1Please note that the "What is RCU?" LWN series is an excellent place 2to start learning about RCU: 3 41. What is RCU, Fundamentally? http://lwn.net/Articles/262464/ 52. What is RCU? Part 2: Usage http://lwn.net/Articles/263130/ 63. RCU part 3: the RCU API http://lwn.net/Articles/264090/ 74. The RCU API, 2010 Edition http://lwn.net/Articles/418853/ 8 2010 Big API Table http://lwn.net/Articles/419086/ 95. The RCU API, 2014 Edition http://lwn.net/Articles/609904/ 10 2014 Big API Table http://lwn.net/Articles/609973/ 11 12 13What is RCU? 14 15RCU is a synchronization mechanism that was added to the Linux kernel 16during the 2.5 development effort that is optimized for read-mostly 17situations. Although RCU is actually quite simple once you understand it, 18getting there can sometimes be a challenge. Part of the problem is that 19most of the past descriptions of RCU have been written with the mistaken 20assumption that there is "one true way" to describe RCU. Instead, 21the experience has been that different people must take different paths 22to arrive at an understanding of RCU. This document provides several 23different paths, as follows: 24 251. RCU OVERVIEW 262. WHAT IS RCU'S CORE API? 273. WHAT ARE SOME EXAMPLE USES OF CORE RCU API? 284. WHAT IF MY UPDATING THREAD CANNOT BLOCK? 295. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU? 306. ANALOGY WITH READER-WRITER LOCKING 317. FULL LIST OF RCU APIs 328. ANSWERS TO QUICK QUIZZES 33 34People who prefer starting with a conceptual overview should focus on 35Section 1, though most readers will profit by reading this section at 36some point. People who prefer to start with an API that they can then 37experiment with should focus on Section 2. People who prefer to start 38with example uses should focus on Sections 3 and 4. People who need to 39understand the RCU implementation should focus on Section 5, then dive 40into the kernel source code. People who reason best by analogy should 41focus on Section 6. Section 7 serves as an index to the docbook API 42documentation, and Section 8 is the traditional answer key. 43 44So, start with the section that makes the most sense to you and your 45preferred method of learning. If you need to know everything about 46everything, feel free to read the whole thing -- but if you are really 47that type of person, you have perused the source code and will therefore 48never need this document anyway. ;-) 49 50 511. RCU OVERVIEW 52 53The basic idea behind RCU is to split updates into "removal" and 54"reclamation" phases. The removal phase removes references to data items 55within a data structure (possibly by replacing them with references to 56new versions of these data items), and can run concurrently with readers. 57The reason that it is safe to run the removal phase concurrently with 58readers is the semantics of modern CPUs guarantee that readers will see 59either the old or the new version of the data structure rather than a 60partially updated reference. The reclamation phase does the work of reclaiming 61(e.g., freeing) the data items removed from the data structure during the 62removal phase. Because reclaiming data items can disrupt any readers 63concurrently referencing those data items, the reclamation phase must 64not start until readers no longer hold references to those data items. 65 66Splitting the update into removal and reclamation phases permits the 67updater to perform the removal phase immediately, and to defer the 68reclamation phase until all readers active during the removal phase have 69completed, either by blocking until they finish or by registering a 70callback that is invoked after they finish. Only readers that are active 71during the removal phase need be considered, because any reader starting 72after the removal phase will be unable to gain a reference to the removed 73data items, and therefore cannot be disrupted by the reclamation phase. 74 75So the typical RCU update sequence goes something like the following: 76 77a. Remove pointers to a data structure, so that subsequent 78 readers cannot gain a reference to it. 79 80b. Wait for all previous readers to complete their RCU read-side 81 critical sections. 82 83c. At this point, there cannot be any readers who hold references 84 to the data structure, so it now may safely be reclaimed 85 (e.g., kfree()d). 86 87Step (b) above is the key idea underlying RCU's deferred destruction. 88The ability to wait until all readers are done allows RCU readers to 89use much lighter-weight synchronization, in some cases, absolutely no 90synchronization at all. In contrast, in more conventional lock-based 91schemes, readers must use heavy-weight synchronization in order to 92prevent an updater from deleting the data structure out from under them. 93This is because lock-based updaters typically update data items in place, 94and must therefore exclude readers. In contrast, RCU-based updaters 95typically take advantage of the fact that writes to single aligned 96pointers are atomic on modern CPUs, allowing atomic insertion, removal, 97and replacement of data items in a linked structure without disrupting 98readers. Concurrent RCU readers can then continue accessing the old 99versions, and can dispense with the atomic operations, memory barriers, 100and communications cache misses that are so expensive on present-day 101SMP computer systems, even in absence of lock contention. 102 103In the three-step procedure shown above, the updater is performing both 104the removal and the reclamation step, but it is often helpful for an 105entirely different thread to do the reclamation, as is in fact the case 106in the Linux kernel's directory-entry cache (dcache). Even if the same 107thread performs both the update step (step (a) above) and the reclamation 108step (step (c) above), it is often helpful to think of them separately. 109For example, RCU readers and updaters need not communicate at all, 110but RCU provides implicit low-overhead communication between readers 111and reclaimers, namely, in step (b) above. 112 113So how the heck can a reclaimer tell when a reader is done, given 114that readers are not doing any sort of synchronization operations??? 115Read on to learn about how RCU's API makes this easy. 116 117 1182. WHAT IS RCU'S CORE API? 119 120The core RCU API is quite small: 121 122a. rcu_read_lock() 123b. rcu_read_unlock() 124c. synchronize_rcu() / call_rcu() 125d. rcu_assign_pointer() 126e. rcu_dereference() 127 128There are many other members of the RCU API, but the rest can be 129expressed in terms of these five, though most implementations instead 130express synchronize_rcu() in terms of the call_rcu() callback API. 131 132The five core RCU APIs are described below, the other 18 will be enumerated 133later. See the kernel docbook documentation for more info, or look directly 134at the function header comments. 135 136rcu_read_lock() 137 138 void rcu_read_lock(void); 139 140 Used by a reader to inform the reclaimer that the reader is 141 entering an RCU read-side critical section. It is illegal 142 to block while in an RCU read-side critical section, though 143 kernels built with CONFIG_PREEMPT_RCU can preempt RCU 144 read-side critical sections. Any RCU-protected data structure 145 accessed during an RCU read-side critical section is guaranteed to 146 remain unreclaimed for the full duration of that critical section. 147 Reference counts may be used in conjunction with RCU to maintain 148 longer-term references to data structures. 149 150rcu_read_unlock() 151 152 void rcu_read_unlock(void); 153 154 Used by a reader to inform the reclaimer that the reader is 155 exiting an RCU read-side critical section. Note that RCU 156 read-side critical sections may be nested and/or overlapping. 157 158synchronize_rcu() 159 160 void synchronize_rcu(void); 161 162 Marks the end of updater code and the beginning of reclaimer 163 code. It does this by blocking until all pre-existing RCU 164 read-side critical sections on all CPUs have completed. 165 Note that synchronize_rcu() will -not- necessarily wait for 166 any subsequent RCU read-side critical sections to complete. 167 For example, consider the following sequence of events: 168 169 CPU 0 CPU 1 CPU 2 170 ----------------- ------------------------- --------------- 171 1. rcu_read_lock() 172 2. enters synchronize_rcu() 173 3. rcu_read_lock() 174 4. rcu_read_unlock() 175 5. exits synchronize_rcu() 176 6. rcu_read_unlock() 177 178 To reiterate, synchronize_rcu() waits only for ongoing RCU 179 read-side critical sections to complete, not necessarily for 180 any that begin after synchronize_rcu() is invoked. 181 182 Of course, synchronize_rcu() does not necessarily return 183 -immediately- after the last pre-existing RCU read-side critical 184 section completes. For one thing, there might well be scheduling 185 delays. For another thing, many RCU implementations process 186 requests in batches in order to improve efficiencies, which can 187 further delay synchronize_rcu(). 188 189 Since synchronize_rcu() is the API that must figure out when 190 readers are done, its implementation is key to RCU. For RCU 191 to be useful in all but the most read-intensive situations, 192 synchronize_rcu()'s overhead must also be quite small. 193 194 The call_rcu() API is a callback form of synchronize_rcu(), 195 and is described in more detail in a later section. Instead of 196 blocking, it registers a function and argument which are invoked 197 after all ongoing RCU read-side critical sections have completed. 198 This callback variant is particularly useful in situations where 199 it is illegal to block or where update-side performance is 200 critically important. 201 202 However, the call_rcu() API should not be used lightly, as use 203 of the synchronize_rcu() API generally results in simpler code. 204 In addition, the synchronize_rcu() API has the nice property 205 of automatically limiting update rate should grace periods 206 be delayed. This property results in system resilience in face 207 of denial-of-service attacks. Code using call_rcu() should limit 208 update rate in order to gain this same sort of resilience. See 209 checklist.txt for some approaches to limiting the update rate. 210 211rcu_assign_pointer() 212 213 typeof(p) rcu_assign_pointer(p, typeof(p) v); 214 215 Yes, rcu_assign_pointer() -is- implemented as a macro, though it 216 would be cool to be able to declare a function in this manner. 217 (Compiler experts will no doubt disagree.) 218 219 The updater uses this function to assign a new value to an 220 RCU-protected pointer, in order to safely communicate the change 221 in value from the updater to the reader. This function returns 222 the new value, and also executes any memory-barrier instructions 223 required for a given CPU architecture. 224 225 Perhaps just as important, it serves to document (1) which 226 pointers are protected by RCU and (2) the point at which a 227 given structure becomes accessible to other CPUs. That said, 228 rcu_assign_pointer() is most frequently used indirectly, via 229 the _rcu list-manipulation primitives such as list_add_rcu(). 230 231rcu_dereference() 232 233 typeof(p) rcu_dereference(p); 234 235 Like rcu_assign_pointer(), rcu_dereference() must be implemented 236 as a macro. 237 238 The reader uses rcu_dereference() to fetch an RCU-protected 239 pointer, which returns a value that may then be safely 240 dereferenced. Note that rcu_deference() does not actually 241 dereference the pointer, instead, it protects the pointer for 242 later dereferencing. It also executes any needed memory-barrier 243 instructions for a given CPU architecture. Currently, only Alpha 244 needs memory barriers within rcu_dereference() -- on other CPUs, 245 it compiles to nothing, not even a compiler directive. 246 247 Common coding practice uses rcu_dereference() to copy an 248 RCU-protected pointer to a local variable, then dereferences 249 this local variable, for example as follows: 250 251 p = rcu_dereference(head.next); 252 return p->data; 253 254 However, in this case, one could just as easily combine these 255 into one statement: 256 257 return rcu_dereference(head.next)->data; 258 259 If you are going to be fetching multiple fields from the 260 RCU-protected structure, using the local variable is of 261 course preferred. Repeated rcu_dereference() calls look 262 ugly, do not guarantee that the same pointer will be returned 263 if an update happened while in the critical section, and incur 264 unnecessary overhead on Alpha CPUs. 265 266 Note that the value returned by rcu_dereference() is valid 267 only within the enclosing RCU read-side critical section. 268 For example, the following is -not- legal: 269 270 rcu_read_lock(); 271 p = rcu_dereference(head.next); 272 rcu_read_unlock(); 273 x = p->address; /* BUG!!! */ 274 rcu_read_lock(); 275 y = p->data; /* BUG!!! */ 276 rcu_read_unlock(); 277 278 Holding a reference from one RCU read-side critical section 279 to another is just as illegal as holding a reference from 280 one lock-based critical section to another! Similarly, 281 using a reference outside of the critical section in which 282 it was acquired is just as illegal as doing so with normal 283 locking. 284 285 As with rcu_assign_pointer(), an important function of 286 rcu_dereference() is to document which pointers are protected by 287 RCU, in particular, flagging a pointer that is subject to changing 288 at any time, including immediately after the rcu_dereference(). 289 And, again like rcu_assign_pointer(), rcu_dereference() is 290 typically used indirectly, via the _rcu list-manipulation 291 primitives, such as list_for_each_entry_rcu(). 292 293The following diagram shows how each API communicates among the 294reader, updater, and reclaimer. 295 296 297 rcu_assign_pointer() 298 +--------+ 299 +---------------------->| reader |---------+ 300 | +--------+ | 301 | | | 302 | | | Protect: 303 | | | rcu_read_lock() 304 | | | rcu_read_unlock() 305 | rcu_dereference() | | 306 +---------+ | | 307 | updater |<---------------------+ | 308 +---------+ V 309 | +-----------+ 310 +----------------------------------->| reclaimer | 311 +-----------+ 312 Defer: 313 synchronize_rcu() & call_rcu() 314 315 316The RCU infrastructure observes the time sequence of rcu_read_lock(), 317rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in 318order to determine when (1) synchronize_rcu() invocations may return 319to their callers and (2) call_rcu() callbacks may be invoked. Efficient 320implementations of the RCU infrastructure make heavy use of batching in 321order to amortize their overhead over many uses of the corresponding APIs. 322 323There are no fewer than three RCU mechanisms in the Linux kernel; the 324diagram above shows the first one, which is by far the most commonly used. 325The rcu_dereference() and rcu_assign_pointer() primitives are used for 326all three mechanisms, but different defer and protect primitives are 327used as follows: 328 329 Defer Protect 330 331a. synchronize_rcu() rcu_read_lock() / rcu_read_unlock() 332 call_rcu() rcu_dereference() 333 334b. synchronize_rcu_bh() rcu_read_lock_bh() / rcu_read_unlock_bh() 335 call_rcu_bh() rcu_dereference_bh() 336 337c. synchronize_sched() rcu_read_lock_sched() / rcu_read_unlock_sched() 338 call_rcu_sched() preempt_disable() / preempt_enable() 339 local_irq_save() / local_irq_restore() 340 hardirq enter / hardirq exit 341 NMI enter / NMI exit 342 rcu_dereference_sched() 343 344These three mechanisms are used as follows: 345 346a. RCU applied to normal data structures. 347 348b. RCU applied to networking data structures that may be subjected 349 to remote denial-of-service attacks. 350 351c. RCU applied to scheduler and interrupt/NMI-handler tasks. 352 353Again, most uses will be of (a). The (b) and (c) cases are important 354for specialized uses, but are relatively uncommon. 355 356 3573. WHAT ARE SOME EXAMPLE USES OF CORE RCU API? 358 359This section shows a simple use of the core RCU API to protect a 360global pointer to a dynamically allocated structure. More-typical 361uses of RCU may be found in listRCU.txt, arrayRCU.txt, and NMI-RCU.txt. 362 363 struct foo { 364 int a; 365 char b; 366 long c; 367 }; 368 DEFINE_SPINLOCK(foo_mutex); 369 370 struct foo __rcu *gbl_foo; 371 372 /* 373 * Create a new struct foo that is the same as the one currently 374 * pointed to by gbl_foo, except that field "a" is replaced 375 * with "new_a". Points gbl_foo to the new structure, and 376 * frees up the old structure after a grace period. 377 * 378 * Uses rcu_assign_pointer() to ensure that concurrent readers 379 * see the initialized version of the new structure. 380 * 381 * Uses synchronize_rcu() to ensure that any readers that might 382 * have references to the old structure complete before freeing 383 * the old structure. 384 */ 385 void foo_update_a(int new_a) 386 { 387 struct foo *new_fp; 388 struct foo *old_fp; 389 390 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL); 391 spin_lock(&foo_mutex); 392 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex)); 393 *new_fp = *old_fp; 394 new_fp->a = new_a; 395 rcu_assign_pointer(gbl_foo, new_fp); 396 spin_unlock(&foo_mutex); 397 synchronize_rcu(); 398 kfree(old_fp); 399 } 400 401 /* 402 * Return the value of field "a" of the current gbl_foo 403 * structure. Use rcu_read_lock() and rcu_read_unlock() 404 * to ensure that the structure does not get deleted out 405 * from under us, and use rcu_dereference() to ensure that 406 * we see the initialized version of the structure (important 407 * for DEC Alpha and for people reading the code). 408 */ 409 int foo_get_a(void) 410 { 411 int retval; 412 413 rcu_read_lock(); 414 retval = rcu_dereference(gbl_foo)->a; 415 rcu_read_unlock(); 416 return retval; 417 } 418 419So, to sum up: 420 421o Use rcu_read_lock() and rcu_read_unlock() to guard RCU 422 read-side critical sections. 423 424o Within an RCU read-side critical section, use rcu_dereference() 425 to dereference RCU-protected pointers. 426 427o Use some solid scheme (such as locks or semaphores) to 428 keep concurrent updates from interfering with each other. 429 430o Use rcu_assign_pointer() to update an RCU-protected pointer. 431 This primitive protects concurrent readers from the updater, 432 -not- concurrent updates from each other! You therefore still 433 need to use locking (or something similar) to keep concurrent 434 rcu_assign_pointer() primitives from interfering with each other. 435 436o Use synchronize_rcu() -after- removing a data element from an 437 RCU-protected data structure, but -before- reclaiming/freeing 438 the data element, in order to wait for the completion of all 439 RCU read-side critical sections that might be referencing that 440 data item. 441 442See checklist.txt for additional rules to follow when using RCU. 443And again, more-typical uses of RCU may be found in listRCU.txt, 444arrayRCU.txt, and NMI-RCU.txt. 445 446 4474. WHAT IF MY UPDATING THREAD CANNOT BLOCK? 448 449In the example above, foo_update_a() blocks until a grace period elapses. 450This is quite simple, but in some cases one cannot afford to wait so 451long -- there might be other high-priority work to be done. 452 453In such cases, one uses call_rcu() rather than synchronize_rcu(). 454The call_rcu() API is as follows: 455 456 void call_rcu(struct rcu_head * head, 457 void (*func)(struct rcu_head *head)); 458 459This function invokes func(head) after a grace period has elapsed. 460This invocation might happen from either softirq or process context, 461so the function is not permitted to block. The foo struct needs to 462have an rcu_head structure added, perhaps as follows: 463 464 struct foo { 465 int a; 466 char b; 467 long c; 468 struct rcu_head rcu; 469 }; 470 471The foo_update_a() function might then be written as follows: 472 473 /* 474 * Create a new struct foo that is the same as the one currently 475 * pointed to by gbl_foo, except that field "a" is replaced 476 * with "new_a". Points gbl_foo to the new structure, and 477 * frees up the old structure after a grace period. 478 * 479 * Uses rcu_assign_pointer() to ensure that concurrent readers 480 * see the initialized version of the new structure. 481 * 482 * Uses call_rcu() to ensure that any readers that might have 483 * references to the old structure complete before freeing the 484 * old structure. 485 */ 486 void foo_update_a(int new_a) 487 { 488 struct foo *new_fp; 489 struct foo *old_fp; 490 491 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL); 492 spin_lock(&foo_mutex); 493 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex)); 494 *new_fp = *old_fp; 495 new_fp->a = new_a; 496 rcu_assign_pointer(gbl_foo, new_fp); 497 spin_unlock(&foo_mutex); 498 call_rcu(&old_fp->rcu, foo_reclaim); 499 } 500 501The foo_reclaim() function might appear as follows: 502 503 void foo_reclaim(struct rcu_head *rp) 504 { 505 struct foo *fp = container_of(rp, struct foo, rcu); 506 507 foo_cleanup(fp->a); 508 509 kfree(fp); 510 } 511 512The container_of() primitive is a macro that, given a pointer into a 513struct, the type of the struct, and the pointed-to field within the 514struct, returns a pointer to the beginning of the struct. 515 516The use of call_rcu() permits the caller of foo_update_a() to 517immediately regain control, without needing to worry further about the 518old version of the newly updated element. It also clearly shows the 519RCU distinction between updater, namely foo_update_a(), and reclaimer, 520namely foo_reclaim(). 521 522The summary of advice is the same as for the previous section, except 523that we are now using call_rcu() rather than synchronize_rcu(): 524 525o Use call_rcu() -after- removing a data element from an 526 RCU-protected data structure in order to register a callback 527 function that will be invoked after the completion of all RCU 528 read-side critical sections that might be referencing that 529 data item. 530 531If the callback for call_rcu() is not doing anything more than calling 532kfree() on the structure, you can use kfree_rcu() instead of call_rcu() 533to avoid having to write your own callback: 534 535 kfree_rcu(old_fp, rcu); 536 537Again, see checklist.txt for additional rules governing the use of RCU. 538 539 5405. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU? 541 542One of the nice things about RCU is that it has extremely simple "toy" 543implementations that are a good first step towards understanding the 544production-quality implementations in the Linux kernel. This section 545presents two such "toy" implementations of RCU, one that is implemented 546in terms of familiar locking primitives, and another that more closely 547resembles "classic" RCU. Both are way too simple for real-world use, 548lacking both functionality and performance. However, they are useful 549in getting a feel for how RCU works. See kernel/rcupdate.c for a 550production-quality implementation, and see: 551 552 http://www.rdrop.com/users/paulmck/RCU 553 554for papers describing the Linux kernel RCU implementation. The OLS'01 555and OLS'02 papers are a good introduction, and the dissertation provides 556more details on the current implementation as of early 2004. 557 558 5595A. "TOY" IMPLEMENTATION #1: LOCKING 560 561This section presents a "toy" RCU implementation that is based on 562familiar locking primitives. Its overhead makes it a non-starter for 563real-life use, as does its lack of scalability. It is also unsuitable 564for realtime use, since it allows scheduling latency to "bleed" from 565one read-side critical section to another. 566 567However, it is probably the easiest implementation to relate to, so is 568a good starting point. 569 570It is extremely simple: 571 572 static DEFINE_RWLOCK(rcu_gp_mutex); 573 574 void rcu_read_lock(void) 575 { 576 read_lock(&rcu_gp_mutex); 577 } 578 579 void rcu_read_unlock(void) 580 { 581 read_unlock(&rcu_gp_mutex); 582 } 583 584 void synchronize_rcu(void) 585 { 586 write_lock(&rcu_gp_mutex); 587 write_unlock(&rcu_gp_mutex); 588 } 589 590[You can ignore rcu_assign_pointer() and rcu_dereference() without 591missing much. But here they are anyway. And whatever you do, don't 592forget about them when submitting patches making use of RCU!] 593 594 #define rcu_assign_pointer(p, v) ({ \ 595 smp_wmb(); \ 596 (p) = (v); \ 597 }) 598 599 #define rcu_dereference(p) ({ \ 600 typeof(p) _________p1 = p; \ 601 smp_read_barrier_depends(); \ 602 (_________p1); \ 603 }) 604 605 606The rcu_read_lock() and rcu_read_unlock() primitive read-acquire 607and release a global reader-writer lock. The synchronize_rcu() 608primitive write-acquires this same lock, then immediately releases 609it. This means that once synchronize_rcu() exits, all RCU read-side 610critical sections that were in progress before synchronize_rcu() was 611called are guaranteed to have completed -- there is no way that 612synchronize_rcu() would have been able to write-acquire the lock 613otherwise. 614 615It is possible to nest rcu_read_lock(), since reader-writer locks may 616be recursively acquired. Note also that rcu_read_lock() is immune 617from deadlock (an important property of RCU). The reason for this is 618that the only thing that can block rcu_read_lock() is a synchronize_rcu(). 619But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex, 620so there can be no deadlock cycle. 621 622Quick Quiz #1: Why is this argument naive? How could a deadlock 623 occur when using this algorithm in a real-world Linux 624 kernel? How could this deadlock be avoided? 625 626 6275B. "TOY" EXAMPLE #2: CLASSIC RCU 628 629This section presents a "toy" RCU implementation that is based on 630"classic RCU". It is also short on performance (but only for updates) and 631on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT 632kernels. The definitions of rcu_dereference() and rcu_assign_pointer() 633are the same as those shown in the preceding section, so they are omitted. 634 635 void rcu_read_lock(void) { } 636 637 void rcu_read_unlock(void) { } 638 639 void synchronize_rcu(void) 640 { 641 int cpu; 642 643 for_each_possible_cpu(cpu) 644 run_on(cpu); 645 } 646 647Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing. 648This is the great strength of classic RCU in a non-preemptive kernel: 649read-side overhead is precisely zero, at least on non-Alpha CPUs. 650And there is absolutely no way that rcu_read_lock() can possibly 651participate in a deadlock cycle! 652 653The implementation of synchronize_rcu() simply schedules itself on each 654CPU in turn. The run_on() primitive can be implemented straightforwardly 655in terms of the sched_setaffinity() primitive. Of course, a somewhat less 656"toy" implementation would restore the affinity upon completion rather 657than just leaving all tasks running on the last CPU, but when I said 658"toy", I meant -toy-! 659 660So how the heck is this supposed to work??? 661 662Remember that it is illegal to block while in an RCU read-side critical 663section. Therefore, if a given CPU executes a context switch, we know 664that it must have completed all preceding RCU read-side critical sections. 665Once -all- CPUs have executed a context switch, then -all- preceding 666RCU read-side critical sections will have completed. 667 668So, suppose that we remove a data item from its structure and then invoke 669synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed 670that there are no RCU read-side critical sections holding a reference 671to that data item, so we can safely reclaim it. 672 673Quick Quiz #2: Give an example where Classic RCU's read-side 674 overhead is -negative-. 675 676Quick Quiz #3: If it is illegal to block in an RCU read-side 677 critical section, what the heck do you do in 678 PREEMPT_RT, where normal spinlocks can block??? 679 680 6816. ANALOGY WITH READER-WRITER LOCKING 682 683Although RCU can be used in many different ways, a very common use of 684RCU is analogous to reader-writer locking. The following unified 685diff shows how closely related RCU and reader-writer locking can be. 686 687 @@ -5,5 +5,5 @@ struct el { 688 int data; 689 /* Other data fields */ 690 }; 691 -rwlock_t listmutex; 692 +spinlock_t listmutex; 693 struct el head; 694 695 @@ -13,15 +14,15 @@ 696 struct list_head *lp; 697 struct el *p; 698 699 - read_lock(&listmutex); 700 - list_for_each_entry(p, head, lp) { 701 + rcu_read_lock(); 702 + list_for_each_entry_rcu(p, head, lp) { 703 if (p->key == key) { 704 *result = p->data; 705 - read_unlock(&listmutex); 706 + rcu_read_unlock(); 707 return 1; 708 } 709 } 710 - read_unlock(&listmutex); 711 + rcu_read_unlock(); 712 return 0; 713 } 714 715 @@ -29,15 +30,16 @@ 716 { 717 struct el *p; 718 719 - write_lock(&listmutex); 720 + spin_lock(&listmutex); 721 list_for_each_entry(p, head, lp) { 722 if (p->key == key) { 723 - list_del(&p->list); 724 - write_unlock(&listmutex); 725 + list_del_rcu(&p->list); 726 + spin_unlock(&listmutex); 727 + synchronize_rcu(); 728 kfree(p); 729 return 1; 730 } 731 } 732 - write_unlock(&listmutex); 733 + spin_unlock(&listmutex); 734 return 0; 735 } 736 737Or, for those who prefer a side-by-side listing: 738 739 1 struct el { 1 struct el { 740 2 struct list_head list; 2 struct list_head list; 741 3 long key; 3 long key; 742 4 spinlock_t mutex; 4 spinlock_t mutex; 743 5 int data; 5 int data; 744 6 /* Other data fields */ 6 /* Other data fields */ 745 7 }; 7 }; 746 8 rwlock_t listmutex; 8 spinlock_t listmutex; 747 9 struct el head; 9 struct el head; 748 749 1 int search(long key, int *result) 1 int search(long key, int *result) 750 2 { 2 { 751 3 struct list_head *lp; 3 struct list_head *lp; 752 4 struct el *p; 4 struct el *p; 753 5 5 754 6 read_lock(&listmutex); 6 rcu_read_lock(); 755 7 list_for_each_entry(p, head, lp) { 7 list_for_each_entry_rcu(p, head, lp) { 756 8 if (p->key == key) { 8 if (p->key == key) { 757 9 *result = p->data; 9 *result = p->data; 75810 read_unlock(&listmutex); 10 rcu_read_unlock(); 75911 return 1; 11 return 1; 76012 } 12 } 76113 } 13 } 76214 read_unlock(&listmutex); 14 rcu_read_unlock(); 76315 return 0; 15 return 0; 76416 } 16 } 765 766 1 int delete(long key) 1 int delete(long key) 767 2 { 2 { 768 3 struct el *p; 3 struct el *p; 769 4 4 770 5 write_lock(&listmutex); 5 spin_lock(&listmutex); 771 6 list_for_each_entry(p, head, lp) { 6 list_for_each_entry(p, head, lp) { 772 7 if (p->key == key) { 7 if (p->key == key) { 773 8 list_del(&p->list); 8 list_del_rcu(&p->list); 774 9 write_unlock(&listmutex); 9 spin_unlock(&listmutex); 775 10 synchronize_rcu(); 77610 kfree(p); 11 kfree(p); 77711 return 1; 12 return 1; 77812 } 13 } 77913 } 14 } 78014 write_unlock(&listmutex); 15 spin_unlock(&listmutex); 78115 return 0; 16 return 0; 78216 } 17 } 783 784Either way, the differences are quite small. Read-side locking moves 785to rcu_read_lock() and rcu_read_unlock, update-side locking moves from 786a reader-writer lock to a simple spinlock, and a synchronize_rcu() 787precedes the kfree(). 788 789However, there is one potential catch: the read-side and update-side 790critical sections can now run concurrently. In many cases, this will 791not be a problem, but it is necessary to check carefully regardless. 792For example, if multiple independent list updates must be seen as 793a single atomic update, converting to RCU will require special care. 794 795Also, the presence of synchronize_rcu() means that the RCU version of 796delete() can now block. If this is a problem, there is a callback-based 797mechanism that never blocks, namely call_rcu() or kfree_rcu(), that can 798be used in place of synchronize_rcu(). 799 800 8017. FULL LIST OF RCU APIs 802 803The RCU APIs are documented in docbook-format header comments in the 804Linux-kernel source code, but it helps to have a full list of the 805APIs, since there does not appear to be a way to categorize them 806in docbook. Here is the list, by category. 807 808RCU list traversal: 809 810 list_entry_rcu 811 list_first_entry_rcu 812 list_next_rcu 813 list_for_each_entry_rcu 814 list_for_each_entry_continue_rcu 815 hlist_first_rcu 816 hlist_next_rcu 817 hlist_pprev_rcu 818 hlist_for_each_entry_rcu 819 hlist_for_each_entry_rcu_bh 820 hlist_for_each_entry_continue_rcu 821 hlist_for_each_entry_continue_rcu_bh 822 hlist_nulls_first_rcu 823 hlist_nulls_for_each_entry_rcu 824 hlist_bl_first_rcu 825 hlist_bl_for_each_entry_rcu 826 827RCU pointer/list update: 828 829 rcu_assign_pointer 830 list_add_rcu 831 list_add_tail_rcu 832 list_del_rcu 833 list_replace_rcu 834 hlist_add_behind_rcu 835 hlist_add_before_rcu 836 hlist_add_head_rcu 837 hlist_del_rcu 838 hlist_del_init_rcu 839 hlist_replace_rcu 840 list_splice_init_rcu() 841 hlist_nulls_del_init_rcu 842 hlist_nulls_del_rcu 843 hlist_nulls_add_head_rcu 844 hlist_bl_add_head_rcu 845 hlist_bl_del_init_rcu 846 hlist_bl_del_rcu 847 hlist_bl_set_first_rcu 848 849RCU: Critical sections Grace period Barrier 850 851 rcu_read_lock synchronize_net rcu_barrier 852 rcu_read_unlock synchronize_rcu 853 rcu_dereference synchronize_rcu_expedited 854 rcu_read_lock_held call_rcu 855 rcu_dereference_check kfree_rcu 856 rcu_dereference_protected 857 858bh: Critical sections Grace period Barrier 859 860 rcu_read_lock_bh call_rcu_bh rcu_barrier_bh 861 rcu_read_unlock_bh synchronize_rcu_bh 862 rcu_dereference_bh synchronize_rcu_bh_expedited 863 rcu_dereference_bh_check 864 rcu_dereference_bh_protected 865 rcu_read_lock_bh_held 866 867sched: Critical sections Grace period Barrier 868 869 rcu_read_lock_sched synchronize_sched rcu_barrier_sched 870 rcu_read_unlock_sched call_rcu_sched 871 [preempt_disable] synchronize_sched_expedited 872 [and friends] 873 rcu_read_lock_sched_notrace 874 rcu_read_unlock_sched_notrace 875 rcu_dereference_sched 876 rcu_dereference_sched_check 877 rcu_dereference_sched_protected 878 rcu_read_lock_sched_held 879 880 881SRCU: Critical sections Grace period Barrier 882 883 srcu_read_lock synchronize_srcu srcu_barrier 884 srcu_read_unlock call_srcu 885 srcu_dereference synchronize_srcu_expedited 886 srcu_dereference_check 887 srcu_read_lock_held 888 889SRCU: Initialization/cleanup 890 init_srcu_struct 891 cleanup_srcu_struct 892 893All: lockdep-checked RCU-protected pointer access 894 895 rcu_access_pointer 896 rcu_dereference_raw 897 RCU_LOCKDEP_WARN 898 rcu_sleep_check 899 RCU_NONIDLE 900 901See the comment headers in the source code (or the docbook generated 902from them) for more information. 903 904However, given that there are no fewer than four families of RCU APIs 905in the Linux kernel, how do you choose which one to use? The following 906list can be helpful: 907 908a. Will readers need to block? If so, you need SRCU. 909 910b. What about the -rt patchset? If readers would need to block 911 in an non-rt kernel, you need SRCU. If readers would block 912 in a -rt kernel, but not in a non-rt kernel, SRCU is not 913 necessary. 914 915c. Do you need to treat NMI handlers, hardirq handlers, 916 and code segments with preemption disabled (whether 917 via preempt_disable(), local_irq_save(), local_bh_disable(), 918 or some other mechanism) as if they were explicit RCU readers? 919 If so, RCU-sched is the only choice that will work for you. 920 921d. Do you need RCU grace periods to complete even in the face 922 of softirq monopolization of one or more of the CPUs? For 923 example, is your code subject to network-based denial-of-service 924 attacks? If so, you need RCU-bh. 925 926e. Is your workload too update-intensive for normal use of 927 RCU, but inappropriate for other synchronization mechanisms? 928 If so, consider SLAB_DESTROY_BY_RCU. But please be careful! 929 930f. Do you need read-side critical sections that are respected 931 even though they are in the middle of the idle loop, during 932 user-mode execution, or on an offlined CPU? If so, SRCU is the 933 only choice that will work for you. 934 935g. Otherwise, use RCU. 936 937Of course, this all assumes that you have determined that RCU is in fact 938the right tool for your job. 939 940 9418. ANSWERS TO QUICK QUIZZES 942 943Quick Quiz #1: Why is this argument naive? How could a deadlock 944 occur when using this algorithm in a real-world Linux 945 kernel? [Referring to the lock-based "toy" RCU 946 algorithm.] 947 948Answer: Consider the following sequence of events: 949 950 1. CPU 0 acquires some unrelated lock, call it 951 "problematic_lock", disabling irq via 952 spin_lock_irqsave(). 953 954 2. CPU 1 enters synchronize_rcu(), write-acquiring 955 rcu_gp_mutex. 956 957 3. CPU 0 enters rcu_read_lock(), but must wait 958 because CPU 1 holds rcu_gp_mutex. 959 960 4. CPU 1 is interrupted, and the irq handler 961 attempts to acquire problematic_lock. 962 963 The system is now deadlocked. 964 965 One way to avoid this deadlock is to use an approach like 966 that of CONFIG_PREEMPT_RT, where all normal spinlocks 967 become blocking locks, and all irq handlers execute in 968 the context of special tasks. In this case, in step 4 969 above, the irq handler would block, allowing CPU 1 to 970 release rcu_gp_mutex, avoiding the deadlock. 971 972 Even in the absence of deadlock, this RCU implementation 973 allows latency to "bleed" from readers to other 974 readers through synchronize_rcu(). To see this, 975 consider task A in an RCU read-side critical section 976 (thus read-holding rcu_gp_mutex), task B blocked 977 attempting to write-acquire rcu_gp_mutex, and 978 task C blocked in rcu_read_lock() attempting to 979 read_acquire rcu_gp_mutex. Task A's RCU read-side 980 latency is holding up task C, albeit indirectly via 981 task B. 982 983 Realtime RCU implementations therefore use a counter-based 984 approach where tasks in RCU read-side critical sections 985 cannot be blocked by tasks executing synchronize_rcu(). 986 987Quick Quiz #2: Give an example where Classic RCU's read-side 988 overhead is -negative-. 989 990Answer: Imagine a single-CPU system with a non-CONFIG_PREEMPT 991 kernel where a routing table is used by process-context 992 code, but can be updated by irq-context code (for example, 993 by an "ICMP REDIRECT" packet). The usual way of handling 994 this would be to have the process-context code disable 995 interrupts while searching the routing table. Use of 996 RCU allows such interrupt-disabling to be dispensed with. 997 Thus, without RCU, you pay the cost of disabling interrupts, 998 and with RCU you don't. 999 1000 One can argue that the overhead of RCU in this 1001 case is negative with respect to the single-CPU 1002 interrupt-disabling approach. Others might argue that 1003 the overhead of RCU is merely zero, and that replacing 1004 the positive overhead of the interrupt-disabling scheme 1005 with the zero-overhead RCU scheme does not constitute 1006 negative overhead. 1007 1008 In real life, of course, things are more complex. But 1009 even the theoretical possibility of negative overhead for 1010 a synchronization primitive is a bit unexpected. ;-) 1011 1012Quick Quiz #3: If it is illegal to block in an RCU read-side 1013 critical section, what the heck do you do in 1014 PREEMPT_RT, where normal spinlocks can block??? 1015 1016Answer: Just as PREEMPT_RT permits preemption of spinlock 1017 critical sections, it permits preemption of RCU 1018 read-side critical sections. It also permits 1019 spinlocks blocking while in RCU read-side critical 1020 sections. 1021 1022 Why the apparent inconsistency? Because it is it 1023 possible to use priority boosting to keep the RCU 1024 grace periods short if need be (for example, if running 1025 short of memory). In contrast, if blocking waiting 1026 for (say) network reception, there is no way to know 1027 what should be boosted. Especially given that the 1028 process we need to boost might well be a human being 1029 who just went out for a pizza or something. And although 1030 a computer-operated cattle prod might arouse serious 1031 interest, it might also provoke serious objections. 1032 Besides, how does the computer know what pizza parlor 1033 the human being went to??? 1034 1035 1036ACKNOWLEDGEMENTS 1037 1038My thanks to the people who helped make this human-readable, including 1039Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern. 1040 1041 1042For more information, see http://www.rdrop.com/users/paulmck/RCU. 1043