xref: /linux/Documentation/RCU/checklist.rst (revision 4fd18fc38757217c746aa063ba9e4729814dc737)
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 bid 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.	If you don't believe me, see:
133
134			http://www.openvms.compaq.com/wizard/wiz_2637.html
135
136		The rcu_dereference() primitive is also an excellent
137		documentation aid, letting the person reading the
138		code know exactly which pointers are protected by RCU.
139		Please note that compilers can also reorder code, and
140		they are becoming increasingly aggressive about doing
141		just that.  The rcu_dereference() primitive therefore also
142		prevents destructive compiler optimizations.  However,
143		with a bit of devious creativity, it is possible to
144		mishandle the return value from rcu_dereference().
145		Please see rcu_dereference.txt in this directory for
146		more information.
147
148		The rcu_dereference() primitive is used by the
149		various "_rcu()" list-traversal primitives, such
150		as the list_for_each_entry_rcu().  Note that it is
151		perfectly legal (if redundant) for update-side code to
152		use rcu_dereference() and the "_rcu()" list-traversal
153		primitives.  This is particularly useful in code that
154		is common to readers and updaters.  However, lockdep
155		will complain if you access rcu_dereference() outside
156		of an RCU read-side critical section.  See lockdep.txt
157		to learn what to do about this.
158
159		Of course, neither rcu_dereference() nor the "_rcu()"
160		list-traversal primitives can substitute for a good
161		concurrency design coordinating among multiple updaters.
162
163	b.	If the list macros are being used, the list_add_tail_rcu()
164		and list_add_rcu() primitives must be used in order
165		to prevent weakly ordered machines from misordering
166		structure initialization and pointer planting.
167		Similarly, if the hlist macros are being used, the
168		hlist_add_head_rcu() primitive is required.
169
170	c.	If the list macros are being used, the list_del_rcu()
171		primitive must be used to keep list_del()'s pointer
172		poisoning from inflicting toxic effects on concurrent
173		readers.  Similarly, if the hlist macros are being used,
174		the hlist_del_rcu() primitive is required.
175
176		The list_replace_rcu() and hlist_replace_rcu() primitives
177		may be used to replace an old structure with a new one
178		in their respective types of RCU-protected lists.
179
180	d.	Rules similar to (4b) and (4c) apply to the "hlist_nulls"
181		type of RCU-protected linked lists.
182
183	e.	Updates must ensure that initialization of a given
184		structure happens before pointers to that structure are
185		publicized.  Use the rcu_assign_pointer() primitive
186		when publicizing a pointer to a structure that can
187		be traversed by an RCU read-side critical section.
188
1895.	If call_rcu() or call_srcu() is used, the callback function will
190	be called from softirq context.  In particular, it cannot block.
191
1926.	Since synchronize_rcu() can block, it cannot be called
193	from any sort of irq context.  The same rule applies
194	for synchronize_srcu(), synchronize_rcu_expedited(), and
195	synchronize_srcu_expedited().
196
197	The expedited forms of these primitives have the same semantics
198	as the non-expedited forms, but expediting is both expensive and
199	(with the exception of synchronize_srcu_expedited()) unfriendly
200	to real-time workloads.  Use of the expedited primitives should
201	be restricted to rare configuration-change operations that would
202	not normally be undertaken while a real-time workload is running.
203	However, real-time workloads can use rcupdate.rcu_normal kernel
204	boot parameter to completely disable expedited grace periods,
205	though this might have performance implications.
206
207	In particular, if you find yourself invoking one of the expedited
208	primitives repeatedly in a loop, please do everyone a favor:
209	Restructure your code so that it batches the updates, allowing
210	a single non-expedited primitive to cover the entire batch.
211	This will very likely be faster than the loop containing the
212	expedited primitive, and will be much much easier on the rest
213	of the system, especially to real-time workloads running on
214	the rest of the system.
215
2167.	As of v4.20, a given kernel implements only one RCU flavor,
217	which is RCU-sched for PREEMPT=n and RCU-preempt for PREEMPT=y.
218	If the updater uses call_rcu() or synchronize_rcu(),
219	then the corresponding readers my use rcu_read_lock() and
220	rcu_read_unlock(), rcu_read_lock_bh() and rcu_read_unlock_bh(),
221	or any pair of primitives that disables and re-enables preemption,
222	for example, rcu_read_lock_sched() and rcu_read_unlock_sched().
223	If the updater uses synchronize_srcu() or call_srcu(),
224	then the corresponding readers must use srcu_read_lock() and
225	srcu_read_unlock(), and with the same srcu_struct.  The rules for
226	the expedited primitives are the same as for their non-expedited
227	counterparts.  Mixing things up will result in confusion and
228	broken kernels, and has even resulted in an exploitable security
229	issue.
230
231	One exception to this rule: rcu_read_lock() and rcu_read_unlock()
232	may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh()
233	in cases where local bottom halves are already known to be
234	disabled, for example, in irq or softirq context.  Commenting
235	such cases is a must, of course!  And the jury is still out on
236	whether the increased speed is worth it.
237
2388.	Although synchronize_rcu() is slower than is call_rcu(), it
239	usually results in simpler code.  So, unless update performance is
240	critically important, the updaters cannot block, or the latency of
241	synchronize_rcu() is visible from userspace, synchronize_rcu()
242	should be used in preference to call_rcu().  Furthermore,
243	kfree_rcu() usually results in even simpler code than does
244	synchronize_rcu() without synchronize_rcu()'s multi-millisecond
245	latency.  So please take advantage of kfree_rcu()'s "fire and
246	forget" memory-freeing capabilities where it applies.
247
248	An especially important property of the synchronize_rcu()
249	primitive is that it automatically self-limits: if grace periods
250	are delayed for whatever reason, then the synchronize_rcu()
251	primitive will correspondingly delay updates.  In contrast,
252	code using call_rcu() should explicitly limit update rate in
253	cases where grace periods are delayed, as failing to do so can
254	result in excessive realtime latencies or even OOM conditions.
255
256	Ways of gaining this self-limiting property when using call_rcu()
257	include:
258
259	a.	Keeping a count of the number of data-structure elements
260		used by the RCU-protected data structure, including
261		those waiting for a grace period to elapse.  Enforce a
262		limit on this number, stalling updates as needed to allow
263		previously deferred frees to complete.	Alternatively,
264		limit only the number awaiting deferred free rather than
265		the total number of elements.
266
267		One way to stall the updates is to acquire the update-side
268		mutex.	(Don't try this with a spinlock -- other CPUs
269		spinning on the lock could prevent the grace period
270		from ever ending.)  Another way to stall the updates
271		is for the updates to use a wrapper function around
272		the memory allocator, so that this wrapper function
273		simulates OOM when there is too much memory awaiting an
274		RCU grace period.  There are of course many other
275		variations on this theme.
276
277	b.	Limiting update rate.  For example, if updates occur only
278		once per hour, then no explicit rate limiting is
279		required, unless your system is already badly broken.
280		Older versions of the dcache subsystem take this approach,
281		guarding updates with a global lock, limiting their rate.
282
283	c.	Trusted update -- if updates can only be done manually by
284		superuser or some other trusted user, then it might not
285		be necessary to automatically limit them.  The theory
286		here is that superuser already has lots of ways to crash
287		the machine.
288
289	d.	Periodically invoke synchronize_rcu(), permitting a limited
290		number of updates per grace period.
291
292	The same cautions apply to call_srcu() and kfree_rcu().
293
294	Note that although these primitives do take action to avoid memory
295	exhaustion when any given CPU has too many callbacks, a determined
296	user could still exhaust memory.  This is especially the case
297	if a system with a large number of CPUs has been configured to
298	offload all of its RCU callbacks onto a single CPU, or if the
299	system has relatively little free memory.
300
3019.	All RCU list-traversal primitives, which include
302	rcu_dereference(), list_for_each_entry_rcu(), and
303	list_for_each_safe_rcu(), must be either within an RCU read-side
304	critical section or must be protected by appropriate update-side
305	locks.	RCU read-side critical sections are delimited by
306	rcu_read_lock() and rcu_read_unlock(), or by similar primitives
307	such as rcu_read_lock_bh() and rcu_read_unlock_bh(), in which
308	case the matching rcu_dereference() primitive must be used in
309	order to keep lockdep happy, in this case, rcu_dereference_bh().
310
311	The reason that it is permissible to use RCU list-traversal
312	primitives when the update-side lock is held is that doing so
313	can be quite helpful in reducing code bloat when common code is
314	shared between readers and updaters.  Additional primitives
315	are provided for this case, as discussed in lockdep.txt.
316
317	One exception to this rule is when data is only ever added to
318	the linked data structure, and is never removed during any
319	time that readers might be accessing that structure.  In such
320	cases, READ_ONCE() may be used in place of rcu_dereference()
321	and the read-side markers (rcu_read_lock() and rcu_read_unlock(),
322	for example) may be omitted.
323
32410.	Conversely, if you are in an RCU read-side critical section,
325	and you don't hold the appropriate update-side lock, you -must-
326	use the "_rcu()" variants of the list macros.  Failing to do so
327	will break Alpha, cause aggressive compilers to generate bad code,
328	and confuse people trying to read your code.
329
33011.	Any lock acquired by an RCU callback must be acquired elsewhere
331	with softirq disabled, e.g., via spin_lock_irqsave(),
332	spin_lock_bh(), etc.  Failing to disable softirq on a given
333	acquisition of that lock will result in deadlock as soon as
334	the RCU softirq handler happens to run your RCU callback while
335	interrupting that acquisition's critical section.
336
33712.	RCU callbacks can be and are executed in parallel.  In many cases,
338	the callback code simply wrappers around kfree(), so that this
339	is not an issue (or, more accurately, to the extent that it is
340	an issue, the memory-allocator locking handles it).  However,
341	if the callbacks do manipulate a shared data structure, they
342	must use whatever locking or other synchronization is required
343	to safely access and/or modify that data structure.
344
345	Do not assume that RCU callbacks will be executed on the same
346	CPU that executed the corresponding call_rcu() or call_srcu().
347	For example, if a given CPU goes offline while having an RCU
348	callback pending, then that RCU callback will execute on some
349	surviving CPU.	(If this was not the case, a self-spawning RCU
350	callback would prevent the victim CPU from ever going offline.)
351	Furthermore, CPUs designated by rcu_nocbs= might well -always-
352	have their RCU callbacks executed on some other CPUs, in fact,
353	for some  real-time workloads, this is the whole point of using
354	the rcu_nocbs= kernel boot parameter.
355
35613.	Unlike other forms of RCU, it -is- permissible to block in an
357	SRCU read-side critical section (demarked by srcu_read_lock()
358	and srcu_read_unlock()), hence the "SRCU": "sleepable RCU".
359	Please note that if you don't need to sleep in read-side critical
360	sections, you should be using RCU rather than SRCU, because RCU
361	is almost always faster and easier to use than is SRCU.
362
363	Also unlike other forms of RCU, explicit initialization and
364	cleanup is required either at build time via DEFINE_SRCU()
365	or DEFINE_STATIC_SRCU() or at runtime via init_srcu_struct()
366	and cleanup_srcu_struct().  These last two are passed a
367	"struct srcu_struct" that defines the scope of a given
368	SRCU domain.  Once initialized, the srcu_struct is passed
369	to srcu_read_lock(), srcu_read_unlock() synchronize_srcu(),
370	synchronize_srcu_expedited(), and call_srcu().	A given
371	synchronize_srcu() waits only for SRCU read-side critical
372	sections governed by srcu_read_lock() and srcu_read_unlock()
373	calls that have been passed the same srcu_struct.  This property
374	is what makes sleeping read-side critical sections tolerable --
375	a given subsystem delays only its own updates, not those of other
376	subsystems using SRCU.	Therefore, SRCU is less prone to OOM the
377	system than RCU would be if RCU's read-side critical sections
378	were permitted to sleep.
379
380	The ability to sleep in read-side critical sections does not
381	come for free.	First, corresponding srcu_read_lock() and
382	srcu_read_unlock() calls must be passed the same srcu_struct.
383	Second, grace-period-detection overhead is amortized only
384	over those updates sharing a given srcu_struct, rather than
385	being globally amortized as they are for other forms of RCU.
386	Therefore, SRCU should be used in preference to rw_semaphore
387	only in extremely read-intensive situations, or in situations
388	requiring SRCU's read-side deadlock immunity or low read-side
389	realtime latency.  You should also consider percpu_rw_semaphore
390	when you need lightweight readers.
391
392	SRCU's expedited primitive (synchronize_srcu_expedited())
393	never sends IPIs to other CPUs, so it is easier on
394	real-time workloads than is synchronize_rcu_expedited().
395
396	Note that rcu_assign_pointer() relates to SRCU just as it does to
397	other forms of RCU, but instead of rcu_dereference() you should
398	use srcu_dereference() in order to avoid lockdep splats.
399
40014.	The whole point of call_rcu(), synchronize_rcu(), and friends
401	is to wait until all pre-existing readers have finished before
402	carrying out some otherwise-destructive operation.  It is
403	therefore critically important to -first- remove any path
404	that readers can follow that could be affected by the
405	destructive operation, and -only- -then- invoke call_rcu(),
406	synchronize_rcu(), or friends.
407
408	Because these primitives only wait for pre-existing readers, it
409	is the caller's responsibility to guarantee that any subsequent
410	readers will execute safely.
411
41215.	The various RCU read-side primitives do -not- necessarily contain
413	memory barriers.  You should therefore plan for the CPU
414	and the compiler to freely reorder code into and out of RCU
415	read-side critical sections.  It is the responsibility of the
416	RCU update-side primitives to deal with this.
417
418	For SRCU readers, you can use smp_mb__after_srcu_read_unlock()
419	immediately after an srcu_read_unlock() to get a full barrier.
420
42116.	Use CONFIG_PROVE_LOCKING, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and the
422	__rcu sparse checks to validate your RCU code.	These can help
423	find problems as follows:
424
425	CONFIG_PROVE_LOCKING:
426		check that accesses to RCU-protected data
427		structures are carried out under the proper RCU
428		read-side critical section, while holding the right
429		combination of locks, or whatever other conditions
430		are appropriate.
431
432	CONFIG_DEBUG_OBJECTS_RCU_HEAD:
433		check that you don't pass the
434		same object to call_rcu() (or friends) before an RCU
435		grace period has elapsed since the last time that you
436		passed that same object to call_rcu() (or friends).
437
438	__rcu sparse checks:
439		tag the pointer to the RCU-protected data
440		structure with __rcu, and sparse will warn you if you
441		access that pointer without the services of one of the
442		variants of rcu_dereference().
443
444	These debugging aids can help you find problems that are
445	otherwise extremely difficult to spot.
446
44717.	If you register a callback using call_rcu() or call_srcu(), and
448	pass in a function defined within a loadable module, then it in
449	necessary to wait for all pending callbacks to be invoked after
450	the last invocation and before unloading that module.  Note that
451	it is absolutely -not- sufficient to wait for a grace period!
452	The current (say) synchronize_rcu() implementation is -not-
453	guaranteed to wait for callbacks registered on other CPUs.
454	Or even on the current CPU if that CPU recently went offline
455	and came back online.
456
457	You instead need to use one of the barrier functions:
458
459	-	call_rcu() -> rcu_barrier()
460	-	call_srcu() -> srcu_barrier()
461
462	However, these barrier functions are absolutely -not- guaranteed
463	to wait for a grace period.  In fact, if there are no call_rcu()
464	callbacks waiting anywhere in the system, rcu_barrier() is within
465	its rights to return immediately.
466
467	So if you need to wait for both an RCU grace period and for
468	all pre-existing call_rcu() callbacks, you will need to execute
469	both rcu_barrier() and synchronize_rcu(), if necessary, using
470	something like workqueues to to execute them concurrently.
471
472	See rcubarrier.txt for more information.
473