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.
563 with <tt>CONFIG_SPARSE_RCU_POINTER=y</tt>,
564 <tt>sparse</tt> will complain if you
565 define <tt>gp</tt> with <tt>__rcu</tt> and then
566 access it without using
567 either <tt>rcu_access_pointer()</tt> or <tt>rcu_dereference()</tt>.
569 <tr><td> </td></tr>
573 In short, RCU's publish-subscribe guarantee is provided by the combination
574 of <tt>rcu_assign_pointer()</tt> and <tt>rcu_dereference()</tt>.
575 This guarantee allows data elements to be safely added to RCU-protected
576 linked data structures without disrupting RCU readers.
577 This guarantee can be used in combination with the grace-period
578 guarantee to also allow data elements to be removed from RCU-protected
579 linked data structures, again without disrupting RCU readers.
582 This guarantee was only partially premeditated.
583 DYNIX/ptx used an explicit memory barrier for publication, but had nothing
584 resembling <tt>rcu_dereference()</tt> for subscription, nor did it
585 have anything resembling the <tt>smp_read_barrier_depends()</tt>
586 that was later subsumed into <tt>rcu_dereference()</tt>.
587 The need for these operations made itself known quite suddenly at a
588 late-1990s meeting with the DEC Alpha architects, back in the days when
589 DEC was still a free-standing company.
590 It took the Alpha architects a good hour to convince me that any sort
591 of barrier would ever be needed, and it then took me a good <i>two</i> hours
592 to convince them that their documentation did not make this point clear.
593 More recent work with the C and C++ standards committees have provided
594 much education on tricks and traps from the compiler.
595 In short, compilers were much less tricky in the early 1990s, but in
596 2015, don't even think about omitting <tt>rcu_dereference()</tt>!
598 <h3><a name="Memory-Barrier Guarantees">Memory-Barrier Guarantees</a></h3>
601 The previous section's simple linked-data-structure scenario clearly
602 demonstrates the need for RCU's stringent memory-ordering guarantees on
603 systems with more than one CPU:
606 <li> Each CPU that has an RCU read-side critical section that
607 begins before <tt>synchronize_rcu()</tt> starts is
608 guaranteed to execute a full memory barrier between the time
609 that the RCU read-side critical section ends and the time that
610 <tt>synchronize_rcu()</tt> returns.
611 Without this guarantee, a pre-existing RCU read-side critical section
612 might hold a reference to the newly removed <tt>struct foo</tt>
613 after the <tt>kfree()</tt> on line 14 of
614 <tt>remove_gp_synchronous()</tt>.
615 <li> Each CPU that has an RCU read-side critical section that ends
616 after <tt>synchronize_rcu()</tt> returns is guaranteed
617 to execute a full memory barrier between the time that
618 <tt>synchronize_rcu()</tt> begins and the time that the RCU
619 read-side critical section begins.
620 Without this guarantee, a later RCU read-side critical section
621 running after the <tt>kfree()</tt> on line 14 of
622 <tt>remove_gp_synchronous()</tt> might
623 later run <tt>do_something_gp()</tt> and find the
624 newly deleted <tt>struct foo</tt>.
625 <li> If the task invoking <tt>synchronize_rcu()</tt> remains
626 on a given CPU, then that CPU is guaranteed to execute a full
627 memory barrier sometime during the execution of
628 <tt>synchronize_rcu()</tt>.
629 This guarantee ensures that the <tt>kfree()</tt> on
630 line 14 of <tt>remove_gp_synchronous()</tt> really does
631 execute after the removal on line 11.
632 <li> If the task invoking <tt>synchronize_rcu()</tt> migrates
633 among a group of CPUs during that invocation, then each of the
634 CPUs in that group is guaranteed to execute a full memory barrier
635 sometime during the execution of <tt>synchronize_rcu()</tt>.
636 This guarantee also ensures that the <tt>kfree()</tt> on
637 line 14 of <tt>remove_gp_synchronous()</tt> really does
638 execute after the removal on
639 line 11, but also in the case where the thread executing the
640 <tt>synchronize_rcu()</tt> migrates in the meantime.
644 <tr><th> </th></tr>
645 <tr><th align="left">Quick Quiz:</th></tr>
647 Given that multiple CPUs can start RCU read-side critical sections
648 at any time without any ordering whatsoever, how can RCU possibly
649 tell whether or not a given RCU read-side critical section starts
650 before a given instance of <tt>synchronize_rcu()</tt>?
652 <tr><th align="left">Answer:</th></tr>
653 <tr><td bgcolor="#ffffff"><font color="ffffff">
654 If RCU cannot tell whether or not a given
655 RCU read-side critical section starts before a
656 given instance of <tt>synchronize_rcu()</tt>,
657 then it must assume that the RCU read-side critical section
659 In other words, a given instance of <tt>synchronize_rcu()</tt>
660 can avoid waiting on a given RCU read-side critical section only
661 if it can prove that <tt>synchronize_rcu()</tt> started first.
664 <p><font color="ffffff">
665 A related question is “When <tt>rcu_read_lock()</tt>
666 doesn't generate any code, why does it matter how it relates
667 to a grace period?”
668 The answer is that it is not the relationship of
669 <tt>rcu_read_lock()</tt> itself that is important, but rather
670 the relationship of the code within the enclosed RCU read-side
671 critical section to the code preceding and following the
673 If we take this viewpoint, then a given RCU read-side critical
674 section begins before a given grace period when some access
675 preceding the grace period observes the effect of some access
676 within the critical section, in which case none of the accesses
677 within the critical section may observe the effects of any
678 access following the grace period.
681 <p><font color="ffffff">
682 As of late 2016, mathematical models of RCU take this
683 viewpoint, for example, see slides 62 and 63
685 <a href="http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.2016.10.04c.LCE.pdf">2016 LinuxCon EU</a>
688 <tr><td> </td></tr>
692 <tr><th> </th></tr>
693 <tr><th align="left">Quick Quiz:</th></tr>
695 The first and second guarantees require unbelievably strict ordering!
696 Are all these memory barriers <i> really</i> required?
698 <tr><th align="left">Answer:</th></tr>
699 <tr><td bgcolor="#ffffff"><font color="ffffff">
700 Yes, they really are required.
701 To see why the first guarantee is required, consider the following
706 <li> <font color="ffffff">
707 CPU 1: <tt>rcu_read_lock()</tt>
709 <li> <font color="ffffff">
710 CPU 1: <tt>q = rcu_dereference(gp);
711 /* Very likely to return p. */</tt>
713 <li> <font color="ffffff">
714 CPU 0: <tt>list_del_rcu(p);</tt>
716 <li> <font color="ffffff">
717 CPU 0: <tt>synchronize_rcu()</tt> starts.
719 <li> <font color="ffffff">
720 CPU 1: <tt>do_something_with(q->a);
721 /* No smp_mb(), so might happen after kfree(). */</tt>
723 <li> <font color="ffffff">
724 CPU 1: <tt>rcu_read_unlock()</tt>
726 <li> <font color="ffffff">
727 CPU 0: <tt>synchronize_rcu()</tt> returns.
729 <li> <font color="ffffff">
730 CPU 0: <tt>kfree(p);</tt>
734 <p><font color="ffffff">
735 Therefore, there absolutely must be a full memory barrier between the
736 end of the RCU read-side critical section and the end of the
740 <p><font color="ffffff">
741 The sequence of events demonstrating the necessity of the second rule
746 <li> <font color="ffffff">CPU 0: <tt>list_del_rcu(p);</tt>
748 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> starts.
750 <li> <font color="ffffff">CPU 1: <tt>rcu_read_lock()</tt>
752 <li> <font color="ffffff">CPU 1: <tt>q = rcu_dereference(gp);
753 /* Might return p if no memory barrier. */</tt>
755 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> returns.
757 <li> <font color="ffffff">CPU 0: <tt>kfree(p);</tt>
759 <li> <font color="ffffff">
760 CPU 1: <tt>do_something_with(q->a); /* Boom!!! */</tt>
762 <li> <font color="ffffff">CPU 1: <tt>rcu_read_unlock()</tt>
766 <p><font color="ffffff">
767 And similarly, without a memory barrier between the beginning of the
768 grace period and the beginning of the RCU read-side critical section,
769 CPU 1 might end up accessing the freelist.
772 <p><font color="ffffff">
773 The “as if” rule of course applies, so that any
774 implementation that acts as if the appropriate memory barriers
775 were in place is a correct implementation.
776 That said, it is much easier to fool yourself into believing
777 that you have adhered to the as-if rule than it is to actually
780 <tr><td> </td></tr>
784 <tr><th> </th></tr>
785 <tr><th align="left">Quick Quiz:</th></tr>
787 You claim that <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
788 generate absolutely no code in some kernel builds.
789 This means that the compiler might arbitrarily rearrange consecutive
790 RCU read-side critical sections.
791 Given such rearrangement, if a given RCU read-side critical section
792 is done, how can you be sure that all prior RCU read-side critical
794 Won't the compiler rearrangements make that impossible to determine?
796 <tr><th align="left">Answer:</th></tr>
797 <tr><td bgcolor="#ffffff"><font color="ffffff">
798 In cases where <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
799 generate absolutely no code, RCU infers quiescent states only at
800 special locations, for example, within the scheduler.
801 Because calls to <tt>schedule()</tt> had better prevent calling-code
802 accesses to shared variables from being rearranged across the call to
803 <tt>schedule()</tt>, if RCU detects the end of a given RCU read-side
804 critical section, it will necessarily detect the end of all prior
805 RCU read-side critical sections, no matter how aggressively the
806 compiler scrambles the code.
809 <p><font color="ffffff">
810 Again, this all assumes that the compiler cannot scramble code across
811 calls to the scheduler, out of interrupt handlers, into the idle loop,
812 into user-mode code, and so on.
813 But if your kernel build allows that sort of scrambling, you have broken
814 far more than just RCU!
816 <tr><td> </td></tr>
820 Note that these memory-barrier requirements do not replace the fundamental
821 RCU requirement that a grace period wait for all pre-existing readers.
822 On the contrary, the memory barriers called out in this section must operate in
823 such a way as to <i>enforce</i> this fundamental requirement.
824 Of course, different implementations enforce this requirement in different
825 ways, but enforce it they must.
827 <h3><a name="RCU Primitives Guaranteed to Execute Unconditionally">RCU Primitives Guaranteed to Execute Unconditionally</a></h3>
830 The common-case RCU primitives are unconditional.
831 They are invoked, they do their job, and they return, with no possibility
832 of error, and no need to retry.
833 This is a key RCU design philosophy.
836 However, this philosophy is pragmatic rather than pigheaded.
837 If someone comes up with a good justification for a particular conditional
838 RCU primitive, it might well be implemented and added.
839 After all, this guarantee was reverse-engineered, not premeditated.
840 The unconditional nature of the RCU primitives was initially an
841 accident of implementation, and later experience with synchronization
842 primitives with conditional primitives caused me to elevate this
843 accident to a guarantee.
844 Therefore, the justification for adding a conditional primitive to
845 RCU would need to be based on detailed and compelling use cases.
847 <h3><a name="Guaranteed Read-to-Write Upgrade">Guaranteed Read-to-Write Upgrade</a></h3>
850 As far as RCU is concerned, it is always possible to carry out an
851 update within an RCU read-side critical section.
852 For example, that RCU read-side critical section might search for
853 a given data element, and then might acquire the update-side
854 spinlock in order to update that element, all while remaining
855 in that RCU read-side critical section.
856 Of course, it is necessary to exit the RCU read-side critical section
857 before invoking <tt>synchronize_rcu()</tt>, however, this
858 inconvenience can be avoided through use of the
859 <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt> API members
860 described later in this document.
863 <tr><th> </th></tr>
864 <tr><th align="left">Quick Quiz:</th></tr>
866 But how does the upgrade-to-write operation exclude other readers?
868 <tr><th align="left">Answer:</th></tr>
869 <tr><td bgcolor="#ffffff"><font color="ffffff">
870 It doesn't, just like normal RCU updates, which also do not exclude
873 <tr><td> </td></tr>
877 This guarantee allows lookup code to be shared between read-side
878 and update-side code, and was premeditated, appearing in the earliest
879 DYNIX/ptx RCU documentation.
881 <h2><a name="Fundamental Non-Requirements">Fundamental Non-Requirements</a></h2>
884 RCU provides extremely lightweight readers, and its read-side guarantees,
885 though quite useful, are correspondingly lightweight.
886 It is therefore all too easy to assume that RCU is guaranteeing more
888 Of course, the list of things that RCU does not guarantee is infinitely
889 long, however, the following sections list a few non-guarantees that
890 have caused confusion.
891 Except where otherwise noted, these non-guarantees were premeditated.
894 <li> <a href="#Readers Impose Minimal Ordering">
895 Readers Impose Minimal Ordering</a>
896 <li> <a href="#Readers Do Not Exclude Updaters">
897 Readers Do Not Exclude Updaters</a>
898 <li> <a href="#Updaters Only Wait For Old Readers">
899 Updaters Only Wait For Old Readers</a>
900 <li> <a href="#Grace Periods Don't Partition Read-Side Critical Sections">
901 Grace Periods Don't Partition Read-Side Critical Sections</a>
902 <li> <a href="#Read-Side Critical Sections Don't Partition Grace Periods">
903 Read-Side Critical Sections Don't Partition Grace Periods</a>
904 <li> <a href="#Disabling Preemption Does Not Block Grace Periods">
905 Disabling Preemption Does Not Block Grace Periods</a>
908 <h3><a name="Readers Impose Minimal Ordering">Readers Impose Minimal Ordering</a></h3>
911 Reader-side markers such as <tt>rcu_read_lock()</tt> and
912 <tt>rcu_read_unlock()</tt> provide absolutely no ordering guarantees
913 except through their interaction with the grace-period APIs such as
914 <tt>synchronize_rcu()</tt>.
915 To see this, consider the following pair of threads:
929 11 void thread1(void)
932 14 r1 = READ_ONCE(y);
933 15 rcu_read_unlock();
935 17 r2 = READ_ONCE(x);
936 18 rcu_read_unlock();
942 After <tt>thread0()</tt> and <tt>thread1()</tt> execute
943 concurrently, it is quite possible to have
947 (r1 == 1 && r2 == 0)
951 (that is, <tt>y</tt> appears to have been assigned before <tt>x</tt>),
952 which would not be possible if <tt>rcu_read_lock()</tt> and
953 <tt>rcu_read_unlock()</tt> had much in the way of ordering
955 But they do not, so the CPU is within its rights
956 to do significant reordering.
957 This is by design: Any significant ordering constraints would slow down
958 these fast-path APIs.
961 <tr><th> </th></tr>
962 <tr><th align="left">Quick Quiz:</th></tr>
964 Can't the compiler also reorder this code?
966 <tr><th align="left">Answer:</th></tr>
967 <tr><td bgcolor="#ffffff"><font color="ffffff">
968 No, the volatile casts in <tt>READ_ONCE()</tt> and
969 <tt>WRITE_ONCE()</tt> prevent the compiler from reordering in
970 this particular case.
972 <tr><td> </td></tr>
975 <h3><a name="Readers Do Not Exclude Updaters">Readers Do Not Exclude Updaters</a></h3>
978 Neither <tt>rcu_read_lock()</tt> nor <tt>rcu_read_unlock()</tt>
980 All they do is to prevent grace periods from ending.
981 The following example illustrates this:
990 6 do_something_with_nonzero_x();
992 8 WARN_ON(!r2); /* BUG!!! */
994 10 rcu_read_unlock();
997 13 void thread1(void)
999 15 spin_lock(&my_lock);
1000 16 WRITE_ONCE(x, 1);
1001 17 WRITE_ONCE(y, 1);
1002 18 spin_unlock(&my_lock);
1008 If the <tt>thread0()</tt> function's <tt>rcu_read_lock()</tt>
1009 excluded the <tt>thread1()</tt> function's update,
1010 the <tt>WARN_ON()</tt> could never fire.
1011 But the fact is that <tt>rcu_read_lock()</tt> does not exclude
1012 much of anything aside from subsequent grace periods, of which
1013 <tt>thread1()</tt> has none, so the
1014 <tt>WARN_ON()</tt> can and does fire.
1016 <h3><a name="Updaters Only Wait For Old Readers">Updaters Only Wait For Old Readers</a></h3>
1019 It might be tempting to assume that after <tt>synchronize_rcu()</tt>
1020 completes, there are no readers executing.
1021 This temptation must be avoided because
1022 new readers can start immediately after <tt>synchronize_rcu()</tt>
1023 starts, and <tt>synchronize_rcu()</tt> is under no
1024 obligation to wait for these new readers.
1027 <tr><th> </th></tr>
1028 <tr><th align="left">Quick Quiz:</th></tr>
1030 Suppose that synchronize_rcu() did wait until <i>all</i>
1031 readers had completed instead of waiting only on
1032 pre-existing readers.
1033 For how long would the updater be able to rely on there
1036 <tr><th align="left">Answer:</th></tr>
1037 <tr><td bgcolor="#ffffff"><font color="ffffff">
1039 Even if <tt>synchronize_rcu()</tt> were to wait until
1040 all readers had completed, a new reader might start immediately after
1041 <tt>synchronize_rcu()</tt> completed.
1042 Therefore, the code following
1043 <tt>synchronize_rcu()</tt> can <i>never</i> rely on there being
1046 <tr><td> </td></tr>
1049 <h3><a name="Grace Periods Don't Partition Read-Side Critical Sections">
1050 Grace Periods Don't Partition Read-Side Critical Sections</a></h3>
1053 It is tempting to assume that if any part of one RCU read-side critical
1054 section precedes a given grace period, and if any part of another RCU
1055 read-side critical section follows that same grace period, then all of
1056 the first RCU read-side critical section must precede all of the second.
1057 However, this just isn't the case: A single grace period does not
1058 partition the set of RCU read-side critical sections.
1059 An example of this situation can be illustrated as follows, where
1060 <tt>x</tt>, <tt>y</tt>, and <tt>z</tt> are initially all zero:
1064 1 void thread0(void)
1069 6 rcu_read_unlock();
1072 9 void thread1(void)
1074 11 r1 = READ_ONCE(a);
1075 12 synchronize_rcu();
1076 13 WRITE_ONCE(c, 1);
1079 16 void thread2(void)
1082 19 r2 = READ_ONCE(b);
1083 20 r3 = READ_ONCE(c);
1084 21 rcu_read_unlock();
1090 It turns out that the outcome:
1094 (r1 == 1 && r2 == 0 && r3 == 1)
1098 is entirely possible.
1099 The following figure show how this can happen, with each circled
1100 <tt>QS</tt> indicating the point at which RCU recorded a
1101 <i>quiescent state</i> for each thread, that is, a state in which
1102 RCU knows that the thread cannot be in the midst of an RCU read-side
1103 critical section that started before the current grace period:
1105 <p><img src="GPpartitionReaders1.svg" alt="GPpartitionReaders1.svg" width="60%"></p>
1108 If it is necessary to partition RCU read-side critical sections in this
1109 manner, it is necessary to use two grace periods, where the first
1110 grace period is known to end before the second grace period starts:
1114 1 void thread0(void)
1119 6 rcu_read_unlock();
1122 9 void thread1(void)
1124 11 r1 = READ_ONCE(a);
1125 12 synchronize_rcu();
1126 13 WRITE_ONCE(c, 1);
1129 16 void thread2(void)
1131 18 r2 = READ_ONCE(c);
1132 19 synchronize_rcu();
1133 20 WRITE_ONCE(d, 1);
1136 23 void thread3(void)
1139 26 r3 = READ_ONCE(b);
1140 27 r4 = READ_ONCE(d);
1141 28 rcu_read_unlock();
1147 Here, if <tt>(r1 == 1)</tt>, then
1148 <tt>thread0()</tt>'s write to <tt>b</tt> must happen
1149 before the end of <tt>thread1()</tt>'s grace period.
1150 If in addition <tt>(r4 == 1)</tt>, then
1151 <tt>thread3()</tt>'s read from <tt>b</tt> must happen
1152 after the beginning of <tt>thread2()</tt>'s grace period.
1153 If it is also the case that <tt>(r2 == 1)</tt>, then the
1154 end of <tt>thread1()</tt>'s grace period must precede the
1155 beginning of <tt>thread2()</tt>'s grace period.
1156 This mean that the two RCU read-side critical sections cannot overlap,
1157 guaranteeing that <tt>(r3 == 1)</tt>.
1158 As a result, the outcome:
1162 (r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
1169 This non-requirement was also non-premeditated, but became apparent
1170 when studying RCU's interaction with memory ordering.
1172 <h3><a name="Read-Side Critical Sections Don't Partition Grace Periods">
1173 Read-Side Critical Sections Don't Partition Grace Periods</a></h3>
1176 It is also tempting to assume that if an RCU read-side critical section
1177 happens between a pair of grace periods, then those grace periods cannot
1179 However, this temptation leads nowhere good, as can be illustrated by
1180 the following, with all variables initially zero:
1184 1 void thread0(void)
1189 6 rcu_read_unlock();
1192 9 void thread1(void)
1194 11 r1 = READ_ONCE(a);
1195 12 synchronize_rcu();
1196 13 WRITE_ONCE(c, 1);
1199 16 void thread2(void)
1202 19 WRITE_ONCE(d, 1);
1203 20 r2 = READ_ONCE(c);
1204 21 rcu_read_unlock();
1207 24 void thread3(void)
1209 26 r3 = READ_ONCE(d);
1210 27 synchronize_rcu();
1211 28 WRITE_ONCE(e, 1);
1214 31 void thread4(void)
1217 34 r4 = READ_ONCE(b);
1218 35 r5 = READ_ONCE(e);
1219 36 rcu_read_unlock();
1225 In this case, the outcome:
1229 (r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
1233 is entirely possible, as illustrated below:
1235 <p><img src="ReadersPartitionGP1.svg" alt="ReadersPartitionGP1.svg" width="100%"></p>
1238 Again, an RCU read-side critical section can overlap almost all of a
1239 given grace period, just so long as it does not overlap the entire
1241 As a result, an RCU read-side critical section cannot partition a pair
1242 of RCU grace periods.
1245 <tr><th> </th></tr>
1246 <tr><th align="left">Quick Quiz:</th></tr>
1248 How long a sequence of grace periods, each separated by an RCU
1249 read-side critical section, would be required to partition the RCU
1250 read-side critical sections at the beginning and end of the chain?
1252 <tr><th align="left">Answer:</th></tr>
1253 <tr><td bgcolor="#ffffff"><font color="ffffff">
1254 In theory, an infinite number.
1255 In practice, an unknown number that is sensitive to both implementation
1256 details and timing considerations.
1257 Therefore, even in practice, RCU users must abide by the
1258 theoretical rather than the practical answer.
1260 <tr><td> </td></tr>
1263 <h3><a name="Disabling Preemption Does Not Block Grace Periods">
1264 Disabling Preemption Does Not Block Grace Periods</a></h3>
1267 There was a time when disabling preemption on any given CPU would block
1268 subsequent grace periods.
1269 However, this was an accident of implementation and is not a requirement.
1270 And in the current Linux-kernel implementation, disabling preemption
1271 on a given CPU in fact does not block grace periods, as Oleg Nesterov
1272 <a href="https://lkml.kernel.org/g/20150614193825.GA19582@redhat.com">demonstrated</a>.
1275 If you need a preempt-disable region to block grace periods, you need to add
1276 <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>, for example
1281 1 preempt_disable();
1284 4 rcu_read_unlock();
1287 7 /* Spinlocks implicitly disable preemption. */
1288 8 spin_lock(&mylock);
1291 11 rcu_read_unlock();
1292 12 spin_unlock(&mylock);
1297 In theory, you could enter the RCU read-side critical section first,
1298 but it is more efficient to keep the entire RCU read-side critical
1299 section contained in the preempt-disable region as shown above.
1300 Of course, RCU read-side critical sections that extend outside of
1301 preempt-disable regions will work correctly, but such critical sections
1302 can be preempted, which forces <tt>rcu_read_unlock()</tt> to do
1304 And no, this is <i>not</i> an invitation to enclose all of your RCU
1305 read-side critical sections within preempt-disable regions, because
1306 doing so would degrade real-time response.
1309 This non-requirement appeared with preemptible RCU.
1310 If you need a grace period that waits on non-preemptible code regions, use
1311 <a href="#Sched Flavor">RCU-sched</a>.
1313 <h2><a name="Parallelism Facts of Life">Parallelism Facts of Life</a></h2>
1316 These parallelism facts of life are by no means specific to RCU, but
1317 the RCU implementation must abide by them.
1318 They therefore bear repeating:
1321 <li> Any CPU or task may be delayed at any time,
1322 and any attempts to avoid these delays by disabling
1323 preemption, interrupts, or whatever are completely futile.
1324 This is most obvious in preemptible user-level
1325 environments and in virtualized environments (where
1326 a given guest OS's VCPUs can be preempted at any time by
1327 the underlying hypervisor), but can also happen in bare-metal
1328 environments due to ECC errors, NMIs, and other hardware
1330 Although a delay of more than about 20 seconds can result
1331 in splats, the RCU implementation is obligated to use
1332 algorithms that can tolerate extremely long delays, but where
1333 “extremely long” is not long enough to allow
1334 wrap-around when incrementing a 64-bit counter.
1335 <li> Both the compiler and the CPU can reorder memory accesses.
1336 Where it matters, RCU must use compiler directives and
1337 memory-barrier instructions to preserve ordering.
1338 <li> Conflicting writes to memory locations in any given cache line
1339 will result in expensive cache misses.
1340 Greater numbers of concurrent writes and more-frequent
1341 concurrent writes will result in more dramatic slowdowns.
1342 RCU is therefore obligated to use algorithms that have
1343 sufficient locality to avoid significant performance and
1344 scalability problems.
1345 <li> As a rough rule of thumb, only one CPU's worth of processing
1346 may be carried out under the protection of any given exclusive
1348 RCU must therefore use scalable locking designs.
1349 <li> Counters are finite, especially on 32-bit systems.
1350 RCU's use of counters must therefore tolerate counter wrap,
1351 or be designed such that counter wrap would take way more
1352 time than a single system is likely to run.
1353 An uptime of ten years is quite possible, a runtime
1354 of a century much less so.
1355 As an example of the latter, RCU's dyntick-idle nesting counter
1356 allows 54 bits for interrupt nesting level (this counter
1357 is 64 bits even on a 32-bit system).
1358 Overflowing this counter requires 2<sup>54</sup>
1359 half-interrupts on a given CPU without that CPU ever going idle.
1360 If a half-interrupt happened every microsecond, it would take
1361 570 years of runtime to overflow this counter, which is currently
1362 believed to be an acceptably long time.
1363 <li> Linux systems can have thousands of CPUs running a single
1364 Linux kernel in a single shared-memory environment.
1365 RCU must therefore pay close attention to high-end scalability.
1369 This last parallelism fact of life means that RCU must pay special
1370 attention to the preceding facts of life.
1371 The idea that Linux might scale to systems with thousands of CPUs would
1372 have been met with some skepticism in the 1990s, but these requirements
1373 would have otherwise have been unsurprising, even in the early 1990s.
1375 <h2><a name="Quality-of-Implementation Requirements">Quality-of-Implementation Requirements</a></h2>
1378 These sections list quality-of-implementation requirements.
1379 Although an RCU implementation that ignores these requirements could
1380 still be used, it would likely be subject to limitations that would
1381 make it inappropriate for industrial-strength production use.
1382 Classes of quality-of-implementation requirements are as follows:
1385 <li> <a href="#Specialization">Specialization</a>
1386 <li> <a href="#Performance and Scalability">Performance and Scalability</a>
1387 <li> <a href="#Composability">Composability</a>
1388 <li> <a href="#Corner Cases">Corner Cases</a>
1392 These classes is covered in the following sections.
1394 <h3><a name="Specialization">Specialization</a></h3>
1397 RCU is and always has been intended primarily for read-mostly situations,
1398 which means that RCU's read-side primitives are optimized, often at the
1399 expense of its update-side primitives.
1400 Experience thus far is captured by the following list of situations:
1403 <li> Read-mostly data, where stale and inconsistent data is not
1404 a problem: RCU works great!
1405 <li> Read-mostly data, where data must be consistent:
1407 <li> Read-write data, where data must be consistent:
1408 RCU <i>might</i> work OK.
1410 <li> Write-mostly data, where data must be consistent:
1411 RCU is very unlikely to be the right tool for the job,
1412 with the following exceptions, where RCU can provide:
1414 <li> Existence guarantees for update-friendly mechanisms.
1415 <li> Wait-free read-side primitives for real-time use.
1420 This focus on read-mostly situations means that RCU must interoperate
1421 with other synchronization primitives.
1422 For example, the <tt>add_gp()</tt> and <tt>remove_gp_synchronous()</tt>
1423 examples discussed earlier use RCU to protect readers and locking to
1424 coordinate updaters.
1425 However, the need extends much farther, requiring that a variety of
1426 synchronization primitives be legal within RCU read-side critical sections,
1427 including spinlocks, sequence locks, atomic operations, reference
1428 counters, and memory barriers.
1431 <tr><th> </th></tr>
1432 <tr><th align="left">Quick Quiz:</th></tr>
1434 What about sleeping locks?
1436 <tr><th align="left">Answer:</th></tr>
1437 <tr><td bgcolor="#ffffff"><font color="ffffff">
1438 These are forbidden within Linux-kernel RCU read-side critical
1439 sections because it is not legal to place a quiescent state
1440 (in this case, voluntary context switch) within an RCU read-side
1442 However, sleeping locks may be used within userspace RCU read-side
1443 critical sections, and also within Linux-kernel sleepable RCU
1444 <a href="#Sleepable RCU"><font color="ffffff">(SRCU)</font></a>
1445 read-side critical sections.
1446 In addition, the -rt patchset turns spinlocks into a
1447 sleeping locks so that the corresponding critical sections
1448 can be preempted, which also means that these sleeplockified
1449 spinlocks (but not other sleeping locks!) may be acquire within
1450 -rt-Linux-kernel RCU read-side critical sections.
1453 <p><font color="ffffff">
1454 Note that it <i>is</i> legal for a normal RCU read-side
1455 critical section to conditionally acquire a sleeping locks
1456 (as in <tt>mutex_trylock()</tt>), but only as long as it does
1457 not loop indefinitely attempting to conditionally acquire that
1459 The key point is that things like <tt>mutex_trylock()</tt>
1460 either return with the mutex held, or return an error indication if
1461 the mutex was not immediately available.
1462 Either way, <tt>mutex_trylock()</tt> returns immediately without
1465 <tr><td> </td></tr>
1469 It often comes as a surprise that many algorithms do not require a
1470 consistent view of data, but many can function in that mode,
1471 with network routing being the poster child.
1472 Internet routing algorithms take significant time to propagate
1473 updates, so that by the time an update arrives at a given system,
1474 that system has been sending network traffic the wrong way for
1475 a considerable length of time.
1476 Having a few threads continue to send traffic the wrong way for a
1477 few more milliseconds is clearly not a problem: In the worst case,
1478 TCP retransmissions will eventually get the data where it needs to go.
1479 In general, when tracking the state of the universe outside of the
1480 computer, some level of inconsistency must be tolerated due to
1481 speed-of-light delays if nothing else.
1484 Furthermore, uncertainty about external state is inherent in many cases.
1485 For example, a pair of veterinarians might use heartbeat to determine
1486 whether or not a given cat was alive.
1487 But how long should they wait after the last heartbeat to decide that
1488 the cat is in fact dead?
1489 Waiting less than 400 milliseconds makes no sense because this would
1490 mean that a relaxed cat would be considered to cycle between death
1491 and life more than 100 times per minute.
1492 Moreover, just as with human beings, a cat's heart might stop for
1493 some period of time, so the exact wait period is a judgment call.
1494 One of our pair of veterinarians might wait 30 seconds before pronouncing
1495 the cat dead, while the other might insist on waiting a full minute.
1496 The two veterinarians would then disagree on the state of the cat during
1497 the final 30 seconds of the minute following the last heartbeat.
1500 Interestingly enough, this same situation applies to hardware.
1501 When push comes to shove, how do we tell whether or not some
1502 external server has failed?
1503 We send messages to it periodically, and declare it failed if we
1504 don't receive a response within a given period of time.
1505 Policy decisions can usually tolerate short
1506 periods of inconsistency.
1507 The policy was decided some time ago, and is only now being put into
1508 effect, so a few milliseconds of delay is normally inconsequential.
1511 However, there are algorithms that absolutely must see consistent data.
1512 For example, the translation between a user-level SystemV semaphore
1513 ID to the corresponding in-kernel data structure is protected by RCU,
1514 but it is absolutely forbidden to update a semaphore that has just been
1516 In the Linux kernel, this need for consistency is accommodated by acquiring
1517 spinlocks located in the in-kernel data structure from within
1518 the RCU read-side critical section, and this is indicated by the
1519 green box in the figure above.
1520 Many other techniques may be used, and are in fact used within the
1524 In short, RCU is not required to maintain consistency, and other
1525 mechanisms may be used in concert with RCU when consistency is required.
1526 RCU's specialization allows it to do its job extremely well, and its
1527 ability to interoperate with other synchronization mechanisms allows
1528 the right mix of synchronization tools to be used for a given job.
1530 <h3><a name="Performance and Scalability">Performance and Scalability</a></h3>
1533 Energy efficiency is a critical component of performance today,
1534 and Linux-kernel RCU implementations must therefore avoid unnecessarily
1535 awakening idle CPUs.
1536 I cannot claim that this requirement was premeditated.
1537 In fact, I learned of it during a telephone conversation in which I
1538 was given “frank and open” feedback on the importance
1539 of energy efficiency in battery-powered systems and on specific
1540 energy-efficiency shortcomings of the Linux-kernel RCU implementation.
1541 In my experience, the battery-powered embedded community will consider
1542 any unnecessary wakeups to be extremely unfriendly acts.
1543 So much so that mere Linux-kernel-mailing-list posts are
1544 insufficient to vent their ire.
1547 Memory consumption is not particularly important for in most
1548 situations, and has become decreasingly
1549 so as memory sizes have expanded and memory
1550 costs have plummeted.
1551 However, as I learned from Matt Mackall's
1552 <a href="http://elinux.org/Linux_Tiny-FAQ">bloatwatch</a>
1553 efforts, memory footprint is critically important on single-CPU systems with
1554 non-preemptible (<tt>CONFIG_PREEMPT=n</tt>) kernels, and thus
1555 <a href="https://lkml.kernel.org/g/20090113221724.GA15307@linux.vnet.ibm.com">tiny RCU</a>
1557 Josh Triplett has since taken over the small-memory banner with his
1558 <a href="https://tiny.wiki.kernel.org/">Linux kernel tinification</a>
1559 project, which resulted in
1560 <a href="#Sleepable RCU">SRCU</a>
1561 becoming optional for those kernels not needing it.
1564 The remaining performance requirements are, for the most part,
1566 For example, in keeping with RCU's read-side specialization,
1567 <tt>rcu_dereference()</tt> should have negligible overhead (for
1568 example, suppression of a few minor compiler optimizations).
1569 Similarly, in non-preemptible environments, <tt>rcu_read_lock()</tt> and
1570 <tt>rcu_read_unlock()</tt> should have exactly zero overhead.
1573 In preemptible environments, in the case where the RCU read-side
1574 critical section was not preempted (as will be the case for the
1575 highest-priority real-time process), <tt>rcu_read_lock()</tt> and
1576 <tt>rcu_read_unlock()</tt> should have minimal overhead.
1577 In particular, they should not contain atomic read-modify-write
1578 operations, memory-barrier instructions, preemption disabling,
1579 interrupt disabling, or backwards branches.
1580 However, in the case where the RCU read-side critical section was preempted,
1581 <tt>rcu_read_unlock()</tt> may acquire spinlocks and disable interrupts.
1582 This is why it is better to nest an RCU read-side critical section
1583 within a preempt-disable region than vice versa, at least in cases
1584 where that critical section is short enough to avoid unduly degrading
1585 real-time latencies.
1588 The <tt>synchronize_rcu()</tt> grace-period-wait primitive is
1589 optimized for throughput.
1590 It may therefore incur several milliseconds of latency in addition to
1591 the duration of the longest RCU read-side critical section.
1592 On the other hand, multiple concurrent invocations of
1593 <tt>synchronize_rcu()</tt> are required to use batching optimizations
1594 so that they can be satisfied by a single underlying grace-period-wait
1596 For example, in the Linux kernel, it is not unusual for a single
1597 grace-period-wait operation to serve more than
1598 <a href="https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response">1,000 separate invocations</a>
1599 of <tt>synchronize_rcu()</tt>, thus amortizing the per-invocation
1600 overhead down to nearly zero.
1601 However, the grace-period optimization is also required to avoid
1602 measurable degradation of real-time scheduling and interrupt latencies.
1605 In some cases, the multi-millisecond <tt>synchronize_rcu()</tt>
1606 latencies are unacceptable.
1607 In these cases, <tt>synchronize_rcu_expedited()</tt> may be used
1608 instead, reducing the grace-period latency down to a few tens of
1609 microseconds on small systems, at least in cases where the RCU read-side
1610 critical sections are short.
1611 There are currently no special latency requirements for
1612 <tt>synchronize_rcu_expedited()</tt> on large systems, but,
1613 consistent with the empirical nature of the RCU specification,
1614 that is subject to change.
1615 However, there most definitely are scalability requirements:
1616 A storm of <tt>synchronize_rcu_expedited()</tt> invocations on 4096
1617 CPUs should at least make reasonable forward progress.
1618 In return for its shorter latencies, <tt>synchronize_rcu_expedited()</tt>
1619 is permitted to impose modest degradation of real-time latency
1620 on non-idle online CPUs.
1621 Here, “modest” means roughly the same latency
1622 degradation as a scheduling-clock interrupt.
1625 There are a number of situations where even
1626 <tt>synchronize_rcu_expedited()</tt>'s reduced grace-period
1627 latency is unacceptable.
1628 In these situations, the asynchronous <tt>call_rcu()</tt> can be
1629 used in place of <tt>synchronize_rcu()</tt> as follows:
1636 4 struct rcu_head rh;
1639 7 static void remove_gp_cb(struct rcu_head *rhp)
1641 9 struct foo *p = container_of(rhp, struct foo, rh);
1646 14 bool remove_gp_asynchronous(void)
1650 18 spin_lock(&gp_lock);
1651 19 p = rcu_dereference(gp);
1653 21 spin_unlock(&gp_lock);
1656 24 rcu_assign_pointer(gp, NULL);
1657 25 call_rcu(&p->rh, remove_gp_cb);
1658 26 spin_unlock(&gp_lock);
1665 A definition of <tt>struct foo</tt> is finally needed, and appears
1667 The function <tt>remove_gp_cb()</tt> is passed to <tt>call_rcu()</tt>
1668 on line 25, and will be invoked after the end of a subsequent
1670 This gets the same effect as <tt>remove_gp_synchronous()</tt>,
1671 but without forcing the updater to wait for a grace period to elapse.
1672 The <tt>call_rcu()</tt> function may be used in a number of
1673 situations where neither <tt>synchronize_rcu()</tt> nor
1674 <tt>synchronize_rcu_expedited()</tt> would be legal,
1675 including within preempt-disable code, <tt>local_bh_disable()</tt> code,
1676 interrupt-disable code, and interrupt handlers.
1677 However, even <tt>call_rcu()</tt> is illegal within NMI handlers
1678 and from idle and offline CPUs.
1679 The callback function (<tt>remove_gp_cb()</tt> in this case) will be
1680 executed within softirq (software interrupt) environment within the
1682 either within a real softirq handler or under the protection
1683 of <tt>local_bh_disable()</tt>.
1684 In both the Linux kernel and in userspace, it is bad practice to
1685 write an RCU callback function that takes too long.
1686 Long-running operations should be relegated to separate threads or
1687 (in the Linux kernel) workqueues.
1690 <tr><th> </th></tr>
1691 <tr><th align="left">Quick Quiz:</th></tr>
1693 Why does line 19 use <tt>rcu_access_pointer()</tt>?
1694 After all, <tt>call_rcu()</tt> on line 25 stores into the
1695 structure, which would interact badly with concurrent insertions.
1696 Doesn't this mean that <tt>rcu_dereference()</tt> is required?
1698 <tr><th align="left">Answer:</th></tr>
1699 <tr><td bgcolor="#ffffff"><font color="ffffff">
1700 Presumably the <tt>->gp_lock</tt> acquired on line 18 excludes
1701 any changes, including any insertions that <tt>rcu_dereference()</tt>
1702 would protect against.
1703 Therefore, any insertions will be delayed until after
1704 <tt>->gp_lock</tt>
1705 is released on line 25, which in turn means that
1706 <tt>rcu_access_pointer()</tt> suffices.
1708 <tr><td> </td></tr>
1712 However, all that <tt>remove_gp_cb()</tt> is doing is
1713 invoking <tt>kfree()</tt> on the data element.
1714 This is a common idiom, and is supported by <tt>kfree_rcu()</tt>,
1715 which allows “fire and forget” operation as shown below:
1722 4 struct rcu_head rh;
1725 7 bool remove_gp_faf(void)
1729 11 spin_lock(&gp_lock);
1730 12 p = rcu_dereference(gp);
1732 14 spin_unlock(&gp_lock);
1735 17 rcu_assign_pointer(gp, NULL);
1736 18 kfree_rcu(p, rh);
1737 19 spin_unlock(&gp_lock);
1744 Note that <tt>remove_gp_faf()</tt> simply invokes
1745 <tt>kfree_rcu()</tt> and proceeds, without any need to pay any
1746 further attention to the subsequent grace period and <tt>kfree()</tt>.
1747 It is permissible to invoke <tt>kfree_rcu()</tt> from the same
1748 environments as for <tt>call_rcu()</tt>.
1749 Interestingly enough, DYNIX/ptx had the equivalents of
1750 <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>, but not
1751 <tt>synchronize_rcu()</tt>.
1752 This was due to the fact that RCU was not heavily used within DYNIX/ptx,
1753 so the very few places that needed something like
1754 <tt>synchronize_rcu()</tt> simply open-coded it.
1757 <tr><th> </th></tr>
1758 <tr><th align="left">Quick Quiz:</th></tr>
1760 Earlier it was claimed that <tt>call_rcu()</tt> and
1761 <tt>kfree_rcu()</tt> allowed updaters to avoid being blocked
1763 But how can that be correct, given that the invocation of the callback
1764 and the freeing of the memory (respectively) must still wait for
1765 a grace period to elapse?
1767 <tr><th align="left">Answer:</th></tr>
1768 <tr><td bgcolor="#ffffff"><font color="ffffff">
1769 We could define things this way, but keep in mind that this sort of
1770 definition would say that updates in garbage-collected languages
1771 cannot complete until the next time the garbage collector runs,
1772 which does not seem at all reasonable.
1773 The key point is that in most cases, an updater using either
1774 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> can proceed to the
1775 next update as soon as it has invoked <tt>call_rcu()</tt> or
1776 <tt>kfree_rcu()</tt>, without having to wait for a subsequent
1779 <tr><td> </td></tr>
1783 But what if the updater must wait for the completion of code to be
1784 executed after the end of the grace period, but has other tasks
1785 that can be carried out in the meantime?
1786 The polling-style <tt>get_state_synchronize_rcu()</tt> and
1787 <tt>cond_synchronize_rcu()</tt> functions may be used for this
1788 purpose, as shown below:
1792 1 bool remove_gp_poll(void)
1797 6 spin_lock(&gp_lock);
1798 7 p = rcu_access_pointer(gp);
1800 9 spin_unlock(&gp_lock);
1803 12 rcu_assign_pointer(gp, NULL);
1804 13 spin_unlock(&gp_lock);
1805 14 s = get_state_synchronize_rcu();
1806 15 do_something_while_waiting();
1807 16 cond_synchronize_rcu(s);
1815 On line 14, <tt>get_state_synchronize_rcu()</tt> obtains a
1816 “cookie” from RCU,
1817 then line 15 carries out other tasks,
1818 and finally, line 16 returns immediately if a grace period has
1819 elapsed in the meantime, but otherwise waits as required.
1820 The need for <tt>get_state_synchronize_rcu</tt> and
1821 <tt>cond_synchronize_rcu()</tt> has appeared quite recently,
1822 so it is too early to tell whether they will stand the test of time.
1825 RCU thus provides a range of tools to allow updaters to strike the
1826 required tradeoff between latency, flexibility and CPU overhead.
1828 <h3><a name="Composability">Composability</a></h3>
1831 Composability has received much attention in recent years, perhaps in part
1832 due to the collision of multicore hardware with object-oriented techniques
1833 designed in single-threaded environments for single-threaded use.
1834 And in theory, RCU read-side critical sections may be composed, and in
1835 fact may be nested arbitrarily deeply.
1836 In practice, as with all real-world implementations of composable
1837 constructs, there are limitations.
1840 Implementations of RCU for which <tt>rcu_read_lock()</tt>
1841 and <tt>rcu_read_unlock()</tt> generate no code, such as
1842 Linux-kernel RCU when <tt>CONFIG_PREEMPT=n</tt>, can be
1843 nested arbitrarily deeply.
1844 After all, there is no overhead.
1845 Except that if all these instances of <tt>rcu_read_lock()</tt>
1846 and <tt>rcu_read_unlock()</tt> are visible to the compiler,
1847 compilation will eventually fail due to exhausting memory,
1848 mass storage, or user patience, whichever comes first.
1849 If the nesting is not visible to the compiler, as is the case with
1850 mutually recursive functions each in its own translation unit,
1851 stack overflow will result.
1852 If the nesting takes the form of loops, either the control variable
1853 will overflow or (in the Linux kernel) you will get an RCU CPU stall warning.
1854 Nevertheless, this class of RCU implementations is one
1855 of the most composable constructs in existence.
1858 RCU implementations that explicitly track nesting depth
1859 are limited by the nesting-depth counter.
1860 For example, the Linux kernel's preemptible RCU limits nesting to
1862 This should suffice for almost all practical purposes.
1863 That said, a consecutive pair of RCU read-side critical sections
1864 between which there is an operation that waits for a grace period
1865 cannot be enclosed in another RCU read-side critical section.
1866 This is because it is not legal to wait for a grace period within
1867 an RCU read-side critical section: To do so would result either
1869 in RCU implicitly splitting the enclosing RCU read-side critical
1870 section, neither of which is conducive to a long-lived and prosperous
1874 It is worth noting that RCU is not alone in limiting composability.
1875 For example, many transactional-memory implementations prohibit
1876 composing a pair of transactions separated by an irrevocable
1877 operation (for example, a network receive operation).
1878 For another example, lock-based critical sections can be composed
1879 surprisingly freely, but only if deadlock is avoided.
1882 In short, although RCU read-side critical sections are highly composable,
1883 care is required in some situations, just as is the case for any other
1884 composable synchronization mechanism.
1886 <h3><a name="Corner Cases">Corner Cases</a></h3>
1889 A given RCU workload might have an endless and intense stream of
1890 RCU read-side critical sections, perhaps even so intense that there
1891 was never a point in time during which there was not at least one
1892 RCU read-side critical section in flight.
1893 RCU cannot allow this situation to block grace periods: As long as
1894 all the RCU read-side critical sections are finite, grace periods
1895 must also be finite.
1898 That said, preemptible RCU implementations could potentially result
1899 in RCU read-side critical sections being preempted for long durations,
1900 which has the effect of creating a long-duration RCU read-side
1902 This situation can arise only in heavily loaded systems, but systems using
1903 real-time priorities are of course more vulnerable.
1904 Therefore, RCU priority boosting is provided to help deal with this
1906 That said, the exact requirements on RCU priority boosting will likely
1907 evolve as more experience accumulates.
1910 Other workloads might have very high update rates.
1911 Although one can argue that such workloads should instead use
1912 something other than RCU, the fact remains that RCU must
1913 handle such workloads gracefully.
1914 This requirement is another factor driving batching of grace periods,
1915 but it is also the driving force behind the checks for large numbers
1916 of queued RCU callbacks in the <tt>call_rcu()</tt> code path.
1917 Finally, high update rates should not delay RCU read-side critical
1918 sections, although some small read-side delays can occur when using
1919 <tt>synchronize_rcu_expedited()</tt>, courtesy of this function's use
1920 of <tt>smp_call_function_single()</tt>.
1923 Although all three of these corner cases were understood in the early
1924 1990s, a simple user-level test consisting of <tt>close(open(path))</tt>
1926 in the early 2000s suddenly provided a much deeper appreciation of the
1927 high-update-rate corner case.
1928 This test also motivated addition of some RCU code to react to high update
1929 rates, for example, if a given CPU finds itself with more than 10,000
1930 RCU callbacks queued, it will cause RCU to take evasive action by
1931 more aggressively starting grace periods and more aggressively forcing
1932 completion of grace-period processing.
1933 This evasive action causes the grace period to complete more quickly,
1934 but at the cost of restricting RCU's batching optimizations, thus
1935 increasing the CPU overhead incurred by that grace period.
1937 <h2><a name="Software-Engineering Requirements">
1938 Software-Engineering Requirements</a></h2>
1941 Between Murphy's Law and “To err is human”, it is necessary to
1942 guard against mishaps and misuse:
1945 <li> It is all too easy to forget to use <tt>rcu_read_lock()</tt>
1946 everywhere that it is needed, so kernels built with
1947 <tt>CONFIG_PROVE_RCU=y</tt> will splat if
1948 <tt>rcu_dereference()</tt> is used outside of an
1949 RCU read-side critical section.
1950 Update-side code can use <tt>rcu_dereference_protected()</tt>,
1952 <a href="https://lwn.net/Articles/371986/">lockdep expression</a>
1953 to indicate what is providing the protection.
1954 If the indicated protection is not provided, a lockdep splat
1958 Code shared between readers and updaters can use
1959 <tt>rcu_dereference_check()</tt>, which also takes a
1960 lockdep expression, and emits a lockdep splat if neither
1961 <tt>rcu_read_lock()</tt> nor the indicated protection
1963 In addition, <tt>rcu_dereference_raw()</tt> is used in those
1964 (hopefully rare) cases where the required protection cannot
1965 be easily described.
1966 Finally, <tt>rcu_read_lock_held()</tt> is provided to
1967 allow a function to verify that it has been invoked within
1968 an RCU read-side critical section.
1969 I was made aware of this set of requirements shortly after Thomas
1970 Gleixner audited a number of RCU uses.
1971 <li> A given function might wish to check for RCU-related preconditions
1972 upon entry, before using any other RCU API.
1973 The <tt>rcu_lockdep_assert()</tt> does this job,
1974 asserting the expression in kernels having lockdep enabled
1975 and doing nothing otherwise.
1976 <li> It is also easy to forget to use <tt>rcu_assign_pointer()</tt>
1977 and <tt>rcu_dereference()</tt>, perhaps (incorrectly)
1978 substituting a simple assignment.
1979 To catch this sort of error, a given RCU-protected pointer may be
1980 tagged with <tt>__rcu</tt>, after which running sparse
1981 with <tt>CONFIG_SPARSE_RCU_POINTER=y</tt> will complain
1982 about simple-assignment accesses to that pointer.
1983 Arnd Bergmann made me aware of this requirement, and also
1985 <a href="https://lwn.net/Articles/376011/">patch series</a>.
1986 <li> Kernels built with <tt>CONFIG_DEBUG_OBJECTS_RCU_HEAD=y</tt>
1987 will splat if a data element is passed to <tt>call_rcu()</tt>
1988 twice in a row, without a grace period in between.
1989 (This error is similar to a double free.)
1990 The corresponding <tt>rcu_head</tt> structures that are
1991 dynamically allocated are automatically tracked, but
1992 <tt>rcu_head</tt> structures allocated on the stack
1993 must be initialized with <tt>init_rcu_head_on_stack()</tt>
1994 and cleaned up with <tt>destroy_rcu_head_on_stack()</tt>.
1995 Similarly, statically allocated non-stack <tt>rcu_head</tt>
1996 structures must be initialized with <tt>init_rcu_head()</tt>
1997 and cleaned up with <tt>destroy_rcu_head()</tt>.
1998 Mathieu Desnoyers made me aware of this requirement, and also
2000 <a href="https://lkml.kernel.org/g/20100319013024.GA28456@Krystal">patch</a>.
2001 <li> An infinite loop in an RCU read-side critical section will
2002 eventually trigger an RCU CPU stall warning splat, with
2003 the duration of “eventually” being controlled by the
2004 <tt>RCU_CPU_STALL_TIMEOUT</tt> <tt>Kconfig</tt> option, or,
2005 alternatively, by the
2006 <tt>rcupdate.rcu_cpu_stall_timeout</tt> boot/sysfs
2008 However, RCU is not obligated to produce this splat
2009 unless there is a grace period waiting on that particular
2010 RCU read-side critical section.
2012 Some extreme workloads might intentionally delay
2013 RCU grace periods, and systems running those workloads can
2014 be booted with <tt>rcupdate.rcu_cpu_stall_suppress</tt>
2015 to suppress the splats.
2016 This kernel parameter may also be set via <tt>sysfs</tt>.
2017 Furthermore, RCU CPU stall warnings are counter-productive
2018 during sysrq dumps and during panics.
2019 RCU therefore supplies the <tt>rcu_sysrq_start()</tt> and
2020 <tt>rcu_sysrq_end()</tt> API members to be called before
2021 and after long sysrq dumps.
2022 RCU also supplies the <tt>rcu_panic()</tt> notifier that is
2023 automatically invoked at the beginning of a panic to suppress
2024 further RCU CPU stall warnings.
2027 This requirement made itself known in the early 1990s, pretty
2028 much the first time that it was necessary to debug a CPU stall.
2029 That said, the initial implementation in DYNIX/ptx was quite
2030 generic in comparison with that of Linux.
2031 <li> Although it would be very good to detect pointers leaking out
2032 of RCU read-side critical sections, there is currently no
2033 good way of doing this.
2034 One complication is the need to distinguish between pointers
2035 leaking and pointers that have been handed off from RCU to
2036 some other synchronization mechanism, for example, reference
2038 <li> In kernels built with <tt>CONFIG_RCU_TRACE=y</tt>, RCU-related
2039 information is provided via both debugfs and event tracing.
2040 <li> Open-coded use of <tt>rcu_assign_pointer()</tt> and
2041 <tt>rcu_dereference()</tt> to create typical linked
2042 data structures can be surprisingly error-prone.
2043 Therefore, RCU-protected
2044 <a href="https://lwn.net/Articles/609973/#RCU List APIs">linked lists</a>
2045 and, more recently, RCU-protected
2046 <a href="https://lwn.net/Articles/612100/">hash tables</a>
2048 Many other special-purpose RCU-protected data structures are
2049 available in the Linux kernel and the userspace RCU library.
2050 <li> Some linked structures are created at compile time, but still
2051 require <tt>__rcu</tt> checking.
2052 The <tt>RCU_POINTER_INITIALIZER()</tt> macro serves this
2054 <li> It is not necessary to use <tt>rcu_assign_pointer()</tt>
2055 when creating linked structures that are to be published via
2056 a single external pointer.
2057 The <tt>RCU_INIT_POINTER()</tt> macro is provided for
2058 this task and also for assigning <tt>NULL</tt> pointers
2063 This not a hard-and-fast list: RCU's diagnostic capabilities will
2064 continue to be guided by the number and type of usage bugs found
2065 in real-world RCU usage.
2067 <h2><a name="Linux Kernel Complications">Linux Kernel Complications</a></h2>
2070 The Linux kernel provides an interesting environment for all kinds of
2071 software, including RCU.
2072 Some of the relevant points of interest are as follows:
2075 <li> <a href="#Configuration">Configuration</a>.
2076 <li> <a href="#Firmware Interface">Firmware Interface</a>.
2077 <li> <a href="#Early Boot">Early Boot</a>.
2078 <li> <a href="#Interrupts and NMIs">
2079 Interrupts and non-maskable interrupts (NMIs)</a>.
2080 <li> <a href="#Loadable Modules">Loadable Modules</a>.
2081 <li> <a href="#Hotplug CPU">Hotplug CPU</a>.
2082 <li> <a href="#Scheduler and RCU">Scheduler and RCU</a>.
2083 <li> <a href="#Tracing and RCU">Tracing and RCU</a>.
2084 <li> <a href="#Energy Efficiency">Energy Efficiency</a>.
2085 <li> <a href="#Memory Efficiency">Memory Efficiency</a>.
2086 <li> <a href="#Performance, Scalability, Response Time, and Reliability">
2087 Performance, Scalability, Response Time, and Reliability</a>.
2091 This list is probably incomplete, but it does give a feel for the
2092 most notable Linux-kernel complications.
2093 Each of the following sections covers one of the above topics.
2095 <h3><a name="Configuration">Configuration</a></h3>
2098 RCU's goal is automatic configuration, so that almost nobody
2099 needs to worry about RCU's <tt>Kconfig</tt> options.
2100 And for almost all users, RCU does in fact work well
2101 “out of the box.”
2104 However, there are specialized use cases that are handled by
2105 kernel boot parameters and <tt>Kconfig</tt> options.
2106 Unfortunately, the <tt>Kconfig</tt> system will explicitly ask users
2107 about new <tt>Kconfig</tt> options, which requires almost all of them
2108 be hidden behind a <tt>CONFIG_RCU_EXPERT</tt> <tt>Kconfig</tt> option.
2111 This all should be quite obvious, but the fact remains that
2112 Linus Torvalds recently had to
2113 <a href="https://lkml.kernel.org/g/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com">remind</a>
2114 me of this requirement.
2116 <h3><a name="Firmware Interface">Firmware Interface</a></h3>
2119 In many cases, kernel obtains information about the system from the
2120 firmware, and sometimes things are lost in translation.
2121 Or the translation is accurate, but the original message is bogus.
2124 For example, some systems' firmware overreports the number of CPUs,
2125 sometimes by a large factor.
2126 If RCU naively believed the firmware, as it used to do,
2127 it would create too many per-CPU kthreads.
2128 Although the resulting system will still run correctly, the extra
2129 kthreads needlessly consume memory and can cause confusion
2130 when they show up in <tt>ps</tt> listings.
2133 RCU must therefore wait for a given CPU to actually come online before
2134 it can allow itself to believe that the CPU actually exists.
2135 The resulting “ghost CPUs” (which are never going to
2136 come online) cause a number of
2137 <a href="https://paulmck.livejournal.com/37494.html">interesting complications</a>.
2139 <h3><a name="Early Boot">Early Boot</a></h3>
2142 The Linux kernel's boot sequence is an interesting process,
2143 and RCU is used early, even before <tt>rcu_init()</tt>
2145 In fact, a number of RCU's primitives can be used as soon as the
2146 initial task's <tt>task_struct</tt> is available and the
2147 boot CPU's per-CPU variables are set up.
2148 The read-side primitives (<tt>rcu_read_lock()</tt>,
2149 <tt>rcu_read_unlock()</tt>, <tt>rcu_dereference()</tt>,
2150 and <tt>rcu_access_pointer()</tt>) will operate normally very early on,
2151 as will <tt>rcu_assign_pointer()</tt>.
2154 Although <tt>call_rcu()</tt> may be invoked at any
2155 time during boot, callbacks are not guaranteed to be invoked until after
2156 all of RCU's kthreads have been spawned, which occurs at
2157 <tt>early_initcall()</tt> time.
2158 This delay in callback invocation is due to the fact that RCU does not
2159 invoke callbacks until it is fully initialized, and this full initialization
2160 cannot occur until after the scheduler has initialized itself to the
2161 point where RCU can spawn and run its kthreads.
2162 In theory, it would be possible to invoke callbacks earlier,
2163 however, this is not a panacea because there would be severe restrictions
2164 on what operations those callbacks could invoke.
2167 Perhaps surprisingly, <tt>synchronize_rcu()</tt>,
2168 <a href="#Bottom-Half Flavor"><tt>synchronize_rcu_bh()</tt></a>
2169 (<a href="#Bottom-Half Flavor">discussed below</a>),
2170 <a href="#Sched Flavor"><tt>synchronize_sched()</tt></a>,
2171 <tt>synchronize_rcu_expedited()</tt>,
2172 <tt>synchronize_rcu_bh_expedited()</tt>, and
2173 <tt>synchronize_sched_expedited()</tt>
2174 will all operate normally
2175 during very early boot, the reason being that there is only one CPU
2176 and preemption is disabled.
2177 This means that the call <tt>synchronize_rcu()</tt> (or friends)
2178 itself is a quiescent
2179 state and thus a grace period, so the early-boot implementation can
2183 However, once the scheduler has spawned its first kthread, this early
2184 boot trick fails for <tt>synchronize_rcu()</tt> (as well as for
2185 <tt>synchronize_rcu_expedited()</tt>) in <tt>CONFIG_PREEMPT=y</tt>
2187 The reason is that an RCU read-side critical section might be preempted,
2188 which means that a subsequent <tt>synchronize_rcu()</tt> really does have
2189 to wait for something, as opposed to simply returning immediately.
2190 Unfortunately, <tt>synchronize_rcu()</tt> can't do this until all of
2191 its kthreads are spawned, which doesn't happen until some time during
2192 <tt>early_initcalls()</tt> time.
2193 But this is no excuse: RCU is nevertheless required to correctly handle
2194 synchronous grace periods during this time period.
2195 Once all of its kthreads are up and running, RCU starts running
2199 <tr><th> </th></tr>
2200 <tr><th align="left">Quick Quiz:</th></tr>
2202 How can RCU possibly handle grace periods before all of its
2203 kthreads have been spawned???
2205 <tr><th align="left">Answer:</th></tr>
2206 <tr><td bgcolor="#ffffff"><font color="ffffff">
2210 <p><font color="ffffff">
2211 During the “dead zone” between the time that the
2212 scheduler spawns the first task and the time that all of RCU's
2213 kthreads have been spawned, all synchronous grace periods are
2214 handled by the expedited grace-period mechanism.
2215 At runtime, this expedited mechanism relies on workqueues, but
2216 during the dead zone the requesting task itself drives the
2217 desired expedited grace period.
2218 Because dead-zone execution takes place within task context,
2220 Once the dead zone ends, expedited grace periods go back to
2221 using workqueues, as is required to avoid problems that would
2222 otherwise occur when a user task received a POSIX signal while
2223 driving an expedited grace period.
2226 <p><font color="ffffff">
2227 And yes, this does mean that it is unhelpful to send POSIX
2228 signals to random tasks between the time that the scheduler
2229 spawns its first kthread and the time that RCU's kthreads
2230 have all been spawned.
2231 If there ever turns out to be a good reason for sending POSIX
2232 signals during that time, appropriate adjustments will be made.
2233 (If it turns out that POSIX signals are sent during this time for
2234 no good reason, other adjustments will be made, appropriate
2237 <tr><td> </td></tr>
2241 I learned of these boot-time requirements as a result of a series of
2244 <h3><a name="Interrupts and NMIs">Interrupts and NMIs</a></h3>
2247 The Linux kernel has interrupts, and RCU read-side critical sections are
2248 legal within interrupt handlers and within interrupt-disabled regions
2249 of code, as are invocations of <tt>call_rcu()</tt>.
2252 Some Linux-kernel architectures can enter an interrupt handler from
2253 non-idle process context, and then just never leave it, instead stealthily
2254 transitioning back to process context.
2255 This trick is sometimes used to invoke system calls from inside the kernel.
2256 These “half-interrupts” mean that RCU has to be very careful
2257 about how it counts interrupt nesting levels.
2258 I learned of this requirement the hard way during a rewrite
2259 of RCU's dyntick-idle code.
2262 The Linux kernel has non-maskable interrupts (NMIs), and
2263 RCU read-side critical sections are legal within NMI handlers.
2264 Thankfully, RCU update-side primitives, including
2265 <tt>call_rcu()</tt>, are prohibited within NMI handlers.
2268 The name notwithstanding, some Linux-kernel architectures
2269 can have nested NMIs, which RCU must handle correctly.
2271 <a href="https://lkml.kernel.org/g/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com">surprised me</a>
2272 with this requirement;
2273 he also kindly surprised me with
2274 <a href="https://lkml.kernel.org/g/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com">an algorithm</a>
2275 that meets this requirement.
2277 <h3><a name="Loadable Modules">Loadable Modules</a></h3>
2280 The Linux kernel has loadable modules, and these modules can
2282 After a given module has been unloaded, any attempt to call
2283 one of its functions results in a segmentation fault.
2284 The module-unload functions must therefore cancel any
2285 delayed calls to loadable-module functions, for example,
2286 any outstanding <tt>mod_timer()</tt> must be dealt with
2287 via <tt>del_timer_sync()</tt> or similar.
2290 Unfortunately, there is no way to cancel an RCU callback;
2291 once you invoke <tt>call_rcu()</tt>, the callback function is
2292 going to eventually be invoked, unless the system goes down first.
2293 Because it is normally considered socially irresponsible to crash the system
2294 in response to a module unload request, we need some other way
2295 to deal with in-flight RCU callbacks.
2298 RCU therefore provides
2299 <tt><a href="https://lwn.net/Articles/217484/">rcu_barrier()</a></tt>,
2300 which waits until all in-flight RCU callbacks have been invoked.
2301 If a module uses <tt>call_rcu()</tt>, its exit function should therefore
2302 prevent any future invocation of <tt>call_rcu()</tt>, then invoke
2303 <tt>rcu_barrier()</tt>.
2304 In theory, the underlying module-unload code could invoke
2305 <tt>rcu_barrier()</tt> unconditionally, but in practice this would
2306 incur unacceptable latencies.
2309 Nikita Danilov noted this requirement for an analogous filesystem-unmount
2310 situation, and Dipankar Sarma incorporated <tt>rcu_barrier()</tt> into RCU.
2311 The need for <tt>rcu_barrier()</tt> for module unloading became
2315 <b>Important note</b>: The <tt>rcu_barrier()</tt> function is not,
2316 repeat, <i>not</i>, obligated to wait for a grace period.
2317 It is instead only required to wait for RCU callbacks that have
2318 already been posted.
2319 Therefore, if there are no RCU callbacks posted anywhere in the system,
2320 <tt>rcu_barrier()</tt> is within its rights to return immediately.
2321 Even if there are callbacks posted, <tt>rcu_barrier()</tt> does not
2322 necessarily need to wait for a grace period.
2325 <tr><th> </th></tr>
2326 <tr><th align="left">Quick Quiz:</th></tr>
2329 Each RCU callbacks must wait for a grace period to complete,
2330 and <tt>rcu_barrier()</tt> must wait for each pre-existing
2331 callback to be invoked.
2332 Doesn't <tt>rcu_barrier()</tt> therefore need to wait for
2333 a full grace period if there is even one callback posted anywhere
2336 <tr><th align="left">Answer:</th></tr>
2337 <tr><td bgcolor="#ffffff"><font color="ffffff">
2341 <p><font color="ffffff">
2342 Yes, each RCU callbacks must wait for a grace period to complete,
2343 but it might well be partly (or even completely) finished waiting
2344 by the time <tt>rcu_barrier()</tt> is invoked.
2345 In that case, <tt>rcu_barrier()</tt> need only wait for the
2346 remaining portion of the grace period to elapse.
2347 So even if there are quite a few callbacks posted,
2348 <tt>rcu_barrier()</tt> might well return quite quickly.
2351 <p><font color="ffffff">
2352 So if you need to wait for a grace period as well as for all
2353 pre-existing callbacks, you will need to invoke both
2354 <tt>synchronize_rcu()</tt> and <tt>rcu_barrier()</tt>.
2355 If latency is a concern, you can always use workqueues
2356 to invoke them concurrently.
2358 <tr><td> </td></tr>
2361 <h3><a name="Hotplug CPU">Hotplug CPU</a></h3>
2364 The Linux kernel supports CPU hotplug, which means that CPUs
2366 It is of course illegal to use any RCU API member from an offline CPU,
2367 with the exception of <a href="#Sleepable RCU">SRCU</a> read-side
2369 This requirement was present from day one in DYNIX/ptx, but
2370 on the other hand, the Linux kernel's CPU-hotplug implementation
2371 is “interesting.”
2374 The Linux-kernel CPU-hotplug implementation has notifiers that
2375 are used to allow the various kernel subsystems (including RCU)
2376 to respond appropriately to a given CPU-hotplug operation.
2377 Most RCU operations may be invoked from CPU-hotplug notifiers,
2378 including even synchronous grace-period operations such as
2379 <tt>synchronize_rcu()</tt> and <tt>synchronize_rcu_expedited()</tt>.
2382 However, all-callback-wait operations such as
2383 <tt>rcu_barrier()</tt> are also not supported, due to the
2384 fact that there are phases of CPU-hotplug operations where
2385 the outgoing CPU's callbacks will not be invoked until after
2386 the CPU-hotplug operation ends, which could also result in deadlock.
2387 Furthermore, <tt>rcu_barrier()</tt> blocks CPU-hotplug operations
2388 during its execution, which results in another type of deadlock
2389 when invoked from a CPU-hotplug notifier.
2391 <h3><a name="Scheduler and RCU">Scheduler and RCU</a></h3>
2394 RCU depends on the scheduler, and the scheduler uses RCU to
2395 protect some of its data structures.
2396 This means the scheduler is forbidden from acquiring
2397 the runqueue locks and the priority-inheritance locks
2398 in the middle of an outermost RCU read-side critical section unless either
2399 (1) it releases them before exiting that same
2400 RCU read-side critical section, or
2401 (2) interrupts are disabled across
2402 that entire RCU read-side critical section.
2403 This same prohibition also applies (recursively!) to any lock that is acquired
2404 while holding any lock to which this prohibition applies.
2405 Adhering to this rule prevents preemptible RCU from invoking
2406 <tt>rcu_read_unlock_special()</tt> while either runqueue or
2407 priority-inheritance locks are held, thus avoiding deadlock.
2410 Prior to v4.4, it was only necessary to disable preemption across
2411 RCU read-side critical sections that acquired scheduler locks.
2412 In v4.4, expedited grace periods started using IPIs, and these
2413 IPIs could force a <tt>rcu_read_unlock()</tt> to take the slowpath.
2414 Therefore, this expedited-grace-period change required disabling of
2415 interrupts, not just preemption.
2418 For RCU's part, the preemptible-RCU <tt>rcu_read_unlock()</tt>
2419 implementation must be written carefully to avoid similar deadlocks.
2420 In particular, <tt>rcu_read_unlock()</tt> must tolerate an
2421 interrupt where the interrupt handler invokes both
2422 <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
2423 This possibility requires <tt>rcu_read_unlock()</tt> to use
2424 negative nesting levels to avoid destructive recursion via
2425 interrupt handler's use of RCU.
2428 This pair of mutual scheduler-RCU requirements came as a
2429 <a href="https://lwn.net/Articles/453002/">complete surprise</a>.
2432 As noted above, RCU makes use of kthreads, and it is necessary to
2433 avoid excessive CPU-time accumulation by these kthreads.
2434 This requirement was no surprise, but RCU's violation of it
2435 when running context-switch-heavy workloads when built with
2436 <tt>CONFIG_NO_HZ_FULL=y</tt>
2437 <a href="http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf">did come as a surprise [PDF]</a>.
2438 RCU has made good progress towards meeting this requirement, even
2439 for context-switch-have <tt>CONFIG_NO_HZ_FULL=y</tt> workloads,
2440 but there is room for further improvement.
2442 <h3><a name="Tracing and RCU">Tracing and RCU</a></h3>
2445 It is possible to use tracing on RCU code, but tracing itself
2447 For this reason, <tt>rcu_dereference_raw_notrace()</tt>
2448 is provided for use by tracing, which avoids the destructive
2449 recursion that could otherwise ensue.
2450 This API is also used by virtualization in some architectures,
2451 where RCU readers execute in environments in which tracing
2453 The tracing folks both located the requirement and provided the
2454 needed fix, so this surprise requirement was relatively painless.
2456 <h3><a name="Energy Efficiency">Energy Efficiency</a></h3>
2459 Interrupting idle CPUs is considered socially unacceptable,
2460 especially by people with battery-powered embedded systems.
2461 RCU therefore conserves energy by detecting which CPUs are
2462 idle, including tracking CPUs that have been interrupted from idle.
2463 This is a large part of the energy-efficiency requirement,
2464 so I learned of this via an irate phone call.
2467 Because RCU avoids interrupting idle CPUs, it is illegal to
2468 execute an RCU read-side critical section on an idle CPU.
2469 (Kernels built with <tt>CONFIG_PROVE_RCU=y</tt> will splat
2471 The <tt>RCU_NONIDLE()</tt> macro and <tt>_rcuidle</tt>
2472 event tracing is provided to work around this restriction.
2473 In addition, <tt>rcu_is_watching()</tt> may be used to
2474 test whether or not it is currently legal to run RCU read-side
2475 critical sections on this CPU.
2476 I learned of the need for diagnostics on the one hand
2477 and <tt>RCU_NONIDLE()</tt> on the other while inspecting
2479 Steven Rostedt supplied <tt>_rcuidle</tt> event tracing,
2480 which is used quite heavily in the idle loop.
2481 However, there are some restrictions on the code placed within
2482 <tt>RCU_NONIDLE()</tt>:
2485 <li> Blocking is prohibited.
2486 In practice, this is not a serious restriction given that idle
2487 tasks are prohibited from blocking to begin with.
2488 <li> Although nesting <tt>RCU_NONIDLE()</tt> is permitted, they cannot
2489 nest indefinitely deeply.
2490 However, given that they can be nested on the order of a million
2491 deep, even on 32-bit systems, this should not be a serious
2493 This nesting limit would probably be reached long after the
2494 compiler OOMed or the stack overflowed.
2495 <li> Any code path that enters <tt>RCU_NONIDLE()</tt> must sequence
2496 out of that same <tt>RCU_NONIDLE()</tt>.
2497 For example, the following is grossly illegal:
2503 3 goto bad_idea; /* BUG!!! */
2504 4 do_something_else();});
2510 It is just as illegal to transfer control into the middle of
2511 <tt>RCU_NONIDLE()</tt>'s argument.
2512 Yes, in theory, you could transfer in as long as you also
2513 transferred out, but in practice you could also expect to get sharply
2514 worded review comments.
2518 It is similarly socially unacceptable to interrupt an
2519 <tt>nohz_full</tt> CPU running in userspace.
2520 RCU must therefore track <tt>nohz_full</tt> userspace
2523 <a href="https://lwn.net/Articles/558284/"><tt>CONFIG_NO_HZ_FULL_SYSIDLE=y</tt></a>
2524 kernels, RCU must separately track idle CPUs on the one hand and
2525 CPUs that are either idle or executing in userspace on the other.
2526 In both cases, RCU must be able to sample state at two points in
2527 time, and be able to determine whether or not some other CPU spent
2528 any time idle and/or executing in userspace.
2531 These energy-efficiency requirements have proven quite difficult to
2532 understand and to meet, for example, there have been more than five
2533 clean-sheet rewrites of RCU's energy-efficiency code, the last of
2534 which was finally able to demonstrate
2535 <a href="http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf">real energy savings running on real hardware [PDF]</a>.
2537 I learned of many of these requirements via angry phone calls:
2538 Flaming me on the Linux-kernel mailing list was apparently not
2539 sufficient to fully vent their ire at RCU's energy-efficiency bugs!
2541 <h3><a name="Memory Efficiency">Memory Efficiency</a></h3>
2544 Although small-memory non-realtime systems can simply use Tiny RCU,
2545 code size is only one aspect of memory efficiency.
2546 Another aspect is the size of the <tt>rcu_head</tt> structure
2547 used by <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>.
2548 Although this structure contains nothing more than a pair of pointers,
2549 it does appear in many RCU-protected data structures, including
2550 some that are size critical.
2551 The <tt>page</tt> structure is a case in point, as evidenced by
2552 the many occurrences of the <tt>union</tt> keyword within that structure.
2555 This need for memory efficiency is one reason that RCU uses hand-crafted
2556 singly linked lists to track the <tt>rcu_head</tt> structures that
2557 are waiting for a grace period to elapse.
2558 It is also the reason why <tt>rcu_head</tt> structures do not contain
2559 debug information, such as fields tracking the file and line of the
2560 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> that posted them.
2561 Although this information might appear in debug-only kernel builds at some
2562 point, in the meantime, the <tt>->func</tt> field will often provide
2563 the needed debug information.
2566 However, in some cases, the need for memory efficiency leads to even
2567 more extreme measures.
2568 Returning to the <tt>page</tt> structure, the <tt>rcu_head</tt> field
2569 shares storage with a great many other structures that are used at
2570 various points in the corresponding page's lifetime.
2571 In order to correctly resolve certain
2572 <a href="https://lkml.kernel.org/g/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com">race conditions</a>,
2573 the Linux kernel's memory-management subsystem needs a particular bit
2574 to remain zero during all phases of grace-period processing,
2575 and that bit happens to map to the bottom bit of the
2576 <tt>rcu_head</tt> structure's <tt>->next</tt> field.
2577 RCU makes this guarantee as long as <tt>call_rcu()</tt>
2578 is used to post the callback, as opposed to <tt>kfree_rcu()</tt>
2579 or some future “lazy”
2580 variant of <tt>call_rcu()</tt> that might one day be created for
2581 energy-efficiency purposes.
2584 That said, there are limits.
2585 RCU requires that the <tt>rcu_head</tt> structure be aligned to a
2586 two-byte boundary, and passing a misaligned <tt>rcu_head</tt>
2587 structure to one of the <tt>call_rcu()</tt> family of functions
2588 will result in a splat.
2589 It is therefore necessary to exercise caution when packing
2590 structures containing fields of type <tt>rcu_head</tt>.
2591 Why not a four-byte or even eight-byte alignment requirement?
2592 Because the m68k architecture provides only two-byte alignment,
2593 and thus acts as alignment's least common denominator.
2596 The reason for reserving the bottom bit of pointers to
2597 <tt>rcu_head</tt> structures is to leave the door open to
2598 “lazy” callbacks whose invocations can safely be deferred.
2599 Deferring invocation could potentially have energy-efficiency
2600 benefits, but only if the rate of non-lazy callbacks decreases
2601 significantly for some important workload.
2602 In the meantime, reserving the bottom bit keeps this option open
2603 in case it one day becomes useful.
2605 <h3><a name="Performance, Scalability, Response Time, and Reliability">
2606 Performance, Scalability, Response Time, and Reliability</a></h3>
2610 <a href="#Performance and Scalability">earlier discussion</a>,
2611 RCU is used heavily by hot code paths in performance-critical
2612 portions of the Linux kernel's networking, security, virtualization,
2613 and scheduling code paths.
2614 RCU must therefore use efficient implementations, especially in its
2615 read-side primitives.
2616 To that end, it would be good if preemptible RCU's implementation
2617 of <tt>rcu_read_lock()</tt> could be inlined, however, doing
2618 this requires resolving <tt>#include</tt> issues with the
2619 <tt>task_struct</tt> structure.
2622 The Linux kernel supports hardware configurations with up to
2623 4096 CPUs, which means that RCU must be extremely scalable.
2624 Algorithms that involve frequent acquisitions of global locks or
2625 frequent atomic operations on global variables simply cannot be
2626 tolerated within the RCU implementation.
2627 RCU therefore makes heavy use of a combining tree based on the
2628 <tt>rcu_node</tt> structure.
2629 RCU is required to tolerate all CPUs continuously invoking any
2630 combination of RCU's runtime primitives with minimal per-operation
2632 In fact, in many cases, increasing load must <i>decrease</i> the
2633 per-operation overhead, witness the batching optimizations for
2634 <tt>synchronize_rcu()</tt>, <tt>call_rcu()</tt>,
2635 <tt>synchronize_rcu_expedited()</tt>, and <tt>rcu_barrier()</tt>.
2636 As a general rule, RCU must cheerfully accept whatever the
2637 rest of the Linux kernel decides to throw at it.
2640 The Linux kernel is used for real-time workloads, especially
2641 in conjunction with the
2642 <a href="https://rt.wiki.kernel.org/index.php/Main_Page">-rt patchset</a>.
2643 The real-time-latency response requirements are such that the
2644 traditional approach of disabling preemption across RCU
2645 read-side critical sections is inappropriate.
2646 Kernels built with <tt>CONFIG_PREEMPT=y</tt> therefore
2647 use an RCU implementation that allows RCU read-side critical
2648 sections to be preempted.
2649 This requirement made its presence known after users made it
2650 clear that an earlier
2651 <a href="https://lwn.net/Articles/107930/">real-time patch</a>
2652 did not meet their needs, in conjunction with some
2653 <a href="https://lkml.kernel.org/g/20050318002026.GA2693@us.ibm.com">RCU issues</a>
2654 encountered by a very early version of the -rt patchset.
2657 In addition, RCU must make do with a sub-100-microsecond real-time latency
2659 In fact, on smaller systems with the -rt patchset, the Linux kernel
2660 provides sub-20-microsecond real-time latencies for the whole kernel,
2662 RCU's scalability and latency must therefore be sufficient for
2663 these sorts of configurations.
2664 To my surprise, the sub-100-microsecond real-time latency budget
2665 <a href="http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf">
2666 applies to even the largest systems [PDF]</a>,
2667 up to and including systems with 4096 CPUs.
2668 This real-time requirement motivated the grace-period kthread, which
2669 also simplified handling of a number of race conditions.
2672 RCU must avoid degrading real-time response for CPU-bound threads, whether
2673 executing in usermode (which is one use case for
2674 <tt>CONFIG_NO_HZ_FULL=y</tt>) or in the kernel.
2675 That said, CPU-bound loops in the kernel must execute
2676 <tt>cond_resched_rcu_qs()</tt> at least once per few tens of milliseconds
2677 in order to avoid receiving an IPI from RCU.
2680 Finally, RCU's status as a synchronization primitive means that
2681 any RCU failure can result in arbitrary memory corruption that can be
2682 extremely difficult to debug.
2683 This means that RCU must be extremely reliable, which in
2684 practice also means that RCU must have an aggressive stress-test
2686 This stress-test suite is called <tt>rcutorture</tt>.
2689 Although the need for <tt>rcutorture</tt> was no surprise,
2690 the current immense popularity of the Linux kernel is posing
2691 interesting—and perhaps unprecedented—validation
2693 To see this, keep in mind that there are well over one billion
2694 instances of the Linux kernel running today, given Android
2695 smartphones, Linux-powered televisions, and servers.
2696 This number can be expected to increase sharply with the advent of
2697 the celebrated Internet of Things.
2700 Suppose that RCU contains a race condition that manifests on average
2701 once per million years of runtime.
2702 This bug will be occurring about three times per <i>day</i> across
2704 RCU could simply hide behind hardware error rates, given that no one
2705 should really expect their smartphone to last for a million years.
2706 However, anyone taking too much comfort from this thought should
2707 consider the fact that in most jurisdictions, a successful multi-year
2708 test of a given mechanism, which might include a Linux kernel,
2709 suffices for a number of types of safety-critical certifications.
2710 In fact, rumor has it that the Linux kernel is already being used
2711 in production for safety-critical applications.
2712 I don't know about you, but I would feel quite bad if a bug in RCU
2714 Which might explain my recent focus on validation and verification.
2716 <h2><a name="Other RCU Flavors">Other RCU Flavors</a></h2>
2719 One of the more surprising things about RCU is that there are now
2720 no fewer than five <i>flavors</i>, or API families.
2721 In addition, the primary flavor that has been the sole focus up to
2722 this point has two different implementations, non-preemptible and
2724 The other four flavors are listed below, with requirements for each
2725 described in a separate section.
2728 <li> <a href="#Bottom-Half Flavor">Bottom-Half Flavor</a>
2729 <li> <a href="#Sched Flavor">Sched Flavor</a>
2730 <li> <a href="#Sleepable RCU">Sleepable RCU</a>
2731 <li> <a href="#Tasks RCU">Tasks RCU</a>
2732 <li> <a href="#Waiting for Multiple Grace Periods">
2733 Waiting for Multiple Grace Periods</a>
2736 <h3><a name="Bottom-Half Flavor">Bottom-Half Flavor</a></h3>
2739 The softirq-disable (AKA “bottom-half”,
2740 hence the “_bh” abbreviations)
2741 flavor of RCU, or <i>RCU-bh</i>, was developed by
2742 Dipankar Sarma to provide a flavor of RCU that could withstand the
2743 network-based denial-of-service attacks researched by Robert
2745 These attacks placed so much networking load on the system
2746 that some of the CPUs never exited softirq execution,
2747 which in turn prevented those CPUs from ever executing a context switch,
2748 which, in the RCU implementation of that time, prevented grace periods
2750 The result was an out-of-memory condition and a system hang.
2753 The solution was the creation of RCU-bh, which does
2754 <tt>local_bh_disable()</tt>
2755 across its read-side critical sections, and which uses the transition
2756 from one type of softirq processing to another as a quiescent state
2757 in addition to context switch, idle, user mode, and offline.
2758 This means that RCU-bh grace periods can complete even when some of
2759 the CPUs execute in softirq indefinitely, thus allowing algorithms
2760 based on RCU-bh to withstand network-based denial-of-service attacks.
2764 <tt>rcu_read_lock_bh()</tt> and <tt>rcu_read_unlock_bh()</tt>
2765 disable and re-enable softirq handlers, any attempt to start a softirq
2767 RCU-bh read-side critical section will be deferred.
2768 In this case, <tt>rcu_read_unlock_bh()</tt>
2769 will invoke softirq processing, which can take considerable time.
2770 One can of course argue that this softirq overhead should be associated
2771 with the code following the RCU-bh read-side critical section rather
2772 than <tt>rcu_read_unlock_bh()</tt>, but the fact
2773 is that most profiling tools cannot be expected to make this sort
2774 of fine distinction.
2775 For example, suppose that a three-millisecond-long RCU-bh read-side
2776 critical section executes during a time of heavy networking load.
2777 There will very likely be an attempt to invoke at least one softirq
2778 handler during that three milliseconds, but any such invocation will
2779 be delayed until the time of the <tt>rcu_read_unlock_bh()</tt>.
2780 This can of course make it appear at first glance as if
2781 <tt>rcu_read_unlock_bh()</tt> was executing very slowly.
2785 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-bh API</a>
2787 <tt>rcu_read_lock_bh()</tt>,
2788 <tt>rcu_read_unlock_bh()</tt>,
2789 <tt>rcu_dereference_bh()</tt>,
2790 <tt>rcu_dereference_bh_check()</tt>,
2791 <tt>synchronize_rcu_bh()</tt>,
2792 <tt>synchronize_rcu_bh_expedited()</tt>,
2793 <tt>call_rcu_bh()</tt>,
2794 <tt>rcu_barrier_bh()</tt>, and
2795 <tt>rcu_read_lock_bh_held()</tt>.
2797 <h3><a name="Sched Flavor">Sched Flavor</a></h3>
2800 Before preemptible RCU, waiting for an RCU grace period had the
2801 side effect of also waiting for all pre-existing interrupt
2803 However, there are legitimate preemptible-RCU implementations that
2804 do not have this property, given that any point in the code outside
2805 of an RCU read-side critical section can be a quiescent state.
2806 Therefore, <i>RCU-sched</i> was created, which follows “classic”
2807 RCU in that an RCU-sched grace period waits for for pre-existing
2808 interrupt and NMI handlers.
2809 In kernels built with <tt>CONFIG_PREEMPT=n</tt>, the RCU and RCU-sched
2810 APIs have identical implementations, while kernels built with
2811 <tt>CONFIG_PREEMPT=y</tt> provide a separate implementation for each.
2814 Note well that in <tt>CONFIG_PREEMPT=y</tt> kernels,
2815 <tt>rcu_read_lock_sched()</tt> and <tt>rcu_read_unlock_sched()</tt>
2816 disable and re-enable preemption, respectively.
2817 This means that if there was a preemption attempt during the
2818 RCU-sched read-side critical section, <tt>rcu_read_unlock_sched()</tt>
2819 will enter the scheduler, with all the latency and overhead entailed.
2820 Just as with <tt>rcu_read_unlock_bh()</tt>, this can make it look
2821 as if <tt>rcu_read_unlock_sched()</tt> was executing very slowly.
2822 However, the highest-priority task won't be preempted, so that task
2823 will enjoy low-overhead <tt>rcu_read_unlock_sched()</tt> invocations.
2827 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-sched API</a>
2829 <tt>rcu_read_lock_sched()</tt>,
2830 <tt>rcu_read_unlock_sched()</tt>,
2831 <tt>rcu_read_lock_sched_notrace()</tt>,
2832 <tt>rcu_read_unlock_sched_notrace()</tt>,
2833 <tt>rcu_dereference_sched()</tt>,
2834 <tt>rcu_dereference_sched_check()</tt>,
2835 <tt>synchronize_sched()</tt>,
2836 <tt>synchronize_rcu_sched_expedited()</tt>,
2837 <tt>call_rcu_sched()</tt>,
2838 <tt>rcu_barrier_sched()</tt>, and
2839 <tt>rcu_read_lock_sched_held()</tt>.
2840 However, anything that disables preemption also marks an RCU-sched
2841 read-side critical section, including
2842 <tt>preempt_disable()</tt> and <tt>preempt_enable()</tt>,
2843 <tt>local_irq_save()</tt> and <tt>local_irq_restore()</tt>,
2846 <h3><a name="Sleepable RCU">Sleepable RCU</a></h3>
2849 For well over a decade, someone saying “I need to block within
2850 an RCU read-side critical section” was a reliable indication
2851 that this someone did not understand RCU.
2852 After all, if you are always blocking in an RCU read-side critical
2853 section, you can probably afford to use a higher-overhead synchronization
2855 However, that changed with the advent of the Linux kernel's notifiers,
2856 whose RCU read-side critical
2857 sections almost never sleep, but sometimes need to.
2858 This resulted in the introduction of
2859 <a href="https://lwn.net/Articles/202847/">sleepable RCU</a>,
2863 SRCU allows different domains to be defined, with each such domain
2864 defined by an instance of an <tt>srcu_struct</tt> structure.
2865 A pointer to this structure must be passed in to each SRCU function,
2866 for example, <tt>synchronize_srcu(&ss)</tt>, where
2867 <tt>ss</tt> is the <tt>srcu_struct</tt> structure.
2868 The key benefit of these domains is that a slow SRCU reader in one
2869 domain does not delay an SRCU grace period in some other domain.
2870 That said, one consequence of these domains is that read-side code
2871 must pass a “cookie” from <tt>srcu_read_lock()</tt>
2872 to <tt>srcu_read_unlock()</tt>, for example, as follows:
2878 3 idx = srcu_read_lock(&ss);
2880 5 srcu_read_unlock(&ss, idx);
2885 As noted above, it is legal to block within SRCU read-side critical sections,
2886 however, with great power comes great responsibility.
2887 If you block forever in one of a given domain's SRCU read-side critical
2888 sections, then that domain's grace periods will also be blocked forever.
2889 Of course, one good way to block forever is to deadlock, which can
2890 happen if any operation in a given domain's SRCU read-side critical
2891 section can block waiting, either directly or indirectly, for that domain's
2892 grace period to elapse.
2893 For example, this results in a self-deadlock:
2899 3 idx = srcu_read_lock(&ss);
2901 5 synchronize_srcu(&ss);
2902 6 srcu_read_unlock(&ss, idx);
2907 However, if line 5 acquired a mutex that was held across
2908 a <tt>synchronize_srcu()</tt> for domain <tt>ss</tt>,
2909 deadlock would still be possible.
2910 Furthermore, if line 5 acquired a mutex that was held across
2911 a <tt>synchronize_srcu()</tt> for some other domain <tt>ss1</tt>,
2912 and if an <tt>ss1</tt>-domain SRCU read-side critical section
2913 acquired another mutex that was held across as <tt>ss</tt>-domain
2914 <tt>synchronize_srcu()</tt>,
2915 deadlock would again be possible.
2916 Such a deadlock cycle could extend across an arbitrarily large number
2917 of different SRCU domains.
2918 Again, with great power comes great responsibility.
2921 Unlike the other RCU flavors, SRCU read-side critical sections can
2922 run on idle and even offline CPUs.
2923 This ability requires that <tt>srcu_read_lock()</tt> and
2924 <tt>srcu_read_unlock()</tt> contain memory barriers, which means
2925 that SRCU readers will run a bit slower than would RCU readers.
2926 It also motivates the <tt>smp_mb__after_srcu_read_unlock()</tt>
2927 API, which, in combination with <tt>srcu_read_unlock()</tt>,
2928 guarantees a full memory barrier.
2931 Also unlike other RCU flavors, SRCU's callbacks-wait function
2932 <tt>srcu_barrier()</tt> may be invoked from CPU-hotplug notifiers,
2933 though this is not necessarily a good idea.
2934 The reason that this is possible is that SRCU is insensitive
2935 to whether or not a CPU is online, which means that <tt>srcu_barrier()</tt>
2936 need not exclude CPU-hotplug operations.
2939 As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating
2940 a locking bottleneck present in prior kernel versions.
2941 Although this will allow users to put much heavier stress on
2942 <tt>call_srcu()</tt>, it is important to note that SRCU does not
2943 yet take any special steps to deal with callback flooding.
2944 So if you are posting (say) 10,000 SRCU callbacks per second per CPU,
2945 you are probably totally OK, but if you intend to post (say) 1,000,000
2946 SRCU callbacks per second per CPU, please run some tests first.
2947 SRCU just might need a few adjustment to deal with that sort of load.
2948 Of course, your mileage may vary based on the speed of your CPUs and
2949 the size of your memory.
2953 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">SRCU API</a>
2955 <tt>srcu_read_lock()</tt>,
2956 <tt>srcu_read_unlock()</tt>,
2957 <tt>srcu_dereference()</tt>,
2958 <tt>srcu_dereference_check()</tt>,
2959 <tt>synchronize_srcu()</tt>,
2960 <tt>synchronize_srcu_expedited()</tt>,
2961 <tt>call_srcu()</tt>,
2962 <tt>srcu_barrier()</tt>, and
2963 <tt>srcu_read_lock_held()</tt>.
2965 <tt>DEFINE_SRCU()</tt>,
2966 <tt>DEFINE_STATIC_SRCU()</tt>, and
2967 <tt>init_srcu_struct()</tt>
2968 APIs for defining and initializing <tt>srcu_struct</tt> structures.
2970 <h3><a name="Tasks RCU">Tasks RCU</a></h3>
2973 Some forms of tracing use “trampolines” to handle the
2974 binary rewriting required to install different types of probes.
2975 It would be good to be able to free old trampolines, which sounds
2976 like a job for some form of RCU.
2977 However, because it is necessary to be able to install a trace
2978 anywhere in the code, it is not possible to use read-side markers
2979 such as <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
2980 In addition, it does not work to have these markers in the trampoline
2981 itself, because there would need to be instructions following
2982 <tt>rcu_read_unlock()</tt>.
2983 Although <tt>synchronize_rcu()</tt> would guarantee that execution
2984 reached the <tt>rcu_read_unlock()</tt>, it would not be able to
2985 guarantee that execution had completely left the trampoline.
2988 The solution, in the form of
2989 <a href="https://lwn.net/Articles/607117/"><i>Tasks RCU</i></a>,
2991 read-side critical sections that are delimited by voluntary context
2992 switches, that is, calls to <tt>schedule()</tt>,
2993 <tt>cond_resched_rcu_qs()</tt>, and
2994 <tt>synchronize_rcu_tasks()</tt>.
2995 In addition, transitions to and from userspace execution also delimit
2996 tasks-RCU read-side critical sections.
2999 The tasks-RCU API is quite compact, consisting only of
3000 <tt>call_rcu_tasks()</tt>,
3001 <tt>synchronize_rcu_tasks()</tt>, and
3002 <tt>rcu_barrier_tasks()</tt>.
3004 <h3><a name="Waiting for Multiple Grace Periods">
3005 Waiting for Multiple Grace Periods</a></h3>
3008 Perhaps you have an RCU protected data structure that is accessed from
3009 RCU read-side critical sections, from softirq handlers, and from
3010 hardware interrupt handlers.
3011 That is three flavors of RCU, the normal flavor, the bottom-half flavor,
3012 and the sched flavor.
3013 How to wait for a compound grace period?
3016 The best approach is usually to “just say no!” and
3017 insert <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
3018 around each RCU read-side critical section, regardless of what
3019 environment it happens to be in.
3020 But suppose that some of the RCU read-side critical sections are
3021 on extremely hot code paths, and that use of <tt>CONFIG_PREEMPT=n</tt>
3022 is not a viable option, so that <tt>rcu_read_lock()</tt> and
3023 <tt>rcu_read_unlock()</tt> are not free.
3027 You <i>could</i> wait on all three grace periods in succession, as follows:
3031 1 synchronize_rcu();
3032 2 synchronize_rcu_bh();
3033 3 synchronize_sched();
3038 This works, but triples the update-side latency penalty.
3039 In cases where this is not acceptable, <tt>synchronize_rcu_mult()</tt>
3040 may be used to wait on all three flavors of grace period concurrently:
3044 1 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched);
3049 But what if it is necessary to also wait on SRCU?
3050 This can be done as follows:
3054 1 static void call_my_srcu(struct rcu_head *head,
3055 2 void (*func)(struct rcu_head *head))
3057 4 call_srcu(&my_srcu, head, func);
3060 7 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched, call_my_srcu);
3065 If you needed to wait on multiple different flavors of SRCU
3066 (but why???), you would need to create a wrapper function resembling
3067 <tt>call_my_srcu()</tt> for each SRCU flavor.
3070 <tr><th> </th></tr>
3071 <tr><th align="left">Quick Quiz:</th></tr>
3073 But what if I need to wait for multiple RCU flavors, but I also need
3074 the grace periods to be expedited?
3076 <tr><th align="left">Answer:</th></tr>
3077 <tr><td bgcolor="#ffffff"><font color="ffffff">
3078 If you are using expedited grace periods, there should be less penalty
3079 for waiting on them in succession.
3080 But if that is nevertheless a problem, you can use workqueues
3081 or multiple kthreads to wait on the various expedited grace
3082 periods concurrently.
3084 <tr><td> </td></tr>
3088 Again, it is usually better to adjust the RCU read-side critical sections
3089 to use a single flavor of RCU, but when this is not feasible, you can use
3090 <tt>synchronize_rcu_mult()</tt>.
3092 <h2><a name="Possible Future Changes">Possible Future Changes</a></h2>
3095 One of the tricks that RCU uses to attain update-side scalability is
3096 to increase grace-period latency with increasing numbers of CPUs.
3097 If this becomes a serious problem, it will be necessary to rework the
3098 grace-period state machine so as to avoid the need for the additional
3102 Expedited grace periods scan the CPUs, so their latency and overhead
3103 increases with increasing numbers of CPUs.
3104 If this becomes a serious problem on large systems, it will be necessary
3105 to do some redesign to avoid this scalability problem.
3108 RCU disables CPU hotplug in a few places, perhaps most notably in the
3109 <tt>rcu_barrier()</tt> operations.
3110 If there is a strong reason to use <tt>rcu_barrier()</tt> in CPU-hotplug
3111 notifiers, it will be necessary to avoid disabling CPU hotplug.
3112 This would introduce some complexity, so there had better be a <i>very</i>
3116 The tradeoff between grace-period latency on the one hand and interruptions
3117 of other CPUs on the other hand may need to be re-examined.
3118 The desire is of course for zero grace-period latency as well as zero
3119 interprocessor interrupts undertaken during an expedited grace period
3121 While this ideal is unlikely to be achievable, it is quite possible that
3122 further improvements can be made.
3125 The multiprocessor implementations of RCU use a combining tree that
3126 groups CPUs so as to reduce lock contention and increase cache locality.
3127 However, this combining tree does not spread its memory across NUMA
3128 nodes nor does it align the CPU groups with hardware features such
3129 as sockets or cores.
3130 Such spreading and alignment is currently believed to be unnecessary
3131 because the hotpath read-side primitives do not access the combining
3132 tree, nor does <tt>call_rcu()</tt> in the common case.
3133 If you believe that your architecture needs such spreading and alignment,
3134 then your architecture should also benefit from the
3135 <tt>rcutree.rcu_fanout_leaf</tt> boot parameter, which can be set
3136 to the number of CPUs in a socket, NUMA node, or whatever.
3137 If the number of CPUs is too large, use a fraction of the number of
3139 If the number of CPUs is a large prime number, well, that certainly
3140 is an “interesting” architectural choice!
3141 More flexible arrangements might be considered, but only if
3142 <tt>rcutree.rcu_fanout_leaf</tt> has proven inadequate, and only
3143 if the inadequacy has been demonstrated by a carefully run and
3144 realistic system-level workload.
3147 Please note that arrangements that require RCU to remap CPU numbers will
3148 require extremely good demonstration of need and full exploration of
3152 There is an embarrassingly large number of flavors of RCU, and this
3153 number has been increasing over time.
3154 Perhaps it will be possible to combine some at some future date.
3157 RCU's various kthreads are reasonably recent additions.
3158 It is quite likely that adjustments will be required to more gracefully
3159 handle extreme loads.
3160 It might also be necessary to be able to relate CPU utilization by
3161 RCU's kthreads and softirq handlers to the code that instigated this
3163 For example, RCU callback overhead might be charged back to the
3164 originating <tt>call_rcu()</tt> instance, though probably not
3165 in production kernels.
3167 <h2><a name="Summary">Summary</a></h2>
3170 This document has presented more than two decade's worth of RCU
3172 Given that the requirements keep changing, this will not be the last
3173 word on this subject, but at least it serves to get an important
3174 subset of the requirements set forth.
3176 <h2><a name="Acknowledgments">Acknowledgments</a></h2>
3178 I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar,
3179 Oleg Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and
3180 Andy Lutomirski for their help in rendering
3181 this article human readable, and to Michelle Rankin for her support
3183 Other contributions are acknowledged in the Linux kernel's git archive.
git.samba.org / sfrench / cifs-2.6.git / blob