1<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN" 2 "http://www.w3.org/TR/html4/loose.dtd"> 3 <html> 4 <head><title>A Tour Through RCU's Requirements [LWN.net]</title> 5 <meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=utf-8"> 6 7<h1>A Tour Through RCU's Requirements</h1> 8 9<p>Copyright IBM Corporation, 2015</p> 10<p>Author: Paul E. McKenney</p> 11<p><i>The initial version of this document appeared in the 12<a href="https://lwn.net/">LWN</a> articles 13<a href="https://lwn.net/Articles/652156/">here</a>, 14<a href="https://lwn.net/Articles/652677/">here</a>, and 15<a href="https://lwn.net/Articles/653326/">here</a>.</i></p> 16 17<h2>Introduction</h2> 18 19<p> 20Read-copy update (RCU) is a synchronization mechanism that is often 21used as a replacement for reader-writer locking. 22RCU is unusual in that updaters do not block readers, 23which means that RCU's read-side primitives can be exceedingly fast 24and scalable. 25In addition, updaters can make useful forward progress concurrently 26with readers. 27However, all this concurrency between RCU readers and updaters does raise 28the question of exactly what RCU readers are doing, which in turn 29raises the question of exactly what RCU's requirements are. 30 31<p> 32This document therefore summarizes RCU's requirements, and can be thought 33of as an informal, high-level specification for RCU. 34It is important to understand that RCU's specification is primarily 35empirical in nature; 36in fact, I learned about many of these requirements the hard way. 37This situation might cause some consternation, however, not only 38has this learning process been a lot of fun, but it has also been 39a great privilege to work with so many people willing to apply 40technologies in interesting new ways. 41 42<p> 43All that aside, here are the categories of currently known RCU requirements: 44</p> 45 46<ol> 47<li> <a href="#Fundamental Requirements"> 48 Fundamental Requirements</a> 49<li> <a href="#Fundamental Non-Requirements">Fundamental Non-Requirements</a> 50<li> <a href="#Parallelism Facts of Life"> 51 Parallelism Facts of Life</a> 52<li> <a href="#Quality-of-Implementation Requirements"> 53 Quality-of-Implementation Requirements</a> 54<li> <a href="#Linux Kernel Complications"> 55 Linux Kernel Complications</a> 56<li> <a href="#Software-Engineering Requirements"> 57 Software-Engineering Requirements</a> 58<li> <a href="#Other RCU Flavors"> 59 Other RCU Flavors</a> 60<li> <a href="#Possible Future Changes"> 61 Possible Future Changes</a> 62</ol> 63 64<p> 65This is followed by a <a href="#Summary">summary</a>, 66however, the answers to each quick quiz immediately follows the quiz. 67Select the big white space with your mouse to see the answer. 68 69<h2><a name="Fundamental Requirements">Fundamental Requirements</a></h2> 70 71<p> 72RCU's fundamental requirements are the closest thing RCU has to hard 73mathematical requirements. 74These are: 75 76<ol> 77<li> <a href="#Grace-Period Guarantee"> 78 Grace-Period Guarantee</a> 79<li> <a href="#Publish-Subscribe Guarantee"> 80 Publish-Subscribe Guarantee</a> 81<li> <a href="#Memory-Barrier Guarantees"> 82 Memory-Barrier Guarantees</a> 83<li> <a href="#RCU Primitives Guaranteed to Execute Unconditionally"> 84 RCU Primitives Guaranteed to Execute Unconditionally</a> 85<li> <a href="#Guaranteed Read-to-Write Upgrade"> 86 Guaranteed Read-to-Write Upgrade</a> 87</ol> 88 89<h3><a name="Grace-Period Guarantee">Grace-Period Guarantee</a></h3> 90 91<p> 92RCU's grace-period guarantee is unusual in being premeditated: 93Jack Slingwine and I had this guarantee firmly in mind when we started 94work on RCU (then called “rclock”) in the early 1990s. 95That said, the past two decades of experience with RCU have produced 96a much more detailed understanding of this guarantee. 97 98<p> 99RCU's grace-period guarantee allows updaters to wait for the completion 100of all pre-existing RCU read-side critical sections. 101An RCU read-side critical section 102begins with the marker <tt>rcu_read_lock()</tt> and ends with 103the marker <tt>rcu_read_unlock()</tt>. 104These markers may be nested, and RCU treats a nested set as one 105big RCU read-side critical section. 106Production-quality implementations of <tt>rcu_read_lock()</tt> and 107<tt>rcu_read_unlock()</tt> are extremely lightweight, and in 108fact have exactly zero overhead in Linux kernels built for production 109use with <tt>CONFIG_PREEMPT=n</tt>. 110 111<p> 112This guarantee allows ordering to be enforced with extremely low 113overhead to readers, for example: 114 115<blockquote> 116<pre> 117 1 int x, y; 118 2 119 3 void thread0(void) 120 4 { 121 5 rcu_read_lock(); 122 6 r1 = READ_ONCE(x); 123 7 r2 = READ_ONCE(y); 124 8 rcu_read_unlock(); 125 9 } 12610 12711 void thread1(void) 12812 { 12913 WRITE_ONCE(x, 1); 13014 synchronize_rcu(); 13115 WRITE_ONCE(y, 1); 13216 } 133</pre> 134</blockquote> 135 136<p> 137Because the <tt>synchronize_rcu()</tt> on line 14 waits for 138all pre-existing readers, any instance of <tt>thread0()</tt> that 139loads a value of zero from <tt>x</tt> must complete before 140<tt>thread1()</tt> stores to <tt>y</tt>, so that instance must 141also load a value of zero from <tt>y</tt>. 142Similarly, any instance of <tt>thread0()</tt> that loads a value of 143one from <tt>y</tt> must have started after the 144<tt>synchronize_rcu()</tt> started, and must therefore also load 145a value of one from <tt>x</tt>. 146Therefore, the outcome: 147<blockquote> 148<pre> 149(r1 == 0 && r2 == 1) 150</pre> 151</blockquote> 152cannot happen. 153 154<table> 155<tr><th> </th></tr> 156<tr><th align="left">Quick Quiz:</th></tr> 157<tr><td> 158 Wait a minute! 159 You said that updaters can make useful forward progress concurrently 160 with readers, but pre-existing readers will block 161 <tt>synchronize_rcu()</tt>!!! 162 Just who are you trying to fool??? 163</td></tr> 164<tr><th align="left">Answer:</th></tr> 165<tr><td bgcolor="#ffffff"><font color="ffffff"> 166 First, if updaters do not wish to be blocked by readers, they can use 167 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt>, which will 168 be discussed later. 169 Second, even when using <tt>synchronize_rcu()</tt>, the other 170 update-side code does run concurrently with readers, whether 171 pre-existing or not. 172</font></td></tr> 173<tr><td> </td></tr> 174</table> 175 176<p> 177This scenario resembles one of the first uses of RCU in 178<a href="https://en.wikipedia.org/wiki/DYNIX">DYNIX/ptx</a>, 179which managed a distributed lock manager's transition into 180a state suitable for handling recovery from node failure, 181more or less as follows: 182 183<blockquote> 184<pre> 185 1 #define STATE_NORMAL 0 186 2 #define STATE_WANT_RECOVERY 1 187 3 #define STATE_RECOVERING 2 188 4 #define STATE_WANT_NORMAL 3 189 5 190 6 int state = STATE_NORMAL; 191 7 192 8 void do_something_dlm(void) 193 9 { 19410 int state_snap; 19511 19612 rcu_read_lock(); 19713 state_snap = READ_ONCE(state); 19814 if (state_snap == STATE_NORMAL) 19915 do_something(); 20016 else 20117 do_something_carefully(); 20218 rcu_read_unlock(); 20319 } 20420 20521 void start_recovery(void) 20622 { 20723 WRITE_ONCE(state, STATE_WANT_RECOVERY); 20824 synchronize_rcu(); 20925 WRITE_ONCE(state, STATE_RECOVERING); 21026 recovery(); 21127 WRITE_ONCE(state, STATE_WANT_NORMAL); 21228 synchronize_rcu(); 21329 WRITE_ONCE(state, STATE_NORMAL); 21430 } 215</pre> 216</blockquote> 217 218<p> 219The RCU read-side critical section in <tt>do_something_dlm()</tt> 220works with the <tt>synchronize_rcu()</tt> in <tt>start_recovery()</tt> 221to guarantee that <tt>do_something()</tt> never runs concurrently 222with <tt>recovery()</tt>, but with little or no synchronization 223overhead in <tt>do_something_dlm()</tt>. 224 225<table> 226<tr><th> </th></tr> 227<tr><th align="left">Quick Quiz:</th></tr> 228<tr><td> 229 Why is the <tt>synchronize_rcu()</tt> on line 28 needed? 230</td></tr> 231<tr><th align="left">Answer:</th></tr> 232<tr><td bgcolor="#ffffff"><font color="ffffff"> 233 Without that extra grace period, memory reordering could result in 234 <tt>do_something_dlm()</tt> executing <tt>do_something()</tt> 235 concurrently with the last bits of <tt>recovery()</tt>. 236</font></td></tr> 237<tr><td> </td></tr> 238</table> 239 240<p> 241In order to avoid fatal problems such as deadlocks, 242an RCU read-side critical section must not contain calls to 243<tt>synchronize_rcu()</tt>. 244Similarly, an RCU read-side critical section must not 245contain anything that waits, directly or indirectly, on completion of 246an invocation of <tt>synchronize_rcu()</tt>. 247 248<p> 249Although RCU's grace-period guarantee is useful in and of itself, with 250<a href="https://lwn.net/Articles/573497/">quite a few use cases</a>, 251it would be good to be able to use RCU to coordinate read-side 252access to linked data structures. 253For this, the grace-period guarantee is not sufficient, as can 254be seen in function <tt>add_gp_buggy()</tt> below. 255We will look at the reader's code later, but in the meantime, just think of 256the reader as locklessly picking up the <tt>gp</tt> pointer, 257and, if the value loaded is non-<tt>NULL</tt>, locklessly accessing the 258<tt>->a</tt> and <tt>->b</tt> fields. 259 260<blockquote> 261<pre> 262 1 bool add_gp_buggy(int a, int b) 263 2 { 264 3 p = kmalloc(sizeof(*p), GFP_KERNEL); 265 4 if (!p) 266 5 return -ENOMEM; 267 6 spin_lock(&gp_lock); 268 7 if (rcu_access_pointer(gp)) { 269 8 spin_unlock(&gp_lock); 270 9 return false; 27110 } 27211 p->a = a; 27312 p->b = a; 27413 gp = p; /* ORDERING BUG */ 27514 spin_unlock(&gp_lock); 27615 return true; 27716 } 278</pre> 279</blockquote> 280 281<p> 282The problem is that both the compiler and weakly ordered CPUs are within 283their rights to reorder this code as follows: 284 285<blockquote> 286<pre> 287 1 bool add_gp_buggy_optimized(int a, int b) 288 2 { 289 3 p = kmalloc(sizeof(*p), GFP_KERNEL); 290 4 if (!p) 291 5 return -ENOMEM; 292 6 spin_lock(&gp_lock); 293 7 if (rcu_access_pointer(gp)) { 294 8 spin_unlock(&gp_lock); 295 9 return false; 29610 } 297<b>11 gp = p; /* ORDERING BUG */ 29812 p->a = a; 29913 p->b = a;</b> 30014 spin_unlock(&gp_lock); 30115 return true; 30216 } 303</pre> 304</blockquote> 305 306<p> 307If an RCU reader fetches <tt>gp</tt> just after 308<tt>add_gp_buggy_optimized</tt> executes line 11, 309it will see garbage in the <tt>->a</tt> and <tt>->b</tt> 310fields. 311And this is but one of many ways in which compiler and hardware optimizations 312could cause trouble. 313Therefore, we clearly need some way to prevent the compiler and the CPU from 314reordering in this manner, which brings us to the publish-subscribe 315guarantee discussed in the next section. 316 317<h3><a name="Publish-Subscribe Guarantee">Publish/Subscribe Guarantee</a></h3> 318 319<p> 320RCU's publish-subscribe guarantee allows data to be inserted 321into a linked data structure without disrupting RCU readers. 322The updater uses <tt>rcu_assign_pointer()</tt> to insert the 323new data, and readers use <tt>rcu_dereference()</tt> to 324access data, whether new or old. 325The following shows an example of insertion: 326 327<blockquote> 328<pre> 329 1 bool add_gp(int a, int b) 330 2 { 331 3 p = kmalloc(sizeof(*p), GFP_KERNEL); 332 4 if (!p) 333 5 return -ENOMEM; 334 6 spin_lock(&gp_lock); 335 7 if (rcu_access_pointer(gp)) { 336 8 spin_unlock(&gp_lock); 337 9 return false; 33810 } 33911 p->a = a; 34012 p->b = a; 34113 rcu_assign_pointer(gp, p); 34214 spin_unlock(&gp_lock); 34315 return true; 34416 } 345</pre> 346</blockquote> 347 348<p> 349The <tt>rcu_assign_pointer()</tt> on line 13 is conceptually 350equivalent to a simple assignment statement, but also guarantees 351that its assignment will 352happen after the two assignments in lines 11 and 12, 353similar to the C11 <tt>memory_order_release</tt> store operation. 354It also prevents any number of “interesting” compiler 355optimizations, for example, the use of <tt>gp</tt> as a scratch 356location immediately preceding the assignment. 357 358<table> 359<tr><th> </th></tr> 360<tr><th align="left">Quick Quiz:</th></tr> 361<tr><td> 362 But <tt>rcu_assign_pointer()</tt> does nothing to prevent the 363 two assignments to <tt>p->a</tt> and <tt>p->b</tt> 364 from being reordered. 365 Can't that also cause problems? 366</td></tr> 367<tr><th align="left">Answer:</th></tr> 368<tr><td bgcolor="#ffffff"><font color="ffffff"> 369 No, it cannot. 370 The readers cannot see either of these two fields until 371 the assignment to <tt>gp</tt>, by which time both fields are 372 fully initialized. 373 So reordering the assignments 374 to <tt>p->a</tt> and <tt>p->b</tt> cannot possibly 375 cause any problems. 376</font></td></tr> 377<tr><td> </td></tr> 378</table> 379 380<p> 381It is tempting to assume that the reader need not do anything special 382to control its accesses to the RCU-protected data, 383as shown in <tt>do_something_gp_buggy()</tt> below: 384 385<blockquote> 386<pre> 387 1 bool do_something_gp_buggy(void) 388 2 { 389 3 rcu_read_lock(); 390 4 p = gp; /* OPTIMIZATIONS GALORE!!! */ 391 5 if (p) { 392 6 do_something(p->a, p->b); 393 7 rcu_read_unlock(); 394 8 return true; 395 9 } 39610 rcu_read_unlock(); 39711 return false; 39812 } 399</pre> 400</blockquote> 401 402<p> 403However, this temptation must be resisted because there are a 404surprisingly large number of ways that the compiler 405(to say nothing of 406<a href="https://h71000.www7.hp.com/wizard/wiz_2637.html">DEC Alpha CPUs</a>) 407can trip this code up. 408For but one example, if the compiler were short of registers, it 409might choose to refetch from <tt>gp</tt> rather than keeping 410a separate copy in <tt>p</tt> as follows: 411 412<blockquote> 413<pre> 414 1 bool do_something_gp_buggy_optimized(void) 415 2 { 416 3 rcu_read_lock(); 417 4 if (gp) { /* OPTIMIZATIONS GALORE!!! */ 418<b> 5 do_something(gp->a, gp->b);</b> 419 6 rcu_read_unlock(); 420 7 return true; 421 8 } 422 9 rcu_read_unlock(); 42310 return false; 42411 } 425</pre> 426</blockquote> 427 428<p> 429If this function ran concurrently with a series of updates that 430replaced the current structure with a new one, 431the fetches of <tt>gp->a</tt> 432and <tt>gp->b</tt> might well come from two different structures, 433which could cause serious confusion. 434To prevent this (and much else besides), <tt>do_something_gp()</tt> uses 435<tt>rcu_dereference()</tt> to fetch from <tt>gp</tt>: 436 437<blockquote> 438<pre> 439 1 bool do_something_gp(void) 440 2 { 441 3 rcu_read_lock(); 442 4 p = rcu_dereference(gp); 443 5 if (p) { 444 6 do_something(p->a, p->b); 445 7 rcu_read_unlock(); 446 8 return true; 447 9 } 44810 rcu_read_unlock(); 44911 return false; 45012 } 451</pre> 452</blockquote> 453 454<p> 455The <tt>rcu_dereference()</tt> uses volatile casts and (for DEC Alpha) 456memory barriers in the Linux kernel. 457Should a 458<a href="http://www.rdrop.com/users/paulmck/RCU/consume.2015.07.13a.pdf">high-quality implementation of C11 <tt>memory_order_consume</tt> [PDF]</a> 459ever appear, then <tt>rcu_dereference()</tt> could be implemented 460as a <tt>memory_order_consume</tt> load. 461Regardless of the exact implementation, a pointer fetched by 462<tt>rcu_dereference()</tt> may not be used outside of the 463outermost RCU read-side critical section containing that 464<tt>rcu_dereference()</tt>, unless protection of 465the corresponding data element has been passed from RCU to some 466other synchronization mechanism, most commonly locking or 467<a href="https://www.kernel.org/doc/Documentation/RCU/rcuref.txt">reference counting</a>. 468 469<p> 470In short, updaters use <tt>rcu_assign_pointer()</tt> and readers 471use <tt>rcu_dereference()</tt>, and these two RCU API elements 472work together to ensure that readers have a consistent view of 473newly added data elements. 474 475<p> 476Of course, it is also necessary to remove elements from RCU-protected 477data structures, for example, using the following process: 478 479<ol> 480<li> Remove the data element from the enclosing structure. 481<li> Wait for all pre-existing RCU read-side critical sections 482 to complete (because only pre-existing readers can possibly have 483 a reference to the newly removed data element). 484<li> At this point, only the updater has a reference to the 485 newly removed data element, so it can safely reclaim 486 the data element, for example, by passing it to <tt>kfree()</tt>. 487</ol> 488 489This process is implemented by <tt>remove_gp_synchronous()</tt>: 490 491<blockquote> 492<pre> 493 1 bool remove_gp_synchronous(void) 494 2 { 495 3 struct foo *p; 496 4 497 5 spin_lock(&gp_lock); 498 6 p = rcu_access_pointer(gp); 499 7 if (!p) { 500 8 spin_unlock(&gp_lock); 501 9 return false; 50210 } 50311 rcu_assign_pointer(gp, NULL); 50412 spin_unlock(&gp_lock); 50513 synchronize_rcu(); 50614 kfree(p); 50715 return true; 50816 } 509</pre> 510</blockquote> 511 512<p> 513This function is straightforward, with line 13 waiting for a grace 514period before line 14 frees the old data element. 515This waiting ensures that readers will reach line 7 of 516<tt>do_something_gp()</tt> before the data element referenced by 517<tt>p</tt> is freed. 518The <tt>rcu_access_pointer()</tt> on line 6 is similar to 519<tt>rcu_dereference()</tt>, except that: 520 521<ol> 522<li> The value returned by <tt>rcu_access_pointer()</tt> 523 cannot be dereferenced. 524 If you want to access the value pointed to as well as 525 the pointer itself, use <tt>rcu_dereference()</tt> 526 instead of <tt>rcu_access_pointer()</tt>. 527<li> The call to <tt>rcu_access_pointer()</tt> need not be 528 protected. 529 In contrast, <tt>rcu_dereference()</tt> must either be 530 within an RCU read-side critical section or in a code 531 segment where the pointer cannot change, for example, in 532 code protected by the corresponding update-side lock. 533</ol> 534 535<table> 536<tr><th> </th></tr> 537<tr><th align="left">Quick Quiz:</th></tr> 538<tr><td> 539 Without the <tt>rcu_dereference()</tt> or the 540 <tt>rcu_access_pointer()</tt>, what destructive optimizations 541 might the compiler make use of? 542</td></tr> 543<tr><th align="left">Answer:</th></tr> 544<tr><td bgcolor="#ffffff"><font color="ffffff"> 545 Let's start with what happens to <tt>do_something_gp()</tt> 546 if it fails to use <tt>rcu_dereference()</tt>. 547 It could reuse a value formerly fetched from this same pointer. 548 It could also fetch the pointer from <tt>gp</tt> in a byte-at-a-time 549 manner, resulting in <i>load tearing</i>, in turn resulting a bytewise 550 mash-up of two distince pointer values. 551 It might even use value-speculation optimizations, where it makes 552 a wrong guess, but by the time it gets around to checking the 553 value, an update has changed the pointer to match the wrong guess. 554 Too bad about any dereferences that returned pre-initialization garbage 555 in the meantime! 556 </font> 557 558 <p><font color="ffffff"> 559 For <tt>remove_gp_synchronous()</tt>, as long as all modifications 560 to <tt>gp</tt> are carried out while holding <tt>gp_lock</tt>, 561 the above optimizations are harmless. 562 However, 563 with <tt>CONFIG_SPARSE_RCU_POINTER=y</tt>, 564 <tt>sparse</tt> will complain if you 565 define <tt>gp</tt> with <tt>__rcu</tt> and then 566 access it without using 567 either <tt>rcu_access_pointer()</tt> or <tt>rcu_dereference()</tt>. 568</font></td></tr> 569<tr><td> </td></tr> 570</table> 571 572<p> 573In short, RCU's publish-subscribe guarantee is provided by the combination 574of <tt>rcu_assign_pointer()</tt> and <tt>rcu_dereference()</tt>. 575This guarantee allows data elements to be safely added to RCU-protected 576linked data structures without disrupting RCU readers. 577This guarantee can be used in combination with the grace-period 578guarantee to also allow data elements to be removed from RCU-protected 579linked data structures, again without disrupting RCU readers. 580 581<p> 582This guarantee was only partially premeditated. 583DYNIX/ptx used an explicit memory barrier for publication, but had nothing 584resembling <tt>rcu_dereference()</tt> for subscription, nor did it 585have anything resembling the <tt>smp_read_barrier_depends()</tt> 586that was later subsumed into <tt>rcu_dereference()</tt>. 587The need for these operations made itself known quite suddenly at a 588late-1990s meeting with the DEC Alpha architects, back in the days when 589DEC was still a free-standing company. 590It took the Alpha architects a good hour to convince me that any sort 591of barrier would ever be needed, and it then took me a good <i>two</i> hours 592to convince them that their documentation did not make this point clear. 593More recent work with the C and C++ standards committees have provided 594much education on tricks and traps from the compiler. 595In short, compilers were much less tricky in the early 1990s, but in 5962015, don't even think about omitting <tt>rcu_dereference()</tt>! 597 598<h3><a name="Memory-Barrier Guarantees">Memory-Barrier Guarantees</a></h3> 599 600<p> 601The previous section's simple linked-data-structure scenario clearly 602demonstrates the need for RCU's stringent memory-ordering guarantees on 603systems with more than one CPU: 604 605<ol> 606<li> Each CPU that has an RCU read-side critical section that 607 begins before <tt>synchronize_rcu()</tt> starts is 608 guaranteed to execute a full memory barrier between the time 609 that the RCU read-side critical section ends and the time that 610 <tt>synchronize_rcu()</tt> returns. 611 Without this guarantee, a pre-existing RCU read-side critical section 612 might hold a reference to the newly removed <tt>struct foo</tt> 613 after the <tt>kfree()</tt> on line 14 of 614 <tt>remove_gp_synchronous()</tt>. 615<li> Each CPU that has an RCU read-side critical section that ends 616 after <tt>synchronize_rcu()</tt> returns is guaranteed 617 to execute a full memory barrier between the time that 618 <tt>synchronize_rcu()</tt> begins and the time that the RCU 619 read-side critical section begins. 620 Without this guarantee, a later RCU read-side critical section 621 running after the <tt>kfree()</tt> on line 14 of 622 <tt>remove_gp_synchronous()</tt> might 623 later run <tt>do_something_gp()</tt> and find the 624 newly deleted <tt>struct foo</tt>. 625<li> If the task invoking <tt>synchronize_rcu()</tt> remains 626 on a given CPU, then that CPU is guaranteed to execute a full 627 memory barrier sometime during the execution of 628 <tt>synchronize_rcu()</tt>. 629 This guarantee ensures that the <tt>kfree()</tt> on 630 line 14 of <tt>remove_gp_synchronous()</tt> really does 631 execute after the removal on line 11. 632<li> If the task invoking <tt>synchronize_rcu()</tt> migrates 633 among a group of CPUs during that invocation, then each of the 634 CPUs in that group is guaranteed to execute a full memory barrier 635 sometime during the execution of <tt>synchronize_rcu()</tt>. 636 This guarantee also ensures that the <tt>kfree()</tt> on 637 line 14 of <tt>remove_gp_synchronous()</tt> really does 638 execute after the removal on 639 line 11, but also in the case where the thread executing the 640 <tt>synchronize_rcu()</tt> migrates in the meantime. 641</ol> 642 643<table> 644<tr><th> </th></tr> 645<tr><th align="left">Quick Quiz:</th></tr> 646<tr><td> 647 Given that multiple CPUs can start RCU read-side critical sections 648 at any time without any ordering whatsoever, how can RCU possibly 649 tell whether or not a given RCU read-side critical section starts 650 before a given instance of <tt>synchronize_rcu()</tt>? 651</td></tr> 652<tr><th align="left">Answer:</th></tr> 653<tr><td bgcolor="#ffffff"><font color="ffffff"> 654 If RCU cannot tell whether or not a given 655 RCU read-side critical section starts before a 656 given instance of <tt>synchronize_rcu()</tt>, 657 then it must assume that the RCU read-side critical section 658 started first. 659 In other words, a given instance of <tt>synchronize_rcu()</tt> 660 can avoid waiting on a given RCU read-side critical section only 661 if it can prove that <tt>synchronize_rcu()</tt> started first. 662</font></td></tr> 663<tr><td> </td></tr> 664</table> 665 666<table> 667<tr><th> </th></tr> 668<tr><th align="left">Quick Quiz:</th></tr> 669<tr><td> 670 The first and second guarantees require unbelievably strict ordering! 671 Are all these memory barriers <i> really</i> required? 672</td></tr> 673<tr><th align="left">Answer:</th></tr> 674<tr><td bgcolor="#ffffff"><font color="ffffff"> 675 Yes, they really are required. 676 To see why the first guarantee is required, consider the following 677 sequence of events: 678 </font> 679 680 <ol> 681 <li> <font color="ffffff"> 682 CPU 1: <tt>rcu_read_lock()</tt> 683 </font> 684 <li> <font color="ffffff"> 685 CPU 1: <tt>q = rcu_dereference(gp); 686 /* Very likely to return p. */</tt> 687 </font> 688 <li> <font color="ffffff"> 689 CPU 0: <tt>list_del_rcu(p);</tt> 690 </font> 691 <li> <font color="ffffff"> 692 CPU 0: <tt>synchronize_rcu()</tt> starts. 693 </font> 694 <li> <font color="ffffff"> 695 CPU 1: <tt>do_something_with(q->a); 696 /* No smp_mb(), so might happen after kfree(). */</tt> 697 </font> 698 <li> <font color="ffffff"> 699 CPU 1: <tt>rcu_read_unlock()</tt> 700 </font> 701 <li> <font color="ffffff"> 702 CPU 0: <tt>synchronize_rcu()</tt> returns. 703 </font> 704 <li> <font color="ffffff"> 705 CPU 0: <tt>kfree(p);</tt> 706 </font> 707 </ol> 708 709 <p><font color="ffffff"> 710 Therefore, there absolutely must be a full memory barrier between the 711 end of the RCU read-side critical section and the end of the 712 grace period. 713 </font> 714 715 <p><font color="ffffff"> 716 The sequence of events demonstrating the necessity of the second rule 717 is roughly similar: 718 </font> 719 720 <ol> 721 <li> <font color="ffffff">CPU 0: <tt>list_del_rcu(p);</tt> 722 </font> 723 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> starts. 724 </font> 725 <li> <font color="ffffff">CPU 1: <tt>rcu_read_lock()</tt> 726 </font> 727 <li> <font color="ffffff">CPU 1: <tt>q = rcu_dereference(gp); 728 /* Might return p if no memory barrier. */</tt> 729 </font> 730 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> returns. 731 </font> 732 <li> <font color="ffffff">CPU 0: <tt>kfree(p);</tt> 733 </font> 734 <li> <font color="ffffff"> 735 CPU 1: <tt>do_something_with(q->a); /* Boom!!! */</tt> 736 </font> 737 <li> <font color="ffffff">CPU 1: <tt>rcu_read_unlock()</tt> 738 </font> 739 </ol> 740 741 <p><font color="ffffff"> 742 And similarly, without a memory barrier between the beginning of the 743 grace period and the beginning of the RCU read-side critical section, 744 CPU 1 might end up accessing the freelist. 745 </font> 746 747 <p><font color="ffffff"> 748 The “as if” rule of course applies, so that any 749 implementation that acts as if the appropriate memory barriers 750 were in place is a correct implementation. 751 That said, it is much easier to fool yourself into believing 752 that you have adhered to the as-if rule than it is to actually 753 adhere to it! 754</font></td></tr> 755<tr><td> </td></tr> 756</table> 757 758<table> 759<tr><th> </th></tr> 760<tr><th align="left">Quick Quiz:</th></tr> 761<tr><td> 762 You claim that <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt> 763 generate absolutely no code in some kernel builds. 764 This means that the compiler might arbitrarily rearrange consecutive 765 RCU read-side critical sections. 766 Given such rearrangement, if a given RCU read-side critical section 767 is done, how can you be sure that all prior RCU read-side critical 768 sections are done? 769 Won't the compiler rearrangements make that impossible to determine? 770</td></tr> 771<tr><th align="left">Answer:</th></tr> 772<tr><td bgcolor="#ffffff"><font color="ffffff"> 773 In cases where <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt> 774 generate absolutely no code, RCU infers quiescent states only at 775 special locations, for example, within the scheduler. 776 Because calls to <tt>schedule()</tt> had better prevent calling-code 777 accesses to shared variables from being rearranged across the call to 778 <tt>schedule()</tt>, if RCU detects the end of a given RCU read-side 779 critical section, it will necessarily detect the end of all prior 780 RCU read-side critical sections, no matter how aggressively the 781 compiler scrambles the code. 782 </font> 783 784 <p><font color="ffffff"> 785 Again, this all assumes that the compiler cannot scramble code across 786 calls to the scheduler, out of interrupt handlers, into the idle loop, 787 into user-mode code, and so on. 788 But if your kernel build allows that sort of scrambling, you have broken 789 far more than just RCU! 790</font></td></tr> 791<tr><td> </td></tr> 792</table> 793 794<p> 795Note that these memory-barrier requirements do not replace the fundamental 796RCU requirement that a grace period wait for all pre-existing readers. 797On the contrary, the memory barriers called out in this section must operate in 798such a way as to <i>enforce</i> this fundamental requirement. 799Of course, different implementations enforce this requirement in different 800ways, but enforce it they must. 801 802<h3><a name="RCU Primitives Guaranteed to Execute Unconditionally">RCU Primitives Guaranteed to Execute Unconditionally</a></h3> 803 804<p> 805The common-case RCU primitives are unconditional. 806They are invoked, they do their job, and they return, with no possibility 807of error, and no need to retry. 808This is a key RCU design philosophy. 809 810<p> 811However, this philosophy is pragmatic rather than pigheaded. 812If someone comes up with a good justification for a particular conditional 813RCU primitive, it might well be implemented and added. 814After all, this guarantee was reverse-engineered, not premeditated. 815The unconditional nature of the RCU primitives was initially an 816accident of implementation, and later experience with synchronization 817primitives with conditional primitives caused me to elevate this 818accident to a guarantee. 819Therefore, the justification for adding a conditional primitive to 820RCU would need to be based on detailed and compelling use cases. 821 822<h3><a name="Guaranteed Read-to-Write Upgrade">Guaranteed Read-to-Write Upgrade</a></h3> 823 824<p> 825As far as RCU is concerned, it is always possible to carry out an 826update within an RCU read-side critical section. 827For example, that RCU read-side critical section might search for 828a given data element, and then might acquire the update-side 829spinlock in order to update that element, all while remaining 830in that RCU read-side critical section. 831Of course, it is necessary to exit the RCU read-side critical section 832before invoking <tt>synchronize_rcu()</tt>, however, this 833inconvenience can be avoided through use of the 834<tt>call_rcu()</tt> and <tt>kfree_rcu()</tt> API members 835described later in this document. 836 837<table> 838<tr><th> </th></tr> 839<tr><th align="left">Quick Quiz:</th></tr> 840<tr><td> 841 But how does the upgrade-to-write operation exclude other readers? 842</td></tr> 843<tr><th align="left">Answer:</th></tr> 844<tr><td bgcolor="#ffffff"><font color="ffffff"> 845 It doesn't, just like normal RCU updates, which also do not exclude 846 RCU readers. 847</font></td></tr> 848<tr><td> </td></tr> 849</table> 850 851<p> 852This guarantee allows lookup code to be shared between read-side 853and update-side code, and was premeditated, appearing in the earliest 854DYNIX/ptx RCU documentation. 855 856<h2><a name="Fundamental Non-Requirements">Fundamental Non-Requirements</a></h2> 857 858<p> 859RCU provides extremely lightweight readers, and its read-side guarantees, 860though quite useful, are correspondingly lightweight. 861It is therefore all too easy to assume that RCU is guaranteeing more 862than it really is. 863Of course, the list of things that RCU does not guarantee is infinitely 864long, however, the following sections list a few non-guarantees that 865have caused confusion. 866Except where otherwise noted, these non-guarantees were premeditated. 867 868<ol> 869<li> <a href="#Readers Impose Minimal Ordering"> 870 Readers Impose Minimal Ordering</a> 871<li> <a href="#Readers Do Not Exclude Updaters"> 872 Readers Do Not Exclude Updaters</a> 873<li> <a href="#Updaters Only Wait For Old Readers"> 874 Updaters Only Wait For Old Readers</a> 875<li> <a href="#Grace Periods Don't Partition Read-Side Critical Sections"> 876 Grace Periods Don't Partition Read-Side Critical Sections</a> 877<li> <a href="#Read-Side Critical Sections Don't Partition Grace Periods"> 878 Read-Side Critical Sections Don't Partition Grace Periods</a> 879<li> <a href="#Disabling Preemption Does Not Block Grace Periods"> 880 Disabling Preemption Does Not Block Grace Periods</a> 881</ol> 882 883<h3><a name="Readers Impose Minimal Ordering">Readers Impose Minimal Ordering</a></h3> 884 885<p> 886Reader-side markers such as <tt>rcu_read_lock()</tt> and 887<tt>rcu_read_unlock()</tt> provide absolutely no ordering guarantees 888except through their interaction with the grace-period APIs such as 889<tt>synchronize_rcu()</tt>. 890To see this, consider the following pair of threads: 891 892<blockquote> 893<pre> 894 1 void thread0(void) 895 2 { 896 3 rcu_read_lock(); 897 4 WRITE_ONCE(x, 1); 898 5 rcu_read_unlock(); 899 6 rcu_read_lock(); 900 7 WRITE_ONCE(y, 1); 901 8 rcu_read_unlock(); 902 9 } 90310 90411 void thread1(void) 90512 { 90613 rcu_read_lock(); 90714 r1 = READ_ONCE(y); 90815 rcu_read_unlock(); 90916 rcu_read_lock(); 91017 r2 = READ_ONCE(x); 91118 rcu_read_unlock(); 91219 } 913</pre> 914</blockquote> 915 916<p> 917After <tt>thread0()</tt> and <tt>thread1()</tt> execute 918concurrently, it is quite possible to have 919 920<blockquote> 921<pre> 922(r1 == 1 && r2 == 0) 923</pre> 924</blockquote> 925 926(that is, <tt>y</tt> appears to have been assigned before <tt>x</tt>), 927which would not be possible if <tt>rcu_read_lock()</tt> and 928<tt>rcu_read_unlock()</tt> had much in the way of ordering 929properties. 930But they do not, so the CPU is within its rights 931to do significant reordering. 932This is by design: Any significant ordering constraints would slow down 933these fast-path APIs. 934 935<table> 936<tr><th> </th></tr> 937<tr><th align="left">Quick Quiz:</th></tr> 938<tr><td> 939 Can't the compiler also reorder this code? 940</td></tr> 941<tr><th align="left">Answer:</th></tr> 942<tr><td bgcolor="#ffffff"><font color="ffffff"> 943 No, the volatile casts in <tt>READ_ONCE()</tt> and 944 <tt>WRITE_ONCE()</tt> prevent the compiler from reordering in 945 this particular case. 946</font></td></tr> 947<tr><td> </td></tr> 948</table> 949 950<h3><a name="Readers Do Not Exclude Updaters">Readers Do Not Exclude Updaters</a></h3> 951 952<p> 953Neither <tt>rcu_read_lock()</tt> nor <tt>rcu_read_unlock()</tt> 954exclude updates. 955All they do is to prevent grace periods from ending. 956The following example illustrates this: 957 958<blockquote> 959<pre> 960 1 void thread0(void) 961 2 { 962 3 rcu_read_lock(); 963 4 r1 = READ_ONCE(y); 964 5 if (r1) { 965 6 do_something_with_nonzero_x(); 966 7 r2 = READ_ONCE(x); 967 8 WARN_ON(!r2); /* BUG!!! */ 968 9 } 96910 rcu_read_unlock(); 97011 } 97112 97213 void thread1(void) 97314 { 97415 spin_lock(&my_lock); 97516 WRITE_ONCE(x, 1); 97617 WRITE_ONCE(y, 1); 97718 spin_unlock(&my_lock); 97819 } 979</pre> 980</blockquote> 981 982<p> 983If the <tt>thread0()</tt> function's <tt>rcu_read_lock()</tt> 984excluded the <tt>thread1()</tt> function's update, 985the <tt>WARN_ON()</tt> could never fire. 986But the fact is that <tt>rcu_read_lock()</tt> does not exclude 987much of anything aside from subsequent grace periods, of which 988<tt>thread1()</tt> has none, so the 989<tt>WARN_ON()</tt> can and does fire. 990 991<h3><a name="Updaters Only Wait For Old Readers">Updaters Only Wait For Old Readers</a></h3> 992 993<p> 994It might be tempting to assume that after <tt>synchronize_rcu()</tt> 995completes, there are no readers executing. 996This temptation must be avoided because 997new readers can start immediately after <tt>synchronize_rcu()</tt> 998starts, and <tt>synchronize_rcu()</tt> is under no 999obligation to wait for these new readers. 1000 1001<table> 1002<tr><th> </th></tr> 1003<tr><th align="left">Quick Quiz:</th></tr> 1004<tr><td> 1005 Suppose that synchronize_rcu() did wait until <i>all</i> 1006 readers had completed instead of waiting only on 1007 pre-existing readers. 1008 For how long would the updater be able to rely on there 1009 being no readers? 1010</td></tr> 1011<tr><th align="left">Answer:</th></tr> 1012<tr><td bgcolor="#ffffff"><font color="ffffff"> 1013 For no time at all. 1014 Even if <tt>synchronize_rcu()</tt> were to wait until 1015 all readers had completed, a new reader might start immediately after 1016 <tt>synchronize_rcu()</tt> completed. 1017 Therefore, the code following 1018 <tt>synchronize_rcu()</tt> can <i>never</i> rely on there being 1019 no readers. 1020</font></td></tr> 1021<tr><td> </td></tr> 1022</table> 1023 1024<h3><a name="Grace Periods Don't Partition Read-Side Critical Sections"> 1025Grace Periods Don't Partition Read-Side Critical Sections</a></h3> 1026 1027<p> 1028It is tempting to assume that if any part of one RCU read-side critical 1029section precedes a given grace period, and if any part of another RCU 1030read-side critical section follows that same grace period, then all of 1031the first RCU read-side critical section must precede all of the second. 1032However, this just isn't the case: A single grace period does not 1033partition the set of RCU read-side critical sections. 1034An example of this situation can be illustrated as follows, where 1035<tt>x</tt>, <tt>y</tt>, and <tt>z</tt> are initially all zero: 1036 1037<blockquote> 1038<pre> 1039 1 void thread0(void) 1040 2 { 1041 3 rcu_read_lock(); 1042 4 WRITE_ONCE(a, 1); 1043 5 WRITE_ONCE(b, 1); 1044 6 rcu_read_unlock(); 1045 7 } 1046 8 1047 9 void thread1(void) 104810 { 104911 r1 = READ_ONCE(a); 105012 synchronize_rcu(); 105113 WRITE_ONCE(c, 1); 105214 } 105315 105416 void thread2(void) 105517 { 105618 rcu_read_lock(); 105719 r2 = READ_ONCE(b); 105820 r3 = READ_ONCE(c); 105921 rcu_read_unlock(); 106022 } 1061</pre> 1062</blockquote> 1063 1064<p> 1065It turns out that the outcome: 1066 1067<blockquote> 1068<pre> 1069(r1 == 1 && r2 == 0 && r3 == 1) 1070</pre> 1071</blockquote> 1072 1073is entirely possible. 1074The following figure show how this can happen, with each circled 1075<tt>QS</tt> indicating the point at which RCU recorded a 1076<i>quiescent state</i> for each thread, that is, a state in which 1077RCU knows that the thread cannot be in the midst of an RCU read-side 1078critical section that started before the current grace period: 1079 1080<p><img src="GPpartitionReaders1.svg" alt="GPpartitionReaders1.svg" width="60%"></p> 1081 1082<p> 1083If it is necessary to partition RCU read-side critical sections in this 1084manner, it is necessary to use two grace periods, where the first 1085grace period is known to end before the second grace period starts: 1086 1087<blockquote> 1088<pre> 1089 1 void thread0(void) 1090 2 { 1091 3 rcu_read_lock(); 1092 4 WRITE_ONCE(a, 1); 1093 5 WRITE_ONCE(b, 1); 1094 6 rcu_read_unlock(); 1095 7 } 1096 8 1097 9 void thread1(void) 109810 { 109911 r1 = READ_ONCE(a); 110012 synchronize_rcu(); 110113 WRITE_ONCE(c, 1); 110214 } 110315 110416 void thread2(void) 110517 { 110618 r2 = READ_ONCE(c); 110719 synchronize_rcu(); 110820 WRITE_ONCE(d, 1); 110921 } 111022 111123 void thread3(void) 111224 { 111325 rcu_read_lock(); 111426 r3 = READ_ONCE(b); 111527 r4 = READ_ONCE(d); 111628 rcu_read_unlock(); 111729 } 1118</pre> 1119</blockquote> 1120 1121<p> 1122Here, if <tt>(r1 == 1)</tt>, then 1123<tt>thread0()</tt>'s write to <tt>b</tt> must happen 1124before the end of <tt>thread1()</tt>'s grace period. 1125If in addition <tt>(r4 == 1)</tt>, then 1126<tt>thread3()</tt>'s read from <tt>b</tt> must happen 1127after the beginning of <tt>thread2()</tt>'s grace period. 1128If it is also the case that <tt>(r2 == 1)</tt>, then the 1129end of <tt>thread1()</tt>'s grace period must precede the 1130beginning of <tt>thread2()</tt>'s grace period. 1131This mean that the two RCU read-side critical sections cannot overlap, 1132guaranteeing that <tt>(r3 == 1)</tt>. 1133As a result, the outcome: 1134 1135<blockquote> 1136<pre> 1137(r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1) 1138</pre> 1139</blockquote> 1140 1141cannot happen. 1142 1143<p> 1144This non-requirement was also non-premeditated, but became apparent 1145when studying RCU's interaction with memory ordering. 1146 1147<h3><a name="Read-Side Critical Sections Don't Partition Grace Periods"> 1148Read-Side Critical Sections Don't Partition Grace Periods</a></h3> 1149 1150<p> 1151It is also tempting to assume that if an RCU read-side critical section 1152happens between a pair of grace periods, then those grace periods cannot 1153overlap. 1154However, this temptation leads nowhere good, as can be illustrated by 1155the following, with all variables initially zero: 1156 1157<blockquote> 1158<pre> 1159 1 void thread0(void) 1160 2 { 1161 3 rcu_read_lock(); 1162 4 WRITE_ONCE(a, 1); 1163 5 WRITE_ONCE(b, 1); 1164 6 rcu_read_unlock(); 1165 7 } 1166 8 1167 9 void thread1(void) 116810 { 116911 r1 = READ_ONCE(a); 117012 synchronize_rcu(); 117113 WRITE_ONCE(c, 1); 117214 } 117315 117416 void thread2(void) 117517 { 117618 rcu_read_lock(); 117719 WRITE_ONCE(d, 1); 117820 r2 = READ_ONCE(c); 117921 rcu_read_unlock(); 118022 } 118123 118224 void thread3(void) 118325 { 118426 r3 = READ_ONCE(d); 118527 synchronize_rcu(); 118628 WRITE_ONCE(e, 1); 118729 } 118830 118931 void thread4(void) 119032 { 119133 rcu_read_lock(); 119234 r4 = READ_ONCE(b); 119335 r5 = READ_ONCE(e); 119436 rcu_read_unlock(); 119537 } 1196</pre> 1197</blockquote> 1198 1199<p> 1200In this case, the outcome: 1201 1202<blockquote> 1203<pre> 1204(r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1) 1205</pre> 1206</blockquote> 1207 1208is entirely possible, as illustrated below: 1209 1210<p><img src="ReadersPartitionGP1.svg" alt="ReadersPartitionGP1.svg" width="100%"></p> 1211 1212<p> 1213Again, an RCU read-side critical section can overlap almost all of a 1214given grace period, just so long as it does not overlap the entire 1215grace period. 1216As a result, an RCU read-side critical section cannot partition a pair 1217of RCU grace periods. 1218 1219<table> 1220<tr><th> </th></tr> 1221<tr><th align="left">Quick Quiz:</th></tr> 1222<tr><td> 1223 How long a sequence of grace periods, each separated by an RCU 1224 read-side critical section, would be required to partition the RCU 1225 read-side critical sections at the beginning and end of the chain? 1226</td></tr> 1227<tr><th align="left">Answer:</th></tr> 1228<tr><td bgcolor="#ffffff"><font color="ffffff"> 1229 In theory, an infinite number. 1230 In practice, an unknown number that is sensitive to both implementation 1231 details and timing considerations. 1232 Therefore, even in practice, RCU users must abide by the 1233 theoretical rather than the practical answer. 1234</font></td></tr> 1235<tr><td> </td></tr> 1236</table> 1237 1238<h3><a name="Disabling Preemption Does Not Block Grace Periods"> 1239Disabling Preemption Does Not Block Grace Periods</a></h3> 1240 1241<p> 1242There was a time when disabling preemption on any given CPU would block 1243subsequent grace periods. 1244However, this was an accident of implementation and is not a requirement. 1245And in the current Linux-kernel implementation, disabling preemption 1246on a given CPU in fact does not block grace periods, as Oleg Nesterov 1247<a href="https://lkml.kernel.org/g/20150614193825.GA19582@redhat.com">demonstrated</a>. 1248 1249<p> 1250If you need a preempt-disable region to block grace periods, you need to add 1251<tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>, for example 1252as follows: 1253 1254<blockquote> 1255<pre> 1256 1 preempt_disable(); 1257 2 rcu_read_lock(); 1258 3 do_something(); 1259 4 rcu_read_unlock(); 1260 5 preempt_enable(); 1261 6 1262 7 /* Spinlocks implicitly disable preemption. */ 1263 8 spin_lock(&mylock); 1264 9 rcu_read_lock(); 126510 do_something(); 126611 rcu_read_unlock(); 126712 spin_unlock(&mylock); 1268</pre> 1269</blockquote> 1270 1271<p> 1272In theory, you could enter the RCU read-side critical section first, 1273but it is more efficient to keep the entire RCU read-side critical 1274section contained in the preempt-disable region as shown above. 1275Of course, RCU read-side critical sections that extend outside of 1276preempt-disable regions will work correctly, but such critical sections 1277can be preempted, which forces <tt>rcu_read_unlock()</tt> to do 1278more work. 1279And no, this is <i>not</i> an invitation to enclose all of your RCU 1280read-side critical sections within preempt-disable regions, because 1281doing so would degrade real-time response. 1282 1283<p> 1284This non-requirement appeared with preemptible RCU. 1285If you need a grace period that waits on non-preemptible code regions, use 1286<a href="#Sched Flavor">RCU-sched</a>. 1287 1288<h2><a name="Parallelism Facts of Life">Parallelism Facts of Life</a></h2> 1289 1290<p> 1291These parallelism facts of life are by no means specific to RCU, but 1292the RCU implementation must abide by them. 1293They therefore bear repeating: 1294 1295<ol> 1296<li> Any CPU or task may be delayed at any time, 1297 and any attempts to avoid these delays by disabling 1298 preemption, interrupts, or whatever are completely futile. 1299 This is most obvious in preemptible user-level 1300 environments and in virtualized environments (where 1301 a given guest OS's VCPUs can be preempted at any time by 1302 the underlying hypervisor), but can also happen in bare-metal 1303 environments due to ECC errors, NMIs, and other hardware 1304 events. 1305 Although a delay of more than about 20 seconds can result 1306 in splats, the RCU implementation is obligated to use 1307 algorithms that can tolerate extremely long delays, but where 1308 “extremely long” is not long enough to allow 1309 wrap-around when incrementing a 64-bit counter. 1310<li> Both the compiler and the CPU can reorder memory accesses. 1311 Where it matters, RCU must use compiler directives and 1312 memory-barrier instructions to preserve ordering. 1313<li> Conflicting writes to memory locations in any given cache line 1314 will result in expensive cache misses. 1315 Greater numbers of concurrent writes and more-frequent 1316 concurrent writes will result in more dramatic slowdowns. 1317 RCU is therefore obligated to use algorithms that have 1318 sufficient locality to avoid significant performance and 1319 scalability problems. 1320<li> As a rough rule of thumb, only one CPU's worth of processing 1321 may be carried out under the protection of any given exclusive 1322 lock. 1323 RCU must therefore use scalable locking designs. 1324<li> Counters are finite, especially on 32-bit systems. 1325 RCU's use of counters must therefore tolerate counter wrap, 1326 or be designed such that counter wrap would take way more 1327 time than a single system is likely to run. 1328 An uptime of ten years is quite possible, a runtime 1329 of a century much less so. 1330 As an example of the latter, RCU's dyntick-idle nesting counter 1331 allows 54 bits for interrupt nesting level (this counter 1332 is 64 bits even on a 32-bit system). 1333 Overflowing this counter requires 2<sup>54</sup> 1334 half-interrupts on a given CPU without that CPU ever going idle. 1335 If a half-interrupt happened every microsecond, it would take 1336 570 years of runtime to overflow this counter, which is currently 1337 believed to be an acceptably long time. 1338<li> Linux systems can have thousands of CPUs running a single 1339 Linux kernel in a single shared-memory environment. 1340 RCU must therefore pay close attention to high-end scalability. 1341</ol> 1342 1343<p> 1344This last parallelism fact of life means that RCU must pay special 1345attention to the preceding facts of life. 1346The idea that Linux might scale to systems with thousands of CPUs would 1347have been met with some skepticism in the 1990s, but these requirements 1348would have otherwise have been unsurprising, even in the early 1990s. 1349 1350<h2><a name="Quality-of-Implementation Requirements">Quality-of-Implementation Requirements</a></h2> 1351 1352<p> 1353These sections list quality-of-implementation requirements. 1354Although an RCU implementation that ignores these requirements could 1355still be used, it would likely be subject to limitations that would 1356make it inappropriate for industrial-strength production use. 1357Classes of quality-of-implementation requirements are as follows: 1358 1359<ol> 1360<li> <a href="#Specialization">Specialization</a> 1361<li> <a href="#Performance and Scalability">Performance and Scalability</a> 1362<li> <a href="#Composability">Composability</a> 1363<li> <a href="#Corner Cases">Corner Cases</a> 1364</ol> 1365 1366<p> 1367These classes is covered in the following sections. 1368 1369<h3><a name="Specialization">Specialization</a></h3> 1370 1371<p> 1372RCU is and always has been intended primarily for read-mostly situations, 1373which means that RCU's read-side primitives are optimized, often at the 1374expense of its update-side primitives. 1375Experience thus far is captured by the following list of situations: 1376 1377<ol> 1378<li> Read-mostly data, where stale and inconsistent data is not 1379 a problem: RCU works great! 1380<li> Read-mostly data, where data must be consistent: 1381 RCU works well. 1382<li> Read-write data, where data must be consistent: 1383 RCU <i>might</i> work OK. 1384 Or not. 1385<li> Write-mostly data, where data must be consistent: 1386 RCU is very unlikely to be the right tool for the job, 1387 with the following exceptions, where RCU can provide: 1388 <ol type=a> 1389 <li> Existence guarantees for update-friendly mechanisms. 1390 <li> Wait-free read-side primitives for real-time use. 1391 </ol> 1392</ol> 1393 1394<p> 1395This focus on read-mostly situations means that RCU must interoperate 1396with other synchronization primitives. 1397For example, the <tt>add_gp()</tt> and <tt>remove_gp_synchronous()</tt> 1398examples discussed earlier use RCU to protect readers and locking to 1399coordinate updaters. 1400However, the need extends much farther, requiring that a variety of 1401synchronization primitives be legal within RCU read-side critical sections, 1402including spinlocks, sequence locks, atomic operations, reference 1403counters, and memory barriers. 1404 1405<table> 1406<tr><th> </th></tr> 1407<tr><th align="left">Quick Quiz:</th></tr> 1408<tr><td> 1409 What about sleeping locks? 1410</td></tr> 1411<tr><th align="left">Answer:</th></tr> 1412<tr><td bgcolor="#ffffff"><font color="ffffff"> 1413 These are forbidden within Linux-kernel RCU read-side critical 1414 sections because it is not legal to place a quiescent state 1415 (in this case, voluntary context switch) within an RCU read-side 1416 critical section. 1417 However, sleeping locks may be used within userspace RCU read-side 1418 critical sections, and also within Linux-kernel sleepable RCU 1419 <a href="#Sleepable RCU"><font color="ffffff">(SRCU)</font></a> 1420 read-side critical sections. 1421 In addition, the -rt patchset turns spinlocks into a 1422 sleeping locks so that the corresponding critical sections 1423 can be preempted, which also means that these sleeplockified 1424 spinlocks (but not other sleeping locks!) may be acquire within 1425 -rt-Linux-kernel RCU read-side critical sections. 1426 </font> 1427 1428 <p><font color="ffffff"> 1429 Note that it <i>is</i> legal for a normal RCU read-side 1430 critical section to conditionally acquire a sleeping locks 1431 (as in <tt>mutex_trylock()</tt>), but only as long as it does 1432 not loop indefinitely attempting to conditionally acquire that 1433 sleeping locks. 1434 The key point is that things like <tt>mutex_trylock()</tt> 1435 either return with the mutex held, or return an error indication if 1436 the mutex was not immediately available. 1437 Either way, <tt>mutex_trylock()</tt> returns immediately without 1438 sleeping. 1439</font></td></tr> 1440<tr><td> </td></tr> 1441</table> 1442 1443<p> 1444It often comes as a surprise that many algorithms do not require a 1445consistent view of data, but many can function in that mode, 1446with network routing being the poster child. 1447Internet routing algorithms take significant time to propagate 1448updates, so that by the time an update arrives at a given system, 1449that system has been sending network traffic the wrong way for 1450a considerable length of time. 1451Having a few threads continue to send traffic the wrong way for a 1452few more milliseconds is clearly not a problem: In the worst case, 1453TCP retransmissions will eventually get the data where it needs to go. 1454In general, when tracking the state of the universe outside of the 1455computer, some level of inconsistency must be tolerated due to 1456speed-of-light delays if nothing else. 1457 1458<p> 1459Furthermore, uncertainty about external state is inherent in many cases. 1460For example, a pair of veternarians might use heartbeat to determine 1461whether or not a given cat was alive. 1462But how long should they wait after the last heartbeat to decide that 1463the cat is in fact dead? 1464Waiting less than 400 milliseconds makes no sense because this would 1465mean that a relaxed cat would be considered to cycle between death 1466and life more than 100 times per minute. 1467Moreover, just as with human beings, a cat's heart might stop for 1468some period of time, so the exact wait period is a judgment call. 1469One of our pair of veternarians might wait 30 seconds before pronouncing 1470the cat dead, while the other might insist on waiting a full minute. 1471The two veternarians would then disagree on the state of the cat during 1472the final 30 seconds of the minute following the last heartbeat. 1473 1474<p> 1475Interestingly enough, this same situation applies to hardware. 1476When push comes to shove, how do we tell whether or not some 1477external server has failed? 1478We send messages to it periodically, and declare it failed if we 1479don't receive a response within a given period of time. 1480Policy decisions can usually tolerate short 1481periods of inconsistency. 1482The policy was decided some time ago, and is only now being put into 1483effect, so a few milliseconds of delay is normally inconsequential. 1484 1485<p> 1486However, there are algorithms that absolutely must see consistent data. 1487For example, the translation between a user-level SystemV semaphore 1488ID to the corresponding in-kernel data structure is protected by RCU, 1489but it is absolutely forbidden to update a semaphore that has just been 1490removed. 1491In the Linux kernel, this need for consistency is accommodated by acquiring 1492spinlocks located in the in-kernel data structure from within 1493the RCU read-side critical section, and this is indicated by the 1494green box in the figure above. 1495Many other techniques may be used, and are in fact used within the 1496Linux kernel. 1497 1498<p> 1499In short, RCU is not required to maintain consistency, and other 1500mechanisms may be used in concert with RCU when consistency is required. 1501RCU's specialization allows it to do its job extremely well, and its 1502ability to interoperate with other synchronization mechanisms allows 1503the right mix of synchronization tools to be used for a given job. 1504 1505<h3><a name="Performance and Scalability">Performance and Scalability</a></h3> 1506 1507<p> 1508Energy efficiency is a critical component of performance today, 1509and Linux-kernel RCU implementations must therefore avoid unnecessarily 1510awakening idle CPUs. 1511I cannot claim that this requirement was premeditated. 1512In fact, I learned of it during a telephone conversation in which I 1513was given “frank and open” feedback on the importance 1514of energy efficiency in battery-powered systems and on specific 1515energy-efficiency shortcomings of the Linux-kernel RCU implementation. 1516In my experience, the battery-powered embedded community will consider 1517any unnecessary wakeups to be extremely unfriendly acts. 1518So much so that mere Linux-kernel-mailing-list posts are 1519insufficient to vent their ire. 1520 1521<p> 1522Memory consumption is not particularly important for in most 1523situations, and has become decreasingly 1524so as memory sizes have expanded and memory 1525costs have plummeted. 1526However, as I learned from Matt Mackall's 1527<a href="http://elinux.org/Linux_Tiny-FAQ">bloatwatch</a> 1528efforts, memory footprint is critically important on single-CPU systems with 1529non-preemptible (<tt>CONFIG_PREEMPT=n</tt>) kernels, and thus 1530<a href="https://lkml.kernel.org/g/20090113221724.GA15307@linux.vnet.ibm.com">tiny RCU</a> 1531was born. 1532Josh Triplett has since taken over the small-memory banner with his 1533<a href="https://tiny.wiki.kernel.org/">Linux kernel tinification</a> 1534project, which resulted in 1535<a href="#Sleepable RCU">SRCU</a> 1536becoming optional for those kernels not needing it. 1537 1538<p> 1539The remaining performance requirements are, for the most part, 1540unsurprising. 1541For example, in keeping with RCU's read-side specialization, 1542<tt>rcu_dereference()</tt> should have negligible overhead (for 1543example, suppression of a few minor compiler optimizations). 1544Similarly, in non-preemptible environments, <tt>rcu_read_lock()</tt> and 1545<tt>rcu_read_unlock()</tt> should have exactly zero overhead. 1546 1547<p> 1548In preemptible environments, in the case where the RCU read-side 1549critical section was not preempted (as will be the case for the 1550highest-priority real-time process), <tt>rcu_read_lock()</tt> and 1551<tt>rcu_read_unlock()</tt> should have minimal overhead. 1552In particular, they should not contain atomic read-modify-write 1553operations, memory-barrier instructions, preemption disabling, 1554interrupt disabling, or backwards branches. 1555However, in the case where the RCU read-side critical section was preempted, 1556<tt>rcu_read_unlock()</tt> may acquire spinlocks and disable interrupts. 1557This is why it is better to nest an RCU read-side critical section 1558within a preempt-disable region than vice versa, at least in cases 1559where that critical section is short enough to avoid unduly degrading 1560real-time latencies. 1561 1562<p> 1563The <tt>synchronize_rcu()</tt> grace-period-wait primitive is 1564optimized for throughput. 1565It may therefore incur several milliseconds of latency in addition to 1566the duration of the longest RCU read-side critical section. 1567On the other hand, multiple concurrent invocations of 1568<tt>synchronize_rcu()</tt> are required to use batching optimizations 1569so that they can be satisfied by a single underlying grace-period-wait 1570operation. 1571For example, in the Linux kernel, it is not unusual for a single 1572grace-period-wait operation to serve more than 1573<a href="https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response">1,000 separate invocations</a> 1574of <tt>synchronize_rcu()</tt>, thus amortizing the per-invocation 1575overhead down to nearly zero. 1576However, the grace-period optimization is also required to avoid 1577measurable degradation of real-time scheduling and interrupt latencies. 1578 1579<p> 1580In some cases, the multi-millisecond <tt>synchronize_rcu()</tt> 1581latencies are unacceptable. 1582In these cases, <tt>synchronize_rcu_expedited()</tt> may be used 1583instead, reducing the grace-period latency down to a few tens of 1584microseconds on small systems, at least in cases where the RCU read-side 1585critical sections are short. 1586There are currently no special latency requirements for 1587<tt>synchronize_rcu_expedited()</tt> on large systems, but, 1588consistent with the empirical nature of the RCU specification, 1589that is subject to change. 1590However, there most definitely are scalability requirements: 1591A storm of <tt>synchronize_rcu_expedited()</tt> invocations on 4096 1592CPUs should at least make reasonable forward progress. 1593In return for its shorter latencies, <tt>synchronize_rcu_expedited()</tt> 1594is permitted to impose modest degradation of real-time latency 1595on non-idle online CPUs. 1596That said, it will likely be necessary to take further steps to reduce this 1597degradation, hopefully to roughly that of a scheduling-clock interrupt. 1598 1599<p> 1600There are a number of situations where even 1601<tt>synchronize_rcu_expedited()</tt>'s reduced grace-period 1602latency is unacceptable. 1603In these situations, the asynchronous <tt>call_rcu()</tt> can be 1604used in place of <tt>synchronize_rcu()</tt> as follows: 1605 1606<blockquote> 1607<pre> 1608 1 struct foo { 1609 2 int a; 1610 3 int b; 1611 4 struct rcu_head rh; 1612 5 }; 1613 6 1614 7 static void remove_gp_cb(struct rcu_head *rhp) 1615 8 { 1616 9 struct foo *p = container_of(rhp, struct foo, rh); 161710 161811 kfree(p); 161912 } 162013 162114 bool remove_gp_asynchronous(void) 162215 { 162316 struct foo *p; 162417 162518 spin_lock(&gp_lock); 162619 p = rcu_dereference(gp); 162720 if (!p) { 162821 spin_unlock(&gp_lock); 162922 return false; 163023 } 163124 rcu_assign_pointer(gp, NULL); 163225 call_rcu(&p->rh, remove_gp_cb); 163326 spin_unlock(&gp_lock); 163427 return true; 163528 } 1636</pre> 1637</blockquote> 1638 1639<p> 1640A definition of <tt>struct foo</tt> is finally needed, and appears 1641on lines 1-5. 1642The function <tt>remove_gp_cb()</tt> is passed to <tt>call_rcu()</tt> 1643on line 25, and will be invoked after the end of a subsequent 1644grace period. 1645This gets the same effect as <tt>remove_gp_synchronous()</tt>, 1646but without forcing the updater to wait for a grace period to elapse. 1647The <tt>call_rcu()</tt> function may be used in a number of 1648situations where neither <tt>synchronize_rcu()</tt> nor 1649<tt>synchronize_rcu_expedited()</tt> would be legal, 1650including within preempt-disable code, <tt>local_bh_disable()</tt> code, 1651interrupt-disable code, and interrupt handlers. 1652However, even <tt>call_rcu()</tt> is illegal within NMI handlers 1653and from idle and offline CPUs. 1654The callback function (<tt>remove_gp_cb()</tt> in this case) will be 1655executed within softirq (software interrupt) environment within the 1656Linux kernel, 1657either within a real softirq handler or under the protection 1658of <tt>local_bh_disable()</tt>. 1659In both the Linux kernel and in userspace, it is bad practice to 1660write an RCU callback function that takes too long. 1661Long-running operations should be relegated to separate threads or 1662(in the Linux kernel) workqueues. 1663 1664<table> 1665<tr><th> </th></tr> 1666<tr><th align="left">Quick Quiz:</th></tr> 1667<tr><td> 1668 Why does line 19 use <tt>rcu_access_pointer()</tt>? 1669 After all, <tt>call_rcu()</tt> on line 25 stores into the 1670 structure, which would interact badly with concurrent insertions. 1671 Doesn't this mean that <tt>rcu_dereference()</tt> is required? 1672</td></tr> 1673<tr><th align="left">Answer:</th></tr> 1674<tr><td bgcolor="#ffffff"><font color="ffffff"> 1675 Presumably the <tt>->gp_lock</tt> acquired on line 18 excludes 1676 any changes, including any insertions that <tt>rcu_dereference()</tt> 1677 would protect against. 1678 Therefore, any insertions will be delayed until after 1679 <tt>->gp_lock</tt> 1680 is released on line 25, which in turn means that 1681 <tt>rcu_access_pointer()</tt> suffices. 1682</font></td></tr> 1683<tr><td> </td></tr> 1684</table> 1685 1686<p> 1687However, all that <tt>remove_gp_cb()</tt> is doing is 1688invoking <tt>kfree()</tt> on the data element. 1689This is a common idiom, and is supported by <tt>kfree_rcu()</tt>, 1690which allows “fire and forget” operation as shown below: 1691 1692<blockquote> 1693<pre> 1694 1 struct foo { 1695 2 int a; 1696 3 int b; 1697 4 struct rcu_head rh; 1698 5 }; 1699 6 1700 7 bool remove_gp_faf(void) 1701 8 { 1702 9 struct foo *p; 170310 170411 spin_lock(&gp_lock); 170512 p = rcu_dereference(gp); 170613 if (!p) { 170714 spin_unlock(&gp_lock); 170815 return false; 170916 } 171017 rcu_assign_pointer(gp, NULL); 171118 kfree_rcu(p, rh); 171219 spin_unlock(&gp_lock); 171320 return true; 171421 } 1715</pre> 1716</blockquote> 1717 1718<p> 1719Note that <tt>remove_gp_faf()</tt> simply invokes 1720<tt>kfree_rcu()</tt> and proceeds, without any need to pay any 1721further attention to the subsequent grace period and <tt>kfree()</tt>. 1722It is permissible to invoke <tt>kfree_rcu()</tt> from the same 1723environments as for <tt>call_rcu()</tt>. 1724Interestingly enough, DYNIX/ptx had the equivalents of 1725<tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>, but not 1726<tt>synchronize_rcu()</tt>. 1727This was due to the fact that RCU was not heavily used within DYNIX/ptx, 1728so the very few places that needed something like 1729<tt>synchronize_rcu()</tt> simply open-coded it. 1730 1731<table> 1732<tr><th> </th></tr> 1733<tr><th align="left">Quick Quiz:</th></tr> 1734<tr><td> 1735 Earlier it was claimed that <tt>call_rcu()</tt> and 1736 <tt>kfree_rcu()</tt> allowed updaters to avoid being blocked 1737 by readers. 1738 But how can that be correct, given that the invocation of the callback 1739 and the freeing of the memory (respectively) must still wait for 1740 a grace period to elapse? 1741</td></tr> 1742<tr><th align="left">Answer:</th></tr> 1743<tr><td bgcolor="#ffffff"><font color="ffffff"> 1744 We could define things this way, but keep in mind that this sort of 1745 definition would say that updates in garbage-collected languages 1746 cannot complete until the next time the garbage collector runs, 1747 which does not seem at all reasonable. 1748 The key point is that in most cases, an updater using either 1749 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> can proceed to the 1750 next update as soon as it has invoked <tt>call_rcu()</tt> or 1751 <tt>kfree_rcu()</tt>, without having to wait for a subsequent 1752 grace period. 1753</font></td></tr> 1754<tr><td> </td></tr> 1755</table> 1756 1757<p> 1758But what if the updater must wait for the completion of code to be 1759executed after the end of the grace period, but has other tasks 1760that can be carried out in the meantime? 1761The polling-style <tt>get_state_synchronize_rcu()</tt> and 1762<tt>cond_synchronize_rcu()</tt> functions may be used for this 1763purpose, as shown below: 1764 1765<blockquote> 1766<pre> 1767 1 bool remove_gp_poll(void) 1768 2 { 1769 3 struct foo *p; 1770 4 unsigned long s; 1771 5 1772 6 spin_lock(&gp_lock); 1773 7 p = rcu_access_pointer(gp); 1774 8 if (!p) { 1775 9 spin_unlock(&gp_lock); 177610 return false; 177711 } 177812 rcu_assign_pointer(gp, NULL); 177913 spin_unlock(&gp_lock); 178014 s = get_state_synchronize_rcu(); 178115 do_something_while_waiting(); 178216 cond_synchronize_rcu(s); 178317 kfree(p); 178418 return true; 178519 } 1786</pre> 1787</blockquote> 1788 1789<p> 1790On line 14, <tt>get_state_synchronize_rcu()</tt> obtains a 1791“cookie” from RCU, 1792then line 15 carries out other tasks, 1793and finally, line 16 returns immediately if a grace period has 1794elapsed in the meantime, but otherwise waits as required. 1795The need for <tt>get_state_synchronize_rcu</tt> and 1796<tt>cond_synchronize_rcu()</tt> has appeared quite recently, 1797so it is too early to tell whether they will stand the test of time. 1798 1799<p> 1800RCU thus provides a range of tools to allow updaters to strike the 1801required tradeoff between latency, flexibility and CPU overhead. 1802 1803<h3><a name="Composability">Composability</a></h3> 1804 1805<p> 1806Composability has received much attention in recent years, perhaps in part 1807due to the collision of multicore hardware with object-oriented techniques 1808designed in single-threaded environments for single-threaded use. 1809And in theory, RCU read-side critical sections may be composed, and in 1810fact may be nested arbitrarily deeply. 1811In practice, as with all real-world implementations of composable 1812constructs, there are limitations. 1813 1814<p> 1815Implementations of RCU for which <tt>rcu_read_lock()</tt> 1816and <tt>rcu_read_unlock()</tt> generate no code, such as 1817Linux-kernel RCU when <tt>CONFIG_PREEMPT=n</tt>, can be 1818nested arbitrarily deeply. 1819After all, there is no overhead. 1820Except that if all these instances of <tt>rcu_read_lock()</tt> 1821and <tt>rcu_read_unlock()</tt> are visible to the compiler, 1822compilation will eventually fail due to exhausting memory, 1823mass storage, or user patience, whichever comes first. 1824If the nesting is not visible to the compiler, as is the case with 1825mutually recursive functions each in its own translation unit, 1826stack overflow will result. 1827If the nesting takes the form of loops, either the control variable 1828will overflow or (in the Linux kernel) you will get an RCU CPU stall warning. 1829Nevertheless, this class of RCU implementations is one 1830of the most composable constructs in existence. 1831 1832<p> 1833RCU implementations that explicitly track nesting depth 1834are limited by the nesting-depth counter. 1835For example, the Linux kernel's preemptible RCU limits nesting to 1836<tt>INT_MAX</tt>. 1837This should suffice for almost all practical purposes. 1838That said, a consecutive pair of RCU read-side critical sections 1839between which there is an operation that waits for a grace period 1840cannot be enclosed in another RCU read-side critical section. 1841This is because it is not legal to wait for a grace period within 1842an RCU read-side critical section: To do so would result either 1843in deadlock or 1844in RCU implicitly splitting the enclosing RCU read-side critical 1845section, neither of which is conducive to a long-lived and prosperous 1846kernel. 1847 1848<p> 1849It is worth noting that RCU is not alone in limiting composability. 1850For example, many transactional-memory implementations prohibit 1851composing a pair of transactions separated by an irrevocable 1852operation (for example, a network receive operation). 1853For another example, lock-based critical sections can be composed 1854surprisingly freely, but only if deadlock is avoided. 1855 1856<p> 1857In short, although RCU read-side critical sections are highly composable, 1858care is required in some situations, just as is the case for any other 1859composable synchronization mechanism. 1860 1861<h3><a name="Corner Cases">Corner Cases</a></h3> 1862 1863<p> 1864A given RCU workload might have an endless and intense stream of 1865RCU read-side critical sections, perhaps even so intense that there 1866was never a point in time during which there was not at least one 1867RCU read-side critical section in flight. 1868RCU cannot allow this situation to block grace periods: As long as 1869all the RCU read-side critical sections are finite, grace periods 1870must also be finite. 1871 1872<p> 1873That said, preemptible RCU implementations could potentially result 1874in RCU read-side critical sections being preempted for long durations, 1875which has the effect of creating a long-duration RCU read-side 1876critical section. 1877This situation can arise only in heavily loaded systems, but systems using 1878real-time priorities are of course more vulnerable. 1879Therefore, RCU priority boosting is provided to help deal with this 1880case. 1881That said, the exact requirements on RCU priority boosting will likely 1882evolve as more experience accumulates. 1883 1884<p> 1885Other workloads might have very high update rates. 1886Although one can argue that such workloads should instead use 1887something other than RCU, the fact remains that RCU must 1888handle such workloads gracefully. 1889This requirement is another factor driving batching of grace periods, 1890but it is also the driving force behind the checks for large numbers 1891of queued RCU callbacks in the <tt>call_rcu()</tt> code path. 1892Finally, high update rates should not delay RCU read-side critical 1893sections, although some read-side delays can occur when using 1894<tt>synchronize_rcu_expedited()</tt>, courtesy of this function's use 1895of <tt>try_stop_cpus()</tt>. 1896(In the future, <tt>synchronize_rcu_expedited()</tt> will be 1897converted to use lighter-weight inter-processor interrupts (IPIs), 1898but this will still disturb readers, though to a much smaller degree.) 1899 1900<p> 1901Although all three of these corner cases were understood in the early 19021990s, a simple user-level test consisting of <tt>close(open(path))</tt> 1903in a tight loop 1904in the early 2000s suddenly provided a much deeper appreciation of the 1905high-update-rate corner case. 1906This test also motivated addition of some RCU code to react to high update 1907rates, for example, if a given CPU finds itself with more than 10,000 1908RCU callbacks queued, it will cause RCU to take evasive action by 1909more aggressively starting grace periods and more aggressively forcing 1910completion of grace-period processing. 1911This evasive action causes the grace period to complete more quickly, 1912but at the cost of restricting RCU's batching optimizations, thus 1913increasing the CPU overhead incurred by that grace period. 1914 1915<h2><a name="Software-Engineering Requirements"> 1916Software-Engineering Requirements</a></h2> 1917 1918<p> 1919Between Murphy's Law and “To err is human”, it is necessary to 1920guard against mishaps and misuse: 1921 1922<ol> 1923<li> It is all too easy to forget to use <tt>rcu_read_lock()</tt> 1924 everywhere that it is needed, so kernels built with 1925 <tt>CONFIG_PROVE_RCU=y</tt> will spat if 1926 <tt>rcu_dereference()</tt> is used outside of an 1927 RCU read-side critical section. 1928 Update-side code can use <tt>rcu_dereference_protected()</tt>, 1929 which takes a 1930 <a href="https://lwn.net/Articles/371986/">lockdep expression</a> 1931 to indicate what is providing the protection. 1932 If the indicated protection is not provided, a lockdep splat 1933 is emitted. 1934 1935 <p> 1936 Code shared between readers and updaters can use 1937 <tt>rcu_dereference_check()</tt>, which also takes a 1938 lockdep expression, and emits a lockdep splat if neither 1939 <tt>rcu_read_lock()</tt> nor the indicated protection 1940 is in place. 1941 In addition, <tt>rcu_dereference_raw()</tt> is used in those 1942 (hopefully rare) cases where the required protection cannot 1943 be easily described. 1944 Finally, <tt>rcu_read_lock_held()</tt> is provided to 1945 allow a function to verify that it has been invoked within 1946 an RCU read-side critical section. 1947 I was made aware of this set of requirements shortly after Thomas 1948 Gleixner audited a number of RCU uses. 1949<li> A given function might wish to check for RCU-related preconditions 1950 upon entry, before using any other RCU API. 1951 The <tt>rcu_lockdep_assert()</tt> does this job, 1952 asserting the expression in kernels having lockdep enabled 1953 and doing nothing otherwise. 1954<li> It is also easy to forget to use <tt>rcu_assign_pointer()</tt> 1955 and <tt>rcu_dereference()</tt>, perhaps (incorrectly) 1956 substituting a simple assignment. 1957 To catch this sort of error, a given RCU-protected pointer may be 1958 tagged with <tt>__rcu</tt>, after which running sparse 1959 with <tt>CONFIG_SPARSE_RCU_POINTER=y</tt> will complain 1960 about simple-assignment accesses to that pointer. 1961 Arnd Bergmann made me aware of this requirement, and also 1962 supplied the needed 1963 <a href="https://lwn.net/Articles/376011/">patch series</a>. 1964<li> Kernels built with <tt>CONFIG_DEBUG_OBJECTS_RCU_HEAD=y</tt> 1965 will splat if a data element is passed to <tt>call_rcu()</tt> 1966 twice in a row, without a grace period in between. 1967 (This error is similar to a double free.) 1968 The corresponding <tt>rcu_head</tt> structures that are 1969 dynamically allocated are automatically tracked, but 1970 <tt>rcu_head</tt> structures allocated on the stack 1971 must be initialized with <tt>init_rcu_head_on_stack()</tt> 1972 and cleaned up with <tt>destroy_rcu_head_on_stack()</tt>. 1973 Similarly, statically allocated non-stack <tt>rcu_head</tt> 1974 structures must be initialized with <tt>init_rcu_head()</tt> 1975 and cleaned up with <tt>destroy_rcu_head()</tt>. 1976 Mathieu Desnoyers made me aware of this requirement, and also 1977 supplied the needed 1978 <a href="https://lkml.kernel.org/g/20100319013024.GA28456@Krystal">patch</a>. 1979<li> An infinite loop in an RCU read-side critical section will 1980 eventually trigger an RCU CPU stall warning splat, with 1981 the duration of “eventually” being controlled by the 1982 <tt>RCU_CPU_STALL_TIMEOUT</tt> <tt>Kconfig</tt> option, or, 1983 alternatively, by the 1984 <tt>rcupdate.rcu_cpu_stall_timeout</tt> boot/sysfs 1985 parameter. 1986 However, RCU is not obligated to produce this splat 1987 unless there is a grace period waiting on that particular 1988 RCU read-side critical section. 1989 <p> 1990 Some extreme workloads might intentionally delay 1991 RCU grace periods, and systems running those workloads can 1992 be booted with <tt>rcupdate.rcu_cpu_stall_suppress</tt> 1993 to suppress the splats. 1994 This kernel parameter may also be set via <tt>sysfs</tt>. 1995 Furthermore, RCU CPU stall warnings are counter-productive 1996 during sysrq dumps and during panics. 1997 RCU therefore supplies the <tt>rcu_sysrq_start()</tt> and 1998 <tt>rcu_sysrq_end()</tt> API members to be called before 1999 and after long sysrq dumps. 2000 RCU also supplies the <tt>rcu_panic()</tt> notifier that is 2001 automatically invoked at the beginning of a panic to suppress 2002 further RCU CPU stall warnings. 2003 2004 <p> 2005 This requirement made itself known in the early 1990s, pretty 2006 much the first time that it was necessary to debug a CPU stall. 2007 That said, the initial implementation in DYNIX/ptx was quite 2008 generic in comparison with that of Linux. 2009<li> Although it would be very good to detect pointers leaking out 2010 of RCU read-side critical sections, there is currently no 2011 good way of doing this. 2012 One complication is the need to distinguish between pointers 2013 leaking and pointers that have been handed off from RCU to 2014 some other synchronization mechanism, for example, reference 2015 counting. 2016<li> In kernels built with <tt>CONFIG_RCU_TRACE=y</tt>, RCU-related 2017 information is provided via both debugfs and event tracing. 2018<li> Open-coded use of <tt>rcu_assign_pointer()</tt> and 2019 <tt>rcu_dereference()</tt> to create typical linked 2020 data structures can be surprisingly error-prone. 2021 Therefore, RCU-protected 2022 <a href="https://lwn.net/Articles/609973/#RCU List APIs">linked lists</a> 2023 and, more recently, RCU-protected 2024 <a href="https://lwn.net/Articles/612100/">hash tables</a> 2025 are available. 2026 Many other special-purpose RCU-protected data structures are 2027 available in the Linux kernel and the userspace RCU library. 2028<li> Some linked structures are created at compile time, but still 2029 require <tt>__rcu</tt> checking. 2030 The <tt>RCU_POINTER_INITIALIZER()</tt> macro serves this 2031 purpose. 2032<li> It is not necessary to use <tt>rcu_assign_pointer()</tt> 2033 when creating linked structures that are to be published via 2034 a single external pointer. 2035 The <tt>RCU_INIT_POINTER()</tt> macro is provided for 2036 this task and also for assigning <tt>NULL</tt> pointers 2037 at runtime. 2038</ol> 2039 2040<p> 2041This not a hard-and-fast list: RCU's diagnostic capabilities will 2042continue to be guided by the number and type of usage bugs found 2043in real-world RCU usage. 2044 2045<h2><a name="Linux Kernel Complications">Linux Kernel Complications</a></h2> 2046 2047<p> 2048The Linux kernel provides an interesting environment for all kinds of 2049software, including RCU. 2050Some of the relevant points of interest are as follows: 2051 2052<ol> 2053<li> <a href="#Configuration">Configuration</a>. 2054<li> <a href="#Firmware Interface">Firmware Interface</a>. 2055<li> <a href="#Early Boot">Early Boot</a>. 2056<li> <a href="#Interrupts and NMIs"> 2057 Interrupts and non-maskable interrupts (NMIs)</a>. 2058<li> <a href="#Loadable Modules">Loadable Modules</a>. 2059<li> <a href="#Hotplug CPU">Hotplug CPU</a>. 2060<li> <a href="#Scheduler and RCU">Scheduler and RCU</a>. 2061<li> <a href="#Tracing and RCU">Tracing and RCU</a>. 2062<li> <a href="#Energy Efficiency">Energy Efficiency</a>. 2063<li> <a href="#Memory Efficiency">Memory Efficiency</a>. 2064<li> <a href="#Performance, Scalability, Response Time, and Reliability"> 2065 Performance, Scalability, Response Time, and Reliability</a>. 2066</ol> 2067 2068<p> 2069This list is probably incomplete, but it does give a feel for the 2070most notable Linux-kernel complications. 2071Each of the following sections covers one of the above topics. 2072 2073<h3><a name="Configuration">Configuration</a></h3> 2074 2075<p> 2076RCU's goal is automatic configuration, so that almost nobody 2077needs to worry about RCU's <tt>Kconfig</tt> options. 2078And for almost all users, RCU does in fact work well 2079“out of the box.” 2080 2081<p> 2082However, there are specialized use cases that are handled by 2083kernel boot parameters and <tt>Kconfig</tt> options. 2084Unfortunately, the <tt>Kconfig</tt> system will explicitly ask users 2085about new <tt>Kconfig</tt> options, which requires almost all of them 2086be hidden behind a <tt>CONFIG_RCU_EXPERT</tt> <tt>Kconfig</tt> option. 2087 2088<p> 2089This all should be quite obvious, but the fact remains that 2090Linus Torvalds recently had to 2091<a href="https://lkml.kernel.org/g/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com">remind</a> 2092me of this requirement. 2093 2094<h3><a name="Firmware Interface">Firmware Interface</a></h3> 2095 2096<p> 2097In many cases, kernel obtains information about the system from the 2098firmware, and sometimes things are lost in translation. 2099Or the translation is accurate, but the original message is bogus. 2100 2101<p> 2102For example, some systems' firmware overreports the number of CPUs, 2103sometimes by a large factor. 2104If RCU naively believed the firmware, as it used to do, 2105it would create too many per-CPU kthreads. 2106Although the resulting system will still run correctly, the extra 2107kthreads needlessly consume memory and can cause confusion 2108when they show up in <tt>ps</tt> listings. 2109 2110<p> 2111RCU must therefore wait for a given CPU to actually come online before 2112it can allow itself to believe that the CPU actually exists. 2113The resulting “ghost CPUs” (which are never going to 2114come online) cause a number of 2115<a href="https://paulmck.livejournal.com/37494.html">interesting complications</a>. 2116 2117<h3><a name="Early Boot">Early Boot</a></h3> 2118 2119<p> 2120The Linux kernel's boot sequence is an interesting process, 2121and RCU is used early, even before <tt>rcu_init()</tt> 2122is invoked. 2123In fact, a number of RCU's primitives can be used as soon as the 2124initial task's <tt>task_struct</tt> is available and the 2125boot CPU's per-CPU variables are set up. 2126The read-side primitives (<tt>rcu_read_lock()</tt>, 2127<tt>rcu_read_unlock()</tt>, <tt>rcu_dereference()</tt>, 2128and <tt>rcu_access_pointer()</tt>) will operate normally very early on, 2129as will <tt>rcu_assign_pointer()</tt>. 2130 2131<p> 2132Although <tt>call_rcu()</tt> may be invoked at any 2133time during boot, callbacks are not guaranteed to be invoked until after 2134the scheduler is fully up and running. 2135This delay in callback invocation is due to the fact that RCU does not 2136invoke callbacks until it is fully initialized, and this full initialization 2137cannot occur until after the scheduler has initialized itself to the 2138point where RCU can spawn and run its kthreads. 2139In theory, it would be possible to invoke callbacks earlier, 2140however, this is not a panacea because there would be severe restrictions 2141on what operations those callbacks could invoke. 2142 2143<p> 2144Perhaps surprisingly, <tt>synchronize_rcu()</tt>, 2145<a href="#Bottom-Half Flavor"><tt>synchronize_rcu_bh()</tt></a> 2146(<a href="#Bottom-Half Flavor">discussed below</a>), 2147and 2148<a href="#Sched Flavor"><tt>synchronize_sched()</tt></a> 2149will all operate normally 2150during very early boot, the reason being that there is only one CPU 2151and preemption is disabled. 2152This means that the call <tt>synchronize_rcu()</tt> (or friends) 2153itself is a quiescent 2154state and thus a grace period, so the early-boot implementation can 2155be a no-op. 2156 2157<p> 2158Both <tt>synchronize_rcu_bh()</tt> and <tt>synchronize_sched()</tt> 2159continue to operate normally through the remainder of boot, courtesy 2160of the fact that preemption is disabled across their RCU read-side 2161critical sections and also courtesy of the fact that there is still 2162only one CPU. 2163However, once the scheduler starts initializing, preemption is enabled. 2164There is still only a single CPU, but the fact that preemption is enabled 2165means that the no-op implementation of <tt>synchronize_rcu()</tt> no 2166longer works in <tt>CONFIG_PREEMPT=y</tt> kernels. 2167Therefore, as soon as the scheduler starts initializing, the early-boot 2168fastpath is disabled. 2169This means that <tt>synchronize_rcu()</tt> switches to its runtime 2170mode of operation where it posts callbacks, which in turn means that 2171any call to <tt>synchronize_rcu()</tt> will block until the corresponding 2172callback is invoked. 2173Unfortunately, the callback cannot be invoked until RCU's runtime 2174grace-period machinery is up and running, which cannot happen until 2175the scheduler has initialized itself sufficiently to allow RCU's 2176kthreads to be spawned. 2177Therefore, invoking <tt>synchronize_rcu()</tt> during scheduler 2178initialization can result in deadlock. 2179 2180<table> 2181<tr><th> </th></tr> 2182<tr><th align="left">Quick Quiz:</th></tr> 2183<tr><td> 2184 So what happens with <tt>synchronize_rcu()</tt> during 2185 scheduler initialization for <tt>CONFIG_PREEMPT=n</tt> 2186 kernels? 2187</td></tr> 2188<tr><th align="left">Answer:</th></tr> 2189<tr><td bgcolor="#ffffff"><font color="ffffff"> 2190 In <tt>CONFIG_PREEMPT=n</tt> kernel, <tt>synchronize_rcu()</tt> 2191 maps directly to <tt>synchronize_sched()</tt>. 2192 Therefore, <tt>synchronize_rcu()</tt> works normally throughout 2193 boot in <tt>CONFIG_PREEMPT=n</tt> kernels. 2194 However, your code must also work in <tt>CONFIG_PREEMPT=y</tt> kernels, 2195 so it is still necessary to avoid invoking <tt>synchronize_rcu()</tt> 2196 during scheduler initialization. 2197</font></td></tr> 2198<tr><td> </td></tr> 2199</table> 2200 2201<p> 2202I learned of these boot-time requirements as a result of a series of 2203system hangs. 2204 2205<h3><a name="Interrupts and NMIs">Interrupts and NMIs</a></h3> 2206 2207<p> 2208The Linux kernel has interrupts, and RCU read-side critical sections are 2209legal within interrupt handlers and within interrupt-disabled regions 2210of code, as are invocations of <tt>call_rcu()</tt>. 2211 2212<p> 2213Some Linux-kernel architectures can enter an interrupt handler from 2214non-idle process context, and then just never leave it, instead stealthily 2215transitioning back to process context. 2216This trick is sometimes used to invoke system calls from inside the kernel. 2217These “half-interrupts” mean that RCU has to be very careful 2218about how it counts interrupt nesting levels. 2219I learned of this requirement the hard way during a rewrite 2220of RCU's dyntick-idle code. 2221 2222<p> 2223The Linux kernel has non-maskable interrupts (NMIs), and 2224RCU read-side critical sections are legal within NMI handlers. 2225Thankfully, RCU update-side primitives, including 2226<tt>call_rcu()</tt>, are prohibited within NMI handlers. 2227 2228<p> 2229The name notwithstanding, some Linux-kernel architectures 2230can have nested NMIs, which RCU must handle correctly. 2231Andy Lutomirski 2232<a href="https://lkml.kernel.org/g/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com">surprised me</a> 2233with this requirement; 2234he also kindly surprised me with 2235<a href="https://lkml.kernel.org/g/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com">an algorithm</a> 2236that meets this requirement. 2237 2238<h3><a name="Loadable Modules">Loadable Modules</a></h3> 2239 2240<p> 2241The Linux kernel has loadable modules, and these modules can 2242also be unloaded. 2243After a given module has been unloaded, any attempt to call 2244one of its functions results in a segmentation fault. 2245The module-unload functions must therefore cancel any 2246delayed calls to loadable-module functions, for example, 2247any outstanding <tt>mod_timer()</tt> must be dealt with 2248via <tt>del_timer_sync()</tt> or similar. 2249 2250<p> 2251Unfortunately, there is no way to cancel an RCU callback; 2252once you invoke <tt>call_rcu()</tt>, the callback function is 2253going to eventually be invoked, unless the system goes down first. 2254Because it is normally considered socially irresponsible to crash the system 2255in response to a module unload request, we need some other way 2256to deal with in-flight RCU callbacks. 2257 2258<p> 2259RCU therefore provides 2260<tt><a href="https://lwn.net/Articles/217484/">rcu_barrier()</a></tt>, 2261which waits until all in-flight RCU callbacks have been invoked. 2262If a module uses <tt>call_rcu()</tt>, its exit function should therefore 2263prevent any future invocation of <tt>call_rcu()</tt>, then invoke 2264<tt>rcu_barrier()</tt>. 2265In theory, the underlying module-unload code could invoke 2266<tt>rcu_barrier()</tt> unconditionally, but in practice this would 2267incur unacceptable latencies. 2268 2269<p> 2270Nikita Danilov noted this requirement for an analogous filesystem-unmount 2271situation, and Dipankar Sarma incorporated <tt>rcu_barrier()</tt> into RCU. 2272The need for <tt>rcu_barrier()</tt> for module unloading became 2273apparent later. 2274 2275<h3><a name="Hotplug CPU">Hotplug CPU</a></h3> 2276 2277<p> 2278The Linux kernel supports CPU hotplug, which means that CPUs 2279can come and go. 2280It is of course illegal to use any RCU API member from an offline CPU. 2281This requirement was present from day one in DYNIX/ptx, but 2282on the other hand, the Linux kernel's CPU-hotplug implementation 2283is “interesting.” 2284 2285<p> 2286The Linux-kernel CPU-hotplug implementation has notifiers that 2287are used to allow the various kernel subsystems (including RCU) 2288to respond appropriately to a given CPU-hotplug operation. 2289Most RCU operations may be invoked from CPU-hotplug notifiers, 2290including even normal synchronous grace-period operations 2291such as <tt>synchronize_rcu()</tt>. 2292However, expedited grace-period operations such as 2293<tt>synchronize_rcu_expedited()</tt> are not supported, 2294due to the fact that current implementations block CPU-hotplug 2295operations, which could result in deadlock. 2296 2297<p> 2298In addition, all-callback-wait operations such as 2299<tt>rcu_barrier()</tt> are also not supported, due to the 2300fact that there are phases of CPU-hotplug operations where 2301the outgoing CPU's callbacks will not be invoked until after 2302the CPU-hotplug operation ends, which could also result in deadlock. 2303 2304<h3><a name="Scheduler and RCU">Scheduler and RCU</a></h3> 2305 2306<p> 2307RCU depends on the scheduler, and the scheduler uses RCU to 2308protect some of its data structures. 2309This means the scheduler is forbidden from acquiring 2310the runqueue locks and the priority-inheritance locks 2311in the middle of an outermost RCU read-side critical section unless either 2312(1) it releases them before exiting that same 2313RCU read-side critical section, or 2314(2) interrupts are disabled across 2315that entire RCU read-side critical section. 2316This same prohibition also applies (recursively!) to any lock that is acquired 2317while holding any lock to which this prohibition applies. 2318Adhering to this rule prevents preemptible RCU from invoking 2319<tt>rcu_read_unlock_special()</tt> while either runqueue or 2320priority-inheritance locks are held, thus avoiding deadlock. 2321 2322<p> 2323Prior to v4.4, it was only necessary to disable preemption across 2324RCU read-side critical sections that acquired scheduler locks. 2325In v4.4, expedited grace periods started using IPIs, and these 2326IPIs could force a <tt>rcu_read_unlock()</tt> to take the slowpath. 2327Therefore, this expedited-grace-period change required disabling of 2328interrupts, not just preemption. 2329 2330<p> 2331For RCU's part, the preemptible-RCU <tt>rcu_read_unlock()</tt> 2332implementation must be written carefully to avoid similar deadlocks. 2333In particular, <tt>rcu_read_unlock()</tt> must tolerate an 2334interrupt where the interrupt handler invokes both 2335<tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>. 2336This possibility requires <tt>rcu_read_unlock()</tt> to use 2337negative nesting levels to avoid destructive recursion via 2338interrupt handler's use of RCU. 2339 2340<p> 2341This pair of mutual scheduler-RCU requirements came as a 2342<a href="https://lwn.net/Articles/453002/">complete surprise</a>. 2343 2344<p> 2345As noted above, RCU makes use of kthreads, and it is necessary to 2346avoid excessive CPU-time accumulation by these kthreads. 2347This requirement was no surprise, but RCU's violation of it 2348when running context-switch-heavy workloads when built with 2349<tt>CONFIG_NO_HZ_FULL=y</tt> 2350<a href="http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf">did come as a surprise [PDF]</a>. 2351RCU has made good progress towards meeting this requirement, even 2352for context-switch-have <tt>CONFIG_NO_HZ_FULL=y</tt> workloads, 2353but there is room for further improvement. 2354 2355<h3><a name="Tracing and RCU">Tracing and RCU</a></h3> 2356 2357<p> 2358It is possible to use tracing on RCU code, but tracing itself 2359uses RCU. 2360For this reason, <tt>rcu_dereference_raw_notrace()</tt> 2361is provided for use by tracing, which avoids the destructive 2362recursion that could otherwise ensue. 2363This API is also used by virtualization in some architectures, 2364where RCU readers execute in environments in which tracing 2365cannot be used. 2366The tracing folks both located the requirement and provided the 2367needed fix, so this surprise requirement was relatively painless. 2368 2369<h3><a name="Energy Efficiency">Energy Efficiency</a></h3> 2370 2371<p> 2372Interrupting idle CPUs is considered socially unacceptable, 2373especially by people with battery-powered embedded systems. 2374RCU therefore conserves energy by detecting which CPUs are 2375idle, including tracking CPUs that have been interrupted from idle. 2376This is a large part of the energy-efficiency requirement, 2377so I learned of this via an irate phone call. 2378 2379<p> 2380Because RCU avoids interrupting idle CPUs, it is illegal to 2381execute an RCU read-side critical section on an idle CPU. 2382(Kernels built with <tt>CONFIG_PROVE_RCU=y</tt> will splat 2383if you try it.) 2384The <tt>RCU_NONIDLE()</tt> macro and <tt>_rcuidle</tt> 2385event tracing is provided to work around this restriction. 2386In addition, <tt>rcu_is_watching()</tt> may be used to 2387test whether or not it is currently legal to run RCU read-side 2388critical sections on this CPU. 2389I learned of the need for diagnostics on the one hand 2390and <tt>RCU_NONIDLE()</tt> on the other while inspecting 2391idle-loop code. 2392Steven Rostedt supplied <tt>_rcuidle</tt> event tracing, 2393which is used quite heavily in the idle loop. 2394However, there are some restrictions on the code placed within 2395<tt>RCU_NONIDLE()</tt>: 2396 2397<ol> 2398<li> Blocking is prohibited. 2399 In practice, this is not a serious restriction given that idle 2400 tasks are prohibited from blocking to begin with. 2401<li> Although nesting <tt>RCU_NONIDLE()</tt> is permited, they cannot 2402 nest indefinitely deeply. 2403 However, given that they can be nested on the order of a million 2404 deep, even on 32-bit systems, this should not be a serious 2405 restriction. 2406 This nesting limit would probably be reached long after the 2407 compiler OOMed or the stack overflowed. 2408<li> Any code path that enters <tt>RCU_NONIDLE()</tt> must sequence 2409 out of that same <tt>RCU_NONIDLE()</tt>. 2410 For example, the following is grossly illegal: 2411 2412 <blockquote> 2413 <pre> 2414 1 RCU_NONIDLE({ 2415 2 do_something(); 2416 3 goto bad_idea; /* BUG!!! */ 2417 4 do_something_else();}); 2418 5 bad_idea: 2419 </pre> 2420 </blockquote> 2421 2422 <p> 2423 It is just as illegal to transfer control into the middle of 2424 <tt>RCU_NONIDLE()</tt>'s argument. 2425 Yes, in theory, you could transfer in as long as you also 2426 transferred out, but in practice you could also expect to get sharply 2427 worded review comments. 2428</ol> 2429 2430<p> 2431It is similarly socially unacceptable to interrupt an 2432<tt>nohz_full</tt> CPU running in userspace. 2433RCU must therefore track <tt>nohz_full</tt> userspace 2434execution. 2435And in 2436<a href="https://lwn.net/Articles/558284/"><tt>CONFIG_NO_HZ_FULL_SYSIDLE=y</tt></a> 2437kernels, RCU must separately track idle CPUs on the one hand and 2438CPUs that are either idle or executing in userspace on the other. 2439In both cases, RCU must be able to sample state at two points in 2440time, and be able to determine whether or not some other CPU spent 2441any time idle and/or executing in userspace. 2442 2443<p> 2444These energy-efficiency requirements have proven quite difficult to 2445understand and to meet, for example, there have been more than five 2446clean-sheet rewrites of RCU's energy-efficiency code, the last of 2447which was finally able to demonstrate 2448<a href="http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf">real energy savings running on real hardware [PDF]</a>. 2449As noted earlier, 2450I learned of many of these requirements via angry phone calls: 2451Flaming me on the Linux-kernel mailing list was apparently not 2452sufficient to fully vent their ire at RCU's energy-efficiency bugs! 2453 2454<h3><a name="Memory Efficiency">Memory Efficiency</a></h3> 2455 2456<p> 2457Although small-memory non-realtime systems can simply use Tiny RCU, 2458code size is only one aspect of memory efficiency. 2459Another aspect is the size of the <tt>rcu_head</tt> structure 2460used by <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>. 2461Although this structure contains nothing more than a pair of pointers, 2462it does appear in many RCU-protected data structures, including 2463some that are size critical. 2464The <tt>page</tt> structure is a case in point, as evidenced by 2465the many occurrences of the <tt>union</tt> keyword within that structure. 2466 2467<p> 2468This need for memory efficiency is one reason that RCU uses hand-crafted 2469singly linked lists to track the <tt>rcu_head</tt> structures that 2470are waiting for a grace period to elapse. 2471It is also the reason why <tt>rcu_head</tt> structures do not contain 2472debug information, such as fields tracking the file and line of the 2473<tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> that posted them. 2474Although this information might appear in debug-only kernel builds at some 2475point, in the meantime, the <tt>->func</tt> field will often provide 2476the needed debug information. 2477 2478<p> 2479However, in some cases, the need for memory efficiency leads to even 2480more extreme measures. 2481Returning to the <tt>page</tt> structure, the <tt>rcu_head</tt> field 2482shares storage with a great many other structures that are used at 2483various points in the corresponding page's lifetime. 2484In order to correctly resolve certain 2485<a href="https://lkml.kernel.org/g/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com">race conditions</a>, 2486the Linux kernel's memory-management subsystem needs a particular bit 2487to remain zero during all phases of grace-period processing, 2488and that bit happens to map to the bottom bit of the 2489<tt>rcu_head</tt> structure's <tt>->next</tt> field. 2490RCU makes this guarantee as long as <tt>call_rcu()</tt> 2491is used to post the callback, as opposed to <tt>kfree_rcu()</tt> 2492or some future “lazy” 2493variant of <tt>call_rcu()</tt> that might one day be created for 2494energy-efficiency purposes. 2495 2496<p> 2497That said, there are limits. 2498RCU requires that the <tt>rcu_head</tt> structure be aligned to a 2499two-byte boundary, and passing a misaligned <tt>rcu_head</tt> 2500structure to one of the <tt>call_rcu()</tt> family of functions 2501will result in a splat. 2502It is therefore necessary to exercise caution when packing 2503structures containing fields of type <tt>rcu_head</tt>. 2504Why not a four-byte or even eight-byte alignment requirement? 2505Because the m68k architecture provides only two-byte alignment, 2506and thus acts as alignment's least common denominator. 2507 2508<p> 2509The reason for reserving the bottom bit of pointers to 2510<tt>rcu_head</tt> structures is to leave the door open to 2511“lazy” callbacks whose invocations can safely be deferred. 2512Deferring invocation could potentially have energy-efficiency 2513benefits, but only if the rate of non-lazy callbacks decreases 2514significantly for some important workload. 2515In the meantime, reserving the bottom bit keeps this option open 2516in case it one day becomes useful. 2517 2518<h3><a name="Performance, Scalability, Response Time, and Reliability"> 2519Performance, Scalability, Response Time, and Reliability</a></h3> 2520 2521<p> 2522Expanding on the 2523<a href="#Performance and Scalability">earlier discussion</a>, 2524RCU is used heavily by hot code paths in performance-critical 2525portions of the Linux kernel's networking, security, virtualization, 2526and scheduling code paths. 2527RCU must therefore use efficient implementations, especially in its 2528read-side primitives. 2529To that end, it would be good if preemptible RCU's implementation 2530of <tt>rcu_read_lock()</tt> could be inlined, however, doing 2531this requires resolving <tt>#include</tt> issues with the 2532<tt>task_struct</tt> structure. 2533 2534<p> 2535The Linux kernel supports hardware configurations with up to 25364096 CPUs, which means that RCU must be extremely scalable. 2537Algorithms that involve frequent acquisitions of global locks or 2538frequent atomic operations on global variables simply cannot be 2539tolerated within the RCU implementation. 2540RCU therefore makes heavy use of a combining tree based on the 2541<tt>rcu_node</tt> structure. 2542RCU is required to tolerate all CPUs continuously invoking any 2543combination of RCU's runtime primitives with minimal per-operation 2544overhead. 2545In fact, in many cases, increasing load must <i>decrease</i> the 2546per-operation overhead, witness the batching optimizations for 2547<tt>synchronize_rcu()</tt>, <tt>call_rcu()</tt>, 2548<tt>synchronize_rcu_expedited()</tt>, and <tt>rcu_barrier()</tt>. 2549As a general rule, RCU must cheerfully accept whatever the 2550rest of the Linux kernel decides to throw at it. 2551 2552<p> 2553The Linux kernel is used for real-time workloads, especially 2554in conjunction with the 2555<a href="https://rt.wiki.kernel.org/index.php/Main_Page">-rt patchset</a>. 2556The real-time-latency response requirements are such that the 2557traditional approach of disabling preemption across RCU 2558read-side critical sections is inappropriate. 2559Kernels built with <tt>CONFIG_PREEMPT=y</tt> therefore 2560use an RCU implementation that allows RCU read-side critical 2561sections to be preempted. 2562This requirement made its presence known after users made it 2563clear that an earlier 2564<a href="https://lwn.net/Articles/107930/">real-time patch</a> 2565did not meet their needs, in conjunction with some 2566<a href="https://lkml.kernel.org/g/20050318002026.GA2693@us.ibm.com">RCU issues</a> 2567encountered by a very early version of the -rt patchset. 2568 2569<p> 2570In addition, RCU must make do with a sub-100-microsecond real-time latency 2571budget. 2572In fact, on smaller systems with the -rt patchset, the Linux kernel 2573provides sub-20-microsecond real-time latencies for the whole kernel, 2574including RCU. 2575RCU's scalability and latency must therefore be sufficient for 2576these sorts of configurations. 2577To my surprise, the sub-100-microsecond real-time latency budget 2578<a href="http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf"> 2579applies to even the largest systems [PDF]</a>, 2580up to and including systems with 4096 CPUs. 2581This real-time requirement motivated the grace-period kthread, which 2582also simplified handling of a number of race conditions. 2583 2584<p> 2585RCU must avoid degrading real-time response for CPU-bound threads, whether 2586executing in usermode (which is one use case for 2587<tt>CONFIG_NO_HZ_FULL=y</tt>) or in the kernel. 2588That said, CPU-bound loops in the kernel must execute 2589<tt>cond_resched_rcu_qs()</tt> at least once per few tens of milliseconds 2590in order to avoid receiving an IPI from RCU. 2591 2592<p> 2593Finally, RCU's status as a synchronization primitive means that 2594any RCU failure can result in arbitrary memory corruption that can be 2595extremely difficult to debug. 2596This means that RCU must be extremely reliable, which in 2597practice also means that RCU must have an aggressive stress-test 2598suite. 2599This stress-test suite is called <tt>rcutorture</tt>. 2600 2601<p> 2602Although the need for <tt>rcutorture</tt> was no surprise, 2603the current immense popularity of the Linux kernel is posing 2604interesting—and perhaps unprecedented—validation 2605challenges. 2606To see this, keep in mind that there are well over one billion 2607instances of the Linux kernel running today, given Android 2608smartphones, Linux-powered televisions, and servers. 2609This number can be expected to increase sharply with the advent of 2610the celebrated Internet of Things. 2611 2612<p> 2613Suppose that RCU contains a race condition that manifests on average 2614once per million years of runtime. 2615This bug will be occurring about three times per <i>day</i> across 2616the installed base. 2617RCU could simply hide behind hardware error rates, given that no one 2618should really expect their smartphone to last for a million years. 2619However, anyone taking too much comfort from this thought should 2620consider the fact that in most jurisdictions, a successful multi-year 2621test of a given mechanism, which might include a Linux kernel, 2622suffices for a number of types of safety-critical certifications. 2623In fact, rumor has it that the Linux kernel is already being used 2624in production for safety-critical applications. 2625I don't know about you, but I would feel quite bad if a bug in RCU 2626killed someone. 2627Which might explain my recent focus on validation and verification. 2628 2629<h2><a name="Other RCU Flavors">Other RCU Flavors</a></h2> 2630 2631<p> 2632One of the more surprising things about RCU is that there are now 2633no fewer than five <i>flavors</i>, or API families. 2634In addition, the primary flavor that has been the sole focus up to 2635this point has two different implementations, non-preemptible and 2636preemptible. 2637The other four flavors are listed below, with requirements for each 2638described in a separate section. 2639 2640<ol> 2641<li> <a href="#Bottom-Half Flavor">Bottom-Half Flavor</a> 2642<li> <a href="#Sched Flavor">Sched Flavor</a> 2643<li> <a href="#Sleepable RCU">Sleepable RCU</a> 2644<li> <a href="#Tasks RCU">Tasks RCU</a> 2645<li> <a href="#Waiting for Multiple Grace Periods"> 2646 Waiting for Multiple Grace Periods</a> 2647</ol> 2648 2649<h3><a name="Bottom-Half Flavor">Bottom-Half Flavor</a></h3> 2650 2651<p> 2652The softirq-disable (AKA “bottom-half”, 2653hence the “_bh” abbreviations) 2654flavor of RCU, or <i>RCU-bh</i>, was developed by 2655Dipankar Sarma to provide a flavor of RCU that could withstand the 2656network-based denial-of-service attacks researched by Robert 2657Olsson. 2658These attacks placed so much networking load on the system 2659that some of the CPUs never exited softirq execution, 2660which in turn prevented those CPUs from ever executing a context switch, 2661which, in the RCU implementation of that time, prevented grace periods 2662from ever ending. 2663The result was an out-of-memory condition and a system hang. 2664 2665<p> 2666The solution was the creation of RCU-bh, which does 2667<tt>local_bh_disable()</tt> 2668across its read-side critical sections, and which uses the transition 2669from one type of softirq processing to another as a quiescent state 2670in addition to context switch, idle, user mode, and offline. 2671This means that RCU-bh grace periods can complete even when some of 2672the CPUs execute in softirq indefinitely, thus allowing algorithms 2673based on RCU-bh to withstand network-based denial-of-service attacks. 2674 2675<p> 2676Because 2677<tt>rcu_read_lock_bh()</tt> and <tt>rcu_read_unlock_bh()</tt> 2678disable and re-enable softirq handlers, any attempt to start a softirq 2679handlers during the 2680RCU-bh read-side critical section will be deferred. 2681In this case, <tt>rcu_read_unlock_bh()</tt> 2682will invoke softirq processing, which can take considerable time. 2683One can of course argue that this softirq overhead should be associated 2684with the code following the RCU-bh read-side critical section rather 2685than <tt>rcu_read_unlock_bh()</tt>, but the fact 2686is that most profiling tools cannot be expected to make this sort 2687of fine distinction. 2688For example, suppose that a three-millisecond-long RCU-bh read-side 2689critical section executes during a time of heavy networking load. 2690There will very likely be an attempt to invoke at least one softirq 2691handler during that three milliseconds, but any such invocation will 2692be delayed until the time of the <tt>rcu_read_unlock_bh()</tt>. 2693This can of course make it appear at first glance as if 2694<tt>rcu_read_unlock_bh()</tt> was executing very slowly. 2695 2696<p> 2697The 2698<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-bh API</a> 2699includes 2700<tt>rcu_read_lock_bh()</tt>, 2701<tt>rcu_read_unlock_bh()</tt>, 2702<tt>rcu_dereference_bh()</tt>, 2703<tt>rcu_dereference_bh_check()</tt>, 2704<tt>synchronize_rcu_bh()</tt>, 2705<tt>synchronize_rcu_bh_expedited()</tt>, 2706<tt>call_rcu_bh()</tt>, 2707<tt>rcu_barrier_bh()</tt>, and 2708<tt>rcu_read_lock_bh_held()</tt>. 2709 2710<h3><a name="Sched Flavor">Sched Flavor</a></h3> 2711 2712<p> 2713Before preemptible RCU, waiting for an RCU grace period had the 2714side effect of also waiting for all pre-existing interrupt 2715and NMI handlers. 2716However, there are legitimate preemptible-RCU implementations that 2717do not have this property, given that any point in the code outside 2718of an RCU read-side critical section can be a quiescent state. 2719Therefore, <i>RCU-sched</i> was created, which follows “classic” 2720RCU in that an RCU-sched grace period waits for for pre-existing 2721interrupt and NMI handlers. 2722In kernels built with <tt>CONFIG_PREEMPT=n</tt>, the RCU and RCU-sched 2723APIs have identical implementations, while kernels built with 2724<tt>CONFIG_PREEMPT=y</tt> provide a separate implementation for each. 2725 2726<p> 2727Note well that in <tt>CONFIG_PREEMPT=y</tt> kernels, 2728<tt>rcu_read_lock_sched()</tt> and <tt>rcu_read_unlock_sched()</tt> 2729disable and re-enable preemption, respectively. 2730This means that if there was a preemption attempt during the 2731RCU-sched read-side critical section, <tt>rcu_read_unlock_sched()</tt> 2732will enter the scheduler, with all the latency and overhead entailed. 2733Just as with <tt>rcu_read_unlock_bh()</tt>, this can make it look 2734as if <tt>rcu_read_unlock_sched()</tt> was executing very slowly. 2735However, the highest-priority task won't be preempted, so that task 2736will enjoy low-overhead <tt>rcu_read_unlock_sched()</tt> invocations. 2737 2738<p> 2739The 2740<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-sched API</a> 2741includes 2742<tt>rcu_read_lock_sched()</tt>, 2743<tt>rcu_read_unlock_sched()</tt>, 2744<tt>rcu_read_lock_sched_notrace()</tt>, 2745<tt>rcu_read_unlock_sched_notrace()</tt>, 2746<tt>rcu_dereference_sched()</tt>, 2747<tt>rcu_dereference_sched_check()</tt>, 2748<tt>synchronize_sched()</tt>, 2749<tt>synchronize_rcu_sched_expedited()</tt>, 2750<tt>call_rcu_sched()</tt>, 2751<tt>rcu_barrier_sched()</tt>, and 2752<tt>rcu_read_lock_sched_held()</tt>. 2753However, anything that disables preemption also marks an RCU-sched 2754read-side critical section, including 2755<tt>preempt_disable()</tt> and <tt>preempt_enable()</tt>, 2756<tt>local_irq_save()</tt> and <tt>local_irq_restore()</tt>, 2757and so on. 2758 2759<h3><a name="Sleepable RCU">Sleepable RCU</a></h3> 2760 2761<p> 2762For well over a decade, someone saying “I need to block within 2763an RCU read-side critical section” was a reliable indication 2764that this someone did not understand RCU. 2765After all, if you are always blocking in an RCU read-side critical 2766section, you can probably afford to use a higher-overhead synchronization 2767mechanism. 2768However, that changed with the advent of the Linux kernel's notifiers, 2769whose RCU read-side critical 2770sections almost never sleep, but sometimes need to. 2771This resulted in the introduction of 2772<a href="https://lwn.net/Articles/202847/">sleepable RCU</a>, 2773or <i>SRCU</i>. 2774 2775<p> 2776SRCU allows different domains to be defined, with each such domain 2777defined by an instance of an <tt>srcu_struct</tt> structure. 2778A pointer to this structure must be passed in to each SRCU function, 2779for example, <tt>synchronize_srcu(&ss)</tt>, where 2780<tt>ss</tt> is the <tt>srcu_struct</tt> structure. 2781The key benefit of these domains is that a slow SRCU reader in one 2782domain does not delay an SRCU grace period in some other domain. 2783That said, one consequence of these domains is that read-side code 2784must pass a “cookie” from <tt>srcu_read_lock()</tt> 2785to <tt>srcu_read_unlock()</tt>, for example, as follows: 2786 2787<blockquote> 2788<pre> 2789 1 int idx; 2790 2 2791 3 idx = srcu_read_lock(&ss); 2792 4 do_something(); 2793 5 srcu_read_unlock(&ss, idx); 2794</pre> 2795</blockquote> 2796 2797<p> 2798As noted above, it is legal to block within SRCU read-side critical sections, 2799however, with great power comes great responsibility. 2800If you block forever in one of a given domain's SRCU read-side critical 2801sections, then that domain's grace periods will also be blocked forever. 2802Of course, one good way to block forever is to deadlock, which can 2803happen if any operation in a given domain's SRCU read-side critical 2804section can block waiting, either directly or indirectly, for that domain's 2805grace period to elapse. 2806For example, this results in a self-deadlock: 2807 2808<blockquote> 2809<pre> 2810 1 int idx; 2811 2 2812 3 idx = srcu_read_lock(&ss); 2813 4 do_something(); 2814 5 synchronize_srcu(&ss); 2815 6 srcu_read_unlock(&ss, idx); 2816</pre> 2817</blockquote> 2818 2819<p> 2820However, if line 5 acquired a mutex that was held across 2821a <tt>synchronize_srcu()</tt> for domain <tt>ss</tt>, 2822deadlock would still be possible. 2823Furthermore, if line 5 acquired a mutex that was held across 2824a <tt>synchronize_srcu()</tt> for some other domain <tt>ss1</tt>, 2825and if an <tt>ss1</tt>-domain SRCU read-side critical section 2826acquired another mutex that was held across as <tt>ss</tt>-domain 2827<tt>synchronize_srcu()</tt>, 2828deadlock would again be possible. 2829Such a deadlock cycle could extend across an arbitrarily large number 2830of different SRCU domains. 2831Again, with great power comes great responsibility. 2832 2833<p> 2834Unlike the other RCU flavors, SRCU read-side critical sections can 2835run on idle and even offline CPUs. 2836This ability requires that <tt>srcu_read_lock()</tt> and 2837<tt>srcu_read_unlock()</tt> contain memory barriers, which means 2838that SRCU readers will run a bit slower than would RCU readers. 2839It also motivates the <tt>smp_mb__after_srcu_read_unlock()</tt> 2840API, which, in combination with <tt>srcu_read_unlock()</tt>, 2841guarantees a full memory barrier. 2842 2843<p> 2844The 2845<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">SRCU API</a> 2846includes 2847<tt>srcu_read_lock()</tt>, 2848<tt>srcu_read_unlock()</tt>, 2849<tt>srcu_dereference()</tt>, 2850<tt>srcu_dereference_check()</tt>, 2851<tt>synchronize_srcu()</tt>, 2852<tt>synchronize_srcu_expedited()</tt>, 2853<tt>call_srcu()</tt>, 2854<tt>srcu_barrier()</tt>, and 2855<tt>srcu_read_lock_held()</tt>. 2856It also includes 2857<tt>DEFINE_SRCU()</tt>, 2858<tt>DEFINE_STATIC_SRCU()</tt>, and 2859<tt>init_srcu_struct()</tt> 2860APIs for defining and initializing <tt>srcu_struct</tt> structures. 2861 2862<h3><a name="Tasks RCU">Tasks RCU</a></h3> 2863 2864<p> 2865Some forms of tracing use “tramopolines” to handle the 2866binary rewriting required to install different types of probes. 2867It would be good to be able to free old trampolines, which sounds 2868like a job for some form of RCU. 2869However, because it is necessary to be able to install a trace 2870anywhere in the code, it is not possible to use read-side markers 2871such as <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>. 2872In addition, it does not work to have these markers in the trampoline 2873itself, because there would need to be instructions following 2874<tt>rcu_read_unlock()</tt>. 2875Although <tt>synchronize_rcu()</tt> would guarantee that execution 2876reached the <tt>rcu_read_unlock()</tt>, it would not be able to 2877guarantee that execution had completely left the trampoline. 2878 2879<p> 2880The solution, in the form of 2881<a href="https://lwn.net/Articles/607117/"><i>Tasks RCU</i></a>, 2882is to have implicit 2883read-side critical sections that are delimited by voluntary context 2884switches, that is, calls to <tt>schedule()</tt>, 2885<tt>cond_resched_rcu_qs()</tt>, and 2886<tt>synchronize_rcu_tasks()</tt>. 2887In addition, transitions to and from userspace execution also delimit 2888tasks-RCU read-side critical sections. 2889 2890<p> 2891The tasks-RCU API is quite compact, consisting only of 2892<tt>call_rcu_tasks()</tt>, 2893<tt>synchronize_rcu_tasks()</tt>, and 2894<tt>rcu_barrier_tasks()</tt>. 2895 2896<h3><a name="Waiting for Multiple Grace Periods"> 2897Waiting for Multiple Grace Periods</a></h3> 2898 2899<p> 2900Perhaps you have an RCU protected data structure that is accessed from 2901RCU read-side critical sections, from softirq handlers, and from 2902hardware interrupt handlers. 2903That is three flavors of RCU, the normal flavor, the bottom-half flavor, 2904and the sched flavor. 2905How to wait for a compound grace period? 2906 2907<p> 2908The best approach is usually to “just say no!” and 2909insert <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt> 2910around each RCU read-side critical section, regardless of what 2911environment it happens to be in. 2912But suppose that some of the RCU read-side critical sections are 2913on extremely hot code paths, and that use of <tt>CONFIG_PREEMPT=n</tt> 2914is not a viable option, so that <tt>rcu_read_lock()</tt> and 2915<tt>rcu_read_unlock()</tt> are not free. 2916What then? 2917 2918<p> 2919You <i>could</i> wait on all three grace periods in succession, as follows: 2920 2921<blockquote> 2922<pre> 2923 1 synchronize_rcu(); 2924 2 synchronize_rcu_bh(); 2925 3 synchronize_sched(); 2926</pre> 2927</blockquote> 2928 2929<p> 2930This works, but triples the update-side latency penalty. 2931In cases where this is not acceptable, <tt>synchronize_rcu_mult()</tt> 2932may be used to wait on all three flavors of grace period concurrently: 2933 2934<blockquote> 2935<pre> 2936 1 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched); 2937</pre> 2938</blockquote> 2939 2940<p> 2941But what if it is necessary to also wait on SRCU? 2942This can be done as follows: 2943 2944<blockquote> 2945<pre> 2946 1 static void call_my_srcu(struct rcu_head *head, 2947 2 void (*func)(struct rcu_head *head)) 2948 3 { 2949 4 call_srcu(&my_srcu, head, func); 2950 5 } 2951 6 2952 7 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched, call_my_srcu); 2953</pre> 2954</blockquote> 2955 2956<p> 2957If you needed to wait on multiple different flavors of SRCU 2958(but why???), you would need to create a wrapper function resembling 2959<tt>call_my_srcu()</tt> for each SRCU flavor. 2960 2961<table> 2962<tr><th> </th></tr> 2963<tr><th align="left">Quick Quiz:</th></tr> 2964<tr><td> 2965 But what if I need to wait for multiple RCU flavors, but I also need 2966 the grace periods to be expedited? 2967</td></tr> 2968<tr><th align="left">Answer:</th></tr> 2969<tr><td bgcolor="#ffffff"><font color="ffffff"> 2970 If you are using expedited grace periods, there should be less penalty 2971 for waiting on them in succession. 2972 But if that is nevertheless a problem, you can use workqueues 2973 or multiple kthreads to wait on the various expedited grace 2974 periods concurrently. 2975</font></td></tr> 2976<tr><td> </td></tr> 2977</table> 2978 2979<p> 2980Again, it is usually better to adjust the RCU read-side critical sections 2981to use a single flavor of RCU, but when this is not feasible, you can use 2982<tt>synchronize_rcu_mult()</tt>. 2983 2984<h2><a name="Possible Future Changes">Possible Future Changes</a></h2> 2985 2986<p> 2987One of the tricks that RCU uses to attain update-side scalability is 2988to increase grace-period latency with increasing numbers of CPUs. 2989If this becomes a serious problem, it will be necessary to rework the 2990grace-period state machine so as to avoid the need for the additional 2991latency. 2992 2993<p> 2994Expedited grace periods scan the CPUs, so their latency and overhead 2995increases with increasing numbers of CPUs. 2996If this becomes a serious problem on large systems, it will be necessary 2997to do some redesign to avoid this scalability problem. 2998 2999<p> 3000RCU disables CPU hotplug in a few places, perhaps most notably in the 3001expedited grace-period and <tt>rcu_barrier()</tt> operations. 3002If there is a strong reason to use expedited grace periods in CPU-hotplug 3003notifiers, it will be necessary to avoid disabling CPU hotplug. 3004This would introduce some complexity, so there had better be a <i>very</i> 3005good reason. 3006 3007<p> 3008The tradeoff between grace-period latency on the one hand and interruptions 3009of other CPUs on the other hand may need to be re-examined. 3010The desire is of course for zero grace-period latency as well as zero 3011interprocessor interrupts undertaken during an expedited grace period 3012operation. 3013While this ideal is unlikely to be achievable, it is quite possible that 3014further improvements can be made. 3015 3016<p> 3017The multiprocessor implementations of RCU use a combining tree that 3018groups CPUs so as to reduce lock contention and increase cache locality. 3019However, this combining tree does not spread its memory across NUMA 3020nodes nor does it align the CPU groups with hardware features such 3021as sockets or cores. 3022Such spreading and alignment is currently believed to be unnecessary 3023because the hotpath read-side primitives do not access the combining 3024tree, nor does <tt>call_rcu()</tt> in the common case. 3025If you believe that your architecture needs such spreading and alignment, 3026then your architecture should also benefit from the 3027<tt>rcutree.rcu_fanout_leaf</tt> boot parameter, which can be set 3028to the number of CPUs in a socket, NUMA node, or whatever. 3029If the number of CPUs is too large, use a fraction of the number of 3030CPUs. 3031If the number of CPUs is a large prime number, well, that certainly 3032is an “interesting” architectural choice! 3033More flexible arrangements might be considered, but only if 3034<tt>rcutree.rcu_fanout_leaf</tt> has proven inadequate, and only 3035if the inadequacy has been demonstrated by a carefully run and 3036realistic system-level workload. 3037 3038<p> 3039Please note that arrangements that require RCU to remap CPU numbers will 3040require extremely good demonstration of need and full exploration of 3041alternatives. 3042 3043<p> 3044There is an embarrassingly large number of flavors of RCU, and this 3045number has been increasing over time. 3046Perhaps it will be possible to combine some at some future date. 3047 3048<p> 3049RCU's various kthreads are reasonably recent additions. 3050It is quite likely that adjustments will be required to more gracefully 3051handle extreme loads. 3052It might also be necessary to be able to relate CPU utilization by 3053RCU's kthreads and softirq handlers to the code that instigated this 3054CPU utilization. 3055For example, RCU callback overhead might be charged back to the 3056originating <tt>call_rcu()</tt> instance, though probably not 3057in production kernels. 3058 3059<h2><a name="Summary">Summary</a></h2> 3060 3061<p> 3062This document has presented more than two decade's worth of RCU 3063requirements. 3064Given that the requirements keep changing, this will not be the last 3065word on this subject, but at least it serves to get an important 3066subset of the requirements set forth. 3067 3068<h2><a name="Acknowledgments">Acknowledgments</a></h2> 3069 3070I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar, 3071Oleg Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and 3072Andy Lutomirski for their help in rendering 3073this article human readable, and to Michelle Rankin for her support 3074of this effort. 3075Other contributions are acknowledged in the Linux kernel's git archive. 3076The cartoon is copyright (c) 2013 by Melissa Broussard, 3077and is provided 3078under the terms of the Creative Commons Attribution-Share Alike 3.0 3079United States license. 3080 3081</body></html> 3082
