1 Review Checklist for RCU Patches
4 This document contains a checklist for producing and reviewing patches
5 that make use of RCU. Violating any of the rules listed below will
6 result in the same sorts of problems that leaving out a locking primitive
7 would cause. This list is based on experiences reviewing such patches
8 over a rather long period of time, but improvements are always welcome!
10 0. Is RCU being applied to a read-mostly situation? If the data
11 structure is updated more than about 10% of the time, then
12 you should strongly consider some other approach, unless
13 detailed performance measurements show that RCU is nonetheless
14 the right tool for the job.
16 The other exception would be where performance is not an issue,
17 and RCU provides a simpler implementation. An example of this
18 situation is the dynamic NMI code in the Linux 2.6 kernel,
19 at least on architectures where NMIs are rare.
21 1. Does the update code have proper mutual exclusion?
23 RCU does allow -readers- to run (almost) naked, but -writers- must
24 still use some sort of mutual exclusion, such as:
27 b. atomic operations, or
28 c. restricting updates to a single task.
30 If you choose #b, be prepared to describe how you have handled
31 memory barriers on weakly ordered machines (pretty much all of
32 them -- even x86 allows reads to be reordered), and be prepared
33 to explain why this added complexity is worthwhile. If you
34 choose #c, be prepared to explain how this single task does not
35 become a major bottleneck on big multiprocessor machines (for
36 example, if the task is updating information relating to itself
37 that other tasks can read, there by definition can be no
40 2. Do the RCU read-side critical sections make proper use of
41 rcu_read_lock() and friends? These primitives are needed
42 to suppress preemption (or bottom halves, in the case of
43 rcu_read_lock_bh()) in the read-side critical sections,
44 and are also an excellent aid to readability.
46 As a rough rule of thumb, any dereference of an RCU-protected
47 pointer must be covered by rcu_read_lock() or rcu_read_lock_bh()
48 or by the appropriate update-side lock.
50 3. Does the update code tolerate concurrent accesses?
52 The whole point of RCU is to permit readers to run without
53 any locks or atomic operations. This means that readers will
54 be running while updates are in progress. There are a number
55 of ways to handle this concurrency, depending on the situation:
57 a. Make updates appear atomic to readers. For example,
58 pointer updates to properly aligned fields will appear
59 atomic, as will individual atomic primitives. Operations
60 performed under a lock and sequences of multiple atomic
61 primitives will -not- appear to be atomic.
63 This is almost always the best approach.
65 b. Carefully order the updates and the reads so that
66 readers see valid data at all phases of the update.
67 This is often more difficult than it sounds, especially
68 given modern CPUs' tendency to reorder memory references.
69 One must usually liberally sprinkle memory barriers
70 (smp_wmb(), smp_rmb(), smp_mb()) through the code,
71 making it difficult to understand and to test.
73 It is usually better to group the changing data into
74 a separate structure, so that the change may be made
75 to appear atomic by updating a pointer to reference
76 a new structure containing updated values.
78 4. Weakly ordered CPUs pose special challenges. Almost all CPUs
79 are weakly ordered -- even i386 CPUs allow reads to be reordered.
80 RCU code must take all of the following measures to prevent
81 memory-corruption problems:
83 a. Readers must maintain proper ordering of their memory
84 accesses. The rcu_dereference() primitive ensures that
85 the CPU picks up the pointer before it picks up the data
86 that the pointer points to. This really is necessary
87 on Alpha CPUs. If you don't believe me, see:
89 http://www.openvms.compaq.com/wizard/wiz_2637.html
91 The rcu_dereference() primitive is also an excellent
92 documentation aid, letting the person reading the code
93 know exactly which pointers are protected by RCU.
95 The rcu_dereference() primitive is used by the various
96 "_rcu()" list-traversal primitives, such as the
97 list_for_each_entry_rcu(). Note that it is perfectly
98 legal (if redundant) for update-side code to use
99 rcu_dereference() and the "_rcu()" list-traversal
100 primitives. This is particularly useful in code
101 that is common to readers and updaters.
103 b. If the list macros are being used, the list_add_tail_rcu()
104 and list_add_rcu() primitives must be used in order
105 to prevent weakly ordered machines from misordering
106 structure initialization and pointer planting.
107 Similarly, if the hlist macros are being used, the
108 hlist_add_head_rcu() primitive is required.
110 c. If the list macros are being used, the list_del_rcu()
111 primitive must be used to keep list_del()'s pointer
112 poisoning from inflicting toxic effects on concurrent
113 readers. Similarly, if the hlist macros are being used,
114 the hlist_del_rcu() primitive is required.
116 The list_replace_rcu() primitive may be used to
117 replace an old structure with a new one in an
120 d. Updates must ensure that initialization of a given
121 structure happens before pointers to that structure are
122 publicized. Use the rcu_assign_pointer() primitive
123 when publicizing a pointer to a structure that can
124 be traversed by an RCU read-side critical section.
126 5. If call_rcu(), or a related primitive such as call_rcu_bh(),
127 is used, the callback function must be written to be called
128 from softirq context. In particular, it cannot block.
130 6. Since synchronize_rcu() can block, it cannot be called from
131 any sort of irq context.
133 7. If the updater uses call_rcu(), then the corresponding readers
134 must use rcu_read_lock() and rcu_read_unlock(). If the updater
135 uses call_rcu_bh(), then the corresponding readers must use
136 rcu_read_lock_bh() and rcu_read_unlock_bh(). Mixing things up
137 will result in confusion and broken kernels.
139 One exception to this rule: rcu_read_lock() and rcu_read_unlock()
140 may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh()
141 in cases where local bottom halves are already known to be
142 disabled, for example, in irq or softirq context. Commenting
143 such cases is a must, of course! And the jury is still out on
144 whether the increased speed is worth it.
146 8. Although synchronize_rcu() is a bit slower than is call_rcu(),
147 it usually results in simpler code. So, unless update
148 performance is critically important or the updaters cannot block,
149 synchronize_rcu() should be used in preference to call_rcu().
151 An especially important property of the synchronize_rcu()
152 primitive is that it automatically self-limits: if grace periods
153 are delayed for whatever reason, then the synchronize_rcu()
154 primitive will correspondingly delay updates. In contrast,
155 code using call_rcu() should explicitly limit update rate in
156 cases where grace periods are delayed, as failing to do so can
157 result in excessive realtime latencies or even OOM conditions.
159 Ways of gaining this self-limiting property when using call_rcu()
162 a. Keeping a count of the number of data-structure elements
163 used by the RCU-protected data structure, including those
164 waiting for a grace period to elapse. Enforce a limit
165 on this number, stalling updates as needed to allow
166 previously deferred frees to complete.
168 Alternatively, limit only the number awaiting deferred
169 free rather than the total number of elements.
171 b. Limiting update rate. For example, if updates occur only
172 once per hour, then no explicit rate limiting is required,
173 unless your system is already badly broken. The dcache
174 subsystem takes this approach -- updates are guarded
175 by a global lock, limiting their rate.
177 c. Trusted update -- if updates can only be done manually by
178 superuser or some other trusted user, then it might not
179 be necessary to automatically limit them. The theory
180 here is that superuser already has lots of ways to crash
183 d. Use call_rcu_bh() rather than call_rcu(), in order to take
184 advantage of call_rcu_bh()'s faster grace periods.
186 e. Periodically invoke synchronize_rcu(), permitting a limited
187 number of updates per grace period.
189 9. All RCU list-traversal primitives, which include
190 list_for_each_rcu(), list_for_each_entry_rcu(),
191 list_for_each_continue_rcu(), and list_for_each_safe_rcu(),
192 must be within an RCU read-side critical section. RCU
193 read-side critical sections are delimited by rcu_read_lock()
194 and rcu_read_unlock(), or by similar primitives such as
195 rcu_read_lock_bh() and rcu_read_unlock_bh().
197 Use of the _rcu() list-traversal primitives outside of an
198 RCU read-side critical section causes no harm other than
199 a slight performance degradation on Alpha CPUs. It can
200 also be quite helpful in reducing code bloat when common
201 code is shared between readers and updaters.
203 10. Conversely, if you are in an RCU read-side critical section,
204 you -must- use the "_rcu()" variants of the list macros.
205 Failing to do so will break Alpha and confuse people reading
208 11. Note that synchronize_rcu() -only- guarantees to wait until
209 all currently executing rcu_read_lock()-protected RCU read-side
210 critical sections complete. It does -not- necessarily guarantee
211 that all currently running interrupts, NMIs, preempt_disable()
212 code, or idle loops will complete. Therefore, if you do not have
213 rcu_read_lock()-protected read-side critical sections, do -not-
214 use synchronize_rcu().
216 If you want to wait for some of these other things, you might
217 instead need to use synchronize_irq() or synchronize_sched().
219 12. Any lock acquired by an RCU callback must be acquired elsewhere
220 with irq disabled, e.g., via spin_lock_irqsave(). Failing to
221 disable irq on a given acquisition of that lock will result in
222 deadlock as soon as the RCU callback happens to interrupt that
223 acquisition's critical section.
225 13. RCU callbacks can be and are executed in parallel. In many cases,
226 the callback code simply wrappers around kfree(), so that this
227 is not an issue (or, more accurately, to the extent that it is
228 an issue, the memory-allocator locking handles it). However,
229 if the callbacks do manipulate a shared data structure, they
230 must use whatever locking or other synchronization is required
231 to safely access and/or modify that data structure.
233 14. SRCU (srcu_read_lock(), srcu_read_unlock(), and synchronize_srcu())
234 may only be invoked from process context. Unlike other forms of
235 RCU, it -is- permissible to block in an SRCU read-side critical
236 section (demarked by srcu_read_lock() and srcu_read_unlock()),
237 hence the "SRCU": "sleepable RCU". Please note that if you
238 don't need to sleep in read-side critical sections, you should
239 be using RCU rather than SRCU, because RCU is almost always
240 faster and easier to use than is SRCU.
242 Also unlike other forms of RCU, explicit initialization
243 and cleanup is required via init_srcu_struct() and
244 cleanup_srcu_struct(). These are passed a "struct srcu_struct"
245 that defines the scope of a given SRCU domain. Once initialized,
246 the srcu_struct is passed to srcu_read_lock(), srcu_read_unlock()
247 and synchronize_srcu(). A given synchronize_srcu() waits only
248 for SRCU read-side critical sections governed by srcu_read_lock()
249 and srcu_read_unlock() calls that have been passd the same
250 srcu_struct. This property is what makes sleeping read-side
251 critical sections tolerable -- a given subsystem delays only
252 its own updates, not those of other subsystems using SRCU.
253 Therefore, SRCU is less prone to OOM the system than RCU would
254 be if RCU's read-side critical sections were permitted to
257 The ability to sleep in read-side critical sections does not
258 come for free. First, corresponding srcu_read_lock() and
259 srcu_read_unlock() calls must be passed the same srcu_struct.
260 Second, grace-period-detection overhead is amortized only
261 over those updates sharing a given srcu_struct, rather than
262 being globally amortized as they are for other forms of RCU.
263 Therefore, SRCU should be used in preference to rw_semaphore
264 only in extremely read-intensive situations, or in situations
265 requiring SRCU's read-side deadlock immunity or low read-side
268 Note that, rcu_assign_pointer() and rcu_dereference() relate to
269 SRCU just as they do to other forms of RCU.
projects / linux-2.6 / blob