tt>.
3012 <h3><a name="Waiting for Multiple Grace Periods">
3013 Waiting for Multiple Grace Periods</a></h3>
3016 Perhaps you have an RCU protected data structure that is accessed from
3017 RCU read-side critical sections, from softirq handlers, and from
3018 hardware interrupt handlers.
3019 That is three flavors of RCU, the normal flavor, the bottom-half flavor,
3020 and the sched flavor.
3021 How to wait for a compound grace period?
3024 The best approach is usually to “just say no!” and
3025 insert <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
3026 around each RCU read-side critical section, regardless of what
3027 environment it happens to be in.
3028 But suppose that some of the RCU read-side critical sections are
3029 on extremely hot code paths, and that use of <tt>CONFIG_PREEMPT=n</tt>
3030 is not a viable option, so that <tt>rcu_read_lock()</tt> and
3031 <tt>rcu_read_unlock()</tt> are not free.
3035 You <i>could</i> wait on all three grace periods in succession, as follows:
3039 1 synchronize_rcu();
3040 2 synchronize_rcu_bh();
3041 3 synchronize_sched();
3046 This works, but triples the update-side latency penalty.
3047 In cases where this is not acceptable, <tt>synchronize_rcu_mult()</tt>
3048 may be used to wait on all three flavors of grace period concurrently:
3052 1 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched);
3057 But what if it is necessary to also wait on SRCU?
3058 This can be done as follows:
3062 1 static void call_my_srcu(struct rcu_head *head,
3063 2 void (*func)(struct rcu_head *head))
3065 4 call_srcu(&my_srcu, head, func);
3068 7 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched, call_my_srcu);
3073 If you needed to wait on multiple different flavors of SRCU
3074 (but why???), you would need to create a wrapper function resembling
3075 <tt>call_my_srcu()</tt> for each SRCU flavor.
3078 <tr><th> </th></tr>
3079 <tr><th align="left">Quick Quiz:</th></tr>
3081 But what if I need to wait for multiple RCU flavors, but I also need
3082 the grace periods to be expedited?
3084 <tr><th align="left">Answer:</th></tr>
3085 <tr><td bgcolor="#ffffff"><font color="ffffff">
3086 If you are using expedited grace periods, there should be less penalty
3087 for waiting on them in succession.
3088 But if that is nevertheless a problem, you can use workqueues
3089 or multiple kthreads to wait on the various expedited grace
3090 periods concurrently.
3092 <tr><td> </td></tr>
3096 Again, it is usually better to adjust the RCU read-side critical sections
3097 to use a single flavor of RCU, but when this is not feasible, you can use
3098 <tt>synchronize_rcu_mult()</tt>.
3100 <h2><a name="Possible Future Changes">Possible Future Changes</a></h2>
3103 One of the tricks that RCU uses to attain update-side scalability is
3104 to increase grace-period latency with increasing numbers of CPUs.
3105 If this becomes a serious problem, it will be necessary to rework the
3106 grace-period state machine so as to avoid the need for the additional
3110 Expedited grace periods scan the CPUs, so their latency and overhead
3111 increases with increasing numbers of CPUs.
3112 If this becomes a serious problem on large systems, it will be necessary
3113 to do some redesign to avoid this scalability problem.
3116 RCU disables CPU hotplug in a few places, perhaps most notably in the
3117 <tt>rcu_barrier()</tt> operations.
3118 If there is a strong reason to use <tt>rcu_barrier()</tt> in CPU-hotplug
3119 notifiers, it will be necessary to avoid disabling CPU hotplug.
3120 This would introduce some complexity, so there had better be a <i>very</i>
3124 The tradeoff between grace-period latency on the one hand and interruptions
3125 of other CPUs on the other hand may need to be re-examined.
3126 The desire is of course for zero grace-period latency as well as zero
3127 interprocessor interrupts undertaken during an expedited grace period
3129 While this ideal is unlikely to be achievable, it is quite possible that
3130 further improvements can be made.
3133 The multiprocessor implementations of RCU use a combining tree that
3134 groups CPUs so as to reduce lock contention and increase cache locality.
3135 However, this combining tree does not spread its memory across NUMA
3136 nodes nor does it align the CPU groups with hardware features such
3137 as sockets or cores.
3138 Such spreading and alignment is currently believed to be unnecessary
3139 because the hotpath read-side primitives do not access the combining
3140 tree, nor does <tt>call_rcu()</tt> in the common case.
3141 If you believe that your architecture needs such spreading and alignment,
3142 then your architecture should also benefit from the
3143 <tt>rcutree.rcu_fanout_leaf</tt> boot parameter, which can be set
3144 to the number of CPUs in a socket, NUMA node, or whatever.
3145 If the number of CPUs is too large, use a fraction of the number of
3147 If the number of CPUs is a large prime number, well, that certainly
3148 is an “interesting” architectural choice!
3149 More flexible arrangements might be considered, but only if
3150 <tt>rcutree.rcu_fanout_leaf</tt> has proven inadequate, and only
3151 if the inadequacy has been demonstrated by a carefully run and
3152 realistic system-level workload.
3155 Please note that arrangements that require RCU to remap CPU numbers will
3156 require extremely good demonstration of need and full exploration of
3160 There is an embarrassingly large number of flavors of RCU, and this
3161 number has been increasing over time.
3162 Perhaps it will be possible to combine some at some future date.
3165 RCU's various kthreads are reasonably recent additions.
3166 It is quite likely that adjustments will be required to more gracefully
3167 handle extreme loads.
3168 It might also be necessary to be able to relate CPU utilization by
3169 RCU's kthreads and softirq handlers to the code that instigated this
3171 For example, RCU callback overhead might be charged back to the
3172 originating <tt>call_rcu()</tt> instance, though probably not
3173 in production kernels.
3175 <h2><a name="Summary">Summary</a></h2>
3178 This document has presented more than two decade's worth of RCU
3180 Given that the requirements keep changing, this will not be the last
3181 word on this subject, but at least it serves to get an important
3182 subset of the requirements set forth.
3184 <h2><a name="Acknowledgments">Acknowledgments</a></h2>
3186 I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar,
3187 Oleg Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and
3188 Andy Lutomirski for their help in rendering
3189 this article human readable, and to Michelle Rankin for her support
3191 Other contributions are acknowledged in the Linux kernel's git archive.