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 can provide the same simplifications to certain types 36 of lockless algorithms that garbage 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 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. 58 592. 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 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. 72 73 Please note that you *cannot* rely on code known to be built 74 only in non-preemptible kernels. Such code can and will break, 75 especially in kernels built with CONFIG_PREEMPT_COUNT=y. 76 77 Letting RCU-protected pointers "leak" out of an RCU read-side 78 critical section is every bit as bad as letting them leak out 79 from under a lock. Unless, of course, you have arranged some 80 other means of protection, such as a lock or a reference count 81 *before* letting them out of the RCU read-side critical section. 82 833. Does the update code tolerate concurrent accesses? 84 85 The whole point of RCU is to permit readers to run without 86 any locks or atomic operations. This means that readers will 87 be running while updates are in progress. There are a number 88 of ways to handle this concurrency, depending on the situation: 89 90 a. Use the RCU variants of the list and hlist update 91 primitives to add, remove, and replace elements on 92 an RCU-protected list. Alternatively, use the other 93 RCU-protected data structures that have been added to 94 the Linux kernel. 95 96 This is almost always the best approach. 97 98 b. Proceed as in (a) above, but also maintain per-element 99 locks (that are acquired by both readers and writers) 100 that guard per-element state. Fields that the readers 101 refrain from accessing can be guarded by some other lock 102 acquired only by updaters, if desired. 103 104 This also works quite well. 105 106 c. Make updates appear atomic to readers. For example, 107 pointer updates to properly aligned fields will 108 appear atomic, as will individual atomic primitives. 109 Sequences of operations performed under a lock will *not* 110 appear to be atomic to RCU readers, nor will sequences 111 of multiple atomic primitives. One alternative is to 112 move multiple individual fields to a separate structure, 113 thus solving the multiple-field problem by imposing an 114 additional level of indirection. 115 116 This can work, but is starting to get a bit tricky. 117 118 d. Carefully order the updates and the reads so that readers 119 see valid data at all phases of the update. This is often 120 more difficult than it sounds, especially given modern 121 CPUs' tendency to reorder memory references. One must 122 usually liberally sprinkle memory-ordering operations 123 through the code, making it difficult to understand and 124 to test. Where it works, it is better to use things 125 like smp_store_release() and smp_load_acquire(), but in 126 some cases the smp_mb() full memory barrier is required. 127 128 As noted earlier, it is usually better to group the 129 changing data into a separate structure, so that the 130 change may be made to appear atomic by updating a pointer 131 to reference a new structure containing updated values. 132 1334. Weakly ordered CPUs pose special challenges. Almost all CPUs 134 are weakly ordered -- even x86 CPUs allow later loads to be 135 reordered to precede earlier stores. RCU code must take all of 136 the following measures to prevent memory-corruption problems: 137 138 a. Readers must maintain proper ordering of their memory 139 accesses. The rcu_dereference() primitive ensures that 140 the CPU picks up the pointer before it picks up the data 141 that the pointer points to. This really is necessary 142 on Alpha CPUs. 143 144 The rcu_dereference() primitive is also an excellent 145 documentation aid, letting the person reading the 146 code know exactly which pointers are protected by RCU. 147 Please note that compilers can also reorder code, and 148 they are becoming increasingly aggressive about doing 149 just that. The rcu_dereference() primitive therefore also 150 prevents destructive compiler optimizations. However, 151 with a bit of devious creativity, it is possible to 152 mishandle the return value from rcu_dereference(). 153 Please see rcu_dereference.rst for more information. 154 155 The rcu_dereference() primitive is used by the 156 various "_rcu()" list-traversal primitives, such 157 as the list_for_each_entry_rcu(). Note that it is 158 perfectly legal (if redundant) for update-side code to 159 use rcu_dereference() and the "_rcu()" list-traversal 160 primitives. This is particularly useful in code that 161 is common to readers and updaters. However, lockdep 162 will complain if you access rcu_dereference() outside 163 of an RCU read-side critical section. See lockdep.rst 164 to learn what to do about this. 165 166 Of course, neither rcu_dereference() nor the "_rcu()" 167 list-traversal primitives can substitute for a good 168 concurrency design coordinating among multiple updaters. 169 170 b. If the list macros are being used, the list_add_tail_rcu() 171 and list_add_rcu() primitives must be used in order 172 to prevent weakly ordered machines from misordering 173 structure initialization and pointer planting. 174 Similarly, if the hlist macros are being used, the 175 hlist_add_head_rcu() primitive is required. 176 177 c. If the list macros are being used, the list_del_rcu() 178 primitive must be used to keep list_del()'s pointer 179 poisoning from inflicting toxic effects on concurrent 180 readers. Similarly, if the hlist macros are being used, 181 the hlist_del_rcu() primitive is required. 182 183 The list_replace_rcu() and hlist_replace_rcu() primitives 184 may be used to replace an old structure with a new one 185 in their respective types of RCU-protected lists. 186 187 d. Rules similar to (4b) and (4c) apply to the "hlist_nulls" 188 type of RCU-protected linked lists. 189 190 e. Updates must ensure that initialization of a given 191 structure happens before pointers to that structure are 192 publicized. Use the rcu_assign_pointer() primitive 193 when publicizing a pointer to a structure that can 194 be traversed by an RCU read-side critical section. 195 1965. If any of call_rcu(), call_srcu(), call_rcu_tasks(), 197 call_rcu_tasks_rude(), or call_rcu_tasks_trace() is used, 198 the callback function may be invoked from softirq context, 199 and in any case with bottom halves disabled. In particular, 200 this callback function cannot block. If you need the callback 201 to block, run that code in a workqueue handler scheduled from 202 the callback. The queue_rcu_work() function does this for you 203 in the case of call_rcu(). 204 2056. 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 2307. 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 If the updater uses call_rcu_tasks() or synchronize_rcu_tasks(), 245 then the readers must refrain from executing voluntary 246 context switches, that is, from blocking. If the updater uses 247 call_rcu_tasks_trace() or synchronize_rcu_tasks_trace(), then 248 the corresponding readers must use rcu_read_lock_trace() and 249 rcu_read_unlock_trace(). If an updater uses call_rcu_tasks_rude() 250 or synchronize_rcu_tasks_rude(), then the corresponding readers 251 must use anything that disables preemption, for example, 252 preempt_disable() and preempt_enable(). 253 254 Mixing things up will result in confusion and broken kernels, and 255 has even resulted in an exploitable security issue. Therefore, 256 when using non-obvious pairs of primitives, commenting is 257 of course a must. One example of non-obvious pairing is 258 the XDP feature in networking, which calls BPF programs from 259 network-driver NAPI (softirq) context. BPF relies heavily on RCU 260 protection for its data structures, but because the BPF program 261 invocation happens entirely within a single local_bh_disable() 262 section in a NAPI poll cycle, this usage is safe. The reason 263 that this usage is safe is that readers can use anything that 264 disables BH when updaters use call_rcu() or synchronize_rcu(). 265 2668. Although synchronize_rcu() is slower than is call_rcu(), 267 it usually results in simpler code. So, unless update 268 performance is critically important, the updaters cannot block, 269 or the latency of synchronize_rcu() is visible from userspace, 270 synchronize_rcu() should be used in preference to call_rcu(). 271 Furthermore, kfree_rcu() and kvfree_rcu() usually result 272 in even simpler code than does synchronize_rcu() without 273 synchronize_rcu()'s multi-millisecond latency. So please take 274 advantage of kfree_rcu()'s and kvfree_rcu()'s "fire and forget" 275 memory-freeing capabilities where it applies. 276 277 An especially important property of the synchronize_rcu() 278 primitive is that it automatically self-limits: if grace periods 279 are delayed for whatever reason, then the synchronize_rcu() 280 primitive will correspondingly delay updates. In contrast, 281 code using call_rcu() should explicitly limit update rate in 282 cases where grace periods are delayed, as failing to do so can 283 result in excessive realtime latencies or even OOM conditions. 284 285 Ways of gaining this self-limiting property when using call_rcu(), 286 kfree_rcu(), or kvfree_rcu() include: 287 288 a. Keeping a count of the number of data-structure elements 289 used by the RCU-protected data structure, including 290 those waiting for a grace period to elapse. Enforce a 291 limit on this number, stalling updates as needed to allow 292 previously deferred frees to complete. Alternatively, 293 limit only the number awaiting deferred free rather than 294 the total number of elements. 295 296 One way to stall the updates is to acquire the update-side 297 mutex. (Don't try this with a spinlock -- other CPUs 298 spinning on the lock could prevent the grace period 299 from ever ending.) Another way to stall the updates 300 is for the updates to use a wrapper function around 301 the memory allocator, so that this wrapper function 302 simulates OOM when there is too much memory awaiting an 303 RCU grace period. There are of course many other 304 variations on this theme. 305 306 b. Limiting update rate. For example, if updates occur only 307 once per hour, then no explicit rate limiting is 308 required, unless your system is already badly broken. 309 Older versions of the dcache subsystem take this approach, 310 guarding updates with a global lock, limiting their rate. 311 312 c. Trusted update -- if updates can only be done manually by 313 superuser or some other trusted user, then it might not 314 be necessary to automatically limit them. The theory 315 here is that superuser already has lots of ways to crash 316 the machine. 317 318 d. Periodically invoke rcu_barrier(), permitting a limited 319 number of updates per grace period. 320 321 The same cautions apply to call_srcu(), call_rcu_tasks(), 322 call_rcu_tasks_rude(), and call_rcu_tasks_trace(). This is 323 why there is an srcu_barrier(), rcu_barrier_tasks(), 324 rcu_barrier_tasks_rude(), and rcu_barrier_tasks_rude(), 325 respectively. 326 327 Note that although these primitives do take action to avoid 328 memory exhaustion when any given CPU has too many callbacks, 329 a determined user or administrator can still exhaust memory. 330 This is especially the case if a system with a large number of 331 CPUs has been configured to offload all of its RCU callbacks onto 332 a single CPU, or if the system has relatively little free memory. 333 3349. All RCU list-traversal primitives, which include 335 rcu_dereference(), list_for_each_entry_rcu(), and 336 list_for_each_safe_rcu(), must be either within an RCU read-side 337 critical section or must be protected by appropriate update-side 338 locks. RCU read-side critical sections are delimited by 339 rcu_read_lock() and rcu_read_unlock(), or by similar primitives 340 such as rcu_read_lock_bh() and rcu_read_unlock_bh(), in which 341 case the matching rcu_dereference() primitive must be used in 342 order to keep lockdep happy, in this case, rcu_dereference_bh(). 343 344 The reason that it is permissible to use RCU list-traversal 345 primitives when the update-side lock is held is that doing so 346 can be quite helpful in reducing code bloat when common code is 347 shared between readers and updaters. Additional primitives 348 are provided for this case, as discussed in lockdep.rst. 349 350 One exception to this rule is when data is only ever added to 351 the linked data structure, and is never removed during any 352 time that readers might be accessing that structure. In such 353 cases, READ_ONCE() may be used in place of rcu_dereference() 354 and the read-side markers (rcu_read_lock() and rcu_read_unlock(), 355 for example) may be omitted. 356 35710. Conversely, if you are in an RCU read-side critical section, 358 and you don't hold the appropriate update-side lock, you *must* 359 use the "_rcu()" variants of the list macros. Failing to do so 360 will break Alpha, cause aggressive compilers to generate bad code, 361 and confuse people trying to understand your code. 362 36311. Any lock acquired by an RCU callback must be acquired elsewhere 364 with softirq disabled, e.g., via spin_lock_bh(). Failing to 365 disable softirq on a given acquisition of that lock will result 366 in deadlock as soon as the RCU softirq handler happens to run 367 your RCU callback while interrupting that acquisition's critical 368 section. 369 37012. RCU callbacks can be and are executed in parallel. In many cases, 371 the callback code simply wrappers around kfree(), so that this 372 is not an issue (or, more accurately, to the extent that it is 373 an issue, the memory-allocator locking handles it). However, 374 if the callbacks do manipulate a shared data structure, they 375 must use whatever locking or other synchronization is required 376 to safely access and/or modify that data structure. 377 378 Do not assume that RCU callbacks will be executed on the same 379 CPU that executed the corresponding call_rcu() or call_srcu(). 380 For example, if a given CPU goes offline while having an RCU 381 callback pending, then that RCU callback will execute on some 382 surviving CPU. (If this was not the case, a self-spawning RCU 383 callback would prevent the victim CPU from ever going offline.) 384 Furthermore, CPUs designated by rcu_nocbs= might well *always* 385 have their RCU callbacks executed on some other CPUs, in fact, 386 for some real-time workloads, this is the whole point of using 387 the rcu_nocbs= kernel boot parameter. 388 389 In addition, do not assume that callbacks queued in a given order 390 will be invoked in that order, even if they all are queued on the 391 same CPU. Furthermore, do not assume that same-CPU callbacks will 392 be invoked serially. For example, in recent kernels, CPUs can be 393 switched between offloaded and de-offloaded callback invocation, 394 and while a given CPU is undergoing such a switch, its callbacks 395 might be concurrently invoked by that CPU's softirq handler and 396 that CPU's rcuo kthread. At such times, that CPU's callbacks 397 might be executed both concurrently and out of order. 398 39913. Unlike most flavors of RCU, it *is* permissible to block in an 400 SRCU read-side critical section (demarked by srcu_read_lock() 401 and srcu_read_unlock()), hence the "SRCU": "sleepable RCU". 402 Please note that if you don't need to sleep in read-side critical 403 sections, you should be using RCU rather than SRCU, because RCU 404 is almost always faster and easier to use than is SRCU. 405 406 Also unlike other forms of RCU, explicit initialization and 407 cleanup is required either at build time via DEFINE_SRCU() 408 or DEFINE_STATIC_SRCU() or at runtime via init_srcu_struct() 409 and cleanup_srcu_struct(). These last two are passed a 410 "struct srcu_struct" that defines the scope of a given 411 SRCU domain. Once initialized, the srcu_struct is passed 412 to srcu_read_lock(), srcu_read_unlock() synchronize_srcu(), 413 synchronize_srcu_expedited(), and call_srcu(). A given 414 synchronize_srcu() waits only for SRCU read-side critical 415 sections governed by srcu_read_lock() and srcu_read_unlock() 416 calls that have been passed the same srcu_struct. This property 417 is what makes sleeping read-side critical sections tolerable -- 418 a given subsystem delays only its own updates, not those of other 419 subsystems using SRCU. Therefore, SRCU is less prone to OOM the 420 system than RCU would be if RCU's read-side critical sections 421 were permitted to sleep. 422 423 The ability to sleep in read-side critical sections does not 424 come for free. First, corresponding srcu_read_lock() and 425 srcu_read_unlock() calls must be passed the same srcu_struct. 426 Second, grace-period-detection overhead is amortized only 427 over those updates sharing a given srcu_struct, rather than 428 being globally amortized as they are for other forms of RCU. 429 Therefore, SRCU should be used in preference to rw_semaphore 430 only in extremely read-intensive situations, or in situations 431 requiring SRCU's read-side deadlock immunity or low read-side 432 realtime latency. You should also consider percpu_rw_semaphore 433 when you need lightweight readers. 434 435 SRCU's expedited primitive (synchronize_srcu_expedited()) 436 never sends IPIs to other CPUs, so it is easier on 437 real-time workloads than is synchronize_rcu_expedited(). 438 439 It is also permissible to sleep in RCU Tasks Trace read-side 440 critical, which are delimited by rcu_read_lock_trace() and 441 rcu_read_unlock_trace(). However, this is a specialized flavor 442 of RCU, and you should not use it without first checking with 443 its current users. In most cases, you should instead use SRCU. 444 445 Note that rcu_assign_pointer() relates to SRCU just as it does to 446 other forms of RCU, but instead of rcu_dereference() you should 447 use srcu_dereference() in order to avoid lockdep splats. 448 44914. The whole point of call_rcu(), synchronize_rcu(), and friends 450 is to wait until all pre-existing readers have finished before 451 carrying out some otherwise-destructive operation. It is 452 therefore critically important to *first* remove any path 453 that readers can follow that could be affected by the 454 destructive operation, and *only then* invoke call_rcu(), 455 synchronize_rcu(), or friends. 456 457 Because these primitives only wait for pre-existing readers, it 458 is the caller's responsibility to guarantee that any subsequent 459 readers will execute safely. 460 46115. The various RCU read-side primitives do *not* necessarily contain 462 memory barriers. You should therefore plan for the CPU 463 and the compiler to freely reorder code into and out of RCU 464 read-side critical sections. It is the responsibility of the 465 RCU update-side primitives to deal with this. 466 467 For SRCU readers, you can use smp_mb__after_srcu_read_unlock() 468 immediately after an srcu_read_unlock() to get a full barrier. 469 47016. Use CONFIG_PROVE_LOCKING, CONFIG_DEBUG_OBJECTS_RCU_HEAD, and the 471 __rcu sparse checks to validate your RCU code. These can help 472 find problems as follows: 473 474 CONFIG_PROVE_LOCKING: 475 check that accesses to RCU-protected data structures 476 are carried out under the proper RCU read-side critical 477 section, while holding the right combination of locks, 478 or whatever other conditions are appropriate. 479 480 CONFIG_DEBUG_OBJECTS_RCU_HEAD: 481 check that you don't pass the same object to call_rcu() 482 (or friends) before an RCU grace period has elapsed 483 since the last time that you passed that same object to 484 call_rcu() (or friends). 485 486 __rcu sparse checks: 487 tag the pointer to the RCU-protected data structure 488 with __rcu, and sparse will warn you if you access that 489 pointer without the services of one of the variants 490 of rcu_dereference(). 491 492 These debugging aids can help you find problems that are 493 otherwise extremely difficult to spot. 494 49517. If you pass a callback function defined within a module to one of 496 call_rcu(), call_srcu(), call_rcu_tasks(), call_rcu_tasks_rude(), 497 or call_rcu_tasks_trace(), then it is necessary to wait for all 498 pending callbacks to be invoked before unloading that module. 499 Note that it is absolutely *not* sufficient to wait for a grace 500 period! For example, synchronize_rcu() implementation is *not* 501 guaranteed to wait for callbacks registered on other CPUs via 502 call_rcu(). Or even on the current CPU if that CPU recently 503 went offline and came back online. 504 505 You instead need to use one of the barrier functions: 506 507 - call_rcu() -> rcu_barrier() 508 - call_srcu() -> srcu_barrier() 509 - call_rcu_tasks() -> rcu_barrier_tasks() 510 - call_rcu_tasks_rude() -> rcu_barrier_tasks_rude() 511 - call_rcu_tasks_trace() -> rcu_barrier_tasks_trace() 512 513 However, these barrier functions are absolutely *not* guaranteed 514 to wait for a grace period. For example, if there are no 515 call_rcu() callbacks queued anywhere in the system, rcu_barrier() 516 can and will return immediately. 517 518 So if you need to wait for both a grace period and for all 519 pre-existing callbacks, you will need to invoke both functions, 520 with the pair depending on the flavor of RCU: 521 522 - Either synchronize_rcu() or synchronize_rcu_expedited(), 523 together with rcu_barrier() 524 - Either synchronize_srcu() or synchronize_srcu_expedited(), 525 together with and srcu_barrier() 526 - synchronize_rcu_tasks() and rcu_barrier_tasks() 527 - synchronize_tasks_rude() and rcu_barrier_tasks_rude() 528 - synchronize_tasks_trace() and rcu_barrier_tasks_trace() 529 530 If necessary, you can use something like workqueues to execute 531 the requisite pair of functions concurrently. 532 533 See rcubarrier.rst for more information. 534