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