1=================================
2A Tour Through RCU's Requirements
3=================================
4
5Copyright IBM Corporation, 2015
6
7Author: Paul E. McKenney
8
9The initial version of this document appeared in the
10`LWN <https://lwn.net/>`_ on those articles:
11`part 1 <https://lwn.net/Articles/652156/>`_,
12`part 2 <https://lwn.net/Articles/652677/>`_, and
13`part 3 <https://lwn.net/Articles/653326/>`_.
14
15Introduction
16------------
17
18Read-copy update (RCU) is a synchronization mechanism that is often used
19as a replacement for reader-writer locking. RCU is unusual in that
20updaters do not block readers, which means that RCU's read-side
21primitives can be exceedingly fast and scalable. In addition, updaters
22can make useful forward progress concurrently with readers. However, all
23this concurrency between RCU readers and updaters does raise the
24question of exactly what RCU readers are doing, which in turn raises the
25question of exactly what RCU's requirements are.
26
27This document therefore summarizes RCU's requirements, and can be
28thought of as an informal, high-level specification for RCU. It is
29important to understand that RCU's specification is primarily empirical
30in nature; in fact, I learned about many of these requirements the hard
31way. This situation might cause some consternation, however, not only
32has this learning process been a lot of fun, but it has also been a
33great privilege to work with so many people willing to apply
34technologies in interesting new ways.
35
36All that aside, here are the categories of currently known RCU
37requirements:
38
39#. `Fundamental Requirements`_
40#. `Fundamental Non-Requirements`_
41#. `Parallelism Facts of Life`_
42#. `Quality-of-Implementation Requirements`_
43#. `Linux Kernel Complications`_
44#. `Software-Engineering Requirements`_
45#. `Other RCU Flavors`_
46#. `Possible Future Changes`_
47
48This is followed by a summary_, however, the answers to
49each quick quiz immediately follows the quiz. Select the big white space
50with your mouse to see the answer.
51
52Fundamental Requirements
53------------------------
54
55RCU's fundamental requirements are the closest thing RCU has to hard
56mathematical requirements. These are:
57
58#. `Grace-Period Guarantee`_
59#. `Publish/Subscribe Guarantee`_
60#. `Memory-Barrier Guarantees`_
61#. `RCU Primitives Guaranteed to Execute Unconditionally`_
62#. `Guaranteed Read-to-Write Upgrade`_
63
64Grace-Period Guarantee
65~~~~~~~~~~~~~~~~~~~~~~
66
67RCU's grace-period guarantee is unusual in being premeditated: Jack
68Slingwine and I had this guarantee firmly in mind when we started work
69on RCU (then called “rclock”) in the early 1990s. That said, the past
70two decades of experience with RCU have produced a much more detailed
71understanding of this guarantee.
72
73RCU's grace-period guarantee allows updaters to wait for the completion
74of all pre-existing RCU read-side critical sections. An RCU read-side
75critical section begins with the marker rcu_read_lock() and ends
76with the marker rcu_read_unlock(). These markers may be nested, and
77RCU treats a nested set as one big RCU read-side critical section.
78Production-quality implementations of rcu_read_lock() and
79rcu_read_unlock() are extremely lightweight, and in fact have
80exactly zero overhead in Linux kernels built for production use with
81``CONFIG_PREEMPTION=n``.
82
83This guarantee allows ordering to be enforced with extremely low
84overhead to readers, for example:
85
86   ::
87
88       1 int x, y;
89       2
90       3 void thread0(void)
91       4 {
92       5   rcu_read_lock();
93       6   r1 = READ_ONCE(x);
94       7   r2 = READ_ONCE(y);
95       8   rcu_read_unlock();
96       9 }
97      10
98      11 void thread1(void)
99      12 {
100      13   WRITE_ONCE(x, 1);
101      14   synchronize_rcu();
102      15   WRITE_ONCE(y, 1);
103      16 }
104
105Because the synchronize_rcu() on line 14 waits for all pre-existing
106readers, any instance of thread0() that loads a value of zero from
107``x`` must complete before thread1() stores to ``y``, so that
108instance must also load a value of zero from ``y``. Similarly, any
109instance of thread0() that loads a value of one from ``y`` must have
110started after the synchronize_rcu() started, and must therefore also
111load a value of one from ``x``. Therefore, the outcome:
112
113   ::
114
115      (r1 == 0 && r2 == 1)
116
117cannot happen.
118
119+-----------------------------------------------------------------------+
120| **Quick Quiz**:                                                       |
121+-----------------------------------------------------------------------+
122| Wait a minute! You said that updaters can make useful forward         |
123| progress concurrently with readers, but pre-existing readers will     |
124| block synchronize_rcu()!!!                                            |
125| Just who are you trying to fool???                                    |
126+-----------------------------------------------------------------------+
127| **Answer**:                                                           |
128+-----------------------------------------------------------------------+
129| First, if updaters do not wish to be blocked by readers, they can use |
130| call_rcu() or kfree_rcu(), which will be discussed later.             |
131| Second, even when using synchronize_rcu(), the other update-side      |
132| code does run concurrently with readers, whether pre-existing or not. |
133+-----------------------------------------------------------------------+
134
135This scenario resembles one of the first uses of RCU in
136`DYNIX/ptx <https://en.wikipedia.org/wiki/DYNIX>`__, which managed a
137distributed lock manager's transition into a state suitable for handling
138recovery from node failure, more or less as follows:
139
140   ::
141
142       1 #define STATE_NORMAL        0
143       2 #define STATE_WANT_RECOVERY 1
144       3 #define STATE_RECOVERING    2
145       4 #define STATE_WANT_NORMAL   3
146       5
147       6 int state = STATE_NORMAL;
148       7
149       8 void do_something_dlm(void)
150       9 {
151      10   int state_snap;
152      11
153      12   rcu_read_lock();
154      13   state_snap = READ_ONCE(state);
155      14   if (state_snap == STATE_NORMAL)
156      15     do_something();
157      16   else
158      17     do_something_carefully();
159      18   rcu_read_unlock();
160      19 }
161      20
162      21 void start_recovery(void)
163      22 {
164      23   WRITE_ONCE(state, STATE_WANT_RECOVERY);
165      24   synchronize_rcu();
166      25   WRITE_ONCE(state, STATE_RECOVERING);
167      26   recovery();
168      27   WRITE_ONCE(state, STATE_WANT_NORMAL);
169      28   synchronize_rcu();
170      29   WRITE_ONCE(state, STATE_NORMAL);
171      30 }
172
173The RCU read-side critical section in do_something_dlm() works with
174the synchronize_rcu() in start_recovery() to guarantee that
175do_something() never runs concurrently with recovery(), but with
176little or no synchronization overhead in do_something_dlm().
177
178+-----------------------------------------------------------------------+
179| **Quick Quiz**:                                                       |
180+-----------------------------------------------------------------------+
181| Why is the synchronize_rcu() on line 28 needed?                       |
182+-----------------------------------------------------------------------+
183| **Answer**:                                                           |
184+-----------------------------------------------------------------------+
185| Without that extra grace period, memory reordering could result in    |
186| do_something_dlm() executing do_something() concurrently with         |
187| the last bits of recovery().                                          |
188+-----------------------------------------------------------------------+
189
190In order to avoid fatal problems such as deadlocks, an RCU read-side
191critical section must not contain calls to synchronize_rcu().
192Similarly, an RCU read-side critical section must not contain anything
193that waits, directly or indirectly, on completion of an invocation of
194synchronize_rcu().
195
196Although RCU's grace-period guarantee is useful in and of itself, with
197`quite a few use cases <https://lwn.net/Articles/573497/>`__, it would
198be good to be able to use RCU to coordinate read-side access to linked
199data structures. For this, the grace-period guarantee is not sufficient,
200as can be seen in function add_gp_buggy() below. We will look at the
201reader's code later, but in the meantime, just think of the reader as
202locklessly picking up the ``gp`` pointer, and, if the value loaded is
203non-\ ``NULL``, locklessly accessing the ``->a`` and ``->b`` fields.
204
205   ::
206
207       1 bool add_gp_buggy(int a, int b)
208       2 {
209       3   p = kmalloc(sizeof(*p), GFP_KERNEL);
210       4   if (!p)
211       5     return -ENOMEM;
212       6   spin_lock(&gp_lock);
213       7   if (rcu_access_pointer(gp)) {
214       8     spin_unlock(&gp_lock);
215       9     return false;
216      10   }
217      11   p->a = a;
218      12   p->b = a;
219      13   gp = p; /* ORDERING BUG */
220      14   spin_unlock(&gp_lock);
221      15   return true;
222      16 }
223
224The problem is that both the compiler and weakly ordered CPUs are within
225their rights to reorder this code as follows:
226
227   ::
228
229       1 bool add_gp_buggy_optimized(int a, int b)
230       2 {
231       3   p = kmalloc(sizeof(*p), GFP_KERNEL);
232       4   if (!p)
233       5     return -ENOMEM;
234       6   spin_lock(&gp_lock);
235       7   if (rcu_access_pointer(gp)) {
236       8     spin_unlock(&gp_lock);
237       9     return false;
238      10   }
239      11   gp = p; /* ORDERING BUG */
240      12   p->a = a;
241      13   p->b = a;
242      14   spin_unlock(&gp_lock);
243      15   return true;
244      16 }
245
246If an RCU reader fetches ``gp`` just after ``add_gp_buggy_optimized``
247executes line 11, it will see garbage in the ``->a`` and ``->b`` fields.
248And this is but one of many ways in which compiler and hardware
249optimizations could cause trouble. Therefore, we clearly need some way
250to prevent the compiler and the CPU from reordering in this manner,
251which brings us to the publish-subscribe guarantee discussed in the next
252section.
253
254Publish/Subscribe Guarantee
255~~~~~~~~~~~~~~~~~~~~~~~~~~~
256
257RCU's publish-subscribe guarantee allows data to be inserted into a
258linked data structure without disrupting RCU readers. The updater uses
259rcu_assign_pointer() to insert the new data, and readers use
260rcu_dereference() to access data, whether new or old. The following
261shows an example of insertion:
262
263   ::
264
265       1 bool add_gp(int a, int b)
266       2 {
267       3   p = kmalloc(sizeof(*p), GFP_KERNEL);
268       4   if (!p)
269       5     return -ENOMEM;
270       6   spin_lock(&gp_lock);
271       7   if (rcu_access_pointer(gp)) {
272       8     spin_unlock(&gp_lock);
273       9     return false;
274      10   }
275      11   p->a = a;
276      12   p->b = a;
277      13   rcu_assign_pointer(gp, p);
278      14   spin_unlock(&gp_lock);
279      15   return true;
280      16 }
281
282The rcu_assign_pointer() on line 13 is conceptually equivalent to a
283simple assignment statement, but also guarantees that its assignment
284will happen after the two assignments in lines 11 and 12, similar to the
285C11 ``memory_order_release`` store operation. It also prevents any
286number of “interesting” compiler optimizations, for example, the use of
287``gp`` as a scratch location immediately preceding the assignment.
288
289+-----------------------------------------------------------------------+
290| **Quick Quiz**:                                                       |
291+-----------------------------------------------------------------------+
292| But rcu_assign_pointer() does nothing to prevent the two              |
293| assignments to ``p->a`` and ``p->b`` from being reordered. Can't that |
294| also cause problems?                                                  |
295+-----------------------------------------------------------------------+
296| **Answer**:                                                           |
297+-----------------------------------------------------------------------+
298| No, it cannot. The readers cannot see either of these two fields      |
299| until the assignment to ``gp``, by which time both fields are fully   |
300| initialized. So reordering the assignments to ``p->a`` and ``p->b``   |
301| cannot possibly cause any problems.                                   |
302+-----------------------------------------------------------------------+
303
304It is tempting to assume that the reader need not do anything special to
305control its accesses to the RCU-protected data, as shown in
306do_something_gp_buggy() below:
307
308   ::
309
310       1 bool do_something_gp_buggy(void)
311       2 {
312       3   rcu_read_lock();
313       4   p = gp;  /* OPTIMIZATIONS GALORE!!! */
314       5   if (p) {
315       6     do_something(p->a, p->b);
316       7     rcu_read_unlock();
317       8     return true;
318       9   }
319      10   rcu_read_unlock();
320      11   return false;
321      12 }
322
323However, this temptation must be resisted because there are a
324surprisingly large number of ways that the compiler (or weak ordering
325CPUs like the DEC Alpha) can trip this code up. For but one example, if
326the compiler were short of registers, it might choose to refetch from
327``gp`` rather than keeping a separate copy in ``p`` as follows:
328
329   ::
330
331       1 bool do_something_gp_buggy_optimized(void)
332       2 {
333       3   rcu_read_lock();
334       4   if (gp) { /* OPTIMIZATIONS GALORE!!! */
335       5     do_something(gp->a, gp->b);
336       6     rcu_read_unlock();
337       7     return true;
338       8   }
339       9   rcu_read_unlock();
340      10   return false;
341      11 }
342
343If this function ran concurrently with a series of updates that replaced
344the current structure with a new one, the fetches of ``gp->a`` and
345``gp->b`` might well come from two different structures, which could
346cause serious confusion. To prevent this (and much else besides),
347do_something_gp() uses rcu_dereference() to fetch from ``gp``:
348
349   ::
350
351       1 bool do_something_gp(void)
352       2 {
353       3   rcu_read_lock();
354       4   p = rcu_dereference(gp);
355       5   if (p) {
356       6     do_something(p->a, p->b);
357       7     rcu_read_unlock();
358       8     return true;
359       9   }
360      10   rcu_read_unlock();
361      11   return false;
362      12 }
363
364The rcu_dereference() uses volatile casts and (for DEC Alpha) memory
365barriers in the Linux kernel. Should a |high-quality implementation of
366C11 memory_order_consume [PDF]|_
367ever appear, then rcu_dereference() could be implemented as a
368``memory_order_consume`` load. Regardless of the exact implementation, a
369pointer fetched by rcu_dereference() may not be used outside of the
370outermost RCU read-side critical section containing that
371rcu_dereference(), unless protection of the corresponding data
372element has been passed from RCU to some other synchronization
373mechanism, most commonly locking or `reference
374counting <https://www.kernel.org/doc/Documentation/RCU/rcuref.txt>`__.
375
376.. |high-quality implementation of C11 memory_order_consume [PDF]| replace:: high-quality implementation of C11 ``memory_order_consume`` [PDF]
377.. _high-quality implementation of C11 memory_order_consume [PDF]: http://www.rdrop.com/users/paulmck/RCU/consume.2015.07.13a.pdf
378
379In short, updaters use rcu_assign_pointer() and readers use
380rcu_dereference(), and these two RCU API elements work together to
381ensure that readers have a consistent view of newly added data elements.
382
383Of course, it is also necessary to remove elements from RCU-protected
384data structures, for example, using the following process:
385
386#. Remove the data element from the enclosing structure.
387#. Wait for all pre-existing RCU read-side critical sections to complete
388   (because only pre-existing readers can possibly have a reference to
389   the newly removed data element).
390#. At this point, only the updater has a reference to the newly removed
391   data element, so it can safely reclaim the data element, for example,
392   by passing it to kfree().
393
394This process is implemented by remove_gp_synchronous():
395
396   ::
397
398       1 bool remove_gp_synchronous(void)
399       2 {
400       3   struct foo *p;
401       4
402       5   spin_lock(&gp_lock);
403       6   p = rcu_access_pointer(gp);
404       7   if (!p) {
405       8     spin_unlock(&gp_lock);
406       9     return false;
407      10   }
408      11   rcu_assign_pointer(gp, NULL);
409      12   spin_unlock(&gp_lock);
410      13   synchronize_rcu();
411      14   kfree(p);
412      15   return true;
413      16 }
414
415This function is straightforward, with line 13 waiting for a grace
416period before line 14 frees the old data element. This waiting ensures
417that readers will reach line 7 of do_something_gp() before the data
418element referenced by ``p`` is freed. The rcu_access_pointer() on
419line 6 is similar to rcu_dereference(), except that:
420
421#. The value returned by rcu_access_pointer() cannot be
422   dereferenced. If you want to access the value pointed to as well as
423   the pointer itself, use rcu_dereference() instead of
424   rcu_access_pointer().
425#. The call to rcu_access_pointer() need not be protected. In
426   contrast, rcu_dereference() must either be within an RCU
427   read-side critical section or in a code segment where the pointer
428   cannot change, for example, in code protected by the corresponding
429   update-side lock.
430
431+-----------------------------------------------------------------------+
432| **Quick Quiz**:                                                       |
433+-----------------------------------------------------------------------+
434| Without the rcu_dereference() or the rcu_access_pointer(),            |
435| what destructive optimizations might the compiler make use of?        |
436+-----------------------------------------------------------------------+
437| **Answer**:                                                           |
438+-----------------------------------------------------------------------+
439| Let's start with what happens to do_something_gp() if it fails to     |
440| use rcu_dereference(). It could reuse a value formerly fetched        |
441| from this same pointer. It could also fetch the pointer from ``gp``   |
442| in a byte-at-a-time manner, resulting in *load tearing*, in turn      |
443| resulting a bytewise mash-up of two distinct pointer values. It might |
444| even use value-speculation optimizations, where it makes a wrong      |
445| guess, but by the time it gets around to checking the value, an       |
446| update has changed the pointer to match the wrong guess. Too bad      |
447| about any dereferences that returned pre-initialization garbage in    |
448| the meantime!                                                         |
449| For remove_gp_synchronous(), as long as all modifications to          |
450| ``gp`` are carried out while holding ``gp_lock``, the above           |
451| optimizations are harmless. However, ``sparse`` will complain if you  |
452| define ``gp`` with ``__rcu`` and then access it without using either  |
453| rcu_access_pointer() or rcu_dereference().                            |
454+-----------------------------------------------------------------------+
455
456In short, RCU's publish-subscribe guarantee is provided by the
457combination of rcu_assign_pointer() and rcu_dereference(). This
458guarantee allows data elements to be safely added to RCU-protected
459linked data structures without disrupting RCU readers. This guarantee
460can be used in combination with the grace-period guarantee to also allow
461data elements to be removed from RCU-protected linked data structures,
462again without disrupting RCU readers.
463
464This guarantee was only partially premeditated. DYNIX/ptx used an
465explicit memory barrier for publication, but had nothing resembling
466rcu_dereference() for subscription, nor did it have anything
467resembling the dependency-ordering barrier that was later subsumed
468into rcu_dereference() and later still into READ_ONCE(). The
469need for these operations made itself known quite suddenly at a
470late-1990s meeting with the DEC Alpha architects, back in the days when
471DEC was still a free-standing company. It took the Alpha architects a
472good hour to convince me that any sort of barrier would ever be needed,
473and it then took me a good *two* hours to convince them that their
474documentation did not make this point clear. More recent work with the C
475and C++ standards committees have provided much education on tricks and
476traps from the compiler. In short, compilers were much less tricky in
477the early 1990s, but in 2015, don't even think about omitting
478rcu_dereference()!
479
480Memory-Barrier Guarantees
481~~~~~~~~~~~~~~~~~~~~~~~~~
482
483The previous section's simple linked-data-structure scenario clearly
484demonstrates the need for RCU's stringent memory-ordering guarantees on
485systems with more than one CPU:
486
487#. Each CPU that has an RCU read-side critical section that begins
488   before synchronize_rcu() starts is guaranteed to execute a full
489   memory barrier between the time that the RCU read-side critical
490   section ends and the time that synchronize_rcu() returns. Without
491   this guarantee, a pre-existing RCU read-side critical section might
492   hold a reference to the newly removed ``struct foo`` after the
493   kfree() on line 14 of remove_gp_synchronous().
494#. Each CPU that has an RCU read-side critical section that ends after
495   synchronize_rcu() returns is guaranteed to execute a full memory
496   barrier between the time that synchronize_rcu() begins and the
497   time that the RCU read-side critical section begins. Without this
498   guarantee, a later RCU read-side critical section running after the
499   kfree() on line 14 of remove_gp_synchronous() might later run
500   do_something_gp() and find the newly deleted ``struct foo``.
501#. If the task invoking synchronize_rcu() remains on a given CPU,
502   then that CPU is guaranteed to execute a full memory barrier sometime
503   during the execution of synchronize_rcu(). This guarantee ensures
504   that the kfree() on line 14 of remove_gp_synchronous() really
505   does execute after the removal on line 11.
506#. If the task invoking synchronize_rcu() migrates among a group of
507   CPUs during that invocation, then each of the CPUs in that group is
508   guaranteed to execute a full memory barrier sometime during the
509   execution of synchronize_rcu(). This guarantee also ensures that
510   the kfree() on line 14 of remove_gp_synchronous() really does
511   execute after the removal on line 11, but also in the case where the
512   thread executing the synchronize_rcu() migrates in the meantime.
513
514+-----------------------------------------------------------------------+
515| **Quick Quiz**:                                                       |
516+-----------------------------------------------------------------------+
517| Given that multiple CPUs can start RCU read-side critical sections at |
518| any time without any ordering whatsoever, how can RCU possibly tell   |
519| whether or not a given RCU read-side critical section starts before a |
520| given instance of synchronize_rcu()?                                  |
521+-----------------------------------------------------------------------+
522| **Answer**:                                                           |
523+-----------------------------------------------------------------------+
524| If RCU cannot tell whether or not a given RCU read-side critical      |
525| section starts before a given instance of synchronize_rcu(), then     |
526| it must assume that the RCU read-side critical section started first. |
527| In other words, a given instance of synchronize_rcu() can avoid       |
528| waiting on a given RCU read-side critical section only if it can      |
529| prove that synchronize_rcu() started first.                           |
530| A related question is “When rcu_read_lock() doesn't generate any      |
531| code, why does it matter how it relates to a grace period?” The       |
532| answer is that it is not the relationship of rcu_read_lock()          |
533| itself that is important, but rather the relationship of the code     |
534| within the enclosed RCU read-side critical section to the code        |
535| preceding and following the grace period. If we take this viewpoint,  |
536| then a given RCU read-side critical section begins before a given     |
537| grace period when some access preceding the grace period observes the |
538| effect of some access within the critical section, in which case none |
539| of the accesses within the critical section may observe the effects   |
540| of any access following the grace period.                             |
541|                                                                       |
542| As of late 2016, mathematical models of RCU take this viewpoint, for  |
543| example, see slides 62 and 63 of the `2016 LinuxCon                   |
544| EU <http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.201 |
545| 6.10.04c.LCE.pdf>`__                                                  |
546| presentation.                                                         |
547+-----------------------------------------------------------------------+
548
549+-----------------------------------------------------------------------+
550| **Quick Quiz**:                                                       |
551+-----------------------------------------------------------------------+
552| The first and second guarantees require unbelievably strict ordering! |
553| Are all these memory barriers *really* required?                      |
554+-----------------------------------------------------------------------+
555| **Answer**:                                                           |
556+-----------------------------------------------------------------------+
557| Yes, they really are required. To see why the first guarantee is      |
558| required, consider the following sequence of events:                  |
559|                                                                       |
560| #. CPU 1: rcu_read_lock()                                             |
561| #. CPU 1: ``q = rcu_dereference(gp); /* Very likely to return p. */`` |
562| #. CPU 0: ``list_del_rcu(p);``                                        |
563| #. CPU 0: synchronize_rcu() starts.                                   |
564| #. CPU 1: ``do_something_with(q->a);``                                |
565|    ``/* No smp_mb(), so might happen after kfree(). */``              |
566| #. CPU 1: rcu_read_unlock()                                           |
567| #. CPU 0: synchronize_rcu() returns.                                  |
568| #. CPU 0: ``kfree(p);``                                               |
569|                                                                       |
570| Therefore, there absolutely must be a full memory barrier between the |
571| end of the RCU read-side critical section and the end of the grace    |
572| period.                                                               |
573|                                                                       |
574| The sequence of events demonstrating the necessity of the second rule |
575| is roughly similar:                                                   |
576|                                                                       |
577| #. CPU 0: ``list_del_rcu(p);``                                        |
578| #. CPU 0: synchronize_rcu() starts.                                   |
579| #. CPU 1: rcu_read_lock()                                             |
580| #. CPU 1: ``q = rcu_dereference(gp);``                                |
581|    ``/* Might return p if no memory barrier. */``                     |
582| #. CPU 0: synchronize_rcu() returns.                                  |
583| #. CPU 0: ``kfree(p);``                                               |
584| #. CPU 1: ``do_something_with(q->a); /* Boom!!! */``                  |
585| #. CPU 1: rcu_read_unlock()                                           |
586|                                                                       |
587| And similarly, without a memory barrier between the beginning of the  |
588| grace period and the beginning of the RCU read-side critical section, |
589| CPU 1 might end up accessing the freelist.                            |
590|                                                                       |
591| The “as if” rule of course applies, so that any implementation that   |
592| acts as if the appropriate memory barriers were in place is a correct |
593| implementation. That said, it is much easier to fool yourself into    |
594| believing that you have adhered to the as-if rule than it is to       |
595| actually adhere to it!                                                |
596+-----------------------------------------------------------------------+
597
598+-----------------------------------------------------------------------+
599| **Quick Quiz**:                                                       |
600+-----------------------------------------------------------------------+
601| You claim that rcu_read_lock() and rcu_read_unlock() generate         |
602| absolutely no code in some kernel builds. This means that the         |
603| compiler might arbitrarily rearrange consecutive RCU read-side        |
604| critical sections. Given such rearrangement, if a given RCU read-side |
605| critical section is done, how can you be sure that all prior RCU      |
606| read-side critical sections are done? Won't the compiler              |
607| rearrangements make that impossible to determine?                     |
608+-----------------------------------------------------------------------+
609| **Answer**:                                                           |
610+-----------------------------------------------------------------------+
611| In cases where rcu_read_lock() and rcu_read_unlock() generate         |
612| absolutely no code, RCU infers quiescent states only at special       |
613| locations, for example, within the scheduler. Because calls to        |
614| schedule() had better prevent calling-code accesses to shared         |
615| variables from being rearranged across the call to schedule(), if     |
616| RCU detects the end of a given RCU read-side critical section, it     |
617| will necessarily detect the end of all prior RCU read-side critical   |
618| sections, no matter how aggressively the compiler scrambles the code. |
619| Again, this all assumes that the compiler cannot scramble code across |
620| calls to the scheduler, out of interrupt handlers, into the idle      |
621| loop, into user-mode code, and so on. But if your kernel build allows |
622| that sort of scrambling, you have broken far more than just RCU!      |
623+-----------------------------------------------------------------------+
624
625Note that these memory-barrier requirements do not replace the
626fundamental RCU requirement that a grace period wait for all
627pre-existing readers. On the contrary, the memory barriers called out in
628this section must operate in such a way as to *enforce* this fundamental
629requirement. Of course, different implementations enforce this
630requirement in different ways, but enforce it they must.
631
632RCU Primitives Guaranteed to Execute Unconditionally
633~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
634
635The common-case RCU primitives are unconditional. They are invoked, they
636do their job, and they return, with no possibility of error, and no need
637to retry. This is a key RCU design philosophy.
638
639However, this philosophy is pragmatic rather than pigheaded. If someone
640comes up with a good justification for a particular conditional RCU
641primitive, it might well be implemented and added. After all, this
642guarantee was reverse-engineered, not premeditated. The unconditional
643nature of the RCU primitives was initially an accident of
644implementation, and later experience with synchronization primitives
645with conditional primitives caused me to elevate this accident to a
646guarantee. Therefore, the justification for adding a conditional
647primitive to RCU would need to be based on detailed and compelling use
648cases.
649
650Guaranteed Read-to-Write Upgrade
651~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
652
653As far as RCU is concerned, it is always possible to carry out an update
654within an RCU read-side critical section. For example, that RCU
655read-side critical section might search for a given data element, and
656then might acquire the update-side spinlock in order to update that
657element, all while remaining in that RCU read-side critical section. Of
658course, it is necessary to exit the RCU read-side critical section
659before invoking synchronize_rcu(), however, this inconvenience can
660be avoided through use of the call_rcu() and kfree_rcu() API
661members described later in this document.
662
663+-----------------------------------------------------------------------+
664| **Quick Quiz**:                                                       |
665+-----------------------------------------------------------------------+
666| But how does the upgrade-to-write operation exclude other readers?    |
667+-----------------------------------------------------------------------+
668| **Answer**:                                                           |
669+-----------------------------------------------------------------------+
670| It doesn't, just like normal RCU updates, which also do not exclude   |
671| RCU readers.                                                          |
672+-----------------------------------------------------------------------+
673
674This guarantee allows lookup code to be shared between read-side and
675update-side code, and was premeditated, appearing in the earliest
676DYNIX/ptx RCU documentation.
677
678Fundamental Non-Requirements
679----------------------------
680
681RCU provides extremely lightweight readers, and its read-side
682guarantees, though quite useful, are correspondingly lightweight. It is
683therefore all too easy to assume that RCU is guaranteeing more than it
684really is. Of course, the list of things that RCU does not guarantee is
685infinitely long, however, the following sections list a few
686non-guarantees that have caused confusion. Except where otherwise noted,
687these non-guarantees were premeditated.
688
689#. `Readers Impose Minimal Ordering`_
690#. `Readers Do Not Exclude Updaters`_
691#. `Updaters Only Wait For Old Readers`_
692#. `Grace Periods Don't Partition Read-Side Critical Sections`_
693#. `Read-Side Critical Sections Don't Partition Grace Periods`_
694
695Readers Impose Minimal Ordering
696~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
697
698Reader-side markers such as rcu_read_lock() and
699rcu_read_unlock() provide absolutely no ordering guarantees except
700through their interaction with the grace-period APIs such as
701synchronize_rcu(). To see this, consider the following pair of
702threads:
703
704   ::
705
706       1 void thread0(void)
707       2 {
708       3   rcu_read_lock();
709       4   WRITE_ONCE(x, 1);
710       5   rcu_read_unlock();
711       6   rcu_read_lock();
712       7   WRITE_ONCE(y, 1);
713       8   rcu_read_unlock();
714       9 }
715      10
716      11 void thread1(void)
717      12 {
718      13   rcu_read_lock();
719      14   r1 = READ_ONCE(y);
720      15   rcu_read_unlock();
721      16   rcu_read_lock();
722      17   r2 = READ_ONCE(x);
723      18   rcu_read_unlock();
724      19 }
725
726After thread0() and thread1() execute concurrently, it is quite
727possible to have
728
729   ::
730
731      (r1 == 1 && r2 == 0)
732
733(that is, ``y`` appears to have been assigned before ``x``), which would
734not be possible if rcu_read_lock() and rcu_read_unlock() had
735much in the way of ordering properties. But they do not, so the CPU is
736within its rights to do significant reordering. This is by design: Any
737significant ordering constraints would slow down these fast-path APIs.
738
739+-----------------------------------------------------------------------+
740| **Quick Quiz**:                                                       |
741+-----------------------------------------------------------------------+
742| Can't the compiler also reorder this code?                            |
743+-----------------------------------------------------------------------+
744| **Answer**:                                                           |
745+-----------------------------------------------------------------------+
746| No, the volatile casts in READ_ONCE() and WRITE_ONCE()                |
747| prevent the compiler from reordering in this particular case.         |
748+-----------------------------------------------------------------------+
749
750Readers Do Not Exclude Updaters
751~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
752
753Neither rcu_read_lock() nor rcu_read_unlock() exclude updates.
754All they do is to prevent grace periods from ending. The following
755example illustrates this:
756
757   ::
758
759       1 void thread0(void)
760       2 {
761       3   rcu_read_lock();
762       4   r1 = READ_ONCE(y);
763       5   if (r1) {
764       6     do_something_with_nonzero_x();
765       7     r2 = READ_ONCE(x);
766       8     WARN_ON(!r2); /* BUG!!! */
767       9   }
768      10   rcu_read_unlock();
769      11 }
770      12
771      13 void thread1(void)
772      14 {
773      15   spin_lock(&my_lock);
774      16   WRITE_ONCE(x, 1);
775      17   WRITE_ONCE(y, 1);
776      18   spin_unlock(&my_lock);
777      19 }
778
779If the thread0() function's rcu_read_lock() excluded the
780thread1() function's update, the WARN_ON() could never fire. But
781the fact is that rcu_read_lock() does not exclude much of anything
782aside from subsequent grace periods, of which thread1() has none, so
783the WARN_ON() can and does fire.
784
785Updaters Only Wait For Old Readers
786~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
787
788It might be tempting to assume that after synchronize_rcu()
789completes, there are no readers executing. This temptation must be
790avoided because new readers can start immediately after
791synchronize_rcu() starts, and synchronize_rcu() is under no
792obligation to wait for these new readers.
793
794+-----------------------------------------------------------------------+
795| **Quick Quiz**:                                                       |
796+-----------------------------------------------------------------------+
797| Suppose that synchronize_rcu() did wait until *all* readers had       |
798| completed instead of waiting only on pre-existing readers. For how    |
799| long would the updater be able to rely on there being no readers?     |
800+-----------------------------------------------------------------------+
801| **Answer**:                                                           |
802+-----------------------------------------------------------------------+
803| For no time at all. Even if synchronize_rcu() were to wait until      |
804| all readers had completed, a new reader might start immediately after |
805| synchronize_rcu() completed. Therefore, the code following            |
806| synchronize_rcu() can *never* rely on there being no readers.         |
807+-----------------------------------------------------------------------+
808
809Grace Periods Don't Partition Read-Side Critical Sections
810~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
811
812It is tempting to assume that if any part of one RCU read-side critical
813section precedes a given grace period, and if any part of another RCU
814read-side critical section follows that same grace period, then all of
815the first RCU read-side critical section must precede all of the second.
816However, this just isn't the case: A single grace period does not
817partition the set of RCU read-side critical sections. An example of this
818situation can be illustrated as follows, where ``x``, ``y``, and ``z``
819are initially all zero:
820
821   ::
822
823       1 void thread0(void)
824       2 {
825       3   rcu_read_lock();
826       4   WRITE_ONCE(a, 1);
827       5   WRITE_ONCE(b, 1);
828       6   rcu_read_unlock();
829       7 }
830       8
831       9 void thread1(void)
832      10 {
833      11   r1 = READ_ONCE(a);
834      12   synchronize_rcu();
835      13   WRITE_ONCE(c, 1);
836      14 }
837      15
838      16 void thread2(void)
839      17 {
840      18   rcu_read_lock();
841      19   r2 = READ_ONCE(b);
842      20   r3 = READ_ONCE(c);
843      21   rcu_read_unlock();
844      22 }
845
846It turns out that the outcome:
847
848   ::
849
850      (r1 == 1 && r2 == 0 && r3 == 1)
851
852is entirely possible. The following figure show how this can happen,
853with each circled ``QS`` indicating the point at which RCU recorded a
854*quiescent state* for each thread, that is, a state in which RCU knows
855that the thread cannot be in the midst of an RCU read-side critical
856section that started before the current grace period:
857
858.. kernel-figure:: GPpartitionReaders1.svg
859
860If it is necessary to partition RCU read-side critical sections in this
861manner, it is necessary to use two grace periods, where the first grace
862period is known to end before the second grace period starts:
863
864   ::
865
866       1 void thread0(void)
867       2 {
868       3   rcu_read_lock();
869       4   WRITE_ONCE(a, 1);
870       5   WRITE_ONCE(b, 1);
871       6   rcu_read_unlock();
872       7 }
873       8
874       9 void thread1(void)
875      10 {
876      11   r1 = READ_ONCE(a);
877      12   synchronize_rcu();
878      13   WRITE_ONCE(c, 1);
879      14 }
880      15
881      16 void thread2(void)
882      17 {
883      18   r2 = READ_ONCE(c);
884      19   synchronize_rcu();
885      20   WRITE_ONCE(d, 1);
886      21 }
887      22
888      23 void thread3(void)
889      24 {
890      25   rcu_read_lock();
891      26   r3 = READ_ONCE(b);
892      27   r4 = READ_ONCE(d);
893      28   rcu_read_unlock();
894      29 }
895
896Here, if ``(r1 == 1)``, then thread0()'s write to ``b`` must happen
897before the end of thread1()'s grace period. If in addition
898``(r4 == 1)``, then thread3()'s read from ``b`` must happen after
899the beginning of thread2()'s grace period. If it is also the case
900that ``(r2 == 1)``, then the end of thread1()'s grace period must
901precede the beginning of thread2()'s grace period. This mean that
902the two RCU read-side critical sections cannot overlap, guaranteeing
903that ``(r3 == 1)``. As a result, the outcome:
904
905   ::
906
907      (r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1)
908
909cannot happen.
910
911This non-requirement was also non-premeditated, but became apparent when
912studying RCU's interaction with memory ordering.
913
914Read-Side Critical Sections Don't Partition Grace Periods
915~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
916
917It is also tempting to assume that if an RCU read-side critical section
918happens between a pair of grace periods, then those grace periods cannot
919overlap. However, this temptation leads nowhere good, as can be
920illustrated by the following, with all variables initially zero:
921
922   ::
923
924       1 void thread0(void)
925       2 {
926       3   rcu_read_lock();
927       4   WRITE_ONCE(a, 1);
928       5   WRITE_ONCE(b, 1);
929       6   rcu_read_unlock();
930       7 }
931       8
932       9 void thread1(void)
933      10 {
934      11   r1 = READ_ONCE(a);
935      12   synchronize_rcu();
936      13   WRITE_ONCE(c, 1);
937      14 }
938      15
939      16 void thread2(void)
940      17 {
941      18   rcu_read_lock();
942      19   WRITE_ONCE(d, 1);
943      20   r2 = READ_ONCE(c);
944      21   rcu_read_unlock();
945      22 }
946      23
947      24 void thread3(void)
948      25 {
949      26   r3 = READ_ONCE(d);
950      27   synchronize_rcu();
951      28   WRITE_ONCE(e, 1);
952      29 }
953      30
954      31 void thread4(void)
955      32 {
956      33   rcu_read_lock();
957      34   r4 = READ_ONCE(b);
958      35   r5 = READ_ONCE(e);
959      36   rcu_read_unlock();
960      37 }
961
962In this case, the outcome:
963
964   ::
965
966      (r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1)
967
968is entirely possible, as illustrated below:
969
970.. kernel-figure:: ReadersPartitionGP1.svg
971
972Again, an RCU read-side critical section can overlap almost all of a
973given grace period, just so long as it does not overlap the entire grace
974period. As a result, an RCU read-side critical section cannot partition
975a pair of RCU grace periods.
976
977+-----------------------------------------------------------------------+
978| **Quick Quiz**:                                                       |
979+-----------------------------------------------------------------------+
980| How long a sequence of grace periods, each separated by an RCU        |
981| read-side critical section, would be required to partition the RCU    |
982| read-side critical sections at the beginning and end of the chain?    |
983+-----------------------------------------------------------------------+
984| **Answer**:                                                           |
985+-----------------------------------------------------------------------+
986| In theory, an infinite number. In practice, an unknown number that is |
987| sensitive to both implementation details and timing considerations.   |
988| Therefore, even in practice, RCU users must abide by the theoretical  |
989| rather than the practical answer.                                     |
990+-----------------------------------------------------------------------+
991
992Parallelism Facts of Life
993-------------------------
994
995These parallelism facts of life are by no means specific to RCU, but the
996RCU implementation must abide by them. They therefore bear repeating:
997
998#. Any CPU or task may be delayed at any time, and any attempts to avoid
999   these delays by disabling preemption, interrupts, or whatever are
1000   completely futile. This is most obvious in preemptible user-level
1001   environments and in virtualized environments (where a given guest
1002   OS's VCPUs can be preempted at any time by the underlying
1003   hypervisor), but can also happen in bare-metal environments due to
1004   ECC errors, NMIs, and other hardware events. Although a delay of more
1005   than about 20 seconds can result in splats, the RCU implementation is
1006   obligated to use algorithms that can tolerate extremely long delays,
1007   but where “extremely long” is not long enough to allow wrap-around
1008   when incrementing a 64-bit counter.
1009#. Both the compiler and the CPU can reorder memory accesses. Where it
1010   matters, RCU must use compiler directives and memory-barrier
1011   instructions to preserve ordering.
1012#. Conflicting writes to memory locations in any given cache line will
1013   result in expensive cache misses. Greater numbers of concurrent
1014   writes and more-frequent concurrent writes will result in more
1015   dramatic slowdowns. RCU is therefore obligated to use algorithms that
1016   have sufficient locality to avoid significant performance and
1017   scalability problems.
1018#. As a rough rule of thumb, only one CPU's worth of processing may be
1019   carried out under the protection of any given exclusive lock. RCU
1020   must therefore use scalable locking designs.
1021#. Counters are finite, especially on 32-bit systems. RCU's use of
1022   counters must therefore tolerate counter wrap, or be designed such
1023   that counter wrap would take way more time than a single system is
1024   likely to run. An uptime of ten years is quite possible, a runtime of
1025   a century much less so. As an example of the latter, RCU's
1026   dyntick-idle nesting counter allows 54 bits for interrupt nesting
1027   level (this counter is 64 bits even on a 32-bit system). Overflowing
1028   this counter requires 2\ :sup:`54` half-interrupts on a given CPU
1029   without that CPU ever going idle. If a half-interrupt happened every
1030   microsecond, it would take 570 years of runtime to overflow this
1031   counter, which is currently believed to be an acceptably long time.
1032#. Linux systems can have thousands of CPUs running a single Linux
1033   kernel in a single shared-memory environment. RCU must therefore pay
1034   close attention to high-end scalability.
1035
1036This last parallelism fact of life means that RCU must pay special
1037attention to the preceding facts of life. The idea that Linux might
1038scale to systems with thousands of CPUs would have been met with some
1039skepticism in the 1990s, but these requirements would have otherwise
1040have been unsurprising, even in the early 1990s.
1041
1042Quality-of-Implementation Requirements
1043--------------------------------------
1044
1045These sections list quality-of-implementation requirements. Although an
1046RCU implementation that ignores these requirements could still be used,
1047it would likely be subject to limitations that would make it
1048inappropriate for industrial-strength production use. Classes of
1049quality-of-implementation requirements are as follows:
1050
1051#. `Specialization`_
1052#. `Performance and Scalability`_
1053#. `Forward Progress`_
1054#. `Composability`_
1055#. `Corner Cases`_
1056
1057These classes is covered in the following sections.
1058
1059Specialization
1060~~~~~~~~~~~~~~
1061
1062RCU is and always has been intended primarily for read-mostly
1063situations, which means that RCU's read-side primitives are optimized,
1064often at the expense of its update-side primitives. Experience thus far
1065is captured by the following list of situations:
1066
1067#. Read-mostly data, where stale and inconsistent data is not a problem:
1068   RCU works great!
1069#. Read-mostly data, where data must be consistent: RCU works well.
1070#. Read-write data, where data must be consistent: RCU *might* work OK.
1071   Or not.
1072#. Write-mostly data, where data must be consistent: RCU is very
1073   unlikely to be the right tool for the job, with the following
1074   exceptions, where RCU can provide:
1075
1076   a. Existence guarantees for update-friendly mechanisms.
1077   b. Wait-free read-side primitives for real-time use.
1078
1079This focus on read-mostly situations means that RCU must interoperate
1080with other synchronization primitives. For example, the add_gp() and
1081remove_gp_synchronous() examples discussed earlier use RCU to
1082protect readers and locking to coordinate updaters. However, the need
1083extends much farther, requiring that a variety of synchronization
1084primitives be legal within RCU read-side critical sections, including
1085spinlocks, sequence locks, atomic operations, reference counters, and
1086memory barriers.
1087
1088+-----------------------------------------------------------------------+
1089| **Quick Quiz**:                                                       |
1090+-----------------------------------------------------------------------+
1091| What about sleeping locks?                                            |
1092+-----------------------------------------------------------------------+
1093| **Answer**:                                                           |
1094+-----------------------------------------------------------------------+
1095| These are forbidden within Linux-kernel RCU read-side critical        |
1096| sections because it is not legal to place a quiescent state (in this  |
1097| case, voluntary context switch) within an RCU read-side critical      |
1098| section. However, sleeping locks may be used within userspace RCU     |
1099| read-side critical sections, and also within Linux-kernel sleepable   |
1100| RCU `(SRCU) <Sleepable RCU_>`__ read-side critical sections. In       |
1101| addition, the -rt patchset turns spinlocks into a sleeping locks so   |
1102| that the corresponding critical sections can be preempted, which also |
1103| means that these sleeplockified spinlocks (but not other sleeping     |
1104| locks!) may be acquire within -rt-Linux-kernel RCU read-side critical |
1105| sections.                                                             |
1106| Note that it *is* legal for a normal RCU read-side critical section   |
1107| to conditionally acquire a sleeping locks (as in                      |
1108| mutex_trylock()), but only as long as it does not loop                |
1109| indefinitely attempting to conditionally acquire that sleeping locks. |
1110| The key point is that things like mutex_trylock() either return       |
1111| with the mutex held, or return an error indication if the mutex was   |
1112| not immediately available. Either way, mutex_trylock() returns        |
1113| immediately without sleeping.                                         |
1114+-----------------------------------------------------------------------+
1115
1116It often comes as a surprise that many algorithms do not require a
1117consistent view of data, but many can function in that mode, with
1118network routing being the poster child. Internet routing algorithms take
1119significant time to propagate updates, so that by the time an update
1120arrives at a given system, that system has been sending network traffic
1121the wrong way for a considerable length of time. Having a few threads
1122continue to send traffic the wrong way for a few more milliseconds is
1123clearly not a problem: In the worst case, TCP retransmissions will
1124eventually get the data where it needs to go. In general, when tracking
1125the state of the universe outside of the computer, some level of
1126inconsistency must be tolerated due to speed-of-light delays if nothing
1127else.
1128
1129Furthermore, uncertainty about external state is inherent in many cases.
1130For example, a pair of veterinarians might use heartbeat to determine
1131whether or not a given cat was alive. But how long should they wait
1132after the last heartbeat to decide that the cat is in fact dead? Waiting
1133less than 400 milliseconds makes no sense because this would mean that a
1134relaxed cat would be considered to cycle between death and life more
1135than 100 times per minute. Moreover, just as with human beings, a cat's
1136heart might stop for some period of time, so the exact wait period is a
1137judgment call. One of our pair of veterinarians might wait 30 seconds
1138before pronouncing the cat dead, while the other might insist on waiting
1139a full minute. The two veterinarians would then disagree on the state of
1140the cat during the final 30 seconds of the minute following the last
1141heartbeat.
1142
1143Interestingly enough, this same situation applies to hardware. When push
1144comes to shove, how do we tell whether or not some external server has
1145failed? We send messages to it periodically, and declare it failed if we
1146don't receive a response within a given period of time. Policy decisions
1147can usually tolerate short periods of inconsistency. The policy was
1148decided some time ago, and is only now being put into effect, so a few
1149milliseconds of delay is normally inconsequential.
1150
1151However, there are algorithms that absolutely must see consistent data.
1152For example, the translation between a user-level SystemV semaphore ID
1153to the corresponding in-kernel data structure is protected by RCU, but
1154it is absolutely forbidden to update a semaphore that has just been
1155removed. In the Linux kernel, this need for consistency is accommodated
1156by acquiring spinlocks located in the in-kernel data structure from
1157within the RCU read-side critical section, and this is indicated by the
1158green box in the figure above. Many other techniques may be used, and
1159are in fact used within the Linux kernel.
1160
1161In short, RCU is not required to maintain consistency, and other
1162mechanisms may be used in concert with RCU when consistency is required.
1163RCU's specialization allows it to do its job extremely well, and its
1164ability to interoperate with other synchronization mechanisms allows the
1165right mix of synchronization tools to be used for a given job.
1166
1167Performance and Scalability
1168~~~~~~~~~~~~~~~~~~~~~~~~~~~
1169
1170Energy efficiency is a critical component of performance today, and
1171Linux-kernel RCU implementations must therefore avoid unnecessarily
1172awakening idle CPUs. I cannot claim that this requirement was
1173premeditated. In fact, I learned of it during a telephone conversation
1174in which I was given “frank and open” feedback on the importance of
1175energy efficiency in battery-powered systems and on specific
1176energy-efficiency shortcomings of the Linux-kernel RCU implementation.
1177In my experience, the battery-powered embedded community will consider
1178any unnecessary wakeups to be extremely unfriendly acts. So much so that
1179mere Linux-kernel-mailing-list posts are insufficient to vent their ire.
1180
1181Memory consumption is not particularly important for in most situations,
1182and has become decreasingly so as memory sizes have expanded and memory
1183costs have plummeted. However, as I learned from Matt Mackall's
1184`bloatwatch <http://elinux.org/Linux_Tiny-FAQ>`__ efforts, memory
1185footprint is critically important on single-CPU systems with
1186non-preemptible (``CONFIG_PREEMPTION=n``) kernels, and thus `tiny
1187RCU <https://lore.kernel.org/r/20090113221724.GA15307@linux.vnet.ibm.com>`__
1188was born. Josh Triplett has since taken over the small-memory banner
1189with his `Linux kernel tinification <https://tiny.wiki.kernel.org/>`__
1190project, which resulted in `SRCU <Sleepable RCU_>`__ becoming optional
1191for those kernels not needing it.
1192
1193The remaining performance requirements are, for the most part,
1194unsurprising. For example, in keeping with RCU's read-side
1195specialization, rcu_dereference() should have negligible overhead
1196(for example, suppression of a few minor compiler optimizations).
1197Similarly, in non-preemptible environments, rcu_read_lock() and
1198rcu_read_unlock() should have exactly zero overhead.
1199
1200In preemptible environments, in the case where the RCU read-side
1201critical section was not preempted (as will be the case for the
1202highest-priority real-time process), rcu_read_lock() and
1203rcu_read_unlock() should have minimal overhead. In particular, they
1204should not contain atomic read-modify-write operations, memory-barrier
1205instructions, preemption disabling, interrupt disabling, or backwards
1206branches. However, in the case where the RCU read-side critical section
1207was preempted, rcu_read_unlock() may acquire spinlocks and disable
1208interrupts. This is why it is better to nest an RCU read-side critical
1209section within a preempt-disable region than vice versa, at least in
1210cases where that critical section is short enough to avoid unduly
1211degrading real-time latencies.
1212
1213The synchronize_rcu() grace-period-wait primitive is optimized for
1214throughput. It may therefore incur several milliseconds of latency in
1215addition to the duration of the longest RCU read-side critical section.
1216On the other hand, multiple concurrent invocations of
1217synchronize_rcu() are required to use batching optimizations so that
1218they can be satisfied by a single underlying grace-period-wait
1219operation. For example, in the Linux kernel, it is not unusual for a
1220single grace-period-wait operation to serve more than `1,000 separate
1221invocations <https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response>`__
1222of synchronize_rcu(), thus amortizing the per-invocation overhead
1223down to nearly zero. However, the grace-period optimization is also
1224required to avoid measurable degradation of real-time scheduling and
1225interrupt latencies.
1226
1227In some cases, the multi-millisecond synchronize_rcu() latencies are
1228unacceptable. In these cases, synchronize_rcu_expedited() may be
1229used instead, reducing the grace-period latency down to a few tens of
1230microseconds on small systems, at least in cases where the RCU read-side
1231critical sections are short. There are currently no special latency
1232requirements for synchronize_rcu_expedited() on large systems, but,
1233consistent with the empirical nature of the RCU specification, that is
1234subject to change. However, there most definitely are scalability
1235requirements: A storm of synchronize_rcu_expedited() invocations on
12364096 CPUs should at least make reasonable forward progress. In return
1237for its shorter latencies, synchronize_rcu_expedited() is permitted
1238to impose modest degradation of real-time latency on non-idle online
1239CPUs. Here, “modest” means roughly the same latency degradation as a
1240scheduling-clock interrupt.
1241
1242There are a number of situations where even
1243synchronize_rcu_expedited()'s reduced grace-period latency is
1244unacceptable. In these situations, the asynchronous call_rcu() can
1245be used in place of synchronize_rcu() as follows:
1246
1247   ::
1248
1249       1 struct foo {
1250       2   int a;
1251       3   int b;
1252       4   struct rcu_head rh;
1253       5 };
1254       6
1255       7 static void remove_gp_cb(struct rcu_head *rhp)
1256       8 {
1257       9   struct foo *p = container_of(rhp, struct foo, rh);
1258      10
1259      11   kfree(p);
1260      12 }
1261      13
1262      14 bool remove_gp_asynchronous(void)
1263      15 {
1264      16   struct foo *p;
1265      17
1266      18   spin_lock(&gp_lock);
1267      19   p = rcu_access_pointer(gp);
1268      20   if (!p) {
1269      21     spin_unlock(&gp_lock);
1270      22     return false;
1271      23   }
1272      24   rcu_assign_pointer(gp, NULL);
1273      25   call_rcu(&p->rh, remove_gp_cb);
1274      26   spin_unlock(&gp_lock);
1275      27   return true;
1276      28 }
1277
1278A definition of ``struct foo`` is finally needed, and appears on
1279lines 1-5. The function remove_gp_cb() is passed to call_rcu()
1280on line 25, and will be invoked after the end of a subsequent grace
1281period. This gets the same effect as remove_gp_synchronous(), but
1282without forcing the updater to wait for a grace period to elapse. The
1283call_rcu() function may be used in a number of situations where
1284neither synchronize_rcu() nor synchronize_rcu_expedited() would
1285be legal, including within preempt-disable code, local_bh_disable()
1286code, interrupt-disable code, and interrupt handlers. However, even
1287call_rcu() is illegal within NMI handlers and from idle and offline
1288CPUs. The callback function (remove_gp_cb() in this case) will be
1289executed within softirq (software interrupt) environment within the
1290Linux kernel, either within a real softirq handler or under the
1291protection of local_bh_disable(). In both the Linux kernel and in
1292userspace, it is bad practice to write an RCU callback function that
1293takes too long. Long-running operations should be relegated to separate
1294threads or (in the Linux kernel) workqueues.
1295
1296+-----------------------------------------------------------------------+
1297| **Quick Quiz**:                                                       |
1298+-----------------------------------------------------------------------+
1299| Why does line 19 use rcu_access_pointer()? After all,                 |
1300| call_rcu() on line 25 stores into the structure, which would          |
1301| interact badly with concurrent insertions. Doesn't this mean that     |
1302| rcu_dereference() is required?                                        |
1303+-----------------------------------------------------------------------+
1304| **Answer**:                                                           |
1305+-----------------------------------------------------------------------+
1306| Presumably the ``->gp_lock`` acquired on line 18 excludes any         |
1307| changes, including any insertions that rcu_dereference() would        |
1308| protect against. Therefore, any insertions will be delayed until      |
1309| after ``->gp_lock`` is released on line 25, which in turn means that  |
1310| rcu_access_pointer() suffices.                                        |
1311+-----------------------------------------------------------------------+
1312
1313However, all that remove_gp_cb() is doing is invoking kfree() on
1314the data element. This is a common idiom, and is supported by
1315kfree_rcu(), which allows “fire and forget” operation as shown
1316below:
1317
1318   ::
1319
1320       1 struct foo {
1321       2   int a;
1322       3   int b;
1323       4   struct rcu_head rh;
1324       5 };
1325       6
1326       7 bool remove_gp_faf(void)
1327       8 {
1328       9   struct foo *p;
1329      10
1330      11   spin_lock(&gp_lock);
1331      12   p = rcu_dereference(gp);
1332      13   if (!p) {
1333      14     spin_unlock(&gp_lock);
1334      15     return false;
1335      16   }
1336      17   rcu_assign_pointer(gp, NULL);
1337      18   kfree_rcu(p, rh);
1338      19   spin_unlock(&gp_lock);
1339      20   return true;
1340      21 }
1341
1342Note that remove_gp_faf() simply invokes kfree_rcu() and
1343proceeds, without any need to pay any further attention to the
1344subsequent grace period and kfree(). It is permissible to invoke
1345kfree_rcu() from the same environments as for call_rcu().
1346Interestingly enough, DYNIX/ptx had the equivalents of call_rcu()
1347and kfree_rcu(), but not synchronize_rcu(). This was due to the
1348fact that RCU was not heavily used within DYNIX/ptx, so the very few
1349places that needed something like synchronize_rcu() simply
1350open-coded it.
1351
1352+-----------------------------------------------------------------------+
1353| **Quick Quiz**:                                                       |
1354+-----------------------------------------------------------------------+
1355| Earlier it was claimed that call_rcu() and kfree_rcu()                |
1356| allowed updaters to avoid being blocked by readers. But how can that  |
1357| be correct, given that the invocation of the callback and the freeing |
1358| of the memory (respectively) must still wait for a grace period to    |
1359| elapse?                                                               |
1360+-----------------------------------------------------------------------+
1361| **Answer**:                                                           |
1362+-----------------------------------------------------------------------+
1363| We could define things this way, but keep in mind that this sort of   |
1364| definition would say that updates in garbage-collected languages      |
1365| cannot complete until the next time the garbage collector runs, which |
1366| does not seem at all reasonable. The key point is that in most cases, |
1367| an updater using either call_rcu() or kfree_rcu() can proceed         |
1368| to the next update as soon as it has invoked call_rcu() or            |
1369| kfree_rcu(), without having to wait for a subsequent grace            |
1370| period.                                                               |
1371+-----------------------------------------------------------------------+
1372
1373But what if the updater must wait for the completion of code to be
1374executed after the end of the grace period, but has other tasks that can
1375be carried out in the meantime? The polling-style
1376get_state_synchronize_rcu() and cond_synchronize_rcu() functions
1377may be used for this purpose, as shown below:
1378
1379   ::
1380
1381       1 bool remove_gp_poll(void)
1382       2 {
1383       3   struct foo *p;
1384       4   unsigned long s;
1385       5
1386       6   spin_lock(&gp_lock);
1387       7   p = rcu_access_pointer(gp);
1388       8   if (!p) {
1389       9     spin_unlock(&gp_lock);
1390      10     return false;
1391      11   }
1392      12   rcu_assign_pointer(gp, NULL);
1393      13   spin_unlock(&gp_lock);
1394      14   s = get_state_synchronize_rcu();
1395      15   do_something_while_waiting();
1396      16   cond_synchronize_rcu(s);
1397      17   kfree(p);
1398      18   return true;
1399      19 }
1400
1401On line 14, get_state_synchronize_rcu() obtains a “cookie” from RCU,
1402then line 15 carries out other tasks, and finally, line 16 returns
1403immediately if a grace period has elapsed in the meantime, but otherwise
1404waits as required. The need for ``get_state_synchronize_rcu`` and
1405cond_synchronize_rcu() has appeared quite recently, so it is too
1406early to tell whether they will stand the test of time.
1407
1408RCU thus provides a range of tools to allow updaters to strike the
1409required tradeoff between latency, flexibility and CPU overhead.
1410
1411Forward Progress
1412~~~~~~~~~~~~~~~~
1413
1414In theory, delaying grace-period completion and callback invocation is
1415harmless. In practice, not only are memory sizes finite but also
1416callbacks sometimes do wakeups, and sufficiently deferred wakeups can be
1417difficult to distinguish from system hangs. Therefore, RCU must provide
1418a number of mechanisms to promote forward progress.
1419
1420These mechanisms are not foolproof, nor can they be. For one simple
1421example, an infinite loop in an RCU read-side critical section must by
1422definition prevent later grace periods from ever completing. For a more
1423involved example, consider a 64-CPU system built with
1424``CONFIG_RCU_NOCB_CPU=y`` and booted with ``rcu_nocbs=1-63``, where
1425CPUs 1 through 63 spin in tight loops that invoke call_rcu(). Even
1426if these tight loops also contain calls to cond_resched() (thus
1427allowing grace periods to complete), CPU 0 simply will not be able to
1428invoke callbacks as fast as the other 63 CPUs can register them, at
1429least not until the system runs out of memory. In both of these
1430examples, the Spiderman principle applies: With great power comes great
1431responsibility. However, short of this level of abuse, RCU is required
1432to ensure timely completion of grace periods and timely invocation of
1433callbacks.
1434
1435RCU takes the following steps to encourage timely completion of grace
1436periods:
1437
1438#. If a grace period fails to complete within 100 milliseconds, RCU
1439   causes future invocations of cond_resched() on the holdout CPUs
1440   to provide an RCU quiescent state. RCU also causes those CPUs'
1441   need_resched() invocations to return ``true``, but only after the
1442   corresponding CPU's next scheduling-clock.
1443#. CPUs mentioned in the ``nohz_full`` kernel boot parameter can run
1444   indefinitely in the kernel without scheduling-clock interrupts, which
1445   defeats the above need_resched() strategem. RCU will therefore
1446   invoke resched_cpu() on any ``nohz_full`` CPUs still holding out
1447   after 109 milliseconds.
1448#. In kernels built with ``CONFIG_RCU_BOOST=y``, if a given task that
1449   has been preempted within an RCU read-side critical section is
1450   holding out for more than 500 milliseconds, RCU will resort to
1451   priority boosting.
1452#. If a CPU is still holding out 10 seconds into the grace period, RCU
1453   will invoke resched_cpu() on it regardless of its ``nohz_full``
1454   state.
1455
1456The above values are defaults for systems running with ``HZ=1000``. They
1457will vary as the value of ``HZ`` varies, and can also be changed using
1458the relevant Kconfig options and kernel boot parameters. RCU currently
1459does not do much sanity checking of these parameters, so please use
1460caution when changing them. Note that these forward-progress measures
1461are provided only for RCU, not for `SRCU <Sleepable RCU_>`__ or `Tasks
1462RCU`_.
1463
1464RCU takes the following steps in call_rcu() to encourage timely
1465invocation of callbacks when any given non-\ ``rcu_nocbs`` CPU has
146610,000 callbacks, or has 10,000 more callbacks than it had the last time
1467encouragement was provided:
1468
1469#. Starts a grace period, if one is not already in progress.
1470#. Forces immediate checking for quiescent states, rather than waiting
1471   for three milliseconds to have elapsed since the beginning of the
1472   grace period.
1473#. Immediately tags the CPU's callbacks with their grace period
1474   completion numbers, rather than waiting for the ``RCU_SOFTIRQ``
1475   handler to get around to it.
1476#. Lifts callback-execution batch limits, which speeds up callback
1477   invocation at the expense of degrading realtime response.
1478
1479Again, these are default values when running at ``HZ=1000``, and can be
1480overridden. Again, these forward-progress measures are provided only for
1481RCU, not for `SRCU <Sleepable RCU_>`__ or `Tasks
1482RCU`_. Even for RCU, callback-invocation forward
1483progress for ``rcu_nocbs`` CPUs is much less well-developed, in part
1484because workloads benefiting from ``rcu_nocbs`` CPUs tend to invoke
1485call_rcu() relatively infrequently. If workloads emerge that need
1486both ``rcu_nocbs`` CPUs and high call_rcu() invocation rates, then
1487additional forward-progress work will be required.
1488
1489Composability
1490~~~~~~~~~~~~~
1491
1492Composability has received much attention in recent years, perhaps in
1493part due to the collision of multicore hardware with object-oriented
1494techniques designed in single-threaded environments for single-threaded
1495use. And in theory, RCU read-side critical sections may be composed, and
1496in fact may be nested arbitrarily deeply. In practice, as with all
1497real-world implementations of composable constructs, there are
1498limitations.
1499
1500Implementations of RCU for which rcu_read_lock() and
1501rcu_read_unlock() generate no code, such as Linux-kernel RCU when
1502``CONFIG_PREEMPTION=n``, can be nested arbitrarily deeply. After all, there
1503is no overhead. Except that if all these instances of
1504rcu_read_lock() and rcu_read_unlock() are visible to the
1505compiler, compilation will eventually fail due to exhausting memory,
1506mass storage, or user patience, whichever comes first. If the nesting is
1507not visible to the compiler, as is the case with mutually recursive
1508functions each in its own translation unit, stack overflow will result.
1509If the nesting takes the form of loops, perhaps in the guise of tail
1510recursion, either the control variable will overflow or (in the Linux
1511kernel) you will get an RCU CPU stall warning. Nevertheless, this class
1512of RCU implementations is one of the most composable constructs in
1513existence.
1514
1515RCU implementations that explicitly track nesting depth are limited by
1516the nesting-depth counter. For example, the Linux kernel's preemptible
1517RCU limits nesting to ``INT_MAX``. This should suffice for almost all
1518practical purposes. That said, a consecutive pair of RCU read-side
1519critical sections between which there is an operation that waits for a
1520grace period cannot be enclosed in another RCU read-side critical
1521section. This is because it is not legal to wait for a grace period
1522within an RCU read-side critical section: To do so would result either
1523in deadlock or in RCU implicitly splitting the enclosing RCU read-side
1524critical section, neither of which is conducive to a long-lived and
1525prosperous kernel.
1526
1527It is worth noting that RCU is not alone in limiting composability. For
1528example, many transactional-memory implementations prohibit composing a
1529pair of transactions separated by an irrevocable operation (for example,
1530a network receive operation). For another example, lock-based critical
1531sections can be composed surprisingly freely, but only if deadlock is
1532avoided.
1533
1534In short, although RCU read-side critical sections are highly
1535composable, care is required in some situations, just as is the case for
1536any other composable synchronization mechanism.
1537
1538Corner Cases
1539~~~~~~~~~~~~
1540
1541A given RCU workload might have an endless and intense stream of RCU
1542read-side critical sections, perhaps even so intense that there was
1543never a point in time during which there was not at least one RCU
1544read-side critical section in flight. RCU cannot allow this situation to
1545block grace periods: As long as all the RCU read-side critical sections
1546are finite, grace periods must also be finite.
1547
1548That said, preemptible RCU implementations could potentially result in
1549RCU read-side critical sections being preempted for long durations,
1550which has the effect of creating a long-duration RCU read-side critical
1551section. This situation can arise only in heavily loaded systems, but
1552systems using real-time priorities are of course more vulnerable.
1553Therefore, RCU priority boosting is provided to help deal with this
1554case. That said, the exact requirements on RCU priority boosting will
1555likely evolve as more experience accumulates.
1556
1557Other workloads might have very high update rates. Although one can
1558argue that such workloads should instead use something other than RCU,
1559the fact remains that RCU must handle such workloads gracefully. This
1560requirement is another factor driving batching of grace periods, but it
1561is also the driving force behind the checks for large numbers of queued
1562RCU callbacks in the call_rcu() code path. Finally, high update
1563rates should not delay RCU read-side critical sections, although some
1564small read-side delays can occur when using
1565synchronize_rcu_expedited(), courtesy of this function's use of
1566smp_call_function_single().
1567
1568Although all three of these corner cases were understood in the early
15691990s, a simple user-level test consisting of ``close(open(path))`` in a
1570tight loop in the early 2000s suddenly provided a much deeper
1571appreciation of the high-update-rate corner case. This test also
1572motivated addition of some RCU code to react to high update rates, for
1573example, if a given CPU finds itself with more than 10,000 RCU callbacks
1574queued, it will cause RCU to take evasive action by more aggressively
1575starting grace periods and more aggressively forcing completion of
1576grace-period processing. This evasive action causes the grace period to
1577complete more quickly, but at the cost of restricting RCU's batching
1578optimizations, thus increasing the CPU overhead incurred by that grace
1579period.
1580
1581Software-Engineering Requirements
1582---------------------------------
1583
1584Between Murphy's Law and “To err is human”, it is necessary to guard
1585against mishaps and misuse:
1586
1587#. It is all too easy to forget to use rcu_read_lock() everywhere
1588   that it is needed, so kernels built with ``CONFIG_PROVE_RCU=y`` will
1589   splat if rcu_dereference() is used outside of an RCU read-side
1590   critical section. Update-side code can use
1591   rcu_dereference_protected(), which takes a `lockdep
1592   expression <https://lwn.net/Articles/371986/>`__ to indicate what is
1593   providing the protection. If the indicated protection is not
1594   provided, a lockdep splat is emitted.
1595   Code shared between readers and updaters can use
1596   rcu_dereference_check(), which also takes a lockdep expression,
1597   and emits a lockdep splat if neither rcu_read_lock() nor the
1598   indicated protection is in place. In addition,
1599   rcu_dereference_raw() is used in those (hopefully rare) cases
1600   where the required protection cannot be easily described. Finally,
1601   rcu_read_lock_held() is provided to allow a function to verify
1602   that it has been invoked within an RCU read-side critical section. I
1603   was made aware of this set of requirements shortly after Thomas
1604   Gleixner audited a number of RCU uses.
1605#. A given function might wish to check for RCU-related preconditions
1606   upon entry, before using any other RCU API. The
1607   rcu_lockdep_assert() does this job, asserting the expression in
1608   kernels having lockdep enabled and doing nothing otherwise.
1609#. It is also easy to forget to use rcu_assign_pointer() and
1610   rcu_dereference(), perhaps (incorrectly) substituting a simple
1611   assignment. To catch this sort of error, a given RCU-protected
1612   pointer may be tagged with ``__rcu``, after which sparse will
1613   complain about simple-assignment accesses to that pointer. Arnd
1614   Bergmann made me aware of this requirement, and also supplied the
1615   needed `patch series <https://lwn.net/Articles/376011/>`__.
1616#. Kernels built with ``CONFIG_DEBUG_OBJECTS_RCU_HEAD=y`` will splat if
1617   a data element is passed to call_rcu() twice in a row, without a
1618   grace period in between. (This error is similar to a double free.)
1619   The corresponding ``rcu_head`` structures that are dynamically
1620   allocated are automatically tracked, but ``rcu_head`` structures
1621   allocated on the stack must be initialized with
1622   init_rcu_head_on_stack() and cleaned up with
1623   destroy_rcu_head_on_stack(). Similarly, statically allocated
1624   non-stack ``rcu_head`` structures must be initialized with
1625   init_rcu_head() and cleaned up with destroy_rcu_head().
1626   Mathieu Desnoyers made me aware of this requirement, and also
1627   supplied the needed
1628   `patch <https://lore.kernel.org/r/20100319013024.GA28456@Krystal>`__.
1629#. An infinite loop in an RCU read-side critical section will eventually
1630   trigger an RCU CPU stall warning splat, with the duration of
1631   “eventually” being controlled by the ``RCU_CPU_STALL_TIMEOUT``
1632   ``Kconfig`` option, or, alternatively, by the
1633   ``rcupdate.rcu_cpu_stall_timeout`` boot/sysfs parameter. However, RCU
1634   is not obligated to produce this splat unless there is a grace period
1635   waiting on that particular RCU read-side critical section.
1636
1637   Some extreme workloads might intentionally delay RCU grace periods,
1638   and systems running those workloads can be booted with
1639   ``rcupdate.rcu_cpu_stall_suppress`` to suppress the splats. This
1640   kernel parameter may also be set via ``sysfs``. Furthermore, RCU CPU
1641   stall warnings are counter-productive during sysrq dumps and during
1642   panics. RCU therefore supplies the rcu_sysrq_start() and
1643   rcu_sysrq_end() API members to be called before and after long
1644   sysrq dumps. RCU also supplies the rcu_panic() notifier that is
1645   automatically invoked at the beginning of a panic to suppress further
1646   RCU CPU stall warnings.
1647
1648   This requirement made itself known in the early 1990s, pretty much
1649   the first time that it was necessary to debug a CPU stall. That said,
1650   the initial implementation in DYNIX/ptx was quite generic in
1651   comparison with that of Linux.
1652
1653#. Although it would be very good to detect pointers leaking out of RCU
1654   read-side critical sections, there is currently no good way of doing
1655   this. One complication is the need to distinguish between pointers
1656   leaking and pointers that have been handed off from RCU to some other
1657   synchronization mechanism, for example, reference counting.
1658#. In kernels built with ``CONFIG_RCU_TRACE=y``, RCU-related information
1659   is provided via event tracing.
1660#. Open-coded use of rcu_assign_pointer() and rcu_dereference()
1661   to create typical linked data structures can be surprisingly
1662   error-prone. Therefore, RCU-protected `linked
1663   lists <https://lwn.net/Articles/609973/#RCU%20List%20APIs>`__ and,
1664   more recently, RCU-protected `hash
1665   tables <https://lwn.net/Articles/612100/>`__ are available. Many
1666   other special-purpose RCU-protected data structures are available in
1667   the Linux kernel and the userspace RCU library.
1668#. Some linked structures are created at compile time, but still require
1669   ``__rcu`` checking. The RCU_POINTER_INITIALIZER() macro serves
1670   this purpose.
1671#. It is not necessary to use rcu_assign_pointer() when creating
1672   linked structures that are to be published via a single external
1673   pointer. The RCU_INIT_POINTER() macro is provided for this task.
1674
1675This not a hard-and-fast list: RCU's diagnostic capabilities will
1676continue to be guided by the number and type of usage bugs found in
1677real-world RCU usage.
1678
1679Linux Kernel Complications
1680--------------------------
1681
1682The Linux kernel provides an interesting environment for all kinds of
1683software, including RCU. Some of the relevant points of interest are as
1684follows:
1685
1686#. `Configuration`_
1687#. `Firmware Interface`_
1688#. `Early Boot`_
1689#. `Interrupts and NMIs`_
1690#. `Loadable Modules`_
1691#. `Hotplug CPU`_
1692#. `Scheduler and RCU`_
1693#. `Tracing and RCU`_
1694#. `Accesses to User Memory and RCU`_
1695#. `Energy Efficiency`_
1696#. `Scheduling-Clock Interrupts and RCU`_
1697#. `Memory Efficiency`_
1698#. `Performance, Scalability, Response Time, and Reliability`_
1699
1700This list is probably incomplete, but it does give a feel for the most
1701notable Linux-kernel complications. Each of the following sections
1702covers one of the above topics.
1703
1704Configuration
1705~~~~~~~~~~~~~
1706
1707RCU's goal is automatic configuration, so that almost nobody needs to
1708worry about RCU's ``Kconfig`` options. And for almost all users, RCU
1709does in fact work well “out of the box.”
1710
1711However, there are specialized use cases that are handled by kernel boot
1712parameters and ``Kconfig`` options. Unfortunately, the ``Kconfig``
1713system will explicitly ask users about new ``Kconfig`` options, which
1714requires almost all of them be hidden behind a ``CONFIG_RCU_EXPERT``
1715``Kconfig`` option.
1716
1717This all should be quite obvious, but the fact remains that Linus
1718Torvalds recently had to
1719`remind <https://lore.kernel.org/r/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com>`__
1720me of this requirement.
1721
1722Firmware Interface
1723~~~~~~~~~~~~~~~~~~
1724
1725In many cases, kernel obtains information about the system from the
1726firmware, and sometimes things are lost in translation. Or the
1727translation is accurate, but the original message is bogus.
1728
1729For example, some systems' firmware overreports the number of CPUs,
1730sometimes by a large factor. If RCU naively believed the firmware, as it
1731used to do, it would create too many per-CPU kthreads. Although the
1732resulting system will still run correctly, the extra kthreads needlessly
1733consume memory and can cause confusion when they show up in ``ps``
1734listings.
1735
1736RCU must therefore wait for a given CPU to actually come online before
1737it can allow itself to believe that the CPU actually exists. The
1738resulting “ghost CPUs” (which are never going to come online) cause a
1739number of `interesting
1740complications <https://paulmck.livejournal.com/37494.html>`__.
1741
1742Early Boot
1743~~~~~~~~~~
1744
1745The Linux kernel's boot sequence is an interesting process, and RCU is
1746used early, even before rcu_init() is invoked. In fact, a number of
1747RCU's primitives can be used as soon as the initial task's
1748``task_struct`` is available and the boot CPU's per-CPU variables are
1749set up. The read-side primitives (rcu_read_lock(),
1750rcu_read_unlock(), rcu_dereference(), and
1751rcu_access_pointer()) will operate normally very early on, as will
1752rcu_assign_pointer().
1753
1754Although call_rcu() may be invoked at any time during boot,
1755callbacks are not guaranteed to be invoked until after all of RCU's
1756kthreads have been spawned, which occurs at early_initcall() time.
1757This delay in callback invocation is due to the fact that RCU does not
1758invoke callbacks until it is fully initialized, and this full
1759initialization cannot occur until after the scheduler has initialized
1760itself to the point where RCU can spawn and run its kthreads. In theory,
1761it would be possible to invoke callbacks earlier, however, this is not a
1762panacea because there would be severe restrictions on what operations
1763those callbacks could invoke.
1764
1765Perhaps surprisingly, synchronize_rcu() and
1766synchronize_rcu_expedited(), will operate normally during very early
1767boot, the reason being that there is only one CPU and preemption is
1768disabled. This means that the call synchronize_rcu() (or friends)
1769itself is a quiescent state and thus a grace period, so the early-boot
1770implementation can be a no-op.
1771
1772However, once the scheduler has spawned its first kthread, this early
1773boot trick fails for synchronize_rcu() (as well as for
1774synchronize_rcu_expedited()) in ``CONFIG_PREEMPTION=y`` kernels. The
1775reason is that an RCU read-side critical section might be preempted,
1776which means that a subsequent synchronize_rcu() really does have to
1777wait for something, as opposed to simply returning immediately.
1778Unfortunately, synchronize_rcu() can't do this until all of its
1779kthreads are spawned, which doesn't happen until some time during
1780early_initcalls() time. But this is no excuse: RCU is nevertheless
1781required to correctly handle synchronous grace periods during this time
1782period. Once all of its kthreads are up and running, RCU starts running
1783normally.
1784
1785+-----------------------------------------------------------------------+
1786| **Quick Quiz**:                                                       |
1787+-----------------------------------------------------------------------+
1788| How can RCU possibly handle grace periods before all of its kthreads  |
1789| have been spawned???                                                  |
1790+-----------------------------------------------------------------------+
1791| **Answer**:                                                           |
1792+-----------------------------------------------------------------------+
1793| Very carefully!                                                       |
1794| During the “dead zone” between the time that the scheduler spawns the |
1795| first task and the time that all of RCU's kthreads have been spawned, |
1796| all synchronous grace periods are handled by the expedited            |
1797| grace-period mechanism. At runtime, this expedited mechanism relies   |
1798| on workqueues, but during the dead zone the requesting task itself    |
1799| drives the desired expedited grace period. Because dead-zone          |
1800| execution takes place within task context, everything works. Once the |
1801| dead zone ends, expedited grace periods go back to using workqueues,  |
1802| as is required to avoid problems that would otherwise occur when a    |
1803| user task received a POSIX signal while driving an expedited grace    |
1804| period.                                                               |
1805|                                                                       |
1806| And yes, this does mean that it is unhelpful to send POSIX signals to |
1807| random tasks between the time that the scheduler spawns its first     |
1808| kthread and the time that RCU's kthreads have all been spawned. If    |
1809| there ever turns out to be a good reason for sending POSIX signals    |
1810| during that time, appropriate adjustments will be made. (If it turns  |
1811| out that POSIX signals are sent during this time for no good reason,  |
1812| other adjustments will be made, appropriate or otherwise.)            |
1813+-----------------------------------------------------------------------+
1814
1815I learned of these boot-time requirements as a result of a series of
1816system hangs.
1817
1818Interrupts and NMIs
1819~~~~~~~~~~~~~~~~~~~
1820
1821The Linux kernel has interrupts, and RCU read-side critical sections are
1822legal within interrupt handlers and within interrupt-disabled regions of
1823code, as are invocations of call_rcu().
1824
1825Some Linux-kernel architectures can enter an interrupt handler from
1826non-idle process context, and then just never leave it, instead
1827stealthily transitioning back to process context. This trick is
1828sometimes used to invoke system calls from inside the kernel. These
1829“half-interrupts” mean that RCU has to be very careful about how it
1830counts interrupt nesting levels. I learned of this requirement the hard
1831way during a rewrite of RCU's dyntick-idle code.
1832
1833The Linux kernel has non-maskable interrupts (NMIs), and RCU read-side
1834critical sections are legal within NMI handlers. Thankfully, RCU
1835update-side primitives, including call_rcu(), are prohibited within
1836NMI handlers.
1837
1838The name notwithstanding, some Linux-kernel architectures can have
1839nested NMIs, which RCU must handle correctly. Andy Lutomirski `surprised
1840me <https://lore.kernel.org/r/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com>`__
1841with this requirement; he also kindly surprised me with `an
1842algorithm <https://lore.kernel.org/r/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com>`__
1843that meets this requirement.
1844
1845Furthermore, NMI handlers can be interrupted by what appear to RCU to be
1846normal interrupts. One way that this can happen is for code that
1847directly invokes rcu_irq_enter() and rcu_irq_exit() to be called
1848from an NMI handler. This astonishing fact of life prompted the current
1849code structure, which has rcu_irq_enter() invoking
1850rcu_nmi_enter() and rcu_irq_exit() invoking rcu_nmi_exit().
1851And yes, I also learned of this requirement the hard way.
1852
1853Loadable Modules
1854~~~~~~~~~~~~~~~~
1855
1856The Linux kernel has loadable modules, and these modules can also be
1857unloaded. After a given module has been unloaded, any attempt to call
1858one of its functions results in a segmentation fault. The module-unload
1859functions must therefore cancel any delayed calls to loadable-module
1860functions, for example, any outstanding mod_timer() must be dealt
1861with via del_timer_sync() or similar.
1862
1863Unfortunately, there is no way to cancel an RCU callback; once you
1864invoke call_rcu(), the callback function is eventually going to be
1865invoked, unless the system goes down first. Because it is normally
1866considered socially irresponsible to crash the system in response to a
1867module unload request, we need some other way to deal with in-flight RCU
1868callbacks.
1869
1870RCU therefore provides rcu_barrier(), which waits until all
1871in-flight RCU callbacks have been invoked. If a module uses
1872call_rcu(), its exit function should therefore prevent any future
1873invocation of call_rcu(), then invoke rcu_barrier(). In theory,
1874the underlying module-unload code could invoke rcu_barrier()
1875unconditionally, but in practice this would incur unacceptable
1876latencies.
1877
1878Nikita Danilov noted this requirement for an analogous
1879filesystem-unmount situation, and Dipankar Sarma incorporated
1880rcu_barrier() into RCU. The need for rcu_barrier() for module
1881unloading became apparent later.
1882
1883.. important::
1884
1885   The rcu_barrier() function is not, repeat,
1886   *not*, obligated to wait for a grace period. It is instead only required
1887   to wait for RCU callbacks that have already been posted. Therefore, if
1888   there are no RCU callbacks posted anywhere in the system,
1889   rcu_barrier() is within its rights to return immediately. Even if
1890   there are callbacks posted, rcu_barrier() does not necessarily need
1891   to wait for a grace period.
1892
1893+-----------------------------------------------------------------------+
1894| **Quick Quiz**:                                                       |
1895+-----------------------------------------------------------------------+
1896| Wait a minute! Each RCU callbacks must wait for a grace period to     |
1897| complete, and rcu_barrier() must wait for each pre-existing           |
1898| callback to be invoked. Doesn't rcu_barrier() therefore need to       |
1899| wait for a full grace period if there is even one callback posted     |
1900| anywhere in the system?                                               |
1901+-----------------------------------------------------------------------+
1902| **Answer**:                                                           |
1903+-----------------------------------------------------------------------+
1904| Absolutely not!!!                                                     |
1905| Yes, each RCU callbacks must wait for a grace period to complete, but |
1906| it might well be partly (or even completely) finished waiting by the  |
1907| time rcu_barrier() is invoked. In that case, rcu_barrier()            |
1908| need only wait for the remaining portion of the grace period to       |
1909| elapse. So even if there are quite a few callbacks posted,            |
1910| rcu_barrier() might well return quite quickly.                        |
1911|                                                                       |
1912| So if you need to wait for a grace period as well as for all          |
1913| pre-existing callbacks, you will need to invoke both                  |
1914| synchronize_rcu() and rcu_barrier(). If latency is a concern,         |
1915| you can always use workqueues to invoke them concurrently.            |
1916+-----------------------------------------------------------------------+
1917
1918Hotplug CPU
1919~~~~~~~~~~~
1920
1921The Linux kernel supports CPU hotplug, which means that CPUs can come
1922and go. It is of course illegal to use any RCU API member from an
1923offline CPU, with the exception of `SRCU <Sleepable RCU_>`__ read-side
1924critical sections. This requirement was present from day one in
1925DYNIX/ptx, but on the other hand, the Linux kernel's CPU-hotplug
1926implementation is “interesting.”
1927
1928The Linux-kernel CPU-hotplug implementation has notifiers that are used
1929to allow the various kernel subsystems (including RCU) to respond
1930appropriately to a given CPU-hotplug operation. Most RCU operations may
1931be invoked from CPU-hotplug notifiers, including even synchronous
1932grace-period operations such as (synchronize_rcu() and
1933synchronize_rcu_expedited()).  However, these synchronous operations
1934do block and therefore cannot be invoked from notifiers that execute via
1935stop_machine(), specifically those between the ``CPUHP_AP_OFFLINE``
1936and ``CPUHP_AP_ONLINE`` states.
1937
1938In addition, all-callback-wait operations such as rcu_barrier() may
1939not be invoked from any CPU-hotplug notifier.  This restriction is due
1940to the fact that there are phases of CPU-hotplug operations where the
1941outgoing CPU's callbacks will not be invoked until after the CPU-hotplug
1942operation ends, which could also result in deadlock. Furthermore,
1943rcu_barrier() blocks CPU-hotplug operations during its execution,
1944which results in another type of deadlock when invoked from a CPU-hotplug
1945notifier.
1946
1947Finally, RCU must avoid deadlocks due to interaction between hotplug,
1948timers and grace period processing. It does so by maintaining its own set
1949of books that duplicate the centrally maintained ``cpu_online_mask``,
1950and also by reporting quiescent states explicitly when a CPU goes
1951offline.  This explicit reporting of quiescent states avoids any need
1952for the force-quiescent-state loop (FQS) to report quiescent states for
1953offline CPUs.  However, as a debugging measure, the FQS loop does splat
1954if offline CPUs block an RCU grace period for too long.
1955
1956An offline CPU's quiescent state will be reported either:
1957
19581.  As the CPU goes offline using RCU's hotplug notifier (rcu_report_dead()).
19592.  When grace period initialization (rcu_gp_init()) detects a
1960    race either with CPU offlining or with a task unblocking on a leaf
1961    ``rcu_node`` structure whose CPUs are all offline.
1962
1963The CPU-online path (rcu_cpu_starting()) should never need to report
1964a quiescent state for an offline CPU.  However, as a debugging measure,
1965it does emit a warning if a quiescent state was not already reported
1966for that CPU.
1967
1968During the checking/modification of RCU's hotplug bookkeeping, the
1969corresponding CPU's leaf node lock is held. This avoids race conditions
1970between RCU's hotplug notifier hooks, the grace period initialization
1971code, and the FQS loop, all of which refer to or modify this bookkeeping.
1972
1973Scheduler and RCU
1974~~~~~~~~~~~~~~~~~
1975
1976RCU makes use of kthreads, and it is necessary to avoid excessive CPU-time
1977accumulation by these kthreads. This requirement was no surprise, but
1978RCU's violation of it when running context-switch-heavy workloads when
1979built with ``CONFIG_NO_HZ_FULL=y`` `did come as a surprise
1980[PDF] <http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf>`__.
1981RCU has made good progress towards meeting this requirement, even for
1982context-switch-heavy ``CONFIG_NO_HZ_FULL=y`` workloads, but there is
1983room for further improvement.
1984
1985There is no longer any prohibition against holding any of
1986scheduler's runqueue or priority-inheritance spinlocks across an
1987rcu_read_unlock(), even if interrupts and preemption were enabled
1988somewhere within the corresponding RCU read-side critical section.
1989Therefore, it is now perfectly legal to execute rcu_read_lock()
1990with preemption enabled, acquire one of the scheduler locks, and hold
1991that lock across the matching rcu_read_unlock().
1992
1993Similarly, the RCU flavor consolidation has removed the need for negative
1994nesting.  The fact that interrupt-disabled regions of code act as RCU
1995read-side critical sections implicitly avoids earlier issues that used
1996to result in destructive recursion via interrupt handler's use of RCU.
1997
1998Tracing and RCU
1999~~~~~~~~~~~~~~~
2000
2001It is possible to use tracing on RCU code, but tracing itself uses RCU.
2002For this reason, rcu_dereference_raw_check() is provided for use
2003by tracing, which avoids the destructive recursion that could otherwise
2004ensue. This API is also used by virtualization in some architectures,
2005where RCU readers execute in environments in which tracing cannot be
2006used. The tracing folks both located the requirement and provided the
2007needed fix, so this surprise requirement was relatively painless.
2008
2009Accesses to User Memory and RCU
2010~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2011
2012The kernel needs to access user-space memory, for example, to access data
2013referenced by system-call parameters.  The get_user() macro does this job.
2014
2015However, user-space memory might well be paged out, which means that
2016get_user() might well page-fault and thus block while waiting for the
2017resulting I/O to complete.  It would be a very bad thing for the compiler to
2018reorder a get_user() invocation into an RCU read-side critical section.
2019
2020For example, suppose that the source code looked like this:
2021
2022  ::
2023
2024       1 rcu_read_lock();
2025       2 p = rcu_dereference(gp);
2026       3 v = p->value;
2027       4 rcu_read_unlock();
2028       5 get_user(user_v, user_p);
2029       6 do_something_with(v, user_v);
2030
2031The compiler must not be permitted to transform this source code into
2032the following:
2033
2034  ::
2035
2036       1 rcu_read_lock();
2037       2 p = rcu_dereference(gp);
2038       3 get_user(user_v, user_p); // BUG: POSSIBLE PAGE FAULT!!!
2039       4 v = p->value;
2040       5 rcu_read_unlock();
2041       6 do_something_with(v, user_v);
2042
2043If the compiler did make this transformation in a ``CONFIG_PREEMPTION=n`` kernel
2044build, and if get_user() did page fault, the result would be a quiescent
2045state in the middle of an RCU read-side critical section.  This misplaced
2046quiescent state could result in line 4 being a use-after-free access,
2047which could be bad for your kernel's actuarial statistics.  Similar examples
2048can be constructed with the call to get_user() preceding the
2049rcu_read_lock().
2050
2051Unfortunately, get_user() doesn't have any particular ordering properties,
2052and in some architectures the underlying ``asm`` isn't even marked
2053``volatile``.  And even if it was marked ``volatile``, the above access to
2054``p->value`` is not volatile, so the compiler would not have any reason to keep
2055those two accesses in order.
2056
2057Therefore, the Linux-kernel definitions of rcu_read_lock() and
2058rcu_read_unlock() must act as compiler barriers, at least for outermost
2059instances of rcu_read_lock() and rcu_read_unlock() within a nested set
2060of RCU read-side critical sections.
2061
2062Energy Efficiency
2063~~~~~~~~~~~~~~~~~
2064
2065Interrupting idle CPUs is considered socially unacceptable, especially
2066by people with battery-powered embedded systems. RCU therefore conserves
2067energy by detecting which CPUs are idle, including tracking CPUs that
2068have been interrupted from idle. This is a large part of the
2069energy-efficiency requirement, so I learned of this via an irate phone
2070call.
2071
2072Because RCU avoids interrupting idle CPUs, it is illegal to execute an
2073RCU read-side critical section on an idle CPU. (Kernels built with
2074``CONFIG_PROVE_RCU=y`` will splat if you try it.) The RCU_NONIDLE()
2075macro and ``_rcuidle`` event tracing is provided to work around this
2076restriction. In addition, rcu_is_watching() may be used to test
2077whether or not it is currently legal to run RCU read-side critical
2078sections on this CPU. I learned of the need for diagnostics on the one
2079hand and RCU_NONIDLE() on the other while inspecting idle-loop code.
2080Steven Rostedt supplied ``_rcuidle`` event tracing, which is used quite
2081heavily in the idle loop. However, there are some restrictions on the
2082code placed within RCU_NONIDLE():
2083
2084#. Blocking is prohibited. In practice, this is not a serious
2085   restriction given that idle tasks are prohibited from blocking to
2086   begin with.
2087#. Although nesting RCU_NONIDLE() is permitted, they cannot nest
2088   indefinitely deeply. However, given that they can be nested on the
2089   order of a million deep, even on 32-bit systems, this should not be a
2090   serious restriction. This nesting limit would probably be reached
2091   long after the compiler OOMed or the stack overflowed.
2092#. Any code path that enters RCU_NONIDLE() must sequence out of that
2093   same RCU_NONIDLE(). For example, the following is grossly
2094   illegal:
2095
2096      ::
2097
2098	  1     RCU_NONIDLE({
2099	  2       do_something();
2100	  3       goto bad_idea;  /* BUG!!! */
2101	  4       do_something_else();});
2102	  5   bad_idea:
2103
2104
2105   It is just as illegal to transfer control into the middle of
2106   RCU_NONIDLE()'s argument. Yes, in theory, you could transfer in
2107   as long as you also transferred out, but in practice you could also
2108   expect to get sharply worded review comments.
2109
2110It is similarly socially unacceptable to interrupt an ``nohz_full`` CPU
2111running in userspace. RCU must therefore track ``nohz_full`` userspace
2112execution. RCU must therefore be able to sample state at two points in
2113time, and be able to determine whether or not some other CPU spent any
2114time idle and/or executing in userspace.
2115
2116These energy-efficiency requirements have proven quite difficult to
2117understand and to meet, for example, there have been more than five
2118clean-sheet rewrites of RCU's energy-efficiency code, the last of which
2119was finally able to demonstrate `real energy savings running on real
2120hardware
2121[PDF] <http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf>`__.
2122As noted earlier, I learned of many of these requirements via angry
2123phone calls: Flaming me on the Linux-kernel mailing list was apparently
2124not sufficient to fully vent their ire at RCU's energy-efficiency bugs!
2125
2126Scheduling-Clock Interrupts and RCU
2127~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2128
2129The kernel transitions between in-kernel non-idle execution, userspace
2130execution, and the idle loop. Depending on kernel configuration, RCU
2131handles these states differently:
2132
2133+-----------------+------------------+------------------+-----------------+
2134| ``HZ`` Kconfig  | In-Kernel        | Usermode         | Idle            |
2135+=================+==================+==================+=================+
2136| ``HZ_PERIODIC`` | Can rely on      | Can rely on      | Can rely on     |
2137|                 | scheduling-clock | scheduling-clock | RCU's           |
2138|                 | interrupt.       | interrupt and    | dyntick-idle    |
2139|                 |                  | its detection    | detection.      |
2140|                 |                  | of interrupt     |                 |
2141|                 |                  | from usermode.   |                 |
2142+-----------------+------------------+------------------+-----------------+
2143| ``NO_HZ_IDLE``  | Can rely on      | Can rely on      | Can rely on     |
2144|                 | scheduling-clock | scheduling-clock | RCU's           |
2145|                 | interrupt.       | interrupt and    | dyntick-idle    |
2146|                 |                  | its detection    | detection.      |
2147|                 |                  | of interrupt     |                 |
2148|                 |                  | from usermode.   |                 |
2149+-----------------+------------------+------------------+-----------------+
2150| ``NO_HZ_FULL``  | Can only         | Can rely on      | Can rely on     |
2151|                 | sometimes rely   | RCU's            | RCU's           |
2152|                 | on               | dyntick-idle     | dyntick-idle    |
2153|                 | scheduling-clock | detection.       | detection.      |
2154|                 | interrupt. In    |                  |                 |
2155|                 | other cases, it  |                  |                 |
2156|                 | is necessary to  |                  |                 |
2157|                 | bound kernel     |                  |                 |
2158|                 | execution times  |                  |                 |
2159|                 | and/or use       |                  |                 |
2160|                 | IPIs.            |                  |                 |
2161+-----------------+------------------+------------------+-----------------+
2162
2163+-----------------------------------------------------------------------+
2164| **Quick Quiz**:                                                       |
2165+-----------------------------------------------------------------------+
2166| Why can't ``NO_HZ_FULL`` in-kernel execution rely on the              |
2167| scheduling-clock interrupt, just like ``HZ_PERIODIC`` and             |
2168| ``NO_HZ_IDLE`` do?                                                    |
2169+-----------------------------------------------------------------------+
2170| **Answer**:                                                           |
2171+-----------------------------------------------------------------------+
2172| Because, as a performance optimization, ``NO_HZ_FULL`` does not       |
2173| necessarily re-enable the scheduling-clock interrupt on entry to each |
2174| and every system call.                                                |
2175+-----------------------------------------------------------------------+
2176
2177However, RCU must be reliably informed as to whether any given CPU is
2178currently in the idle loop, and, for ``NO_HZ_FULL``, also whether that
2179CPU is executing in usermode, as discussed
2180`earlier <Energy Efficiency_>`__. It also requires that the
2181scheduling-clock interrupt be enabled when RCU needs it to be:
2182
2183#. If a CPU is either idle or executing in usermode, and RCU believes it
2184   is non-idle, the scheduling-clock tick had better be running.
2185   Otherwise, you will get RCU CPU stall warnings. Or at best, very long
2186   (11-second) grace periods, with a pointless IPI waking the CPU from
2187   time to time.
2188#. If a CPU is in a portion of the kernel that executes RCU read-side
2189   critical sections, and RCU believes this CPU to be idle, you will get
2190   random memory corruption. **DON'T DO THIS!!!**
2191   This is one reason to test with lockdep, which will complain about
2192   this sort of thing.
2193#. If a CPU is in a portion of the kernel that is absolutely positively
2194   no-joking guaranteed to never execute any RCU read-side critical
2195   sections, and RCU believes this CPU to be idle, no problem. This
2196   sort of thing is used by some architectures for light-weight
2197   exception handlers, which can then avoid the overhead of
2198   rcu_irq_enter() and rcu_irq_exit() at exception entry and
2199   exit, respectively. Some go further and avoid the entireties of
2200   irq_enter() and irq_exit().
2201   Just make very sure you are running some of your tests with
2202   ``CONFIG_PROVE_RCU=y``, just in case one of your code paths was in
2203   fact joking about not doing RCU read-side critical sections.
2204#. If a CPU is executing in the kernel with the scheduling-clock
2205   interrupt disabled and RCU believes this CPU to be non-idle, and if
2206   the CPU goes idle (from an RCU perspective) every few jiffies, no
2207   problem. It is usually OK for there to be the occasional gap between
2208   idle periods of up to a second or so.
2209   If the gap grows too long, you get RCU CPU stall warnings.
2210#. If a CPU is either idle or executing in usermode, and RCU believes it
2211   to be idle, of course no problem.
2212#. If a CPU is executing in the kernel, the kernel code path is passing
2213   through quiescent states at a reasonable frequency (preferably about
2214   once per few jiffies, but the occasional excursion to a second or so
2215   is usually OK) and the scheduling-clock interrupt is enabled, of
2216   course no problem.
2217   If the gap between a successive pair of quiescent states grows too
2218   long, you get RCU CPU stall warnings.
2219
2220+-----------------------------------------------------------------------+
2221| **Quick Quiz**:                                                       |
2222+-----------------------------------------------------------------------+
2223| But what if my driver has a hardware interrupt handler that can run   |
2224| for many seconds? I cannot invoke schedule() from an hardware         |
2225| interrupt handler, after all!                                         |
2226+-----------------------------------------------------------------------+
2227| **Answer**:                                                           |
2228+-----------------------------------------------------------------------+
2229| One approach is to do ``rcu_irq_exit();rcu_irq_enter();`` every so    |
2230| often. But given that long-running interrupt handlers can cause other |
2231| problems, not least for response time, shouldn't you work to keep     |
2232| your interrupt handler's runtime within reasonable bounds?            |
2233+-----------------------------------------------------------------------+
2234
2235But as long as RCU is properly informed of kernel state transitions
2236between in-kernel execution, usermode execution, and idle, and as long
2237as the scheduling-clock interrupt is enabled when RCU needs it to be,
2238you can rest assured that the bugs you encounter will be in some other
2239part of RCU or some other part of the kernel!
2240
2241Memory Efficiency
2242~~~~~~~~~~~~~~~~~
2243
2244Although small-memory non-realtime systems can simply use Tiny RCU, code
2245size is only one aspect of memory efficiency. Another aspect is the size
2246of the ``rcu_head`` structure used by call_rcu() and
2247kfree_rcu(). Although this structure contains nothing more than a
2248pair of pointers, it does appear in many RCU-protected data structures,
2249including some that are size critical. The ``page`` structure is a case
2250in point, as evidenced by the many occurrences of the ``union`` keyword
2251within that structure.
2252
2253This need for memory efficiency is one reason that RCU uses hand-crafted
2254singly linked lists to track the ``rcu_head`` structures that are
2255waiting for a grace period to elapse. It is also the reason why
2256``rcu_head`` structures do not contain debug information, such as fields
2257tracking the file and line of the call_rcu() or kfree_rcu() that
2258posted them. Although this information might appear in debug-only kernel
2259builds at some point, in the meantime, the ``->func`` field will often
2260provide the needed debug information.
2261
2262However, in some cases, the need for memory efficiency leads to even
2263more extreme measures. Returning to the ``page`` structure, the
2264``rcu_head`` field shares storage with a great many other structures
2265that are used at various points in the corresponding page's lifetime. In
2266order to correctly resolve certain `race
2267conditions <https://lore.kernel.org/r/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com>`__,
2268the Linux kernel's memory-management subsystem needs a particular bit to
2269remain zero during all phases of grace-period processing, and that bit
2270happens to map to the bottom bit of the ``rcu_head`` structure's
2271``->next`` field. RCU makes this guarantee as long as call_rcu() is
2272used to post the callback, as opposed to kfree_rcu() or some future
2273“lazy” variant of call_rcu() that might one day be created for
2274energy-efficiency purposes.
2275
2276That said, there are limits. RCU requires that the ``rcu_head``
2277structure be aligned to a two-byte boundary, and passing a misaligned
2278``rcu_head`` structure to one of the call_rcu() family of functions
2279will result in a splat. It is therefore necessary to exercise caution
2280when packing structures containing fields of type ``rcu_head``. Why not
2281a four-byte or even eight-byte alignment requirement? Because the m68k
2282architecture provides only two-byte alignment, and thus acts as
2283alignment's least common denominator.
2284
2285The reason for reserving the bottom bit of pointers to ``rcu_head``
2286structures is to leave the door open to “lazy” callbacks whose
2287invocations can safely be deferred. Deferring invocation could
2288potentially have energy-efficiency benefits, but only if the rate of
2289non-lazy callbacks decreases significantly for some important workload.
2290In the meantime, reserving the bottom bit keeps this option open in case
2291it one day becomes useful.
2292
2293Performance, Scalability, Response Time, and Reliability
2294~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2295
2296Expanding on the `earlier
2297discussion <Performance and Scalability_>`__, RCU is used heavily by
2298hot code paths in performance-critical portions of the Linux kernel's
2299networking, security, virtualization, and scheduling code paths. RCU
2300must therefore use efficient implementations, especially in its
2301read-side primitives. To that end, it would be good if preemptible RCU's
2302implementation of rcu_read_lock() could be inlined, however, doing
2303this requires resolving ``#include`` issues with the ``task_struct``
2304structure.
2305
2306The Linux kernel supports hardware configurations with up to 4096 CPUs,
2307which means that RCU must be extremely scalable. Algorithms that involve
2308frequent acquisitions of global locks or frequent atomic operations on
2309global variables simply cannot be tolerated within the RCU
2310implementation. RCU therefore makes heavy use of a combining tree based
2311on the ``rcu_node`` structure. RCU is required to tolerate all CPUs
2312continuously invoking any combination of RCU's runtime primitives with
2313minimal per-operation overhead. In fact, in many cases, increasing load
2314must *decrease* the per-operation overhead, witness the batching
2315optimizations for synchronize_rcu(), call_rcu(),
2316synchronize_rcu_expedited(), and rcu_barrier(). As a general
2317rule, RCU must cheerfully accept whatever the rest of the Linux kernel
2318decides to throw at it.
2319
2320The Linux kernel is used for real-time workloads, especially in
2321conjunction with the `-rt
2322patchset <https://wiki.linuxfoundation.org/realtime/>`__. The
2323real-time-latency response requirements are such that the traditional
2324approach of disabling preemption across RCU read-side critical sections
2325is inappropriate. Kernels built with ``CONFIG_PREEMPTION=y`` therefore use
2326an RCU implementation that allows RCU read-side critical sections to be
2327preempted. This requirement made its presence known after users made it
2328clear that an earlier `real-time
2329patch <https://lwn.net/Articles/107930/>`__ did not meet their needs, in
2330conjunction with some `RCU
2331issues <https://lore.kernel.org/r/20050318002026.GA2693@us.ibm.com>`__
2332encountered by a very early version of the -rt patchset.
2333
2334In addition, RCU must make do with a sub-100-microsecond real-time
2335latency budget. In fact, on smaller systems with the -rt patchset, the
2336Linux kernel provides sub-20-microsecond real-time latencies for the
2337whole kernel, including RCU. RCU's scalability and latency must
2338therefore be sufficient for these sorts of configurations. To my
2339surprise, the sub-100-microsecond real-time latency budget `applies to
2340even the largest systems
2341[PDF] <http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf>`__,
2342up to and including systems with 4096 CPUs. This real-time requirement
2343motivated the grace-period kthread, which also simplified handling of a
2344number of race conditions.
2345
2346RCU must avoid degrading real-time response for CPU-bound threads,
2347whether executing in usermode (which is one use case for
2348``CONFIG_NO_HZ_FULL=y``) or in the kernel. That said, CPU-bound loops in
2349the kernel must execute cond_resched() at least once per few tens of
2350milliseconds in order to avoid receiving an IPI from RCU.
2351
2352Finally, RCU's status as a synchronization primitive means that any RCU
2353failure can result in arbitrary memory corruption that can be extremely
2354difficult to debug. This means that RCU must be extremely reliable,
2355which in practice also means that RCU must have an aggressive
2356stress-test suite. This stress-test suite is called ``rcutorture``.
2357
2358Although the need for ``rcutorture`` was no surprise, the current
2359immense popularity of the Linux kernel is posing interesting—and perhaps
2360unprecedented—validation challenges. To see this, keep in mind that
2361there are well over one billion instances of the Linux kernel running
2362today, given Android smartphones, Linux-powered televisions, and
2363servers. This number can be expected to increase sharply with the advent
2364of the celebrated Internet of Things.
2365
2366Suppose that RCU contains a race condition that manifests on average
2367once per million years of runtime. This bug will be occurring about
2368three times per *day* across the installed base. RCU could simply hide
2369behind hardware error rates, given that no one should really expect
2370their smartphone to last for a million years. However, anyone taking too
2371much comfort from this thought should consider the fact that in most
2372jurisdictions, a successful multi-year test of a given mechanism, which
2373might include a Linux kernel, suffices for a number of types of
2374safety-critical certifications. In fact, rumor has it that the Linux
2375kernel is already being used in production for safety-critical
2376applications. I don't know about you, but I would feel quite bad if a
2377bug in RCU killed someone. Which might explain my recent focus on
2378validation and verification.
2379
2380Other RCU Flavors
2381-----------------
2382
2383One of the more surprising things about RCU is that there are now no
2384fewer than five *flavors*, or API families. In addition, the primary
2385flavor that has been the sole focus up to this point has two different
2386implementations, non-preemptible and preemptible. The other four flavors
2387are listed below, with requirements for each described in a separate
2388section.
2389
2390#. `Bottom-Half Flavor (Historical)`_
2391#. `Sched Flavor (Historical)`_
2392#. `Sleepable RCU`_
2393#. `Tasks RCU`_
2394
2395Bottom-Half Flavor (Historical)
2396~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2397
2398The RCU-bh flavor of RCU has since been expressed in terms of the other
2399RCU flavors as part of a consolidation of the three flavors into a
2400single flavor. The read-side API remains, and continues to disable
2401softirq and to be accounted for by lockdep. Much of the material in this
2402section is therefore strictly historical in nature.
2403
2404The softirq-disable (AKA “bottom-half”, hence the “_bh” abbreviations)
2405flavor of RCU, or *RCU-bh*, was developed by Dipankar Sarma to provide a
2406flavor of RCU that could withstand the network-based denial-of-service
2407attacks researched by Robert Olsson. These attacks placed so much
2408networking load on the system that some of the CPUs never exited softirq
2409execution, which in turn prevented those CPUs from ever executing a
2410context switch, which, in the RCU implementation of that time, prevented
2411grace periods from ever ending. The result was an out-of-memory
2412condition and a system hang.
2413
2414The solution was the creation of RCU-bh, which does
2415local_bh_disable() across its read-side critical sections, and which
2416uses the transition from one type of softirq processing to another as a
2417quiescent state in addition to context switch, idle, user mode, and
2418offline. This means that RCU-bh grace periods can complete even when
2419some of the CPUs execute in softirq indefinitely, thus allowing
2420algorithms based on RCU-bh to withstand network-based denial-of-service
2421attacks.
2422
2423Because rcu_read_lock_bh() and rcu_read_unlock_bh() disable and
2424re-enable softirq handlers, any attempt to start a softirq handlers
2425during the RCU-bh read-side critical section will be deferred. In this
2426case, rcu_read_unlock_bh() will invoke softirq processing, which can
2427take considerable time. One can of course argue that this softirq
2428overhead should be associated with the code following the RCU-bh
2429read-side critical section rather than rcu_read_unlock_bh(), but the
2430fact is that most profiling tools cannot be expected to make this sort
2431of fine distinction. For example, suppose that a three-millisecond-long
2432RCU-bh read-side critical section executes during a time of heavy
2433networking load. There will very likely be an attempt to invoke at least
2434one softirq handler during that three milliseconds, but any such
2435invocation will be delayed until the time of the
2436rcu_read_unlock_bh(). This can of course make it appear at first
2437glance as if rcu_read_unlock_bh() was executing very slowly.
2438
2439The `RCU-bh
2440API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2441includes rcu_read_lock_bh(), rcu_read_unlock_bh(), rcu_dereference_bh(),
2442rcu_dereference_bh_check(), and rcu_read_lock_bh_held(). However, the
2443old RCU-bh update-side APIs are now gone, replaced by synchronize_rcu(),
2444synchronize_rcu_expedited(), call_rcu(), and rcu_barrier().  In addition,
2445anything that disables bottom halves also marks an RCU-bh read-side
2446critical section, including local_bh_disable() and local_bh_enable(),
2447local_irq_save() and local_irq_restore(), and so on.
2448
2449Sched Flavor (Historical)
2450~~~~~~~~~~~~~~~~~~~~~~~~~
2451
2452The RCU-sched flavor of RCU has since been expressed in terms of the
2453other RCU flavors as part of a consolidation of the three flavors into a
2454single flavor. The read-side API remains, and continues to disable
2455preemption and to be accounted for by lockdep. Much of the material in
2456this section is therefore strictly historical in nature.
2457
2458Before preemptible RCU, waiting for an RCU grace period had the side
2459effect of also waiting for all pre-existing interrupt and NMI handlers.
2460However, there are legitimate preemptible-RCU implementations that do
2461not have this property, given that any point in the code outside of an
2462RCU read-side critical section can be a quiescent state. Therefore,
2463*RCU-sched* was created, which follows “classic” RCU in that an
2464RCU-sched grace period waits for pre-existing interrupt and NMI
2465handlers. In kernels built with ``CONFIG_PREEMPTION=n``, the RCU and
2466RCU-sched APIs have identical implementations, while kernels built with
2467``CONFIG_PREEMPTION=y`` provide a separate implementation for each.
2468
2469Note well that in ``CONFIG_PREEMPTION=y`` kernels,
2470rcu_read_lock_sched() and rcu_read_unlock_sched() disable and
2471re-enable preemption, respectively. This means that if there was a
2472preemption attempt during the RCU-sched read-side critical section,
2473rcu_read_unlock_sched() will enter the scheduler, with all the
2474latency and overhead entailed. Just as with rcu_read_unlock_bh(),
2475this can make it look as if rcu_read_unlock_sched() was executing
2476very slowly. However, the highest-priority task won't be preempted, so
2477that task will enjoy low-overhead rcu_read_unlock_sched()
2478invocations.
2479
2480The `RCU-sched
2481API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2482includes rcu_read_lock_sched(), rcu_read_unlock_sched(),
2483rcu_read_lock_sched_notrace(), rcu_read_unlock_sched_notrace(),
2484rcu_dereference_sched(), rcu_dereference_sched_check(), and
2485rcu_read_lock_sched_held().  However, the old RCU-sched update-side APIs
2486are now gone, replaced by synchronize_rcu(), synchronize_rcu_expedited(),
2487call_rcu(), and rcu_barrier().  In addition, anything that disables
2488preemption also marks an RCU-sched read-side critical section,
2489including preempt_disable() and preempt_enable(), local_irq_save()
2490and local_irq_restore(), and so on.
2491
2492Sleepable RCU
2493~~~~~~~~~~~~~
2494
2495For well over a decade, someone saying “I need to block within an RCU
2496read-side critical section” was a reliable indication that this someone
2497did not understand RCU. After all, if you are always blocking in an RCU
2498read-side critical section, you can probably afford to use a
2499higher-overhead synchronization mechanism. However, that changed with
2500the advent of the Linux kernel's notifiers, whose RCU read-side critical
2501sections almost never sleep, but sometimes need to. This resulted in the
2502introduction of `sleepable RCU <https://lwn.net/Articles/202847/>`__, or
2503*SRCU*.
2504
2505SRCU allows different domains to be defined, with each such domain
2506defined by an instance of an ``srcu_struct`` structure. A pointer to
2507this structure must be passed in to each SRCU function, for example,
2508``synchronize_srcu(&ss)``, where ``ss`` is the ``srcu_struct``
2509structure. The key benefit of these domains is that a slow SRCU reader
2510in one domain does not delay an SRCU grace period in some other domain.
2511That said, one consequence of these domains is that read-side code must
2512pass a “cookie” from srcu_read_lock() to srcu_read_unlock(), for
2513example, as follows:
2514
2515   ::
2516
2517       1 int idx;
2518       2
2519       3 idx = srcu_read_lock(&ss);
2520       4 do_something();
2521       5 srcu_read_unlock(&ss, idx);
2522
2523As noted above, it is legal to block within SRCU read-side critical
2524sections, however, with great power comes great responsibility. If you
2525block forever in one of a given domain's SRCU read-side critical
2526sections, then that domain's grace periods will also be blocked forever.
2527Of course, one good way to block forever is to deadlock, which can
2528happen if any operation in a given domain's SRCU read-side critical
2529section can wait, either directly or indirectly, for that domain's grace
2530period to elapse. For example, this results in a self-deadlock:
2531
2532   ::
2533
2534       1 int idx;
2535       2
2536       3 idx = srcu_read_lock(&ss);
2537       4 do_something();
2538       5 synchronize_srcu(&ss);
2539       6 srcu_read_unlock(&ss, idx);
2540
2541However, if line 5 acquired a mutex that was held across a
2542synchronize_srcu() for domain ``ss``, deadlock would still be
2543possible. Furthermore, if line 5 acquired a mutex that was held across a
2544synchronize_srcu() for some other domain ``ss1``, and if an
2545``ss1``-domain SRCU read-side critical section acquired another mutex
2546that was held across as ``ss``-domain synchronize_srcu(), deadlock
2547would again be possible. Such a deadlock cycle could extend across an
2548arbitrarily large number of different SRCU domains. Again, with great
2549power comes great responsibility.
2550
2551Unlike the other RCU flavors, SRCU read-side critical sections can run
2552on idle and even offline CPUs. This ability requires that
2553srcu_read_lock() and srcu_read_unlock() contain memory barriers,
2554which means that SRCU readers will run a bit slower than would RCU
2555readers. It also motivates the smp_mb__after_srcu_read_unlock() API,
2556which, in combination with srcu_read_unlock(), guarantees a full
2557memory barrier.
2558
2559Also unlike other RCU flavors, synchronize_srcu() may **not** be
2560invoked from CPU-hotplug notifiers, due to the fact that SRCU grace
2561periods make use of timers and the possibility of timers being
2562temporarily “stranded” on the outgoing CPU. This stranding of timers
2563means that timers posted to the outgoing CPU will not fire until late in
2564the CPU-hotplug process. The problem is that if a notifier is waiting on
2565an SRCU grace period, that grace period is waiting on a timer, and that
2566timer is stranded on the outgoing CPU, then the notifier will never be
2567awakened, in other words, deadlock has occurred. This same situation of
2568course also prohibits srcu_barrier() from being invoked from
2569CPU-hotplug notifiers.
2570
2571SRCU also differs from other RCU flavors in that SRCU's expedited and
2572non-expedited grace periods are implemented by the same mechanism. This
2573means that in the current SRCU implementation, expediting a future grace
2574period has the side effect of expediting all prior grace periods that
2575have not yet completed. (But please note that this is a property of the
2576current implementation, not necessarily of future implementations.) In
2577addition, if SRCU has been idle for longer than the interval specified
2578by the ``srcutree.exp_holdoff`` kernel boot parameter (25 microseconds
2579by default), and if a synchronize_srcu() invocation ends this idle
2580period, that invocation will be automatically expedited.
2581
2582As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating a
2583locking bottleneck present in prior kernel versions. Although this will
2584allow users to put much heavier stress on call_srcu(), it is
2585important to note that SRCU does not yet take any special steps to deal
2586with callback flooding. So if you are posting (say) 10,000 SRCU
2587callbacks per second per CPU, you are probably totally OK, but if you
2588intend to post (say) 1,000,000 SRCU callbacks per second per CPU, please
2589run some tests first. SRCU just might need a few adjustment to deal with
2590that sort of load. Of course, your mileage may vary based on the speed
2591of your CPUs and the size of your memory.
2592
2593The `SRCU
2594API <https://lwn.net/Articles/609973/#RCU%20Per-Flavor%20API%20Table>`__
2595includes srcu_read_lock(), srcu_read_unlock(),
2596srcu_dereference(), srcu_dereference_check(),
2597synchronize_srcu(), synchronize_srcu_expedited(),
2598call_srcu(), srcu_barrier(), and srcu_read_lock_held(). It
2599also includes DEFINE_SRCU(), DEFINE_STATIC_SRCU(), and
2600init_srcu_struct() APIs for defining and initializing
2601``srcu_struct`` structures.
2602
2603More recently, the SRCU API has added polling interfaces:
2604
2605#. start_poll_synchronize_srcu() returns a cookie identifying
2606   the completion of a future SRCU grace period and ensures
2607   that this grace period will be started.
2608#. poll_state_synchronize_srcu() returns ``true`` iff the
2609   specified cookie corresponds to an already-completed
2610   SRCU grace period.
2611#. get_state_synchronize_srcu() returns a cookie just like
2612   start_poll_synchronize_srcu() does, but differs in that
2613   it does nothing to ensure that any future SRCU grace period
2614   will be started.
2615
2616These functions are used to avoid unnecessary SRCU grace periods in
2617certain types of buffer-cache algorithms having multi-stage age-out
2618mechanisms.  The idea is that by the time the block has aged completely
2619from the cache, an SRCU grace period will be very likely to have elapsed.
2620
2621Tasks RCU
2622~~~~~~~~~
2623
2624Some forms of tracing use “trampolines” to handle the binary rewriting
2625required to install different types of probes. It would be good to be
2626able to free old trampolines, which sounds like a job for some form of
2627RCU. However, because it is necessary to be able to install a trace
2628anywhere in the code, it is not possible to use read-side markers such
2629as rcu_read_lock() and rcu_read_unlock(). In addition, it does
2630not work to have these markers in the trampoline itself, because there
2631would need to be instructions following rcu_read_unlock(). Although
2632synchronize_rcu() would guarantee that execution reached the
2633rcu_read_unlock(), it would not be able to guarantee that execution
2634had completely left the trampoline. Worse yet, in some situations
2635the trampoline's protection must extend a few instructions *prior* to
2636execution reaching the trampoline.  For example, these few instructions
2637might calculate the address of the trampoline, so that entering the
2638trampoline would be pre-ordained a surprisingly long time before execution
2639actually reached the trampoline itself.
2640
2641The solution, in the form of `Tasks
2642RCU <https://lwn.net/Articles/607117/>`__, is to have implicit read-side
2643critical sections that are delimited by voluntary context switches, that
2644is, calls to schedule(), cond_resched(), and
2645synchronize_rcu_tasks(). In addition, transitions to and from
2646userspace execution also delimit tasks-RCU read-side critical sections.
2647
2648The tasks-RCU API is quite compact, consisting only of
2649call_rcu_tasks(), synchronize_rcu_tasks(), and
2650rcu_barrier_tasks(). In ``CONFIG_PREEMPTION=n`` kernels, trampolines
2651cannot be preempted, so these APIs map to call_rcu(),
2652synchronize_rcu(), and rcu_barrier(), respectively. In
2653``CONFIG_PREEMPTION=y`` kernels, trampolines can be preempted, and these
2654three APIs are therefore implemented by separate functions that check
2655for voluntary context switches.
2656
2657Possible Future Changes
2658-----------------------
2659
2660One of the tricks that RCU uses to attain update-side scalability is to
2661increase grace-period latency with increasing numbers of CPUs. If this
2662becomes a serious problem, it will be necessary to rework the
2663grace-period state machine so as to avoid the need for the additional
2664latency.
2665
2666RCU disables CPU hotplug in a few places, perhaps most notably in the
2667rcu_barrier() operations. If there is a strong reason to use
2668rcu_barrier() in CPU-hotplug notifiers, it will be necessary to
2669avoid disabling CPU hotplug. This would introduce some complexity, so
2670there had better be a *very* good reason.
2671
2672The tradeoff between grace-period latency on the one hand and
2673interruptions of other CPUs on the other hand may need to be
2674re-examined. The desire is of course for zero grace-period latency as
2675well as zero interprocessor interrupts undertaken during an expedited
2676grace period operation. While this ideal is unlikely to be achievable,
2677it is quite possible that further improvements can be made.
2678
2679The multiprocessor implementations of RCU use a combining tree that
2680groups CPUs so as to reduce lock contention and increase cache locality.
2681However, this combining tree does not spread its memory across NUMA
2682nodes nor does it align the CPU groups with hardware features such as
2683sockets or cores. Such spreading and alignment is currently believed to
2684be unnecessary because the hotpath read-side primitives do not access
2685the combining tree, nor does call_rcu() in the common case. If you
2686believe that your architecture needs such spreading and alignment, then
2687your architecture should also benefit from the
2688``rcutree.rcu_fanout_leaf`` boot parameter, which can be set to the
2689number of CPUs in a socket, NUMA node, or whatever. If the number of
2690CPUs is too large, use a fraction of the number of CPUs. If the number
2691of CPUs is a large prime number, well, that certainly is an
2692“interesting” architectural choice! More flexible arrangements might be
2693considered, but only if ``rcutree.rcu_fanout_leaf`` has proven
2694inadequate, and only if the inadequacy has been demonstrated by a
2695carefully run and realistic system-level workload.
2696
2697Please note that arrangements that require RCU to remap CPU numbers will
2698require extremely good demonstration of need and full exploration of
2699alternatives.
2700
2701RCU's various kthreads are reasonably recent additions. It is quite
2702likely that adjustments will be required to more gracefully handle
2703extreme loads. It might also be necessary to be able to relate CPU
2704utilization by RCU's kthreads and softirq handlers to the code that
2705instigated this CPU utilization. For example, RCU callback overhead
2706might be charged back to the originating call_rcu() instance, though
2707probably not in production kernels.
2708
2709Additional work may be required to provide reasonable forward-progress
2710guarantees under heavy load for grace periods and for callback
2711invocation.
2712
2713Summary
2714-------
2715
2716This document has presented more than two decade's worth of RCU
2717requirements. Given that the requirements keep changing, this will not
2718be the last word on this subject, but at least it serves to get an
2719important subset of the requirements set forth.
2720
2721Acknowledgments
2722---------------
2723
2724I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar, Oleg
2725Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and Andy
2726Lutomirski for their help in rendering this article human readable, and
2727to Michelle Rankin for her support of this effort. Other contributions
2728are acknowledged in the Linux kernel's git archive.
2729