1Review Checklist for RCU Patches 2 3 4This document contains a checklist for producing and reviewing patches 5that make use of RCU. Violating any of the rules listed below will 6result in the same sorts of problems that leaving out a locking primitive 7would cause. This list is based on experiences reviewing such patches 8over a rather long period of time, but improvements are always welcome! 9 100. Is RCU being applied to a read-mostly situation? If the data 11 structure is updated more than about 10% of the time, then you 12 should strongly consider some other approach, unless detailed 13 performance measurements show that RCU is nonetheless the right 14 tool for the job. Yes, RCU does reduce read-side overhead by 15 increasing write-side overhead, which is exactly why normal uses 16 of RCU will do much more reading than updating. 17 18 Another exception is where performance is not an issue, and RCU 19 provides a simpler implementation. An example of this situation 20 is the dynamic NMI code in the Linux 2.6 kernel, at least on 21 architectures where NMIs are rare. 22 23 Yet another exception is where the low real-time latency of RCU's 24 read-side primitives is critically important. 25 261. Does the update code have proper mutual exclusion? 27 28 RCU does allow -readers- to run (almost) naked, but -writers- must 29 still use some sort of mutual exclusion, such as: 30 31 a. locking, 32 b. atomic operations, or 33 c. restricting updates to a single task. 34 35 If you choose #b, be prepared to describe how you have handled 36 memory barriers on weakly ordered machines (pretty much all of 37 them -- even x86 allows later loads to be reordered to precede 38 earlier stores), and be prepared to explain why this added 39 complexity is worthwhile. If you choose #c, be prepared to 40 explain how this single task does not become a major bottleneck on 41 big multiprocessor machines (for example, if the task is updating 42 information relating to itself that other tasks can read, there 43 by definition can be no bottleneck). 44 452. Do the RCU read-side critical sections make proper use of 46 rcu_read_lock() and friends? These primitives are needed 47 to prevent grace periods from ending prematurely, which 48 could result in data being unceremoniously freed out from 49 under your read-side code, which can greatly increase the 50 actuarial risk of your kernel. 51 52 As a rough rule of thumb, any dereference of an RCU-protected 53 pointer must be covered by rcu_read_lock(), rcu_read_lock_bh(), 54 rcu_read_lock_sched(), or by the appropriate update-side lock. 55 Disabling of preemption can serve as rcu_read_lock_sched(), but 56 is less readable. 57 583. Does the update code tolerate concurrent accesses? 59 60 The whole point of RCU is to permit readers to run without 61 any locks or atomic operations. This means that readers will 62 be running while updates are in progress. There are a number 63 of ways to handle this concurrency, depending on the situation: 64 65 a. Use the RCU variants of the list and hlist update 66 primitives to add, remove, and replace elements on 67 an RCU-protected list. Alternatively, use the other 68 RCU-protected data structures that have been added to 69 the Linux kernel. 70 71 This is almost always the best approach. 72 73 b. Proceed as in (a) above, but also maintain per-element 74 locks (that are acquired by both readers and writers) 75 that guard per-element state. Of course, fields that 76 the readers refrain from accessing can be guarded by 77 some other lock acquired only by updaters, if desired. 78 79 This works quite well, also. 80 81 c. Make updates appear atomic to readers. For example, 82 pointer updates to properly aligned fields will 83 appear atomic, as will individual atomic primitives. 84 Sequences of perations performed under a lock will -not- 85 appear to be atomic to RCU readers, nor will sequences 86 of multiple atomic primitives. 87 88 This can work, but is starting to get a bit tricky. 89 90 d. Carefully order the updates and the reads so that 91 readers see valid data at all phases of the update. 92 This is often more difficult than it sounds, especially 93 given modern CPUs' tendency to reorder memory references. 94 One must usually liberally sprinkle memory barriers 95 (smp_wmb(), smp_rmb(), smp_mb()) through the code, 96 making it difficult to understand and to test. 97 98 It is usually better to group the changing data into 99 a separate structure, so that the change may be made 100 to appear atomic by updating a pointer to reference 101 a new structure containing updated values. 102 1034. Weakly ordered CPUs pose special challenges. Almost all CPUs 104 are weakly ordered -- even x86 CPUs allow later loads to be 105 reordered to precede earlier stores. RCU code must take all of 106 the following measures to prevent memory-corruption problems: 107 108 a. Readers must maintain proper ordering of their memory 109 accesses. The rcu_dereference() primitive ensures that 110 the CPU picks up the pointer before it picks up the data 111 that the pointer points to. This really is necessary 112 on Alpha CPUs. If you don't believe me, see: 113 114 http://www.openvms.compaq.com/wizard/wiz_2637.html 115 116 The rcu_dereference() primitive is also an excellent 117 documentation aid, letting the person reading the code 118 know exactly which pointers are protected by RCU. 119 Please note that compilers can also reorder code, and 120 they are becoming increasingly aggressive about doing 121 just that. The rcu_dereference() primitive therefore 122 also prevents destructive compiler optimizations. 123 124 The rcu_dereference() primitive is used by the 125 various "_rcu()" list-traversal primitives, such 126 as the list_for_each_entry_rcu(). Note that it is 127 perfectly legal (if redundant) for update-side code to 128 use rcu_dereference() and the "_rcu()" list-traversal 129 primitives. This is particularly useful in code that 130 is common to readers and updaters. However, lockdep 131 will complain if you access rcu_dereference() outside 132 of an RCU read-side critical section. See lockdep.txt 133 to learn what to do about this. 134 135 Of course, neither rcu_dereference() nor the "_rcu()" 136 list-traversal primitives can substitute for a good 137 concurrency design coordinating among multiple updaters. 138 139 b. If the list macros are being used, the list_add_tail_rcu() 140 and list_add_rcu() primitives must be used in order 141 to prevent weakly ordered machines from misordering 142 structure initialization and pointer planting. 143 Similarly, if the hlist macros are being used, the 144 hlist_add_head_rcu() primitive is required. 145 146 c. If the list macros are being used, the list_del_rcu() 147 primitive must be used to keep list_del()'s pointer 148 poisoning from inflicting toxic effects on concurrent 149 readers. Similarly, if the hlist macros are being used, 150 the hlist_del_rcu() primitive is required. 151 152 The list_replace_rcu() and hlist_replace_rcu() primitives 153 may be used to replace an old structure with a new one 154 in their respective types of RCU-protected lists. 155 156 d. Rules similar to (4b) and (4c) apply to the "hlist_nulls" 157 type of RCU-protected linked lists. 158 159 e. Updates must ensure that initialization of a given 160 structure happens before pointers to that structure are 161 publicized. Use the rcu_assign_pointer() primitive 162 when publicizing a pointer to a structure that can 163 be traversed by an RCU read-side critical section. 164 1655. If call_rcu(), or a related primitive such as call_rcu_bh() or 166 call_rcu_sched(), is used, the callback function must be 167 written to be called from softirq context. In particular, 168 it cannot block. 169 1706. Since synchronize_rcu() can block, it cannot be called from 171 any sort of irq context. The same rule applies for 172 synchronize_rcu_bh(), synchronize_sched(), synchronize_srcu(), 173 synchronize_rcu_expedited(), synchronize_rcu_bh_expedited(), 174 synchronize_sched_expedite(), and synchronize_srcu_expedited(). 175 176 The expedited forms of these primitives have the same semantics 177 as the non-expedited forms, but expediting is both expensive 178 and unfriendly to real-time workloads. Use of the expedited 179 primitives should be restricted to rare configuration-change 180 operations that would not normally be undertaken while a real-time 181 workload is running. 182 1837. If the updater uses call_rcu() or synchronize_rcu(), then the 184 corresponding readers must use rcu_read_lock() and 185 rcu_read_unlock(). If the updater uses call_rcu_bh() or 186 synchronize_rcu_bh(), then the corresponding readers must 187 use rcu_read_lock_bh() and rcu_read_unlock_bh(). If the 188 updater uses call_rcu_sched() or synchronize_sched(), then 189 the corresponding readers must disable preemption, possibly 190 by calling rcu_read_lock_sched() and rcu_read_unlock_sched(). 191 If the updater uses synchronize_srcu(), the the corresponding 192 readers must use srcu_read_lock() and srcu_read_unlock(), 193 and with the same srcu_struct. The rules for the expedited 194 primitives are the same as for their non-expedited counterparts. 195 Mixing things up will result in confusion and broken kernels. 196 197 One exception to this rule: rcu_read_lock() and rcu_read_unlock() 198 may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh() 199 in cases where local bottom halves are already known to be 200 disabled, for example, in irq or softirq context. Commenting 201 such cases is a must, of course! And the jury is still out on 202 whether the increased speed is worth it. 203 2048. Although synchronize_rcu() is slower than is call_rcu(), it 205 usually results in simpler code. So, unless update performance 206 is critically important or the updaters cannot block, 207 synchronize_rcu() should be used in preference to call_rcu(). 208 209 An especially important property of the synchronize_rcu() 210 primitive is that it automatically self-limits: if grace periods 211 are delayed for whatever reason, then the synchronize_rcu() 212 primitive will correspondingly delay updates. In contrast, 213 code using call_rcu() should explicitly limit update rate in 214 cases where grace periods are delayed, as failing to do so can 215 result in excessive realtime latencies or even OOM conditions. 216 217 Ways of gaining this self-limiting property when using call_rcu() 218 include: 219 220 a. Keeping a count of the number of data-structure elements 221 used by the RCU-protected data structure, including those 222 waiting for a grace period to elapse. Enforce a limit 223 on this number, stalling updates as needed to allow 224 previously deferred frees to complete. 225 226 Alternatively, limit only the number awaiting deferred 227 free rather than the total number of elements. 228 229 b. Limiting update rate. For example, if updates occur only 230 once per hour, then no explicit rate limiting is required, 231 unless your system is already badly broken. The dcache 232 subsystem takes this approach -- updates are guarded 233 by a global lock, limiting their rate. 234 235 c. Trusted update -- if updates can only be done manually by 236 superuser or some other trusted user, then it might not 237 be necessary to automatically limit them. The theory 238 here is that superuser already has lots of ways to crash 239 the machine. 240 241 d. Use call_rcu_bh() rather than call_rcu(), in order to take 242 advantage of call_rcu_bh()'s faster grace periods. 243 244 e. Periodically invoke synchronize_rcu(), permitting a limited 245 number of updates per grace period. 246 247 The same cautions apply to call_rcu_bh() and call_rcu_sched(). 248 2499. All RCU list-traversal primitives, which include 250 rcu_dereference(), list_for_each_entry_rcu(), 251 list_for_each_continue_rcu(), and list_for_each_safe_rcu(), 252 must be either within an RCU read-side critical section or 253 must be protected by appropriate update-side locks. RCU 254 read-side critical sections are delimited by rcu_read_lock() 255 and rcu_read_unlock(), or by similar primitives such as 256 rcu_read_lock_bh() and rcu_read_unlock_bh(), in which case 257 the matching rcu_dereference() primitive must be used in order 258 to keep lockdep happy, in this case, rcu_dereference_bh(). 259 260 The reason that it is permissible to use RCU list-traversal 261 primitives when the update-side lock is held is that doing so 262 can be quite helpful in reducing code bloat when common code is 263 shared between readers and updaters. Additional primitives 264 are provided for this case, as discussed in lockdep.txt. 265 26610. Conversely, if you are in an RCU read-side critical section, 267 and you don't hold the appropriate update-side lock, you -must- 268 use the "_rcu()" variants of the list macros. Failing to do so 269 will break Alpha, cause aggressive compilers to generate bad code, 270 and confuse people trying to read your code. 271 27211. Note that synchronize_rcu() -only- guarantees to wait until 273 all currently executing rcu_read_lock()-protected RCU read-side 274 critical sections complete. It does -not- necessarily guarantee 275 that all currently running interrupts, NMIs, preempt_disable() 276 code, or idle loops will complete. Therefore, if you do not have 277 rcu_read_lock()-protected read-side critical sections, do -not- 278 use synchronize_rcu(). 279 280 Similarly, disabling preemption is not an acceptable substitute 281 for rcu_read_lock(). Code that attempts to use preemption 282 disabling where it should be using rcu_read_lock() will break 283 in real-time kernel builds. 284 285 If you want to wait for interrupt handlers, NMI handlers, and 286 code under the influence of preempt_disable(), you instead 287 need to use synchronize_irq() or synchronize_sched(). 288 28912. Any lock acquired by an RCU callback must be acquired elsewhere 290 with softirq disabled, e.g., via spin_lock_irqsave(), 291 spin_lock_bh(), etc. Failing to disable irq on a given 292 acquisition of that lock will result in deadlock as soon as 293 the RCU softirq handler happens to run your RCU callback while 294 interrupting that acquisition's critical section. 295 29613. RCU callbacks can be and are executed in parallel. In many cases, 297 the callback code simply wrappers around kfree(), so that this 298 is not an issue (or, more accurately, to the extent that it is 299 an issue, the memory-allocator locking handles it). However, 300 if the callbacks do manipulate a shared data structure, they 301 must use whatever locking or other synchronization is required 302 to safely access and/or modify that data structure. 303 304 RCU callbacks are -usually- executed on the same CPU that executed 305 the corresponding call_rcu(), call_rcu_bh(), or call_rcu_sched(), 306 but are by -no- means guaranteed to be. For example, if a given 307 CPU goes offline while having an RCU callback pending, then that 308 RCU callback will execute on some surviving CPU. (If this was 309 not the case, a self-spawning RCU callback would prevent the 310 victim CPU from ever going offline.) 311 31214. SRCU (srcu_read_lock(), srcu_read_unlock(), srcu_dereference(), 313 synchronize_srcu(), and synchronize_srcu_expedited()) may only 314 be invoked from process context. Unlike other forms of RCU, it 315 -is- permissible to block in an SRCU read-side critical section 316 (demarked by srcu_read_lock() and srcu_read_unlock()), hence the 317 "SRCU": "sleepable RCU". Please note that if you don't need 318 to sleep in read-side critical sections, you should be using 319 RCU rather than SRCU, because RCU is almost always faster and 320 easier to use than is SRCU. 321 322 Also unlike other forms of RCU, explicit initialization 323 and cleanup is required via init_srcu_struct() and 324 cleanup_srcu_struct(). These are passed a "struct srcu_struct" 325 that defines the scope of a given SRCU domain. Once initialized, 326 the srcu_struct is passed to srcu_read_lock(), srcu_read_unlock() 327 synchronize_srcu(), and synchronize_srcu_expedited(). A given 328 synchronize_srcu() waits only for SRCU read-side critical 329 sections governed by srcu_read_lock() and srcu_read_unlock() 330 calls that have been passed the same srcu_struct. This property 331 is what makes sleeping read-side critical sections tolerable -- 332 a given subsystem delays only its own updates, not those of other 333 subsystems using SRCU. Therefore, SRCU is less prone to OOM the 334 system than RCU would be if RCU's read-side critical sections 335 were permitted to sleep. 336 337 The ability to sleep in read-side critical sections does not 338 come for free. First, corresponding srcu_read_lock() and 339 srcu_read_unlock() calls must be passed the same srcu_struct. 340 Second, grace-period-detection overhead is amortized only 341 over those updates sharing a given srcu_struct, rather than 342 being globally amortized as they are for other forms of RCU. 343 Therefore, SRCU should be used in preference to rw_semaphore 344 only in extremely read-intensive situations, or in situations 345 requiring SRCU's read-side deadlock immunity or low read-side 346 realtime latency. 347 348 Note that, rcu_assign_pointer() relates to SRCU just as they do 349 to other forms of RCU. 350 35115. The whole point of call_rcu(), synchronize_rcu(), and friends 352 is to wait until all pre-existing readers have finished before 353 carrying out some otherwise-destructive operation. It is 354 therefore critically important to -first- remove any path 355 that readers can follow that could be affected by the 356 destructive operation, and -only- -then- invoke call_rcu(), 357 synchronize_rcu(), or friends. 358 359 Because these primitives only wait for pre-existing readers, it 360 is the caller's responsibility to guarantee that any subsequent 361 readers will execute safely. 362 36316. The various RCU read-side primitives do -not- necessarily contain 364 memory barriers. You should therefore plan for the CPU 365 and the compiler to freely reorder code into and out of RCU 366 read-side critical sections. It is the responsibility of the 367 RCU update-side primitives to deal with this. 368