Based on kernel version 4.1. Page generated on 2015-06-28 12:14 EST.
1 Review Checklist for RCU Patches 2 3 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! 9 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 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 26 1. 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 45 2. 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 58 3. 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. 102 103 4. 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 118 code 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 also 122 prevents destructive compiler optimizations. However, 123 with a bit of devious creativity, it is possible to 124 mishandle the return value from rcu_dereference(). 125 Please see rcu_dereference.txt in this directory for 126 more information. 127 128 The rcu_dereference() primitive is used by the 129 various "_rcu()" list-traversal primitives, such 130 as the list_for_each_entry_rcu(). Note that it is 131 perfectly legal (if redundant) for update-side code to 132 use rcu_dereference() and the "_rcu()" list-traversal 133 primitives. This is particularly useful in code that 134 is common to readers and updaters. However, lockdep 135 will complain if you access rcu_dereference() outside 136 of an RCU read-side critical section. See lockdep.txt 137 to learn what to do about this. 138 139 Of course, neither rcu_dereference() nor the "_rcu()" 140 list-traversal primitives can substitute for a good 141 concurrency design coordinating among multiple updaters. 142 143 b. If the list macros are being used, the list_add_tail_rcu() 144 and list_add_rcu() primitives must be used in order 145 to prevent weakly ordered machines from misordering 146 structure initialization and pointer planting. 147 Similarly, if the hlist macros are being used, the 148 hlist_add_head_rcu() primitive is required. 149 150 c. If the list macros are being used, the list_del_rcu() 151 primitive must be used to keep list_del()'s pointer 152 poisoning from inflicting toxic effects on concurrent 153 readers. Similarly, if the hlist macros are being used, 154 the hlist_del_rcu() primitive is required. 155 156 The list_replace_rcu() and hlist_replace_rcu() primitives 157 may be used to replace an old structure with a new one 158 in their respective types of RCU-protected lists. 159 160 d. Rules similar to (4b) and (4c) apply to the "hlist_nulls" 161 type of RCU-protected linked lists. 162 163 e. Updates must ensure that initialization of a given 164 structure happens before pointers to that structure are 165 publicized. Use the rcu_assign_pointer() primitive 166 when publicizing a pointer to a structure that can 167 be traversed by an RCU read-side critical section. 168 169 5. If call_rcu(), or a related primitive such as call_rcu_bh(), 170 call_rcu_sched(), or call_srcu() is used, the callback function 171 must be written to be called from softirq context. In particular, 172 it cannot block. 173 174 6. Since synchronize_rcu() can block, it cannot be called from 175 any sort of irq context. The same rule applies for 176 synchronize_rcu_bh(), synchronize_sched(), synchronize_srcu(), 177 synchronize_rcu_expedited(), synchronize_rcu_bh_expedited(), 178 synchronize_sched_expedite(), and synchronize_srcu_expedited(). 179 180 The expedited forms of these primitives have the same semantics 181 as the non-expedited forms, but expediting is both expensive 182 and unfriendly to real-time workloads. Use of the expedited 183 primitives should be restricted to rare configuration-change 184 operations that would not normally be undertaken while a real-time 185 workload is running. 186 187 In particular, if you find yourself invoking one of the expedited 188 primitives repeatedly in a loop, please do everyone a favor: 189 Restructure your code so that it batches the updates, allowing 190 a single non-expedited primitive to cover the entire batch. 191 This will very likely be faster than the loop containing the 192 expedited primitive, and will be much much easier on the rest 193 of the system, especially to real-time workloads running on 194 the rest of the system. 195 196 In addition, it is illegal to call the expedited forms from 197 a CPU-hotplug notifier, or while holding a lock that is acquired 198 by a CPU-hotplug notifier. Failing to observe this restriction 199 will result in deadlock. 200 201 7. If the updater uses call_rcu() or synchronize_rcu(), then the 202 corresponding readers must use rcu_read_lock() and 203 rcu_read_unlock(). If the updater uses call_rcu_bh() or 204 synchronize_rcu_bh(), then the corresponding readers must 205 use rcu_read_lock_bh() and rcu_read_unlock_bh(). If the 206 updater uses call_rcu_sched() or synchronize_sched(), then 207 the corresponding readers must disable preemption, possibly 208 by calling rcu_read_lock_sched() and rcu_read_unlock_sched(). 209 If the updater uses synchronize_srcu() or call_srcu(), then 210 the corresponding readers must use srcu_read_lock() and 211 srcu_read_unlock(), and with the same srcu_struct. The rules for 212 the expedited primitives are the same as for their non-expedited 213 counterparts. Mixing things up will result in confusion and 214 broken kernels. 215 216 One exception to this rule: rcu_read_lock() and rcu_read_unlock() 217 may be substituted for rcu_read_lock_bh() and rcu_read_unlock_bh() 218 in cases where local bottom halves are already known to be 219 disabled, for example, in irq or softirq context. Commenting 220 such cases is a must, of course! And the jury is still out on 221 whether the increased speed is worth it. 222 223 8. Although synchronize_rcu() is slower than is call_rcu(), it 224 usually results in simpler code. So, unless update performance is 225 critically important, the updaters cannot block, or the latency of 226 synchronize_rcu() is visible from userspace, synchronize_rcu() 227 should be used in preference to call_rcu(). Furthermore, 228 kfree_rcu() usually results in even simpler code than does 229 synchronize_rcu() without synchronize_rcu()'s multi-millisecond 230 latency. So please take advantage of kfree_rcu()'s "fire and 231 forget" memory-freeing capabilities where it applies. 232 233 An especially important property of the synchronize_rcu() 234 primitive is that it automatically self-limits: if grace periods 235 are delayed for whatever reason, then the synchronize_rcu() 236 primitive will correspondingly delay updates. In contrast, 237 code using call_rcu() should explicitly limit update rate in 238 cases where grace periods are delayed, as failing to do so can 239 result in excessive realtime latencies or even OOM conditions. 240 241 Ways of gaining this self-limiting property when using call_rcu() 242 include: 243 244 a. Keeping a count of the number of data-structure elements 245 used by the RCU-protected data structure, including 246 those waiting for a grace period to elapse. Enforce a 247 limit on this number, stalling updates as needed to allow 248 previously deferred frees to complete. Alternatively, 249 limit only the number awaiting deferred free rather than 250 the total number of elements. 251 252 One way to stall the updates is to acquire the update-side 253 mutex. (Don't try this with a spinlock -- other CPUs 254 spinning on the lock could prevent the grace period 255 from ever ending.) Another way to stall the updates 256 is for the updates to use a wrapper function around 257 the memory allocator, so that this wrapper function 258 simulates OOM when there is too much memory awaiting an 259 RCU grace period. There are of course many other 260 variations on this theme. 261 262 b. Limiting update rate. For example, if updates occur only 263 once per hour, then no explicit rate limiting is 264 required, unless your system is already badly broken. 265 Older versions of the dcache subsystem take this approach, 266 guarding updates with a global lock, limiting their rate. 267 268 c. Trusted update -- if updates can only be done manually by 269 superuser or some other trusted user, then it might not 270 be necessary to automatically limit them. The theory 271 here is that superuser already has lots of ways to crash 272 the machine. 273 274 d. Use call_rcu_bh() rather than call_rcu(), in order to take 275 advantage of call_rcu_bh()'s faster grace periods. (This 276 is only a partial solution, though.) 277 278 e. Periodically invoke synchronize_rcu(), permitting a limited 279 number of updates per grace period. 280 281 The same cautions apply to call_rcu_bh(), call_rcu_sched(), 282 call_srcu(), and kfree_rcu(). 283 284 Note that although these primitives do take action to avoid memory 285 exhaustion when any given CPU has too many callbacks, a determined 286 user could still exhaust memory. This is especially the case 287 if a system with a large number of CPUs has been configured to 288 offload all of its RCU callbacks onto a single CPU, or if the 289 system has relatively little free memory. 290 291 9. All RCU list-traversal primitives, which include 292 rcu_dereference(), list_for_each_entry_rcu(), and 293 list_for_each_safe_rcu(), must be either within an RCU read-side 294 critical section or must be protected by appropriate update-side 295 locks. RCU read-side critical sections are delimited by 296 rcu_read_lock() and rcu_read_unlock(), or by similar primitives 297 such as rcu_read_lock_bh() and rcu_read_unlock_bh(), in which 298 case the matching rcu_dereference() primitive must be used in 299 order to keep lockdep happy, in this case, rcu_dereference_bh(). 300 301 The reason that it is permissible to use RCU list-traversal 302 primitives when the update-side lock is held is that doing so 303 can be quite helpful in reducing code bloat when common code is 304 shared between readers and updaters. Additional primitives 305 are provided for this case, as discussed in lockdep.txt. 306 307 10. Conversely, if you are in an RCU read-side critical section, 308 and you don't hold the appropriate update-side lock, you -must- 309 use the "_rcu()" variants of the list macros. Failing to do so 310 will break Alpha, cause aggressive compilers to generate bad code, 311 and confuse people trying to read your code. 312 313 11. Note that synchronize_rcu() -only- guarantees to wait until 314 all currently executing rcu_read_lock()-protected RCU read-side 315 critical sections complete. It does -not- necessarily guarantee 316 that all currently running interrupts, NMIs, preempt_disable() 317 code, or idle loops will complete. Therefore, if your 318 read-side critical sections are protected by something other 319 than rcu_read_lock(), do -not- use synchronize_rcu(). 320 321 Similarly, disabling preemption is not an acceptable substitute 322 for rcu_read_lock(). Code that attempts to use preemption 323 disabling where it should be using rcu_read_lock() will break 324 in real-time kernel builds. 325 326 If you want to wait for interrupt handlers, NMI handlers, and 327 code under the influence of preempt_disable(), you instead 328 need to use synchronize_irq() or synchronize_sched(). 329 330 This same limitation also applies to synchronize_rcu_bh() 331 and synchronize_srcu(), as well as to the asynchronous and 332 expedited forms of the three primitives, namely call_rcu(), 333 call_rcu_bh(), call_srcu(), synchronize_rcu_expedited(), 334 synchronize_rcu_bh_expedited(), and synchronize_srcu_expedited(). 335 336 12. Any lock acquired by an RCU callback must be acquired elsewhere 337 with softirq disabled, e.g., via spin_lock_irqsave(), 338 spin_lock_bh(), etc. Failing to disable irq on a given 339 acquisition of that lock will result in deadlock as soon as 340 the RCU softirq handler happens to run your RCU callback while 341 interrupting that acquisition's critical section. 342 343 13. RCU callbacks can be and are executed in parallel. In many cases, 344 the callback code simply wrappers around kfree(), so that this 345 is not an issue (or, more accurately, to the extent that it is 346 an issue, the memory-allocator locking handles it). However, 347 if the callbacks do manipulate a shared data structure, they 348 must use whatever locking or other synchronization is required 349 to safely access and/or modify that data structure. 350 351 RCU callbacks are -usually- executed on the same CPU that executed 352 the corresponding call_rcu(), call_rcu_bh(), or call_rcu_sched(), 353 but are by -no- means guaranteed to be. For example, if a given 354 CPU goes offline while having an RCU callback pending, then that 355 RCU callback will execute on some surviving CPU. (If this was 356 not the case, a self-spawning RCU callback would prevent the 357 victim CPU from ever going offline.) 358 359 14. SRCU (srcu_read_lock(), srcu_read_unlock(), srcu_dereference(), 360 synchronize_srcu(), synchronize_srcu_expedited(), and call_srcu()) 361 may only be invoked from process context. Unlike other forms of 362 RCU, it -is- permissible to block in an SRCU read-side critical 363 section (demarked by srcu_read_lock() and srcu_read_unlock()), 364 hence the "SRCU": "sleepable RCU". Please note that if you 365 don't need to sleep in read-side critical sections, you should be 366 using RCU rather than SRCU, because RCU is almost always faster 367 and easier to use than is SRCU. 368 369 Also unlike other forms of RCU, explicit initialization 370 and cleanup is required via init_srcu_struct() and 371 cleanup_srcu_struct(). These are passed a "struct srcu_struct" 372 that defines the scope of a given SRCU domain. Once initialized, 373 the srcu_struct is passed to srcu_read_lock(), srcu_read_unlock() 374 synchronize_srcu(), synchronize_srcu_expedited(), and call_srcu(). 375 A given synchronize_srcu() waits only for SRCU read-side critical 376 sections governed by srcu_read_lock() and srcu_read_unlock() 377 calls that have been passed the same srcu_struct. This property 378 is what makes sleeping read-side critical sections tolerable -- 379 a given subsystem delays only its own updates, not those of other 380 subsystems using SRCU. Therefore, SRCU is less prone to OOM the 381 system than RCU would be if RCU's read-side critical sections 382 were permitted to sleep. 383 384 The ability to sleep in read-side critical sections does not 385 come for free. First, corresponding srcu_read_lock() and 386 srcu_read_unlock() calls must be passed the same srcu_struct. 387 Second, grace-period-detection overhead is amortized only 388 over those updates sharing a given srcu_struct, rather than 389 being globally amortized as they are for other forms of RCU. 390 Therefore, SRCU should be used in preference to rw_semaphore 391 only in extremely read-intensive situations, or in situations 392 requiring SRCU's read-side deadlock immunity or low read-side 393 realtime latency. 394 395 Note that, rcu_assign_pointer() relates to SRCU just as it does 396 to other forms of RCU. 397 398 15. The whole point of call_rcu(), synchronize_rcu(), and friends 399 is to wait until all pre-existing readers have finished before 400 carrying out some otherwise-destructive operation. It is 401 therefore critically important to -first- remove any path 402 that readers can follow that could be affected by the 403 destructive operation, and -only- -then- invoke call_rcu(), 404 synchronize_rcu(), or friends. 405 406 Because these primitives only wait for pre-existing readers, it 407 is the caller's responsibility to guarantee that any subsequent 408 readers will execute safely. 409 410 16. The various RCU read-side primitives do -not- necessarily contain 411 memory barriers. You should therefore plan for the CPU 412 and the compiler to freely reorder code into and out of RCU 413 read-side critical sections. It is the responsibility of the 414 RCU update-side primitives to deal with this. 415 416 17. Use CONFIG_PROVE_RCU, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and the 417 __rcu sparse checks (enabled by CONFIG_SPARSE_RCU_POINTER) to 418 validate your RCU code. These can help find problems as follows: 419 420 CONFIG_PROVE_RCU: check that accesses to RCU-protected data 421 structures are carried out under the proper RCU 422 read-side critical section, while holding the right 423 combination of locks, or whatever other conditions 424 are appropriate. 425 426 CONFIG_DEBUG_OBJECTS_RCU_HEAD: check that you don't pass the 427 same object to call_rcu() (or friends) before an RCU 428 grace period has elapsed since the last time that you 429 passed that same object to call_rcu() (or friends). 430 431 __rcu sparse checks: tag the pointer to the RCU-protected data 432 structure with __rcu, and sparse will warn you if you 433 access that pointer without the services of one of the 434 variants of rcu_dereference(). 435 436 These debugging aids can help you find problems that are 437 otherwise extremely difficult to spot.