whatisRCU.txt 38 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031
  1. Please note that the "What is RCU?" LWN series is an excellent place
  2. to start learning about RCU:
  3. 1. What is RCU, Fundamentally? http://lwn.net/Articles/262464/
  4. 2. What is RCU? Part 2: Usage http://lwn.net/Articles/263130/
  5. 3. RCU part 3: the RCU API http://lwn.net/Articles/264090/
  6. 4. The RCU API, 2010 Edition http://lwn.net/Articles/418853/
  7. What is RCU?
  8. RCU is a synchronization mechanism that was added to the Linux kernel
  9. during the 2.5 development effort that is optimized for read-mostly
  10. situations. Although RCU is actually quite simple once you understand it,
  11. getting there can sometimes be a challenge. Part of the problem is that
  12. most of the past descriptions of RCU have been written with the mistaken
  13. assumption that there is "one true way" to describe RCU. Instead,
  14. the experience has been that different people must take different paths
  15. to arrive at an understanding of RCU. This document provides several
  16. different paths, as follows:
  17. 1. RCU OVERVIEW
  18. 2. WHAT IS RCU'S CORE API?
  19. 3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
  20. 4. WHAT IF MY UPDATING THREAD CANNOT BLOCK?
  21. 5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
  22. 6. ANALOGY WITH READER-WRITER LOCKING
  23. 7. FULL LIST OF RCU APIs
  24. 8. ANSWERS TO QUICK QUIZZES
  25. People who prefer starting with a conceptual overview should focus on
  26. Section 1, though most readers will profit by reading this section at
  27. some point. People who prefer to start with an API that they can then
  28. experiment with should focus on Section 2. People who prefer to start
  29. with example uses should focus on Sections 3 and 4. People who need to
  30. understand the RCU implementation should focus on Section 5, then dive
  31. into the kernel source code. People who reason best by analogy should
  32. focus on Section 6. Section 7 serves as an index to the docbook API
  33. documentation, and Section 8 is the traditional answer key.
  34. So, start with the section that makes the most sense to you and your
  35. preferred method of learning. If you need to know everything about
  36. everything, feel free to read the whole thing -- but if you are really
  37. that type of person, you have perused the source code and will therefore
  38. never need this document anyway. ;-)
  39. 1. RCU OVERVIEW
  40. The basic idea behind RCU is to split updates into "removal" and
  41. "reclamation" phases. The removal phase removes references to data items
  42. within a data structure (possibly by replacing them with references to
  43. new versions of these data items), and can run concurrently with readers.
  44. The reason that it is safe to run the removal phase concurrently with
  45. readers is the semantics of modern CPUs guarantee that readers will see
  46. either the old or the new version of the data structure rather than a
  47. partially updated reference. The reclamation phase does the work of reclaiming
  48. (e.g., freeing) the data items removed from the data structure during the
  49. removal phase. Because reclaiming data items can disrupt any readers
  50. concurrently referencing those data items, the reclamation phase must
  51. not start until readers no longer hold references to those data items.
  52. Splitting the update into removal and reclamation phases permits the
  53. updater to perform the removal phase immediately, and to defer the
  54. reclamation phase until all readers active during the removal phase have
  55. completed, either by blocking until they finish or by registering a
  56. callback that is invoked after they finish. Only readers that are active
  57. during the removal phase need be considered, because any reader starting
  58. after the removal phase will be unable to gain a reference to the removed
  59. data items, and therefore cannot be disrupted by the reclamation phase.
  60. So the typical RCU update sequence goes something like the following:
  61. a. Remove pointers to a data structure, so that subsequent
  62. readers cannot gain a reference to it.
  63. b. Wait for all previous readers to complete their RCU read-side
  64. critical sections.
  65. c. At this point, there cannot be any readers who hold references
  66. to the data structure, so it now may safely be reclaimed
  67. (e.g., kfree()d).
  68. Step (b) above is the key idea underlying RCU's deferred destruction.
  69. The ability to wait until all readers are done allows RCU readers to
  70. use much lighter-weight synchronization, in some cases, absolutely no
  71. synchronization at all. In contrast, in more conventional lock-based
  72. schemes, readers must use heavy-weight synchronization in order to
  73. prevent an updater from deleting the data structure out from under them.
  74. This is because lock-based updaters typically update data items in place,
  75. and must therefore exclude readers. In contrast, RCU-based updaters
  76. typically take advantage of the fact that writes to single aligned
  77. pointers are atomic on modern CPUs, allowing atomic insertion, removal,
  78. and replacement of data items in a linked structure without disrupting
  79. readers. Concurrent RCU readers can then continue accessing the old
  80. versions, and can dispense with the atomic operations, memory barriers,
  81. and communications cache misses that are so expensive on present-day
  82. SMP computer systems, even in absence of lock contention.
  83. In the three-step procedure shown above, the updater is performing both
  84. the removal and the reclamation step, but it is often helpful for an
  85. entirely different thread to do the reclamation, as is in fact the case
  86. in the Linux kernel's directory-entry cache (dcache). Even if the same
  87. thread performs both the update step (step (a) above) and the reclamation
  88. step (step (c) above), it is often helpful to think of them separately.
  89. For example, RCU readers and updaters need not communicate at all,
  90. but RCU provides implicit low-overhead communication between readers
  91. and reclaimers, namely, in step (b) above.
  92. So how the heck can a reclaimer tell when a reader is done, given
  93. that readers are not doing any sort of synchronization operations???
  94. Read on to learn about how RCU's API makes this easy.
  95. 2. WHAT IS RCU'S CORE API?
  96. The core RCU API is quite small:
  97. a. rcu_read_lock()
  98. b. rcu_read_unlock()
  99. c. synchronize_rcu() / call_rcu()
  100. d. rcu_assign_pointer()
  101. e. rcu_dereference()
  102. There are many other members of the RCU API, but the rest can be
  103. expressed in terms of these five, though most implementations instead
  104. express synchronize_rcu() in terms of the call_rcu() callback API.
  105. The five core RCU APIs are described below, the other 18 will be enumerated
  106. later. See the kernel docbook documentation for more info, or look directly
  107. at the function header comments.
  108. rcu_read_lock()
  109. void rcu_read_lock(void);
  110. Used by a reader to inform the reclaimer that the reader is
  111. entering an RCU read-side critical section. It is illegal
  112. to block while in an RCU read-side critical section, though
  113. kernels built with CONFIG_PREEMPT_RCU can preempt RCU
  114. read-side critical sections. Any RCU-protected data structure
  115. accessed during an RCU read-side critical section is guaranteed to
  116. remain unreclaimed for the full duration of that critical section.
  117. Reference counts may be used in conjunction with RCU to maintain
  118. longer-term references to data structures.
  119. rcu_read_unlock()
  120. void rcu_read_unlock(void);
  121. Used by a reader to inform the reclaimer that the reader is
  122. exiting an RCU read-side critical section. Note that RCU
  123. read-side critical sections may be nested and/or overlapping.
  124. synchronize_rcu()
  125. void synchronize_rcu(void);
  126. Marks the end of updater code and the beginning of reclaimer
  127. code. It does this by blocking until all pre-existing RCU
  128. read-side critical sections on all CPUs have completed.
  129. Note that synchronize_rcu() will -not- necessarily wait for
  130. any subsequent RCU read-side critical sections to complete.
  131. For example, consider the following sequence of events:
  132. CPU 0 CPU 1 CPU 2
  133. ----------------- ------------------------- ---------------
  134. 1. rcu_read_lock()
  135. 2. enters synchronize_rcu()
  136. 3. rcu_read_lock()
  137. 4. rcu_read_unlock()
  138. 5. exits synchronize_rcu()
  139. 6. rcu_read_unlock()
  140. To reiterate, synchronize_rcu() waits only for ongoing RCU
  141. read-side critical sections to complete, not necessarily for
  142. any that begin after synchronize_rcu() is invoked.
  143. Of course, synchronize_rcu() does not necessarily return
  144. -immediately- after the last pre-existing RCU read-side critical
  145. section completes. For one thing, there might well be scheduling
  146. delays. For another thing, many RCU implementations process
  147. requests in batches in order to improve efficiencies, which can
  148. further delay synchronize_rcu().
  149. Since synchronize_rcu() is the API that must figure out when
  150. readers are done, its implementation is key to RCU. For RCU
  151. to be useful in all but the most read-intensive situations,
  152. synchronize_rcu()'s overhead must also be quite small.
  153. The call_rcu() API is a callback form of synchronize_rcu(),
  154. and is described in more detail in a later section. Instead of
  155. blocking, it registers a function and argument which are invoked
  156. after all ongoing RCU read-side critical sections have completed.
  157. This callback variant is particularly useful in situations where
  158. it is illegal to block or where update-side performance is
  159. critically important.
  160. However, the call_rcu() API should not be used lightly, as use
  161. of the synchronize_rcu() API generally results in simpler code.
  162. In addition, the synchronize_rcu() API has the nice property
  163. of automatically limiting update rate should grace periods
  164. be delayed. This property results in system resilience in face
  165. of denial-of-service attacks. Code using call_rcu() should limit
  166. update rate in order to gain this same sort of resilience. See
  167. checklist.txt for some approaches to limiting the update rate.
  168. rcu_assign_pointer()
  169. typeof(p) rcu_assign_pointer(p, typeof(p) v);
  170. Yes, rcu_assign_pointer() -is- implemented as a macro, though it
  171. would be cool to be able to declare a function in this manner.
  172. (Compiler experts will no doubt disagree.)
  173. The updater uses this function to assign a new value to an
  174. RCU-protected pointer, in order to safely communicate the change
  175. in value from the updater to the reader. This function returns
  176. the new value, and also executes any memory-barrier instructions
  177. required for a given CPU architecture.
  178. Perhaps just as important, it serves to document (1) which
  179. pointers are protected by RCU and (2) the point at which a
  180. given structure becomes accessible to other CPUs. That said,
  181. rcu_assign_pointer() is most frequently used indirectly, via
  182. the _rcu list-manipulation primitives such as list_add_rcu().
  183. rcu_dereference()
  184. typeof(p) rcu_dereference(p);
  185. Like rcu_assign_pointer(), rcu_dereference() must be implemented
  186. as a macro.
  187. The reader uses rcu_dereference() to fetch an RCU-protected
  188. pointer, which returns a value that may then be safely
  189. dereferenced. Note that rcu_deference() does not actually
  190. dereference the pointer, instead, it protects the pointer for
  191. later dereferencing. It also executes any needed memory-barrier
  192. instructions for a given CPU architecture. Currently, only Alpha
  193. needs memory barriers within rcu_dereference() -- on other CPUs,
  194. it compiles to nothing, not even a compiler directive.
  195. Common coding practice uses rcu_dereference() to copy an
  196. RCU-protected pointer to a local variable, then dereferences
  197. this local variable, for example as follows:
  198. p = rcu_dereference(head.next);
  199. return p->data;
  200. However, in this case, one could just as easily combine these
  201. into one statement:
  202. return rcu_dereference(head.next)->data;
  203. If you are going to be fetching multiple fields from the
  204. RCU-protected structure, using the local variable is of
  205. course preferred. Repeated rcu_dereference() calls look
  206. ugly, do not guarantee that the same pointer will be returned
  207. if an update happened while in the critical section, and incur
  208. unnecessary overhead on Alpha CPUs.
  209. Note that the value returned by rcu_dereference() is valid
  210. only within the enclosing RCU read-side critical section.
  211. For example, the following is -not- legal:
  212. rcu_read_lock();
  213. p = rcu_dereference(head.next);
  214. rcu_read_unlock();
  215. x = p->address; /* BUG!!! */
  216. rcu_read_lock();
  217. y = p->data; /* BUG!!! */
  218. rcu_read_unlock();
  219. Holding a reference from one RCU read-side critical section
  220. to another is just as illegal as holding a reference from
  221. one lock-based critical section to another! Similarly,
  222. using a reference outside of the critical section in which
  223. it was acquired is just as illegal as doing so with normal
  224. locking.
  225. As with rcu_assign_pointer(), an important function of
  226. rcu_dereference() is to document which pointers are protected by
  227. RCU, in particular, flagging a pointer that is subject to changing
  228. at any time, including immediately after the rcu_dereference().
  229. And, again like rcu_assign_pointer(), rcu_dereference() is
  230. typically used indirectly, via the _rcu list-manipulation
  231. primitives, such as list_for_each_entry_rcu().
  232. The following diagram shows how each API communicates among the
  233. reader, updater, and reclaimer.
  234. rcu_assign_pointer()
  235. +--------+
  236. +---------------------->| reader |---------+
  237. | +--------+ |
  238. | | |
  239. | | | Protect:
  240. | | | rcu_read_lock()
  241. | | | rcu_read_unlock()
  242. | rcu_dereference() | |
  243. +---------+ | |
  244. | updater |<---------------------+ |
  245. +---------+ V
  246. | +-----------+
  247. +----------------------------------->| reclaimer |
  248. +-----------+
  249. Defer:
  250. synchronize_rcu() & call_rcu()
  251. The RCU infrastructure observes the time sequence of rcu_read_lock(),
  252. rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in
  253. order to determine when (1) synchronize_rcu() invocations may return
  254. to their callers and (2) call_rcu() callbacks may be invoked. Efficient
  255. implementations of the RCU infrastructure make heavy use of batching in
  256. order to amortize their overhead over many uses of the corresponding APIs.
  257. There are no fewer than three RCU mechanisms in the Linux kernel; the
  258. diagram above shows the first one, which is by far the most commonly used.
  259. The rcu_dereference() and rcu_assign_pointer() primitives are used for
  260. all three mechanisms, but different defer and protect primitives are
  261. used as follows:
  262. Defer Protect
  263. a. synchronize_rcu() rcu_read_lock() / rcu_read_unlock()
  264. call_rcu() rcu_dereference()
  265. b. synchronize_rcu_bh() rcu_read_lock_bh() / rcu_read_unlock_bh()
  266. call_rcu_bh() rcu_dereference_bh()
  267. c. synchronize_sched() rcu_read_lock_sched() / rcu_read_unlock_sched()
  268. call_rcu_sched() preempt_disable() / preempt_enable()
  269. local_irq_save() / local_irq_restore()
  270. hardirq enter / hardirq exit
  271. NMI enter / NMI exit
  272. rcu_dereference_sched()
  273. These three mechanisms are used as follows:
  274. a. RCU applied to normal data structures.
  275. b. RCU applied to networking data structures that may be subjected
  276. to remote denial-of-service attacks.
  277. c. RCU applied to scheduler and interrupt/NMI-handler tasks.
  278. Again, most uses will be of (a). The (b) and (c) cases are important
  279. for specialized uses, but are relatively uncommon.
  280. 3. WHAT ARE SOME EXAMPLE USES OF CORE RCU API?
  281. This section shows a simple use of the core RCU API to protect a
  282. global pointer to a dynamically allocated structure. More-typical
  283. uses of RCU may be found in listRCU.txt, arrayRCU.txt, and NMI-RCU.txt.
  284. struct foo {
  285. int a;
  286. char b;
  287. long c;
  288. };
  289. DEFINE_SPINLOCK(foo_mutex);
  290. struct foo __rcu *gbl_foo;
  291. /*
  292. * Create a new struct foo that is the same as the one currently
  293. * pointed to by gbl_foo, except that field "a" is replaced
  294. * with "new_a". Points gbl_foo to the new structure, and
  295. * frees up the old structure after a grace period.
  296. *
  297. * Uses rcu_assign_pointer() to ensure that concurrent readers
  298. * see the initialized version of the new structure.
  299. *
  300. * Uses synchronize_rcu() to ensure that any readers that might
  301. * have references to the old structure complete before freeing
  302. * the old structure.
  303. */
  304. void foo_update_a(int new_a)
  305. {
  306. struct foo *new_fp;
  307. struct foo *old_fp;
  308. new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
  309. spin_lock(&foo_mutex);
  310. old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
  311. *new_fp = *old_fp;
  312. new_fp->a = new_a;
  313. rcu_assign_pointer(gbl_foo, new_fp);
  314. spin_unlock(&foo_mutex);
  315. synchronize_rcu();
  316. kfree(old_fp);
  317. }
  318. /*
  319. * Return the value of field "a" of the current gbl_foo
  320. * structure. Use rcu_read_lock() and rcu_read_unlock()
  321. * to ensure that the structure does not get deleted out
  322. * from under us, and use rcu_dereference() to ensure that
  323. * we see the initialized version of the structure (important
  324. * for DEC Alpha and for people reading the code).
  325. */
  326. int foo_get_a(void)
  327. {
  328. int retval;
  329. rcu_read_lock();
  330. retval = rcu_dereference(gbl_foo)->a;
  331. rcu_read_unlock();
  332. return retval;
  333. }
  334. So, to sum up:
  335. o Use rcu_read_lock() and rcu_read_unlock() to guard RCU
  336. read-side critical sections.
  337. o Within an RCU read-side critical section, use rcu_dereference()
  338. to dereference RCU-protected pointers.
  339. o Use some solid scheme (such as locks or semaphores) to
  340. keep concurrent updates from interfering with each other.
  341. o Use rcu_assign_pointer() to update an RCU-protected pointer.
  342. This primitive protects concurrent readers from the updater,
  343. -not- concurrent updates from each other! You therefore still
  344. need to use locking (or something similar) to keep concurrent
  345. rcu_assign_pointer() primitives from interfering with each other.
  346. o Use synchronize_rcu() -after- removing a data element from an
  347. RCU-protected data structure, but -before- reclaiming/freeing
  348. the data element, in order to wait for the completion of all
  349. RCU read-side critical sections that might be referencing that
  350. data item.
  351. See checklist.txt for additional rules to follow when using RCU.
  352. And again, more-typical uses of RCU may be found in listRCU.txt,
  353. arrayRCU.txt, and NMI-RCU.txt.
  354. 4. WHAT IF MY UPDATING THREAD CANNOT BLOCK?
  355. In the example above, foo_update_a() blocks until a grace period elapses.
  356. This is quite simple, but in some cases one cannot afford to wait so
  357. long -- there might be other high-priority work to be done.
  358. In such cases, one uses call_rcu() rather than synchronize_rcu().
  359. The call_rcu() API is as follows:
  360. void call_rcu(struct rcu_head * head,
  361. void (*func)(struct rcu_head *head));
  362. This function invokes func(head) after a grace period has elapsed.
  363. This invocation might happen from either softirq or process context,
  364. so the function is not permitted to block. The foo struct needs to
  365. have an rcu_head structure added, perhaps as follows:
  366. struct foo {
  367. int a;
  368. char b;
  369. long c;
  370. struct rcu_head rcu;
  371. };
  372. The foo_update_a() function might then be written as follows:
  373. /*
  374. * Create a new struct foo that is the same as the one currently
  375. * pointed to by gbl_foo, except that field "a" is replaced
  376. * with "new_a". Points gbl_foo to the new structure, and
  377. * frees up the old structure after a grace period.
  378. *
  379. * Uses rcu_assign_pointer() to ensure that concurrent readers
  380. * see the initialized version of the new structure.
  381. *
  382. * Uses call_rcu() to ensure that any readers that might have
  383. * references to the old structure complete before freeing the
  384. * old structure.
  385. */
  386. void foo_update_a(int new_a)
  387. {
  388. struct foo *new_fp;
  389. struct foo *old_fp;
  390. new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL);
  391. spin_lock(&foo_mutex);
  392. old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex));
  393. *new_fp = *old_fp;
  394. new_fp->a = new_a;
  395. rcu_assign_pointer(gbl_foo, new_fp);
  396. spin_unlock(&foo_mutex);
  397. call_rcu(&old_fp->rcu, foo_reclaim);
  398. }
  399. The foo_reclaim() function might appear as follows:
  400. void foo_reclaim(struct rcu_head *rp)
  401. {
  402. struct foo *fp = container_of(rp, struct foo, rcu);
  403. foo_cleanup(fp->a);
  404. kfree(fp);
  405. }
  406. The container_of() primitive is a macro that, given a pointer into a
  407. struct, the type of the struct, and the pointed-to field within the
  408. struct, returns a pointer to the beginning of the struct.
  409. The use of call_rcu() permits the caller of foo_update_a() to
  410. immediately regain control, without needing to worry further about the
  411. old version of the newly updated element. It also clearly shows the
  412. RCU distinction between updater, namely foo_update_a(), and reclaimer,
  413. namely foo_reclaim().
  414. The summary of advice is the same as for the previous section, except
  415. that we are now using call_rcu() rather than synchronize_rcu():
  416. o Use call_rcu() -after- removing a data element from an
  417. RCU-protected data structure in order to register a callback
  418. function that will be invoked after the completion of all RCU
  419. read-side critical sections that might be referencing that
  420. data item.
  421. If the callback for call_rcu() is not doing anything more than calling
  422. kfree() on the structure, you can use kfree_rcu() instead of call_rcu()
  423. to avoid having to write your own callback:
  424. kfree_rcu(old_fp, rcu);
  425. Again, see checklist.txt for additional rules governing the use of RCU.
  426. 5. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU?
  427. One of the nice things about RCU is that it has extremely simple "toy"
  428. implementations that are a good first step towards understanding the
  429. production-quality implementations in the Linux kernel. This section
  430. presents two such "toy" implementations of RCU, one that is implemented
  431. in terms of familiar locking primitives, and another that more closely
  432. resembles "classic" RCU. Both are way too simple for real-world use,
  433. lacking both functionality and performance. However, they are useful
  434. in getting a feel for how RCU works. See kernel/rcupdate.c for a
  435. production-quality implementation, and see:
  436. http://www.rdrop.com/users/paulmck/RCU
  437. for papers describing the Linux kernel RCU implementation. The OLS'01
  438. and OLS'02 papers are a good introduction, and the dissertation provides
  439. more details on the current implementation as of early 2004.
  440. 5A. "TOY" IMPLEMENTATION #1: LOCKING
  441. This section presents a "toy" RCU implementation that is based on
  442. familiar locking primitives. Its overhead makes it a non-starter for
  443. real-life use, as does its lack of scalability. It is also unsuitable
  444. for realtime use, since it allows scheduling latency to "bleed" from
  445. one read-side critical section to another.
  446. However, it is probably the easiest implementation to relate to, so is
  447. a good starting point.
  448. It is extremely simple:
  449. static DEFINE_RWLOCK(rcu_gp_mutex);
  450. void rcu_read_lock(void)
  451. {
  452. read_lock(&rcu_gp_mutex);
  453. }
  454. void rcu_read_unlock(void)
  455. {
  456. read_unlock(&rcu_gp_mutex);
  457. }
  458. void synchronize_rcu(void)
  459. {
  460. write_lock(&rcu_gp_mutex);
  461. write_unlock(&rcu_gp_mutex);
  462. }
  463. [You can ignore rcu_assign_pointer() and rcu_dereference() without
  464. missing much. But here they are anyway. And whatever you do, don't
  465. forget about them when submitting patches making use of RCU!]
  466. #define rcu_assign_pointer(p, v) ({ \
  467. smp_wmb(); \
  468. (p) = (v); \
  469. })
  470. #define rcu_dereference(p) ({ \
  471. typeof(p) _________p1 = p; \
  472. smp_read_barrier_depends(); \
  473. (_________p1); \
  474. })
  475. The rcu_read_lock() and rcu_read_unlock() primitive read-acquire
  476. and release a global reader-writer lock. The synchronize_rcu()
  477. primitive write-acquires this same lock, then immediately releases
  478. it. This means that once synchronize_rcu() exits, all RCU read-side
  479. critical sections that were in progress before synchronize_rcu() was
  480. called are guaranteed to have completed -- there is no way that
  481. synchronize_rcu() would have been able to write-acquire the lock
  482. otherwise.
  483. It is possible to nest rcu_read_lock(), since reader-writer locks may
  484. be recursively acquired. Note also that rcu_read_lock() is immune
  485. from deadlock (an important property of RCU). The reason for this is
  486. that the only thing that can block rcu_read_lock() is a synchronize_rcu().
  487. But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex,
  488. so there can be no deadlock cycle.
  489. Quick Quiz #1: Why is this argument naive? How could a deadlock
  490. occur when using this algorithm in a real-world Linux
  491. kernel? How could this deadlock be avoided?
  492. 5B. "TOY" EXAMPLE #2: CLASSIC RCU
  493. This section presents a "toy" RCU implementation that is based on
  494. "classic RCU". It is also short on performance (but only for updates) and
  495. on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT
  496. kernels. The definitions of rcu_dereference() and rcu_assign_pointer()
  497. are the same as those shown in the preceding section, so they are omitted.
  498. void rcu_read_lock(void) { }
  499. void rcu_read_unlock(void) { }
  500. void synchronize_rcu(void)
  501. {
  502. int cpu;
  503. for_each_possible_cpu(cpu)
  504. run_on(cpu);
  505. }
  506. Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing.
  507. This is the great strength of classic RCU in a non-preemptive kernel:
  508. read-side overhead is precisely zero, at least on non-Alpha CPUs.
  509. And there is absolutely no way that rcu_read_lock() can possibly
  510. participate in a deadlock cycle!
  511. The implementation of synchronize_rcu() simply schedules itself on each
  512. CPU in turn. The run_on() primitive can be implemented straightforwardly
  513. in terms of the sched_setaffinity() primitive. Of course, a somewhat less
  514. "toy" implementation would restore the affinity upon completion rather
  515. than just leaving all tasks running on the last CPU, but when I said
  516. "toy", I meant -toy-!
  517. So how the heck is this supposed to work???
  518. Remember that it is illegal to block while in an RCU read-side critical
  519. section. Therefore, if a given CPU executes a context switch, we know
  520. that it must have completed all preceding RCU read-side critical sections.
  521. Once -all- CPUs have executed a context switch, then -all- preceding
  522. RCU read-side critical sections will have completed.
  523. So, suppose that we remove a data item from its structure and then invoke
  524. synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed
  525. that there are no RCU read-side critical sections holding a reference
  526. to that data item, so we can safely reclaim it.
  527. Quick Quiz #2: Give an example where Classic RCU's read-side
  528. overhead is -negative-.
  529. Quick Quiz #3: If it is illegal to block in an RCU read-side
  530. critical section, what the heck do you do in
  531. PREEMPT_RT, where normal spinlocks can block???
  532. 6. ANALOGY WITH READER-WRITER LOCKING
  533. Although RCU can be used in many different ways, a very common use of
  534. RCU is analogous to reader-writer locking. The following unified
  535. diff shows how closely related RCU and reader-writer locking can be.
  536. @@ -13,15 +14,15 @@
  537. struct list_head *lp;
  538. struct el *p;
  539. - read_lock();
  540. - list_for_each_entry(p, head, lp) {
  541. + rcu_read_lock();
  542. + list_for_each_entry_rcu(p, head, lp) {
  543. if (p->key == key) {
  544. *result = p->data;
  545. - read_unlock();
  546. + rcu_read_unlock();
  547. return 1;
  548. }
  549. }
  550. - read_unlock();
  551. + rcu_read_unlock();
  552. return 0;
  553. }
  554. @@ -29,15 +30,16 @@
  555. {
  556. struct el *p;
  557. - write_lock(&listmutex);
  558. + spin_lock(&listmutex);
  559. list_for_each_entry(p, head, lp) {
  560. if (p->key == key) {
  561. - list_del(&p->list);
  562. - write_unlock(&listmutex);
  563. + list_del_rcu(&p->list);
  564. + spin_unlock(&listmutex);
  565. + synchronize_rcu();
  566. kfree(p);
  567. return 1;
  568. }
  569. }
  570. - write_unlock(&listmutex);
  571. + spin_unlock(&listmutex);
  572. return 0;
  573. }
  574. Or, for those who prefer a side-by-side listing:
  575. 1 struct el { 1 struct el {
  576. 2 struct list_head list; 2 struct list_head list;
  577. 3 long key; 3 long key;
  578. 4 spinlock_t mutex; 4 spinlock_t mutex;
  579. 5 int data; 5 int data;
  580. 6 /* Other data fields */ 6 /* Other data fields */
  581. 7 }; 7 };
  582. 8 spinlock_t listmutex; 8 spinlock_t listmutex;
  583. 9 struct el head; 9 struct el head;
  584. 1 int search(long key, int *result) 1 int search(long key, int *result)
  585. 2 { 2 {
  586. 3 struct list_head *lp; 3 struct list_head *lp;
  587. 4 struct el *p; 4 struct el *p;
  588. 5 5
  589. 6 read_lock(); 6 rcu_read_lock();
  590. 7 list_for_each_entry(p, head, lp) { 7 list_for_each_entry_rcu(p, head, lp) {
  591. 8 if (p->key == key) { 8 if (p->key == key) {
  592. 9 *result = p->data; 9 *result = p->data;
  593. 10 read_unlock(); 10 rcu_read_unlock();
  594. 11 return 1; 11 return 1;
  595. 12 } 12 }
  596. 13 } 13 }
  597. 14 read_unlock(); 14 rcu_read_unlock();
  598. 15 return 0; 15 return 0;
  599. 16 } 16 }
  600. 1 int delete(long key) 1 int delete(long key)
  601. 2 { 2 {
  602. 3 struct el *p; 3 struct el *p;
  603. 4 4
  604. 5 write_lock(&listmutex); 5 spin_lock(&listmutex);
  605. 6 list_for_each_entry(p, head, lp) { 6 list_for_each_entry(p, head, lp) {
  606. 7 if (p->key == key) { 7 if (p->key == key) {
  607. 8 list_del(&p->list); 8 list_del_rcu(&p->list);
  608. 9 write_unlock(&listmutex); 9 spin_unlock(&listmutex);
  609. 10 synchronize_rcu();
  610. 10 kfree(p); 11 kfree(p);
  611. 11 return 1; 12 return 1;
  612. 12 } 13 }
  613. 13 } 14 }
  614. 14 write_unlock(&listmutex); 15 spin_unlock(&listmutex);
  615. 15 return 0; 16 return 0;
  616. 16 } 17 }
  617. Either way, the differences are quite small. Read-side locking moves
  618. to rcu_read_lock() and rcu_read_unlock, update-side locking moves from
  619. a reader-writer lock to a simple spinlock, and a synchronize_rcu()
  620. precedes the kfree().
  621. However, there is one potential catch: the read-side and update-side
  622. critical sections can now run concurrently. In many cases, this will
  623. not be a problem, but it is necessary to check carefully regardless.
  624. For example, if multiple independent list updates must be seen as
  625. a single atomic update, converting to RCU will require special care.
  626. Also, the presence of synchronize_rcu() means that the RCU version of
  627. delete() can now block. If this is a problem, there is a callback-based
  628. mechanism that never blocks, namely call_rcu() or kfree_rcu(), that can
  629. be used in place of synchronize_rcu().
  630. 7. FULL LIST OF RCU APIs
  631. The RCU APIs are documented in docbook-format header comments in the
  632. Linux-kernel source code, but it helps to have a full list of the
  633. APIs, since there does not appear to be a way to categorize them
  634. in docbook. Here is the list, by category.
  635. RCU list traversal:
  636. list_entry_rcu
  637. list_first_entry_rcu
  638. list_next_rcu
  639. list_for_each_entry_rcu
  640. list_for_each_entry_continue_rcu
  641. hlist_first_rcu
  642. hlist_next_rcu
  643. hlist_pprev_rcu
  644. hlist_for_each_entry_rcu
  645. hlist_for_each_entry_rcu_bh
  646. hlist_for_each_entry_continue_rcu
  647. hlist_for_each_entry_continue_rcu_bh
  648. hlist_nulls_first_rcu
  649. hlist_nulls_for_each_entry_rcu
  650. hlist_bl_first_rcu
  651. hlist_bl_for_each_entry_rcu
  652. RCU pointer/list update:
  653. rcu_assign_pointer
  654. list_add_rcu
  655. list_add_tail_rcu
  656. list_del_rcu
  657. list_replace_rcu
  658. hlist_add_behind_rcu
  659. hlist_add_before_rcu
  660. hlist_add_head_rcu
  661. hlist_del_rcu
  662. hlist_del_init_rcu
  663. hlist_replace_rcu
  664. list_splice_init_rcu()
  665. hlist_nulls_del_init_rcu
  666. hlist_nulls_del_rcu
  667. hlist_nulls_add_head_rcu
  668. hlist_bl_add_head_rcu
  669. hlist_bl_del_init_rcu
  670. hlist_bl_del_rcu
  671. hlist_bl_set_first_rcu
  672. RCU: Critical sections Grace period Barrier
  673. rcu_read_lock synchronize_net rcu_barrier
  674. rcu_read_unlock synchronize_rcu
  675. rcu_dereference synchronize_rcu_expedited
  676. rcu_read_lock_held call_rcu
  677. rcu_dereference_check kfree_rcu
  678. rcu_dereference_protected
  679. bh: Critical sections Grace period Barrier
  680. rcu_read_lock_bh call_rcu_bh rcu_barrier_bh
  681. rcu_read_unlock_bh synchronize_rcu_bh
  682. rcu_dereference_bh synchronize_rcu_bh_expedited
  683. rcu_dereference_bh_check
  684. rcu_dereference_bh_protected
  685. rcu_read_lock_bh_held
  686. sched: Critical sections Grace period Barrier
  687. rcu_read_lock_sched synchronize_sched rcu_barrier_sched
  688. rcu_read_unlock_sched call_rcu_sched
  689. [preempt_disable] synchronize_sched_expedited
  690. [and friends]
  691. rcu_read_lock_sched_notrace
  692. rcu_read_unlock_sched_notrace
  693. rcu_dereference_sched
  694. rcu_dereference_sched_check
  695. rcu_dereference_sched_protected
  696. rcu_read_lock_sched_held
  697. SRCU: Critical sections Grace period Barrier
  698. srcu_read_lock synchronize_srcu srcu_barrier
  699. srcu_read_unlock call_srcu
  700. srcu_dereference synchronize_srcu_expedited
  701. srcu_dereference_check
  702. srcu_read_lock_held
  703. SRCU: Initialization/cleanup
  704. init_srcu_struct
  705. cleanup_srcu_struct
  706. All: lockdep-checked RCU-protected pointer access
  707. rcu_access_pointer
  708. rcu_dereference_raw
  709. RCU_LOCKDEP_WARN
  710. rcu_sleep_check
  711. RCU_NONIDLE
  712. See the comment headers in the source code (or the docbook generated
  713. from them) for more information.
  714. However, given that there are no fewer than four families of RCU APIs
  715. in the Linux kernel, how do you choose which one to use? The following
  716. list can be helpful:
  717. a. Will readers need to block? If so, you need SRCU.
  718. b. What about the -rt patchset? If readers would need to block
  719. in an non-rt kernel, you need SRCU. If readers would block
  720. in a -rt kernel, but not in a non-rt kernel, SRCU is not
  721. necessary.
  722. c. Do you need to treat NMI handlers, hardirq handlers,
  723. and code segments with preemption disabled (whether
  724. via preempt_disable(), local_irq_save(), local_bh_disable(),
  725. or some other mechanism) as if they were explicit RCU readers?
  726. If so, RCU-sched is the only choice that will work for you.
  727. d. Do you need RCU grace periods to complete even in the face
  728. of softirq monopolization of one or more of the CPUs? For
  729. example, is your code subject to network-based denial-of-service
  730. attacks? If so, you need RCU-bh.
  731. e. Is your workload too update-intensive for normal use of
  732. RCU, but inappropriate for other synchronization mechanisms?
  733. If so, consider SLAB_DESTROY_BY_RCU. But please be careful!
  734. f. Do you need read-side critical sections that are respected
  735. even though they are in the middle of the idle loop, during
  736. user-mode execution, or on an offlined CPU? If so, SRCU is the
  737. only choice that will work for you.
  738. g. Otherwise, use RCU.
  739. Of course, this all assumes that you have determined that RCU is in fact
  740. the right tool for your job.
  741. 8. ANSWERS TO QUICK QUIZZES
  742. Quick Quiz #1: Why is this argument naive? How could a deadlock
  743. occur when using this algorithm in a real-world Linux
  744. kernel? [Referring to the lock-based "toy" RCU
  745. algorithm.]
  746. Answer: Consider the following sequence of events:
  747. 1. CPU 0 acquires some unrelated lock, call it
  748. "problematic_lock", disabling irq via
  749. spin_lock_irqsave().
  750. 2. CPU 1 enters synchronize_rcu(), write-acquiring
  751. rcu_gp_mutex.
  752. 3. CPU 0 enters rcu_read_lock(), but must wait
  753. because CPU 1 holds rcu_gp_mutex.
  754. 4. CPU 1 is interrupted, and the irq handler
  755. attempts to acquire problematic_lock.
  756. The system is now deadlocked.
  757. One way to avoid this deadlock is to use an approach like
  758. that of CONFIG_PREEMPT_RT, where all normal spinlocks
  759. become blocking locks, and all irq handlers execute in
  760. the context of special tasks. In this case, in step 4
  761. above, the irq handler would block, allowing CPU 1 to
  762. release rcu_gp_mutex, avoiding the deadlock.
  763. Even in the absence of deadlock, this RCU implementation
  764. allows latency to "bleed" from readers to other
  765. readers through synchronize_rcu(). To see this,
  766. consider task A in an RCU read-side critical section
  767. (thus read-holding rcu_gp_mutex), task B blocked
  768. attempting to write-acquire rcu_gp_mutex, and
  769. task C blocked in rcu_read_lock() attempting to
  770. read_acquire rcu_gp_mutex. Task A's RCU read-side
  771. latency is holding up task C, albeit indirectly via
  772. task B.
  773. Realtime RCU implementations therefore use a counter-based
  774. approach where tasks in RCU read-side critical sections
  775. cannot be blocked by tasks executing synchronize_rcu().
  776. Quick Quiz #2: Give an example where Classic RCU's read-side
  777. overhead is -negative-.
  778. Answer: Imagine a single-CPU system with a non-CONFIG_PREEMPT
  779. kernel where a routing table is used by process-context
  780. code, but can be updated by irq-context code (for example,
  781. by an "ICMP REDIRECT" packet). The usual way of handling
  782. this would be to have the process-context code disable
  783. interrupts while searching the routing table. Use of
  784. RCU allows such interrupt-disabling to be dispensed with.
  785. Thus, without RCU, you pay the cost of disabling interrupts,
  786. and with RCU you don't.
  787. One can argue that the overhead of RCU in this
  788. case is negative with respect to the single-CPU
  789. interrupt-disabling approach. Others might argue that
  790. the overhead of RCU is merely zero, and that replacing
  791. the positive overhead of the interrupt-disabling scheme
  792. with the zero-overhead RCU scheme does not constitute
  793. negative overhead.
  794. In real life, of course, things are more complex. But
  795. even the theoretical possibility of negative overhead for
  796. a synchronization primitive is a bit unexpected. ;-)
  797. Quick Quiz #3: If it is illegal to block in an RCU read-side
  798. critical section, what the heck do you do in
  799. PREEMPT_RT, where normal spinlocks can block???
  800. Answer: Just as PREEMPT_RT permits preemption of spinlock
  801. critical sections, it permits preemption of RCU
  802. read-side critical sections. It also permits
  803. spinlocks blocking while in RCU read-side critical
  804. sections.
  805. Why the apparent inconsistency? Because it is it
  806. possible to use priority boosting to keep the RCU
  807. grace periods short if need be (for example, if running
  808. short of memory). In contrast, if blocking waiting
  809. for (say) network reception, there is no way to know
  810. what should be boosted. Especially given that the
  811. process we need to boost might well be a human being
  812. who just went out for a pizza or something. And although
  813. a computer-operated cattle prod might arouse serious
  814. interest, it might also provoke serious objections.
  815. Besides, how does the computer know what pizza parlor
  816. the human being went to???
  817. ACKNOWLEDGEMENTS
  818. My thanks to the people who helped make this human-readable, including
  819. Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern.
  820. For more information, see http://www.rdrop.com/users/paulmck/RCU.