1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
2 "http://www.w3.org/TR/html4/loose.dtd">
4 <head><title>A Tour Through RCU's Requirements [LWN.net]</title>
5 <meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=utf-8">
7 <h1>A Tour Through RCU's Requirements</h1>
9 <p>Copyright IBM Corporation, 2015</p>
10 <p>Author: Paul E. McKenney</p>
11 <p><i>The initial version of this document appeared in the
12 <a href="https://lwn.net/">LWN</a> articles
13 <a href="https://lwn.net/Articles/652156/">here</a>,
14 <a href="https://lwn.net/Articles/652677/">here</a>, and
15 <a href="https://lwn.net/Articles/653326/">here</a>.</i></p>
20 Read-copy update (RCU) is a synchronization mechanism that is often
21 used as a replacement for reader-writer locking.
22 RCU is unusual in that updaters do not block readers,
23 which means that RCU's read-side primitives can be exceedingly fast
25 In addition, updaters can make useful forward progress concurrently
27 However, all this concurrency between RCU readers and updaters does raise
28 the question of exactly what RCU readers are doing, which in turn
29 raises the question of exactly what RCU's requirements are.
32 This document therefore summarizes RCU's requirements, and can be thought
33 of as an informal, high-level specification for RCU.
34 It is important to understand that RCU's specification is primarily
36 in fact, I learned about many of these requirements the hard way.
37 This situation might cause some consternation, however, not only
38 has this learning process been a lot of fun, but it has also been
39 a great privilege to work with so many people willing to apply
40 technologies in interesting new ways.
43 All that aside, here are the categories of currently known RCU requirements:
47 <li> <a href="#Fundamental Requirements">
48 Fundamental Requirements</a>
49 <li> <a href="#Fundamental Non-Requirements">Fundamental Non-Requirements</a>
50 <li> <a href="#Parallelism Facts of Life">
51 Parallelism Facts of Life</a>
52 <li> <a href="#Quality-of-Implementation Requirements">
53 Quality-of-Implementation Requirements</a>
54 <li> <a href="#Linux Kernel Complications">
55 Linux Kernel Complications</a>
56 <li> <a href="#Software-Engineering Requirements">
57 Software-Engineering Requirements</a>
58 <li> <a href="#Other RCU Flavors">
60 <li> <a href="#Possible Future Changes">
61 Possible Future Changes</a>
65 This is followed by a <a href="#Summary">summary</a>,
66 however, the answers to each quick quiz immediately follows the quiz.
67 Select the big white space with your mouse to see the answer.
69 <h2><a name="Fundamental Requirements">Fundamental Requirements</a></h2>
72 RCU's fundamental requirements are the closest thing RCU has to hard
73 mathematical requirements.
77 <li> <a href="#Grace-Period Guarantee">
78 Grace-Period Guarantee</a>
79 <li> <a href="#Publish-Subscribe Guarantee">
80 Publish-Subscribe Guarantee</a>
81 <li> <a href="#Memory-Barrier Guarantees">
82 Memory-Barrier Guarantees</a>
83 <li> <a href="#RCU Primitives Guaranteed to Execute Unconditionally">
84 RCU Primitives Guaranteed to Execute Unconditionally</a>
85 <li> <a href="#Guaranteed Read-to-Write Upgrade">
86 Guaranteed Read-to-Write Upgrade</a>
89 <h3><a name="Grace-Period Guarantee">Grace-Period Guarantee</a></h3>
92 RCU's grace-period guarantee is unusual in being premeditated:
93 Jack Slingwine and I had this guarantee firmly in mind when we started
94 work on RCU (then called “rclock”) in the early 1990s.
95 That said, the past two decades of experience with RCU have produced
96 a much more detailed understanding of this guarantee.
99 RCU's grace-period guarantee allows updaters to wait for the completion
100 of all pre-existing RCU read-side critical sections.
101 An RCU read-side critical section
102 begins with the marker <tt>rcu_read_lock()</tt> and ends with
103 the marker <tt>rcu_read_unlock()</tt>.
104 These markers may be nested, and RCU treats a nested set as one
105 big RCU read-side critical section.
106 Production-quality implementations of <tt>rcu_read_lock()</tt> and
107 <tt>rcu_read_unlock()</tt> are extremely lightweight, and in
108 fact have exactly zero overhead in Linux kernels built for production
109 use with <tt>CONFIG_PREEMPT=n</tt>.
112 This guarantee allows ordering to be enforced with extremely low
113 overhead to readers, for example:
127 11 void thread1(void)
130 14 synchronize_rcu();
137 Because the <tt>synchronize_rcu()</tt> on line 14 waits for
138 all pre-existing readers, any instance of <tt>thread0()</tt> that
139 loads a value of zero from <tt>x</tt> must complete before
140 <tt>thread1()</tt> stores to <tt>y</tt>, so that instance must
141 also load a value of zero from <tt>y</tt>.
142 Similarly, any instance of <tt>thread0()</tt> that loads a value of
143 one from <tt>y</tt> must have started after the
144 <tt>synchronize_rcu()</tt> started, and must therefore also load
145 a value of one from <tt>x</tt>.
146 Therefore, the outcome:
149 (r1 == 0 && r2 == 1)
155 <tr><th> </th></tr>
156 <tr><th align="left">Quick Quiz:</th></tr>
159 You said that updaters can make useful forward progress concurrently
160 with readers, but pre-existing readers will block
161 <tt>synchronize_rcu()</tt>!!!
162 Just who are you trying to fool???
164 <tr><th align="left">Answer:</th></tr>
165 <tr><td bgcolor="#ffffff"><font color="ffffff">
166 First, if updaters do not wish to be blocked by readers, they can use
167 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt>, which will
169 Second, even when using <tt>synchronize_rcu()</tt>, the other
170 update-side code does run concurrently with readers, whether
173 <tr><td> </td></tr>
177 This scenario resembles one of the first uses of RCU in
178 <a href="https://en.wikipedia.org/wiki/DYNIX">DYNIX/ptx</a>,
179 which managed a distributed lock manager's transition into
180 a state suitable for handling recovery from node failure,
181 more or less as follows:
185 1 #define STATE_NORMAL 0
186 2 #define STATE_WANT_RECOVERY 1
187 3 #define STATE_RECOVERING 2
188 4 #define STATE_WANT_NORMAL 3
190 6 int state = STATE_NORMAL;
192 8 void do_something_dlm(void)
197 13 state_snap = READ_ONCE(state);
198 14 if (state_snap == STATE_NORMAL)
201 17 do_something_carefully();
202 18 rcu_read_unlock();
205 21 void start_recovery(void)
207 23 WRITE_ONCE(state, STATE_WANT_RECOVERY);
208 24 synchronize_rcu();
209 25 WRITE_ONCE(state, STATE_RECOVERING);
211 27 WRITE_ONCE(state, STATE_WANT_NORMAL);
212 28 synchronize_rcu();
213 29 WRITE_ONCE(state, STATE_NORMAL);
219 The RCU read-side critical section in <tt>do_something_dlm()</tt>
220 works with the <tt>synchronize_rcu()</tt> in <tt>start_recovery()</tt>
221 to guarantee that <tt>do_something()</tt> never runs concurrently
222 with <tt>recovery()</tt>, but with little or no synchronization
223 overhead in <tt>do_something_dlm()</tt>.
226 <tr><th> </th></tr>
227 <tr><th align="left">Quick Quiz:</th></tr>
229 Why is the <tt>synchronize_rcu()</tt> on line 28 needed?
231 <tr><th align="left">Answer:</th></tr>
232 <tr><td bgcolor="#ffffff"><font color="ffffff">
233 Without that extra grace period, memory reordering could result in
234 <tt>do_something_dlm()</tt> executing <tt>do_something()</tt>
235 concurrently with the last bits of <tt>recovery()</tt>.
237 <tr><td> </td></tr>
241 In order to avoid fatal problems such as deadlocks,
242 an RCU read-side critical section must not contain calls to
243 <tt>synchronize_rcu()</tt>.
244 Similarly, an RCU read-side critical section must not
245 contain anything that waits, directly or indirectly, on completion of
246 an invocation of <tt>synchronize_rcu()</tt>.
249 Although RCU's grace-period guarantee is useful in and of itself, with
250 <a href="https://lwn.net/Articles/573497/">quite a few use cases</a>,
251 it would be good to be able to use RCU to coordinate read-side
252 access to linked data structures.
253 For this, the grace-period guarantee is not sufficient, as can
254 be seen in function <tt>add_gp_buggy()</tt> below.
255 We will look at the reader's code later, but in the meantime, just think of
256 the reader as locklessly picking up the <tt>gp</tt> pointer,
257 and, if the value loaded is non-<tt>NULL</tt>, locklessly accessing the
258 <tt>->a</tt> and <tt>->b</tt> fields.
262 1 bool add_gp_buggy(int a, int b)
264 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
267 6 spin_lock(&gp_lock);
268 7 if (rcu_access_pointer(gp)) {
269 8 spin_unlock(&gp_lock);
274 13 gp = p; /* ORDERING BUG */
275 14 spin_unlock(&gp_lock);
282 The problem is that both the compiler and weakly ordered CPUs are within
283 their rights to reorder this code as follows:
287 1 bool add_gp_buggy_optimized(int a, int b)
289 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
292 6 spin_lock(&gp_lock);
293 7 if (rcu_access_pointer(gp)) {
294 8 spin_unlock(&gp_lock);
297 <b>11 gp = p; /* ORDERING BUG */
300 14 spin_unlock(&gp_lock);
307 If an RCU reader fetches <tt>gp</tt> just after
308 <tt>add_gp_buggy_optimized</tt> executes line 11,
309 it will see garbage in the <tt>->a</tt> and <tt>->b</tt>
311 And this is but one of many ways in which compiler and hardware optimizations
313 Therefore, we clearly need some way to prevent the compiler and the CPU from
314 reordering in this manner, which brings us to the publish-subscribe
315 guarantee discussed in the next section.
317 <h3><a name="Publish-Subscribe Guarantee">Publish/Subscribe Guarantee</a></h3>
320 RCU's publish-subscribe guarantee allows data to be inserted
321 into a linked data structure without disrupting RCU readers.
322 The updater uses <tt>rcu_assign_pointer()</tt> to insert the
323 new data, and readers use <tt>rcu_dereference()</tt> to
324 access data, whether new or old.
325 The following shows an example of insertion:
329 1 bool add_gp(int a, int b)
331 3 p = kmalloc(sizeof(*p), GFP_KERNEL);
334 6 spin_lock(&gp_lock);
335 7 if (rcu_access_pointer(gp)) {
336 8 spin_unlock(&gp_lock);
341 13 rcu_assign_pointer(gp, p);
342 14 spin_unlock(&gp_lock);
349 The <tt>rcu_assign_pointer()</tt> on line 13 is conceptually
350 equivalent to a simple assignment statement, but also guarantees
351 that its assignment will
352 happen after the two assignments in lines 11 and 12,
353 similar to the C11 <tt>memory_order_release</tt> store operation.
354 It also prevents any number of “interesting” compiler
355 optimizations, for example, the use of <tt>gp</tt> as a scratch
356 location immediately preceding the assignment.
359 <tr><th> </th></tr>
360 <tr><th align="left">Quick Quiz:</th></tr>
362 But <tt>rcu_assign_pointer()</tt> does nothing to prevent the
363 two assignments to <tt>p->a</tt> and <tt>p->b</tt>
364 from being reordered.
365 Can't that also cause problems?
367 <tr><th align="left">Answer:</th></tr>
368 <tr><td bgcolor="#ffffff"><font color="ffffff">
370 The readers cannot see either of these two fields until
371 the assignment to <tt>gp</tt>, by which time both fields are
373 So reordering the assignments
374 to <tt>p->a</tt> and <tt>p->b</tt> cannot possibly
377 <tr><td> </td></tr>
381 It is tempting to assume that the reader need not do anything special
382 to control its accesses to the RCU-protected data,
383 as shown in <tt>do_something_gp_buggy()</tt> below:
387 1 bool do_something_gp_buggy(void)
390 4 p = gp; /* OPTIMIZATIONS GALORE!!! */
392 6 do_something(p->a, p->b);
396 10 rcu_read_unlock();
403 However, this temptation must be resisted because there are a
404 surprisingly large number of ways that the compiler
406 <a href="https://h71000.www7.hp.com/wizard/wiz_2637.html">DEC Alpha CPUs</a>)
407 can trip this code up.
408 For but one example, if the compiler were short of registers, it
409 might choose to refetch from <tt>gp</tt> rather than keeping
410 a separate copy in <tt>p</tt> as follows:
414 1 bool do_something_gp_buggy_optimized(void)
417 4 if (gp) { /* OPTIMIZATIONS GALORE!!! */
418 <b> 5 do_something(gp->a, gp->b);</b>
429 If this function ran concurrently with a series of updates that
430 replaced the current structure with a new one,
431 the fetches of <tt>gp->a</tt>
432 and <tt>gp->b</tt> might well come from two different structures,
433 which could cause serious confusion.
434 To prevent this (and much else besides), <tt>do_something_gp()</tt> uses
435 <tt>rcu_dereference()</tt> to fetch from <tt>gp</tt>:
439 1 bool do_something_gp(void)
442 4 p = rcu_dereference(gp);
444 6 do_something(p->a, p->b);
448 10 rcu_read_unlock();
455 The <tt>rcu_dereference()</tt> uses volatile casts and (for DEC Alpha)
456 memory barriers in the Linux kernel.
458 <a href="http://www.rdrop.com/users/paulmck/RCU/consume.2015.07.13a.pdf">high-quality implementation of C11 <tt>memory_order_consume</tt> [PDF]</a>
459 ever appear, then <tt>rcu_dereference()</tt> could be implemented
460 as a <tt>memory_order_consume</tt> load.
461 Regardless of the exact implementation, a pointer fetched by
462 <tt>rcu_dereference()</tt> may not be used outside of the
463 outermost RCU read-side critical section containing that
464 <tt>rcu_dereference()</tt>, unless protection of
465 the corresponding data element has been passed from RCU to some
466 other synchronization mechanism, most commonly locking or
467 <a href="https://www.kernel.org/doc/Documentation/RCU/rcuref.txt">reference counting</a>.
470 In short, updaters use <tt>rcu_assign_pointer()</tt> and readers
471 use <tt>rcu_dereference()</tt>, and these two RCU API elements
472 work together to ensure that readers have a consistent view of
473 newly added data elements.
476 Of course, it is also necessary to remove elements from RCU-protected
477 data structures, for example, using the following process:
480 <li> Remove the data element from the enclosing structure.
481 <li> Wait for all pre-existing RCU read-side critical sections
482 to complete (because only pre-existing readers can possibly have
483 a reference to the newly removed data element).
484 <li> At this point, only the updater has a reference to the
485 newly removed data element, so it can safely reclaim
486 the data element, for example, by passing it to <tt>kfree()</tt>.
489 This process is implemented by <tt>remove_gp_synchronous()</tt>:
493 1 bool remove_gp_synchronous(void)
497 5 spin_lock(&gp_lock);
498 6 p = rcu_access_pointer(gp);
500 8 spin_unlock(&gp_lock);
503 11 rcu_assign_pointer(gp, NULL);
504 12 spin_unlock(&gp_lock);
505 13 synchronize_rcu();
513 This function is straightforward, with line 13 waiting for a grace
514 period before line 14 frees the old data element.
515 This waiting ensures that readers will reach line 7 of
516 <tt>do_something_gp()</tt> before the data element referenced by
518 The <tt>rcu_access_pointer()</tt> on line 6 is similar to
519 <tt>rcu_dereference()</tt>, except that:
522 <li> The value returned by <tt>rcu_access_pointer()</tt>
523 cannot be dereferenced.
524 If you want to access the value pointed to as well as
525 the pointer itself, use <tt>rcu_dereference()</tt>
526 instead of <tt>rcu_access_pointer()</tt>.
527 <li> The call to <tt>rcu_access_pointer()</tt> need not be
529 In contrast, <tt>rcu_dereference()</tt> must either be
530 within an RCU read-side critical section or in a code
531 segment where the pointer cannot change, for example, in
532 code protected by the corresponding update-side lock.
536 <tr><th> </th></tr>
537 <tr><th align="left">Quick Quiz:</th></tr>
539 Without the <tt>rcu_dereference()</tt> or the
540 <tt>rcu_access_pointer()</tt>, what destructive optimizations
541 might the compiler make use of?
543 <tr><th align="left">Answer:</th></tr>
544 <tr><td bgcolor="#ffffff"><font color="ffffff">
545 Let's start with what happens to <tt>do_something_gp()</tt>
546 if it fails to use <tt>rcu_dereference()</tt>.
547 It could reuse a value formerly fetched from this same pointer.
548 It could also fetch the pointer from <tt>gp</tt> in a byte-at-a-time
549 manner, resulting in <i>load tearing</i>, in turn resulting a bytewise
550 mash-up of two distinct pointer values.
551 It might even use value-speculation optimizations, where it makes
552 a wrong guess, but by the time it gets around to checking the
553 value, an update has changed the pointer to match the wrong guess.
554 Too bad about any dereferences that returned pre-initialization garbage
558 <p><font color="ffffff">
559 For <tt>remove_gp_synchronous()</tt>, as long as all modifications
560 to <tt>gp</tt> are carried out while holding <tt>gp_lock</tt>,
561 the above optimizations are harmless.
562 However, <tt>sparse</tt> will complain if you
563 define <tt>gp</tt> with <tt>__rcu</tt> and then
564 access it without using
565 either <tt>rcu_access_pointer()</tt> or <tt>rcu_dereference()</tt>.
567 <tr><td> </td></tr>
571 In short, RCU's publish-subscribe guarantee is provided by the combination
572 of <tt>rcu_assign_pointer()</tt> and <tt>rcu_dereference()</tt>.
573 This guarantee allows data elements to be safely added to RCU-protected
574 linked data structures without disrupting RCU readers.
575 This guarantee can be used in combination with the grace-period
576 guarantee to also allow data elements to be removed from RCU-protected
577 linked data structures, again without disrupting RCU readers.
580 This guarantee was only partially premeditated.
581 DYNIX/ptx used an explicit memory barrier for publication, but had nothing
582 resembling <tt>rcu_dereference()</tt> for subscription, nor did it
583 have anything resembling the <tt>smp_read_barrier_depends()</tt>
584 that was later subsumed into <tt>rcu_dereference()</tt> and later
585 still into <tt>READ_ONCE()</tt>.
586 The need for these operations made itself known quite suddenly at a
587 late-1990s meeting with the DEC Alpha architects, back in the days when
588 DEC was still a free-standing company.
589 It took the Alpha architects a good hour to convince me that any sort
590 of barrier would ever be needed, and it then took me a good <i>two</i> hours
591 to convince them that their documentation did not make this point clear.
592 More recent work with the C and C++ standards committees have provided
593 much education on tricks and traps from the compiler.
594 In short, compilers were much less tricky in the early 1990s, but in
595 2015, don't even think about omitting <tt>rcu_dereference()</tt>!
597 <h3><a name="Memory-Barrier Guarantees">Memory-Barrier Guarantees</a></h3>
600 The previous section's simple linked-data-structure scenario clearly
601 demonstrates the need for RCU's stringent memory-ordering guarantees on
602 systems with more than one CPU:
605 <li> Each CPU that has an RCU read-side critical section that
606 begins before <tt>synchronize_rcu()</tt> starts is
607 guaranteed to execute a full memory barrier between the time
608 that the RCU read-side critical section ends and the time that
609 <tt>synchronize_rcu()</tt> returns.
610 Without this guarantee, a pre-existing RCU read-side critical section
611 might hold a reference to the newly removed <tt>struct foo</tt>
612 after the <tt>kfree()</tt> on line 14 of
613 <tt>remove_gp_synchronous()</tt>.
614 <li> Each CPU that has an RCU read-side critical section that ends
615 after <tt>synchronize_rcu()</tt> returns is guaranteed
616 to execute a full memory barrier between the time that
617 <tt>synchronize_rcu()</tt> begins and the time that the RCU
618 read-side critical section begins.
619 Without this guarantee, a later RCU read-side critical section
620 running after the <tt>kfree()</tt> on line 14 of
621 <tt>remove_gp_synchronous()</tt> might
622 later run <tt>do_something_gp()</tt> and find the
623 newly deleted <tt>struct foo</tt>.
624 <li> If the task invoking <tt>synchronize_rcu()</tt> remains
625 on a given CPU, then that CPU is guaranteed to execute a full
626 memory barrier sometime during the execution of
627 <tt>synchronize_rcu()</tt>.
628 This guarantee ensures that the <tt>kfree()</tt> on
629 line 14 of <tt>remove_gp_synchronous()</tt> really does
630 execute after the removal on line 11.
631 <li> If the task invoking <tt>synchronize_rcu()</tt> migrates
632 among a group of CPUs during that invocation, then each of the
633 CPUs in that group is guaranteed to execute a full memory barrier
634 sometime during the execution of <tt>synchronize_rcu()</tt>.
635 This guarantee also ensures that the <tt>kfree()</tt> on
636 line 14 of <tt>remove_gp_synchronous()</tt> really does
637 execute after the removal on
638 line 11, but also in the case where the thread executing the
639 <tt>synchronize_rcu()</tt> migrates in the meantime.
643 <tr><th> </th></tr>
644 <tr><th align="left">Quick Quiz:</th></tr>
646 Given that multiple CPUs can start RCU read-side critical sections
647 at any time without any ordering whatsoever, how can RCU possibly
648 tell whether or not a given RCU read-side critical section starts
649 before a given instance of <tt>synchronize_rcu()</tt>?
651 <tr><th align="left">Answer:</th></tr>
652 <tr><td bgcolor="#ffffff"><font color="ffffff">
653 If RCU cannot tell whether or not a given
654 RCU read-side critical section starts before a
655 given instance of <tt>synchronize_rcu()</tt>,
656 then it must assume that the RCU read-side critical section
658 In other words, a given instance of <tt>synchronize_rcu()</tt>
659 can avoid waiting on a given RCU read-side critical section only
660 if it can prove that <tt>synchronize_rcu()</tt> started first.
663 <p><font color="ffffff">
664 A related question is “When <tt>rcu_read_lock()</tt>
665 doesn't generate any code, why does it matter how it relates
666 to a grace period?”
667 The answer is that it is not the relationship of
668 <tt>rcu_read_lock()</tt> itself that is important, but rather
669 the relationship of the code within the enclosed RCU read-side
670 critical section to the code preceding and following the
672 If we take this viewpoint, then a given RCU read-side critical
673 section begins before a given grace period when some access
674 preceding the grace period observes the effect of some access
675 within the critical section, in which case none of the accesses
676 within the critical section may observe the effects of any
677 access following the grace period.
680 <p><font color="ffffff">
681 As of late 2016, mathematical models of RCU take this
682 viewpoint, for example, see slides 62 and 63
684 <a href="http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.2016.10.04c.LCE.pdf">2016 LinuxCon EU</a>
687 <tr><td> </td></tr>
691 <tr><th> </th></tr>
692 <tr><th align="left">Quick Quiz:</th></tr>
694 The first and second guarantees require unbelievably strict ordering!
695 Are all these memory barriers <i> really</i> required?
697 <tr><th align="left">Answer:</th></tr>
698 <tr><td bgcolor="#ffffff"><font color="ffffff">
699 Yes, they really are required.
700 To see why the first guarantee is required, consider the following
705 <li> <font color="ffffff">
706 CPU 1: <tt>rcu_read_lock()</tt>
708 <li> <font color="ffffff">
709 CPU 1: <tt>q = rcu_dereference(gp);
710 /* Very likely to return p. */</tt>
712 <li> <font color="ffffff">
713 CPU 0: <tt>list_del_rcu(p);</tt>
715 <li> <font color="ffffff">
716 CPU 0: <tt>synchronize_rcu()</tt> starts.
718 <li> <font color="ffffff">
719 CPU 1: <tt>do_something_with(q->a);
720 /* No smp_mb(), so might happen after kfree(). */</tt>
722 <li> <font color="ffffff">
723 CPU 1: <tt>rcu_read_unlock()</tt>
725 <li> <font color="ffffff">
726 CPU 0: <tt>synchronize_rcu()</tt> returns.
728 <li> <font color="ffffff">
729 CPU 0: <tt>kfree(p);</tt>
733 <p><font color="ffffff">
734 Therefore, there absolutely must be a full memory barrier between the
735 end of the RCU read-side critical section and the end of the
739 <p><font color="ffffff">
740 The sequence of events demonstrating the necessity of the second rule
745 <li> <font color="ffffff">CPU 0: <tt>list_del_rcu(p);</tt>
747 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> starts.
749 <li> <font color="ffffff">CPU 1: <tt>rcu_read_lock()</tt>
751 <li> <font color="ffffff">CPU 1: <tt>q = rcu_dereference(gp);
752 /* Might return p if no memory barrier. */</tt>
754 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> returns.
756 <li> <font color="ffffff">CPU 0: <tt>kfree(p);</tt>
758 <li> <font color="ffffff">
759 CPU 1: <tt>do_something_with(q->a); /* Boom!!! */</tt>
761 <li> <font color="ffffff">CPU 1: <tt>rcu_read_unlock()</tt>
765 <p><font color="ffffff">
766 And similarly, without a memory barrier between the beginning of the
767 grace period and the beginning of the RCU read-side critical section,
768 CPU 1 might end up accessing the freelist.
771 <p><font color="ffffff">
772 The “as if” rule of course applies, so that any
773 implementation that acts as if the appropriate memory barriers
774 were in place is a correct implementation.
775 That said, it is much easier to fool yourself into believing
776 that you have adhered to the as-if rule than it is to actually
779 <tr><td> </td></tr>
783 <tr><th> </th></tr>
784 <tr><th align="left">Quick Quiz:</th></tr>
786 You claim that <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
787 generate absolutely no code in some kernel builds.
788 This means that the compiler might arbitrarily rearrange consecutive
789 RCU read-side critical sections.
790 Given such rearrangement, if a given RCU read-side critical section
791 is done, how can you be sure that all prior RCU read-side critical
793 Won't the compiler rearrangements make that impossible to determine?
795 <tr><th align="left">Answer:</th></tr>
796 <tr><td bgcolor="#ffffff"><font color="ffffff">
797 In cases where <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>
798 generate absolutely no code, RCU infers quiescent states only at
799 special locations, for example, within the scheduler.
800 Because calls to <tt>schedule()</tt> had better prevent calling-code
801 accesses to shared variables from being rearranged across the call to
802 <tt>schedule()</tt>, if RCU detects the end of a given RCU read-side
803 critical section, it will necessarily detect the end of all prior
804 RCU read-side critical sections, no matter how aggressively the
805 compiler scrambles the code.
808 <p><font color="ffffff">
809 Again, this all assumes that the compiler cannot scramble code across
810 calls to the scheduler, out of interrupt handlers, into the idle loop,
811 into user-mode code, and so on.
812 But if your kernel build allows that sort of scrambling, you have broken
813 far more than just RCU!
815 <tr><td> </td></tr>
819 Note that these memory-barrier requirements do not replace the fundamental
820 RCU requirement that a grace period wait for all pre-existing readers.
821 On the contrary, the memory barriers called out in this section must operate in
822 such a way as to <i>enforce</i> this fundamental requirement.
823 Of course, different implementations enforce this requirement in different
824 ways, but enforce it they must.
826 <h3><a name="RCU Primitives Guaranteed to Execute Unconditionally">RCU Primitives Guaranteed to Execute Unconditionally</a></h3>
829 The common-case RCU primitives are unconditional.
830 They are invoked, they do their job, and they return, with no possibility
831 of error, and no need to retry.
832 This is a key RCU design philosophy.
835 However, this philosophy is pragmatic rather than pigheaded.
836 If someone comes up with a good justification for a particular conditional
837 RCU primitive, it might well be implemented and added.
838 After all, this guarantee was reverse-engineered, not premeditated.
839 The unconditional nature of the RCU primitives was initially an
840 accident of implementation, and later experience with synchronization
841 primitives with conditional primitives caused me to elevate this
842 accident to a guarantee.
843 Therefore, the justification for adding a conditional primitive to
844 RCU would need to be based on detailed and compelling use cases.
846 <h3><a name="Guaranteed Read-to-Write Upgrade">Guaranteed Read-to-Write Upgrade</a></h3>
849 As far as RCU is concerned, it is always possible to carry out an
850 update within an RCU read-side critical section.
851 For example, that RCU read-side critical section might search for
852 a given data element, and then might acquire the update-side
853 spinlock in order to update that element, all while remaining
854 in that RCU read-side critical section.
855 Of course, it is necessary to exit the RCU read-side critical section
856 before invoking <tt>synchronize_rcu()</tt>, however, this
857 inconvenience can be avoided through use of the
858 <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt> API members
859 described later in this document.
862 <tr><th> </th></tr>
863 <tr><th align="left">Quick Quiz:</th></tr>
865 But how does the upgrade-to-write operation exclude other readers?
867 <tr><th align="left">Answer:</th></tr>
868 <tr><td bgcolor="#ffffff"><font color="ffffff">
869 It doesn't, just like normal RCU updates, which also do not exclude
872 <tr><td> </td></tr>
876 This guarantee allows lookup code to be shared between read-side
877 and update-side code, and was premeditated, appearing in the earliest
878 DYNIX/ptx RCU documentation.
880 <h2><a name="Fundamental Non-Requirements">Fundamental Non-Requirements</a></h2>
883 RCU provides extremely lightweight readers, and its read-side guarantees,
884 though quite useful, are correspondingly lightweight.
885 It is therefore all too easy to assume that RCU is guaranteeing more
887 Of course, the list of things that RCU does not guarantee is infinitely
888 long, however, the following sections list a few non-guarantees that
889 have caused confusion.
890 Except where otherwise noted, these non-guarantees were premeditated.
893 <li> <a href="#Readers Impose Minimal Ordering">
894 Readers Impose Minimal Ordering</a>
895 <li> <a href="#Readers Do Not Exclude Updaters">
896 Readers Do Not Exclude Updaters</a>
897 <li> <a href="#Updaters Only Wait For Old Readers">
898 Updaters Only Wait For Old Readers</a>
899 <li> <a href="#Grace Periods Don't Partition Read-Side Critical Sections">
900 Grace Periods Don't Partition Read-Side Critical Sections</a>
901 <li> <a href="#Read-Side Critical Sections Don't Partition Grace Periods">
902 Read-Side Critical Sections Don't Partition Grace Periods</a>
903 <li> <a href="#Disabling Preemption Does Not Block Grace Periods">
904 Disabling Preemption Does Not Block Grace Periods</a>
907 <h3><a name="Readers Impose Minimal Ordering">Readers Impose Minimal Ordering</a></h3>
910 Reader-side markers such as <tt>rcu_read_lock()</tt> and
911 <tt>rcu_read_unlock()</tt> provide absolutely no ordering guarantees
912 except through their interaction with the grace-period APIs such as
913 <tt>synchronize_rcu()</tt>.
914 To see this, consider the following pair of threads:
928 11 void thread1(void)
931 14 r1 = READ_ONCE(y);
932 15 rcu_read_unlock();
934 17 r2 = READ_ONCE(x);
935 18 rcu_read_unlock();
941 After <tt>thread0()</tt> and <tt>thread1()</tt> execute
942 concurrently, it is quite possible to have
946 (r1 == 1 && r2 == 0)
950 (that is, <tt>y</tt> appears to have been assigned before <tt>x</tt>),
951 which would not be possible if <tt>rcu_read_lock()</tt> and
952 <tt>rcu_read_unlock()</tt> had much in the way of ordering
954 But they do not, so the CPU is within its rights
955 to do significant reordering.
956 This is by design: Any significant ordering constraints would slow down
957 these fast-path APIs.
960 <tr><th> </th></tr>
961 <tr><th align="left">Quick Quiz:</th></tr>
963 Can't the compiler also reorder this code?
965 <tr><th align="left">Answer:</th></tr>
966 <tr><td bgcolor="#ffffff"><font color="ffffff">
967 No, the volatile casts in <tt>READ_ONCE()</tt> and
968 <tt>WRITE_ONCE()</tt> prevent the compiler from reordering in
969 this particular case.
971 <tr><td> </td></tr>
974 <h3><a name="Readers Do Not Exclude Updaters">Readers Do Not Exclude Updaters</a></h3>
977 Neither <tt>rcu_read_lock()</tt> nor <tt>rcu_read_unlock()</tt>
979 All they do is to prevent grace periods from ending.
980 The following example illustrates this:
989 6 do_something_with_nonzero_x();
991 8 WARN_ON(!r2); /* BUG!!! */
993 10 rcu_read_unlock();
996 13 void thread1(void)
998 15 spin_lock(&my_lock);
1000 17 WRITE_ONCE(y, 1);
1001 18 spin_unlock(&my_lock);
1007 If the <tt>thread0()</tt> function's <tt>rcu_read_lock()</tt>
1008 excluded the <tt>thread1()</tt> function's update,
1009 the <tt>WARN_ON()</tt> could never fire.
1010 But the fact is that <tt>rcu_read_lock()</tt> does not exclude
1011 much of anything aside from subsequent grace periods, of which
1012 <tt>thread1()</tt> has none, so the
1013 <tt>WARN_ON()</tt> can and does fire.
1015 <h3><a name="Updaters Only Wait For Old Readers">Updaters Only Wait For Old Readers</a></h3>
1018 It might be tempting to assume that after <tt>synchronize_rcu()</tt>
1019 completes, there are no readers executing.
1020 This temptation must be avoided because
1021 new readers can start immediately after <tt>synchronize_rcu()</tt>
1022 starts, and <tt>synchronize_rcu()</tt> is under no
1023 obligation to wait for these new readers.
1026 <tr><th> </th></tr>
1027 <tr><th align="left">Quick Quiz:</th></tr>
1029 Suppose that synchronize_rcu() did wait until <i>all</i>
1030 readers had completed instead of waiting only on
1031 pre-existing readers.
1032 For how long would the updater be able to rely on there
1035 <tr><th align="left">Answer:</th></tr>
1036 <tr><td bgcolor="#ffffff"><font color="ffffff">
1038 Even if <tt>synchronize_rcu()</tt> were to wait until
1039 all readers had completed, a new reader might start immediately after
1040 <tt>synchronize_rcu()</tt> completed.
1041 Therefore, the code following
1042 <tt>synchronize_rcu()</tt> can <i>never</i> rely on there being
1045 <tr><td> </td></tr>
1048 <h3><a name="Grace Periods Don't Partition Read-Side Critical Sections">
1049 Grace Periods Don't Partition Read-Side Critical Sections</a></h3>
1052 It is tempting to assume that if any part of one RCU read-side critical
1053 section precedes a given grace period, and if any part of another RCU
1054 read-side critical section follows that same grace period, then all of
1055 the first RCU read-side critical section must precede all of the second.
1056 However, this just isn't the case: A single grace period does not
1057 partition the set of RCU read-side critical sections.
1058 An example of this situation can be illustrated as follows, where
1059 <tt>x</tt>, <tt>y</tt>, and <tt>z</tt> are initially all zero:
1063 1 void thread0(void)
1068 6 rcu_read_unlock();
1071 9 void thread1(void)
1073 11 r1 = READ_ONCE(a);
1074 12 synchronize_rcu();
1075 13 WRITE_ONCE(c, 1);
1078 16 void thread2(void)
1081 19 r2 = READ_ONCE(b);
1082 20 r3 = READ_ONCE(c);
1083 21 rcu_read_unlock();
1089 It turns out that the outcome:
1093 (r1 == 1 && r2 == 0 && r3 == 1)
1097 is entirely possible.
1098 The following figure show how this can happen, with each circled
1099 <tt>QS</tt> indicating the point at which RCU recorded a
1100 <i>quiescent state</i> for each thread, that is, a state in which
1101 RCU knows that the thread cannot be in the midst of an RCU read-side
1102 critical section that started before the current grace period:
1104 <p><img src="GPpartitionReaders1.svg" alt="GPpartitionReaders1.svg" width="60%"></p>
1107 If it is necessary to partition RCU read-side critical sections in this
1108 manner, it is necessary to use two grace periods, where the first
1109 grace period is known to end before the second grace period starts:
1113 1 void thread0(void)
1118 6 rcu_read_unlock();
1121 9 void thread1(void)
1123 11 r1 = READ_ONCE(a);
1124 12 synchronize_rcu();
1125 13 WRITE_ONCE(c, 1);
1128 16 void thread2(void)
1130 18 r2 = READ_ONCE(c);
1131 19 synchronize_rcu();
1132 20 WRITE_ONCE(d, 1);
1135 23 void thread3(void)
1138 26 r3 = READ_ONCE(b);
1139 27 r4 = READ_ONCE(d);
1140 28 rcu_read_unlock();
1146 Here, if <tt>(r1 == 1)</tt>, then
1147 <tt>thread0()</tt>'s write to <tt>b</tt> must happen
1148 before the end of <tt>thread1()</tt>'s grace period.
1149 If in addition <tt>(r4 == 1)</tt>, then
1150 <tt>thread3()</tt>'s read from <tt>b</tt> must happen
1151 after the beginning of <tt>thread2()</tt>'s grace period.
1152 If it is also the case that <tt>(r2 == 1)</tt>, then the
1153 end of <tt>thread1()</tt>'s grace period must precede the
1154 beginning of <tt>thread2()</tt>'s grace period.
1155 This mean that the two RCU read-side critical sections cannot overlap,
1156 guaranteeing that <tt>(r3 == 1)</tt>.
1157 As a result, the outcome:
1161 (r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
1168 This non-requirement was also non-premeditated, but became apparent
1169 when studying RCU's interaction with memory ordering.
1171 <h3><a name="Read-Side Critical Sections Don't Partition Grace Periods">
1172 Read-Side Critical Sections Don't Partition Grace Periods</a></h3>
1175 It is also tempting to assume that if an RCU read-side critical section
1176 happens between a pair of grace periods, then those grace periods cannot
1178 However, this temptation leads nowhere good, as can be illustrated by
1179 the following, with all variables initially zero:
1183 1 void thread0(void)
1188 6 rcu_read_unlock();
1191 9 void thread1(void)
1193 11 r1 = READ_ONCE(a);
1194 12 synchronize_rcu();
1195 13 WRITE_ONCE(c, 1);
1198 16 void thread2(void)
1201 19 WRITE_ONCE(d, 1);
1202 20 r2 = READ_ONCE(c);
1203 21 rcu_read_unlock();
1206 24 void thread3(void)
1208 26 r3 = READ_ONCE(d);
1209 27 synchronize_rcu();
1210 28 WRITE_ONCE(e, 1);
1213 31 void thread4(void)
1216 34 r4 = READ_ONCE(b);
1217 35 r5 = READ_ONCE(e);
1218 36 rcu_read_unlock();
1224 In this case, the outcome:
1228 (r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
1232 is entirely possible, as illustrated below:
1234 <p><img src="ReadersPartitionGP1.svg" alt="ReadersPartitionGP1.svg" width="100%"></p>
1237 Again, an RCU read-side critical section can overlap almost all of a
1238 given grace period, just so long as it does not overlap the entire
1240 As a result, an RCU read-side critical section cannot partition a pair
1241 of RCU grace periods.
1244 <tr><th> </th></tr>
1245 <tr><th align="left">Quick Quiz:</th></tr>
1247 How long a sequence of grace periods, each separated by an RCU
1248 read-side critical section, would be required to partition the RCU
1249 read-side critical sections at the beginning and end of the chain?
1251 <tr><th align="left">Answer:</th></tr>
1252 <tr><td bgcolor="#ffffff"><font color="ffffff">
1253 In theory, an infinite number.
1254 In practice, an unknown number that is sensitive to both implementation
1255 details and timing considerations.
1256 Therefore, even in practice, RCU users must abide by the
1257 theoretical rather than the practical answer.
1259 <tr><td> </td></tr>
1262 <h3><a name="Disabling Preemption Does Not Block Grace Periods">
1263 Disabling Preemption Does Not Block Grace Periods</a></h3>
1266 There was a time when disabling preemption on any given CPU would block
1267 subsequent grace periods.
1268 However, this was an accident of implementation and is not a requirement.
1269 And in the current Linux-kernel implementation, disabling preemption
1270 on a given CPU in fact does not block grace periods, as Oleg Nesterov
1271 <a href="https://lkml.kernel.org/g/20150614193825.GA19582@redhat.com">demonstrated</a>.
1274 If you need a preempt-disable region to block grace periods, you need to add
1275 <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>, for example
1280 1 preempt_disable();
1283 4 rcu_read_unlock();
1286 7 /* Spinlocks implicitly disable preemption. */
1287 8 spin_lock(&mylock);
1290 11 rcu_read_unlock();
1291 12 spin_unlock(&mylock);
1296 In theory, you could enter the RCU read-side critical section first,
1297 but it is more efficient to keep the entire RCU read-side critical
1298 section contained in the preempt-disable region as shown above.
1299 Of course, RCU read-side critical sections that extend outside of
1300 preempt-disable regions will work correctly, but such critical sections
1301 can be preempted, which forces <tt>rcu_read_unlock()</tt> to do
1303 And no, this is <i>not</i> an invitation to enclose all of your RCU
1304 read-side critical sections within preempt-disable regions, because
1305 doing so would degrade real-time response.
1308 This non-requirement appeared with preemptible RCU.
1310 <h2><a name="Parallelism Facts of Life">Parallelism Facts of Life</a></h2>
1313 These parallelism facts of life are by no means specific to RCU, but
1314 the RCU implementation must abide by them.
1315 They therefore bear repeating:
1318 <li> Any CPU or task may be delayed at any time,
1319 and any attempts to avoid these delays by disabling
1320 preemption, interrupts, or whatever are completely futile.
1321 This is most obvious in preemptible user-level
1322 environments and in virtualized environments (where
1323 a given guest OS's VCPUs can be preempted at any time by
1324 the underlying hypervisor), but can also happen in bare-metal
1325 environments due to ECC errors, NMIs, and other hardware
1327 Although a delay of more than about 20 seconds can result
1328 in splats, the RCU implementation is obligated to use
1329 algorithms that can tolerate extremely long delays, but where
1330 “extremely long” is not long enough to allow
1331 wrap-around when incrementing a 64-bit counter.
1332 <li> Both the compiler and the CPU can reorder memory accesses.
1333 Where it matters, RCU must use compiler directives and
1334 memory-barrier instructions to preserve ordering.
1335 <li> Conflicting writes to memory locations in any given cache line
1336 will result in expensive cache misses.
1337 Greater numbers of concurrent writes and more-frequent
1338 concurrent writes will result in more dramatic slowdowns.
1339 RCU is therefore obligated to use algorithms that have
1340 sufficient locality to avoid significant performance and
1341 scalability problems.
1342 <li> As a rough rule of thumb, only one CPU's worth of processing
1343 may be carried out under the protection of any given exclusive
1345 RCU must therefore use scalable locking designs.
1346 <li> Counters are finite, especially on 32-bit systems.
1347 RCU's use of counters must therefore tolerate counter wrap,
1348 or be designed such that counter wrap would take way more
1349 time than a single system is likely to run.
1350 An uptime of ten years is quite possible, a runtime
1351 of a century much less so.
1352 As an example of the latter, RCU's dyntick-idle nesting counter
1353 allows 54 bits for interrupt nesting level (this counter
1354 is 64 bits even on a 32-bit system).
1355 Overflowing this counter requires 2<sup>54</sup>
1356 half-interrupts on a given CPU without that CPU ever going idle.
1357 If a half-interrupt happened every microsecond, it would take
1358 570 years of runtime to overflow this counter, which is currently
1359 believed to be an acceptably long time.
1360 <li> Linux systems can have thousands of CPUs running a single
1361 Linux kernel in a single shared-memory environment.
1362 RCU must therefore pay close attention to high-end scalability.
1366 This last parallelism fact of life means that RCU must pay special
1367 attention to the preceding facts of life.
1368 The idea that Linux might scale to systems with thousands of CPUs would
1369 have been met with some skepticism in the 1990s, but these requirements
1370 would have otherwise have been unsurprising, even in the early 1990s.
1372 <h2><a name="Quality-of-Implementation Requirements">Quality-of-Implementation Requirements</a></h2>
1375 These sections list quality-of-implementation requirements.
1376 Although an RCU implementation that ignores these requirements could
1377 still be used, it would likely be subject to limitations that would
1378 make it inappropriate for industrial-strength production use.
1379 Classes of quality-of-implementation requirements are as follows:
1382 <li> <a href="#Specialization">Specialization</a>
1383 <li> <a href="#Performance and Scalability">Performance and Scalability</a>
1384 <li> <a href="#Composability">Composability</a>
1385 <li> <a href="#Corner Cases">Corner Cases</a>
1389 These classes is covered in the following sections.
1391 <h3><a name="Specialization">Specialization</a></h3>
1394 RCU is and always has been intended primarily for read-mostly situations,
1395 which means that RCU's read-side primitives are optimized, often at the
1396 expense of its update-side primitives.
1397 Experience thus far is captured by the following list of situations:
1400 <li> Read-mostly data, where stale and inconsistent data is not
1401 a problem: RCU works great!
1402 <li> Read-mostly data, where data must be consistent:
1404 <li> Read-write data, where data must be consistent:
1405 RCU <i>might</i> work OK.
1407 <li> Write-mostly data, where data must be consistent:
1408 RCU is very unlikely to be the right tool for the job,
1409 with the following exceptions, where RCU can provide:
1411 <li> Existence guarantees for update-friendly mechanisms.
1412 <li> Wait-free read-side primitives for real-time use.
1417 This focus on read-mostly situations means that RCU must interoperate
1418 with other synchronization primitives.
1419 For example, the <tt>add_gp()</tt> and <tt>remove_gp_synchronous()</tt>
1420 examples discussed earlier use RCU to protect readers and locking to
1421 coordinate updaters.
1422 However, the need extends much farther, requiring that a variety of
1423 synchronization primitives be legal within RCU read-side critical sections,
1424 including spinlocks, sequence locks, atomic operations, reference
1425 counters, and memory barriers.
1428 <tr><th> </th></tr>
1429 <tr><th align="left">Quick Quiz:</th></tr>
1431 What about sleeping locks?
1433 <tr><th align="left">Answer:</th></tr>
1434 <tr><td bgcolor="#ffffff"><font color="ffffff">
1435 These are forbidden within Linux-kernel RCU read-side critical
1436 sections because it is not legal to place a quiescent state
1437 (in this case, voluntary context switch) within an RCU read-side
1439 However, sleeping locks may be used within userspace RCU read-side
1440 critical sections, and also within Linux-kernel sleepable RCU
1441 <a href="#Sleepable RCU"><font color="ffffff">(SRCU)</font></a>
1442 read-side critical sections.
1443 In addition, the -rt patchset turns spinlocks into a
1444 sleeping locks so that the corresponding critical sections
1445 can be preempted, which also means that these sleeplockified
1446 spinlocks (but not other sleeping locks!) may be acquire within
1447 -rt-Linux-kernel RCU read-side critical sections.
1450 <p><font color="ffffff">
1451 Note that it <i>is</i> legal for a normal RCU read-side
1452 critical section to conditionally acquire a sleeping locks
1453 (as in <tt>mutex_trylock()</tt>), but only as long as it does
1454 not loop indefinitely attempting to conditionally acquire that
1456 The key point is that things like <tt>mutex_trylock()</tt>
1457 either return with the mutex held, or return an error indication if
1458 the mutex was not immediately available.
1459 Either way, <tt>mutex_trylock()</tt> returns immediately without
1462 <tr><td> </td></tr>
1466 It often comes as a surprise that many algorithms do not require a
1467 consistent view of data, but many can function in that mode,
1468 with network routing being the poster child.
1469 Internet routing algorithms take significant time to propagate
1470 updates, so that by the time an update arrives at a given system,
1471 that system has been sending network traffic the wrong way for
1472 a considerable length of time.
1473 Having a few threads continue to send traffic the wrong way for a
1474 few more milliseconds is clearly not a problem: In the worst case,
1475 TCP retransmissions will eventually get the data where it needs to go.
1476 In general, when tracking the state of the universe outside of the
1477 computer, some level of inconsistency must be tolerated due to
1478 speed-of-light delays if nothing else.
1481 Furthermore, uncertainty about external state is inherent in many cases.
1482 For example, a pair of veterinarians might use heartbeat to determine
1483 whether or not a given cat was alive.
1484 But how long should they wait after the last heartbeat to decide that
1485 the cat is in fact dead?
1486 Waiting less than 400 milliseconds makes no sense because this would
1487 mean that a relaxed cat would be considered to cycle between death
1488 and life more than 100 times per minute.
1489 Moreover, just as with human beings, a cat's heart might stop for
1490 some period of time, so the exact wait period is a judgment call.
1491 One of our pair of veterinarians might wait 30 seconds before pronouncing
1492 the cat dead, while the other might insist on waiting a full minute.
1493 The two veterinarians would then disagree on the state of the cat during
1494 the final 30 seconds of the minute following the last heartbeat.
1497 Interestingly enough, this same situation applies to hardware.
1498 When push comes to shove, how do we tell whether or not some
1499 external server has failed?
1500 We send messages to it periodically, and declare it failed if we
1501 don't receive a response within a given period of time.
1502 Policy decisions can usually tolerate short
1503 periods of inconsistency.
1504 The policy was decided some time ago, and is only now being put into
1505 effect, so a few milliseconds of delay is normally inconsequential.
1508 However, there are algorithms that absolutely must see consistent data.
1509 For example, the translation between a user-level SystemV semaphore
1510 ID to the corresponding in-kernel data structure is protected by RCU,
1511 but it is absolutely forbidden to update a semaphore that has just been
1513 In the Linux kernel, this need for consistency is accommodated by acquiring
1514 spinlocks located in the in-kernel data structure from within
1515 the RCU read-side critical section, and this is indicated by the
1516 green box in the figure above.
1517 Many other techniques may be used, and are in fact used within the
1521 In short, RCU is not required to maintain consistency, and other
1522 mechanisms may be used in concert with RCU when consistency is required.
1523 RCU's specialization allows it to do its job extremely well, and its
1524 ability to interoperate with other synchronization mechanisms allows
1525 the right mix of synchronization tools to be used for a given job.
1527 <h3><a name="Performance and Scalability">Performance and Scalability</a></h3>
1530 Energy efficiency is a critical component of performance today,
1531 and Linux-kernel RCU implementations must therefore avoid unnecessarily
1532 awakening idle CPUs.
1533 I cannot claim that this requirement was premeditated.
1534 In fact, I learned of it during a telephone conversation in which I
1535 was given “frank and open” feedback on the importance
1536 of energy efficiency in battery-powered systems and on specific
1537 energy-efficiency shortcomings of the Linux-kernel RCU implementation.
1538 In my experience, the battery-powered embedded community will consider
1539 any unnecessary wakeups to be extremely unfriendly acts.
1540 So much so that mere Linux-kernel-mailing-list posts are
1541 insufficient to vent their ire.
1544 Memory consumption is not particularly important for in most
1545 situations, and has become decreasingly
1546 so as memory sizes have expanded and memory
1547 costs have plummeted.
1548 However, as I learned from Matt Mackall's
1549 <a href="http://elinux.org/Linux_Tiny-FAQ">bloatwatch</a>
1550 efforts, memory footprint is critically important on single-CPU systems with
1551 non-preemptible (<tt>CONFIG_PREEMPT=n</tt>) kernels, and thus
1552 <a href="https://lkml.kernel.org/g/20090113221724.GA15307@linux.vnet.ibm.com">tiny RCU</a>
1554 Josh Triplett has since taken over the small-memory banner with his
1555 <a href="https://tiny.wiki.kernel.org/">Linux kernel tinification</a>
1556 project, which resulted in
1557 <a href="#Sleepable RCU">SRCU</a>
1558 becoming optional for those kernels not needing it.
1561 The remaining performance requirements are, for the most part,
1563 For example, in keeping with RCU's read-side specialization,
1564 <tt>rcu_dereference()</tt> should have negligible overhead (for
1565 example, suppression of a few minor compiler optimizations).
1566 Similarly, in non-preemptible environments, <tt>rcu_read_lock()</tt> and
1567 <tt>rcu_read_unlock()</tt> should have exactly zero overhead.
1570 In preemptible environments, in the case where the RCU read-side
1571 critical section was not preempted (as will be the case for the
1572 highest-priority real-time process), <tt>rcu_read_lock()</tt> and
1573 <tt>rcu_read_unlock()</tt> should have minimal overhead.
1574 In particular, they should not contain atomic read-modify-write
1575 operations, memory-barrier instructions, preemption disabling,
1576 interrupt disabling, or backwards branches.
1577 However, in the case where the RCU read-side critical section was preempted,
1578 <tt>rcu_read_unlock()</tt> may acquire spinlocks and disable interrupts.
1579 This is why it is better to nest an RCU read-side critical section
1580 within a preempt-disable region than vice versa, at least in cases
1581 where that critical section is short enough to avoid unduly degrading
1582 real-time latencies.
1585 The <tt>synchronize_rcu()</tt> grace-period-wait primitive is
1586 optimized for throughput.
1587 It may therefore incur several milliseconds of latency in addition to
1588 the duration of the longest RCU read-side critical section.
1589 On the other hand, multiple concurrent invocations of
1590 <tt>synchronize_rcu()</tt> are required to use batching optimizations
1591 so that they can be satisfied by a single underlying grace-period-wait
1593 For example, in the Linux kernel, it is not unusual for a single
1594 grace-period-wait operation to serve more than
1595 <a href="https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response">1,000 separate invocations</a>
1596 of <tt>synchronize_rcu()</tt>, thus amortizing the per-invocation
1597 overhead down to nearly zero.
1598 However, the grace-period optimization is also required to avoid
1599 measurable degradation of real-time scheduling and interrupt latencies.
1602 In some cases, the multi-millisecond <tt>synchronize_rcu()</tt>
1603 latencies are unacceptable.
1604 In these cases, <tt>synchronize_rcu_expedited()</tt> may be used
1605 instead, reducing the grace-period latency down to a few tens of
1606 microseconds on small systems, at least in cases where the RCU read-side
1607 critical sections are short.
1608 There are currently no special latency requirements for
1609 <tt>synchronize_rcu_expedited()</tt> on large systems, but,
1610 consistent with the empirical nature of the RCU specification,
1611 that is subject to change.
1612 However, there most definitely are scalability requirements:
1613 A storm of <tt>synchronize_rcu_expedited()</tt> invocations on 4096
1614 CPUs should at least make reasonable forward progress.
1615 In return for its shorter latencies, <tt>synchronize_rcu_expedited()</tt>
1616 is permitted to impose modest degradation of real-time latency
1617 on non-idle online CPUs.
1618 Here, “modest” means roughly the same latency
1619 degradation as a scheduling-clock interrupt.
1622 There are a number of situations where even
1623 <tt>synchronize_rcu_expedited()</tt>'s reduced grace-period
1624 latency is unacceptable.
1625 In these situations, the asynchronous <tt>call_rcu()</tt> can be
1626 used in place of <tt>synchronize_rcu()</tt> as follows:
1633 4 struct rcu_head rh;
1636 7 static void remove_gp_cb(struct rcu_head *rhp)
1638 9 struct foo *p = container_of(rhp, struct foo, rh);
1643 14 bool remove_gp_asynchronous(void)
1647 18 spin_lock(&gp_lock);
1648 19 p = rcu_dereference(gp);
1650 21 spin_unlock(&gp_lock);
1653 24 rcu_assign_pointer(gp, NULL);
1654 25 call_rcu(&p->rh, remove_gp_cb);
1655 26 spin_unlock(&gp_lock);
1662 A definition of <tt>struct foo</tt> is finally needed, and appears
1664 The function <tt>remove_gp_cb()</tt> is passed to <tt>call_rcu()</tt>
1665 on line 25, and will be invoked after the end of a subsequent
1667 This gets the same effect as <tt>remove_gp_synchronous()</tt>,
1668 but without forcing the updater to wait for a grace period to elapse.
1669 The <tt>call_rcu()</tt> function may be used in a number of
1670 situations where neither <tt>synchronize_rcu()</tt> nor
1671 <tt>synchronize_rcu_expedited()</tt> would be legal,
1672 including within preempt-disable code, <tt>local_bh_disable()</tt> code,
1673 interrupt-disable code, and interrupt handlers.
1674 However, even <tt>call_rcu()</tt> is illegal within NMI handlers
1675 and from idle and offline CPUs.
1676 The callback function (<tt>remove_gp_cb()</tt> in this case) will be
1677 executed within softirq (software interrupt) environment within the
1679 either within a real softirq handler or under the protection
1680 of <tt>local_bh_disable()</tt>.
1681 In both the Linux kernel and in userspace, it is bad practice to
1682 write an RCU callback function that takes too long.
1683 Long-running operations should be relegated to separate threads or
1684 (in the Linux kernel) workqueues.
1687 <tr><th> </th></tr>
1688 <tr><th align="left">Quick Quiz:</th></tr>
1690 Why does line 19 use <tt>rcu_access_pointer()</tt>?
1691 After all, <tt>call_rcu()</tt> on line 25 stores into the
1692 structure, which would interact badly with concurrent insertions.
1693 Doesn't this mean that <tt>rcu_dereference()</tt> is required?
1695 <tr><th align="left">Answer:</th></tr>
1696 <tr><td bgcolor="#ffffff"><font color="ffffff">
1697 Presumably the <tt>->gp_lock</tt> acquired on line 18 excludes
1698 any changes, including any insertions that <tt>rcu_dereference()</tt>
1699 would protect against.
1700 Therefore, any insertions will be delayed until after
1701 <tt>->gp_lock</tt>
1702 is released on line 25, which in turn means that
1703 <tt>rcu_access_pointer()</tt> suffices.
1705 <tr><td> </td></tr>
1709 However, all that <tt>remove_gp_cb()</tt> is doing is
1710 invoking <tt>kfree()</tt> on the data element.
1711 This is a common idiom, and is supported by <tt>kfree_rcu()</tt>,
1712 which allows “fire and forget” operation as shown below:
1719 4 struct rcu_head rh;
1722 7 bool remove_gp_faf(void)
1726 11 spin_lock(&gp_lock);
1727 12 p = rcu_dereference(gp);
1729 14 spin_unlock(&gp_lock);
1732 17 rcu_assign_pointer(gp, NULL);
1733 18 kfree_rcu(p, rh);
1734 19 spin_unlock(&gp_lock);
1741 Note that <tt>remove_gp_faf()</tt> simply invokes
1742 <tt>kfree_rcu()</tt> and proceeds, without any need to pay any
1743 further attention to the subsequent grace period and <tt>kfree()</tt>.
1744 It is permissible to invoke <tt>kfree_rcu()</tt> from the same
1745 environments as for <tt>call_rcu()</tt>.
1746 Interestingly enough, DYNIX/ptx had the equivalents of
1747 <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>, but not
1748 <tt>synchronize_rcu()</tt>.
1749 This was due to the fact that RCU was not heavily used within DYNIX/ptx,
1750 so the very few places that needed something like
1751 <tt>synchronize_rcu()</tt> simply open-coded it.
1754 <tr><th> </th></tr>
1755 <tr><th align="left">Quick Quiz:</th></tr>
1757 Earlier it was claimed that <tt>call_rcu()</tt> and
1758 <tt>kfree_rcu()</tt> allowed updaters to avoid being blocked
1760 But how can that be correct, given that the invocation of the callback
1761 and the freeing of the memory (respectively) must still wait for
1762 a grace period to elapse?
1764 <tr><th align="left">Answer:</th></tr>
1765 <tr><td bgcolor="#ffffff"><font color="ffffff">
1766 We could define things this way, but keep in mind that this sort of
1767 definition would say that updates in garbage-collected languages
1768 cannot complete until the next time the garbage collector runs,
1769 which does not seem at all reasonable.
1770 The key point is that in most cases, an updater using either
1771 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> can proceed to the
1772 next update as soon as it has invoked <tt>call_rcu()</tt> or
1773 <tt>kfree_rcu()</tt>, without having to wait for a subsequent
1776 <tr><td> </td></tr>
1780 But what if the updater must wait for the completion of code to be
1781 executed after the end of the grace period, but has other tasks
1782 that can be carried out in the meantime?
1783 The polling-style <tt>get_state_synchronize_rcu()</tt> and
1784 <tt>cond_synchronize_rcu()</tt> functions may be used for this
1785 purpose, as shown below:
1789 1 bool remove_gp_poll(void)
1794 6 spin_lock(&gp_lock);
1795 7 p = rcu_access_pointer(gp);
1797 9 spin_unlock(&gp_lock);
1800 12 rcu_assign_pointer(gp, NULL);
1801 13 spin_unlock(&gp_lock);
1802 14 s = get_state_synchronize_rcu();
1803 15 do_something_while_waiting();
1804 16 cond_synchronize_rcu(s);
1812 On line 14, <tt>get_state_synchronize_rcu()</tt> obtains a
1813 “cookie” from RCU,
1814 then line 15 carries out other tasks,
1815 and finally, line 16 returns immediately if a grace period has
1816 elapsed in the meantime, but otherwise waits as required.
1817 The need for <tt>get_state_synchronize_rcu</tt> and
1818 <tt>cond_synchronize_rcu()</tt> has appeared quite recently,
1819 so it is too early to tell whether they will stand the test of time.
1822 RCU thus provides a range of tools to allow updaters to strike the
1823 required tradeoff between latency, flexibility and CPU overhead.
1825 <h3><a name="Composability">Composability</a></h3>
1828 Composability has received much attention in recent years, perhaps in part
1829 due to the collision of multicore hardware with object-oriented techniques
1830 designed in single-threaded environments for single-threaded use.
1831 And in theory, RCU read-side critical sections may be composed, and in
1832 fact may be nested arbitrarily deeply.
1833 In practice, as with all real-world implementations of composable
1834 constructs, there are limitations.
1837 Implementations of RCU for which <tt>rcu_read_lock()</tt>
1838 and <tt>rcu_read_unlock()</tt> generate no code, such as
1839 Linux-kernel RCU when <tt>CONFIG_PREEMPT=n</tt>, can be
1840 nested arbitrarily deeply.
1841 After all, there is no overhead.
1842 Except that if all these instances of <tt>rcu_read_lock()</tt>
1843 and <tt>rcu_read_unlock()</tt> are visible to the compiler,
1844 compilation will eventually fail due to exhausting memory,
1845 mass storage, or user patience, whichever comes first.
1846 If the nesting is not visible to the compiler, as is the case with
1847 mutually recursive functions each in its own translation unit,
1848 stack overflow will result.
1849 If the nesting takes the form of loops, perhaps in the guise of tail
1850 recursion, either the control variable
1851 will overflow or (in the Linux kernel) you will get an RCU CPU stall warning.
1852 Nevertheless, this class of RCU implementations is one
1853 of the most composable constructs in existence.
1856 RCU implementations that explicitly track nesting depth
1857 are limited by the nesting-depth counter.
1858 For example, the Linux kernel's preemptible RCU limits nesting to
1860 This should suffice for almost all practical purposes.
1861 That said, a consecutive pair of RCU read-side critical sections
1862 between which there is an operation that waits for a grace period
1863 cannot be enclosed in another RCU read-side critical section.
1864 This is because it is not legal to wait for a grace period within
1865 an RCU read-side critical section: To do so would result either
1867 in RCU implicitly splitting the enclosing RCU read-side critical
1868 section, neither of which is conducive to a long-lived and prosperous
1872 It is worth noting that RCU is not alone in limiting composability.
1873 For example, many transactional-memory implementations prohibit
1874 composing a pair of transactions separated by an irrevocable
1875 operation (for example, a network receive operation).
1876 For another example, lock-based critical sections can be composed
1877 surprisingly freely, but only if deadlock is avoided.
1880 In short, although RCU read-side critical sections are highly composable,
1881 care is required in some situations, just as is the case for any other
1882 composable synchronization mechanism.
1884 <h3><a name="Corner Cases">Corner Cases</a></h3>
1887 A given RCU workload might have an endless and intense stream of
1888 RCU read-side critical sections, perhaps even so intense that there
1889 was never a point in time during which there was not at least one
1890 RCU read-side critical section in flight.
1891 RCU cannot allow this situation to block grace periods: As long as
1892 all the RCU read-side critical sections are finite, grace periods
1893 must also be finite.
1896 That said, preemptible RCU implementations could potentially result
1897 in RCU read-side critical sections being preempted for long durations,
1898 which has the effect of creating a long-duration RCU read-side
1900 This situation can arise only in heavily loaded systems, but systems using
1901 real-time priorities are of course more vulnerable.
1902 Therefore, RCU priority boosting is provided to help deal with this
1904 That said, the exact requirements on RCU priority boosting will likely
1905 evolve as more experience accumulates.
1908 Other workloads might have very high update rates.
1909 Although one can argue that such workloads should instead use
1910 something other than RCU, the fact remains that RCU must
1911 handle such workloads gracefully.
1912 This requirement is another factor driving batching of grace periods,
1913 but it is also the driving force behind the checks for large numbers
1914 of queued RCU callbacks in the <tt>call_rcu()</tt> code path.
1915 Finally, high update rates should not delay RCU read-side critical
1916 sections, although some small read-side delays can occur when using
1917 <tt>synchronize_rcu_expedited()</tt>, courtesy of this function's use
1918 of <tt>smp_call_function_single()</tt>.
1921 Although all three of these corner cases were understood in the early
1922 1990s, a simple user-level test consisting of <tt>close(open(path))</tt>
1924 in the early 2000s suddenly provided a much deeper appreciation of the
1925 high-update-rate corner case.
1926 This test also motivated addition of some RCU code to react to high update
1927 rates, for example, if a given CPU finds itself with more than 10,000
1928 RCU callbacks queued, it will cause RCU to take evasive action by
1929 more aggressively starting grace periods and more aggressively forcing
1930 completion of grace-period processing.
1931 This evasive action causes the grace period to complete more quickly,
1932 but at the cost of restricting RCU's batching optimizations, thus
1933 increasing the CPU overhead incurred by that grace period.
1935 <h2><a name="Software-Engineering Requirements">
1936 Software-Engineering Requirements</a></h2>
1939 Between Murphy's Law and “To err is human”, it is necessary to
1940 guard against mishaps and misuse:
1943 <li> It is all too easy to forget to use <tt>rcu_read_lock()</tt>
1944 everywhere that it is needed, so kernels built with
1945 <tt>CONFIG_PROVE_RCU=y</tt> will splat if
1946 <tt>rcu_dereference()</tt> is used outside of an
1947 RCU read-side critical section.
1948 Update-side code can use <tt>rcu_dereference_protected()</tt>,
1950 <a href="https://lwn.net/Articles/371986/">lockdep expression</a>
1951 to indicate what is providing the protection.
1952 If the indicated protection is not provided, a lockdep splat
1956 Code shared between readers and updaters can use
1957 <tt>rcu_dereference_check()</tt>, which also takes a
1958 lockdep expression, and emits a lockdep splat if neither
1959 <tt>rcu_read_lock()</tt> nor the indicated protection
1961 In addition, <tt>rcu_dereference_raw()</tt> is used in those
1962 (hopefully rare) cases where the required protection cannot
1963 be easily described.
1964 Finally, <tt>rcu_read_lock_held()</tt> is provided to
1965 allow a function to verify that it has been invoked within
1966 an RCU read-side critical section.
1967 I was made aware of this set of requirements shortly after Thomas
1968 Gleixner audited a number of RCU uses.
1969 <li> A given function might wish to check for RCU-related preconditions
1970 upon entry, before using any other RCU API.
1971 The <tt>rcu_lockdep_assert()</tt> does this job,
1972 asserting the expression in kernels having lockdep enabled
1973 and doing nothing otherwise.
1974 <li> It is also easy to forget to use <tt>rcu_assign_pointer()</tt>
1975 and <tt>rcu_dereference()</tt>, perhaps (incorrectly)
1976 substituting a simple assignment.
1977 To catch this sort of error, a given RCU-protected pointer may be
1978 tagged with <tt>__rcu</tt>, after which sparse
1979 will complain about simple-assignment accesses to that pointer.
1980 Arnd Bergmann made me aware of this requirement, and also
1982 <a href="https://lwn.net/Articles/376011/">patch series</a>.
1983 <li> Kernels built with <tt>CONFIG_DEBUG_OBJECTS_RCU_HEAD=y</tt>
1984 will splat if a data element is passed to <tt>call_rcu()</tt>
1985 twice in a row, without a grace period in between.
1986 (This error is similar to a double free.)
1987 The corresponding <tt>rcu_head</tt> structures that are
1988 dynamically allocated are automatically tracked, but
1989 <tt>rcu_head</tt> structures allocated on the stack
1990 must be initialized with <tt>init_rcu_head_on_stack()</tt>
1991 and cleaned up with <tt>destroy_rcu_head_on_stack()</tt>.
1992 Similarly, statically allocated non-stack <tt>rcu_head</tt>
1993 structures must be initialized with <tt>init_rcu_head()</tt>
1994 and cleaned up with <tt>destroy_rcu_head()</tt>.
1995 Mathieu Desnoyers made me aware of this requirement, and also
1997 <a href="https://lkml.kernel.org/g/20100319013024.GA28456@Krystal">patch</a>.
1998 <li> An infinite loop in an RCU read-side critical section will
1999 eventually trigger an RCU CPU stall warning splat, with
2000 the duration of “eventually” being controlled by the
2001 <tt>RCU_CPU_STALL_TIMEOUT</tt> <tt>Kconfig</tt> option, or,
2002 alternatively, by the
2003 <tt>rcupdate.rcu_cpu_stall_timeout</tt> boot/sysfs
2005 However, RCU is not obligated to produce this splat
2006 unless there is a grace period waiting on that particular
2007 RCU read-side critical section.
2009 Some extreme workloads might intentionally delay
2010 RCU grace periods, and systems running those workloads can
2011 be booted with <tt>rcupdate.rcu_cpu_stall_suppress</tt>
2012 to suppress the splats.
2013 This kernel parameter may also be set via <tt>sysfs</tt>.
2014 Furthermore, RCU CPU stall warnings are counter-productive
2015 during sysrq dumps and during panics.
2016 RCU therefore supplies the <tt>rcu_sysrq_start()</tt> and
2017 <tt>rcu_sysrq_end()</tt> API members to be called before
2018 and after long sysrq dumps.
2019 RCU also supplies the <tt>rcu_panic()</tt> notifier that is
2020 automatically invoked at the beginning of a panic to suppress
2021 further RCU CPU stall warnings.
2024 This requirement made itself known in the early 1990s, pretty
2025 much the first time that it was necessary to debug a CPU stall.
2026 That said, the initial implementation in DYNIX/ptx was quite
2027 generic in comparison with that of Linux.
2028 <li> Although it would be very good to detect pointers leaking out
2029 of RCU read-side critical sections, there is currently no
2030 good way of doing this.
2031 One complication is the need to distinguish between pointers
2032 leaking and pointers that have been handed off from RCU to
2033 some other synchronization mechanism, for example, reference
2035 <li> In kernels built with <tt>CONFIG_RCU_TRACE=y</tt>, RCU-related
2036 information is provided via event tracing.
2037 <li> Open-coded use of <tt>rcu_assign_pointer()</tt> and
2038 <tt>rcu_dereference()</tt> to create typical linked
2039 data structures can be surprisingly error-prone.
2040 Therefore, RCU-protected
2041 <a href="https://lwn.net/Articles/609973/#RCU List APIs">linked lists</a>
2042 and, more recently, RCU-protected
2043 <a href="https://lwn.net/Articles/612100/">hash tables</a>
2045 Many other special-purpose RCU-protected data structures are
2046 available in the Linux kernel and the userspace RCU library.
2047 <li> Some linked structures are created at compile time, but still
2048 require <tt>__rcu</tt> checking.
2049 The <tt>RCU_POINTER_INITIALIZER()</tt> macro serves this
2051 <li> It is not necessary to use <tt>rcu_assign_pointer()</tt>
2052 when creating linked structures that are to be published via
2053 a single external pointer.
2054 The <tt>RCU_INIT_POINTER()</tt> macro is provided for
2055 this task and also for assigning <tt>NULL</tt> pointers
2060 This not a hard-and-fast list: RCU's diagnostic capabilities will
2061 continue to be guided by the number and type of usage bugs found
2062 in real-world RCU usage.
2064 <h2><a name="Linux Kernel Complications">Linux Kernel Complications</a></h2>
2067 The Linux kernel provides an interesting environment for all kinds of
2068 software, including RCU.
2069 Some of the relevant points of interest are as follows:
2072 <li> <a href="#Configuration">Configuration</a>.
2073 <li> <a href="#Firmware Interface">Firmware Interface</a>.
2074 <li> <a href="#Early Boot">Early Boot</a>.
2075 <li> <a href="#Interrupts and NMIs">
2076 Interrupts and non-maskable interrupts (NMIs)</a>.
2077 <li> <a href="#Loadable Modules">Loadable Modules</a>.
2078 <li> <a href="#Hotplug CPU">Hotplug CPU</a>.
2079 <li> <a href="#Scheduler and RCU">Scheduler and RCU</a>.
2080 <li> <a href="#Tracing and RCU">Tracing and RCU</a>.
2081 <li> <a href="#Energy Efficiency">Energy Efficiency</a>.
2082 <li> <a href="#Scheduling-Clock Interrupts and RCU">
2083 Scheduling-Clock Interrupts and RCU</a>.
2084 <li> <a href="#Memory Efficiency">Memory Efficiency</a>.
2085 <li> <a href="#Performance, Scalability, Response Time, and Reliability">
2086 Performance, Scalability, Response Time, and Reliability</a>.
2090 This list is probably incomplete, but it does give a feel for the
2091 most notable Linux-kernel complications.
2092 Each of the following sections covers one of the above topics.
2094 <h3><a name="Configuration">Configuration</a></h3>
2097 RCU's goal is automatic configuration, so that almost nobody
2098 needs to worry about RCU's <tt>Kconfig</tt> options.
2099 And for almost all users, RCU does in fact work well
2100 “out of the box.”
2103 However, there are specialized use cases that are handled by
2104 kernel boot parameters and <tt>Kconfig</tt> options.
2105 Unfortunately, the <tt>Kconfig</tt> system will explicitly ask users
2106 about new <tt>Kconfig</tt> options, which requires almost all of them
2107 be hidden behind a <tt>CONFIG_RCU_EXPERT</tt> <tt>Kconfig</tt> option.
2110 This all should be quite obvious, but the fact remains that
2111 Linus Torvalds recently had to
2112 <a href="https://lkml.kernel.org/g/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com">remind</a>
2113 me of this requirement.
2115 <h3><a name="Firmware Interface">Firmware Interface</a></h3>
2118 In many cases, kernel obtains information about the system from the
2119 firmware, and sometimes things are lost in translation.
2120 Or the translation is accurate, but the original message is bogus.
2123 For example, some systems' firmware overreports the number of CPUs,
2124 sometimes by a large factor.
2125 If RCU naively believed the firmware, as it used to do,
2126 it would create too many per-CPU kthreads.
2127 Although the resulting system will still run correctly, the extra
2128 kthreads needlessly consume memory and can cause confusion
2129 when they show up in <tt>ps</tt> listings.
2132 RCU must therefore wait for a given CPU to actually come online before
2133 it can allow itself to believe that the CPU actually exists.
2134 The resulting “ghost CPUs” (which are never going to
2135 come online) cause a number of
2136 <a href="https://paulmck.livejournal.com/37494.html">interesting complications</a>.
2138 <h3><a name="Early Boot">Early Boot</a></h3>
2141 The Linux kernel's boot sequence is an interesting process,
2142 and RCU is used early, even before <tt>rcu_init()</tt>
2144 In fact, a number of RCU's primitives can be used as soon as the
2145 initial task's <tt>task_struct</tt> is available and the
2146 boot CPU's per-CPU variables are set up.
2147 The read-side primitives (<tt>rcu_read_lock()</tt>,
2148 <tt>rcu_read_unlock()</tt>, <tt>rcu_dereference()</tt>,
2149 and <tt>rcu_access_pointer()</tt>) will operate normally very early on,
2150 as will <tt>rcu_assign_pointer()</tt>.
2153 Although <tt>call_rcu()</tt> may be invoked at any
2154 time during boot, callbacks are not guaranteed to be invoked until after
2155 all of RCU's kthreads have been spawned, which occurs at
2156 <tt>early_initcall()</tt> time.
2157 This delay in callback invocation is due to the fact that RCU does not
2158 invoke callbacks until it is fully initialized, and this full initialization
2159 cannot occur until after the scheduler has initialized itself to the
2160 point where RCU can spawn and run its kthreads.
2161 In theory, it would be possible to invoke callbacks earlier,
2162 however, this is not a panacea because there would be severe restrictions
2163 on what operations those callbacks could invoke.
2166 Perhaps surprisingly, <tt>synchronize_rcu()</tt> and
2167 <tt>synchronize_rcu_expedited()</tt>,
2168 will operate normally
2169 during very early boot, the reason being that there is only one CPU
2170 and preemption is disabled.
2171 This means that the call <tt>synchronize_rcu()</tt> (or friends)
2172 itself is a quiescent
2173 state and thus a grace period, so the early-boot implementation can
2177 However, once the scheduler has spawned its first kthread, this early
2178 boot trick fails for <tt>synchronize_rcu()</tt> (as well as for
2179 <tt>synchronize_rcu_expedited()</tt>) in <tt>CONFIG_PREEMPT=y</tt>
2181 The reason is that an RCU read-side critical section might be preempted,
2182 which means that a subsequent <tt>synchronize_rcu()</tt> really does have
2183 to wait for something, as opposed to simply returning immediately.
2184 Unfortunately, <tt>synchronize_rcu()</tt> can't do this until all of
2185 its kthreads are spawned, which doesn't happen until some time during
2186 <tt>early_initcalls()</tt> time.
2187 But this is no excuse: RCU is nevertheless required to correctly handle
2188 synchronous grace periods during this time period.
2189 Once all of its kthreads are up and running, RCU starts running
2193 <tr><th> </th></tr>
2194 <tr><th align="left">Quick Quiz:</th></tr>
2196 How can RCU possibly handle grace periods before all of its
2197 kthreads have been spawned???
2199 <tr><th align="left">Answer:</th></tr>
2200 <tr><td bgcolor="#ffffff"><font color="ffffff">
2204 <p><font color="ffffff">
2205 During the “dead zone” between the time that the
2206 scheduler spawns the first task and the time that all of RCU's
2207 kthreads have been spawned, all synchronous grace periods are
2208 handled by the expedited grace-period mechanism.
2209 At runtime, this expedited mechanism relies on workqueues, but
2210 during the dead zone the requesting task itself drives the
2211 desired expedited grace period.
2212 Because dead-zone execution takes place within task context,
2214 Once the dead zone ends, expedited grace periods go back to
2215 using workqueues, as is required to avoid problems that would
2216 otherwise occur when a user task received a POSIX signal while
2217 driving an expedited grace period.
2220 <p><font color="ffffff">
2221 And yes, this does mean that it is unhelpful to send POSIX
2222 signals to random tasks between the time that the scheduler
2223 spawns its first kthread and the time that RCU's kthreads
2224 have all been spawned.
2225 If there ever turns out to be a good reason for sending POSIX
2226 signals during that time, appropriate adjustments will be made.
2227 (If it turns out that POSIX signals are sent during this time for
2228 no good reason, other adjustments will be made, appropriate
2231 <tr><td> </td></tr>
2235 I learned of these boot-time requirements as a result of a series of
2238 <h3><a name="Interrupts and NMIs">Interrupts and NMIs</a></h3>
2241 The Linux kernel has interrupts, and RCU read-side critical sections are
2242 legal within interrupt handlers and within interrupt-disabled regions
2243 of code, as are invocations of <tt>call_rcu()</tt>.
2246 Some Linux-kernel architectures can enter an interrupt handler from
2247 non-idle process context, and then just never leave it, instead stealthily
2248 transitioning back to process context.
2249 This trick is sometimes used to invoke system calls from inside the kernel.
2250 These “half-interrupts” mean that RCU has to be very careful
2251 about how it counts interrupt nesting levels.
2252 I learned of this requirement the hard way during a rewrite
2253 of RCU's dyntick-idle code.
2256 The Linux kernel has non-maskable interrupts (NMIs), and
2257 RCU read-side critical sections are legal within NMI handlers.
2258 Thankfully, RCU update-side primitives, including
2259 <tt>call_rcu()</tt>, are prohibited within NMI handlers.
2262 The name notwithstanding, some Linux-kernel architectures
2263 can have nested NMIs, which RCU must handle correctly.
2265 <a href="https://lkml.kernel.org/r/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com">surprised me</a>
2266 with this requirement;
2267 he also kindly surprised me with
2268 <a href="https://lkml.kernel.org/r/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com">an algorithm</a>
2269 that meets this requirement.
2272 Furthermore, NMI handlers can be interrupted by what appear to RCU
2273 to be normal interrupts.
2274 One way that this can happen is for code that directly invokes
2275 <tt>rcu_irq_enter()</tt> and </tt>rcu_irq_exit()</tt> to be called
2276 from an NMI handler.
2277 This astonishing fact of life prompted the current code structure,
2278 which has <tt>rcu_irq_enter()</tt> invoking <tt>rcu_nmi_enter()</tt>
2279 and <tt>rcu_irq_exit()</tt> invoking <tt>rcu_nmi_exit()</tt>.
2280 And yes, I also learned of this requirement the hard way.
2282 <h3><a name="Loadable Modules">Loadable Modules</a></h3>
2285 The Linux kernel has loadable modules, and these modules can
2287 After a given module has been unloaded, any attempt to call
2288 one of its functions results in a segmentation fault.
2289 The module-unload functions must therefore cancel any
2290 delayed calls to loadable-module functions, for example,
2291 any outstanding <tt>mod_timer()</tt> must be dealt with
2292 via <tt>del_timer_sync()</tt> or similar.
2295 Unfortunately, there is no way to cancel an RCU callback;
2296 once you invoke <tt>call_rcu()</tt>, the callback function is
2297 going to eventually be invoked, unless the system goes down first.
2298 Because it is normally considered socially irresponsible to crash the system
2299 in response to a module unload request, we need some other way
2300 to deal with in-flight RCU callbacks.
2303 RCU therefore provides
2304 <tt><a href="https://lwn.net/Articles/217484/">rcu_barrier()</a></tt>,
2305 which waits until all in-flight RCU callbacks have been invoked.
2306 If a module uses <tt>call_rcu()</tt>, its exit function should therefore
2307 prevent any future invocation of <tt>call_rcu()</tt>, then invoke
2308 <tt>rcu_barrier()</tt>.
2309 In theory, the underlying module-unload code could invoke
2310 <tt>rcu_barrier()</tt> unconditionally, but in practice this would
2311 incur unacceptable latencies.
2314 Nikita Danilov noted this requirement for an analogous filesystem-unmount
2315 situation, and Dipankar Sarma incorporated <tt>rcu_barrier()</tt> into RCU.
2316 The need for <tt>rcu_barrier()</tt> for module unloading became
2320 <b>Important note</b>: The <tt>rcu_barrier()</tt> function is not,
2321 repeat, <i>not</i>, obligated to wait for a grace period.
2322 It is instead only required to wait for RCU callbacks that have
2323 already been posted.
2324 Therefore, if there are no RCU callbacks posted anywhere in the system,
2325 <tt>rcu_barrier()</tt> is within its rights to return immediately.
2326 Even if there are callbacks posted, <tt>rcu_barrier()</tt> does not
2327 necessarily need to wait for a grace period.
2330 <tr><th> </th></tr>
2331 <tr><th align="left">Quick Quiz:</th></tr>
2334 Each RCU callbacks must wait for a grace period to complete,
2335 and <tt>rcu_barrier()</tt> must wait for each pre-existing
2336 callback to be invoked.
2337 Doesn't <tt>rcu_barrier()</tt> therefore need to wait for
2338 a full grace period if there is even one callback posted anywhere
2341 <tr><th align="left">Answer:</th></tr>
2342 <tr><td bgcolor="#ffffff"><font color="ffffff">
2346 <p><font color="ffffff">
2347 Yes, each RCU callbacks must wait for a grace period to complete,
2348 but it might well be partly (or even completely) finished waiting
2349 by the time <tt>rcu_barrier()</tt> is invoked.
2350 In that case, <tt>rcu_barrier()</tt> need only wait for the
2351 remaining portion of the grace period to elapse.
2352 So even if there are quite a few callbacks posted,
2353 <tt>rcu_barrier()</tt> might well return quite quickly.
2356 <p><font color="ffffff">
2357 So if you need to wait for a grace period as well as for all
2358 pre-existing callbacks, you will need to invoke both
2359 <tt>synchronize_rcu()</tt> and <tt>rcu_barrier()</tt>.
2360 If latency is a concern, you can always use workqueues
2361 to invoke them concurrently.
2363 <tr><td> </td></tr>
2366 <h3><a name="Hotplug CPU">Hotplug CPU</a></h3>
2369 The Linux kernel supports CPU hotplug, which means that CPUs
2371 It is of course illegal to use any RCU API member from an offline CPU,
2372 with the exception of <a href="#Sleepable RCU">SRCU</a> read-side
2374 This requirement was present from day one in DYNIX/ptx, but
2375 on the other hand, the Linux kernel's CPU-hotplug implementation
2376 is “interesting.”
2379 The Linux-kernel CPU-hotplug implementation has notifiers that
2380 are used to allow the various kernel subsystems (including RCU)
2381 to respond appropriately to a given CPU-hotplug operation.
2382 Most RCU operations may be invoked from CPU-hotplug notifiers,
2383 including even synchronous grace-period operations such as
2384 <tt>synchronize_rcu()</tt> and <tt>synchronize_rcu_expedited()</tt>.
2387 However, all-callback-wait operations such as
2388 <tt>rcu_barrier()</tt> are also not supported, due to the
2389 fact that there are phases of CPU-hotplug operations where
2390 the outgoing CPU's callbacks will not be invoked until after
2391 the CPU-hotplug operation ends, which could also result in deadlock.
2392 Furthermore, <tt>rcu_barrier()</tt> blocks CPU-hotplug operations
2393 during its execution, which results in another type of deadlock
2394 when invoked from a CPU-hotplug notifier.
2396 <h3><a name="Scheduler and RCU">Scheduler and RCU</a></h3>
2399 RCU depends on the scheduler, and the scheduler uses RCU to
2400 protect some of its data structures.
2401 The preemptible-RCU <tt>rcu_read_unlock()</tt>
2402 implementation must therefore be written carefully to avoid deadlocks
2403 involving the scheduler's runqueue and priority-inheritance locks.
2404 In particular, <tt>rcu_read_unlock()</tt> must tolerate an
2405 interrupt where the interrupt handler invokes both
2406 <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
2407 This possibility requires <tt>rcu_read_unlock()</tt> to use
2408 negative nesting levels to avoid destructive recursion via
2409 interrupt handler's use of RCU.
2412 This scheduler-RCU requirement came as a
2413 <a href="https://lwn.net/Articles/453002/">complete surprise</a>.
2416 As noted above, RCU makes use of kthreads, and it is necessary to
2417 avoid excessive CPU-time accumulation by these kthreads.
2418 This requirement was no surprise, but RCU's violation of it
2419 when running context-switch-heavy workloads when built with
2420 <tt>CONFIG_NO_HZ_FULL=y</tt>
2421 <a href="http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf">did come as a surprise [PDF]</a>.
2422 RCU has made good progress towards meeting this requirement, even
2423 for context-switch-heavy <tt>CONFIG_NO_HZ_FULL=y</tt> workloads,
2424 but there is room for further improvement.
2427 In the past, it was forbidden to disable interrupts across an
2428 <tt>rcu_read_unlock()</tt> unless that interrupt-disabled region
2429 of code also included the matching <tt>rcu_read_lock()</tt>.
2430 Violating this restriction could result in deadlocks involving the
2431 scheduler's runqueue and priority-inheritance spinlocks.
2432 This restriction was lifted when interrupt-disabled calls to
2433 <tt>rcu_read_unlock()</tt> started deferring the reporting of
2434 the resulting RCU-preempt quiescent state until the end of that
2435 interrupts-disabled region.
2436 This deferred reporting means that the scheduler's runqueue and
2437 priority-inheritance locks cannot be held while reporting an RCU-preempt
2438 quiescent state, which lifts the earlier restriction, at least from
2439 a deadlock perspective.
2440 Unfortunately, real-time systems using RCU priority boosting may
2441 need this restriction to remain in effect because deferred
2442 quiescent-state reporting also defers deboosting, which in turn
2443 degrades real-time latencies.
2445 <h3><a name="Tracing and RCU">Tracing and RCU</a></h3>
2448 It is possible to use tracing on RCU code, but tracing itself
2450 For this reason, <tt>rcu_dereference_raw_notrace()</tt>
2451 is provided for use by tracing, which avoids the destructive
2452 recursion that could otherwise ensue.
2453 This API is also used by virtualization in some architectures,
2454 where RCU readers execute in environments in which tracing
2456 The tracing folks both located the requirement and provided the
2457 needed fix, so this surprise requirement was relatively painless.
2459 <h3><a name="Energy Efficiency">Energy Efficiency</a></h3>
2462 Interrupting idle CPUs is considered socially unacceptable,
2463 especially by people with battery-powered embedded systems.
2464 RCU therefore conserves energy by detecting which CPUs are
2465 idle, including tracking CPUs that have been interrupted from idle.
2466 This is a large part of the energy-efficiency requirement,
2467 so I learned of this via an irate phone call.
2470 Because RCU avoids interrupting idle CPUs, it is illegal to
2471 execute an RCU read-side critical section on an idle CPU.
2472 (Kernels built with <tt>CONFIG_PROVE_RCU=y</tt> will splat
2474 The <tt>RCU_NONIDLE()</tt> macro and <tt>_rcuidle</tt>
2475 event tracing is provided to work around this restriction.
2476 In addition, <tt>rcu_is_watching()</tt> may be used to
2477 test whether or not it is currently legal to run RCU read-side
2478 critical sections on this CPU.
2479 I learned of the need for diagnostics on the one hand
2480 and <tt>RCU_NONIDLE()</tt> on the other while inspecting
2482 Steven Rostedt supplied <tt>_rcuidle</tt> event tracing,
2483 which is used quite heavily in the idle loop.
2484 However, there are some restrictions on the code placed within
2485 <tt>RCU_NONIDLE()</tt>:
2488 <li> Blocking is prohibited.
2489 In practice, this is not a serious restriction given that idle
2490 tasks are prohibited from blocking to begin with.
2491 <li> Although nesting <tt>RCU_NONIDLE()</tt> is permitted, they cannot
2492 nest indefinitely deeply.
2493 However, given that they can be nested on the order of a million
2494 deep, even on 32-bit systems, this should not be a serious
2496 This nesting limit would probably be reached long after the
2497 compiler OOMed or the stack overflowed.
2498 <li> Any code path that enters <tt>RCU_NONIDLE()</tt> must sequence
2499 out of that same <tt>RCU_NONIDLE()</tt>.
2500 For example, the following is grossly illegal:
2506 3 goto bad_idea; /* BUG!!! */
2507 4 do_something_else();});
2513 It is just as illegal to transfer control into the middle of
2514 <tt>RCU_NONIDLE()</tt>'s argument.
2515 Yes, in theory, you could transfer in as long as you also
2516 transferred out, but in practice you could also expect to get sharply
2517 worded review comments.
2521 It is similarly socially unacceptable to interrupt an
2522 <tt>nohz_full</tt> CPU running in userspace.
2523 RCU must therefore track <tt>nohz_full</tt> userspace
2525 RCU must therefore be able to sample state at two points in
2526 time, and be able to determine whether or not some other CPU spent
2527 any time idle and/or executing in userspace.
2530 These energy-efficiency requirements have proven quite difficult to
2531 understand and to meet, for example, there have been more than five
2532 clean-sheet rewrites of RCU's energy-efficiency code, the last of
2533 which was finally able to demonstrate
2534 <a href="http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf">real energy savings running on real hardware [PDF]</a>.
2536 I learned of many of these requirements via angry phone calls:
2537 Flaming me on the Linux-kernel mailing list was apparently not
2538 sufficient to fully vent their ire at RCU's energy-efficiency bugs!
2540 <h3><a name="Scheduling-Clock Interrupts and RCU">
2541 Scheduling-Clock Interrupts and RCU</a></h3>
2544 The kernel transitions between in-kernel non-idle execution, userspace
2545 execution, and the idle loop.
2546 Depending on kernel configuration, RCU handles these states differently:
2549 <tr><th><tt>HZ</tt> Kconfig</th>
2553 <tr><th align="left"><tt>HZ_PERIODIC</tt></th>
2554 <td>Can rely on scheduling-clock interrupt.</td>
2555 <td>Can rely on scheduling-clock interrupt and its
2556 detection of interrupt from usermode.</td>
2557 <td>Can rely on RCU's dyntick-idle detection.</td></tr>
2558 <tr><th align="left"><tt>NO_HZ_IDLE</tt></th>
2559 <td>Can rely on scheduling-clock interrupt.</td>
2560 <td>Can rely on scheduling-clock interrupt and its
2561 detection of interrupt from usermode.</td>
2562 <td>Can rely on RCU's dyntick-idle detection.</td></tr>
2563 <tr><th align="left"><tt>NO_HZ_FULL</tt></th>
2564 <td>Can only sometimes rely on scheduling-clock interrupt.
2565 In other cases, it is necessary to bound kernel execution
2566 times and/or use IPIs.</td>
2567 <td>Can rely on RCU's dyntick-idle detection.</td>
2568 <td>Can rely on RCU's dyntick-idle detection.</td></tr>
2572 <tr><th> </th></tr>
2573 <tr><th align="left">Quick Quiz:</th></tr>
2575 Why can't <tt>NO_HZ_FULL</tt> in-kernel execution rely on the
2576 scheduling-clock interrupt, just like <tt>HZ_PERIODIC</tt>
2577 and <tt>NO_HZ_IDLE</tt> do?
2579 <tr><th align="left">Answer:</th></tr>
2580 <tr><td bgcolor="#ffffff"><font color="ffffff">
2581 Because, as a performance optimization, <tt>NO_HZ_FULL</tt>
2582 does not necessarily re-enable the scheduling-clock interrupt
2583 on entry to each and every system call.
2585 <tr><td> </td></tr>
2589 However, RCU must be reliably informed as to whether any given
2590 CPU is currently in the idle loop, and, for <tt>NO_HZ_FULL</tt>,
2591 also whether that CPU is executing in usermode, as discussed
2592 <a href="#Energy Efficiency">earlier</a>.
2593 It also requires that the scheduling-clock interrupt be enabled when
2597 <li> If a CPU is either idle or executing in usermode, and RCU believes
2598 it is non-idle, the scheduling-clock tick had better be running.
2599 Otherwise, you will get RCU CPU stall warnings. Or at best,
2600 very long (11-second) grace periods, with a pointless IPI waking
2601 the CPU from time to time.
2602 <li> If a CPU is in a portion of the kernel that executes RCU read-side
2603 critical sections, and RCU believes this CPU to be idle, you will get
2604 random memory corruption. <b>DON'T DO THIS!!!</b>
2606 <br>This is one reason to test with lockdep, which will complain
2607 about this sort of thing.
2608 <li> If a CPU is in a portion of the kernel that is absolutely
2609 positively no-joking guaranteed to never execute any RCU read-side
2610 critical sections, and RCU believes this CPU to to be idle,
2611 no problem. This sort of thing is used by some architectures
2612 for light-weight exception handlers, which can then avoid the
2613 overhead of <tt>rcu_irq_enter()</tt> and <tt>rcu_irq_exit()</tt>
2614 at exception entry and exit, respectively.
2615 Some go further and avoid the entireties of <tt>irq_enter()</tt>
2616 and <tt>irq_exit()</tt>.
2618 <br>Just make very sure you are running some of your tests with
2619 <tt>CONFIG_PROVE_RCU=y</tt>, just in case one of your code paths
2620 was in fact joking about not doing RCU read-side critical sections.
2621 <li> If a CPU is executing in the kernel with the scheduling-clock
2622 interrupt disabled and RCU believes this CPU to be non-idle,
2623 and if the CPU goes idle (from an RCU perspective) every few
2624 jiffies, no problem. It is usually OK for there to be the
2625 occasional gap between idle periods of up to a second or so.
2627 <br>If the gap grows too long, you get RCU CPU stall warnings.
2628 <li> If a CPU is either idle or executing in usermode, and RCU believes
2629 it to be idle, of course no problem.
2630 <li> If a CPU is executing in the kernel, the kernel code
2631 path is passing through quiescent states at a reasonable
2632 frequency (preferably about once per few jiffies, but the
2633 occasional excursion to a second or so is usually OK) and the
2634 scheduling-clock interrupt is enabled, of course no problem.
2636 <br>If the gap between a successive pair of quiescent states grows
2637 too long, you get RCU CPU stall warnings.
2641 <tr><th> </th></tr>
2642 <tr><th align="left">Quick Quiz:</th></tr>
2644 But what if my driver has a hardware interrupt handler
2645 that can run for many seconds?
2646 I cannot invoke <tt>schedule()</tt> from an hardware
2647 interrupt handler, after all!
2649 <tr><th align="left">Answer:</th></tr>
2650 <tr><td bgcolor="#ffffff"><font color="ffffff">
2651 One approach is to do <tt>rcu_irq_exit();rcu_irq_enter();</tt>
2653 But given that long-running interrupt handlers can cause
2654 other problems, not least for response time, shouldn't you
2655 work to keep your interrupt handler's runtime within reasonable
2658 <tr><td> </td></tr>
2662 But as long as RCU is properly informed of kernel state transitions between
2663 in-kernel execution, usermode execution, and idle, and as long as the
2664 scheduling-clock interrupt is enabled when RCU needs it to be, you
2665 can rest assured that the bugs you encounter will be in some other
2666 part of RCU or some other part of the kernel!
2668 <h3><a name="Memory Efficiency">Memory Efficiency</a></h3>
2671 Although small-memory non-realtime systems can simply use Tiny RCU,
2672 code size is only one aspect of memory efficiency.
2673 Another aspect is the size of the <tt>rcu_head</tt> structure
2674 used by <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>.
2675 Although this structure contains nothing more than a pair of pointers,
2676 it does appear in many RCU-protected data structures, including
2677 some that are size critical.
2678 The <tt>page</tt> structure is a case in point, as evidenced by
2679 the many occurrences of the <tt>union</tt> keyword within that structure.
2682 This need for memory efficiency is one reason that RCU uses hand-crafted
2683 singly linked lists to track the <tt>rcu_head</tt> structures that
2684 are waiting for a grace period to elapse.
2685 It is also the reason why <tt>rcu_head</tt> structures do not contain
2686 debug information, such as fields tracking the file and line of the
2687 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> that posted them.
2688 Although this information might appear in debug-only kernel builds at some
2689 point, in the meantime, the <tt>->func</tt> field will often provide
2690 the needed debug information.
2693 However, in some cases, the need for memory efficiency leads to even
2694 more extreme measures.
2695 Returning to the <tt>page</tt> structure, the <tt>rcu_head</tt> field
2696 shares storage with a great many other structures that are used at
2697 various points in the corresponding page's lifetime.
2698 In order to correctly resolve certain
2699 <a href="https://lkml.kernel.org/g/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com">race conditions</a>,
2700 the Linux kernel's memory-management subsystem needs a particular bit
2701 to remain zero during all phases of grace-period processing,
2702 and that bit happens to map to the bottom bit of the
2703 <tt>rcu_head</tt> structure's <tt>->next</tt> field.
2704 RCU makes this guarantee as long as <tt>call_rcu()</tt>
2705 is used to post the callback, as opposed to <tt>kfree_rcu()</tt>
2706 or some future “lazy”
2707 variant of <tt>call_rcu()</tt> that might one day be created for
2708 energy-efficiency purposes.
2711 That said, there are limits.
2712 RCU requires that the <tt>rcu_head</tt> structure be aligned to a
2713 two-byte boundary, and passing a misaligned <tt>rcu_head</tt>
2714 structure to one of the <tt>call_rcu()</tt> family of functions
2715 will result in a splat.
2716 It is therefore necessary to exercise caution when packing
2717 structures containing fields of type <tt>rcu_head</tt>.
2718 Why not a four-byte or even eight-byte alignment requirement?
2719 Because the m68k architecture provides only two-byte alignment,
2720 and thus acts as alignment's least common denominator.
2723 The reason for reserving the bottom bit of pointers to
2724 <tt>rcu_head</tt> structures is to leave the door open to
2725 “lazy” callbacks whose invocations can safely be deferred.
2726 Deferring invocation could potentially have energy-efficiency
2727 benefits, but only if the rate of non-lazy callbacks decreases
2728 significantly for some important workload.
2729 In the meantime, reserving the bottom bit keeps this option open
2730 in case it one day becomes useful.
2732 <h3><a name="Performance, Scalability, Response Time, and Reliability">
2733 Performance, Scalability, Response Time, and Reliability</a></h3>
2737 <a href="#Performance and Scalability">earlier discussion</a>,
2738 RCU is used heavily by hot code paths in performance-critical
2739 portions of the Linux kernel's networking, security, virtualization,
2740 and scheduling code paths.
2741 RCU must therefore use efficient implementations, especially in its
2742 read-side primitives.
2743 To that end, it would be good if preemptible RCU's implementation
2744 of <tt>rcu_read_lock()</tt> could be inlined, however, doing
2745 this requires resolving <tt>#include</tt> issues with the
2746 <tt>task_struct</tt> structure.
2749 The Linux kernel supports hardware configurations with up to
2750 4096 CPUs, which means that RCU must be extremely scalable.
2751 Algorithms that involve frequent acquisitions of global locks or
2752 frequent atomic operations on global variables simply cannot be
2753 tolerated within the RCU implementation.
2754 RCU therefore makes heavy use of a combining tree based on the
2755 <tt>rcu_node</tt> structure.
2756 RCU is required to tolerate all CPUs continuously invoking any
2757 combination of RCU's runtime primitives with minimal per-operation
2759 In fact, in many cases, increasing load must <i>decrease</i> the
2760 per-operation overhead, witness the batching optimizations for
2761 <tt>synchronize_rcu()</tt>, <tt>call_rcu()</tt>,
2762 <tt>synchronize_rcu_expedited()</tt>, and <tt>rcu_barrier()</tt>.
2763 As a general rule, RCU must cheerfully accept whatever the
2764 rest of the Linux kernel decides to throw at it.
2767 The Linux kernel is used for real-time workloads, especially
2768 in conjunction with the
2769 <a href="https://rt.wiki.kernel.org/index.php/Main_Page">-rt patchset</a>.
2770 The real-time-latency response requirements are such that the
2771 traditional approach of disabling preemption across RCU
2772 read-side critical sections is inappropriate.
2773 Kernels built with <tt>CONFIG_PREEMPT=y</tt> therefore
2774 use an RCU implementation that allows RCU read-side critical
2775 sections to be preempted.
2776 This requirement made its presence known after users made it
2777 clear that an earlier
2778 <a href="https://lwn.net/Articles/107930/">real-time patch</a>
2779 did not meet their needs, in conjunction with some
2780 <a href="https://lkml.kernel.org/g/20050318002026.GA2693@us.ibm.com">RCU issues</a>
2781 encountered by a very early version of the -rt patchset.
2784 In addition, RCU must make do with a sub-100-microsecond real-time latency
2786 In fact, on smaller systems with the -rt patchset, the Linux kernel
2787 provides sub-20-microsecond real-time latencies for the whole kernel,
2789 RCU's scalability and latency must therefore be sufficient for
2790 these sorts of configurations.
2791 To my surprise, the sub-100-microsecond real-time latency budget
2792 <a href="http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf">
2793 applies to even the largest systems [PDF]</a>,
2794 up to and including systems with 4096 CPUs.
2795 This real-time requirement motivated the grace-period kthread, which
2796 also simplified handling of a number of race conditions.
2799 RCU must avoid degrading real-time response for CPU-bound threads, whether
2800 executing in usermode (which is one use case for
2801 <tt>CONFIG_NO_HZ_FULL=y</tt>) or in the kernel.
2802 That said, CPU-bound loops in the kernel must execute
2803 <tt>cond_resched()</tt> at least once per few tens of milliseconds
2804 in order to avoid receiving an IPI from RCU.
2807 Finally, RCU's status as a synchronization primitive means that
2808 any RCU failure can result in arbitrary memory corruption that can be
2809 extremely difficult to debug.
2810 This means that RCU must be extremely reliable, which in
2811 practice also means that RCU must have an aggressive stress-test
2813 This stress-test suite is called <tt>rcutorture</tt>.
2816 Although the need for <tt>rcutorture</tt> was no surprise,
2817 the current immense popularity of the Linux kernel is posing
2818 interesting—and perhaps unprecedented—validation
2820 To see this, keep in mind that there are well over one billion
2821 instances of the Linux kernel running today, given Android
2822 smartphones, Linux-powered televisions, and servers.
2823 This number can be expected to increase sharply with the advent of
2824 the celebrated Internet of Things.
2827 Suppose that RCU contains a race condition that manifests on average
2828 once per million years of runtime.
2829 This bug will be occurring about three times per <i>day</i> across
2831 RCU could simply hide behind hardware error rates, given that no one
2832 should really expect their smartphone to last for a million years.
2833 However, anyone taking too much comfort from this thought should
2834 consider the fact that in most jurisdictions, a successful multi-year
2835 test of a given mechanism, which might include a Linux kernel,
2836 suffices for a number of types of safety-critical certifications.
2837 In fact, rumor has it that the Linux kernel is already being used
2838 in production for safety-critical applications.
2839 I don't know about you, but I would feel quite bad if a bug in RCU
2841 Which might explain my recent focus on validation and verification.
2843 <h2><a name="Other RCU Flavors">Other RCU Flavors</a></h2>
2846 One of the more surprising things about RCU is that there are now
2847 no fewer than five <i>flavors</i>, or API families.
2848 In addition, the primary flavor that has been the sole focus up to
2849 this point has two different implementations, non-preemptible and
2851 The other four flavors are listed below, with requirements for each
2852 described in a separate section.
2855 <li> <a href="#Bottom-Half Flavor">Bottom-Half Flavor (Historical)</a>
2856 <li> <a href="#Sched Flavor">Sched Flavor (Historical)</a>
2857 <li> <a href="#Sleepable RCU">Sleepable RCU</a>
2858 <li> <a href="#Tasks RCU">Tasks RCU</a>
2861 <h3><a name="Bottom-Half Flavor">Bottom-Half Flavor (Historical)</a></h3>
2864 The RCU-bh flavor of RCU has since been expressed in terms of
2865 the other RCU flavors as part of a consolidation of the three
2866 flavors into a single flavor.
2867 The read-side API remains, and continues to disable softirq and to
2868 be accounted for by lockdep.
2869 Much of the material in this section is therefore strictly historical
2873 The softirq-disable (AKA “bottom-half”,
2874 hence the “_bh” abbreviations)
2875 flavor of RCU, or <i>RCU-bh</i>, was developed by
2876 Dipankar Sarma to provide a flavor of RCU that could withstand the
2877 network-based denial-of-service attacks researched by Robert
2879 These attacks placed so much networking load on the system
2880 that some of the CPUs never exited softirq execution,
2881 which in turn prevented those CPUs from ever executing a context switch,
2882 which, in the RCU implementation of that time, prevented grace periods
2884 The result was an out-of-memory condition and a system hang.
2887 The solution was the creation of RCU-bh, which does
2888 <tt>local_bh_disable()</tt>
2889 across its read-side critical sections, and which uses the transition
2890 from one type of softirq processing to another as a quiescent state
2891 in addition to context switch, idle, user mode, and offline.
2892 This means that RCU-bh grace periods can complete even when some of
2893 the CPUs execute in softirq indefinitely, thus allowing algorithms
2894 based on RCU-bh to withstand network-based denial-of-service attacks.
2898 <tt>rcu_read_lock_bh()</tt> and <tt>rcu_read_unlock_bh()</tt>
2899 disable and re-enable softirq handlers, any attempt to start a softirq
2901 RCU-bh read-side critical section will be deferred.
2902 In this case, <tt>rcu_read_unlock_bh()</tt>
2903 will invoke softirq processing, which can take considerable time.
2904 One can of course argue that this softirq overhead should be associated
2905 with the code following the RCU-bh read-side critical section rather
2906 than <tt>rcu_read_unlock_bh()</tt>, but the fact
2907 is that most profiling tools cannot be expected to make this sort
2908 of fine distinction.
2909 For example, suppose that a three-millisecond-long RCU-bh read-side
2910 critical section executes during a time of heavy networking load.
2911 There will very likely be an attempt to invoke at least one softirq
2912 handler during that three milliseconds, but any such invocation will
2913 be delayed until the time of the <tt>rcu_read_unlock_bh()</tt>.
2914 This can of course make it appear at first glance as if
2915 <tt>rcu_read_unlock_bh()</tt> was executing very slowly.
2919 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-bh API</a>
2921 <tt>rcu_read_lock_bh()</tt>,
2922 <tt>rcu_read_unlock_bh()</tt>,
2923 <tt>rcu_dereference_bh()</tt>,
2924 <tt>rcu_dereference_bh_check()</tt>,
2925 <tt>synchronize_rcu_bh()</tt>,
2926 <tt>synchronize_rcu_bh_expedited()</tt>,
2927 <tt>call_rcu_bh()</tt>,
2928 <tt>rcu_barrier_bh()</tt>, and
2929 <tt>rcu_read_lock_bh_held()</tt>.
2930 However, the update-side APIs are now simple wrappers for other RCU
2931 flavors, namely RCU-sched in CONFIG_PREEMPT=n kernels and RCU-preempt
2934 <h3><a name="Sched Flavor">Sched Flavor (Historical)</a></h3>
2937 The RCU-sched flavor of RCU has since been expressed in terms of
2938 the other RCU flavors as part of a consolidation of the three
2939 flavors into a single flavor.
2940 The read-side API remains, and continues to disable preemption and to
2941 be accounted for by lockdep.
2942 Much of the material in this section is therefore strictly historical
2946 Before preemptible RCU, waiting for an RCU grace period had the
2947 side effect of also waiting for all pre-existing interrupt
2949 However, there are legitimate preemptible-RCU implementations that
2950 do not have this property, given that any point in the code outside
2951 of an RCU read-side critical section can be a quiescent state.
2952 Therefore, <i>RCU-sched</i> was created, which follows “classic”
2953 RCU in that an RCU-sched grace period waits for for pre-existing
2954 interrupt and NMI handlers.
2955 In kernels built with <tt>CONFIG_PREEMPT=n</tt>, the RCU and RCU-sched
2956 APIs have identical implementations, while kernels built with
2957 <tt>CONFIG_PREEMPT=y</tt> provide a separate implementation for each.
2960 Note well that in <tt>CONFIG_PREEMPT=y</tt> kernels,
2961 <tt>rcu_read_lock_sched()</tt> and <tt>rcu_read_unlock_sched()</tt>
2962 disable and re-enable preemption, respectively.
2963 This means that if there was a preemption attempt during the
2964 RCU-sched read-side critical section, <tt>rcu_read_unlock_sched()</tt>
2965 will enter the scheduler, with all the latency and overhead entailed.
2966 Just as with <tt>rcu_read_unlock_bh()</tt>, this can make it look
2967 as if <tt>rcu_read_unlock_sched()</tt> was executing very slowly.
2968 However, the highest-priority task won't be preempted, so that task
2969 will enjoy low-overhead <tt>rcu_read_unlock_sched()</tt> invocations.
2973 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-sched API</a>
2975 <tt>rcu_read_lock_sched()</tt>,
2976 <tt>rcu_read_unlock_sched()</tt>,
2977 <tt>rcu_read_lock_sched_notrace()</tt>,
2978 <tt>rcu_read_unlock_sched_notrace()</tt>,
2979 <tt>rcu_dereference_sched()</tt>,
2980 <tt>rcu_dereference_sched_check()</tt>,
2981 <tt>synchronize_sched()</tt>,
2982 <tt>synchronize_rcu_sched_expedited()</tt>,
2983 <tt>call_rcu_sched()</tt>,
2984 <tt>rcu_barrier_sched()</tt>, and
2985 <tt>rcu_read_lock_sched_held()</tt>.
2986 However, anything that disables preemption also marks an RCU-sched
2987 read-side critical section, including
2988 <tt>preempt_disable()</tt> and <tt>preempt_enable()</tt>,
2989 <tt>local_irq_save()</tt> and <tt>local_irq_restore()</tt>,
2992 <h3><a name="Sleepable RCU">Sleepable RCU</a></h3>
2995 For well over a decade, someone saying “I need to block within
2996 an RCU read-side critical section” was a reliable indication
2997 that this someone did not understand RCU.
2998 After all, if you are always blocking in an RCU read-side critical
2999 section, you can probably afford to use a higher-overhead synchronization
3001 However, that changed with the advent of the Linux kernel's notifiers,
3002 whose RCU read-side critical
3003 sections almost never sleep, but sometimes need to.
3004 This resulted in the introduction of
3005 <a href="https://lwn.net/Articles/202847/">sleepable RCU</a>,
3009 SRCU allows different domains to be defined, with each such domain
3010 defined by an instance of an <tt>srcu_struct</tt> structure.
3011 A pointer to this structure must be passed in to each SRCU function,
3012 for example, <tt>synchronize_srcu(&ss)</tt>, where
3013 <tt>ss</tt> is the <tt>srcu_struct</tt> structure.
3014 The key benefit of these domains is that a slow SRCU reader in one
3015 domain does not delay an SRCU grace period in some other domain.
3016 That said, one consequence of these domains is that read-side code
3017 must pass a “cookie” from <tt>srcu_read_lock()</tt>
3018 to <tt>srcu_read_unlock()</tt>, for example, as follows:
3024 3 idx = srcu_read_lock(&ss);
3026 5 srcu_read_unlock(&ss, idx);
3031 As noted above, it is legal to block within SRCU read-side critical sections,
3032 however, with great power comes great responsibility.
3033 If you block forever in one of a given domain's SRCU read-side critical
3034 sections, then that domain's grace periods will also be blocked forever.
3035 Of course, one good way to block forever is to deadlock, which can
3036 happen if any operation in a given domain's SRCU read-side critical
3037 section can block waiting, either directly or indirectly, for that domain's
3038 grace period to elapse.
3039 For example, this results in a self-deadlock:
3045 3 idx = srcu_read_lock(&ss);
3047 5 synchronize_srcu(&ss);
3048 6 srcu_read_unlock(&ss, idx);
3053 However, if line 5 acquired a mutex that was held across
3054 a <tt>synchronize_srcu()</tt> for domain <tt>ss</tt>,
3055 deadlock would still be possible.
3056 Furthermore, if line 5 acquired a mutex that was held across
3057 a <tt>synchronize_srcu()</tt> for some other domain <tt>ss1</tt>,
3058 and if an <tt>ss1</tt>-domain SRCU read-side critical section
3059 acquired another mutex that was held across as <tt>ss</tt>-domain
3060 <tt>synchronize_srcu()</tt>,
3061 deadlock would again be possible.
3062 Such a deadlock cycle could extend across an arbitrarily large number
3063 of different SRCU domains.
3064 Again, with great power comes great responsibility.
3067 Unlike the other RCU flavors, SRCU read-side critical sections can
3068 run on idle and even offline CPUs.
3069 This ability requires that <tt>srcu_read_lock()</tt> and
3070 <tt>srcu_read_unlock()</tt> contain memory barriers, which means
3071 that SRCU readers will run a bit slower than would RCU readers.
3072 It also motivates the <tt>smp_mb__after_srcu_read_unlock()</tt>
3073 API, which, in combination with <tt>srcu_read_unlock()</tt>,
3074 guarantees a full memory barrier.
3077 Also unlike other RCU flavors, SRCU's callbacks-wait function
3078 <tt>srcu_barrier()</tt> may be invoked from CPU-hotplug notifiers,
3079 though this is not necessarily a good idea.
3080 The reason that this is possible is that SRCU is insensitive
3081 to whether or not a CPU is online, which means that <tt>srcu_barrier()</tt>
3082 need not exclude CPU-hotplug operations.
3085 SRCU also differs from other RCU flavors in that SRCU's expedited and
3086 non-expedited grace periods are implemented by the same mechanism.
3087 This means that in the current SRCU implementation, expediting a
3088 future grace period has the side effect of expediting all prior
3089 grace periods that have not yet completed.
3090 (But please note that this is a property of the current implementation,
3091 not necessarily of future implementations.)
3092 In addition, if SRCU has been idle for longer than the interval
3093 specified by the <tt>srcutree.exp_holdoff</tt> kernel boot parameter
3094 (25 microseconds by default),
3095 and if a <tt>synchronize_srcu()</tt> invocation ends this idle period,
3096 that invocation will be automatically expedited.
3099 As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating
3100 a locking bottleneck present in prior kernel versions.
3101 Although this will allow users to put much heavier stress on
3102 <tt>call_srcu()</tt>, it is important to note that SRCU does not
3103 yet take any special steps to deal with callback flooding.
3104 So if you are posting (say) 10,000 SRCU callbacks per second per CPU,
3105 you are probably totally OK, but if you intend to post (say) 1,000,000
3106 SRCU callbacks per second per CPU, please run some tests first.
3107 SRCU just might need a few adjustment to deal with that sort of load.
3108 Of course, your mileage may vary based on the speed of your CPUs and
3109 the size of your memory.
3113 <a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">SRCU API</a>
3115 <tt>srcu_read_lock()</tt>,
3116 <tt>srcu_read_unlock()</tt>,
3117 <tt>srcu_dereference()</tt>,
3118 <tt>srcu_dereference_check()</tt>,
3119 <tt>synchronize_srcu()</tt>,
3120 <tt>synchronize_srcu_expedited()</tt>,
3121 <tt>call_srcu()</tt>,
3122 <tt>srcu_barrier()</tt>, and
3123 <tt>srcu_read_lock_held()</tt>.
3125 <tt>DEFINE_SRCU()</tt>,
3126 <tt>DEFINE_STATIC_SRCU()</tt>, and
3127 <tt>init_srcu_struct()</tt>
3128 APIs for defining and initializing <tt>srcu_struct</tt> structures.
3130 <h3><a name="Tasks RCU">Tasks RCU</a></h3>
3133 Some forms of tracing use “trampolines” to handle the
3134 binary rewriting required to install different types of probes.
3135 It would be good to be able to free old trampolines, which sounds
3136 like a job for some form of RCU.
3137 However, because it is necessary to be able to install a trace
3138 anywhere in the code, it is not possible to use read-side markers
3139 such as <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>.
3140 In addition, it does not work to have these markers in the trampoline
3141 itself, because there would need to be instructions following
3142 <tt>rcu_read_unlock()</tt>.
3143 Although <tt>synchronize_rcu()</tt> would guarantee that execution
3144 reached the <tt>rcu_read_unlock()</tt>, it would not be able to
3145 guarantee that execution had completely left the trampoline.
3148 The solution, in the form of
3149 <a href="https://lwn.net/Articles/607117/"><i>Tasks RCU</i></a>,
3151 read-side critical sections that are delimited by voluntary context
3152 switches, that is, calls to <tt>schedule()</tt>,
3153 <tt>cond_resched()</tt>, and
3154 <tt>synchronize_rcu_tasks()</tt>.
3155 In addition, transitions to and from userspace execution also delimit
3156 tasks-RCU read-side critical sections.
3159 The tasks-RCU API is quite compact, consisting only of
3160 <tt>call_rcu_tasks()</tt>,
3161 <tt>synchronize_rcu_tasks()</tt>, and
3162 <tt>rcu_barrier_tasks()</tt>.
3163 In <tt>CONFIG_PREEMPT=n</tt> kernels, trampolines cannot be preempted,
3164 so these APIs map to
3165 <tt>call_rcu()</tt>,
3166 <tt>synchronize_rcu()</tt>, and
3167 <tt>rcu_barrier()</tt>, respectively.
3168 In <tt>CONFIG_PREEMPT=y</tt> kernels, trampolines can be preempted,
3169 and these three APIs are therefore implemented by separate functions
3170 that check for voluntary context switches.
3172 <h2><a name="Possible Future Changes">Possible Future Changes</a></h2>
3175 One of the tricks that RCU uses to attain update-side scalability is
3176 to increase grace-period latency with increasing numbers of CPUs.
3177 If this becomes a serious problem, it will be necessary to rework the
3178 grace-period state machine so as to avoid the need for the additional
3182 RCU disables CPU hotplug in a few places, perhaps most notably in the
3183 <tt>rcu_barrier()</tt> operations.
3184 If there is a strong reason to use <tt>rcu_barrier()</tt> in CPU-hotplug
3185 notifiers, it will be necessary to avoid disabling CPU hotplug.
3186 This would introduce some complexity, so there had better be a <i>very</i>
3190 The tradeoff between grace-period latency on the one hand and interruptions
3191 of other CPUs on the other hand may need to be re-examined.
3192 The desire is of course for zero grace-period latency as well as zero
3193 interprocessor interrupts undertaken during an expedited grace period
3195 While this ideal is unlikely to be achievable, it is quite possible that
3196 further improvements can be made.
3199 The multiprocessor implementations of RCU use a combining tree that
3200 groups CPUs so as to reduce lock contention and increase cache locality.
3201 However, this combining tree does not spread its memory across NUMA
3202 nodes nor does it align the CPU groups with hardware features such
3203 as sockets or cores.
3204 Such spreading and alignment is currently believed to be unnecessary
3205 because the hotpath read-side primitives do not access the combining
3206 tree, nor does <tt>call_rcu()</tt> in the common case.
3207 If you believe that your architecture needs such spreading and alignment,
3208 then your architecture should also benefit from the
3209 <tt>rcutree.rcu_fanout_leaf</tt> boot parameter, which can be set
3210 to the number of CPUs in a socket, NUMA node, or whatever.
3211 If the number of CPUs is too large, use a fraction of the number of
3213 If the number of CPUs is a large prime number, well, that certainly
3214 is an “interesting” architectural choice!
3215 More flexible arrangements might be considered, but only if
3216 <tt>rcutree.rcu_fanout_leaf</tt> has proven inadequate, and only
3217 if the inadequacy has been demonstrated by a carefully run and
3218 realistic system-level workload.
3221 Please note that arrangements that require RCU to remap CPU numbers will
3222 require extremely good demonstration of need and full exploration of
3226 RCU's various kthreads are reasonably recent additions.
3227 It is quite likely that adjustments will be required to more gracefully
3228 handle extreme loads.
3229 It might also be necessary to be able to relate CPU utilization by
3230 RCU's kthreads and softirq handlers to the code that instigated this
3232 For example, RCU callback overhead might be charged back to the
3233 originating <tt>call_rcu()</tt> instance, though probably not
3234 in production kernels.
3236 <h2><a name="Summary">Summary</a></h2>
3239 This document has presented more than two decade's worth of RCU
3241 Given that the requirements keep changing, this will not be the last
3242 word on this subject, but at least it serves to get an important
3243 subset of the requirements set forth.
3245 <h2><a name="Acknowledgments">Acknowledgments</a></h2>
3247 I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar,
3248 Oleg Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and
3249 Andy Lutomirski for their help in rendering
3250 this article human readable, and to Michelle Rankin for her support
3252 Other contributions are acknowledged in the Linux kernel's git archive.
git.samba.org / sfrench / cifs-2.6.git / blob