1.. SPDX-License-Identifier: GPL-2.0
2
3================================
4Review Checklist for RCU Patches
5================================
6
7
8This document contains a checklist for producing and reviewing patches
9that make use of RCU.  Violating any of the rules listed below will
10result in the same sorts of problems that leaving out a locking primitive
11would cause.  This list is based on experiences reviewing such patches
12over a rather long period of time, but improvements are always welcome!
13
140.	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.
21
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.
26
27	Yet another exception is where the low real-time latency of RCU's
28	read-side primitives is critically important.
29
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
36	collectors do.
37
381.	Does the update code have proper mutual exclusion?
39
40	RCU does allow *readers* to run (almost) naked, but *writers* must
41	still use some sort of mutual exclusion, such as:
42
43	a.	locking,
44	b.	atomic operations, or
45	c.	restricting updates to a single task.
46
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.
58
592.	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.
65
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.
71
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.
77
783.	Does the update code tolerate concurrent accesses?
79
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:
84
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
89		the Linux kernel.
90
91		This is almost always the best approach.
92
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.
98
99		This works quite well, also.
100
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.
107
108		This can work, but is starting to get a bit tricky.
109
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.
117
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.
122
1234.	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:
127
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
132		on Alpha CPUs.
133
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
144		more information.
145
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.
156
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.
160
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.
167
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.
173
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.
177
178	d.	Rules similar to (4b) and (4c) apply to the "hlist_nulls"
179		type of RCU-protected linked lists.
180
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.
186
1875.	If call_rcu() or call_srcu() is used, the callback function will
188	be called from softirq context.  In particular, it cannot block.
189
1906.	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().
194
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.
204
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.
213
2147.	As of v4.20, a given kernel implements only one RCU flavor, which
215	is RCU-sched for PREEMPTION=n and RCU-preempt for PREEMPTION=y.
216	If the updater uses call_rcu() or synchronize_rcu(), then
217	the corresponding readers may use:  (1) rcu_read_lock() and
218	rcu_read_unlock(), (2) any pair of primitives that disables
219	and re-enables softirq, for example, rcu_read_lock_bh() and
220	rcu_read_unlock_bh(), or (3) any pair of primitives that disables
221	and re-enables preemption, for example, rcu_read_lock_sched() and
222	rcu_read_unlock_sched().  If the updater uses synchronize_srcu()
223	or call_srcu(), then the corresponding readers must use
224	srcu_read_lock() and srcu_read_unlock(), and with the same
225	srcu_struct.  The rules for the expedited RCU grace-period-wait
226	primitives are the same as for their non-expedited counterparts.
227
228	If the updater uses call_rcu_tasks() or synchronize_rcu_tasks(),
229	then the readers must refrain from executing voluntary
230	context switches, that is, from blocking.  If the updater uses
231	call_rcu_tasks_trace() or synchronize_rcu_tasks_trace(), then
232	the corresponding readers must use rcu_read_lock_trace() and
233	rcu_read_unlock_trace().  If an updater uses call_rcu_tasks_rude()
234	or synchronize_rcu_tasks_rude(), then the corresponding readers
235	must use anything that disables interrupts.
236
237	Mixing things up will result in confusion and broken kernels, and
238	has even resulted in an exploitable security issue.  Therefore,
239	when using non-obvious pairs of primitives, commenting is
240	of course a must.  One example of non-obvious pairing is
241	the XDP feature in networking, which calls BPF programs from
242	network-driver NAPI (softirq) context.	BPF relies heavily on RCU
243	protection for its data structures, but because the BPF program
244	invocation happens entirely within a single local_bh_disable()
245	section in a NAPI poll cycle, this usage is safe.  The reason
246	that this usage is safe is that readers can use anything that
247	disables BH when updaters use call_rcu() or synchronize_rcu().
248
2498.	Although synchronize_rcu() is slower than is call_rcu(), it
250	usually results in simpler code.  So, unless update performance is
251	critically important, the updaters cannot block, or the latency of
252	synchronize_rcu() is visible from userspace, synchronize_rcu()
253	should be used in preference to call_rcu().  Furthermore,
254	kfree_rcu() usually results in even simpler code than does
255	synchronize_rcu() without synchronize_rcu()'s multi-millisecond
256	latency.  So please take advantage of kfree_rcu()'s "fire and
257	forget" memory-freeing capabilities where it applies.
258
259	An especially important property of the synchronize_rcu()
260	primitive is that it automatically self-limits: if grace periods
261	are delayed for whatever reason, then the synchronize_rcu()
262	primitive will correspondingly delay updates.  In contrast,
263	code using call_rcu() should explicitly limit update rate in
264	cases where grace periods are delayed, as failing to do so can
265	result in excessive realtime latencies or even OOM conditions.
266
267	Ways of gaining this self-limiting property when using call_rcu()
268	include:
269
270	a.	Keeping a count of the number of data-structure elements
271		used by the RCU-protected data structure, including
272		those waiting for a grace period to elapse.  Enforce a
273		limit on this number, stalling updates as needed to allow
274		previously deferred frees to complete.	Alternatively,
275		limit only the number awaiting deferred free rather than
276		the total number of elements.
277
278		One way to stall the updates is to acquire the update-side
279		mutex.	(Don't try this with a spinlock -- other CPUs
280		spinning on the lock could prevent the grace period
281		from ever ending.)  Another way to stall the updates
282		is for the updates to use a wrapper function around
283		the memory allocator, so that this wrapper function
284		simulates OOM when there is too much memory awaiting an
285		RCU grace period.  There are of course many other
286		variations on this theme.
287
288	b.	Limiting update rate.  For example, if updates occur only
289		once per hour, then no explicit rate limiting is
290		required, unless your system is already badly broken.
291		Older versions of the dcache subsystem take this approach,
292		guarding updates with a global lock, limiting their rate.
293
294	c.	Trusted update -- if updates can only be done manually by
295		superuser or some other trusted user, then it might not
296		be necessary to automatically limit them.  The theory
297		here is that superuser already has lots of ways to crash
298		the machine.
299
300	d.	Periodically invoke synchronize_rcu(), permitting a limited
301		number of updates per grace period.
302
303	The same cautions apply to call_srcu() and kfree_rcu().
304
305	Note that although these primitives do take action to avoid memory
306	exhaustion when any given CPU has too many callbacks, a determined
307	user could still exhaust memory.  This is especially the case
308	if a system with a large number of CPUs has been configured to
309	offload all of its RCU callbacks onto a single CPU, or if the
310	system has relatively little free memory.
311
3129.	All RCU list-traversal primitives, which include
313	rcu_dereference(), list_for_each_entry_rcu(), and
314	list_for_each_safe_rcu(), must be either within an RCU read-side
315	critical section or must be protected by appropriate update-side
316	locks.	RCU read-side critical sections are delimited by
317	rcu_read_lock() and rcu_read_unlock(), or by similar primitives
318	such as rcu_read_lock_bh() and rcu_read_unlock_bh(), in which
319	case the matching rcu_dereference() primitive must be used in
320	order to keep lockdep happy, in this case, rcu_dereference_bh().
321
322	The reason that it is permissible to use RCU list-traversal
323	primitives when the update-side lock is held is that doing so
324	can be quite helpful in reducing code bloat when common code is
325	shared between readers and updaters.  Additional primitives
326	are provided for this case, as discussed in lockdep.txt.
327
328	One exception to this rule is when data is only ever added to
329	the linked data structure, and is never removed during any
330	time that readers might be accessing that structure.  In such
331	cases, READ_ONCE() may be used in place of rcu_dereference()
332	and the read-side markers (rcu_read_lock() and rcu_read_unlock(),
333	for example) may be omitted.
334
33510.	Conversely, if you are in an RCU read-side critical section,
336	and you don't hold the appropriate update-side lock, you *must*
337	use the "_rcu()" variants of the list macros.  Failing to do so
338	will break Alpha, cause aggressive compilers to generate bad code,
339	and confuse people trying to read your code.
340
34111.	Any lock acquired by an RCU callback must be acquired elsewhere
342	with softirq disabled, e.g., via spin_lock_irqsave(),
343	spin_lock_bh(), etc.  Failing to disable softirq on a given
344	acquisition of that lock will result in deadlock as soon as
345	the RCU softirq handler happens to run your RCU callback while
346	interrupting that acquisition's critical section.
347
34812.	RCU callbacks can be and are executed in parallel.  In many cases,
349	the callback code simply wrappers around kfree(), so that this
350	is not an issue (or, more accurately, to the extent that it is
351	an issue, the memory-allocator locking handles it).  However,
352	if the callbacks do manipulate a shared data structure, they
353	must use whatever locking or other synchronization is required
354	to safely access and/or modify that data structure.
355
356	Do not assume that RCU callbacks will be executed on the same
357	CPU that executed the corresponding call_rcu() or call_srcu().
358	For example, if a given CPU goes offline while having an RCU
359	callback pending, then that RCU callback will execute on some
360	surviving CPU.	(If this was not the case, a self-spawning RCU
361	callback would prevent the victim CPU from ever going offline.)
362	Furthermore, CPUs designated by rcu_nocbs= might well *always*
363	have their RCU callbacks executed on some other CPUs, in fact,
364	for some  real-time workloads, this is the whole point of using
365	the rcu_nocbs= kernel boot parameter.
366
36713.	Unlike other forms of RCU, it *is* permissible to block in an
368	SRCU read-side critical section (demarked by srcu_read_lock()
369	and srcu_read_unlock()), hence the "SRCU": "sleepable RCU".
370	Please note that if you don't need to sleep in read-side critical
371	sections, you should be using RCU rather than SRCU, because RCU
372	is almost always faster and easier to use than is SRCU.
373
374	Also unlike other forms of RCU, explicit initialization and
375	cleanup is required either at build time via DEFINE_SRCU()
376	or DEFINE_STATIC_SRCU() or at runtime via init_srcu_struct()
377	and cleanup_srcu_struct().  These last two are passed a
378	"struct srcu_struct" that defines the scope of a given
379	SRCU domain.  Once initialized, the srcu_struct is passed
380	to srcu_read_lock(), srcu_read_unlock() synchronize_srcu(),
381	synchronize_srcu_expedited(), and call_srcu().	A given
382	synchronize_srcu() waits only for SRCU read-side critical
383	sections governed by srcu_read_lock() and srcu_read_unlock()
384	calls that have been passed the same srcu_struct.  This property
385	is what makes sleeping read-side critical sections tolerable --
386	a given subsystem delays only its own updates, not those of other
387	subsystems using SRCU.	Therefore, SRCU is less prone to OOM the
388	system than RCU would be if RCU's read-side critical sections
389	were permitted to sleep.
390
391	The ability to sleep in read-side critical sections does not
392	come for free.	First, corresponding srcu_read_lock() and
393	srcu_read_unlock() calls must be passed the same srcu_struct.
394	Second, grace-period-detection overhead is amortized only
395	over those updates sharing a given srcu_struct, rather than
396	being globally amortized as they are for other forms of RCU.
397	Therefore, SRCU should be used in preference to rw_semaphore
398	only in extremely read-intensive situations, or in situations
399	requiring SRCU's read-side deadlock immunity or low read-side
400	realtime latency.  You should also consider percpu_rw_semaphore
401	when you need lightweight readers.
402
403	SRCU's expedited primitive (synchronize_srcu_expedited())
404	never sends IPIs to other CPUs, so it is easier on
405	real-time workloads than is synchronize_rcu_expedited().
406
407	Note that rcu_assign_pointer() relates to SRCU just as it does to
408	other forms of RCU, but instead of rcu_dereference() you should
409	use srcu_dereference() in order to avoid lockdep splats.
410
41114.	The whole point of call_rcu(), synchronize_rcu(), and friends
412	is to wait until all pre-existing readers have finished before
413	carrying out some otherwise-destructive operation.  It is
414	therefore critically important to *first* remove any path
415	that readers can follow that could be affected by the
416	destructive operation, and *only then* invoke call_rcu(),
417	synchronize_rcu(), or friends.
418
419	Because these primitives only wait for pre-existing readers, it
420	is the caller's responsibility to guarantee that any subsequent
421	readers will execute safely.
422
42315.	The various RCU read-side primitives do *not* necessarily contain
424	memory barriers.  You should therefore plan for the CPU
425	and the compiler to freely reorder code into and out of RCU
426	read-side critical sections.  It is the responsibility of the
427	RCU update-side primitives to deal with this.
428
429	For SRCU readers, you can use smp_mb__after_srcu_read_unlock()
430	immediately after an srcu_read_unlock() to get a full barrier.
431
43216.	Use CONFIG_PROVE_LOCKING, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and the
433	__rcu sparse checks to validate your RCU code.	These can help
434	find problems as follows:
435
436	CONFIG_PROVE_LOCKING:
437		check that accesses to RCU-protected data
438		structures are carried out under the proper RCU
439		read-side critical section, while holding the right
440		combination of locks, or whatever other conditions
441		are appropriate.
442
443	CONFIG_DEBUG_OBJECTS_RCU_HEAD:
444		check that you don't pass the
445		same object to call_rcu() (or friends) before an RCU
446		grace period has elapsed since the last time that you
447		passed that same object to call_rcu() (or friends).
448
449	__rcu sparse checks:
450		tag the pointer to the RCU-protected data
451		structure with __rcu, and sparse will warn you if you
452		access that pointer without the services of one of the
453		variants of rcu_dereference().
454
455	These debugging aids can help you find problems that are
456	otherwise extremely difficult to spot.
457
45817.	If you register a callback using call_rcu() or call_srcu(), and
459	pass in a function defined within a loadable module, then it in
460	necessary to wait for all pending callbacks to be invoked after
461	the last invocation and before unloading that module.  Note that
462	it is absolutely *not* sufficient to wait for a grace period!
463	The current (say) synchronize_rcu() implementation is *not*
464	guaranteed to wait for callbacks registered on other CPUs.
465	Or even on the current CPU if that CPU recently went offline
466	and came back online.
467
468	You instead need to use one of the barrier functions:
469
470	-	call_rcu() -> rcu_barrier()
471	-	call_srcu() -> srcu_barrier()
472
473	However, these barrier functions are absolutely *not* guaranteed
474	to wait for a grace period.  In fact, if there are no call_rcu()
475	callbacks waiting anywhere in the system, rcu_barrier() is within
476	its rights to return immediately.
477
478	So if you need to wait for both an RCU grace period and for
479	all pre-existing call_rcu() callbacks, you will need to execute
480	both rcu_barrier() and synchronize_rcu(), if necessary, using
481	something like workqueues to to execute them concurrently.
482
483	See rcubarrier.txt for more information.
484