aboutsummaryrefslogtreecommitdiffstats
path: root/Documentation/RCU/whatisRCU.txt
diff options
context:
space:
mode:
authorPaul E. McKenney <paulmck@us.ibm.com>2005-09-10 03:26:24 -0400
committerLinus Torvalds <torvalds@g5.osdl.org>2005-09-10 13:06:24 -0400
commitdd81eca83c8300c95d8a1eaf0d38f56513711535 (patch)
treea1cb1aee24b703b86f7197cfe4f379529a683c5a /Documentation/RCU/whatisRCU.txt
parent12c62c2e9abf8da804fe1def1f5bb44d023f569f (diff)
[PATCH] Yet another RCU documentation update
Update RCU documentation based on discussions and review of RCU-based tree patches. Add an introductory whatisRCU.txt file. Signed-off-by: <paulmck@us.ibm.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
Diffstat (limited to 'Documentation/RCU/whatisRCU.txt')
-rw-r--r--Documentation/RCU/whatisRCU.txt902
1 files changed, 902 insertions, 0 deletions
diff --git a/Documentation/RCU/whatisRCU.txt b/Documentation/RCU/whatisRCU.txt
new file mode 100644
index 000000000000..354d89c78377
--- /dev/null
+++ b/Documentation/RCU/whatisRCU.txt
@@ -0,0 +1,902 @@
1What is RCU?
2
3RCU is a synchronization mechanism that was added to the Linux kernel
4during the 2.5 development effort that is optimized for read-mostly
5situations. Although RCU is actually quite simple once you understand it,
6getting there can sometimes be a challenge. Part of the problem is that
7most of the past descriptions of RCU have been written with the mistaken
8assumption that there is "one true way" to describe RCU. Instead,
9the experience has been that different people must take different paths
10to arrive at an understanding of RCU. This document provides several
11different paths, as follows:
12
131. RCU OVERVIEW
142. WHAT IS RCU'S CORE API?
153. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
164. WHAT IF MY UPDATING THREAD CANNOT BLOCK?
175. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
186. ANALOGY WITH READER-WRITER LOCKING
197. FULL LIST OF RCU APIs
208. ANSWERS TO QUICK QUIZZES
21
22People who prefer starting with a conceptual overview should focus on
23Section 1, though most readers will profit by reading this section at
24some point. People who prefer to start with an API that they can then
25experiment with should focus on Section 2. People who prefer to start
26with example uses should focus on Sections 3 and 4. People who need to
27understand the RCU implementation should focus on Section 5, then dive
28into the kernel source code. People who reason best by analogy should
29focus on Section 6. Section 7 serves as an index to the docbook API
30documentation, and Section 8 is the traditional answer key.
31
32So, start with the section that makes the most sense to you and your
33preferred method of learning. If you need to know everything about
34everything, feel free to read the whole thing -- but if you are really
35that type of person, you have perused the source code and will therefore
36never need this document anyway. ;-)
37
38
391. RCU OVERVIEW
40
41The basic idea behind RCU is to split updates into "removal" and
42"reclamation" phases. The removal phase removes references to data items
43within a data structure (possibly by replacing them with references to
44new versions of these data items), and can run concurrently with readers.
45The reason that it is safe to run the removal phase concurrently with
46readers is the semantics of modern CPUs guarantee that readers will see
47either the old or the new version of the data structure rather than a
48partially updated reference. The reclamation phase does the work of reclaiming
49(e.g., freeing) the data items removed from the data structure during the
50removal phase. Because reclaiming data items can disrupt any readers
51concurrently referencing those data items, the reclamation phase must
52not start until readers no longer hold references to those data items.
53
54Splitting the update into removal and reclamation phases permits the
55updater to perform the removal phase immediately, and to defer the
56reclamation phase until all readers active during the removal phase have
57completed, either by blocking until they finish or by registering a
58callback that is invoked after they finish. Only readers that are active
59during the removal phase need be considered, because any reader starting
60after the removal phase will be unable to gain a reference to the removed
61data items, and therefore cannot be disrupted by the reclamation phase.
62
63So the typical RCU update sequence goes something like the following:
64
65a. Remove pointers to a data structure, so that subsequent
66 readers cannot gain a reference to it.
67
68b. Wait for all previous readers to complete their RCU read-side
69 critical sections.
70
71c. At this point, there cannot be any readers who hold references
72 to the data structure, so it now may safely be reclaimed
73 (e.g., kfree()d).
74
75Step (b) above is the key idea underlying RCU's deferred destruction.
76The ability to wait until all readers are done allows RCU readers to
77use much lighter-weight synchronization, in some cases, absolutely no
78synchronization at all. In contrast, in more conventional lock-based
79schemes, readers must use heavy-weight synchronization in order to
80prevent an updater from deleting the data structure out from under them.
81This is because lock-based updaters typically update data items in place,
82and must therefore exclude readers. In contrast, RCU-based updaters
83typically take advantage of the fact that writes to single aligned
84pointers are atomic on modern CPUs, allowing atomic insertion, removal,
85and replacement of data items in a linked structure without disrupting
86readers. Concurrent RCU readers can then continue accessing the old
87versions, and can dispense with the atomic operations, memory barriers,
88and communications cache misses that are so expensive on present-day
89SMP computer systems, even in absence of lock contention.
90
91In the three-step procedure shown above, the updater is performing both
92the removal and the reclamation step, but it is often helpful for an
93entirely different thread to do the reclamation, as is in fact the case
94in the Linux kernel's directory-entry cache (dcache). Even if the same
95thread performs both the update step (step (a) above) and the reclamation
96step (step (c) above), it is often helpful to think of them separately.
97For example, RCU readers and updaters need not communicate at all,
98but RCU provides implicit low-overhead communication between readers
99and reclaimers, namely, in step (b) above.
100
101So how the heck can a reclaimer tell when a reader is done, given
102that readers are not doing any sort of synchronization operations???
103Read on to learn about how RCU's API makes this easy.
104
105
1062. WHAT IS RCU'S CORE API?
107
108The core RCU API is quite small:
109
110a. rcu_read_lock()
111b. rcu_read_unlock()
112c. synchronize_rcu() / call_rcu()
113d. rcu_assign_pointer()
114e. rcu_dereference()
115
116There are many other members of the RCU API, but the rest can be
117expressed in terms of these five, though most implementations instead
118express synchronize_rcu() in terms of the call_rcu() callback API.
119
120The five core RCU APIs are described below, the other 18 will be enumerated
121later. See the kernel docbook documentation for more info, or look directly
122at the function header comments.
123
124rcu_read_lock()
125
126 void rcu_read_lock(void);
127
128 Used by a reader to inform the reclaimer that the reader is
129 entering an RCU read-side critical section. It is illegal
130 to block while in an RCU read-side critical section, though
131 kernels built with CONFIG_PREEMPT_RCU can preempt RCU read-side
132 critical sections. Any RCU-protected data structure accessed
133 during an RCU read-side critical section is guaranteed to remain
134 unreclaimed for the full duration of that critical section.
135 Reference counts may be used in conjunction with RCU to maintain
136 longer-term references to data structures.
137
138rcu_read_unlock()
139
140 void rcu_read_unlock(void);
141
142 Used by a reader to inform the reclaimer that the reader is
143 exiting an RCU read-side critical section. Note that RCU
144 read-side critical sections may be nested and/or overlapping.
145
146synchronize_rcu()
147
148 void synchronize_rcu(void);
149
150 Marks the end of updater code and the beginning of reclaimer
151 code. It does this by blocking until all pre-existing RCU
152 read-side critical sections on all CPUs have completed.
153 Note that synchronize_rcu() will -not- necessarily wait for
154 any subsequent RCU read-side critical sections to complete.
155 For example, consider the following sequence of events:
156
157 CPU 0 CPU 1 CPU 2
158 ----------------- ------------------------- ---------------
159 1. rcu_read_lock()
160 2. enters synchronize_rcu()
161 3. rcu_read_lock()
162 4. rcu_read_unlock()
163 5. exits synchronize_rcu()
164 6. rcu_read_unlock()
165
166 To reiterate, synchronize_rcu() waits only for ongoing RCU
167 read-side critical sections to complete, not necessarily for
168 any that begin after synchronize_rcu() is invoked.
169
170 Of course, synchronize_rcu() does not necessarily return
171 -immediately- after the last pre-existing RCU read-side critical
172 section completes. For one thing, there might well be scheduling
173 delays. For another thing, many RCU implementations process
174 requests in batches in order to improve efficiencies, which can
175 further delay synchronize_rcu().
176
177 Since synchronize_rcu() is the API that must figure out when
178 readers are done, its implementation is key to RCU. For RCU
179 to be useful in all but the most read-intensive situations,
180 synchronize_rcu()'s overhead must also be quite small.
181
182 The call_rcu() API is a callback form of synchronize_rcu(),
183 and is described in more detail in a later section. Instead of
184 blocking, it registers a function and argument which are invoked
185 after all ongoing RCU read-side critical sections have completed.
186 This callback variant is particularly useful in situations where
187 it is illegal to block.
188
189rcu_assign_pointer()
190
191 typeof(p) rcu_assign_pointer(p, typeof(p) v);
192
193 Yes, rcu_assign_pointer() -is- implemented as a macro, though it
194 would be cool to be able to declare a function in this manner.
195 (Compiler experts will no doubt disagree.)
196
197 The updater uses this function to assign a new value to an
198 RCU-protected pointer, in order to safely communicate the change
199 in value from the updater to the reader. This function returns
200 the new value, and also executes any memory-barrier instructions
201 required for a given CPU architecture.
202
203 Perhaps more important, it serves to document which pointers
204 are protected by RCU. That said, rcu_assign_pointer() is most
205 frequently used indirectly, via the _rcu list-manipulation
206 primitives such as list_add_rcu().
207
208rcu_dereference()
209
210 typeof(p) rcu_dereference(p);
211
212 Like rcu_assign_pointer(), rcu_dereference() must be implemented
213 as a macro.
214
215 The reader uses rcu_dereference() to fetch an RCU-protected
216 pointer, which returns a value that may then be safely
217 dereferenced. Note that rcu_deference() does not actually
218 dereference the pointer, instead, it protects the pointer for
219 later dereferencing. It also executes any needed memory-barrier
220 instructions for a given CPU architecture. Currently, only Alpha
221 needs memory barriers within rcu_dereference() -- on other CPUs,
222 it compiles to nothing, not even a compiler directive.
223
224 Common coding practice uses rcu_dereference() to copy an
225 RCU-protected pointer to a local variable, then dereferences
226 this local variable, for example as follows:
227
228 p = rcu_dereference(head.next);
229 return p->data;
230
231 However, in this case, one could just as easily combine these
232 into one statement:
233
234 return rcu_dereference(head.next)->data;
235
236 If you are going to be fetching multiple fields from the
237 RCU-protected structure, using the local variable is of
238 course preferred. Repeated rcu_dereference() calls look
239 ugly and incur unnecessary overhead on Alpha CPUs.
240
241 Note that the value returned by rcu_dereference() is valid
242 only within the enclosing RCU read-side critical section.
243 For example, the following is -not- legal:
244
245 rcu_read_lock();
246 p = rcu_dereference(head.next);
247 rcu_read_unlock();
248 x = p->address;
249 rcu_read_lock();
250 y = p->data;
251 rcu_read_unlock();
252
253 Holding a reference from one RCU read-side critical section
254 to another is just as illegal as holding a reference from
255 one lock-based critical section to another! Similarly,
256 using a reference outside of the critical section in which
257 it was acquired is just as illegal as doing so with normal
258 locking.
259
260 As with rcu_assign_pointer(), an important function of
261 rcu_dereference() is to document which pointers are protected
262 by RCU. And, again like rcu_assign_pointer(), rcu_dereference()
263 is typically used indirectly, via the _rcu list-manipulation
264 primitives, such as list_for_each_entry_rcu().
265
266The following diagram shows how each API communicates among the
267reader, updater, and reclaimer.
268
269
270 rcu_assign_pointer()
271 +--------+
272 +---------------------->| reader |---------+
273 | +--------+ |
274 | | |
275 | | | Protect:
276 | | | rcu_read_lock()
277 | | | rcu_read_unlock()
278 | rcu_dereference() | |
279 +---------+ | |
280 | updater |<---------------------+ |
281 +---------+ V
282 | +-----------+
283 +----------------------------------->| reclaimer |
284 +-----------+
285 Defer:
286 synchronize_rcu() & call_rcu()
287
288
289The RCU infrastructure observes the time sequence of rcu_read_lock(),
290rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in
291order to determine when (1) synchronize_rcu() invocations may return
292to their callers and (2) call_rcu() callbacks may be invoked. Efficient
293implementations of the RCU infrastructure make heavy use of batching in
294order to amortize their overhead over many uses of the corresponding APIs.
295
296There are no fewer than three RCU mechanisms in the Linux kernel; the
297diagram above shows the first one, which is by far the most commonly used.
298The rcu_dereference() and rcu_assign_pointer() primitives are used for
299all three mechanisms, but different defer and protect primitives are
300used as follows:
301
302 Defer Protect
303
304a. synchronize_rcu() rcu_read_lock() / rcu_read_unlock()
305 call_rcu()
306
307b. call_rcu_bh() rcu_read_lock_bh() / rcu_read_unlock_bh()
308
309c. synchronize_sched() preempt_disable() / preempt_enable()
310 local_irq_save() / local_irq_restore()
311 hardirq enter / hardirq exit
312 NMI enter / NMI exit
313
314These three mechanisms are used as follows:
315
316a. RCU applied to normal data structures.
317
318b. RCU applied to networking data structures that may be subjected
319 to remote denial-of-service attacks.
320
321c. RCU applied to scheduler and interrupt/NMI-handler tasks.
322
323Again, most uses will be of (a). The (b) and (c) cases are important
324for specialized uses, but are relatively uncommon.
325
326
3273. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
328
329This section shows a simple use of the core RCU API to protect a
330global pointer to a dynamically allocated structure. More typical
331uses of RCU may be found in listRCU.txt, arrayRCU.txt, and NMI-RCU.txt.
332
333 struct foo {
334 int a;
335 char b;
336 long c;
337 };
338 DEFINE_SPINLOCK(foo_mutex);
339
340 struct foo *gbl_foo;
341
342 /*
343 * Create a new struct foo that is the same as the one currently
344 * pointed to by gbl_foo, except that field "a" is replaced
345 * with "new_a". Points gbl_foo to the new structure, and
346 * frees up the old structure after a grace period.
347 *
348 * Uses rcu_assign_pointer() to ensure that concurrent readers
349 * see the initialized version of the new structure.
350 *
351 * Uses synchronize_rcu() to ensure that any readers that might
352 * have references to the old structure complete before freeing
353 * the old structure.
354 */
355 void foo_update_a(int new_a)
356 {
357 struct foo *new_fp;
358 struct foo *old_fp;
359
360 new_fp = kmalloc(sizeof(*fp), GFP_KERNEL);
361 spin_lock(&foo_mutex);
362 old_fp = gbl_foo;
363 *new_fp = *old_fp;
364 new_fp->a = new_a;
365 rcu_assign_pointer(gbl_foo, new_fp);
366 spin_unlock(&foo_mutex);
367 synchronize_rcu();
368 kfree(old_fp);
369 }
370
371 /*
372 * Return the value of field "a" of the current gbl_foo
373 * structure. Use rcu_read_lock() and rcu_read_unlock()
374 * to ensure that the structure does not get deleted out
375 * from under us, and use rcu_dereference() to ensure that
376 * we see the initialized version of the structure (important
377 * for DEC Alpha and for people reading the code).
378 */
379 int foo_get_a(void)
380 {
381 int retval;
382
383 rcu_read_lock();
384 retval = rcu_dereference(gbl_foo)->a;
385 rcu_read_unlock();
386 return retval;
387 }
388
389So, to sum up:
390
391o Use rcu_read_lock() and rcu_read_unlock() to guard RCU
392 read-side critical sections.
393
394o Within an RCU read-side critical section, use rcu_dereference()
395 to dereference RCU-protected pointers.
396
397o Use some solid scheme (such as locks or semaphores) to
398 keep concurrent updates from interfering with each other.
399
400o Use rcu_assign_pointer() to update an RCU-protected pointer.
401 This primitive protects concurrent readers from the updater,
402 -not- concurrent updates from each other! You therefore still
403 need to use locking (or something similar) to keep concurrent
404 rcu_assign_pointer() primitives from interfering with each other.
405
406o Use synchronize_rcu() -after- removing a data element from an
407 RCU-protected data structure, but -before- reclaiming/freeing
408 the data element, in order to wait for the completion of all
409 RCU read-side critical sections that might be referencing that
410 data item.
411
412See checklist.txt for additional rules to follow when using RCU.
413
414
4154. WHAT IF MY UPDATING THREAD CANNOT BLOCK?
416
417In the example above, foo_update_a() blocks until a grace period elapses.
418This is quite simple, but in some cases one cannot afford to wait so
419long -- there might be other high-priority work to be done.
420
421In such cases, one uses call_rcu() rather than synchronize_rcu().
422The call_rcu() API is as follows:
423
424 void call_rcu(struct rcu_head * head,
425 void (*func)(struct rcu_head *head));
426
427This function invokes func(head) after a grace period has elapsed.
428This invocation might happen from either softirq or process context,
429so the function is not permitted to block. The foo struct needs to
430have an rcu_head structure added, perhaps as follows:
431
432 struct foo {
433 int a;
434 char b;
435 long c;
436 struct rcu_head rcu;
437 };
438
439The foo_update_a() function might then be written as follows:
440
441 /*
442 * Create a new struct foo that is the same as the one currently
443 * pointed to by gbl_foo, except that field "a" is replaced
444 * with "new_a". Points gbl_foo to the new structure, and
445 * frees up the old structure after a grace period.
446 *
447 * Uses rcu_assign_pointer() to ensure that concurrent readers
448 * see the initialized version of the new structure.
449 *
450 * Uses call_rcu() to ensure that any readers that might have
451 * references to the old structure complete before freeing the
452 * old structure.
453 */
454 void foo_update_a(int new_a)
455 {
456 struct foo *new_fp;
457 struct foo *old_fp;
458
459 new_fp = kmalloc(sizeof(*fp), GFP_KERNEL);
460 spin_lock(&foo_mutex);
461 old_fp = gbl_foo;
462 *new_fp = *old_fp;
463 new_fp->a = new_a;
464 rcu_assign_pointer(gbl_foo, new_fp);
465 spin_unlock(&foo_mutex);
466 call_rcu(&old_fp->rcu, foo_reclaim);
467 }
468
469The foo_reclaim() function might appear as follows:
470
471 void foo_reclaim(struct rcu_head *rp)
472 {
473 struct foo *fp = container_of(rp, struct foo, rcu);
474
475 kfree(fp);
476 }
477
478The container_of() primitive is a macro that, given a pointer into a
479struct, the type of the struct, and the pointed-to field within the
480struct, returns a pointer to the beginning of the struct.
481
482The use of call_rcu() permits the caller of foo_update_a() to
483immediately regain control, without needing to worry further about the
484old version of the newly updated element. It also clearly shows the
485RCU distinction between updater, namely foo_update_a(), and reclaimer,
486namely foo_reclaim().
487
488The summary of advice is the same as for the previous section, except
489that we are now using call_rcu() rather than synchronize_rcu():
490
491o Use call_rcu() -after- removing a data element from an
492 RCU-protected data structure in order to register a callback
493 function that will be invoked after the completion of all RCU
494 read-side critical sections that might be referencing that
495 data item.
496
497Again, see checklist.txt for additional rules governing the use of RCU.
498
499
5005. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
501
502One of the nice things about RCU is that it has extremely simple "toy"
503implementations that are a good first step towards understanding the
504production-quality implementations in the Linux kernel. This section
505presents two such "toy" implementations of RCU, one that is implemented
506in terms of familiar locking primitives, and another that more closely
507resembles "classic" RCU. Both are way too simple for real-world use,
508lacking both functionality and performance. However, they are useful
509in getting a feel for how RCU works. See kernel/rcupdate.c for a
510production-quality implementation, and see:
511
512 http://www.rdrop.com/users/paulmck/RCU
513
514for papers describing the Linux kernel RCU implementation. The OLS'01
515and OLS'02 papers are a good introduction, and the dissertation provides
516more details on the current implementation.
517
518
5195A. "TOY" IMPLEMENTATION #1: LOCKING
520
521This section presents a "toy" RCU implementation that is based on
522familiar locking primitives. Its overhead makes it a non-starter for
523real-life use, as does its lack of scalability. It is also unsuitable
524for realtime use, since it allows scheduling latency to "bleed" from
525one read-side critical section to another.
526
527However, it is probably the easiest implementation to relate to, so is
528a good starting point.
529
530It is extremely simple:
531
532 static DEFINE_RWLOCK(rcu_gp_mutex);
533
534 void rcu_read_lock(void)
535 {
536 read_lock(&rcu_gp_mutex);
537 }
538
539 void rcu_read_unlock(void)
540 {
541 read_unlock(&rcu_gp_mutex);
542 }
543
544 void synchronize_rcu(void)
545 {
546 write_lock(&rcu_gp_mutex);
547 write_unlock(&rcu_gp_mutex);
548 }
549
550[You can ignore rcu_assign_pointer() and rcu_dereference() without
551missing much. But here they are anyway. And whatever you do, don't
552forget about them when submitting patches making use of RCU!]
553
554 #define rcu_assign_pointer(p, v) ({ \
555 smp_wmb(); \
556 (p) = (v); \
557 })
558
559 #define rcu_dereference(p) ({ \
560 typeof(p) _________p1 = p; \
561 smp_read_barrier_depends(); \
562 (_________p1); \
563 })
564
565
566The rcu_read_lock() and rcu_read_unlock() primitive read-acquire
567and release a global reader-writer lock. The synchronize_rcu()
568primitive write-acquires this same lock, then immediately releases
569it. This means that once synchronize_rcu() exits, all RCU read-side
570critical sections that were in progress before synchonize_rcu() was
571called are guaranteed to have completed -- there is no way that
572synchronize_rcu() would have been able to write-acquire the lock
573otherwise.
574
575It is possible to nest rcu_read_lock(), since reader-writer locks may
576be recursively acquired. Note also that rcu_read_lock() is immune
577from deadlock (an important property of RCU). The reason for this is
578that the only thing that can block rcu_read_lock() is a synchronize_rcu().
579But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex,
580so there can be no deadlock cycle.
581
582Quick Quiz #1: Why is this argument naive? How could a deadlock
583 occur when using this algorithm in a real-world Linux
584 kernel? How could this deadlock be avoided?
585
586
5875B. "TOY" EXAMPLE #2: CLASSIC RCU
588
589This section presents a "toy" RCU implementation that is based on
590"classic RCU". It is also short on performance (but only for updates) and
591on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT
592kernels. The definitions of rcu_dereference() and rcu_assign_pointer()
593are the same as those shown in the preceding section, so they are omitted.
594
595 void rcu_read_lock(void) { }
596
597 void rcu_read_unlock(void) { }
598
599 void synchronize_rcu(void)
600 {
601 int cpu;
602
603 for_each_cpu(cpu)
604 run_on(cpu);
605 }
606
607Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing.
608This is the great strength of classic RCU in a non-preemptive kernel:
609read-side overhead is precisely zero, at least on non-Alpha CPUs.
610And there is absolutely no way that rcu_read_lock() can possibly
611participate in a deadlock cycle!
612
613The implementation of synchronize_rcu() simply schedules itself on each
614CPU in turn. The run_on() primitive can be implemented straightforwardly
615in terms of the sched_setaffinity() primitive. Of course, a somewhat less
616"toy" implementation would restore the affinity upon completion rather
617than just leaving all tasks running on the last CPU, but when I said
618"toy", I meant -toy-!
619
620So how the heck is this supposed to work???
621
622Remember that it is illegal to block while in an RCU read-side critical
623section. Therefore, if a given CPU executes a context switch, we know
624that it must have completed all preceding RCU read-side critical sections.
625Once -all- CPUs have executed a context switch, then -all- preceding
626RCU read-side critical sections will have completed.
627
628So, suppose that we remove a data item from its structure and then invoke
629synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed
630that there are no RCU read-side critical sections holding a reference
631to that data item, so we can safely reclaim it.
632
633Quick Quiz #2: Give an example where Classic RCU's read-side
634 overhead is -negative-.
635
636Quick Quiz #3: If it is illegal to block in an RCU read-side
637 critical section, what the heck do you do in
638 PREEMPT_RT, where normal spinlocks can block???
639
640
6416. ANALOGY WITH READER-WRITER LOCKING
642
643Although RCU can be used in many different ways, a very common use of
644RCU is analogous to reader-writer locking. The following unified
645diff shows how closely related RCU and reader-writer locking can be.
646
647 @@ -13,15 +14,15 @@
648 struct list_head *lp;
649 struct el *p;
650
651 - read_lock();
652 - list_for_each_entry(p, head, lp) {
653 + rcu_read_lock();
654 + list_for_each_entry_rcu(p, head, lp) {
655 if (p->key == key) {
656 *result = p->data;
657 - read_unlock();
658 + rcu_read_unlock();
659 return 1;
660 }
661 }
662 - read_unlock();
663 + rcu_read_unlock();
664 return 0;
665 }
666
667 @@ -29,15 +30,16 @@
668 {
669 struct el *p;
670
671 - write_lock(&listmutex);
672 + spin_lock(&listmutex);
673 list_for_each_entry(p, head, lp) {
674 if (p->key == key) {
675 list_del(&p->list);
676 - write_unlock(&listmutex);
677 + spin_unlock(&listmutex);
678 + synchronize_rcu();
679 kfree(p);
680 return 1;
681 }
682 }
683 - write_unlock(&listmutex);
684 + spin_unlock(&listmutex);
685 return 0;
686 }
687
688Or, for those who prefer a side-by-side listing:
689
690 1 struct el { 1 struct el {
691 2 struct list_head list; 2 struct list_head list;
692 3 long key; 3 long key;
693 4 spinlock_t mutex; 4 spinlock_t mutex;
694 5 int data; 5 int data;
695 6 /* Other data fields */ 6 /* Other data fields */
696 7 }; 7 };
697 8 spinlock_t listmutex; 8 spinlock_t listmutex;
698 9 struct el head; 9 struct el head;
699
700 1 int search(long key, int *result) 1 int search(long key, int *result)
701 2 { 2 {
702 3 struct list_head *lp; 3 struct list_head *lp;
703 4 struct el *p; 4 struct el *p;
704 5 5
705 6 read_lock(); 6 rcu_read_lock();
706 7 list_for_each_entry(p, head, lp) { 7 list_for_each_entry_rcu(p, head, lp) {
707 8 if (p->key == key) { 8 if (p->key == key) {
708 9 *result = p->data; 9 *result = p->data;
70910 read_unlock(); 10 rcu_read_unlock();
71011 return 1; 11 return 1;
71112 } 12 }
71213 } 13 }
71314 read_unlock(); 14 rcu_read_unlock();
71415 return 0; 15 return 0;
71516 } 16 }
716
717 1 int delete(long key) 1 int delete(long key)
718 2 { 2 {
719 3 struct el *p; 3 struct el *p;
720 4 4
721 5 write_lock(&listmutex); 5 spin_lock(&listmutex);
722 6 list_for_each_entry(p, head, lp) { 6 list_for_each_entry(p, head, lp) {
723 7 if (p->key == key) { 7 if (p->key == key) {
724 8 list_del(&p->list); 8 list_del(&p->list);
725 9 write_unlock(&listmutex); 9 spin_unlock(&listmutex);
726 10 synchronize_rcu();
72710 kfree(p); 11 kfree(p);
72811 return 1; 12 return 1;
72912 } 13 }
73013 } 14 }
73114 write_unlock(&listmutex); 15 spin_unlock(&listmutex);
73215 return 0; 16 return 0;
73316 } 17 }
734
735Either way, the differences are quite small. Read-side locking moves
736to rcu_read_lock() and rcu_read_unlock, update-side locking moves from
737from a reader-writer lock to a simple spinlock, and a synchronize_rcu()
738precedes the kfree().
739
740However, there is one potential catch: the read-side and update-side
741critical sections can now run concurrently. In many cases, this will
742not be a problem, but it is necessary to check carefully regardless.
743For example, if multiple independent list updates must be seen as
744a single atomic update, converting to RCU will require special care.
745
746Also, the presence of synchronize_rcu() means that the RCU version of
747delete() can now block. If this is a problem, there is a callback-based
748mechanism that never blocks, namely call_rcu(), that can be used in
749place of synchronize_rcu().
750
751
7527. FULL LIST OF RCU APIs
753
754The RCU APIs are documented in docbook-format header comments in the
755Linux-kernel source code, but it helps to have a full list of the
756APIs, since there does not appear to be a way to categorize them
757in docbook. Here is the list, by category.
758
759Markers for RCU read-side critical sections:
760
761 rcu_read_lock
762 rcu_read_unlock
763 rcu_read_lock_bh
764 rcu_read_unlock_bh
765
766RCU pointer/list traversal:
767
768 rcu_dereference
769 list_for_each_rcu (to be deprecated in favor of
770 list_for_each_entry_rcu)
771 list_for_each_safe_rcu (deprecated, not used)
772 list_for_each_entry_rcu
773 list_for_each_continue_rcu (to be deprecated in favor of new
774 list_for_each_entry_continue_rcu)
775 hlist_for_each_rcu (to be deprecated in favor of
776 hlist_for_each_entry_rcu)
777 hlist_for_each_entry_rcu
778
779RCU pointer update:
780
781 rcu_assign_pointer
782 list_add_rcu
783 list_add_tail_rcu
784 list_del_rcu
785 list_replace_rcu
786 hlist_del_rcu
787 hlist_add_head_rcu
788
789RCU grace period:
790
791 synchronize_kernel (deprecated)
792 synchronize_net
793 synchronize_sched
794 synchronize_rcu
795 call_rcu
796 call_rcu_bh
797
798See the comment headers in the source code (or the docbook generated
799from them) for more information.
800
801
8028. ANSWERS TO QUICK QUIZZES
803
804Quick Quiz #1: Why is this argument naive? How could a deadlock
805 occur when using this algorithm in a real-world Linux
806 kernel? [Referring to the lock-based "toy" RCU
807 algorithm.]
808
809Answer: Consider the following sequence of events:
810
811 1. CPU 0 acquires some unrelated lock, call it
812 "problematic_lock".
813
814 2. CPU 1 enters synchronize_rcu(), write-acquiring
815 rcu_gp_mutex.
816
817 3. CPU 0 enters rcu_read_lock(), but must wait
818 because CPU 1 holds rcu_gp_mutex.
819
820 4. CPU 1 is interrupted, and the irq handler
821 attempts to acquire problematic_lock.
822
823 The system is now deadlocked.
824
825 One way to avoid this deadlock is to use an approach like
826 that of CONFIG_PREEMPT_RT, where all normal spinlocks
827 become blocking locks, and all irq handlers execute in
828 the context of special tasks. In this case, in step 4
829 above, the irq handler would block, allowing CPU 1 to
830 release rcu_gp_mutex, avoiding the deadlock.
831
832 Even in the absence of deadlock, this RCU implementation
833 allows latency to "bleed" from readers to other
834 readers through synchronize_rcu(). To see this,
835 consider task A in an RCU read-side critical section
836 (thus read-holding rcu_gp_mutex), task B blocked
837 attempting to write-acquire rcu_gp_mutex, and
838 task C blocked in rcu_read_lock() attempting to
839 read_acquire rcu_gp_mutex. Task A's RCU read-side
840 latency is holding up task C, albeit indirectly via
841 task B.
842
843 Realtime RCU implementations therefore use a counter-based
844 approach where tasks in RCU read-side critical sections
845 cannot be blocked by tasks executing synchronize_rcu().
846
847Quick Quiz #2: Give an example where Classic RCU's read-side
848 overhead is -negative-.
849
850Answer: Imagine a single-CPU system with a non-CONFIG_PREEMPT
851 kernel where a routing table is used by process-context
852 code, but can be updated by irq-context code (for example,
853 by an "ICMP REDIRECT" packet). The usual way of handling
854 this would be to have the process-context code disable
855 interrupts while searching the routing table. Use of
856 RCU allows such interrupt-disabling to be dispensed with.
857 Thus, without RCU, you pay the cost of disabling interrupts,
858 and with RCU you don't.
859
860 One can argue that the overhead of RCU in this
861 case is negative with respect to the single-CPU
862 interrupt-disabling approach. Others might argue that
863 the overhead of RCU is merely zero, and that replacing
864 the positive overhead of the interrupt-disabling scheme
865 with the zero-overhead RCU scheme does not constitute
866 negative overhead.
867
868 In real life, of course, things are more complex. But
869 even the theoretical possibility of negative overhead for
870 a synchronization primitive is a bit unexpected. ;-)
871
872Quick Quiz #3: If it is illegal to block in an RCU read-side
873 critical section, what the heck do you do in
874 PREEMPT_RT, where normal spinlocks can block???
875
876Answer: Just as PREEMPT_RT permits preemption of spinlock
877 critical sections, it permits preemption of RCU
878 read-side critical sections. It also permits
879 spinlocks blocking while in RCU read-side critical
880 sections.
881
882 Why the apparent inconsistency? Because it is it
883 possible to use priority boosting to keep the RCU
884 grace periods short if need be (for example, if running
885 short of memory). In contrast, if blocking waiting
886 for (say) network reception, there is no way to know
887 what should be boosted. Especially given that the
888 process we need to boost might well be a human being
889 who just went out for a pizza or something. And although
890 a computer-operated cattle prod might arouse serious
891 interest, it might also provoke serious objections.
892 Besides, how does the computer know what pizza parlor
893 the human being went to???
894
895
896ACKNOWLEDGEMENTS
897
898My thanks to the people who helped make this human-readable, including
899Jon Walpole, Josh Triplett, Serge Hallyn, and Suzanne Wood.
900
901
902For more information, see http://www.rdrop.com/users/paulmck/RCU.