blob: 5c8d74968090544ae570105b00a96561ea4acd6c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001Review Checklist for RCU Patches
2
3
4This document contains a checklist for producing and reviewing patches
5that make use of RCU. Violating any of the rules listed below will
6result in the same sorts of problems that leaving out a locking primitive
7would cause. This list is based on experiences reviewing such patches
8over a rather long period of time, but improvements are always welcome!
9
100. Is RCU being applied to a read-mostly situation? If the data
Paul E. McKenney4c540052010-01-14 16:10:57 -080011 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.
Linus Torvalds1da177e2005-04-16 15:20:36 -070017
Paul E. McKenney32300752008-05-12 21:21:05 +020018 Another exception is where performance is not an issue, and RCU
19 provides a simpler implementation. An example of this situation
20 is the dynamic NMI code in the Linux 2.6 kernel, at least on
21 architectures where NMIs are rare.
22
23 Yet another exception is where the low real-time latency of RCU's
24 read-side primitives is critically important.
Linus Torvalds1da177e2005-04-16 15:20:36 -070025
261. Does the update code have proper mutual exclusion?
27
28 RCU does allow -readers- to run (almost) naked, but -writers- must
29 still use some sort of mutual exclusion, such as:
30
31 a. locking,
32 b. atomic operations, or
33 c. restricting updates to a single task.
34
35 If you choose #b, be prepared to describe how you have handled
36 memory barriers on weakly ordered machines (pretty much all of
Paul E. McKenney4c540052010-01-14 16:10:57 -080037 them -- even x86 allows later loads to be reordered to precede
38 earlier stores), and be prepared to explain why this added
39 complexity is worthwhile. If you choose #c, be prepared to
40 explain how this single task does not become a major bottleneck on
41 big multiprocessor machines (for example, if the task is updating
42 information relating to itself that other tasks can read, there
43 by definition can be no bottleneck).
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
452. Do the RCU read-side critical sections make proper use of
46 rcu_read_lock() and friends? These primitives are needed
Paul E. McKenney32300752008-05-12 21:21:05 +020047 to prevent grace periods from ending prematurely, which
48 could result in data being unceremoniously freed out from
49 under your read-side code, which can greatly increase the
50 actuarial risk of your kernel.
Linus Torvalds1da177e2005-04-16 15:20:36 -070051
Paul E. McKenneydd81eca2005-09-10 00:26:24 -070052 As a rough rule of thumb, any dereference of an RCU-protected
Paul E. McKenney4c540052010-01-14 16:10:57 -080053 pointer must be covered by rcu_read_lock(), rcu_read_lock_bh(),
54 rcu_read_lock_sched(), or by the appropriate update-side lock.
55 Disabling of preemption can serve as rcu_read_lock_sched(), but
56 is less readable.
Paul E. McKenneydd81eca2005-09-10 00:26:24 -070057
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583. Does the update code tolerate concurrent accesses?
59
60 The whole point of RCU is to permit readers to run without
61 any locks or atomic operations. This means that readers will
62 be running while updates are in progress. There are a number
63 of ways to handle this concurrency, depending on the situation:
64
Paul E. McKenney32300752008-05-12 21:21:05 +020065 a. Use the RCU variants of the list and hlist update
Paul E. McKenney4c540052010-01-14 16:10:57 -080066 primitives to add, remove, and replace elements on
67 an RCU-protected list. Alternatively, use the other
68 RCU-protected data structures that have been added to
69 the Linux kernel.
Paul E. McKenney32300752008-05-12 21:21:05 +020070
71 This is almost always the best approach.
72
73 b. Proceed as in (a) above, but also maintain per-element
74 locks (that are acquired by both readers and writers)
75 that guard per-element state. Of course, fields that
Paul E. McKenney4c540052010-01-14 16:10:57 -080076 the readers refrain from accessing can be guarded by
77 some other lock acquired only by updaters, if desired.
Paul E. McKenney32300752008-05-12 21:21:05 +020078
79 This works quite well, also.
80
81 c. Make updates appear atomic to readers. For example,
Paul E. McKenney4c540052010-01-14 16:10:57 -080082 pointer updates to properly aligned fields will
83 appear atomic, as will individual atomic primitives.
84 Sequences of perations performed under a lock will -not-
85 appear to be atomic to RCU readers, nor will sequences
86 of multiple atomic primitives.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
Paul E. McKenney32300752008-05-12 21:21:05 +020088 This can work, but is starting to get a bit tricky.
Linus Torvalds1da177e2005-04-16 15:20:36 -070089
Paul E. McKenney32300752008-05-12 21:21:05 +020090 d. Carefully order the updates and the reads so that
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 readers see valid data at all phases of the update.
92 This is often more difficult than it sounds, especially
93 given modern CPUs' tendency to reorder memory references.
94 One must usually liberally sprinkle memory barriers
95 (smp_wmb(), smp_rmb(), smp_mb()) through the code,
96 making it difficult to understand and to test.
97
98 It is usually better to group the changing data into
99 a separate structure, so that the change may be made
100 to appear atomic by updating a pointer to reference
101 a new structure containing updated values.
102
1034. Weakly ordered CPUs pose special challenges. Almost all CPUs
Paul E. McKenney4c540052010-01-14 16:10:57 -0800104 are weakly ordered -- even x86 CPUs allow later loads to be
105 reordered to precede earlier stores. RCU code must take all of
106 the following measures to prevent memory-corruption problems:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700107
108 a. Readers must maintain proper ordering of their memory
109 accesses. The rcu_dereference() primitive ensures that
110 the CPU picks up the pointer before it picks up the data
111 that the pointer points to. This really is necessary
112 on Alpha CPUs. If you don't believe me, see:
113
114 http://www.openvms.compaq.com/wizard/wiz_2637.html
115
116 The rcu_dereference() primitive is also an excellent
117 documentation aid, letting the person reading the code
118 know exactly which pointers are protected by RCU.
Paul E. McKenney4c540052010-01-14 16:10:57 -0800119 Please note that compilers can also reorder code, and
120 they are becoming increasingly aggressive about doing
121 just that. The rcu_dereference() primitive therefore
122 also prevents destructive compiler optimizations.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
Paul E. McKenney4c540052010-01-14 16:10:57 -0800124 The rcu_dereference() primitive is used by the
125 various "_rcu()" list-traversal primitives, such
126 as the list_for_each_entry_rcu(). Note that it is
127 perfectly legal (if redundant) for update-side code to
128 use rcu_dereference() and the "_rcu()" list-traversal
129 primitives. This is particularly useful in code that
Paul E. McKenneyc598a072010-02-22 17:04:57 -0800130 is common to readers and updaters. However, lockdep
131 will complain if you access rcu_dereference() outside
132 of an RCU read-side critical section. See lockdep.txt
133 to learn what to do about this.
134
135 Of course, neither rcu_dereference() nor the "_rcu()"
136 list-traversal primitives can substitute for a good
137 concurrency design coordinating among multiple updaters.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138
Paul E. McKenneya83f1fe2005-05-01 08:59:05 -0700139 b. If the list macros are being used, the list_add_tail_rcu()
140 and list_add_rcu() primitives must be used in order
141 to prevent weakly ordered machines from misordering
142 structure initialization and pointer planting.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 Similarly, if the hlist macros are being used, the
Paul E. McKenneya83f1fe2005-05-01 08:59:05 -0700144 hlist_add_head_rcu() primitive is required.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
Paul E. McKenneya83f1fe2005-05-01 08:59:05 -0700146 c. If the list macros are being used, the list_del_rcu()
147 primitive must be used to keep list_del()'s pointer
148 poisoning from inflicting toxic effects on concurrent
149 readers. Similarly, if the hlist macros are being used,
150 the hlist_del_rcu() primitive is required.
151
Paul E. McKenney4c540052010-01-14 16:10:57 -0800152 The list_replace_rcu() and hlist_replace_rcu() primitives
153 may be used to replace an old structure with a new one
154 in their respective types of RCU-protected lists.
Paul E. McKenneya83f1fe2005-05-01 08:59:05 -0700155
Paul E. McKenney4c540052010-01-14 16:10:57 -0800156 d. Rules similar to (4b) and (4c) apply to the "hlist_nulls"
157 type of RCU-protected linked lists.
158
159 e. Updates must ensure that initialization of a given
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160 structure happens before pointers to that structure are
161 publicized. Use the rcu_assign_pointer() primitive
162 when publicizing a pointer to a structure that can
163 be traversed by an RCU read-side critical section.
164
Paul E. McKenney32300752008-05-12 21:21:05 +02001655. If call_rcu(), or a related primitive such as call_rcu_bh() or
166 call_rcu_sched(), is used, the callback function must be
167 written to be called from softirq context. In particular,
168 it cannot block.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
Paul E. McKenneya83f1fe2005-05-01 08:59:05 -07001706. Since synchronize_rcu() can block, it cannot be called from
Paul E. McKenney4c540052010-01-14 16:10:57 -0800171 any sort of irq context. The same rule applies for
172 synchronize_rcu_bh(), synchronize_sched(), synchronize_srcu(),
173 synchronize_rcu_expedited(), synchronize_rcu_bh_expedited(),
174 synchronize_sched_expedite(), and synchronize_srcu_expedited().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175
Paul E. McKenney4c540052010-01-14 16:10:57 -0800176 The expedited forms of these primitives have the same semantics
177 as the non-expedited forms, but expediting is both expensive
178 and unfriendly to real-time workloads. Use of the expedited
179 primitives should be restricted to rare configuration-change
180 operations that would not normally be undertaken while a real-time
181 workload is running.
182
Paul E. McKenney236fefa2012-01-31 14:00:41 -0800183 In particular, if you find yourself invoking one of the expedited
184 primitives repeatedly in a loop, please do everyone a favor:
185 Restructure your code so that it batches the updates, allowing
186 a single non-expedited primitive to cover the entire batch.
187 This will very likely be faster than the loop containing the
188 expedited primitive, and will be much much easier on the rest
189 of the system, especially to real-time workloads running on
190 the rest of the system.
191
192 In addition, it is illegal to call the expedited forms from
193 a CPU-hotplug notifier, or while holding a lock that is acquired
194 by a CPU-hotplug notifier. Failing to observe this restriction
195 will result in deadlock.
196
Paul E. McKenney4c540052010-01-14 16:10:57 -08001977. If the updater uses call_rcu() or synchronize_rcu(), then the
198 corresponding readers must use rcu_read_lock() and
199 rcu_read_unlock(). If the updater uses call_rcu_bh() or
200 synchronize_rcu_bh(), then the corresponding readers must
201 use rcu_read_lock_bh() and rcu_read_unlock_bh(). If the
202 updater uses call_rcu_sched() or synchronize_sched(), then
203 the corresponding readers must disable preemption, possibly
204 by calling rcu_read_lock_sched() and rcu_read_unlock_sched().
205 If the updater uses synchronize_srcu(), the the corresponding
206 readers must use srcu_read_lock() and srcu_read_unlock(),
207 and with the same srcu_struct. The rules for the expedited
208 primitives are the same as for their non-expedited counterparts.
209 Mixing things up will result in confusion and broken kernels.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210
211 One exception to this rule: rcu_read_lock() and rcu_read_unlock()
212 may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh()
213 in cases where local bottom halves are already known to be
214 disabled, for example, in irq or softirq context. Commenting
215 such cases is a must, of course! And the jury is still out on
216 whether the increased speed is worth it.
217
Paul E. McKenney32300752008-05-12 21:21:05 +02002188. Although synchronize_rcu() is slower than is call_rcu(), it
219 usually results in simpler code. So, unless update performance
220 is critically important or the updaters cannot block,
Paul E. McKenney165d6c72006-06-25 05:48:44 -0700221 synchronize_rcu() should be used in preference to call_rcu().
222
223 An especially important property of the synchronize_rcu()
224 primitive is that it automatically self-limits: if grace periods
225 are delayed for whatever reason, then the synchronize_rcu()
226 primitive will correspondingly delay updates. In contrast,
227 code using call_rcu() should explicitly limit update rate in
228 cases where grace periods are delayed, as failing to do so can
229 result in excessive realtime latencies or even OOM conditions.
230
231 Ways of gaining this self-limiting property when using call_rcu()
232 include:
233
234 a. Keeping a count of the number of data-structure elements
Paul E. McKenney5cc65172010-08-13 16:34:22 -0700235 used by the RCU-protected data structure, including
236 those waiting for a grace period to elapse. Enforce a
237 limit on this number, stalling updates as needed to allow
238 previously deferred frees to complete. Alternatively,
239 limit only the number awaiting deferred free rather than
240 the total number of elements.
Paul E. McKenney165d6c72006-06-25 05:48:44 -0700241
Paul E. McKenney5cc65172010-08-13 16:34:22 -0700242 One way to stall the updates is to acquire the update-side
243 mutex. (Don't try this with a spinlock -- other CPUs
244 spinning on the lock could prevent the grace period
245 from ever ending.) Another way to stall the updates
246 is for the updates to use a wrapper function around
247 the memory allocator, so that this wrapper function
248 simulates OOM when there is too much memory awaiting an
249 RCU grace period. There are of course many other
250 variations on this theme.
Paul E. McKenney165d6c72006-06-25 05:48:44 -0700251
252 b. Limiting update rate. For example, if updates occur only
253 once per hour, then no explicit rate limiting is required,
254 unless your system is already badly broken. The dcache
255 subsystem takes this approach -- updates are guarded
256 by a global lock, limiting their rate.
257
258 c. Trusted update -- if updates can only be done manually by
259 superuser or some other trusted user, then it might not
260 be necessary to automatically limit them. The theory
261 here is that superuser already has lots of ways to crash
262 the machine.
263
264 d. Use call_rcu_bh() rather than call_rcu(), in order to take
265 advantage of call_rcu_bh()'s faster grace periods.
266
267 e. Periodically invoke synchronize_rcu(), permitting a limited
268 number of updates per grace period.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269
Paul E. McKenney4c540052010-01-14 16:10:57 -0800270 The same cautions apply to call_rcu_bh() and call_rcu_sched().
271
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729. All RCU list-traversal primitives, which include
Paul E. McKenney34d7c2b2008-08-01 14:11:05 -0700273 rcu_dereference(), list_for_each_entry_rcu(),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 list_for_each_continue_rcu(), and list_for_each_safe_rcu(),
Paul E. McKenney32300752008-05-12 21:21:05 +0200275 must be either within an RCU read-side critical section or
276 must be protected by appropriate update-side locks. RCU
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277 read-side critical sections are delimited by rcu_read_lock()
278 and rcu_read_unlock(), or by similar primitives such as
Paul E. McKenneyc598a072010-02-22 17:04:57 -0800279 rcu_read_lock_bh() and rcu_read_unlock_bh(), in which case
280 the matching rcu_dereference() primitive must be used in order
281 to keep lockdep happy, in this case, rcu_dereference_bh().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282
Paul E. McKenney32300752008-05-12 21:21:05 +0200283 The reason that it is permissible to use RCU list-traversal
284 primitives when the update-side lock is held is that doing so
285 can be quite helpful in reducing code bloat when common code is
Paul E. McKenney50aec002010-04-09 15:39:12 -0700286 shared between readers and updaters. Additional primitives
287 are provided for this case, as discussed in lockdep.txt.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288
28910. Conversely, if you are in an RCU read-side critical section,
Paul E. McKenney32300752008-05-12 21:21:05 +0200290 and you don't hold the appropriate update-side lock, you -must-
291 use the "_rcu()" variants of the list macros. Failing to do so
Paul E. McKenney4c540052010-01-14 16:10:57 -0800292 will break Alpha, cause aggressive compilers to generate bad code,
293 and confuse people trying to read your code.
Paul E. McKenneya83f1fe2005-05-01 08:59:05 -0700294
29511. Note that synchronize_rcu() -only- guarantees to wait until
296 all currently executing rcu_read_lock()-protected RCU read-side
297 critical sections complete. It does -not- necessarily guarantee
298 that all currently running interrupts, NMIs, preempt_disable()
299 code, or idle loops will complete. Therefore, if you do not have
300 rcu_read_lock()-protected read-side critical sections, do -not-
301 use synchronize_rcu().
302
Paul E. McKenney4c540052010-01-14 16:10:57 -0800303 Similarly, disabling preemption is not an acceptable substitute
304 for rcu_read_lock(). Code that attempts to use preemption
305 disabling where it should be using rcu_read_lock() will break
306 in real-time kernel builds.
307
308 If you want to wait for interrupt handlers, NMI handlers, and
309 code under the influence of preempt_disable(), you instead
310 need to use synchronize_irq() or synchronize_sched().
Paul E. McKenneyd19720a2006-02-01 03:06:42 -0800311
31212. Any lock acquired by an RCU callback must be acquired elsewhere
Paul E. McKenney240ebbf2009-06-25 09:08:18 -0700313 with softirq disabled, e.g., via spin_lock_irqsave(),
314 spin_lock_bh(), etc. Failing to disable irq on a given
Paul E. McKenney4c540052010-01-14 16:10:57 -0800315 acquisition of that lock will result in deadlock as soon as
316 the RCU softirq handler happens to run your RCU callback while
317 interrupting that acquisition's critical section.
Paul E. McKenney621934e2006-10-04 02:17:02 -0700318
Paul E. McKenneyef48bd22007-07-15 23:41:03 -070031913. RCU callbacks can be and are executed in parallel. In many cases,
320 the callback code simply wrappers around kfree(), so that this
321 is not an issue (or, more accurately, to the extent that it is
322 an issue, the memory-allocator locking handles it). However,
323 if the callbacks do manipulate a shared data structure, they
324 must use whatever locking or other synchronization is required
325 to safely access and/or modify that data structure.
326
Paul E. McKenney32300752008-05-12 21:21:05 +0200327 RCU callbacks are -usually- executed on the same CPU that executed
328 the corresponding call_rcu(), call_rcu_bh(), or call_rcu_sched(),
329 but are by -no- means guaranteed to be. For example, if a given
330 CPU goes offline while having an RCU callback pending, then that
331 RCU callback will execute on some surviving CPU. (If this was
332 not the case, a self-spawning RCU callback would prevent the
333 victim CPU from ever going offline.)
334
Paul E. McKenneyc598a072010-02-22 17:04:57 -080033514. SRCU (srcu_read_lock(), srcu_read_unlock(), srcu_dereference(),
336 synchronize_srcu(), and synchronize_srcu_expedited()) may only
337 be invoked from process context. Unlike other forms of RCU, it
338 -is- permissible to block in an SRCU read-side critical section
339 (demarked by srcu_read_lock() and srcu_read_unlock()), hence the
340 "SRCU": "sleepable RCU". Please note that if you don't need
341 to sleep in read-side critical sections, you should be using
342 RCU rather than SRCU, because RCU is almost always faster and
343 easier to use than is SRCU.
Paul E. McKenney621934e2006-10-04 02:17:02 -0700344
Paul E. McKenney9ceae0e2011-11-03 13:43:24 -0700345 If you need to enter your read-side critical section in a
346 hardirq or exception handler, and then exit that same read-side
347 critical section in the task that was interrupted, then you need
348 to srcu_read_lock_raw() and srcu_read_unlock_raw(), which avoid
349 the lockdep checking that would otherwise this practice illegal.
350
Paul E. McKenney621934e2006-10-04 02:17:02 -0700351 Also unlike other forms of RCU, explicit initialization
352 and cleanup is required via init_srcu_struct() and
353 cleanup_srcu_struct(). These are passed a "struct srcu_struct"
354 that defines the scope of a given SRCU domain. Once initialized,
355 the srcu_struct is passed to srcu_read_lock(), srcu_read_unlock()
Paul E. McKenney4c540052010-01-14 16:10:57 -0800356 synchronize_srcu(), and synchronize_srcu_expedited(). A given
357 synchronize_srcu() waits only for SRCU read-side critical
358 sections governed by srcu_read_lock() and srcu_read_unlock()
359 calls that have been passed the same srcu_struct. This property
360 is what makes sleeping read-side critical sections tolerable --
361 a given subsystem delays only its own updates, not those of other
362 subsystems using SRCU. Therefore, SRCU is less prone to OOM the
363 system than RCU would be if RCU's read-side critical sections
364 were permitted to sleep.
Paul E. McKenney621934e2006-10-04 02:17:02 -0700365
366 The ability to sleep in read-side critical sections does not
367 come for free. First, corresponding srcu_read_lock() and
368 srcu_read_unlock() calls must be passed the same srcu_struct.
369 Second, grace-period-detection overhead is amortized only
370 over those updates sharing a given srcu_struct, rather than
371 being globally amortized as they are for other forms of RCU.
372 Therefore, SRCU should be used in preference to rw_semaphore
373 only in extremely read-intensive situations, or in situations
374 requiring SRCU's read-side deadlock immunity or low read-side
375 realtime latency.
376
Paul E. McKenney50aec002010-04-09 15:39:12 -0700377 Note that, rcu_assign_pointer() relates to SRCU just as they do
378 to other forms of RCU.
Paul E. McKenney0612ea02009-03-10 12:55:57 -0700379
38015. The whole point of call_rcu(), synchronize_rcu(), and friends
381 is to wait until all pre-existing readers have finished before
382 carrying out some otherwise-destructive operation. It is
383 therefore critically important to -first- remove any path
384 that readers can follow that could be affected by the
385 destructive operation, and -only- -then- invoke call_rcu(),
386 synchronize_rcu(), or friends.
387
Paul E. McKenney4c540052010-01-14 16:10:57 -0800388 Because these primitives only wait for pre-existing readers, it
389 is the caller's responsibility to guarantee that any subsequent
390 readers will execute safely.
Paul E. McKenney240ebbf2009-06-25 09:08:18 -0700391
Paul E. McKenney4c540052010-01-14 16:10:57 -080039216. The various RCU read-side primitives do -not- necessarily contain
393 memory barriers. You should therefore plan for the CPU
394 and the compiler to freely reorder code into and out of RCU
395 read-side critical sections. It is the responsibility of the
396 RCU update-side primitives to deal with this.
Paul E. McKenney84483ea2010-06-16 16:48:13 -0700397
39817. Use CONFIG_PROVE_RCU, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and
399 the __rcu sparse checks to validate your RCU code. These
400 can help find problems as follows:
401
402 CONFIG_PROVE_RCU: check that accesses to RCU-protected data
403 structures are carried out under the proper RCU
404 read-side critical section, while holding the right
405 combination of locks, or whatever other conditions
406 are appropriate.
407
408 CONFIG_DEBUG_OBJECTS_RCU_HEAD: check that you don't pass the
409 same object to call_rcu() (or friends) before an RCU
410 grace period has elapsed since the last time that you
411 passed that same object to call_rcu() (or friends).
412
413 __rcu sparse checks: tag the pointer to the RCU-protected data
414 structure with __rcu, and sparse will warn you if you
415 access that pointer without the services of one of the
416 variants of rcu_dereference().
417
418 These debugging aids can help you find problems that are
419 otherwise extremely difficult to spot.