Commit | Line | Data |
---|---|---|
32300751 PM |
1 | Please note that the "What is RCU?" LWN series is an excellent place |
2 | to start learning about RCU: | |
3 | ||
4 | 1. What is RCU, Fundamentally? http://lwn.net/Articles/262464/ | |
5 | 2. What is RCU? Part 2: Usage http://lwn.net/Articles/263130/ | |
6 | 3. RCU part 3: the RCU API http://lwn.net/Articles/264090/ | |
7 | ||
8 | ||
dd81eca8 PM |
9 | What is RCU? |
10 | ||
11 | RCU is a synchronization mechanism that was added to the Linux kernel | |
12 | during the 2.5 development effort that is optimized for read-mostly | |
13 | situations. Although RCU is actually quite simple once you understand it, | |
14 | getting there can sometimes be a challenge. Part of the problem is that | |
15 | most of the past descriptions of RCU have been written with the mistaken | |
16 | assumption that there is "one true way" to describe RCU. Instead, | |
17 | the experience has been that different people must take different paths | |
18 | to arrive at an understanding of RCU. This document provides several | |
19 | different paths, as follows: | |
20 | ||
21 | 1. RCU OVERVIEW | |
22 | 2. WHAT IS RCU'S CORE API? | |
23 | 3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API? | |
24 | 4. WHAT IF MY UPDATING THREAD CANNOT BLOCK? | |
25 | 5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU? | |
26 | 6. ANALOGY WITH READER-WRITER LOCKING | |
27 | 7. FULL LIST OF RCU APIs | |
28 | 8. ANSWERS TO QUICK QUIZZES | |
29 | ||
30 | People who prefer starting with a conceptual overview should focus on | |
31 | Section 1, though most readers will profit by reading this section at | |
32 | some point. People who prefer to start with an API that they can then | |
33 | experiment with should focus on Section 2. People who prefer to start | |
34 | with example uses should focus on Sections 3 and 4. People who need to | |
35 | understand the RCU implementation should focus on Section 5, then dive | |
36 | into the kernel source code. People who reason best by analogy should | |
37 | focus on Section 6. Section 7 serves as an index to the docbook API | |
38 | documentation, and Section 8 is the traditional answer key. | |
39 | ||
40 | So, start with the section that makes the most sense to you and your | |
41 | preferred method of learning. If you need to know everything about | |
42 | everything, feel free to read the whole thing -- but if you are really | |
43 | that type of person, you have perused the source code and will therefore | |
44 | never need this document anyway. ;-) | |
45 | ||
46 | ||
47 | 1. RCU OVERVIEW | |
48 | ||
49 | The basic idea behind RCU is to split updates into "removal" and | |
50 | "reclamation" phases. The removal phase removes references to data items | |
51 | within a data structure (possibly by replacing them with references to | |
52 | new versions of these data items), and can run concurrently with readers. | |
53 | The reason that it is safe to run the removal phase concurrently with | |
54 | readers is the semantics of modern CPUs guarantee that readers will see | |
55 | either the old or the new version of the data structure rather than a | |
56 | partially updated reference. The reclamation phase does the work of reclaiming | |
57 | (e.g., freeing) the data items removed from the data structure during the | |
58 | removal phase. Because reclaiming data items can disrupt any readers | |
59 | concurrently referencing those data items, the reclamation phase must | |
60 | not start until readers no longer hold references to those data items. | |
61 | ||
62 | Splitting the update into removal and reclamation phases permits the | |
63 | updater to perform the removal phase immediately, and to defer the | |
64 | reclamation phase until all readers active during the removal phase have | |
65 | completed, either by blocking until they finish or by registering a | |
66 | callback that is invoked after they finish. Only readers that are active | |
67 | during the removal phase need be considered, because any reader starting | |
68 | after the removal phase will be unable to gain a reference to the removed | |
69 | data items, and therefore cannot be disrupted by the reclamation phase. | |
70 | ||
71 | So the typical RCU update sequence goes something like the following: | |
72 | ||
73 | a. Remove pointers to a data structure, so that subsequent | |
74 | readers cannot gain a reference to it. | |
75 | ||
76 | b. Wait for all previous readers to complete their RCU read-side | |
77 | critical sections. | |
78 | ||
79 | c. At this point, there cannot be any readers who hold references | |
80 | to the data structure, so it now may safely be reclaimed | |
81 | (e.g., kfree()d). | |
82 | ||
83 | Step (b) above is the key idea underlying RCU's deferred destruction. | |
84 | The ability to wait until all readers are done allows RCU readers to | |
85 | use much lighter-weight synchronization, in some cases, absolutely no | |
86 | synchronization at all. In contrast, in more conventional lock-based | |
87 | schemes, readers must use heavy-weight synchronization in order to | |
88 | prevent an updater from deleting the data structure out from under them. | |
89 | This is because lock-based updaters typically update data items in place, | |
90 | and must therefore exclude readers. In contrast, RCU-based updaters | |
91 | typically take advantage of the fact that writes to single aligned | |
92 | pointers are atomic on modern CPUs, allowing atomic insertion, removal, | |
93 | and replacement of data items in a linked structure without disrupting | |
94 | readers. Concurrent RCU readers can then continue accessing the old | |
95 | versions, and can dispense with the atomic operations, memory barriers, | |
96 | and communications cache misses that are so expensive on present-day | |
97 | SMP computer systems, even in absence of lock contention. | |
98 | ||
99 | In the three-step procedure shown above, the updater is performing both | |
100 | the removal and the reclamation step, but it is often helpful for an | |
101 | entirely different thread to do the reclamation, as is in fact the case | |
102 | in the Linux kernel's directory-entry cache (dcache). Even if the same | |
103 | thread performs both the update step (step (a) above) and the reclamation | |
104 | step (step (c) above), it is often helpful to think of them separately. | |
105 | For example, RCU readers and updaters need not communicate at all, | |
106 | but RCU provides implicit low-overhead communication between readers | |
107 | and reclaimers, namely, in step (b) above. | |
108 | ||
109 | So how the heck can a reclaimer tell when a reader is done, given | |
110 | that readers are not doing any sort of synchronization operations??? | |
111 | Read on to learn about how RCU's API makes this easy. | |
112 | ||
113 | ||
114 | 2. WHAT IS RCU'S CORE API? | |
115 | ||
116 | The core RCU API is quite small: | |
117 | ||
118 | a. rcu_read_lock() | |
119 | b. rcu_read_unlock() | |
120 | c. synchronize_rcu() / call_rcu() | |
121 | d. rcu_assign_pointer() | |
122 | e. rcu_dereference() | |
123 | ||
124 | There are many other members of the RCU API, but the rest can be | |
125 | expressed in terms of these five, though most implementations instead | |
126 | express synchronize_rcu() in terms of the call_rcu() callback API. | |
127 | ||
128 | The five core RCU APIs are described below, the other 18 will be enumerated | |
129 | later. See the kernel docbook documentation for more info, or look directly | |
130 | at the function header comments. | |
131 | ||
132 | rcu_read_lock() | |
133 | ||
134 | void rcu_read_lock(void); | |
135 | ||
136 | Used by a reader to inform the reclaimer that the reader is | |
137 | entering an RCU read-side critical section. It is illegal | |
138 | to block while in an RCU read-side critical section, though | |
139 | kernels built with CONFIG_PREEMPT_RCU can preempt RCU read-side | |
140 | critical sections. Any RCU-protected data structure accessed | |
141 | during an RCU read-side critical section is guaranteed to remain | |
142 | unreclaimed for the full duration of that critical section. | |
143 | Reference counts may be used in conjunction with RCU to maintain | |
144 | longer-term references to data structures. | |
145 | ||
146 | rcu_read_unlock() | |
147 | ||
148 | void rcu_read_unlock(void); | |
149 | ||
150 | Used by a reader to inform the reclaimer that the reader is | |
151 | exiting an RCU read-side critical section. Note that RCU | |
152 | read-side critical sections may be nested and/or overlapping. | |
153 | ||
154 | synchronize_rcu() | |
155 | ||
156 | void synchronize_rcu(void); | |
157 | ||
158 | Marks the end of updater code and the beginning of reclaimer | |
159 | code. It does this by blocking until all pre-existing RCU | |
160 | read-side critical sections on all CPUs have completed. | |
161 | Note that synchronize_rcu() will -not- necessarily wait for | |
162 | any subsequent RCU read-side critical sections to complete. | |
163 | For example, consider the following sequence of events: | |
164 | ||
165 | CPU 0 CPU 1 CPU 2 | |
166 | ----------------- ------------------------- --------------- | |
167 | 1. rcu_read_lock() | |
168 | 2. enters synchronize_rcu() | |
169 | 3. rcu_read_lock() | |
170 | 4. rcu_read_unlock() | |
171 | 5. exits synchronize_rcu() | |
172 | 6. rcu_read_unlock() | |
173 | ||
174 | To reiterate, synchronize_rcu() waits only for ongoing RCU | |
175 | read-side critical sections to complete, not necessarily for | |
176 | any that begin after synchronize_rcu() is invoked. | |
177 | ||
178 | Of course, synchronize_rcu() does not necessarily return | |
179 | -immediately- after the last pre-existing RCU read-side critical | |
180 | section completes. For one thing, there might well be scheduling | |
181 | delays. For another thing, many RCU implementations process | |
182 | requests in batches in order to improve efficiencies, which can | |
183 | further delay synchronize_rcu(). | |
184 | ||
185 | Since synchronize_rcu() is the API that must figure out when | |
186 | readers are done, its implementation is key to RCU. For RCU | |
187 | to be useful in all but the most read-intensive situations, | |
188 | synchronize_rcu()'s overhead must also be quite small. | |
189 | ||
190 | The call_rcu() API is a callback form of synchronize_rcu(), | |
191 | and is described in more detail in a later section. Instead of | |
192 | blocking, it registers a function and argument which are invoked | |
193 | after all ongoing RCU read-side critical sections have completed. | |
194 | This callback variant is particularly useful in situations where | |
165d6c78 PM |
195 | it is illegal to block or where update-side performance is |
196 | critically important. | |
197 | ||
198 | However, the call_rcu() API should not be used lightly, as use | |
199 | of the synchronize_rcu() API generally results in simpler code. | |
200 | In addition, the synchronize_rcu() API has the nice property | |
201 | of automatically limiting update rate should grace periods | |
202 | be delayed. This property results in system resilience in face | |
203 | of denial-of-service attacks. Code using call_rcu() should limit | |
204 | update rate in order to gain this same sort of resilience. See | |
205 | checklist.txt for some approaches to limiting the update rate. | |
dd81eca8 PM |
206 | |
207 | rcu_assign_pointer() | |
208 | ||
209 | typeof(p) rcu_assign_pointer(p, typeof(p) v); | |
210 | ||
211 | Yes, rcu_assign_pointer() -is- implemented as a macro, though it | |
212 | would be cool to be able to declare a function in this manner. | |
213 | (Compiler experts will no doubt disagree.) | |
214 | ||
215 | The updater uses this function to assign a new value to an | |
216 | RCU-protected pointer, in order to safely communicate the change | |
217 | in value from the updater to the reader. This function returns | |
218 | the new value, and also executes any memory-barrier instructions | |
219 | required for a given CPU architecture. | |
220 | ||
d19720a9 PM |
221 | Perhaps just as important, it serves to document (1) which |
222 | pointers are protected by RCU and (2) the point at which a | |
223 | given structure becomes accessible to other CPUs. That said, | |
224 | rcu_assign_pointer() is most frequently used indirectly, via | |
225 | the _rcu list-manipulation primitives such as list_add_rcu(). | |
dd81eca8 PM |
226 | |
227 | rcu_dereference() | |
228 | ||
229 | typeof(p) rcu_dereference(p); | |
230 | ||
231 | Like rcu_assign_pointer(), rcu_dereference() must be implemented | |
232 | as a macro. | |
233 | ||
234 | The reader uses rcu_dereference() to fetch an RCU-protected | |
235 | pointer, which returns a value that may then be safely | |
236 | dereferenced. Note that rcu_deference() does not actually | |
237 | dereference the pointer, instead, it protects the pointer for | |
238 | later dereferencing. It also executes any needed memory-barrier | |
239 | instructions for a given CPU architecture. Currently, only Alpha | |
240 | needs memory barriers within rcu_dereference() -- on other CPUs, | |
241 | it compiles to nothing, not even a compiler directive. | |
242 | ||
243 | Common coding practice uses rcu_dereference() to copy an | |
244 | RCU-protected pointer to a local variable, then dereferences | |
245 | this local variable, for example as follows: | |
246 | ||
247 | p = rcu_dereference(head.next); | |
248 | return p->data; | |
249 | ||
250 | However, in this case, one could just as easily combine these | |
251 | into one statement: | |
252 | ||
253 | return rcu_dereference(head.next)->data; | |
254 | ||
255 | If you are going to be fetching multiple fields from the | |
256 | RCU-protected structure, using the local variable is of | |
257 | course preferred. Repeated rcu_dereference() calls look | |
258 | ugly and incur unnecessary overhead on Alpha CPUs. | |
259 | ||
260 | Note that the value returned by rcu_dereference() is valid | |
261 | only within the enclosing RCU read-side critical section. | |
262 | For example, the following is -not- legal: | |
263 | ||
264 | rcu_read_lock(); | |
265 | p = rcu_dereference(head.next); | |
266 | rcu_read_unlock(); | |
267 | x = p->address; | |
268 | rcu_read_lock(); | |
269 | y = p->data; | |
270 | rcu_read_unlock(); | |
271 | ||
272 | Holding a reference from one RCU read-side critical section | |
273 | to another is just as illegal as holding a reference from | |
274 | one lock-based critical section to another! Similarly, | |
275 | using a reference outside of the critical section in which | |
276 | it was acquired is just as illegal as doing so with normal | |
277 | locking. | |
278 | ||
279 | As with rcu_assign_pointer(), an important function of | |
d19720a9 PM |
280 | rcu_dereference() is to document which pointers are protected by |
281 | RCU, in particular, flagging a pointer that is subject to changing | |
282 | at any time, including immediately after the rcu_dereference(). | |
283 | And, again like rcu_assign_pointer(), rcu_dereference() is | |
284 | typically used indirectly, via the _rcu list-manipulation | |
dd81eca8 PM |
285 | primitives, such as list_for_each_entry_rcu(). |
286 | ||
287 | The following diagram shows how each API communicates among the | |
288 | reader, updater, and reclaimer. | |
289 | ||
290 | ||
291 | rcu_assign_pointer() | |
292 | +--------+ | |
293 | +---------------------->| reader |---------+ | |
294 | | +--------+ | | |
295 | | | | | |
296 | | | | Protect: | |
297 | | | | rcu_read_lock() | |
298 | | | | rcu_read_unlock() | |
299 | | rcu_dereference() | | | |
300 | +---------+ | | | |
301 | | updater |<---------------------+ | | |
302 | +---------+ V | |
303 | | +-----------+ | |
304 | +----------------------------------->| reclaimer | | |
305 | +-----------+ | |
306 | Defer: | |
307 | synchronize_rcu() & call_rcu() | |
308 | ||
309 | ||
310 | The RCU infrastructure observes the time sequence of rcu_read_lock(), | |
311 | rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in | |
312 | order to determine when (1) synchronize_rcu() invocations may return | |
313 | to their callers and (2) call_rcu() callbacks may be invoked. Efficient | |
314 | implementations of the RCU infrastructure make heavy use of batching in | |
315 | order to amortize their overhead over many uses of the corresponding APIs. | |
316 | ||
317 | There are no fewer than three RCU mechanisms in the Linux kernel; the | |
318 | diagram above shows the first one, which is by far the most commonly used. | |
319 | The rcu_dereference() and rcu_assign_pointer() primitives are used for | |
320 | all three mechanisms, but different defer and protect primitives are | |
321 | used as follows: | |
322 | ||
323 | Defer Protect | |
324 | ||
325 | a. synchronize_rcu() rcu_read_lock() / rcu_read_unlock() | |
326 | call_rcu() | |
327 | ||
328 | b. call_rcu_bh() rcu_read_lock_bh() / rcu_read_unlock_bh() | |
329 | ||
330 | c. synchronize_sched() preempt_disable() / preempt_enable() | |
331 | local_irq_save() / local_irq_restore() | |
332 | hardirq enter / hardirq exit | |
333 | NMI enter / NMI exit | |
334 | ||
335 | These three mechanisms are used as follows: | |
336 | ||
337 | a. RCU applied to normal data structures. | |
338 | ||
339 | b. RCU applied to networking data structures that may be subjected | |
340 | to remote denial-of-service attacks. | |
341 | ||
342 | c. RCU applied to scheduler and interrupt/NMI-handler tasks. | |
343 | ||
344 | Again, most uses will be of (a). The (b) and (c) cases are important | |
345 | for specialized uses, but are relatively uncommon. | |
346 | ||
347 | ||
348 | 3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API? | |
349 | ||
350 | This section shows a simple use of the core RCU API to protect a | |
d19720a9 | 351 | global pointer to a dynamically allocated structure. More-typical |
dd81eca8 PM |
352 | uses of RCU may be found in listRCU.txt, arrayRCU.txt, and NMI-RCU.txt. |
353 | ||
354 | struct foo { | |
355 | int a; | |
356 | char b; | |
357 | long c; | |
358 | }; | |
359 | DEFINE_SPINLOCK(foo_mutex); | |
360 | ||
361 | struct foo *gbl_foo; | |
362 | ||
363 | /* | |
364 | * Create a new struct foo that is the same as the one currently | |
365 | * pointed to by gbl_foo, except that field "a" is replaced | |
366 | * with "new_a". Points gbl_foo to the new structure, and | |
367 | * frees up the old structure after a grace period. | |
368 | * | |
369 | * Uses rcu_assign_pointer() to ensure that concurrent readers | |
370 | * see the initialized version of the new structure. | |
371 | * | |
372 | * Uses synchronize_rcu() to ensure that any readers that might | |
373 | * have references to the old structure complete before freeing | |
374 | * the old structure. | |
375 | */ | |
376 | void foo_update_a(int new_a) | |
377 | { | |
378 | struct foo *new_fp; | |
379 | struct foo *old_fp; | |
380 | ||
de0dfcdf | 381 | new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL); |
dd81eca8 PM |
382 | spin_lock(&foo_mutex); |
383 | old_fp = gbl_foo; | |
384 | *new_fp = *old_fp; | |
385 | new_fp->a = new_a; | |
386 | rcu_assign_pointer(gbl_foo, new_fp); | |
387 | spin_unlock(&foo_mutex); | |
388 | synchronize_rcu(); | |
389 | kfree(old_fp); | |
390 | } | |
391 | ||
392 | /* | |
393 | * Return the value of field "a" of the current gbl_foo | |
394 | * structure. Use rcu_read_lock() and rcu_read_unlock() | |
395 | * to ensure that the structure does not get deleted out | |
396 | * from under us, and use rcu_dereference() to ensure that | |
397 | * we see the initialized version of the structure (important | |
398 | * for DEC Alpha and for people reading the code). | |
399 | */ | |
400 | int foo_get_a(void) | |
401 | { | |
402 | int retval; | |
403 | ||
404 | rcu_read_lock(); | |
405 | retval = rcu_dereference(gbl_foo)->a; | |
406 | rcu_read_unlock(); | |
407 | return retval; | |
408 | } | |
409 | ||
410 | So, to sum up: | |
411 | ||
412 | o Use rcu_read_lock() and rcu_read_unlock() to guard RCU | |
413 | read-side critical sections. | |
414 | ||
415 | o Within an RCU read-side critical section, use rcu_dereference() | |
416 | to dereference RCU-protected pointers. | |
417 | ||
418 | o Use some solid scheme (such as locks or semaphores) to | |
419 | keep concurrent updates from interfering with each other. | |
420 | ||
421 | o Use rcu_assign_pointer() to update an RCU-protected pointer. | |
422 | This primitive protects concurrent readers from the updater, | |
423 | -not- concurrent updates from each other! You therefore still | |
424 | need to use locking (or something similar) to keep concurrent | |
425 | rcu_assign_pointer() primitives from interfering with each other. | |
426 | ||
427 | o Use synchronize_rcu() -after- removing a data element from an | |
428 | RCU-protected data structure, but -before- reclaiming/freeing | |
429 | the data element, in order to wait for the completion of all | |
430 | RCU read-side critical sections that might be referencing that | |
431 | data item. | |
432 | ||
433 | See checklist.txt for additional rules to follow when using RCU. | |
d19720a9 PM |
434 | And again, more-typical uses of RCU may be found in listRCU.txt, |
435 | arrayRCU.txt, and NMI-RCU.txt. | |
dd81eca8 PM |
436 | |
437 | ||
438 | 4. WHAT IF MY UPDATING THREAD CANNOT BLOCK? | |
439 | ||
440 | In the example above, foo_update_a() blocks until a grace period elapses. | |
441 | This is quite simple, but in some cases one cannot afford to wait so | |
442 | long -- there might be other high-priority work to be done. | |
443 | ||
444 | In such cases, one uses call_rcu() rather than synchronize_rcu(). | |
445 | The call_rcu() API is as follows: | |
446 | ||
447 | void call_rcu(struct rcu_head * head, | |
448 | void (*func)(struct rcu_head *head)); | |
449 | ||
450 | This function invokes func(head) after a grace period has elapsed. | |
451 | This invocation might happen from either softirq or process context, | |
452 | so the function is not permitted to block. The foo struct needs to | |
453 | have an rcu_head structure added, perhaps as follows: | |
454 | ||
455 | struct foo { | |
456 | int a; | |
457 | char b; | |
458 | long c; | |
459 | struct rcu_head rcu; | |
460 | }; | |
461 | ||
462 | The foo_update_a() function might then be written as follows: | |
463 | ||
464 | /* | |
465 | * Create a new struct foo that is the same as the one currently | |
466 | * pointed to by gbl_foo, except that field "a" is replaced | |
467 | * with "new_a". Points gbl_foo to the new structure, and | |
468 | * frees up the old structure after a grace period. | |
469 | * | |
470 | * Uses rcu_assign_pointer() to ensure that concurrent readers | |
471 | * see the initialized version of the new structure. | |
472 | * | |
473 | * Uses call_rcu() to ensure that any readers that might have | |
474 | * references to the old structure complete before freeing the | |
475 | * old structure. | |
476 | */ | |
477 | void foo_update_a(int new_a) | |
478 | { | |
479 | struct foo *new_fp; | |
480 | struct foo *old_fp; | |
481 | ||
de0dfcdf | 482 | new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL); |
dd81eca8 PM |
483 | spin_lock(&foo_mutex); |
484 | old_fp = gbl_foo; | |
485 | *new_fp = *old_fp; | |
486 | new_fp->a = new_a; | |
487 | rcu_assign_pointer(gbl_foo, new_fp); | |
488 | spin_unlock(&foo_mutex); | |
489 | call_rcu(&old_fp->rcu, foo_reclaim); | |
490 | } | |
491 | ||
492 | The foo_reclaim() function might appear as follows: | |
493 | ||
494 | void foo_reclaim(struct rcu_head *rp) | |
495 | { | |
496 | struct foo *fp = container_of(rp, struct foo, rcu); | |
497 | ||
498 | kfree(fp); | |
499 | } | |
500 | ||
501 | The container_of() primitive is a macro that, given a pointer into a | |
502 | struct, the type of the struct, and the pointed-to field within the | |
503 | struct, returns a pointer to the beginning of the struct. | |
504 | ||
505 | The use of call_rcu() permits the caller of foo_update_a() to | |
506 | immediately regain control, without needing to worry further about the | |
507 | old version of the newly updated element. It also clearly shows the | |
508 | RCU distinction between updater, namely foo_update_a(), and reclaimer, | |
509 | namely foo_reclaim(). | |
510 | ||
511 | The summary of advice is the same as for the previous section, except | |
512 | that we are now using call_rcu() rather than synchronize_rcu(): | |
513 | ||
514 | o Use call_rcu() -after- removing a data element from an | |
515 | RCU-protected data structure in order to register a callback | |
516 | function that will be invoked after the completion of all RCU | |
517 | read-side critical sections that might be referencing that | |
518 | data item. | |
519 | ||
520 | Again, see checklist.txt for additional rules governing the use of RCU. | |
521 | ||
522 | ||
523 | 5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU? | |
524 | ||
525 | One of the nice things about RCU is that it has extremely simple "toy" | |
526 | implementations that are a good first step towards understanding the | |
527 | production-quality implementations in the Linux kernel. This section | |
528 | presents two such "toy" implementations of RCU, one that is implemented | |
529 | in terms of familiar locking primitives, and another that more closely | |
530 | resembles "classic" RCU. Both are way too simple for real-world use, | |
531 | lacking both functionality and performance. However, they are useful | |
532 | in getting a feel for how RCU works. See kernel/rcupdate.c for a | |
533 | production-quality implementation, and see: | |
534 | ||
535 | http://www.rdrop.com/users/paulmck/RCU | |
536 | ||
537 | for papers describing the Linux kernel RCU implementation. The OLS'01 | |
538 | and OLS'02 papers are a good introduction, and the dissertation provides | |
d19720a9 | 539 | more details on the current implementation as of early 2004. |
dd81eca8 PM |
540 | |
541 | ||
542 | 5A. "TOY" IMPLEMENTATION #1: LOCKING | |
543 | ||
544 | This section presents a "toy" RCU implementation that is based on | |
545 | familiar locking primitives. Its overhead makes it a non-starter for | |
546 | real-life use, as does its lack of scalability. It is also unsuitable | |
547 | for realtime use, since it allows scheduling latency to "bleed" from | |
548 | one read-side critical section to another. | |
549 | ||
550 | However, it is probably the easiest implementation to relate to, so is | |
551 | a good starting point. | |
552 | ||
553 | It is extremely simple: | |
554 | ||
555 | static DEFINE_RWLOCK(rcu_gp_mutex); | |
556 | ||
557 | void rcu_read_lock(void) | |
558 | { | |
559 | read_lock(&rcu_gp_mutex); | |
560 | } | |
561 | ||
562 | void rcu_read_unlock(void) | |
563 | { | |
564 | read_unlock(&rcu_gp_mutex); | |
565 | } | |
566 | ||
567 | void synchronize_rcu(void) | |
568 | { | |
569 | write_lock(&rcu_gp_mutex); | |
570 | write_unlock(&rcu_gp_mutex); | |
571 | } | |
572 | ||
573 | [You can ignore rcu_assign_pointer() and rcu_dereference() without | |
574 | missing much. But here they are anyway. And whatever you do, don't | |
575 | forget about them when submitting patches making use of RCU!] | |
576 | ||
577 | #define rcu_assign_pointer(p, v) ({ \ | |
578 | smp_wmb(); \ | |
579 | (p) = (v); \ | |
580 | }) | |
581 | ||
582 | #define rcu_dereference(p) ({ \ | |
583 | typeof(p) _________p1 = p; \ | |
584 | smp_read_barrier_depends(); \ | |
585 | (_________p1); \ | |
586 | }) | |
587 | ||
588 | ||
589 | The rcu_read_lock() and rcu_read_unlock() primitive read-acquire | |
590 | and release a global reader-writer lock. The synchronize_rcu() | |
591 | primitive write-acquires this same lock, then immediately releases | |
592 | it. This means that once synchronize_rcu() exits, all RCU read-side | |
53cb4726 | 593 | critical sections that were in progress before synchronize_rcu() was |
dd81eca8 PM |
594 | called are guaranteed to have completed -- there is no way that |
595 | synchronize_rcu() would have been able to write-acquire the lock | |
596 | otherwise. | |
597 | ||
598 | It is possible to nest rcu_read_lock(), since reader-writer locks may | |
599 | be recursively acquired. Note also that rcu_read_lock() is immune | |
600 | from deadlock (an important property of RCU). The reason for this is | |
601 | that the only thing that can block rcu_read_lock() is a synchronize_rcu(). | |
602 | But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex, | |
603 | so there can be no deadlock cycle. | |
604 | ||
605 | Quick Quiz #1: Why is this argument naive? How could a deadlock | |
606 | occur when using this algorithm in a real-world Linux | |
607 | kernel? How could this deadlock be avoided? | |
608 | ||
609 | ||
610 | 5B. "TOY" EXAMPLE #2: CLASSIC RCU | |
611 | ||
612 | This section presents a "toy" RCU implementation that is based on | |
613 | "classic RCU". It is also short on performance (but only for updates) and | |
614 | on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT | |
615 | kernels. The definitions of rcu_dereference() and rcu_assign_pointer() | |
616 | are the same as those shown in the preceding section, so they are omitted. | |
617 | ||
618 | void rcu_read_lock(void) { } | |
619 | ||
620 | void rcu_read_unlock(void) { } | |
621 | ||
622 | void synchronize_rcu(void) | |
623 | { | |
624 | int cpu; | |
625 | ||
3c30a752 | 626 | for_each_possible_cpu(cpu) |
dd81eca8 PM |
627 | run_on(cpu); |
628 | } | |
629 | ||
630 | Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing. | |
631 | This is the great strength of classic RCU in a non-preemptive kernel: | |
632 | read-side overhead is precisely zero, at least on non-Alpha CPUs. | |
633 | And there is absolutely no way that rcu_read_lock() can possibly | |
634 | participate in a deadlock cycle! | |
635 | ||
636 | The implementation of synchronize_rcu() simply schedules itself on each | |
637 | CPU in turn. The run_on() primitive can be implemented straightforwardly | |
638 | in terms of the sched_setaffinity() primitive. Of course, a somewhat less | |
639 | "toy" implementation would restore the affinity upon completion rather | |
640 | than just leaving all tasks running on the last CPU, but when I said | |
641 | "toy", I meant -toy-! | |
642 | ||
643 | So how the heck is this supposed to work??? | |
644 | ||
645 | Remember that it is illegal to block while in an RCU read-side critical | |
646 | section. Therefore, if a given CPU executes a context switch, we know | |
647 | that it must have completed all preceding RCU read-side critical sections. | |
648 | Once -all- CPUs have executed a context switch, then -all- preceding | |
649 | RCU read-side critical sections will have completed. | |
650 | ||
651 | So, suppose that we remove a data item from its structure and then invoke | |
652 | synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed | |
653 | that there are no RCU read-side critical sections holding a reference | |
654 | to that data item, so we can safely reclaim it. | |
655 | ||
656 | Quick Quiz #2: Give an example where Classic RCU's read-side | |
657 | overhead is -negative-. | |
658 | ||
659 | Quick Quiz #3: If it is illegal to block in an RCU read-side | |
660 | critical section, what the heck do you do in | |
661 | PREEMPT_RT, where normal spinlocks can block??? | |
662 | ||
663 | ||
664 | 6. ANALOGY WITH READER-WRITER LOCKING | |
665 | ||
666 | Although RCU can be used in many different ways, a very common use of | |
667 | RCU is analogous to reader-writer locking. The following unified | |
668 | diff shows how closely related RCU and reader-writer locking can be. | |
669 | ||
670 | @@ -13,15 +14,15 @@ | |
671 | struct list_head *lp; | |
672 | struct el *p; | |
673 | ||
674 | - read_lock(); | |
675 | - list_for_each_entry(p, head, lp) { | |
676 | + rcu_read_lock(); | |
677 | + list_for_each_entry_rcu(p, head, lp) { | |
678 | if (p->key == key) { | |
679 | *result = p->data; | |
680 | - read_unlock(); | |
681 | + rcu_read_unlock(); | |
682 | return 1; | |
683 | } | |
684 | } | |
685 | - read_unlock(); | |
686 | + rcu_read_unlock(); | |
687 | return 0; | |
688 | } | |
689 | ||
690 | @@ -29,15 +30,16 @@ | |
691 | { | |
692 | struct el *p; | |
693 | ||
694 | - write_lock(&listmutex); | |
695 | + spin_lock(&listmutex); | |
696 | list_for_each_entry(p, head, lp) { | |
697 | if (p->key == key) { | |
82a854ec | 698 | - list_del(&p->list); |
dd81eca8 | 699 | - write_unlock(&listmutex); |
82a854ec | 700 | + list_del_rcu(&p->list); |
dd81eca8 PM |
701 | + spin_unlock(&listmutex); |
702 | + synchronize_rcu(); | |
703 | kfree(p); | |
704 | return 1; | |
705 | } | |
706 | } | |
707 | - write_unlock(&listmutex); | |
708 | + spin_unlock(&listmutex); | |
709 | return 0; | |
710 | } | |
711 | ||
712 | Or, for those who prefer a side-by-side listing: | |
713 | ||
714 | 1 struct el { 1 struct el { | |
715 | 2 struct list_head list; 2 struct list_head list; | |
716 | 3 long key; 3 long key; | |
717 | 4 spinlock_t mutex; 4 spinlock_t mutex; | |
718 | 5 int data; 5 int data; | |
719 | 6 /* Other data fields */ 6 /* Other data fields */ | |
720 | 7 }; 7 }; | |
721 | 8 spinlock_t listmutex; 8 spinlock_t listmutex; | |
722 | 9 struct el head; 9 struct el head; | |
723 | ||
724 | 1 int search(long key, int *result) 1 int search(long key, int *result) | |
725 | 2 { 2 { | |
726 | 3 struct list_head *lp; 3 struct list_head *lp; | |
727 | 4 struct el *p; 4 struct el *p; | |
728 | 5 5 | |
729 | 6 read_lock(); 6 rcu_read_lock(); | |
730 | 7 list_for_each_entry(p, head, lp) { 7 list_for_each_entry_rcu(p, head, lp) { | |
731 | 8 if (p->key == key) { 8 if (p->key == key) { | |
732 | 9 *result = p->data; 9 *result = p->data; | |
733 | 10 read_unlock(); 10 rcu_read_unlock(); | |
734 | 11 return 1; 11 return 1; | |
735 | 12 } 12 } | |
736 | 13 } 13 } | |
737 | 14 read_unlock(); 14 rcu_read_unlock(); | |
738 | 15 return 0; 15 return 0; | |
739 | 16 } 16 } | |
740 | ||
741 | 1 int delete(long key) 1 int delete(long key) | |
742 | 2 { 2 { | |
743 | 3 struct el *p; 3 struct el *p; | |
744 | 4 4 | |
745 | 5 write_lock(&listmutex); 5 spin_lock(&listmutex); | |
746 | 6 list_for_each_entry(p, head, lp) { 6 list_for_each_entry(p, head, lp) { | |
747 | 7 if (p->key == key) { 7 if (p->key == key) { | |
82a854ec | 748 | 8 list_del(&p->list); 8 list_del_rcu(&p->list); |
dd81eca8 PM |
749 | 9 write_unlock(&listmutex); 9 spin_unlock(&listmutex); |
750 | 10 synchronize_rcu(); | |
751 | 10 kfree(p); 11 kfree(p); | |
752 | 11 return 1; 12 return 1; | |
753 | 12 } 13 } | |
754 | 13 } 14 } | |
755 | 14 write_unlock(&listmutex); 15 spin_unlock(&listmutex); | |
756 | 15 return 0; 16 return 0; | |
757 | 16 } 17 } | |
758 | ||
759 | Either way, the differences are quite small. Read-side locking moves | |
760 | to rcu_read_lock() and rcu_read_unlock, update-side locking moves from | |
670e9f34 | 761 | a reader-writer lock to a simple spinlock, and a synchronize_rcu() |
dd81eca8 PM |
762 | precedes the kfree(). |
763 | ||
764 | However, there is one potential catch: the read-side and update-side | |
765 | critical sections can now run concurrently. In many cases, this will | |
766 | not be a problem, but it is necessary to check carefully regardless. | |
767 | For example, if multiple independent list updates must be seen as | |
768 | a single atomic update, converting to RCU will require special care. | |
769 | ||
770 | Also, the presence of synchronize_rcu() means that the RCU version of | |
771 | delete() can now block. If this is a problem, there is a callback-based | |
772 | mechanism that never blocks, namely call_rcu(), that can be used in | |
773 | place of synchronize_rcu(). | |
774 | ||
775 | ||
776 | 7. FULL LIST OF RCU APIs | |
777 | ||
778 | The RCU APIs are documented in docbook-format header comments in the | |
779 | Linux-kernel source code, but it helps to have a full list of the | |
780 | APIs, since there does not appear to be a way to categorize them | |
781 | in docbook. Here is the list, by category. | |
782 | ||
dd81eca8 PM |
783 | RCU pointer/list traversal: |
784 | ||
785 | rcu_dereference | |
32300751 PM |
786 | list_for_each_entry_rcu |
787 | hlist_for_each_entry_rcu | |
788 | ||
dd81eca8 PM |
789 | list_for_each_continue_rcu (to be deprecated in favor of new |
790 | list_for_each_entry_continue_rcu) | |
dd81eca8 | 791 | |
32300751 | 792 | RCU pointer/list update: |
dd81eca8 PM |
793 | |
794 | rcu_assign_pointer | |
795 | list_add_rcu | |
796 | list_add_tail_rcu | |
797 | list_del_rcu | |
798 | list_replace_rcu | |
799 | hlist_del_rcu | |
32300751 PM |
800 | hlist_add_after_rcu |
801 | hlist_add_before_rcu | |
dd81eca8 | 802 | hlist_add_head_rcu |
32300751 PM |
803 | hlist_replace_rcu |
804 | list_splice_init_rcu() | |
dd81eca8 | 805 | |
32300751 PM |
806 | RCU: Critical sections Grace period Barrier |
807 | ||
808 | rcu_read_lock synchronize_net rcu_barrier | |
809 | rcu_read_unlock synchronize_rcu | |
810 | call_rcu | |
811 | ||
812 | ||
813 | bh: Critical sections Grace period Barrier | |
814 | ||
815 | rcu_read_lock_bh call_rcu_bh rcu_barrier_bh | |
816 | rcu_read_unlock_bh | |
817 | ||
818 | ||
819 | sched: Critical sections Grace period Barrier | |
820 | ||
821 | [preempt_disable] synchronize_sched rcu_barrier_sched | |
822 | [and friends] call_rcu_sched | |
823 | ||
824 | ||
825 | SRCU: Critical sections Grace period Barrier | |
826 | ||
827 | srcu_read_lock synchronize_srcu N/A | |
828 | srcu_read_unlock | |
dd81eca8 | 829 | |
dd81eca8 PM |
830 | |
831 | See the comment headers in the source code (or the docbook generated | |
832 | from them) for more information. | |
833 | ||
834 | ||
835 | 8. ANSWERS TO QUICK QUIZZES | |
836 | ||
837 | Quick Quiz #1: Why is this argument naive? How could a deadlock | |
838 | occur when using this algorithm in a real-world Linux | |
839 | kernel? [Referring to the lock-based "toy" RCU | |
840 | algorithm.] | |
841 | ||
842 | Answer: Consider the following sequence of events: | |
843 | ||
844 | 1. CPU 0 acquires some unrelated lock, call it | |
d19720a9 PM |
845 | "problematic_lock", disabling irq via |
846 | spin_lock_irqsave(). | |
dd81eca8 PM |
847 | |
848 | 2. CPU 1 enters synchronize_rcu(), write-acquiring | |
849 | rcu_gp_mutex. | |
850 | ||
851 | 3. CPU 0 enters rcu_read_lock(), but must wait | |
852 | because CPU 1 holds rcu_gp_mutex. | |
853 | ||
854 | 4. CPU 1 is interrupted, and the irq handler | |
855 | attempts to acquire problematic_lock. | |
856 | ||
857 | The system is now deadlocked. | |
858 | ||
859 | One way to avoid this deadlock is to use an approach like | |
860 | that of CONFIG_PREEMPT_RT, where all normal spinlocks | |
861 | become blocking locks, and all irq handlers execute in | |
862 | the context of special tasks. In this case, in step 4 | |
863 | above, the irq handler would block, allowing CPU 1 to | |
864 | release rcu_gp_mutex, avoiding the deadlock. | |
865 | ||
866 | Even in the absence of deadlock, this RCU implementation | |
867 | allows latency to "bleed" from readers to other | |
868 | readers through synchronize_rcu(). To see this, | |
869 | consider task A in an RCU read-side critical section | |
870 | (thus read-holding rcu_gp_mutex), task B blocked | |
871 | attempting to write-acquire rcu_gp_mutex, and | |
872 | task C blocked in rcu_read_lock() attempting to | |
873 | read_acquire rcu_gp_mutex. Task A's RCU read-side | |
874 | latency is holding up task C, albeit indirectly via | |
875 | task B. | |
876 | ||
877 | Realtime RCU implementations therefore use a counter-based | |
878 | approach where tasks in RCU read-side critical sections | |
879 | cannot be blocked by tasks executing synchronize_rcu(). | |
880 | ||
881 | Quick Quiz #2: Give an example where Classic RCU's read-side | |
882 | overhead is -negative-. | |
883 | ||
884 | Answer: Imagine a single-CPU system with a non-CONFIG_PREEMPT | |
885 | kernel where a routing table is used by process-context | |
886 | code, but can be updated by irq-context code (for example, | |
887 | by an "ICMP REDIRECT" packet). The usual way of handling | |
888 | this would be to have the process-context code disable | |
889 | interrupts while searching the routing table. Use of | |
890 | RCU allows such interrupt-disabling to be dispensed with. | |
891 | Thus, without RCU, you pay the cost of disabling interrupts, | |
892 | and with RCU you don't. | |
893 | ||
894 | One can argue that the overhead of RCU in this | |
895 | case is negative with respect to the single-CPU | |
896 | interrupt-disabling approach. Others might argue that | |
897 | the overhead of RCU is merely zero, and that replacing | |
898 | the positive overhead of the interrupt-disabling scheme | |
899 | with the zero-overhead RCU scheme does not constitute | |
900 | negative overhead. | |
901 | ||
902 | In real life, of course, things are more complex. But | |
903 | even the theoretical possibility of negative overhead for | |
904 | a synchronization primitive is a bit unexpected. ;-) | |
905 | ||
906 | Quick Quiz #3: If it is illegal to block in an RCU read-side | |
907 | critical section, what the heck do you do in | |
908 | PREEMPT_RT, where normal spinlocks can block??? | |
909 | ||
910 | Answer: Just as PREEMPT_RT permits preemption of spinlock | |
911 | critical sections, it permits preemption of RCU | |
912 | read-side critical sections. It also permits | |
913 | spinlocks blocking while in RCU read-side critical | |
914 | sections. | |
915 | ||
916 | Why the apparent inconsistency? Because it is it | |
917 | possible to use priority boosting to keep the RCU | |
918 | grace periods short if need be (for example, if running | |
919 | short of memory). In contrast, if blocking waiting | |
920 | for (say) network reception, there is no way to know | |
921 | what should be boosted. Especially given that the | |
922 | process we need to boost might well be a human being | |
923 | who just went out for a pizza or something. And although | |
924 | a computer-operated cattle prod might arouse serious | |
925 | interest, it might also provoke serious objections. | |
926 | Besides, how does the computer know what pizza parlor | |
927 | the human being went to??? | |
928 | ||
929 | ||
930 | ACKNOWLEDGEMENTS | |
931 | ||
932 | My thanks to the people who helped make this human-readable, including | |
d19720a9 | 933 | Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern. |
dd81eca8 PM |
934 | |
935 | ||
936 | For more information, see http://www.rdrop.com/users/paulmck/RCU. |