checklist.rst 22 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465
  1. .. SPDX-License-Identifier: GPL-2.0
  2. ================================
  3. Review Checklist for RCU Patches
  4. ================================
  5. This document contains a checklist for producing and reviewing patches
  6. that make use of RCU. Violating any of the rules listed below will
  7. result in the same sorts of problems that leaving out a locking primitive
  8. would cause. This list is based on experiences reviewing such patches
  9. over a rather long period of time, but improvements are always welcome!
  10. 0. Is RCU being applied to a read-mostly situation? If the data
  11. structure is updated more than about 10% of the time, then you
  12. should strongly consider some other approach, unless detailed
  13. performance measurements show that RCU is nonetheless the right
  14. tool for the job. Yes, RCU does reduce read-side overhead by
  15. increasing write-side overhead, which is exactly why normal uses
  16. of RCU will do much more reading than updating.
  17. Another exception is where performance is not an issue, and RCU
  18. provides a simpler implementation. An example of this situation
  19. is the dynamic NMI code in the Linux 2.6 kernel, at least on
  20. architectures where NMIs are rare.
  21. Yet another exception is where the low real-time latency of RCU's
  22. read-side primitives is critically important.
  23. One final exception is where RCU readers are used to prevent
  24. the ABA problem (https://en.wikipedia.org/wiki/ABA_problem)
  25. for lockless updates. This does result in the mildly
  26. counter-intuitive situation where rcu_read_lock() and
  27. rcu_read_unlock() are used to protect updates, however, this
  28. approach provides the same potential simplifications that garbage
  29. collectors do.
  30. 1. Does the update code have proper mutual exclusion?
  31. RCU does allow -readers- to run (almost) naked, but -writers- must
  32. still use some sort of mutual exclusion, such as:
  33. a. locking,
  34. b. atomic operations, or
  35. c. restricting updates to a single task.
  36. If you choose #b, be prepared to describe how you have handled
  37. memory barriers on weakly ordered machines (pretty much all of
  38. them -- even x86 allows later loads to be reordered to precede
  39. earlier stores), and be prepared to explain why this added
  40. complexity is worthwhile. If you choose #c, be prepared to
  41. explain how this single task does not become a major bottleneck on
  42. big multiprocessor machines (for example, if the task is updating
  43. information relating to itself that other tasks can read, there
  44. by definition can be no bottleneck). Note that the definition
  45. of "large" has changed significantly: Eight CPUs was "large"
  46. in the year 2000, but a hundred CPUs was unremarkable in 2017.
  47. 2. Do the RCU read-side critical sections make proper use of
  48. rcu_read_lock() and friends? These primitives are needed
  49. to prevent grace periods from ending prematurely, which
  50. could result in data being unceremoniously freed out from
  51. under your read-side code, which can greatly increase the
  52. actuarial risk of your kernel.
  53. As a rough rule of thumb, any dereference of an RCU-protected
  54. pointer must be covered by rcu_read_lock(), rcu_read_lock_bh(),
  55. rcu_read_lock_sched(), or by the appropriate update-side lock.
  56. Disabling of preemption can serve as rcu_read_lock_sched(), but
  57. is less readable and prevents lockdep from detecting locking issues.
  58. Letting RCU-protected pointers "leak" out of an RCU read-side
  59. critical section is every bid as bad as letting them leak out
  60. from under a lock. Unless, of course, you have arranged some
  61. other means of protection, such as a lock or a reference count
  62. -before- letting them out of the RCU read-side critical section.
  63. 3. Does the update code tolerate concurrent accesses?
  64. The whole point of RCU is to permit readers to run without
  65. any locks or atomic operations. This means that readers will
  66. be running while updates are in progress. There are a number
  67. of ways to handle this concurrency, depending on the situation:
  68. a. Use the RCU variants of the list and hlist update
  69. primitives to add, remove, and replace elements on
  70. an RCU-protected list. Alternatively, use the other
  71. RCU-protected data structures that have been added to
  72. the Linux kernel.
  73. This is almost always the best approach.
  74. b. Proceed as in (a) above, but also maintain per-element
  75. locks (that are acquired by both readers and writers)
  76. that guard per-element state. Of course, fields that
  77. the readers refrain from accessing can be guarded by
  78. some other lock acquired only by updaters, if desired.
  79. This works quite well, also.
  80. c. Make updates appear atomic to readers. For example,
  81. pointer updates to properly aligned fields will
  82. appear atomic, as will individual atomic primitives.
  83. Sequences of operations performed under a lock will -not-
  84. appear to be atomic to RCU readers, nor will sequences
  85. of multiple atomic primitives.
  86. This can work, but is starting to get a bit tricky.
  87. d. Carefully order the updates and the reads so that
  88. readers see valid data at all phases of the update.
  89. This is often more difficult than it sounds, especially
  90. given modern CPUs' tendency to reorder memory references.
  91. One must usually liberally sprinkle memory barriers
  92. (smp_wmb(), smp_rmb(), smp_mb()) through the code,
  93. making it difficult to understand and to test.
  94. It is usually better to group the changing data into
  95. a separate structure, so that the change may be made
  96. to appear atomic by updating a pointer to reference
  97. a new structure containing updated values.
  98. 4. Weakly ordered CPUs pose special challenges. Almost all CPUs
  99. are weakly ordered -- even x86 CPUs allow later loads to be
  100. reordered to precede earlier stores. RCU code must take all of
  101. the following measures to prevent memory-corruption problems:
  102. a. Readers must maintain proper ordering of their memory
  103. accesses. The rcu_dereference() primitive ensures that
  104. the CPU picks up the pointer before it picks up the data
  105. that the pointer points to. This really is necessary
  106. on Alpha CPUs. If you don't believe me, see:
  107. http://www.openvms.compaq.com/wizard/wiz_2637.html
  108. The rcu_dereference() primitive is also an excellent
  109. documentation aid, letting the person reading the
  110. code know exactly which pointers are protected by RCU.
  111. Please note that compilers can also reorder code, and
  112. they are becoming increasingly aggressive about doing
  113. just that. The rcu_dereference() primitive therefore also
  114. prevents destructive compiler optimizations. However,
  115. with a bit of devious creativity, it is possible to
  116. mishandle the return value from rcu_dereference().
  117. Please see rcu_dereference.txt in this directory for
  118. more information.
  119. The rcu_dereference() primitive is used by the
  120. various "_rcu()" list-traversal primitives, such
  121. as the list_for_each_entry_rcu(). Note that it is
  122. perfectly legal (if redundant) for update-side code to
  123. use rcu_dereference() and the "_rcu()" list-traversal
  124. primitives. This is particularly useful in code that
  125. is common to readers and updaters. However, lockdep
  126. will complain if you access rcu_dereference() outside
  127. of an RCU read-side critical section. See lockdep.txt
  128. to learn what to do about this.
  129. Of course, neither rcu_dereference() nor the "_rcu()"
  130. list-traversal primitives can substitute for a good
  131. concurrency design coordinating among multiple updaters.
  132. b. If the list macros are being used, the list_add_tail_rcu()
  133. and list_add_rcu() primitives must be used in order
  134. to prevent weakly ordered machines from misordering
  135. structure initialization and pointer planting.
  136. Similarly, if the hlist macros are being used, the
  137. hlist_add_head_rcu() primitive is required.
  138. c. If the list macros are being used, the list_del_rcu()
  139. primitive must be used to keep list_del()'s pointer
  140. poisoning from inflicting toxic effects on concurrent
  141. readers. Similarly, if the hlist macros are being used,
  142. the hlist_del_rcu() primitive is required.
  143. The list_replace_rcu() and hlist_replace_rcu() primitives
  144. may be used to replace an old structure with a new one
  145. in their respective types of RCU-protected lists.
  146. d. Rules similar to (4b) and (4c) apply to the "hlist_nulls"
  147. type of RCU-protected linked lists.
  148. e. Updates must ensure that initialization of a given
  149. structure happens before pointers to that structure are
  150. publicized. Use the rcu_assign_pointer() primitive
  151. when publicizing a pointer to a structure that can
  152. be traversed by an RCU read-side critical section.
  153. 5. If call_rcu() or call_srcu() is used, the callback function will
  154. be called from softirq context. In particular, it cannot block.
  155. 6. Since synchronize_rcu() can block, it cannot be called
  156. from any sort of irq context. The same rule applies
  157. for synchronize_srcu(), synchronize_rcu_expedited(), and
  158. synchronize_srcu_expedited().
  159. The expedited forms of these primitives have the same semantics
  160. as the non-expedited forms, but expediting is both expensive and
  161. (with the exception of synchronize_srcu_expedited()) unfriendly
  162. to real-time workloads. Use of the expedited primitives should
  163. be restricted to rare configuration-change operations that would
  164. not normally be undertaken while a real-time workload is running.
  165. However, real-time workloads can use rcupdate.rcu_normal kernel
  166. boot parameter to completely disable expedited grace periods,
  167. though this might have performance implications.
  168. In particular, if you find yourself invoking one of the expedited
  169. primitives repeatedly in a loop, please do everyone a favor:
  170. Restructure your code so that it batches the updates, allowing
  171. a single non-expedited primitive to cover the entire batch.
  172. This will very likely be faster than the loop containing the
  173. expedited primitive, and will be much much easier on the rest
  174. of the system, especially to real-time workloads running on
  175. the rest of the system.
  176. 7. As of v4.20, a given kernel implements only one RCU flavor,
  177. which is RCU-sched for PREEMPT=n and RCU-preempt for PREEMPT=y.
  178. If the updater uses call_rcu() or synchronize_rcu(),
  179. then the corresponding readers my use rcu_read_lock() and
  180. rcu_read_unlock(), rcu_read_lock_bh() and rcu_read_unlock_bh(),
  181. or any pair of primitives that disables and re-enables preemption,
  182. for example, rcu_read_lock_sched() and rcu_read_unlock_sched().
  183. If the updater uses synchronize_srcu() or call_srcu(),
  184. then the corresponding readers must use srcu_read_lock() and
  185. srcu_read_unlock(), and with the same srcu_struct. The rules for
  186. the expedited primitives are the same as for their non-expedited
  187. counterparts. Mixing things up will result in confusion and
  188. broken kernels, and has even resulted in an exploitable security
  189. issue.
  190. One exception to this rule: rcu_read_lock() and rcu_read_unlock()
  191. may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh()
  192. in cases where local bottom halves are already known to be
  193. disabled, for example, in irq or softirq context. Commenting
  194. such cases is a must, of course! And the jury is still out on
  195. whether the increased speed is worth it.
  196. 8. Although synchronize_rcu() is slower than is call_rcu(), it
  197. usually results in simpler code. So, unless update performance is
  198. critically important, the updaters cannot block, or the latency of
  199. synchronize_rcu() is visible from userspace, synchronize_rcu()
  200. should be used in preference to call_rcu(). Furthermore,
  201. kfree_rcu() usually results in even simpler code than does
  202. synchronize_rcu() without synchronize_rcu()'s multi-millisecond
  203. latency. So please take advantage of kfree_rcu()'s "fire and
  204. forget" memory-freeing capabilities where it applies.
  205. An especially important property of the synchronize_rcu()
  206. primitive is that it automatically self-limits: if grace periods
  207. are delayed for whatever reason, then the synchronize_rcu()
  208. primitive will correspondingly delay updates. In contrast,
  209. code using call_rcu() should explicitly limit update rate in
  210. cases where grace periods are delayed, as failing to do so can
  211. result in excessive realtime latencies or even OOM conditions.
  212. Ways of gaining this self-limiting property when using call_rcu()
  213. include:
  214. a. Keeping a count of the number of data-structure elements
  215. used by the RCU-protected data structure, including
  216. those waiting for a grace period to elapse. Enforce a
  217. limit on this number, stalling updates as needed to allow
  218. previously deferred frees to complete. Alternatively,
  219. limit only the number awaiting deferred free rather than
  220. the total number of elements.
  221. One way to stall the updates is to acquire the update-side
  222. mutex. (Don't try this with a spinlock -- other CPUs
  223. spinning on the lock could prevent the grace period
  224. from ever ending.) Another way to stall the updates
  225. is for the updates to use a wrapper function around
  226. the memory allocator, so that this wrapper function
  227. simulates OOM when there is too much memory awaiting an
  228. RCU grace period. There are of course many other
  229. variations on this theme.
  230. b. Limiting update rate. For example, if updates occur only
  231. once per hour, then no explicit rate limiting is
  232. required, unless your system is already badly broken.
  233. Older versions of the dcache subsystem take this approach,
  234. guarding updates with a global lock, limiting their rate.
  235. c. Trusted update -- if updates can only be done manually by
  236. superuser or some other trusted user, then it might not
  237. be necessary to automatically limit them. The theory
  238. here is that superuser already has lots of ways to crash
  239. the machine.
  240. d. Periodically invoke synchronize_rcu(), permitting a limited
  241. number of updates per grace period.
  242. The same cautions apply to call_srcu() and kfree_rcu().
  243. Note that although these primitives do take action to avoid memory
  244. exhaustion when any given CPU has too many callbacks, a determined
  245. user could still exhaust memory. This is especially the case
  246. if a system with a large number of CPUs has been configured to
  247. offload all of its RCU callbacks onto a single CPU, or if the
  248. system has relatively little free memory.
  249. 9. All RCU list-traversal primitives, which include
  250. rcu_dereference(), list_for_each_entry_rcu(), and
  251. list_for_each_safe_rcu(), must be either within an RCU read-side
  252. critical section or must be protected by appropriate update-side
  253. locks. RCU read-side critical sections are delimited by
  254. rcu_read_lock() and rcu_read_unlock(), or by similar primitives
  255. such as rcu_read_lock_bh() and rcu_read_unlock_bh(), in which
  256. case the matching rcu_dereference() primitive must be used in
  257. order to keep lockdep happy, in this case, rcu_dereference_bh().
  258. The reason that it is permissible to use RCU list-traversal
  259. primitives when the update-side lock is held is that doing so
  260. can be quite helpful in reducing code bloat when common code is
  261. shared between readers and updaters. Additional primitives
  262. are provided for this case, as discussed in lockdep.txt.
  263. 10. Conversely, if you are in an RCU read-side critical section,
  264. and you don't hold the appropriate update-side lock, you -must-
  265. use the "_rcu()" variants of the list macros. Failing to do so
  266. will break Alpha, cause aggressive compilers to generate bad code,
  267. and confuse people trying to read your code.
  268. 11. Any lock acquired by an RCU callback must be acquired elsewhere
  269. with softirq disabled, e.g., via spin_lock_irqsave(),
  270. spin_lock_bh(), etc. Failing to disable softirq on a given
  271. acquisition of that lock will result in deadlock as soon as
  272. the RCU softirq handler happens to run your RCU callback while
  273. interrupting that acquisition's critical section.
  274. 12. RCU callbacks can be and are executed in parallel. In many cases,
  275. the callback code simply wrappers around kfree(), so that this
  276. is not an issue (or, more accurately, to the extent that it is
  277. an issue, the memory-allocator locking handles it). However,
  278. if the callbacks do manipulate a shared data structure, they
  279. must use whatever locking or other synchronization is required
  280. to safely access and/or modify that data structure.
  281. Do not assume that RCU callbacks will be executed on the same
  282. CPU that executed the corresponding call_rcu() or call_srcu().
  283. For example, if a given CPU goes offline while having an RCU
  284. callback pending, then that RCU callback will execute on some
  285. surviving CPU. (If this was not the case, a self-spawning RCU
  286. callback would prevent the victim CPU from ever going offline.)
  287. Furthermore, CPUs designated by rcu_nocbs= might well -always-
  288. have their RCU callbacks executed on some other CPUs, in fact,
  289. for some real-time workloads, this is the whole point of using
  290. the rcu_nocbs= kernel boot parameter.
  291. 13. Unlike other forms of RCU, it -is- permissible to block in an
  292. SRCU read-side critical section (demarked by srcu_read_lock()
  293. and srcu_read_unlock()), hence the "SRCU": "sleepable RCU".
  294. Please note that if you don't need to sleep in read-side critical
  295. sections, you should be using RCU rather than SRCU, because RCU
  296. is almost always faster and easier to use than is SRCU.
  297. Also unlike other forms of RCU, explicit initialization and
  298. cleanup is required either at build time via DEFINE_SRCU()
  299. or DEFINE_STATIC_SRCU() or at runtime via init_srcu_struct()
  300. and cleanup_srcu_struct(). These last two are passed a
  301. "struct srcu_struct" that defines the scope of a given
  302. SRCU domain. Once initialized, the srcu_struct is passed
  303. to srcu_read_lock(), srcu_read_unlock() synchronize_srcu(),
  304. synchronize_srcu_expedited(), and call_srcu(). A given
  305. synchronize_srcu() waits only for SRCU read-side critical
  306. sections governed by srcu_read_lock() and srcu_read_unlock()
  307. calls that have been passed the same srcu_struct. This property
  308. is what makes sleeping read-side critical sections tolerable --
  309. a given subsystem delays only its own updates, not those of other
  310. subsystems using SRCU. Therefore, SRCU is less prone to OOM the
  311. system than RCU would be if RCU's read-side critical sections
  312. were permitted to sleep.
  313. The ability to sleep in read-side critical sections does not
  314. come for free. First, corresponding srcu_read_lock() and
  315. srcu_read_unlock() calls must be passed the same srcu_struct.
  316. Second, grace-period-detection overhead is amortized only
  317. over those updates sharing a given srcu_struct, rather than
  318. being globally amortized as they are for other forms of RCU.
  319. Therefore, SRCU should be used in preference to rw_semaphore
  320. only in extremely read-intensive situations, or in situations
  321. requiring SRCU's read-side deadlock immunity or low read-side
  322. realtime latency. You should also consider percpu_rw_semaphore
  323. when you need lightweight readers.
  324. SRCU's expedited primitive (synchronize_srcu_expedited())
  325. never sends IPIs to other CPUs, so it is easier on
  326. real-time workloads than is synchronize_rcu_expedited().
  327. Note that rcu_assign_pointer() relates to SRCU just as it does to
  328. other forms of RCU, but instead of rcu_dereference() you should
  329. use srcu_dereference() in order to avoid lockdep splats.
  330. 14. The whole point of call_rcu(), synchronize_rcu(), and friends
  331. is to wait until all pre-existing readers have finished before
  332. carrying out some otherwise-destructive operation. It is
  333. therefore critically important to -first- remove any path
  334. that readers can follow that could be affected by the
  335. destructive operation, and -only- -then- invoke call_rcu(),
  336. synchronize_rcu(), or friends.
  337. Because these primitives only wait for pre-existing readers, it
  338. is the caller's responsibility to guarantee that any subsequent
  339. readers will execute safely.
  340. 15. The various RCU read-side primitives do -not- necessarily contain
  341. memory barriers. You should therefore plan for the CPU
  342. and the compiler to freely reorder code into and out of RCU
  343. read-side critical sections. It is the responsibility of the
  344. RCU update-side primitives to deal with this.
  345. For SRCU readers, you can use smp_mb__after_srcu_read_unlock()
  346. immediately after an srcu_read_unlock() to get a full barrier.
  347. 16. Use CONFIG_PROVE_LOCKING, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and the
  348. __rcu sparse checks to validate your RCU code. These can help
  349. find problems as follows:
  350. CONFIG_PROVE_LOCKING:
  351. check that accesses to RCU-protected data
  352. structures are carried out under the proper RCU
  353. read-side critical section, while holding the right
  354. combination of locks, or whatever other conditions
  355. are appropriate.
  356. CONFIG_DEBUG_OBJECTS_RCU_HEAD:
  357. check that you don't pass the
  358. same object to call_rcu() (or friends) before an RCU
  359. grace period has elapsed since the last time that you
  360. passed that same object to call_rcu() (or friends).
  361. __rcu sparse checks:
  362. tag the pointer to the RCU-protected data
  363. structure with __rcu, and sparse will warn you if you
  364. access that pointer without the services of one of the
  365. variants of rcu_dereference().
  366. These debugging aids can help you find problems that are
  367. otherwise extremely difficult to spot.
  368. 17. If you register a callback using call_rcu() or call_srcu(), and
  369. pass in a function defined within a loadable module, then it in
  370. necessary to wait for all pending callbacks to be invoked after
  371. the last invocation and before unloading that module. Note that
  372. it is absolutely -not- sufficient to wait for a grace period!
  373. The current (say) synchronize_rcu() implementation is -not-
  374. guaranteed to wait for callbacks registered on other CPUs.
  375. Or even on the current CPU if that CPU recently went offline
  376. and came back online.
  377. You instead need to use one of the barrier functions:
  378. - call_rcu() -> rcu_barrier()
  379. - call_srcu() -> srcu_barrier()
  380. However, these barrier functions are absolutely -not- guaranteed
  381. to wait for a grace period. In fact, if there are no call_rcu()
  382. callbacks waiting anywhere in the system, rcu_barrier() is within
  383. its rights to return immediately.
  384. So if you need to wait for both an RCU grace period and for
  385. all pre-existing call_rcu() callbacks, you will need to execute
  386. both rcu_barrier() and synchronize_rcu(), if necessary, using
  387. something like workqueues to to execute them concurrently.
  388. See rcubarrier.txt for more information.