1 .. SPDX-License-Identifier: GPL-2.0 2 3 ================================ 4 Review Checklist for RCU Patches 5 ================================ 6 7 8 This document contains a checklist for producing and reviewing patches 9 that make use of RCU. Violating any of the rules listed below will 10 result in the same sorts of problems that leaving out a locking primitive 11 would cause. This list is based on experiences reviewing such patches 12 over a rather long period of time, but improvements are always welcome! 13 14 0. 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 can provide the same simplifications to certain types 36 of lockless algorithms that garbage collectors do. 37 38 1. 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 53 on large systems (for example, if the task is updating information 54 relating to itself that other tasks can read, there by definition 55 can be no bottleneck). Note that the definition of "large" has 56 changed significantly: Eight CPUs was "large" in the year 2000, 57 but a hundred CPUs was unremarkable in 2017. 58 59 2. 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 Explicit disabling of preemption (preempt_disable(), for example) 70 can serve as rcu_read_lock_sched(), but is less readable and 71 prevents lockdep from detecting locking issues. Acquiring a 72 spinlock also enters an RCU read-side critical section. 73 74 Please note that you *cannot* rely on code known to be built 75 only in non-preemptible kernels. Such code can and will break, 76 especially in kernels built with CONFIG_PREEMPT_COUNT=y. 77 78 Letting RCU-protected pointers "leak" out of an RCU read-side 79 critical section is every bit as bad as letting them leak out 80 from under a lock. Unless, of course, you have arranged some 81 other means of protection, such as a lock or a reference count 82 *before* letting them out of the RCU read-side critical section. 83 84 3. Does the update code tolerate concurrent accesses? 85 86 The whole point of RCU is to permit readers to run without 87 any locks or atomic operations. This means that readers will 88 be running while updates are in progress. There are a number 89 of ways to handle this concurrency, depending on the situation: 90 91 a. Use the RCU variants of the list and hlist update 92 primitives to add, remove, and replace elements on 93 an RCU-protected list. Alternatively, use the other 94 RCU-protected data structures that have been added to 95 the Linux kernel. 96 97 This is almost always the best approach. 98 99 b. Proceed as in (a) above, but also maintain per-element 100 locks (that are acquired by both readers and writers) 101 that guard per-element state. Fields that the readers 102 refrain from accessing can be guarded by some other lock 103 acquired only by updaters, if desired. 104 105 This also works quite well. 106 107 c. Make updates appear atomic to readers. For example, 108 pointer updates to properly aligned fields will 109 appear atomic, as will individual atomic primitives. 110 Sequences of operations performed under a lock will *not* 111 appear to be atomic to RCU readers, nor will sequences 112 of multiple atomic primitives. One alternative is to 113 move multiple individual fields to a separate structure, 114 thus solving the multiple-field problem by imposing an 115 additional level of indirection. 116 117 This can work, but is starting to get a bit tricky. 118 119 d. Carefully order the updates and the reads so that readers 120 see valid data at all phases of the update. This is often 121 more difficult than it sounds, especially given modern 122 CPUs' tendency to reorder memory references. One must 123 usually liberally sprinkle memory-ordering operations 124 through the code, making it difficult to understand and 125 to test. Where it works, it is better to use things 126 like smp_store_release() and smp_load_acquire(), but in 127 some cases the smp_mb() full memory barrier is required. 128 129 As noted earlier, it is usually better to group the 130 changing data into a separate structure, so that the 131 change may be made to appear atomic by updating a pointer 132 to reference a new structure containing updated values. 133 134 4. Weakly ordered CPUs pose special challenges. Almost all CPUs 135 are weakly ordered -- even x86 CPUs allow later loads to be 136 reordered to precede earlier stores. RCU code must take all of 137 the following measures to prevent memory-corruption problems: 138 139 a. Readers must maintain proper ordering of their memory 140 accesses. The rcu_dereference() primitive ensures that 141 the CPU picks up the pointer before it picks up the data 142 that the pointer points to. This really is necessary 143 on Alpha CPUs. 144 145 The rcu_dereference() primitive is also an excellent 146 documentation aid, letting the person reading the 147 code know exactly which pointers are protected by RCU. 148 Please note that compilers can also reorder code, and 149 they are becoming increasingly aggressive about doing 150 just that. The rcu_dereference() primitive therefore also 151 prevents destructive compiler optimizations. However, 152 with a bit of devious creativity, it is possible to 153 mishandle the return value from rcu_dereference(). 154 Please see rcu_dereference.rst for more information. 155 156 The rcu_dereference() primitive is used by the 157 various "_rcu()" list-traversal primitives, such 158 as the list_for_each_entry_rcu(). Note that it is 159 perfectly legal (if redundant) for update-side code to 160 use rcu_dereference() and the "_rcu()" list-traversal 161 primitives. This is particularly useful in code that 162 is common to readers and updaters. However, lockdep 163 will complain if you access rcu_dereference() outside 164 of an RCU read-side critical section. See lockdep.rst 165 to learn what to do about this. 166 167 Of course, neither rcu_dereference() nor the "_rcu()" 168 list-traversal primitives can substitute for a good 169 concurrency design coordinating among multiple updaters. 170 171 b. If the list macros are being used, the list_add_tail_rcu() 172 and list_add_rcu() primitives must be used in order 173 to prevent weakly ordered machines from misordering 174 structure initialization and pointer planting. 175 Similarly, if the hlist macros are being used, the 176 hlist_add_head_rcu() primitive is required. 177 178 c. If the list macros are being used, the list_del_rcu() 179 primitive must be used to keep list_del()'s pointer 180 poisoning from inflicting toxic effects on concurrent 181 readers. Similarly, if the hlist macros are being used, 182 the hlist_del_rcu() primitive is required. 183 184 The list_replace_rcu() and hlist_replace_rcu() primitives 185 may be used to replace an old structure with a new one 186 in their respective types of RCU-protected lists. 187 188 d. Rules similar to (4b) and (4c) apply to the "hlist_nulls" 189 type of RCU-protected linked lists. 190 191 e. Updates must ensure that initialization of a given 192 structure happens before pointers to that structure are 193 publicized. Use the rcu_assign_pointer() primitive 194 when publicizing a pointer to a structure that can 195 be traversed by an RCU read-side critical section. 196 197 5. If any of call_rcu(), call_srcu(), call_rcu_tasks(), or 198 call_rcu_tasks_trace() is used, the callback function may be 199 invoked from softirq context, and in any case with bottom halves 200 disabled. In particular, this callback function cannot block. 201 If you need the callback to block, run that code in a workqueue 202 handler scheduled from the callback. The queue_rcu_work() 203 function does this for you in the case of call_rcu(). 204 205 6. Since synchronize_rcu() can block, it cannot be called 206 from any sort of irq context. The same rule applies 207 for synchronize_srcu(), synchronize_rcu_expedited(), 208 synchronize_srcu_expedited(), synchronize_rcu_tasks(), 209 synchronize_rcu_tasks_rude(), and synchronize_rcu_tasks_trace(). 210 211 The expedited forms of these primitives have the same semantics 212 as the non-expedited forms, but expediting is more CPU intensive. 213 Use of the expedited primitives should be restricted to rare 214 configuration-change operations that would not normally be 215 undertaken while a real-time workload is running. Note that 216 IPI-sensitive real-time workloads can use the rcupdate.rcu_normal 217 kernel boot parameter to completely disable expedited grace 218 periods, though this might have performance implications. 219 220 In particular, if you find yourself invoking one of the expedited 221 primitives repeatedly in a loop, please do everyone a favor: 222 Restructure your code so that it batches the updates, allowing 223 a single non-expedited primitive to cover the entire batch. 224 This will very likely be faster than the loop containing the 225 expedited primitive, and will be much much easier on the rest 226 of the system, especially to real-time workloads running on the 227 rest of the system. Alternatively, instead use asynchronous 228 primitives such as call_rcu(). 229 230 7. As of v4.20, a given kernel implements only one RCU flavor, which 231 is RCU-sched for PREEMPTION=n and RCU-preempt for PREEMPTION=y. 232 If the updater uses call_rcu() or synchronize_rcu(), then 233 the corresponding readers may use: (1) rcu_read_lock() and 234 rcu_read_unlock(), (2) any pair of primitives that disables 235 and re-enables softirq, for example, rcu_read_lock_bh() and 236 rcu_read_unlock_bh(), or (3) any pair of primitives that disables 237 and re-enables preemption, for example, rcu_read_lock_sched() and 238 rcu_read_unlock_sched(). If the updater uses synchronize_srcu() 239 or call_srcu(), then the corresponding readers must use 240 srcu_read_lock() and srcu_read_unlock(), and with the same 241 srcu_struct. The rules for the expedited RCU grace-period-wait 242 primitives are the same as for their non-expedited counterparts. 243 244 Similarly, it is necessary to correctly use the RCU Tasks flavors: 245 246 a. If the updater uses synchronize_rcu_tasks() or 247 call_rcu_tasks(), then the readers must refrain from 248 executing voluntary context switches, that is, from 249 blocking. 250 251 b. If the updater uses call_rcu_tasks_trace() 252 or synchronize_rcu_tasks_trace(), then the 253 corresponding readers must use rcu_read_lock_trace() 254 and rcu_read_unlock_trace(). 255 256 c. If an updater uses synchronize_rcu_tasks_rude(), 257 then the corresponding readers must use anything that 258 disables preemption, for example, preempt_disable() 259 and preempt_enable(). 260 261 Mixing things up will result in confusion and broken kernels, and 262 has even resulted in an exploitable security issue. Therefore, 263 when using non-obvious pairs of primitives, commenting is 264 of course a must. One example of non-obvious pairing is 265 the XDP feature in networking, which calls BPF programs from 266 network-driver NAPI (softirq) context. BPF relies heavily on RCU 267 protection for its data structures, but because the BPF program 268 invocation happens entirely within a single local_bh_disable() 269 section in a NAPI poll cycle, this usage is safe. The reason 270 that this usage is safe is that readers can use anything that 271 disables BH when updaters use call_rcu() or synchronize_rcu(). 272 273 8. Although synchronize_rcu() is slower than is call_rcu(), 274 it usually results in simpler code. So, unless update 275 performance is critically important, the updaters cannot block, 276 or the latency of synchronize_rcu() is visible from userspace, 277 synchronize_rcu() should be used in preference to call_rcu(). 278 Furthermore, kfree_rcu() and kvfree_rcu() usually result 279 in even simpler code than does synchronize_rcu() without 280 synchronize_rcu()'s multi-millisecond latency. So please take 281 advantage of kfree_rcu()'s and kvfree_rcu()'s "fire and forget" 282 memory-freeing capabilities where it applies. 283 284 An especially important property of the synchronize_rcu() 285 primitive is that it automatically self-limits: if grace periods 286 are delayed for whatever reason, then the synchronize_rcu() 287 primitive will correspondingly delay updates. In contrast, 288 code using call_rcu() should explicitly limit update rate in 289 cases where grace periods are delayed, as failing to do so can 290 result in excessive realtime latencies or even OOM conditions. 291 292 Ways of gaining this self-limiting property when using call_rcu(), 293 kfree_rcu(), or kvfree_rcu() include: 294 295 a. Keeping a count of the number of data-structure elements 296 used by the RCU-protected data structure, including 297 those waiting for a grace period to elapse. Enforce a 298 limit on this number, stalling updates as needed to allow 299 previously deferred frees to complete. Alternatively, 300 limit only the number awaiting deferred free rather than 301 the total number of elements. 302 303 One way to stall the updates is to acquire the update-side 304 mutex. (Don't try this with a spinlock -- other CPUs 305 spinning on the lock could prevent the grace period 306 from ever ending.) Another way to stall the updates 307 is for the updates to use a wrapper function around 308 the memory allocator, so that this wrapper function 309 simulates OOM when there is too much memory awaiting an 310 RCU grace period. There are of course many other 311 variations on this theme. 312 313 b. Limiting update rate. For example, if updates occur only 314 once per hour, then no explicit rate limiting is 315 required, unless your system is already badly broken. 316 Older versions of the dcache subsystem take this approach, 317 guarding updates with a global lock, limiting their rate. 318 319 c. Trusted update -- if updates can only be done manually by 320 superuser or some other trusted user, then it might not 321 be necessary to automatically limit them. The theory 322 here is that superuser already has lots of ways to crash 323 the machine. 324 325 d. Periodically invoke rcu_barrier(), permitting a limited 326 number of updates per grace period. 327 328 The same cautions apply to call_srcu(), call_rcu_tasks(), and 329 call_rcu_tasks_trace(). This is why there is an srcu_barrier(), 330 rcu_barrier_tasks(), and rcu_barrier_tasks_trace(), respectively. 331 332 Note that although these primitives do take action to avoid 333 memory exhaustion when any given CPU has too many callbacks, 334 a determined user or administrator can still exhaust memory. 335 This is especially the case if a system with a large number of 336 CPUs has been configured to offload all of its RCU callbacks onto 337 a single CPU, or if the system has relatively little free memory. 338 339 9. All RCU list-traversal primitives, which include 340 rcu_dereference(), list_for_each_entry_rcu(), and 341 list_for_each_safe_rcu(), must be either within an RCU read-side 342 critical section or must be protected by appropriate update-side 343 locks. RCU read-side critical sections are delimited by 344 rcu_read_lock() and rcu_read_unlock(), or by similar primitives 345 such as rcu_read_lock_bh() and rcu_read_unlock_bh(), in which 346 case the matching rcu_dereference() primitive must be used in 347 order to keep lockdep happy, in this case, rcu_dereference_bh(). 348 349 The reason that it is permissible to use RCU list-traversal 350 primitives when the update-side lock is held is that doing so 351 can be quite helpful in reducing code bloat when common code is 352 shared between readers and updaters. Additional primitives 353 are provided for this case, as discussed in lockdep.rst. 354 355 One exception to this rule is when data is only ever added to 356 the linked data structure, and is never removed during any 357 time that readers might be accessing that structure. In such 358 cases, READ_ONCE() may be used in place of rcu_dereference() 359 and the read-side markers (rcu_read_lock() and rcu_read_unlock(), 360 for example) may be omitted. 361 362 10. Conversely, if you are in an RCU read-side critical section, 363 and you don't hold the appropriate update-side lock, you *must* 364 use the "_rcu()" variants of the list macros. Failing to do so 365 will break Alpha, cause aggressive compilers to generate bad code, 366 and confuse people trying to understand your code. 367 368 11. Any lock acquired by an RCU callback must be acquired elsewhere 369 with softirq disabled, e.g., via spin_lock_bh(). Failing to 370 disable softirq on a given acquisition of that lock will result 371 in deadlock as soon as the RCU softirq handler happens to run 372 your RCU callback while interrupting that acquisition's critical 373 section. 374 375 12. RCU callbacks can be and are executed in parallel. In many cases, 376 the callback code simply wrappers around kfree(), so that this 377 is not an issue (or, more accurately, to the extent that it is 378 an issue, the memory-allocator locking handles it). However, 379 if the callbacks do manipulate a shared data structure, they 380 must use whatever locking or other synchronization is required 381 to safely access and/or modify that data structure. 382 383 Do not assume that RCU callbacks will be executed on the same 384 CPU that executed the corresponding call_rcu(), call_srcu(), 385 call_rcu_tasks(), or call_rcu_tasks_trace(). For example, if 386 a given CPU goes offline while having an RCU callback pending, 387 then that RCU callback will execute on some surviving CPU. 388 (If this was not the case, a self-spawning RCU callback would 389 prevent the victim CPU from ever going offline.) Furthermore, 390 CPUs designated by rcu_nocbs= might well *always* have their 391 RCU callbacks executed on some other CPUs, in fact, for some 392 real-time workloads, this is the whole point of using the 393 rcu_nocbs= kernel boot parameter. 394 395 In addition, do not assume that callbacks queued in a given order 396 will be invoked in that order, even if they all are queued on the 397 same CPU. Furthermore, do not assume that same-CPU callbacks will 398 be invoked serially. For example, in recent kernels, CPUs can be 399 switched between offloaded and de-offloaded callback invocation, 400 and while a given CPU is undergoing such a switch, its callbacks 401 might be concurrently invoked by that CPU's softirq handler and 402 that CPU's rcuo kthread. At such times, that CPU's callbacks 403 might be executed both concurrently and out of order. 404 405 13. Unlike most flavors of RCU, it *is* permissible to block in an 406 SRCU read-side critical section (demarked by srcu_read_lock() 407 and srcu_read_unlock()), hence the "SRCU": "sleepable RCU". 408 Please note that if you don't need to sleep in read-side critical 409 sections, you should be using RCU rather than SRCU, because RCU 410 is almost always faster and easier to use than is SRCU. 411 412 Also unlike other forms of RCU, explicit initialization and 413 cleanup is required either at build time via DEFINE_SRCU() 414 or DEFINE_STATIC_SRCU() or at runtime via init_srcu_struct() 415 and cleanup_srcu_struct(). These last two are passed a 416 "struct srcu_struct" that defines the scope of a given 417 SRCU domain. Once initialized, the srcu_struct is passed 418 to srcu_read_lock(), srcu_read_unlock() synchronize_srcu(), 419 synchronize_srcu_expedited(), and call_srcu(). A given 420 synchronize_srcu() waits only for SRCU read-side critical 421 sections governed by srcu_read_lock() and srcu_read_unlock() 422 calls that have been passed the same srcu_struct. This property 423 is what makes sleeping read-side critical sections tolerable -- 424 a given subsystem delays only its own updates, not those of other 425 subsystems using SRCU. Therefore, SRCU is less prone to OOM the 426 system than RCU would be if RCU's read-side critical sections 427 were permitted to sleep. 428 429 The ability to sleep in read-side critical sections does not 430 come for free. First, corresponding srcu_read_lock() and 431 srcu_read_unlock() calls must be passed the same srcu_struct. 432 Second, grace-period-detection overhead is amortized only 433 over those updates sharing a given srcu_struct, rather than 434 being globally amortized as they are for other forms of RCU. 435 Therefore, SRCU should be used in preference to rw_semaphore 436 only in extremely read-intensive situations, or in situations 437 requiring SRCU's read-side deadlock immunity or low read-side 438 realtime latency. You should also consider percpu_rw_semaphore 439 when you need lightweight readers. 440 441 SRCU's expedited primitive (synchronize_srcu_expedited()) 442 never sends IPIs to other CPUs, so it is easier on 443 real-time workloads than is synchronize_rcu_expedited(). 444 445 It is also permissible to sleep in RCU Tasks Trace read-side 446 critical section, which are delimited by rcu_read_lock_trace() and 447 rcu_read_unlock_trace(). However, this is a specialized flavor 448 of RCU, and you should not use it without first checking with 449 its current users. In most cases, you should instead use SRCU. 450 451 Note that rcu_assign_pointer() relates to SRCU just as it does to 452 other forms of RCU, but instead of rcu_dereference() you should 453 use srcu_dereference() in order to avoid lockdep splats. 454 455 14. The whole point of call_rcu(), synchronize_rcu(), and friends 456 is to wait until all pre-existing readers have finished before 457 carrying out some otherwise-destructive operation. It is 458 therefore critically important to *first* remove any path 459 that readers can follow that could be affected by the 460 destructive operation, and *only then* invoke call_rcu(), 461 synchronize_rcu(), or friends. 462 463 Because these primitives only wait for pre-existing readers, it 464 is the caller's responsibility to guarantee that any subsequent 465 readers will execute safely. 466 467 15. The various RCU read-side primitives do *not* necessarily contain 468 memory barriers. You should therefore plan for the CPU 469 and the compiler to freely reorder code into and out of RCU 470 read-side critical sections. It is the responsibility of the 471 RCU update-side primitives to deal with this. 472 473 For SRCU readers, you can use smp_mb__after_srcu_read_unlock() 474 immediately after an srcu_read_unlock() to get a full barrier. 475 476 16. Use CONFIG_PROVE_LOCKING, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and the 477 __rcu sparse checks to validate your RCU code. These can help 478 find problems as follows: 479 480 CONFIG_PROVE_LOCKING: 481 check that accesses to RCU-protected data structures 482 are carried out under the proper RCU read-side critical 483 section, while holding the right combination of locks, 484 or whatever other conditions are appropriate. 485 486 CONFIG_DEBUG_OBJECTS_RCU_HEAD: 487 check that you don't pass the same object to call_rcu() 488 (or friends) before an RCU grace period has elapsed 489 since the last time that you passed that same object to 490 call_rcu() (or friends). 491 492 CONFIG_RCU_STRICT_GRACE_PERIOD: 493 combine with KASAN to check for pointers leaked out 494 of RCU read-side critical sections. This Kconfig 495 option is tough on both performance and scalability, 496 and so is limited to four-CPU systems. 497 498 __rcu sparse checks: 499 tag the pointer to the RCU-protected data structure 500 with __rcu, and sparse will warn you if you access that 501 pointer without the services of one of the variants 502 of rcu_dereference(). 503 504 These debugging aids can help you find problems that are 505 otherwise extremely difficult to spot. 506 507 17. If you pass a callback function defined within a module 508 to one of call_rcu(), call_srcu(), call_rcu_tasks(), or 509 call_rcu_tasks_trace(), then it is necessary to wait for all 510 pending callbacks to be invoked before unloading that module. 511 Note that it is absolutely *not* sufficient to wait for a grace 512 period! For example, synchronize_rcu() implementation is *not* 513 guaranteed to wait for callbacks registered on other CPUs via 514 call_rcu(). Or even on the current CPU if that CPU recently 515 went offline and came back online. 516 517 You instead need to use one of the barrier functions: 518 519 - call_rcu() -> rcu_barrier() 520 - call_srcu() -> srcu_barrier() 521 - call_rcu_tasks() -> rcu_barrier_tasks() 522 - call_rcu_tasks_trace() -> rcu_barrier_tasks_trace() 523 524 However, these barrier functions are absolutely *not* guaranteed 525 to wait for a grace period. For example, if there are no 526 call_rcu() callbacks queued anywhere in the system, rcu_barrier() 527 can and will return immediately. 528 529 So if you need to wait for both a grace period and for all 530 pre-existing callbacks, you will need to invoke both functions, 531 with the pair depending on the flavor of RCU: 532 533 - Either synchronize_rcu() or synchronize_rcu_expedited(), 534 together with rcu_barrier() 535 - Either synchronize_srcu() or synchronize_srcu_expedited(), 536 together with and srcu_barrier() 537 - synchronize_rcu_tasks() and rcu_barrier_tasks() 538 - synchronize_tasks_trace() and rcu_barrier_tasks_trace() 539 540 If necessary, you can use something like workqueues to execute 541 the requisite pair of functions concurrently. 542 543 See rcubarrier.rst for more information.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.