checklist.txt 12 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261
  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
  9. you should strongly consider some other approach, unless
  10. detailed performance measurements show that RCU is nonetheless
  11. the right tool for the job.
  12. The other exception would be where performance is not an issue,
  13. and RCU provides a simpler implementation. An example of this
  14. situation is the dynamic NMI code in the Linux 2.6 kernel,
  15. at least on architectures where NMIs are rare.
  16. 1. Does the update code have proper mutual exclusion?
  17. RCU does allow -readers- to run (almost) naked, but -writers- must
  18. still use some sort of mutual exclusion, such as:
  19. a. locking,
  20. b. atomic operations, or
  21. c. restricting updates to a single task.
  22. If you choose #b, be prepared to describe how you have handled
  23. memory barriers on weakly ordered machines (pretty much all of
  24. them -- even x86 allows reads to be reordered), and be prepared
  25. to explain why this added complexity is worthwhile. If you
  26. choose #c, be prepared to explain how this single task does not
  27. become a major bottleneck on big multiprocessor machines (for
  28. example, if the task is updating information relating to itself
  29. that other tasks can read, there by definition can be no
  30. bottleneck).
  31. 2. Do the RCU read-side critical sections make proper use of
  32. rcu_read_lock() and friends? These primitives are needed
  33. to suppress preemption (or bottom halves, in the case of
  34. rcu_read_lock_bh()) in the read-side critical sections,
  35. and are also an excellent aid to readability.
  36. As a rough rule of thumb, any dereference of an RCU-protected
  37. pointer must be covered by rcu_read_lock() or rcu_read_lock_bh()
  38. or by the appropriate update-side lock.
  39. 3. Does the update code tolerate concurrent accesses?
  40. The whole point of RCU is to permit readers to run without
  41. any locks or atomic operations. This means that readers will
  42. be running while updates are in progress. There are a number
  43. of ways to handle this concurrency, depending on the situation:
  44. a. Make updates appear atomic to readers. For example,
  45. pointer updates to properly aligned fields will appear
  46. atomic, as will individual atomic primitives. Operations
  47. performed under a lock and sequences of multiple atomic
  48. primitives will -not- appear to be atomic.
  49. This is almost always the best approach.
  50. b. Carefully order the updates and the reads so that
  51. readers see valid data at all phases of the update.
  52. This is often more difficult than it sounds, especially
  53. given modern CPUs' tendency to reorder memory references.
  54. One must usually liberally sprinkle memory barriers
  55. (smp_wmb(), smp_rmb(), smp_mb()) through the code,
  56. making it difficult to understand and to test.
  57. It is usually better to group the changing data into
  58. a separate structure, so that the change may be made
  59. to appear atomic by updating a pointer to reference
  60. a new structure containing updated values.
  61. 4. Weakly ordered CPUs pose special challenges. Almost all CPUs
  62. are weakly ordered -- even i386 CPUs allow reads to be reordered.
  63. RCU code must take all of the following measures to prevent
  64. memory-corruption problems:
  65. a. Readers must maintain proper ordering of their memory
  66. accesses. The rcu_dereference() primitive ensures that
  67. the CPU picks up the pointer before it picks up the data
  68. that the pointer points to. This really is necessary
  69. on Alpha CPUs. If you don't believe me, see:
  70. http://www.openvms.compaq.com/wizard/wiz_2637.html
  71. The rcu_dereference() primitive is also an excellent
  72. documentation aid, letting the person reading the code
  73. know exactly which pointers are protected by RCU.
  74. The rcu_dereference() primitive is used by the various
  75. "_rcu()" list-traversal primitives, such as the
  76. list_for_each_entry_rcu(). Note that it is perfectly
  77. legal (if redundant) for update-side code to use
  78. rcu_dereference() and the "_rcu()" list-traversal
  79. primitives. This is particularly useful in code
  80. that is common to readers and updaters.
  81. b. If the list macros are being used, the list_add_tail_rcu()
  82. and list_add_rcu() primitives must be used in order
  83. to prevent weakly ordered machines from misordering
  84. structure initialization and pointer planting.
  85. Similarly, if the hlist macros are being used, the
  86. hlist_add_head_rcu() primitive is required.
  87. c. If the list macros are being used, the list_del_rcu()
  88. primitive must be used to keep list_del()'s pointer
  89. poisoning from inflicting toxic effects on concurrent
  90. readers. Similarly, if the hlist macros are being used,
  91. the hlist_del_rcu() primitive is required.
  92. The list_replace_rcu() primitive may be used to
  93. replace an old structure with a new one in an
  94. RCU-protected list.
  95. d. Updates must ensure that initialization of a given
  96. structure happens before pointers to that structure are
  97. publicized. Use the rcu_assign_pointer() primitive
  98. when publicizing a pointer to a structure that can
  99. be traversed by an RCU read-side critical section.
  100. 5. If call_rcu(), or a related primitive such as call_rcu_bh(),
  101. is used, the callback function must be written to be called
  102. from softirq context. In particular, it cannot block.
  103. 6. Since synchronize_rcu() can block, it cannot be called from
  104. any sort of irq context.
  105. 7. If the updater uses call_rcu(), then the corresponding readers
  106. must use rcu_read_lock() and rcu_read_unlock(). If the updater
  107. uses call_rcu_bh(), then the corresponding readers must use
  108. rcu_read_lock_bh() and rcu_read_unlock_bh(). Mixing things up
  109. will result in confusion and broken kernels.
  110. One exception to this rule: rcu_read_lock() and rcu_read_unlock()
  111. may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh()
  112. in cases where local bottom halves are already known to be
  113. disabled, for example, in irq or softirq context. Commenting
  114. such cases is a must, of course! And the jury is still out on
  115. whether the increased speed is worth it.
  116. 8. Although synchronize_rcu() is a bit slower than is call_rcu(),
  117. it usually results in simpler code. So, unless update
  118. performance is critically important or the updaters cannot block,
  119. synchronize_rcu() should be used in preference to call_rcu().
  120. An especially important property of the synchronize_rcu()
  121. primitive is that it automatically self-limits: if grace periods
  122. are delayed for whatever reason, then the synchronize_rcu()
  123. primitive will correspondingly delay updates. In contrast,
  124. code using call_rcu() should explicitly limit update rate in
  125. cases where grace periods are delayed, as failing to do so can
  126. result in excessive realtime latencies or even OOM conditions.
  127. Ways of gaining this self-limiting property when using call_rcu()
  128. include:
  129. a. Keeping a count of the number of data-structure elements
  130. used by the RCU-protected data structure, including those
  131. waiting for a grace period to elapse. Enforce a limit
  132. on this number, stalling updates as needed to allow
  133. previously deferred frees to complete.
  134. Alternatively, limit only the number awaiting deferred
  135. free rather than the total number of elements.
  136. b. Limiting update rate. For example, if updates occur only
  137. once per hour, then no explicit rate limiting is required,
  138. unless your system is already badly broken. The dcache
  139. subsystem takes this approach -- updates are guarded
  140. by a global lock, limiting their rate.
  141. c. Trusted update -- if updates can only be done manually by
  142. superuser or some other trusted user, then it might not
  143. be necessary to automatically limit them. The theory
  144. here is that superuser already has lots of ways to crash
  145. the machine.
  146. d. Use call_rcu_bh() rather than call_rcu(), in order to take
  147. advantage of call_rcu_bh()'s faster grace periods.
  148. e. Periodically invoke synchronize_rcu(), permitting a limited
  149. number of updates per grace period.
  150. 9. All RCU list-traversal primitives, which include
  151. list_for_each_rcu(), list_for_each_entry_rcu(),
  152. list_for_each_continue_rcu(), and list_for_each_safe_rcu(),
  153. must be within an RCU read-side critical section. RCU
  154. read-side critical sections are delimited by rcu_read_lock()
  155. and rcu_read_unlock(), or by similar primitives such as
  156. rcu_read_lock_bh() and rcu_read_unlock_bh().
  157. Use of the _rcu() list-traversal primitives outside of an
  158. RCU read-side critical section causes no harm other than
  159. a slight performance degradation on Alpha CPUs. It can
  160. also be quite helpful in reducing code bloat when common
  161. code is shared between readers and updaters.
  162. 10. Conversely, if you are in an RCU read-side critical section,
  163. you -must- use the "_rcu()" variants of the list macros.
  164. Failing to do so will break Alpha and confuse people reading
  165. your code.
  166. 11. Note that synchronize_rcu() -only- guarantees to wait until
  167. all currently executing rcu_read_lock()-protected RCU read-side
  168. critical sections complete. It does -not- necessarily guarantee
  169. that all currently running interrupts, NMIs, preempt_disable()
  170. code, or idle loops will complete. Therefore, if you do not have
  171. rcu_read_lock()-protected read-side critical sections, do -not-
  172. use synchronize_rcu().
  173. If you want to wait for some of these other things, you might
  174. instead need to use synchronize_irq() or synchronize_sched().
  175. 12. Any lock acquired by an RCU callback must be acquired elsewhere
  176. with irq disabled, e.g., via spin_lock_irqsave(). Failing to
  177. disable irq on a given acquisition of that lock will result in
  178. deadlock as soon as the RCU callback happens to interrupt that
  179. acquisition's critical section.
  180. 13. SRCU (srcu_read_lock(), srcu_read_unlock(), and synchronize_srcu())
  181. may only be invoked from process context. Unlike other forms of
  182. RCU, it -is- permissible to block in an SRCU read-side critical
  183. section (demarked by srcu_read_lock() and srcu_read_unlock()),
  184. hence the "SRCU": "sleepable RCU". Please note that if you
  185. don't need to sleep in read-side critical sections, you should
  186. be using RCU rather than SRCU, because RCU is almost always
  187. faster and easier to use than is SRCU.
  188. Also unlike other forms of RCU, explicit initialization
  189. and cleanup is required via init_srcu_struct() and
  190. cleanup_srcu_struct(). These are passed a "struct srcu_struct"
  191. that defines the scope of a given SRCU domain. Once initialized,
  192. the srcu_struct is passed to srcu_read_lock(), srcu_read_unlock()
  193. and synchronize_srcu(). A given synchronize_srcu() waits only
  194. for SRCU read-side critical sections governed by srcu_read_lock()
  195. and srcu_read_unlock() calls that have been passd the same
  196. srcu_struct. This property is what makes sleeping read-side
  197. critical sections tolerable -- a given subsystem delays only
  198. its own updates, not those of other subsystems using SRCU.
  199. Therefore, SRCU is less prone to OOM the system than RCU would
  200. be if RCU's read-side critical sections were permitted to
  201. sleep.
  202. The ability to sleep in read-side critical sections does not
  203. come for free. First, corresponding srcu_read_lock() and
  204. srcu_read_unlock() calls must be passed the same srcu_struct.
  205. Second, grace-period-detection overhead is amortized only
  206. over those updates sharing a given srcu_struct, rather than
  207. being globally amortized as they are for other forms of RCU.
  208. Therefore, SRCU should be used in preference to rw_semaphore
  209. only in extremely read-intensive situations, or in situations
  210. requiring SRCU's read-side deadlock immunity or low read-side
  211. realtime latency.
  212. Note that, rcu_assign_pointer() and rcu_dereference() relate to
  213. SRCU just as they do to other forms of RCU.