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 9 10What is RCU? 11 12RCU is a synchronization mechanism that was added to the Linux kernel 13during the 2.5 development effort that is optimized for read-mostly 14situations. Although RCU is actually quite simple once you understand it, 15getting there can sometimes be a challenge. Part of the problem is that 16most of the past descriptions of RCU have been written with the mistaken 17assumption that there is "one true way" to describe RCU. Instead, 18the experience has been that different people must take different paths 19to arrive at an understanding of RCU. This document provides several 20different paths, as follows: 21 221. RCU OVERVIEW 232. WHAT IS RCU'S CORE API? 243. WHAT ARE SOME EXAMPLE USES OF CORE RCU API? 254. WHAT IF MY UPDATING THREAD CANNOT BLOCK? 265. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU? 276. ANALOGY WITH READER-WRITER LOCKING 287. FULL LIST OF RCU APIs 298. ANSWERS TO QUICK QUIZZES 30 31People who prefer starting with a conceptual overview should focus on 32Section 1, though most readers will profit by reading this section at 33some point. People who prefer to start with an API that they can then 34experiment with should focus on Section 2. People who prefer to start 35with example uses should focus on Sections 3 and 4. People who need to 36understand the RCU implementation should focus on Section 5, then dive 37into the kernel source code. People who reason best by analogy should 38focus on Section 6. Section 7 serves as an index to the docbook API 39documentation, and Section 8 is the traditional answer key. 40 41So, start with the section that makes the most sense to you and your 42preferred method of learning. If you need to know everything about 43everything, feel free to read the whole thing -- but if you are really 44that type of person, you have perused the source code and will therefore 45never need this document anyway. ;-) 46 47 481. RCU OVERVIEW 49 50The basic idea behind RCU is to split updates into "removal" and 51"reclamation" phases. The removal phase removes references to data items 52within a data structure (possibly by replacing them with references to 53new versions of these data items), and can run concurrently with readers. 54The reason that it is safe to run the removal phase concurrently with 55readers is the semantics of modern CPUs guarantee that readers will see 56either the old or the new version of the data structure rather than a 57partially updated reference. The reclamation phase does the work of reclaiming 58(e.g., freeing) the data items removed from the data structure during the 59removal phase. Because reclaiming data items can disrupt any readers 60concurrently referencing those data items, the reclamation phase must 61not start until readers no longer hold references to those data items. 62 63Splitting the update into removal and reclamation phases permits the 64updater to perform the removal phase immediately, and to defer the 65reclamation phase until all readers active during the removal phase have 66completed, either by blocking until they finish or by registering a 67callback that is invoked after they finish. Only readers that are active 68during the removal phase need be considered, because any reader starting 69after the removal phase will be unable to gain a reference to the removed 70data items, and therefore cannot be disrupted by the reclamation phase. 71 72So the typical RCU update sequence goes something like the following: 73 74a. Remove pointers to a data structure, so that subsequent 75 readers cannot gain a reference to it. 76 77b. Wait for all previous readers to complete their RCU read-side 78 critical sections. 79 80c. At this point, there cannot be any readers who hold references 81 to the data structure, so it now may safely be reclaimed 82 (e.g., kfree()d). 83 84Step (b) above is the key idea underlying RCU's deferred destruction. 85The ability to wait until all readers are done allows RCU readers to 86use much lighter-weight synchronization, in some cases, absolutely no 87synchronization at all. In contrast, in more conventional lock-based 88schemes, readers must use heavy-weight synchronization in order to 89prevent an updater from deleting the data structure out from under them. 90This is because lock-based updaters typically update data items in place, 91and must therefore exclude readers. In contrast, RCU-based updaters 92typically take advantage of the fact that writes to single aligned 93pointers are atomic on modern CPUs, allowing atomic insertion, removal, 94and replacement of data items in a linked structure without disrupting 95readers. Concurrent RCU readers can then continue accessing the old 96versions, and can dispense with the atomic operations, memory barriers, 97and communications cache misses that are so expensive on present-day 98SMP computer systems, even in absence of lock contention. 99 100In the three-step procedure shown above, the updater is performing both 101the removal and the reclamation step, but it is often helpful for an 102entirely different thread to do the reclamation, as is in fact the case 103in the Linux kernel's directory-entry cache (dcache). Even if the same 104thread performs both the update step (step (a) above) and the reclamation 105step (step (c) above), it is often helpful to think of them separately. 106For example, RCU readers and updaters need not communicate at all, 107but RCU provides implicit low-overhead communication between readers 108and reclaimers, namely, in step (b) above. 109 110So how the heck can a reclaimer tell when a reader is done, given 111that readers are not doing any sort of synchronization operations??? 112Read on to learn about how RCU's API makes this easy. 113 114 1152. WHAT IS RCU'S CORE API? 116 117The core RCU API is quite small: 118 119a. rcu_read_lock() 120b. rcu_read_unlock() 121c. synchronize_rcu() / call_rcu() 122d. rcu_assign_pointer() 123e. rcu_dereference() 124 125There are many other members of the RCU API, but the rest can be 126expressed in terms of these five, though most implementations instead 127express synchronize_rcu() in terms of the call_rcu() callback API. 128 129The five core RCU APIs are described below, the other 18 will be enumerated 130later. See the kernel docbook documentation for more info, or look directly 131at the function header comments. 132 133rcu_read_lock() 134 135 void rcu_read_lock(void); 136 137 Used by a reader to inform the reclaimer that the reader is 138 entering an RCU read-side critical section. It is illegal 139 to block while in an RCU read-side critical section, though 140 kernels built with CONFIG_PREEMPT_RCU can preempt RCU 141 read-side critical sections. Any RCU-protected data structure 142 accessed during an RCU read-side critical section is guaranteed to 143 remain unreclaimed for the full duration of that critical section. 144 Reference counts may be used in conjunction with RCU to maintain 145 longer-term references to data structures. 146 147rcu_read_unlock() 148 149 void rcu_read_unlock(void); 150 151 Used by a reader to inform the reclaimer that the reader is 152 exiting an RCU read-side critical section. Note that RCU 153 read-side critical sections may be nested and/or overlapping. 154 155synchronize_rcu() 156 157 void synchronize_rcu(void); 158 159 Marks the end of updater code and the beginning of reclaimer 160 code. It does this by blocking until all pre-existing RCU 161 read-side critical sections on all CPUs have completed. 162 Note that synchronize_rcu() will -not- necessarily wait for 163 any subsequent RCU read-side critical sections to complete. 164 For example, consider the following sequence of events: 165 166 CPU 0 CPU 1 CPU 2 167 ----------------- ------------------------- --------------- 168 1. rcu_read_lock() 169 2. enters synchronize_rcu() 170 3. rcu_read_lock() 171 4. rcu_read_unlock() 172 5. exits synchronize_rcu() 173 6. rcu_read_unlock() 174 175 To reiterate, synchronize_rcu() waits only for ongoing RCU 176 read-side critical sections to complete, not necessarily for 177 any that begin after synchronize_rcu() is invoked. 178 179 Of course, synchronize_rcu() does not necessarily return 180 -immediately- after the last pre-existing RCU read-side critical 181 section completes. For one thing, there might well be scheduling 182 delays. For another thing, many RCU implementations process 183 requests in batches in order to improve efficiencies, which can 184 further delay synchronize_rcu(). 185 186 Since synchronize_rcu() is the API that must figure out when 187 readers are done, its implementation is key to RCU. For RCU 188 to be useful in all but the most read-intensive situations, 189 synchronize_rcu()'s overhead must also be quite small. 190 191 The call_rcu() API is a callback form of synchronize_rcu(), 192 and is described in more detail in a later section. Instead of 193 blocking, it registers a function and argument which are invoked 194 after all ongoing RCU read-side critical sections have completed. 195 This callback variant is particularly useful in situations where 196 it is illegal to block or where update-side performance is 197 critically important. 198 199 However, the call_rcu() API should not be used lightly, as use 200 of the synchronize_rcu() API generally results in simpler code. 201 In addition, the synchronize_rcu() API has the nice property 202 of automatically limiting update rate should grace periods 203 be delayed. This property results in system resilience in face 204 of denial-of-service attacks. Code using call_rcu() should limit 205 update rate in order to gain this same sort of resilience. See 206 checklist.txt for some approaches to limiting the update rate. 207 208rcu_assign_pointer() 209 210 typeof(p) rcu_assign_pointer(p, typeof(p) v); 211 212 Yes, rcu_assign_pointer() -is- implemented as a macro, though it 213 would be cool to be able to declare a function in this manner. 214 (Compiler experts will no doubt disagree.) 215 216 The updater uses this function to assign a new value to an 217 RCU-protected pointer, in order to safely communicate the change 218 in value from the updater to the reader. This function returns 219 the new value, and also executes any memory-barrier instructions 220 required for a given CPU architecture. 221 222 Perhaps just as important, it serves to document (1) which 223 pointers are protected by RCU and (2) the point at which a 224 given structure becomes accessible to other CPUs. That said, 225 rcu_assign_pointer() is most frequently used indirectly, via 226 the _rcu list-manipulation primitives such as list_add_rcu(). 227 228rcu_dereference() 229 230 typeof(p) rcu_dereference(p); 231 232 Like rcu_assign_pointer(), rcu_dereference() must be implemented 233 as a macro. 234 235 The reader uses rcu_dereference() to fetch an RCU-protected 236 pointer, which returns a value that may then be safely 237 dereferenced. Note that rcu_deference() does not actually 238 dereference the pointer, instead, it protects the pointer for 239 later dereferencing. It also executes any needed memory-barrier 240 instructions for a given CPU architecture. Currently, only Alpha 241 needs memory barriers within rcu_dereference() -- on other CPUs, 242 it compiles to nothing, not even a compiler directive. 243 244 Common coding practice uses rcu_dereference() to copy an 245 RCU-protected pointer to a local variable, then dereferences 246 this local variable, for example as follows: 247 248 p = rcu_dereference(head.next); 249 return p->data; 250 251 However, in this case, one could just as easily combine these 252 into one statement: 253 254 return rcu_dereference(head.next)->data; 255 256 If you are going to be fetching multiple fields from the 257 RCU-protected structure, using the local variable is of 258 course preferred. Repeated rcu_dereference() calls look 259 ugly, do not guarantee that the same pointer will be returned 260 if an update happened while in the critical section, and incur 261 unnecessary overhead on Alpha CPUs. 262 263 Note that the value returned by rcu_dereference() is valid 264 only within the enclosing RCU read-side critical section. 265 For example, the following is -not- legal: 266 267 rcu_read_lock(); 268 p = rcu_dereference(head.next); 269 rcu_read_unlock(); 270 x = p->address; /* BUG!!! */ 271 rcu_read_lock(); 272 y = p->data; /* BUG!!! */ 273 rcu_read_unlock(); 274 275 Holding a reference from one RCU read-side critical section 276 to another is just as illegal as holding a reference from 277 one lock-based critical section to another! Similarly, 278 using a reference outside of the critical section in which 279 it was acquired is just as illegal as doing so with normal 280 locking. 281 282 As with rcu_assign_pointer(), an important function of 283 rcu_dereference() is to document which pointers are protected by 284 RCU, in particular, flagging a pointer that is subject to changing 285 at any time, including immediately after the rcu_dereference(). 286 And, again like rcu_assign_pointer(), rcu_dereference() is 287 typically used indirectly, via the _rcu list-manipulation 288 primitives, such as list_for_each_entry_rcu(). 289 290The following diagram shows how each API communicates among the 291reader, updater, and reclaimer. 292 293 294 rcu_assign_pointer() 295 +--------+ 296 +---------------------->| reader |---------+ 297 | +--------+ | 298 | | | 299 | | | Protect: 300 | | | rcu_read_lock() 301 | | | rcu_read_unlock() 302 | rcu_dereference() | | 303 +---------+ | | 304 | updater |<---------------------+ | 305 +---------+ V 306 | +-----------+ 307 +----------------------------------->| reclaimer | 308 +-----------+ 309 Defer: 310 synchronize_rcu() & call_rcu() 311 312 313The RCU infrastructure observes the time sequence of rcu_read_lock(), 314rcu_read_unlock(), synchronize_rcu(), and call_rcu() invocations in 315order to determine when (1) synchronize_rcu() invocations may return 316to their callers and (2) call_rcu() callbacks may be invoked. Efficient 317implementations of the RCU infrastructure make heavy use of batching in 318order to amortize their overhead over many uses of the corresponding APIs. 319 320There are no fewer than three RCU mechanisms in the Linux kernel; the 321diagram above shows the first one, which is by far the most commonly used. 322The rcu_dereference() and rcu_assign_pointer() primitives are used for 323all three mechanisms, but different defer and protect primitives are 324used as follows: 325 326 Defer Protect 327 328a. synchronize_rcu() rcu_read_lock() / rcu_read_unlock() 329 call_rcu() rcu_dereference() 330 331b. synchronize_rcu_bh() rcu_read_lock_bh() / rcu_read_unlock_bh() 332 call_rcu_bh() rcu_dereference_bh() 333 334c. synchronize_sched() rcu_read_lock_sched() / rcu_read_unlock_sched() 335 call_rcu_sched() preempt_disable() / preempt_enable() 336 local_irq_save() / local_irq_restore() 337 hardirq enter / hardirq exit 338 NMI enter / NMI exit 339 rcu_dereference_sched() 340 341These three mechanisms are used as follows: 342 343a. RCU applied to normal data structures. 344 345b. RCU applied to networking data structures that may be subjected 346 to remote denial-of-service attacks. 347 348c. RCU applied to scheduler and interrupt/NMI-handler tasks. 349 350Again, most uses will be of (a). The (b) and (c) cases are important 351for specialized uses, but are relatively uncommon. 352 353 3543. WHAT ARE SOME EXAMPLE USES OF CORE RCU API? 355 356This section shows a simple use of the core RCU API to protect a 357global pointer to a dynamically allocated structure. More-typical 358uses of RCU may be found in listRCU.txt, arrayRCU.txt, and NMI-RCU.txt. 359 360 struct foo { 361 int a; 362 char b; 363 long c; 364 }; 365 DEFINE_SPINLOCK(foo_mutex); 366 367 struct foo __rcu *gbl_foo; 368 369 /* 370 * Create a new struct foo that is the same as the one currently 371 * pointed to by gbl_foo, except that field "a" is replaced 372 * with "new_a". Points gbl_foo to the new structure, and 373 * frees up the old structure after a grace period. 374 * 375 * Uses rcu_assign_pointer() to ensure that concurrent readers 376 * see the initialized version of the new structure. 377 * 378 * Uses synchronize_rcu() to ensure that any readers that might 379 * have references to the old structure complete before freeing 380 * the old structure. 381 */ 382 void foo_update_a(int new_a) 383 { 384 struct foo *new_fp; 385 struct foo *old_fp; 386 387 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL); 388 spin_lock(&foo_mutex); 389 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex)); 390 *new_fp = *old_fp; 391 new_fp->a = new_a; 392 rcu_assign_pointer(gbl_foo, new_fp); 393 spin_unlock(&foo_mutex); 394 synchronize_rcu(); 395 kfree(old_fp); 396 } 397 398 /* 399 * Return the value of field "a" of the current gbl_foo 400 * structure. Use rcu_read_lock() and rcu_read_unlock() 401 * to ensure that the structure does not get deleted out 402 * from under us, and use rcu_dereference() to ensure that 403 * we see the initialized version of the structure (important 404 * for DEC Alpha and for people reading the code). 405 */ 406 int foo_get_a(void) 407 { 408 int retval; 409 410 rcu_read_lock(); 411 retval = rcu_dereference(gbl_foo)->a; 412 rcu_read_unlock(); 413 return retval; 414 } 415 416So, to sum up: 417 418o Use rcu_read_lock() and rcu_read_unlock() to guard RCU 419 read-side critical sections. 420 421o Within an RCU read-side critical section, use rcu_dereference() 422 to dereference RCU-protected pointers. 423 424o Use some solid scheme (such as locks or semaphores) to 425 keep concurrent updates from interfering with each other. 426 427o Use rcu_assign_pointer() to update an RCU-protected pointer. 428 This primitive protects concurrent readers from the updater, 429 -not- concurrent updates from each other! You therefore still 430 need to use locking (or something similar) to keep concurrent 431 rcu_assign_pointer() primitives from interfering with each other. 432 433o Use synchronize_rcu() -after- removing a data element from an 434 RCU-protected data structure, but -before- reclaiming/freeing 435 the data element, in order to wait for the completion of all 436 RCU read-side critical sections that might be referencing that 437 data item. 438 439See checklist.txt for additional rules to follow when using RCU. 440And again, more-typical uses of RCU may be found in listRCU.txt, 441arrayRCU.txt, and NMI-RCU.txt. 442 443 4444. WHAT IF MY UPDATING THREAD CANNOT BLOCK? 445 446In the example above, foo_update_a() blocks until a grace period elapses. 447This is quite simple, but in some cases one cannot afford to wait so 448long -- there might be other high-priority work to be done. 449 450In such cases, one uses call_rcu() rather than synchronize_rcu(). 451The call_rcu() API is as follows: 452 453 void call_rcu(struct rcu_head * head, 454 void (*func)(struct rcu_head *head)); 455 456This function invokes func(head) after a grace period has elapsed. 457This invocation might happen from either softirq or process context, 458so the function is not permitted to block. The foo struct needs to 459have an rcu_head structure added, perhaps as follows: 460 461 struct foo { 462 int a; 463 char b; 464 long c; 465 struct rcu_head rcu; 466 }; 467 468The foo_update_a() function might then be written as follows: 469 470 /* 471 * Create a new struct foo that is the same as the one currently 472 * pointed to by gbl_foo, except that field "a" is replaced 473 * with "new_a". Points gbl_foo to the new structure, and 474 * frees up the old structure after a grace period. 475 * 476 * Uses rcu_assign_pointer() to ensure that concurrent readers 477 * see the initialized version of the new structure. 478 * 479 * Uses call_rcu() to ensure that any readers that might have 480 * references to the old structure complete before freeing the 481 * old structure. 482 */ 483 void foo_update_a(int new_a) 484 { 485 struct foo *new_fp; 486 struct foo *old_fp; 487 488 new_fp = kmalloc(sizeof(*new_fp), GFP_KERNEL); 489 spin_lock(&foo_mutex); 490 old_fp = rcu_dereference_protected(gbl_foo, lockdep_is_held(&foo_mutex)); 491 *new_fp = *old_fp; 492 new_fp->a = new_a; 493 rcu_assign_pointer(gbl_foo, new_fp); 494 spin_unlock(&foo_mutex); 495 call_rcu(&old_fp->rcu, foo_reclaim); 496 } 497 498The foo_reclaim() function might appear as follows: 499 500 void foo_reclaim(struct rcu_head *rp) 501 { 502 struct foo *fp = container_of(rp, struct foo, rcu); 503 504 foo_cleanup(fp->a); 505 506 kfree(fp); 507 } 508 509The container_of() primitive is a macro that, given a pointer into a 510struct, the type of the struct, and the pointed-to field within the 511struct, returns a pointer to the beginning of the struct. 512 513The use of call_rcu() permits the caller of foo_update_a() to 514immediately regain control, without needing to worry further about the 515old version of the newly updated element. It also clearly shows the 516RCU distinction between updater, namely foo_update_a(), and reclaimer, 517namely foo_reclaim(). 518 519The summary of advice is the same as for the previous section, except 520that we are now using call_rcu() rather than synchronize_rcu(): 521 522o Use call_rcu() -after- removing a data element from an 523 RCU-protected data structure in order to register a callback 524 function that will be invoked after the completion of all RCU 525 read-side critical sections that might be referencing that 526 data item. 527 528If the callback for call_rcu() is not doing anything more than calling 529kfree() on the structure, you can use kfree_rcu() instead of call_rcu() 530to avoid having to write your own callback: 531 532 kfree_rcu(old_fp, rcu); 533 534Again, see checklist.txt for additional rules governing the use of RCU. 535 536 5375. WHAT ARE SOME SIMPLE IMPLEMENTATIONS OF RCU? 538 539One of the nice things about RCU is that it has extremely simple "toy" 540implementations that are a good first step towards understanding the 541production-quality implementations in the Linux kernel. This section 542presents two such "toy" implementations of RCU, one that is implemented 543in terms of familiar locking primitives, and another that more closely 544resembles "classic" RCU. Both are way too simple for real-world use, 545lacking both functionality and performance. However, they are useful 546in getting a feel for how RCU works. See kernel/rcupdate.c for a 547production-quality implementation, and see: 548 549 http://www.rdrop.com/users/paulmck/RCU 550 551for papers describing the Linux kernel RCU implementation. The OLS'01 552and OLS'02 papers are a good introduction, and the dissertation provides 553more details on the current implementation as of early 2004. 554 555 5565A. "TOY" IMPLEMENTATION #1: LOCKING 557 558This section presents a "toy" RCU implementation that is based on 559familiar locking primitives. Its overhead makes it a non-starter for 560real-life use, as does its lack of scalability. It is also unsuitable 561for realtime use, since it allows scheduling latency to "bleed" from 562one read-side critical section to another. 563 564However, it is probably the easiest implementation to relate to, so is 565a good starting point. 566 567It is extremely simple: 568 569 static DEFINE_RWLOCK(rcu_gp_mutex); 570 571 void rcu_read_lock(void) 572 { 573 read_lock(&rcu_gp_mutex); 574 } 575 576 void rcu_read_unlock(void) 577 { 578 read_unlock(&rcu_gp_mutex); 579 } 580 581 void synchronize_rcu(void) 582 { 583 write_lock(&rcu_gp_mutex); 584 write_unlock(&rcu_gp_mutex); 585 } 586 587[You can ignore rcu_assign_pointer() and rcu_dereference() without 588missing much. But here they are anyway. And whatever you do, don't 589forget about them when submitting patches making use of RCU!] 590 591 #define rcu_assign_pointer(p, v) ({ \ 592 smp_wmb(); \ 593 (p) = (v); \ 594 }) 595 596 #define rcu_dereference(p) ({ \ 597 typeof(p) _________p1 = p; \ 598 smp_read_barrier_depends(); \ 599 (_________p1); \ 600 }) 601 602 603The rcu_read_lock() and rcu_read_unlock() primitive read-acquire 604and release a global reader-writer lock. The synchronize_rcu() 605primitive write-acquires this same lock, then immediately releases 606it. This means that once synchronize_rcu() exits, all RCU read-side 607critical sections that were in progress before synchronize_rcu() was 608called are guaranteed to have completed -- there is no way that 609synchronize_rcu() would have been able to write-acquire the lock 610otherwise. 611 612It is possible to nest rcu_read_lock(), since reader-writer locks may 613be recursively acquired. Note also that rcu_read_lock() is immune 614from deadlock (an important property of RCU). The reason for this is 615that the only thing that can block rcu_read_lock() is a synchronize_rcu(). 616But synchronize_rcu() does not acquire any locks while holding rcu_gp_mutex, 617so there can be no deadlock cycle. 618 619Quick Quiz #1: Why is this argument naive? How could a deadlock 620 occur when using this algorithm in a real-world Linux 621 kernel? How could this deadlock be avoided? 622 623 6245B. "TOY" EXAMPLE #2: CLASSIC RCU 625 626This section presents a "toy" RCU implementation that is based on 627"classic RCU". It is also short on performance (but only for updates) and 628on features such as hotplug CPU and the ability to run in CONFIG_PREEMPT 629kernels. The definitions of rcu_dereference() and rcu_assign_pointer() 630are the same as those shown in the preceding section, so they are omitted. 631 632 void rcu_read_lock(void) { } 633 634 void rcu_read_unlock(void) { } 635 636 void synchronize_rcu(void) 637 { 638 int cpu; 639 640 for_each_possible_cpu(cpu) 641 run_on(cpu); 642 } 643 644Note that rcu_read_lock() and rcu_read_unlock() do absolutely nothing. 645This is the great strength of classic RCU in a non-preemptive kernel: 646read-side overhead is precisely zero, at least on non-Alpha CPUs. 647And there is absolutely no way that rcu_read_lock() can possibly 648participate in a deadlock cycle! 649 650The implementation of synchronize_rcu() simply schedules itself on each 651CPU in turn. The run_on() primitive can be implemented straightforwardly 652in terms of the sched_setaffinity() primitive. Of course, a somewhat less 653"toy" implementation would restore the affinity upon completion rather 654than just leaving all tasks running on the last CPU, but when I said 655"toy", I meant -toy-! 656 657So how the heck is this supposed to work??? 658 659Remember that it is illegal to block while in an RCU read-side critical 660section. Therefore, if a given CPU executes a context switch, we know 661that it must have completed all preceding RCU read-side critical sections. 662Once -all- CPUs have executed a context switch, then -all- preceding 663RCU read-side critical sections will have completed. 664 665So, suppose that we remove a data item from its structure and then invoke 666synchronize_rcu(). Once synchronize_rcu() returns, we are guaranteed 667that there are no RCU read-side critical sections holding a reference 668to that data item, so we can safely reclaim it. 669 670Quick Quiz #2: Give an example where Classic RCU's read-side 671 overhead is -negative-. 672 673Quick Quiz #3: If it is illegal to block in an RCU read-side 674 critical section, what the heck do you do in 675 PREEMPT_RT, where normal spinlocks can block??? 676 677 6786. ANALOGY WITH READER-WRITER LOCKING 679 680Although RCU can be used in many different ways, a very common use of 681RCU is analogous to reader-writer locking. The following unified 682diff shows how closely related RCU and reader-writer locking can be. 683 684 @@ -13,15 +14,15 @@ 685 struct list_head *lp; 686 struct el *p; 687 688 - read_lock(); 689 - list_for_each_entry(p, head, lp) { 690 + rcu_read_lock(); 691 + list_for_each_entry_rcu(p, head, lp) { 692 if (p->key == key) { 693 *result = p->data; 694 - read_unlock(); 695 + rcu_read_unlock(); 696 return 1; 697 } 698 } 699 - read_unlock(); 700 + rcu_read_unlock(); 701 return 0; 702 } 703 704 @@ -29,15 +30,16 @@ 705 { 706 struct el *p; 707 708 - write_lock(&listmutex); 709 + spin_lock(&listmutex); 710 list_for_each_entry(p, head, lp) { 711 if (p->key == key) { 712 - list_del(&p->list); 713 - write_unlock(&listmutex); 714 + list_del_rcu(&p->list); 715 + spin_unlock(&listmutex); 716 + synchronize_rcu(); 717 kfree(p); 718 return 1; 719 } 720 } 721 - write_unlock(&listmutex); 722 + spin_unlock(&listmutex); 723 return 0; 724 } 725 726Or, for those who prefer a side-by-side listing: 727 728 1 struct el { 1 struct el { 729 2 struct list_head list; 2 struct list_head list; 730 3 long key; 3 long key; 731 4 spinlock_t mutex; 4 spinlock_t mutex; 732 5 int data; 5 int data; 733 6 /* Other data fields */ 6 /* Other data fields */ 734 7 }; 7 }; 735 8 spinlock_t listmutex; 8 spinlock_t listmutex; 736 9 struct el head; 9 struct el head; 737 738 1 int search(long key, int *result) 1 int search(long key, int *result) 739 2 { 2 { 740 3 struct list_head *lp; 3 struct list_head *lp; 741 4 struct el *p; 4 struct el *p; 742 5 5 743 6 read_lock(); 6 rcu_read_lock(); 744 7 list_for_each_entry(p, head, lp) { 7 list_for_each_entry_rcu(p, head, lp) { 745 8 if (p->key == key) { 8 if (p->key == key) { 746 9 *result = p->data; 9 *result = p->data; 74710 read_unlock(); 10 rcu_read_unlock(); 74811 return 1; 11 return 1; 74912 } 12 } 75013 } 13 } 75114 read_unlock(); 14 rcu_read_unlock(); 75215 return 0; 15 return 0; 75316 } 16 } 754 755 1 int delete(long key) 1 int delete(long key) 756 2 { 2 { 757 3 struct el *p; 3 struct el *p; 758 4 4 759 5 write_lock(&listmutex); 5 spin_lock(&listmutex); 760 6 list_for_each_entry(p, head, lp) { 6 list_for_each_entry(p, head, lp) { 761 7 if (p->key == key) { 7 if (p->key == key) { 762 8 list_del(&p->list); 8 list_del_rcu(&p->list); 763 9 write_unlock(&listmutex); 9 spin_unlock(&listmutex); 764 10 synchronize_rcu(); 76510 kfree(p); 11 kfree(p); 76611 return 1; 12 return 1; 76712 } 13 } 76813 } 14 } 76914 write_unlock(&listmutex); 15 spin_unlock(&listmutex); 77015 return 0; 16 return 0; 77116 } 17 } 772 773Either way, the differences are quite small. Read-side locking moves 774to rcu_read_lock() and rcu_read_unlock, update-side locking moves from 775a reader-writer lock to a simple spinlock, and a synchronize_rcu() 776precedes the kfree(). 777 778However, there is one potential catch: the read-side and update-side 779critical sections can now run concurrently. In many cases, this will 780not be a problem, but it is necessary to check carefully regardless. 781For example, if multiple independent list updates must be seen as 782a single atomic update, converting to RCU will require special care. 783 784Also, the presence of synchronize_rcu() means that the RCU version of 785delete() can now block. If this is a problem, there is a callback-based 786mechanism that never blocks, namely call_rcu() or kfree_rcu(), that can 787be used in place of synchronize_rcu(). 788 789 7907. FULL LIST OF RCU APIs 791 792The RCU APIs are documented in docbook-format header comments in the 793Linux-kernel source code, but it helps to have a full list of the 794APIs, since there does not appear to be a way to categorize them 795in docbook. Here is the list, by category. 796 797RCU list traversal: 798 799 list_entry_rcu 800 list_first_entry_rcu 801 list_next_rcu 802 list_for_each_entry_rcu 803 list_for_each_entry_continue_rcu 804 hlist_first_rcu 805 hlist_next_rcu 806 hlist_pprev_rcu 807 hlist_for_each_entry_rcu 808 hlist_for_each_entry_rcu_bh 809 hlist_for_each_entry_continue_rcu 810 hlist_for_each_entry_continue_rcu_bh 811 hlist_nulls_first_rcu 812 hlist_nulls_for_each_entry_rcu 813 hlist_bl_first_rcu 814 hlist_bl_for_each_entry_rcu 815 816RCU pointer/list update: 817 818 rcu_assign_pointer 819 list_add_rcu 820 list_add_tail_rcu 821 list_del_rcu 822 list_replace_rcu 823 hlist_add_behind_rcu 824 hlist_add_before_rcu 825 hlist_add_head_rcu 826 hlist_del_rcu 827 hlist_del_init_rcu 828 hlist_replace_rcu 829 list_splice_init_rcu() 830 hlist_nulls_del_init_rcu 831 hlist_nulls_del_rcu 832 hlist_nulls_add_head_rcu 833 hlist_bl_add_head_rcu 834 hlist_bl_del_init_rcu 835 hlist_bl_del_rcu 836 hlist_bl_set_first_rcu 837 838RCU: Critical sections Grace period Barrier 839 840 rcu_read_lock synchronize_net rcu_barrier 841 rcu_read_unlock synchronize_rcu 842 rcu_dereference synchronize_rcu_expedited 843 rcu_read_lock_held call_rcu 844 rcu_dereference_check kfree_rcu 845 rcu_dereference_protected 846 847bh: Critical sections Grace period Barrier 848 849 rcu_read_lock_bh call_rcu_bh rcu_barrier_bh 850 rcu_read_unlock_bh synchronize_rcu_bh 851 rcu_dereference_bh synchronize_rcu_bh_expedited 852 rcu_dereference_bh_check 853 rcu_dereference_bh_protected 854 rcu_read_lock_bh_held 855 856sched: Critical sections Grace period Barrier 857 858 rcu_read_lock_sched synchronize_sched rcu_barrier_sched 859 rcu_read_unlock_sched call_rcu_sched 860 [preempt_disable] synchronize_sched_expedited 861 [and friends] 862 rcu_read_lock_sched_notrace 863 rcu_read_unlock_sched_notrace 864 rcu_dereference_sched 865 rcu_dereference_sched_check 866 rcu_dereference_sched_protected 867 rcu_read_lock_sched_held 868 869 870SRCU: Critical sections Grace period Barrier 871 872 srcu_read_lock synchronize_srcu srcu_barrier 873 srcu_read_unlock call_srcu 874 srcu_dereference synchronize_srcu_expedited 875 srcu_dereference_check 876 srcu_read_lock_held 877 878SRCU: Initialization/cleanup 879 init_srcu_struct 880 cleanup_srcu_struct 881 882All: lockdep-checked RCU-protected pointer access 883 884 rcu_access_pointer 885 rcu_dereference_raw 886 RCU_LOCKDEP_WARN 887 rcu_sleep_check 888 RCU_NONIDLE 889 890See the comment headers in the source code (or the docbook generated 891from them) for more information. 892 893However, given that there are no fewer than four families of RCU APIs 894in the Linux kernel, how do you choose which one to use? The following 895list can be helpful: 896 897a. Will readers need to block? If so, you need SRCU. 898 899b. What about the -rt patchset? If readers would need to block 900 in an non-rt kernel, you need SRCU. If readers would block 901 in a -rt kernel, but not in a non-rt kernel, SRCU is not 902 necessary. 903 904c. Do you need to treat NMI handlers, hardirq handlers, 905 and code segments with preemption disabled (whether 906 via preempt_disable(), local_irq_save(), local_bh_disable(), 907 or some other mechanism) as if they were explicit RCU readers? 908 If so, RCU-sched is the only choice that will work for you. 909 910d. Do you need RCU grace periods to complete even in the face 911 of softirq monopolization of one or more of the CPUs? For 912 example, is your code subject to network-based denial-of-service 913 attacks? If so, you need RCU-bh. 914 915e. Is your workload too update-intensive for normal use of 916 RCU, but inappropriate for other synchronization mechanisms? 917 If so, consider SLAB_DESTROY_BY_RCU. But please be careful! 918 919f. Do you need read-side critical sections that are respected 920 even though they are in the middle of the idle loop, during 921 user-mode execution, or on an offlined CPU? If so, SRCU is the 922 only choice that will work for you. 923 924g. Otherwise, use RCU. 925 926Of course, this all assumes that you have determined that RCU is in fact 927the right tool for your job. 928 929 9308. ANSWERS TO QUICK QUIZZES 931 932Quick Quiz #1: Why is this argument naive? How could a deadlock 933 occur when using this algorithm in a real-world Linux 934 kernel? [Referring to the lock-based "toy" RCU 935 algorithm.] 936 937Answer: Consider the following sequence of events: 938 939 1. CPU 0 acquires some unrelated lock, call it 940 "problematic_lock", disabling irq via 941 spin_lock_irqsave(). 942 943 2. CPU 1 enters synchronize_rcu(), write-acquiring 944 rcu_gp_mutex. 945 946 3. CPU 0 enters rcu_read_lock(), but must wait 947 because CPU 1 holds rcu_gp_mutex. 948 949 4. CPU 1 is interrupted, and the irq handler 950 attempts to acquire problematic_lock. 951 952 The system is now deadlocked. 953 954 One way to avoid this deadlock is to use an approach like 955 that of CONFIG_PREEMPT_RT, where all normal spinlocks 956 become blocking locks, and all irq handlers execute in 957 the context of special tasks. In this case, in step 4 958 above, the irq handler would block, allowing CPU 1 to 959 release rcu_gp_mutex, avoiding the deadlock. 960 961 Even in the absence of deadlock, this RCU implementation 962 allows latency to "bleed" from readers to other 963 readers through synchronize_rcu(). To see this, 964 consider task A in an RCU read-side critical section 965 (thus read-holding rcu_gp_mutex), task B blocked 966 attempting to write-acquire rcu_gp_mutex, and 967 task C blocked in rcu_read_lock() attempting to 968 read_acquire rcu_gp_mutex. Task A's RCU read-side 969 latency is holding up task C, albeit indirectly via 970 task B. 971 972 Realtime RCU implementations therefore use a counter-based 973 approach where tasks in RCU read-side critical sections 974 cannot be blocked by tasks executing synchronize_rcu(). 975 976Quick Quiz #2: Give an example where Classic RCU's read-side 977 overhead is -negative-. 978 979Answer: Imagine a single-CPU system with a non-CONFIG_PREEMPT 980 kernel where a routing table is used by process-context 981 code, but can be updated by irq-context code (for example, 982 by an "ICMP REDIRECT" packet). The usual way of handling 983 this would be to have the process-context code disable 984 interrupts while searching the routing table. Use of 985 RCU allows such interrupt-disabling to be dispensed with. 986 Thus, without RCU, you pay the cost of disabling interrupts, 987 and with RCU you don't. 988 989 One can argue that the overhead of RCU in this 990 case is negative with respect to the single-CPU 991 interrupt-disabling approach. Others might argue that 992 the overhead of RCU is merely zero, and that replacing 993 the positive overhead of the interrupt-disabling scheme 994 with the zero-overhead RCU scheme does not constitute 995 negative overhead. 996 997 In real life, of course, things are more complex. But 998 even the theoretical possibility of negative overhead for 999 a synchronization primitive is a bit unexpected. ;-) 1000 1001Quick Quiz #3: If it is illegal to block in an RCU read-side 1002 critical section, what the heck do you do in 1003 PREEMPT_RT, where normal spinlocks can block??? 1004 1005Answer: Just as PREEMPT_RT permits preemption of spinlock 1006 critical sections, it permits preemption of RCU 1007 read-side critical sections. It also permits 1008 spinlocks blocking while in RCU read-side critical 1009 sections. 1010 1011 Why the apparent inconsistency? Because it is it 1012 possible to use priority boosting to keep the RCU 1013 grace periods short if need be (for example, if running 1014 short of memory). In contrast, if blocking waiting 1015 for (say) network reception, there is no way to know 1016 what should be boosted. Especially given that the 1017 process we need to boost might well be a human being 1018 who just went out for a pizza or something. And although 1019 a computer-operated cattle prod might arouse serious 1020 interest, it might also provoke serious objections. 1021 Besides, how does the computer know what pizza parlor 1022 the human being went to??? 1023 1024 1025ACKNOWLEDGEMENTS 1026 1027My thanks to the people who helped make this human-readable, including 1028Jon Walpole, Josh Triplett, Serge Hallyn, Suzanne Wood, and Alan Stern. 1029 1030 1031For more information, see http://www.rdrop.com/users/paulmck/RCU. 1032