/Documentation/RCU/checklist.txt

https://bitbucket.org/sammyz/iscream_thunderc-2.6.35-rebase · Plain Text · 367 lines · 304 code · 63 blank · 0 comment · 0 complexity · af4a0ce2ac70a603066c6ecdacf79494 MD5 · raw file

  1. Review Checklist for RCU Patches
  2. This document contains a checklist for producing and reviewing patches
  3. that make use of RCU. Violating any of the rules listed below will
  4. result in the same sorts of problems that leaving out a locking primitive
  5. would cause. This list is based on experiences reviewing such patches
  6. over a rather long period of time, but improvements are always welcome!
  7. 0. Is RCU being applied to a read-mostly situation? If the data
  8. structure is updated more than about 10% of the time, then you
  9. should strongly consider some other approach, unless detailed
  10. performance measurements show that RCU is nonetheless the right
  11. tool for the job. Yes, RCU does reduce read-side overhead by
  12. increasing write-side overhead, which is exactly why normal uses
  13. of RCU will do much more reading than updating.
  14. Another exception is where performance is not an issue, and RCU
  15. provides a simpler implementation. An example of this situation
  16. is the dynamic NMI code in the Linux 2.6 kernel, at least on
  17. architectures where NMIs are rare.
  18. Yet another exception is where the low real-time latency of RCU's
  19. read-side primitives is critically important.
  20. 1. Does the update code have proper mutual exclusion?
  21. RCU does allow -readers- to run (almost) naked, but -writers- must
  22. still use some sort of mutual exclusion, such as:
  23. a. locking,
  24. b. atomic operations, or
  25. c. restricting updates to a single task.
  26. If you choose #b, be prepared to describe how you have handled
  27. memory barriers on weakly ordered machines (pretty much all of
  28. them -- even x86 allows later loads to be reordered to precede
  29. earlier stores), and be prepared to explain why this added
  30. complexity is worthwhile. If you choose #c, be prepared to
  31. explain how this single task does not become a major bottleneck on
  32. big multiprocessor machines (for example, if the task is updating
  33. information relating to itself that other tasks can read, there
  34. by definition can be no bottleneck).
  35. 2. Do the RCU read-side critical sections make proper use of
  36. rcu_read_lock() and friends? These primitives are needed
  37. to prevent grace periods from ending prematurely, which
  38. could result in data being unceremoniously freed out from
  39. under your read-side code, which can greatly increase the
  40. actuarial risk of your kernel.
  41. As a rough rule of thumb, any dereference of an RCU-protected
  42. pointer must be covered by rcu_read_lock(), rcu_read_lock_bh(),
  43. rcu_read_lock_sched(), or by the appropriate update-side lock.
  44. Disabling of preemption can serve as rcu_read_lock_sched(), but
  45. is less readable.
  46. 3. Does the update code tolerate concurrent accesses?
  47. The whole point of RCU is to permit readers to run without
  48. any locks or atomic operations. This means that readers will
  49. be running while updates are in progress. There are a number
  50. of ways to handle this concurrency, depending on the situation:
  51. a. Use the RCU variants of the list and hlist update
  52. primitives to add, remove, and replace elements on
  53. an RCU-protected list. Alternatively, use the other
  54. RCU-protected data structures that have been added to
  55. the Linux kernel.
  56. This is almost always the best approach.
  57. b. Proceed as in (a) above, but also maintain per-element
  58. locks (that are acquired by both readers and writers)
  59. that guard per-element state. Of course, fields that
  60. the readers refrain from accessing can be guarded by
  61. some other lock acquired only by updaters, if desired.
  62. This works quite well, also.
  63. c. Make updates appear atomic to readers. For example,
  64. pointer updates to properly aligned fields will
  65. appear atomic, as will individual atomic primitives.
  66. Sequences of perations performed under a lock will -not-
  67. appear to be atomic to RCU readers, nor will sequences
  68. of multiple atomic primitives.
  69. This can work, but is starting to get a bit tricky.
  70. d. Carefully order the updates and the reads so that
  71. readers see valid data at all phases of the update.
  72. This is often more difficult than it sounds, especially
  73. given modern CPUs' tendency to reorder memory references.
  74. One must usually liberally sprinkle memory barriers
  75. (smp_wmb(), smp_rmb(), smp_mb()) through the code,
  76. making it difficult to understand and to test.
  77. It is usually better to group the changing data into
  78. a separate structure, so that the change may be made
  79. to appear atomic by updating a pointer to reference
  80. a new structure containing updated values.
  81. 4. Weakly ordered CPUs pose special challenges. Almost all CPUs
  82. are weakly ordered -- even x86 CPUs allow later loads to be
  83. reordered to precede earlier stores. RCU code must take all of
  84. the following measures to prevent memory-corruption problems:
  85. a. Readers must maintain proper ordering of their memory
  86. accesses. The rcu_dereference() primitive ensures that
  87. the CPU picks up the pointer before it picks up the data
  88. that the pointer points to. This really is necessary
  89. on Alpha CPUs. If you don't believe me, see:
  90. http://www.openvms.compaq.com/wizard/wiz_2637.html
  91. The rcu_dereference() primitive is also an excellent
  92. documentation aid, letting the person reading the code
  93. know exactly which pointers are protected by RCU.
  94. Please note that compilers can also reorder code, and
  95. they are becoming increasingly aggressive about doing
  96. just that. The rcu_dereference() primitive therefore
  97. also prevents destructive compiler optimizations.
  98. The rcu_dereference() primitive is used by the
  99. various "_rcu()" list-traversal primitives, such
  100. as the list_for_each_entry_rcu(). Note that it is
  101. perfectly legal (if redundant) for update-side code to
  102. use rcu_dereference() and the "_rcu()" list-traversal
  103. primitives. This is particularly useful in code that
  104. is common to readers and updaters. However, lockdep
  105. will complain if you access rcu_dereference() outside
  106. of an RCU read-side critical section. See lockdep.txt
  107. to learn what to do about this.
  108. Of course, neither rcu_dereference() nor the "_rcu()"
  109. list-traversal primitives can substitute for a good
  110. concurrency design coordinating among multiple updaters.
  111. b. If the list macros are being used, the list_add_tail_rcu()
  112. and list_add_rcu() primitives must be used in order
  113. to prevent weakly ordered machines from misordering
  114. structure initialization and pointer planting.
  115. Similarly, if the hlist macros are being used, the
  116. hlist_add_head_rcu() primitive is required.
  117. c. If the list macros are being used, the list_del_rcu()
  118. primitive must be used to keep list_del()'s pointer
  119. poisoning from inflicting toxic effects on concurrent
  120. readers. Similarly, if the hlist macros are being used,
  121. the hlist_del_rcu() primitive is required.
  122. The list_replace_rcu() and hlist_replace_rcu() primitives
  123. may be used to replace an old structure with a new one
  124. in their respective types of RCU-protected lists.
  125. d. Rules similar to (4b) and (4c) apply to the "hlist_nulls"
  126. type of RCU-protected linked lists.
  127. e. Updates must ensure that initialization of a given
  128. structure happens before pointers to that structure are
  129. publicized. Use the rcu_assign_pointer() primitive
  130. when publicizing a pointer to a structure that can
  131. be traversed by an RCU read-side critical section.
  132. 5. If call_rcu(), or a related primitive such as call_rcu_bh() or
  133. call_rcu_sched(), is used, the callback function must be
  134. written to be called from softirq context. In particular,
  135. it cannot block.
  136. 6. Since synchronize_rcu() can block, it cannot be called from
  137. any sort of irq context. The same rule applies for
  138. synchronize_rcu_bh(), synchronize_sched(), synchronize_srcu(),
  139. synchronize_rcu_expedited(), synchronize_rcu_bh_expedited(),
  140. synchronize_sched_expedite(), and synchronize_srcu_expedited().
  141. The expedited forms of these primitives have the same semantics
  142. as the non-expedited forms, but expediting is both expensive
  143. and unfriendly to real-time workloads. Use of the expedited
  144. primitives should be restricted to rare configuration-change
  145. operations that would not normally be undertaken while a real-time
  146. workload is running.
  147. 7. If the updater uses call_rcu() or synchronize_rcu(), then the
  148. corresponding readers must use rcu_read_lock() and
  149. rcu_read_unlock(). If the updater uses call_rcu_bh() or
  150. synchronize_rcu_bh(), then the corresponding readers must
  151. use rcu_read_lock_bh() and rcu_read_unlock_bh(). If the
  152. updater uses call_rcu_sched() or synchronize_sched(), then
  153. the corresponding readers must disable preemption, possibly
  154. by calling rcu_read_lock_sched() and rcu_read_unlock_sched().
  155. If the updater uses synchronize_srcu(), the the corresponding
  156. readers must use srcu_read_lock() and srcu_read_unlock(),
  157. and with the same srcu_struct. The rules for the expedited
  158. primitives are the same as for their non-expedited counterparts.
  159. Mixing things up will result in confusion and broken kernels.
  160. One exception to this rule: rcu_read_lock() and rcu_read_unlock()
  161. may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh()
  162. in cases where local bottom halves are already known to be
  163. disabled, for example, in irq or softirq context. Commenting
  164. such cases is a must, of course! And the jury is still out on
  165. whether the increased speed is worth it.
  166. 8. Although synchronize_rcu() is slower than is call_rcu(), it
  167. usually results in simpler code. So, unless update performance
  168. is critically important or the updaters cannot block,
  169. synchronize_rcu() should be used in preference to call_rcu().
  170. An especially important property of the synchronize_rcu()
  171. primitive is that it automatically self-limits: if grace periods
  172. are delayed for whatever reason, then the synchronize_rcu()
  173. primitive will correspondingly delay updates. In contrast,
  174. code using call_rcu() should explicitly limit update rate in
  175. cases where grace periods are delayed, as failing to do so can
  176. result in excessive realtime latencies or even OOM conditions.
  177. Ways of gaining this self-limiting property when using call_rcu()
  178. include:
  179. a. Keeping a count of the number of data-structure elements
  180. used by the RCU-protected data structure, including those
  181. waiting for a grace period to elapse. Enforce a limit
  182. on this number, stalling updates as needed to allow
  183. previously deferred frees to complete.
  184. Alternatively, limit only the number awaiting deferred
  185. free rather than the total number of elements.
  186. b. Limiting update rate. For example, if updates occur only
  187. once per hour, then no explicit rate limiting is required,
  188. unless your system is already badly broken. The dcache
  189. subsystem takes this approach -- updates are guarded
  190. by a global lock, limiting their rate.
  191. c. Trusted update -- if updates can only be done manually by
  192. superuser or some other trusted user, then it might not
  193. be necessary to automatically limit them. The theory
  194. here is that superuser already has lots of ways to crash
  195. the machine.
  196. d. Use call_rcu_bh() rather than call_rcu(), in order to take
  197. advantage of call_rcu_bh()'s faster grace periods.
  198. e. Periodically invoke synchronize_rcu(), permitting a limited
  199. number of updates per grace period.
  200. The same cautions apply to call_rcu_bh() and call_rcu_sched().
  201. 9. All RCU list-traversal primitives, which include
  202. rcu_dereference(), list_for_each_entry_rcu(),
  203. list_for_each_continue_rcu(), and list_for_each_safe_rcu(),
  204. must be either within an RCU read-side critical section or
  205. must be protected by appropriate update-side locks. RCU
  206. read-side critical sections are delimited by rcu_read_lock()
  207. and rcu_read_unlock(), or by similar primitives such as
  208. rcu_read_lock_bh() and rcu_read_unlock_bh(), in which case
  209. the matching rcu_dereference() primitive must be used in order
  210. to keep lockdep happy, in this case, rcu_dereference_bh().
  211. The reason that it is permissible to use RCU list-traversal
  212. primitives when the update-side lock is held is that doing so
  213. can be quite helpful in reducing code bloat when common code is
  214. shared between readers and updaters. Additional primitives
  215. are provided for this case, as discussed in lockdep.txt.
  216. 10. Conversely, if you are in an RCU read-side critical section,
  217. and you don't hold the appropriate update-side lock, you -must-
  218. use the "_rcu()" variants of the list macros. Failing to do so
  219. will break Alpha, cause aggressive compilers to generate bad code,
  220. and confuse people trying to read your code.
  221. 11. Note that synchronize_rcu() -only- guarantees to wait until
  222. all currently executing rcu_read_lock()-protected RCU read-side
  223. critical sections complete. It does -not- necessarily guarantee
  224. that all currently running interrupts, NMIs, preempt_disable()
  225. code, or idle loops will complete. Therefore, if you do not have
  226. rcu_read_lock()-protected read-side critical sections, do -not-
  227. use synchronize_rcu().
  228. Similarly, disabling preemption is not an acceptable substitute
  229. for rcu_read_lock(). Code that attempts to use preemption
  230. disabling where it should be using rcu_read_lock() will break
  231. in real-time kernel builds.
  232. If you want to wait for interrupt handlers, NMI handlers, and
  233. code under the influence of preempt_disable(), you instead
  234. need to use synchronize_irq() or synchronize_sched().
  235. 12. Any lock acquired by an RCU callback must be acquired elsewhere
  236. with softirq disabled, e.g., via spin_lock_irqsave(),
  237. spin_lock_bh(), etc. Failing to disable irq on a given
  238. acquisition of that lock will result in deadlock as soon as
  239. the RCU softirq handler happens to run your RCU callback while
  240. interrupting that acquisition's critical section.
  241. 13. RCU callbacks can be and are executed in parallel. In many cases,
  242. the callback code simply wrappers around kfree(), so that this
  243. is not an issue (or, more accurately, to the extent that it is
  244. an issue, the memory-allocator locking handles it). However,
  245. if the callbacks do manipulate a shared data structure, they
  246. must use whatever locking or other synchronization is required
  247. to safely access and/or modify that data structure.
  248. RCU callbacks are -usually- executed on the same CPU that executed
  249. the corresponding call_rcu(), call_rcu_bh(), or call_rcu_sched(),
  250. but are by -no- means guaranteed to be. For example, if a given
  251. CPU goes offline while having an RCU callback pending, then that
  252. RCU callback will execute on some surviving CPU. (If this was
  253. not the case, a self-spawning RCU callback would prevent the
  254. victim CPU from ever going offline.)
  255. 14. SRCU (srcu_read_lock(), srcu_read_unlock(), srcu_dereference(),
  256. synchronize_srcu(), and synchronize_srcu_expedited()) may only
  257. be invoked from process context. Unlike other forms of RCU, it
  258. -is- permissible to block in an SRCU read-side critical section
  259. (demarked by srcu_read_lock() and srcu_read_unlock()), hence the
  260. "SRCU": "sleepable RCU". Please note that if you don't need
  261. to sleep in read-side critical sections, you should be using
  262. RCU rather than SRCU, because RCU is almost always faster and
  263. easier to use than is SRCU.
  264. Also unlike other forms of RCU, explicit initialization
  265. and cleanup is required via init_srcu_struct() and
  266. cleanup_srcu_struct(). These are passed a "struct srcu_struct"
  267. that defines the scope of a given SRCU domain. Once initialized,
  268. the srcu_struct is passed to srcu_read_lock(), srcu_read_unlock()
  269. synchronize_srcu(), and synchronize_srcu_expedited(). A given
  270. synchronize_srcu() waits only for SRCU read-side critical
  271. sections governed by srcu_read_lock() and srcu_read_unlock()
  272. calls that have been passed the same srcu_struct. This property
  273. is what makes sleeping read-side critical sections tolerable --
  274. a given subsystem delays only its own updates, not those of other
  275. subsystems using SRCU. Therefore, SRCU is less prone to OOM the
  276. system than RCU would be if RCU's read-side critical sections
  277. were permitted to sleep.
  278. The ability to sleep in read-side critical sections does not
  279. come for free. First, corresponding srcu_read_lock() and
  280. srcu_read_unlock() calls must be passed the same srcu_struct.
  281. Second, grace-period-detection overhead is amortized only
  282. over those updates sharing a given srcu_struct, rather than
  283. being globally amortized as they are for other forms of RCU.
  284. Therefore, SRCU should be used in preference to rw_semaphore
  285. only in extremely read-intensive situations, or in situations
  286. requiring SRCU's read-side deadlock immunity or low read-side
  287. realtime latency.
  288. Note that, rcu_assign_pointer() relates to SRCU just as they do
  289. to other forms of RCU.
  290. 15. The whole point of call_rcu(), synchronize_rcu(), and friends
  291. is to wait until all pre-existing readers have finished before
  292. carrying out some otherwise-destructive operation. It is
  293. therefore critically important to -first- remove any path
  294. that readers can follow that could be affected by the
  295. destructive operation, and -only- -then- invoke call_rcu(),
  296. synchronize_rcu(), or friends.
  297. Because these primitives only wait for pre-existing readers, it
  298. is the caller's responsibility to guarantee that any subsequent
  299. readers will execute safely.
  300. 16. The various RCU read-side primitives do -not- necessarily contain
  301. memory barriers. You should therefore plan for the CPU
  302. and the compiler to freely reorder code into and out of RCU
  303. read-side critical sections. It is the responsibility of the
  304. RCU update-side primitives to deal with this.