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.rst for more information.
145 The rcu_dereference() primitive is used by the
146 various "_rcu()" list-traversal primitives, such
147 as the list_for_each_entry_rcu(). Note that it is
148 perfectly legal (if redundant) for update-side code to
149 use rcu_dereference() and the "_rcu()" list-traversal
150 primitives. This is particularly useful in code that
151 is common to readers and updaters. However, lockdep
152 will complain if you access rcu_dereference() outside
153 of an RCU read-side critical section. See lockdep.rst
154 to learn what to do about this.
156 Of course, neither rcu_dereference() nor the "_rcu()"
157 list-traversal primitives can substitute for a good
158 concurrency design coordinating among multiple updaters.
160 b. If the list macros are being used, the list_add_tail_rcu()
161 and list_add_rcu() primitives must be used in order
162 to prevent weakly ordered machines from misordering
163 structure initialization and pointer planting.
164 Similarly, if the hlist macros are being used, the
165 hlist_add_head_rcu() primitive is required.
167 c. If the list macros are being used, the list_del_rcu()
168 primitive must be used to keep list_del()'s pointer
169 poisoning from inflicting toxic effects on concurrent
170 readers. Similarly, if the hlist macros are being used,
171 the hlist_del_rcu() primitive is required.
173 The list_replace_rcu() and hlist_replace_rcu() primitives
174 may be used to replace an old structure with a new one
175 in their respective types of RCU-protected lists.
177 d. Rules similar to (4b) and (4c) apply to the "hlist_nulls"
178 type of RCU-protected linked lists.
180 e. Updates must ensure that initialization of a given
181 structure happens before pointers to that structure are
182 publicized. Use the rcu_assign_pointer() primitive
183 when publicizing a pointer to a structure that can
184 be traversed by an RCU read-side critical section.
186 5. If call_rcu() or call_srcu() is used, the callback function will
187 be called from softirq context. In particular, it cannot block.
189 6. Since synchronize_rcu() can block, it cannot be called
190 from any sort of irq context. The same rule applies
191 for synchronize_srcu(), synchronize_rcu_expedited(), and
192 synchronize_srcu_expedited().
194 The expedited forms of these primitives have the same semantics
195 as the non-expedited forms, but expediting is both expensive and
196 (with the exception of synchronize_srcu_expedited()) unfriendly
197 to real-time workloads. Use of the expedited primitives should
198 be restricted to rare configuration-change operations that would
199 not normally be undertaken while a real-time workload is running.
200 However, real-time workloads can use rcupdate.rcu_normal kernel
201 boot parameter to completely disable expedited grace periods,
202 though this might have performance implications.
204 In particular, if you find yourself invoking one of the expedited
205 primitives repeatedly in a loop, please do everyone a favor:
206 Restructure your code so that it batches the updates, allowing
207 a single non-expedited primitive to cover the entire batch.
208 This will very likely be faster than the loop containing the
209 expedited primitive, and will be much much easier on the rest
210 of the system, especially to real-time workloads running on
211 the rest of the system.
213 7. As of v4.20, a given kernel implements only one RCU flavor, which
214 is RCU-sched for PREEMPTION=n and RCU-preempt for PREEMPTION=y.
215 If the updater uses call_rcu() or synchronize_rcu(), then
216 the corresponding readers may use: (1) rcu_read_lock() and
217 rcu_read_unlock(), (2) any pair of primitives that disables
218 and re-enables softirq, for example, rcu_read_lock_bh() and
219 rcu_read_unlock_bh(), or (3) any pair of primitives that disables
220 and re-enables preemption, for example, rcu_read_lock_sched() and
221 rcu_read_unlock_sched(). If the updater uses synchronize_srcu()
222 or call_srcu(), then the corresponding readers must use
223 srcu_read_lock() and srcu_read_unlock(), and with the same
224 srcu_struct. The rules for the expedited RCU grace-period-wait
225 primitives are the same as for their non-expedited counterparts.
227 If the updater uses call_rcu_tasks() or synchronize_rcu_tasks(),
228 then the readers must refrain from executing voluntary
229 context switches, that is, from blocking. If the updater uses
230 call_rcu_tasks_trace() or synchronize_rcu_tasks_trace(), then
231 the corresponding readers must use rcu_read_lock_trace() and
232 rcu_read_unlock_trace(). If an updater uses call_rcu_tasks_rude()
233 or synchronize_rcu_tasks_rude(), then the corresponding readers
234 must use anything that disables interrupts.
236 Mixing things up will result in confusion and broken kernels, and
237 has even resulted in an exploitable security issue. Therefore,
238 when using non-obvious pairs of primitives, commenting is
239 of course a must. One example of non-obvious pairing is
240 the XDP feature in networking, which calls BPF programs from
241 network-driver NAPI (softirq) context. BPF relies heavily on RCU
242 protection for its data structures, but because the BPF program
243 invocation happens entirely within a single local_bh_disable()
244 section in a NAPI poll cycle, this usage is safe. The reason
245 that this usage is safe is that readers can use anything that
246 disables BH when updaters use call_rcu() or synchronize_rcu().
248 8. Although synchronize_rcu() is slower than is call_rcu(), it
249 usually results in simpler code. So, unless update performance is
250 critically important, the updaters cannot block, or the latency of
251 synchronize_rcu() is visible from userspace, synchronize_rcu()
252 should be used in preference to call_rcu(). Furthermore,
253 kfree_rcu() usually results in even simpler code than does
254 synchronize_rcu() without synchronize_rcu()'s multi-millisecond
255 latency. So please take advantage of kfree_rcu()'s "fire and
256 forget" memory-freeing capabilities where it applies.
258 An especially important property of the synchronize_rcu()
259 primitive is that it automatically self-limits: if grace periods
260 are delayed for whatever reason, then the synchronize_rcu()
261 primitive will correspondingly delay updates. In contrast,
262 code using call_rcu() should explicitly limit update rate in
263 cases where grace periods are delayed, as failing to do so can
264 result in excessive realtime latencies or even OOM conditions.
266 Ways of gaining this self-limiting property when using call_rcu()
269 a. Keeping a count of the number of data-structure elements
270 used by the RCU-protected data structure, including
271 those waiting for a grace period to elapse. Enforce a
272 limit on this number, stalling updates as needed to allow
273 previously deferred frees to complete. Alternatively,
274 limit only the number awaiting deferred free rather than
275 the total number of elements.
277 One way to stall the updates is to acquire the update-side
278 mutex. (Don't try this with a spinlock -- other CPUs
279 spinning on the lock could prevent the grace period
280 from ever ending.) Another way to stall the updates
281 is for the updates to use a wrapper function around
282 the memory allocator, so that this wrapper function
283 simulates OOM when there is too much memory awaiting an
284 RCU grace period. There are of course many other
285 variations on this theme.
287 b. Limiting update rate. For example, if updates occur only
288 once per hour, then no explicit rate limiting is
289 required, unless your system is already badly broken.
290 Older versions of the dcache subsystem take this approach,
291 guarding updates with a global lock, limiting their rate.
293 c. Trusted update -- if updates can only be done manually by
294 superuser or some other trusted user, then it might not
295 be necessary to automatically limit them. The theory
296 here is that superuser already has lots of ways to crash
299 d. Periodically invoke synchronize_rcu(), permitting a limited
300 number of updates per grace period.
302 The same cautions apply to call_srcu() and kfree_rcu().
304 Note that although these primitives do take action to avoid memory
305 exhaustion when any given CPU has too many callbacks, a determined
306 user could still exhaust memory. This is especially the case
307 if a system with a large number of CPUs has been configured to
308 offload all of its RCU callbacks onto a single CPU, or if the
309 system has relatively little free memory.
311 9. All RCU list-traversal primitives, which include
312 rcu_dereference(), list_for_each_entry_rcu(), and
313 list_for_each_safe_rcu(), must be either within an RCU read-side
314 critical section or must be protected by appropriate update-side
315 locks. RCU read-side critical sections are delimited by
316 rcu_read_lock() and rcu_read_unlock(), or by similar primitives
317 such as rcu_read_lock_bh() and rcu_read_unlock_bh(), in which
318 case the matching rcu_dereference() primitive must be used in
319 order to keep lockdep happy, in this case, rcu_dereference_bh().
321 The reason that it is permissible to use RCU list-traversal
322 primitives when the update-side lock is held is that doing so
323 can be quite helpful in reducing code bloat when common code is
324 shared between readers and updaters. Additional primitives
325 are provided for this case, as discussed in lockdep.rst.
327 One exception to this rule is when data is only ever added to
328 the linked data structure, and is never removed during any
329 time that readers might be accessing that structure. In such
330 cases, READ_ONCE() may be used in place of rcu_dereference()
331 and the read-side markers (rcu_read_lock() and rcu_read_unlock(),
332 for example) may be omitted.
334 10. Conversely, if you are in an RCU read-side critical section,
335 and you don't hold the appropriate update-side lock, you *must*
336 use the "_rcu()" variants of the list macros. Failing to do so
337 will break Alpha, cause aggressive compilers to generate bad code,
338 and confuse people trying to read your code.
340 11. Any lock acquired by an RCU callback must be acquired elsewhere
341 with softirq disabled, e.g., via spin_lock_irqsave(),
342 spin_lock_bh(), etc. Failing to disable softirq on a given
343 acquisition of that lock will result in deadlock as soon as
344 the RCU softirq handler happens to run your RCU callback while
345 interrupting that acquisition's critical section.
347 12. RCU callbacks can be and are executed in parallel. In many cases,
348 the callback code simply wrappers around kfree(), so that this
349 is not an issue (or, more accurately, to the extent that it is
350 an issue, the memory-allocator locking handles it). However,
351 if the callbacks do manipulate a shared data structure, they
352 must use whatever locking or other synchronization is required
353 to safely access and/or modify that data structure.
355 Do not assume that RCU callbacks will be executed on the same
356 CPU that executed the corresponding call_rcu() or call_srcu().
357 For example, if a given CPU goes offline while having an RCU
358 callback pending, then that RCU callback will execute on some
359 surviving CPU. (If this was not the case, a self-spawning RCU
360 callback would prevent the victim CPU from ever going offline.)
361 Furthermore, CPUs designated by rcu_nocbs= might well *always*
362 have their RCU callbacks executed on some other CPUs, in fact,
363 for some real-time workloads, this is the whole point of using
364 the rcu_nocbs= kernel boot parameter.
366 13. Unlike other forms of RCU, it *is* permissible to block in an
367 SRCU read-side critical section (demarked by srcu_read_lock()
368 and srcu_read_unlock()), hence the "SRCU": "sleepable RCU".
369 Please note that if you don't need to sleep in read-side critical
370 sections, you should be using RCU rather than SRCU, because RCU
371 is almost always faster and easier to use than is SRCU.
373 Also unlike other forms of RCU, explicit initialization and
374 cleanup is required either at build time via DEFINE_SRCU()
375 or DEFINE_STATIC_SRCU() or at runtime via init_srcu_struct()
376 and cleanup_srcu_struct(). These last two are passed a
377 "struct srcu_struct" that defines the scope of a given
378 SRCU domain. Once initialized, the srcu_struct is passed
379 to srcu_read_lock(), srcu_read_unlock() synchronize_srcu(),
380 synchronize_srcu_expedited(), and call_srcu(). A given
381 synchronize_srcu() waits only for SRCU read-side critical
382 sections governed by srcu_read_lock() and srcu_read_unlock()
383 calls that have been passed the same srcu_struct. This property
384 is what makes sleeping read-side critical sections tolerable --
385 a given subsystem delays only its own updates, not those of other
386 subsystems using SRCU. Therefore, SRCU is less prone to OOM the
387 system than RCU would be if RCU's read-side critical sections
388 were permitted to sleep.
390 The ability to sleep in read-side critical sections does not
391 come for free. First, corresponding srcu_read_lock() and
392 srcu_read_unlock() calls must be passed the same srcu_struct.
393 Second, grace-period-detection overhead is amortized only
394 over those updates sharing a given srcu_struct, rather than
395 being globally amortized as they are for other forms of RCU.
396 Therefore, SRCU should be used in preference to rw_semaphore
397 only in extremely read-intensive situations, or in situations
398 requiring SRCU's read-side deadlock immunity or low read-side
399 realtime latency. You should also consider percpu_rw_semaphore
400 when you need lightweight readers.
402 SRCU's expedited primitive (synchronize_srcu_expedited())
403 never sends IPIs to other CPUs, so it is easier on
404 real-time workloads than is synchronize_rcu_expedited().
406 Note that rcu_assign_pointer() relates to SRCU just as it does to
407 other forms of RCU, but instead of rcu_dereference() you should
408 use srcu_dereference() in order to avoid lockdep splats.
410 14. The whole point of call_rcu(), synchronize_rcu(), and friends
411 is to wait until all pre-existing readers have finished before
412 carrying out some otherwise-destructive operation. It is
413 therefore critically important to *first* remove any path
414 that readers can follow that could be affected by the
415 destructive operation, and *only then* invoke call_rcu(),
416 synchronize_rcu(), or friends.
418 Because these primitives only wait for pre-existing readers, it
419 is the caller's responsibility to guarantee that any subsequent
420 readers will execute safely.
422 15. The various RCU read-side primitives do *not* necessarily contain
423 memory barriers. You should therefore plan for the CPU
424 and the compiler to freely reorder code into and out of RCU
425 read-side critical sections. It is the responsibility of the
426 RCU update-side primitives to deal with this.
428 For SRCU readers, you can use smp_mb__after_srcu_read_unlock()
429 immediately after an srcu_read_unlock() to get a full barrier.
431 16. Use CONFIG_PROVE_LOCKING, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and the
432 __rcu sparse checks to validate your RCU code. These can help
433 find problems as follows:
435 CONFIG_PROVE_LOCKING:
436 check that accesses to RCU-protected data
437 structures are carried out under the proper RCU
438 read-side critical section, while holding the right
439 combination of locks, or whatever other conditions
442 CONFIG_DEBUG_OBJECTS_RCU_HEAD:
443 check that you don't pass the
444 same object to call_rcu() (or friends) before an RCU
445 grace period has elapsed since the last time that you
446 passed that same object to call_rcu() (or friends).
449 tag the pointer to the RCU-protected data
450 structure with __rcu, and sparse will warn you if you
451 access that pointer without the services of one of the
452 variants of rcu_dereference().
454 These debugging aids can help you find problems that are
455 otherwise extremely difficult to spot.
457 17. If you register a callback using call_rcu() or call_srcu(), and
458 pass in a function defined within a loadable module, then it in
459 necessary to wait for all pending callbacks to be invoked after
460 the last invocation and before unloading that module. Note that
461 it is absolutely *not* sufficient to wait for a grace period!
462 The current (say) synchronize_rcu() implementation is *not*
463 guaranteed to wait for callbacks registered on other CPUs.
464 Or even on the current CPU if that CPU recently went offline
465 and came back online.
467 You instead need to use one of the barrier functions:
469 - call_rcu() -> rcu_barrier()
470 - call_srcu() -> srcu_barrier()
472 However, these barrier functions are absolutely *not* guaranteed
473 to wait for a grace period. In fact, if there are no call_rcu()
474 callbacks waiting anywhere in the system, rcu_barrier() is within
475 its rights to return immediately.
477 So if you need to wait for both an RCU grace period and for
478 all pre-existing call_rcu() callbacks, you will need to execute
479 both rcu_barrier() and synchronize_rcu(), if necessary, using
480 something like workqueues to to execute them concurrently.
482 See rcubarrier.rst for more information.