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 distinct 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, <tt>sparse</tt> will complain if you 563 define <tt>gp</tt> with <tt>__rcu</tt> and then 564 access it without using 565 either <tt>rcu_access_pointer()</tt> or <tt>rcu_dereference()</tt>. 566</font></td></tr> 567<tr><td> </td></tr> 568</table> 569 570<p> 571In short, RCU's publish-subscribe guarantee is provided by the combination 572of <tt>rcu_assign_pointer()</tt> and <tt>rcu_dereference()</tt>. 573This guarantee allows data elements to be safely added to RCU-protected 574linked data structures without disrupting RCU readers. 575This guarantee can be used in combination with the grace-period 576guarantee to also allow data elements to be removed from RCU-protected 577linked data structures, again without disrupting RCU readers. 578 579<p> 580This guarantee was only partially premeditated. 581DYNIX/ptx used an explicit memory barrier for publication, but had nothing 582resembling <tt>rcu_dereference()</tt> for subscription, nor did it 583have anything resembling the <tt>smp_read_barrier_depends()</tt> 584that was later subsumed into <tt>rcu_dereference()</tt> and later 585still into <tt>READ_ONCE()</tt>. 586The need for these operations made itself known quite suddenly at a 587late-1990s meeting with the DEC Alpha architects, back in the days when 588DEC was still a free-standing company. 589It took the Alpha architects a good hour to convince me that any sort 590of barrier would ever be needed, and it then took me a good <i>two</i> hours 591to convince them that their documentation did not make this point clear. 592More recent work with the C and C++ standards committees have provided 593much education on tricks and traps from the compiler. 594In short, compilers were much less tricky in the early 1990s, but in 5952015, don't even think about omitting <tt>rcu_dereference()</tt>! 596 597<h3><a name="Memory-Barrier Guarantees">Memory-Barrier Guarantees</a></h3> 598 599<p> 600The previous section's simple linked-data-structure scenario clearly 601demonstrates the need for RCU's stringent memory-ordering guarantees on 602systems with more than one CPU: 603 604<ol> 605<li> Each CPU that has an RCU read-side critical section that 606 begins before <tt>synchronize_rcu()</tt> starts is 607 guaranteed to execute a full memory barrier between the time 608 that the RCU read-side critical section ends and the time that 609 <tt>synchronize_rcu()</tt> returns. 610 Without this guarantee, a pre-existing RCU read-side critical section 611 might hold a reference to the newly removed <tt>struct foo</tt> 612 after the <tt>kfree()</tt> on line 14 of 613 <tt>remove_gp_synchronous()</tt>. 614<li> Each CPU that has an RCU read-side critical section that ends 615 after <tt>synchronize_rcu()</tt> returns is guaranteed 616 to execute a full memory barrier between the time that 617 <tt>synchronize_rcu()</tt> begins and the time that the RCU 618 read-side critical section begins. 619 Without this guarantee, a later RCU read-side critical section 620 running after the <tt>kfree()</tt> on line 14 of 621 <tt>remove_gp_synchronous()</tt> might 622 later run <tt>do_something_gp()</tt> and find the 623 newly deleted <tt>struct foo</tt>. 624<li> If the task invoking <tt>synchronize_rcu()</tt> remains 625 on a given CPU, then that CPU is guaranteed to execute a full 626 memory barrier sometime during the execution of 627 <tt>synchronize_rcu()</tt>. 628 This guarantee ensures that the <tt>kfree()</tt> on 629 line 14 of <tt>remove_gp_synchronous()</tt> really does 630 execute after the removal on line 11. 631<li> If the task invoking <tt>synchronize_rcu()</tt> migrates 632 among a group of CPUs during that invocation, then each of the 633 CPUs in that group is guaranteed to execute a full memory barrier 634 sometime during the execution of <tt>synchronize_rcu()</tt>. 635 This guarantee also ensures that the <tt>kfree()</tt> on 636 line 14 of <tt>remove_gp_synchronous()</tt> really does 637 execute after the removal on 638 line 11, but also in the case where the thread executing the 639 <tt>synchronize_rcu()</tt> migrates in the meantime. 640</ol> 641 642<table> 643<tr><th> </th></tr> 644<tr><th align="left">Quick Quiz:</th></tr> 645<tr><td> 646 Given that multiple CPUs can start RCU read-side critical sections 647 at any time without any ordering whatsoever, how can RCU possibly 648 tell whether or not a given RCU read-side critical section starts 649 before a given instance of <tt>synchronize_rcu()</tt>? 650</td></tr> 651<tr><th align="left">Answer:</th></tr> 652<tr><td bgcolor="#ffffff"><font color="ffffff"> 653 If RCU cannot tell whether or not a given 654 RCU read-side critical section starts before a 655 given instance of <tt>synchronize_rcu()</tt>, 656 then it must assume that the RCU read-side critical section 657 started first. 658 In other words, a given instance of <tt>synchronize_rcu()</tt> 659 can avoid waiting on a given RCU read-side critical section only 660 if it can prove that <tt>synchronize_rcu()</tt> started first. 661 </font> 662 663 <p><font color="ffffff"> 664 A related question is “When <tt>rcu_read_lock()</tt> 665 doesn't generate any code, why does it matter how it relates 666 to a grace period?” 667 The answer is that it is not the relationship of 668 <tt>rcu_read_lock()</tt> itself that is important, but rather 669 the relationship of the code within the enclosed RCU read-side 670 critical section to the code preceding and following the 671 grace period. 672 If we take this viewpoint, then a given RCU read-side critical 673 section begins before a given grace period when some access 674 preceding the grace period observes the effect of some access 675 within the critical section, in which case none of the accesses 676 within the critical section may observe the effects of any 677 access following the grace period. 678 </font> 679 680 <p><font color="ffffff"> 681 As of late 2016, mathematical models of RCU take this 682 viewpoint, for example, see slides 62 and 63 683 of the 684 <a href="http://www2.rdrop.com/users/paulmck/scalability/paper/LinuxMM.2016.10.04c.LCE.pdf">2016 LinuxCon EU</a> 685 presentation. 686</font></td></tr> 687<tr><td> </td></tr> 688</table> 689 690<table> 691<tr><th> </th></tr> 692<tr><th align="left">Quick Quiz:</th></tr> 693<tr><td> 694 The first and second guarantees require unbelievably strict ordering! 695 Are all these memory barriers <i> really</i> required? 696</td></tr> 697<tr><th align="left">Answer:</th></tr> 698<tr><td bgcolor="#ffffff"><font color="ffffff"> 699 Yes, they really are required. 700 To see why the first guarantee is required, consider the following 701 sequence of events: 702 </font> 703 704 <ol> 705 <li> <font color="ffffff"> 706 CPU 1: <tt>rcu_read_lock()</tt> 707 </font> 708 <li> <font color="ffffff"> 709 CPU 1: <tt>q = rcu_dereference(gp); 710 /* Very likely to return p. */</tt> 711 </font> 712 <li> <font color="ffffff"> 713 CPU 0: <tt>list_del_rcu(p);</tt> 714 </font> 715 <li> <font color="ffffff"> 716 CPU 0: <tt>synchronize_rcu()</tt> starts. 717 </font> 718 <li> <font color="ffffff"> 719 CPU 1: <tt>do_something_with(q->a); 720 /* No smp_mb(), so might happen after kfree(). */</tt> 721 </font> 722 <li> <font color="ffffff"> 723 CPU 1: <tt>rcu_read_unlock()</tt> 724 </font> 725 <li> <font color="ffffff"> 726 CPU 0: <tt>synchronize_rcu()</tt> returns. 727 </font> 728 <li> <font color="ffffff"> 729 CPU 0: <tt>kfree(p);</tt> 730 </font> 731 </ol> 732 733 <p><font color="ffffff"> 734 Therefore, there absolutely must be a full memory barrier between the 735 end of the RCU read-side critical section and the end of the 736 grace period. 737 </font> 738 739 <p><font color="ffffff"> 740 The sequence of events demonstrating the necessity of the second rule 741 is roughly similar: 742 </font> 743 744 <ol> 745 <li> <font color="ffffff">CPU 0: <tt>list_del_rcu(p);</tt> 746 </font> 747 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> starts. 748 </font> 749 <li> <font color="ffffff">CPU 1: <tt>rcu_read_lock()</tt> 750 </font> 751 <li> <font color="ffffff">CPU 1: <tt>q = rcu_dereference(gp); 752 /* Might return p if no memory barrier. */</tt> 753 </font> 754 <li> <font color="ffffff">CPU 0: <tt>synchronize_rcu()</tt> returns. 755 </font> 756 <li> <font color="ffffff">CPU 0: <tt>kfree(p);</tt> 757 </font> 758 <li> <font color="ffffff"> 759 CPU 1: <tt>do_something_with(q->a); /* Boom!!! */</tt> 760 </font> 761 <li> <font color="ffffff">CPU 1: <tt>rcu_read_unlock()</tt> 762 </font> 763 </ol> 764 765 <p><font color="ffffff"> 766 And similarly, without a memory barrier between the beginning of the 767 grace period and the beginning of the RCU read-side critical section, 768 CPU 1 might end up accessing the freelist. 769 </font> 770 771 <p><font color="ffffff"> 772 The “as if” rule of course applies, so that any 773 implementation that acts as if the appropriate memory barriers 774 were in place is a correct implementation. 775 That said, it is much easier to fool yourself into believing 776 that you have adhered to the as-if rule than it is to actually 777 adhere to it! 778</font></td></tr> 779<tr><td> </td></tr> 780</table> 781 782<table> 783<tr><th> </th></tr> 784<tr><th align="left">Quick Quiz:</th></tr> 785<tr><td> 786 You claim that <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt> 787 generate absolutely no code in some kernel builds. 788 This means that the compiler might arbitrarily rearrange consecutive 789 RCU read-side critical sections. 790 Given such rearrangement, if a given RCU read-side critical section 791 is done, how can you be sure that all prior RCU read-side critical 792 sections are done? 793 Won't the compiler rearrangements make that impossible to determine? 794</td></tr> 795<tr><th align="left">Answer:</th></tr> 796<tr><td bgcolor="#ffffff"><font color="ffffff"> 797 In cases where <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt> 798 generate absolutely no code, RCU infers quiescent states only at 799 special locations, for example, within the scheduler. 800 Because calls to <tt>schedule()</tt> had better prevent calling-code 801 accesses to shared variables from being rearranged across the call to 802 <tt>schedule()</tt>, if RCU detects the end of a given RCU read-side 803 critical section, it will necessarily detect the end of all prior 804 RCU read-side critical sections, no matter how aggressively the 805 compiler scrambles the code. 806 </font> 807 808 <p><font color="ffffff"> 809 Again, this all assumes that the compiler cannot scramble code across 810 calls to the scheduler, out of interrupt handlers, into the idle loop, 811 into user-mode code, and so on. 812 But if your kernel build allows that sort of scrambling, you have broken 813 far more than just RCU! 814</font></td></tr> 815<tr><td> </td></tr> 816</table> 817 818<p> 819Note that these memory-barrier requirements do not replace the fundamental 820RCU requirement that a grace period wait for all pre-existing readers. 821On the contrary, the memory barriers called out in this section must operate in 822such a way as to <i>enforce</i> this fundamental requirement. 823Of course, different implementations enforce this requirement in different 824ways, but enforce it they must. 825 826<h3><a name="RCU Primitives Guaranteed to Execute Unconditionally">RCU Primitives Guaranteed to Execute Unconditionally</a></h3> 827 828<p> 829The common-case RCU primitives are unconditional. 830They are invoked, they do their job, and they return, with no possibility 831of error, and no need to retry. 832This is a key RCU design philosophy. 833 834<p> 835However, this philosophy is pragmatic rather than pigheaded. 836If someone comes up with a good justification for a particular conditional 837RCU primitive, it might well be implemented and added. 838After all, this guarantee was reverse-engineered, not premeditated. 839The unconditional nature of the RCU primitives was initially an 840accident of implementation, and later experience with synchronization 841primitives with conditional primitives caused me to elevate this 842accident to a guarantee. 843Therefore, the justification for adding a conditional primitive to 844RCU would need to be based on detailed and compelling use cases. 845 846<h3><a name="Guaranteed Read-to-Write Upgrade">Guaranteed Read-to-Write Upgrade</a></h3> 847 848<p> 849As far as RCU is concerned, it is always possible to carry out an 850update within an RCU read-side critical section. 851For example, that RCU read-side critical section might search for 852a given data element, and then might acquire the update-side 853spinlock in order to update that element, all while remaining 854in that RCU read-side critical section. 855Of course, it is necessary to exit the RCU read-side critical section 856before invoking <tt>synchronize_rcu()</tt>, however, this 857inconvenience can be avoided through use of the 858<tt>call_rcu()</tt> and <tt>kfree_rcu()</tt> API members 859described later in this document. 860 861<table> 862<tr><th> </th></tr> 863<tr><th align="left">Quick Quiz:</th></tr> 864<tr><td> 865 But how does the upgrade-to-write operation exclude other readers? 866</td></tr> 867<tr><th align="left">Answer:</th></tr> 868<tr><td bgcolor="#ffffff"><font color="ffffff"> 869 It doesn't, just like normal RCU updates, which also do not exclude 870 RCU readers. 871</font></td></tr> 872<tr><td> </td></tr> 873</table> 874 875<p> 876This guarantee allows lookup code to be shared between read-side 877and update-side code, and was premeditated, appearing in the earliest 878DYNIX/ptx RCU documentation. 879 880<h2><a name="Fundamental Non-Requirements">Fundamental Non-Requirements</a></h2> 881 882<p> 883RCU provides extremely lightweight readers, and its read-side guarantees, 884though quite useful, are correspondingly lightweight. 885It is therefore all too easy to assume that RCU is guaranteeing more 886than it really is. 887Of course, the list of things that RCU does not guarantee is infinitely 888long, however, the following sections list a few non-guarantees that 889have caused confusion. 890Except where otherwise noted, these non-guarantees were premeditated. 891 892<ol> 893<li> <a href="#Readers Impose Minimal Ordering"> 894 Readers Impose Minimal Ordering</a> 895<li> <a href="#Readers Do Not Exclude Updaters"> 896 Readers Do Not Exclude Updaters</a> 897<li> <a href="#Updaters Only Wait For Old Readers"> 898 Updaters Only Wait For Old Readers</a> 899<li> <a href="#Grace Periods Don't Partition Read-Side Critical Sections"> 900 Grace Periods Don't Partition Read-Side Critical Sections</a> 901<li> <a href="#Read-Side Critical Sections Don't Partition Grace Periods"> 902 Read-Side Critical Sections Don't Partition Grace Periods</a> 903<li> <a href="#Disabling Preemption Does Not Block Grace Periods"> 904 Disabling Preemption Does Not Block Grace Periods</a> 905</ol> 906 907<h3><a name="Readers Impose Minimal Ordering">Readers Impose Minimal Ordering</a></h3> 908 909<p> 910Reader-side markers such as <tt>rcu_read_lock()</tt> and 911<tt>rcu_read_unlock()</tt> provide absolutely no ordering guarantees 912except through their interaction with the grace-period APIs such as 913<tt>synchronize_rcu()</tt>. 914To see this, consider the following pair of threads: 915 916<blockquote> 917<pre> 918 1 void thread0(void) 919 2 { 920 3 rcu_read_lock(); 921 4 WRITE_ONCE(x, 1); 922 5 rcu_read_unlock(); 923 6 rcu_read_lock(); 924 7 WRITE_ONCE(y, 1); 925 8 rcu_read_unlock(); 926 9 } 92710 92811 void thread1(void) 92912 { 93013 rcu_read_lock(); 93114 r1 = READ_ONCE(y); 93215 rcu_read_unlock(); 93316 rcu_read_lock(); 93417 r2 = READ_ONCE(x); 93518 rcu_read_unlock(); 93619 } 937</pre> 938</blockquote> 939 940<p> 941After <tt>thread0()</tt> and <tt>thread1()</tt> execute 942concurrently, it is quite possible to have 943 944<blockquote> 945<pre> 946(r1 == 1 && r2 == 0) 947</pre> 948</blockquote> 949 950(that is, <tt>y</tt> appears to have been assigned before <tt>x</tt>), 951which would not be possible if <tt>rcu_read_lock()</tt> and 952<tt>rcu_read_unlock()</tt> had much in the way of ordering 953properties. 954But they do not, so the CPU is within its rights 955to do significant reordering. 956This is by design: Any significant ordering constraints would slow down 957these fast-path APIs. 958 959<table> 960<tr><th> </th></tr> 961<tr><th align="left">Quick Quiz:</th></tr> 962<tr><td> 963 Can't the compiler also reorder this code? 964</td></tr> 965<tr><th align="left">Answer:</th></tr> 966<tr><td bgcolor="#ffffff"><font color="ffffff"> 967 No, the volatile casts in <tt>READ_ONCE()</tt> and 968 <tt>WRITE_ONCE()</tt> prevent the compiler from reordering in 969 this particular case. 970</font></td></tr> 971<tr><td> </td></tr> 972</table> 973 974<h3><a name="Readers Do Not Exclude Updaters">Readers Do Not Exclude Updaters</a></h3> 975 976<p> 977Neither <tt>rcu_read_lock()</tt> nor <tt>rcu_read_unlock()</tt> 978exclude updates. 979All they do is to prevent grace periods from ending. 980The following example illustrates this: 981 982<blockquote> 983<pre> 984 1 void thread0(void) 985 2 { 986 3 rcu_read_lock(); 987 4 r1 = READ_ONCE(y); 988 5 if (r1) { 989 6 do_something_with_nonzero_x(); 990 7 r2 = READ_ONCE(x); 991 8 WARN_ON(!r2); /* BUG!!! */ 992 9 } 99310 rcu_read_unlock(); 99411 } 99512 99613 void thread1(void) 99714 { 99815 spin_lock(&my_lock); 99916 WRITE_ONCE(x, 1); 100017 WRITE_ONCE(y, 1); 100118 spin_unlock(&my_lock); 100219 } 1003</pre> 1004</blockquote> 1005 1006<p> 1007If the <tt>thread0()</tt> function's <tt>rcu_read_lock()</tt> 1008excluded the <tt>thread1()</tt> function's update, 1009the <tt>WARN_ON()</tt> could never fire. 1010But the fact is that <tt>rcu_read_lock()</tt> does not exclude 1011much of anything aside from subsequent grace periods, of which 1012<tt>thread1()</tt> has none, so the 1013<tt>WARN_ON()</tt> can and does fire. 1014 1015<h3><a name="Updaters Only Wait For Old Readers">Updaters Only Wait For Old Readers</a></h3> 1016 1017<p> 1018It might be tempting to assume that after <tt>synchronize_rcu()</tt> 1019completes, there are no readers executing. 1020This temptation must be avoided because 1021new readers can start immediately after <tt>synchronize_rcu()</tt> 1022starts, and <tt>synchronize_rcu()</tt> is under no 1023obligation to wait for these new readers. 1024 1025<table> 1026<tr><th> </th></tr> 1027<tr><th align="left">Quick Quiz:</th></tr> 1028<tr><td> 1029 Suppose that synchronize_rcu() did wait until <i>all</i> 1030 readers had completed instead of waiting only on 1031 pre-existing readers. 1032 For how long would the updater be able to rely on there 1033 being no readers? 1034</td></tr> 1035<tr><th align="left">Answer:</th></tr> 1036<tr><td bgcolor="#ffffff"><font color="ffffff"> 1037 For no time at all. 1038 Even if <tt>synchronize_rcu()</tt> were to wait until 1039 all readers had completed, a new reader might start immediately after 1040 <tt>synchronize_rcu()</tt> completed. 1041 Therefore, the code following 1042 <tt>synchronize_rcu()</tt> can <i>never</i> rely on there being 1043 no readers. 1044</font></td></tr> 1045<tr><td> </td></tr> 1046</table> 1047 1048<h3><a name="Grace Periods Don't Partition Read-Side Critical Sections"> 1049Grace Periods Don't Partition Read-Side Critical Sections</a></h3> 1050 1051<p> 1052It is tempting to assume that if any part of one RCU read-side critical 1053section precedes a given grace period, and if any part of another RCU 1054read-side critical section follows that same grace period, then all of 1055the first RCU read-side critical section must precede all of the second. 1056However, this just isn't the case: A single grace period does not 1057partition the set of RCU read-side critical sections. 1058An example of this situation can be illustrated as follows, where 1059<tt>x</tt>, <tt>y</tt>, and <tt>z</tt> are initially all zero: 1060 1061<blockquote> 1062<pre> 1063 1 void thread0(void) 1064 2 { 1065 3 rcu_read_lock(); 1066 4 WRITE_ONCE(a, 1); 1067 5 WRITE_ONCE(b, 1); 1068 6 rcu_read_unlock(); 1069 7 } 1070 8 1071 9 void thread1(void) 107210 { 107311 r1 = READ_ONCE(a); 107412 synchronize_rcu(); 107513 WRITE_ONCE(c, 1); 107614 } 107715 107816 void thread2(void) 107917 { 108018 rcu_read_lock(); 108119 r2 = READ_ONCE(b); 108220 r3 = READ_ONCE(c); 108321 rcu_read_unlock(); 108422 } 1085</pre> 1086</blockquote> 1087 1088<p> 1089It turns out that the outcome: 1090 1091<blockquote> 1092<pre> 1093(r1 == 1 && r2 == 0 && r3 == 1) 1094</pre> 1095</blockquote> 1096 1097is entirely possible. 1098The following figure show how this can happen, with each circled 1099<tt>QS</tt> indicating the point at which RCU recorded a 1100<i>quiescent state</i> for each thread, that is, a state in which 1101RCU knows that the thread cannot be in the midst of an RCU read-side 1102critical section that started before the current grace period: 1103 1104<p><img src="GPpartitionReaders1.svg" alt="GPpartitionReaders1.svg" width="60%"></p> 1105 1106<p> 1107If it is necessary to partition RCU read-side critical sections in this 1108manner, it is necessary to use two grace periods, where the first 1109grace period is known to end before the second grace period starts: 1110 1111<blockquote> 1112<pre> 1113 1 void thread0(void) 1114 2 { 1115 3 rcu_read_lock(); 1116 4 WRITE_ONCE(a, 1); 1117 5 WRITE_ONCE(b, 1); 1118 6 rcu_read_unlock(); 1119 7 } 1120 8 1121 9 void thread1(void) 112210 { 112311 r1 = READ_ONCE(a); 112412 synchronize_rcu(); 112513 WRITE_ONCE(c, 1); 112614 } 112715 112816 void thread2(void) 112917 { 113018 r2 = READ_ONCE(c); 113119 synchronize_rcu(); 113220 WRITE_ONCE(d, 1); 113321 } 113422 113523 void thread3(void) 113624 { 113725 rcu_read_lock(); 113826 r3 = READ_ONCE(b); 113927 r4 = READ_ONCE(d); 114028 rcu_read_unlock(); 114129 } 1142</pre> 1143</blockquote> 1144 1145<p> 1146Here, if <tt>(r1 == 1)</tt>, then 1147<tt>thread0()</tt>'s write to <tt>b</tt> must happen 1148before the end of <tt>thread1()</tt>'s grace period. 1149If in addition <tt>(r4 == 1)</tt>, then 1150<tt>thread3()</tt>'s read from <tt>b</tt> must happen 1151after the beginning of <tt>thread2()</tt>'s grace period. 1152If it is also the case that <tt>(r2 == 1)</tt>, then the 1153end of <tt>thread1()</tt>'s grace period must precede the 1154beginning of <tt>thread2()</tt>'s grace period. 1155This mean that the two RCU read-side critical sections cannot overlap, 1156guaranteeing that <tt>(r3 == 1)</tt>. 1157As a result, the outcome: 1158 1159<blockquote> 1160<pre> 1161(r1 == 1 && r2 == 1 && r3 == 0 && r4 == 1) 1162</pre> 1163</blockquote> 1164 1165cannot happen. 1166 1167<p> 1168This non-requirement was also non-premeditated, but became apparent 1169when studying RCU's interaction with memory ordering. 1170 1171<h3><a name="Read-Side Critical Sections Don't Partition Grace Periods"> 1172Read-Side Critical Sections Don't Partition Grace Periods</a></h3> 1173 1174<p> 1175It is also tempting to assume that if an RCU read-side critical section 1176happens between a pair of grace periods, then those grace periods cannot 1177overlap. 1178However, this temptation leads nowhere good, as can be illustrated by 1179the following, with all variables initially zero: 1180 1181<blockquote> 1182<pre> 1183 1 void thread0(void) 1184 2 { 1185 3 rcu_read_lock(); 1186 4 WRITE_ONCE(a, 1); 1187 5 WRITE_ONCE(b, 1); 1188 6 rcu_read_unlock(); 1189 7 } 1190 8 1191 9 void thread1(void) 119210 { 119311 r1 = READ_ONCE(a); 119412 synchronize_rcu(); 119513 WRITE_ONCE(c, 1); 119614 } 119715 119816 void thread2(void) 119917 { 120018 rcu_read_lock(); 120119 WRITE_ONCE(d, 1); 120220 r2 = READ_ONCE(c); 120321 rcu_read_unlock(); 120422 } 120523 120624 void thread3(void) 120725 { 120826 r3 = READ_ONCE(d); 120927 synchronize_rcu(); 121028 WRITE_ONCE(e, 1); 121129 } 121230 121331 void thread4(void) 121432 { 121533 rcu_read_lock(); 121634 r4 = READ_ONCE(b); 121735 r5 = READ_ONCE(e); 121836 rcu_read_unlock(); 121937 } 1220</pre> 1221</blockquote> 1222 1223<p> 1224In this case, the outcome: 1225 1226<blockquote> 1227<pre> 1228(r1 == 1 && r2 == 1 && r3 == 1 && r4 == 0 && r5 == 1) 1229</pre> 1230</blockquote> 1231 1232is entirely possible, as illustrated below: 1233 1234<p><img src="ReadersPartitionGP1.svg" alt="ReadersPartitionGP1.svg" width="100%"></p> 1235 1236<p> 1237Again, an RCU read-side critical section can overlap almost all of a 1238given grace period, just so long as it does not overlap the entire 1239grace period. 1240As a result, an RCU read-side critical section cannot partition a pair 1241of RCU grace periods. 1242 1243<table> 1244<tr><th> </th></tr> 1245<tr><th align="left">Quick Quiz:</th></tr> 1246<tr><td> 1247 How long a sequence of grace periods, each separated by an RCU 1248 read-side critical section, would be required to partition the RCU 1249 read-side critical sections at the beginning and end of the chain? 1250</td></tr> 1251<tr><th align="left">Answer:</th></tr> 1252<tr><td bgcolor="#ffffff"><font color="ffffff"> 1253 In theory, an infinite number. 1254 In practice, an unknown number that is sensitive to both implementation 1255 details and timing considerations. 1256 Therefore, even in practice, RCU users must abide by the 1257 theoretical rather than the practical answer. 1258</font></td></tr> 1259<tr><td> </td></tr> 1260</table> 1261 1262<h3><a name="Disabling Preemption Does Not Block Grace Periods"> 1263Disabling Preemption Does Not Block Grace Periods</a></h3> 1264 1265<p> 1266There was a time when disabling preemption on any given CPU would block 1267subsequent grace periods. 1268However, this was an accident of implementation and is not a requirement. 1269And in the current Linux-kernel implementation, disabling preemption 1270on a given CPU in fact does not block grace periods, as Oleg Nesterov 1271<a href="https://lkml.kernel.org/g/20150614193825.GA19582@redhat.com">demonstrated</a>. 1272 1273<p> 1274If you need a preempt-disable region to block grace periods, you need to add 1275<tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>, for example 1276as follows: 1277 1278<blockquote> 1279<pre> 1280 1 preempt_disable(); 1281 2 rcu_read_lock(); 1282 3 do_something(); 1283 4 rcu_read_unlock(); 1284 5 preempt_enable(); 1285 6 1286 7 /* Spinlocks implicitly disable preemption. */ 1287 8 spin_lock(&mylock); 1288 9 rcu_read_lock(); 128910 do_something(); 129011 rcu_read_unlock(); 129112 spin_unlock(&mylock); 1292</pre> 1293</blockquote> 1294 1295<p> 1296In theory, you could enter the RCU read-side critical section first, 1297but it is more efficient to keep the entire RCU read-side critical 1298section contained in the preempt-disable region as shown above. 1299Of course, RCU read-side critical sections that extend outside of 1300preempt-disable regions will work correctly, but such critical sections 1301can be preempted, which forces <tt>rcu_read_unlock()</tt> to do 1302more work. 1303And no, this is <i>not</i> an invitation to enclose all of your RCU 1304read-side critical sections within preempt-disable regions, because 1305doing so would degrade real-time response. 1306 1307<p> 1308This non-requirement appeared with preemptible RCU. 1309If you need a grace period that waits on non-preemptible code regions, use 1310<a href="#Sched Flavor">RCU-sched</a>. 1311 1312<h2><a name="Parallelism Facts of Life">Parallelism Facts of Life</a></h2> 1313 1314<p> 1315These parallelism facts of life are by no means specific to RCU, but 1316the RCU implementation must abide by them. 1317They therefore bear repeating: 1318 1319<ol> 1320<li> Any CPU or task may be delayed at any time, 1321 and any attempts to avoid these delays by disabling 1322 preemption, interrupts, or whatever are completely futile. 1323 This is most obvious in preemptible user-level 1324 environments and in virtualized environments (where 1325 a given guest OS's VCPUs can be preempted at any time by 1326 the underlying hypervisor), but can also happen in bare-metal 1327 environments due to ECC errors, NMIs, and other hardware 1328 events. 1329 Although a delay of more than about 20 seconds can result 1330 in splats, the RCU implementation is obligated to use 1331 algorithms that can tolerate extremely long delays, but where 1332 “extremely long” is not long enough to allow 1333 wrap-around when incrementing a 64-bit counter. 1334<li> Both the compiler and the CPU can reorder memory accesses. 1335 Where it matters, RCU must use compiler directives and 1336 memory-barrier instructions to preserve ordering. 1337<li> Conflicting writes to memory locations in any given cache line 1338 will result in expensive cache misses. 1339 Greater numbers of concurrent writes and more-frequent 1340 concurrent writes will result in more dramatic slowdowns. 1341 RCU is therefore obligated to use algorithms that have 1342 sufficient locality to avoid significant performance and 1343 scalability problems. 1344<li> As a rough rule of thumb, only one CPU's worth of processing 1345 may be carried out under the protection of any given exclusive 1346 lock. 1347 RCU must therefore use scalable locking designs. 1348<li> Counters are finite, especially on 32-bit systems. 1349 RCU's use of counters must therefore tolerate counter wrap, 1350 or be designed such that counter wrap would take way more 1351 time than a single system is likely to run. 1352 An uptime of ten years is quite possible, a runtime 1353 of a century much less so. 1354 As an example of the latter, RCU's dyntick-idle nesting counter 1355 allows 54 bits for interrupt nesting level (this counter 1356 is 64 bits even on a 32-bit system). 1357 Overflowing this counter requires 2<sup>54</sup> 1358 half-interrupts on a given CPU without that CPU ever going idle. 1359 If a half-interrupt happened every microsecond, it would take 1360 570 years of runtime to overflow this counter, which is currently 1361 believed to be an acceptably long time. 1362<li> Linux systems can have thousands of CPUs running a single 1363 Linux kernel in a single shared-memory environment. 1364 RCU must therefore pay close attention to high-end scalability. 1365</ol> 1366 1367<p> 1368This last parallelism fact of life means that RCU must pay special 1369attention to the preceding facts of life. 1370The idea that Linux might scale to systems with thousands of CPUs would 1371have been met with some skepticism in the 1990s, but these requirements 1372would have otherwise have been unsurprising, even in the early 1990s. 1373 1374<h2><a name="Quality-of-Implementation Requirements">Quality-of-Implementation Requirements</a></h2> 1375 1376<p> 1377These sections list quality-of-implementation requirements. 1378Although an RCU implementation that ignores these requirements could 1379still be used, it would likely be subject to limitations that would 1380make it inappropriate for industrial-strength production use. 1381Classes of quality-of-implementation requirements are as follows: 1382 1383<ol> 1384<li> <a href="#Specialization">Specialization</a> 1385<li> <a href="#Performance and Scalability">Performance and Scalability</a> 1386<li> <a href="#Composability">Composability</a> 1387<li> <a href="#Corner Cases">Corner Cases</a> 1388</ol> 1389 1390<p> 1391These classes is covered in the following sections. 1392 1393<h3><a name="Specialization">Specialization</a></h3> 1394 1395<p> 1396RCU is and always has been intended primarily for read-mostly situations, 1397which means that RCU's read-side primitives are optimized, often at the 1398expense of its update-side primitives. 1399Experience thus far is captured by the following list of situations: 1400 1401<ol> 1402<li> Read-mostly data, where stale and inconsistent data is not 1403 a problem: RCU works great! 1404<li> Read-mostly data, where data must be consistent: 1405 RCU works well. 1406<li> Read-write data, where data must be consistent: 1407 RCU <i>might</i> work OK. 1408 Or not. 1409<li> Write-mostly data, where data must be consistent: 1410 RCU is very unlikely to be the right tool for the job, 1411 with the following exceptions, where RCU can provide: 1412 <ol type=a> 1413 <li> Existence guarantees for update-friendly mechanisms. 1414 <li> Wait-free read-side primitives for real-time use. 1415 </ol> 1416</ol> 1417 1418<p> 1419This focus on read-mostly situations means that RCU must interoperate 1420with other synchronization primitives. 1421For example, the <tt>add_gp()</tt> and <tt>remove_gp_synchronous()</tt> 1422examples discussed earlier use RCU to protect readers and locking to 1423coordinate updaters. 1424However, the need extends much farther, requiring that a variety of 1425synchronization primitives be legal within RCU read-side critical sections, 1426including spinlocks, sequence locks, atomic operations, reference 1427counters, and memory barriers. 1428 1429<table> 1430<tr><th> </th></tr> 1431<tr><th align="left">Quick Quiz:</th></tr> 1432<tr><td> 1433 What about sleeping locks? 1434</td></tr> 1435<tr><th align="left">Answer:</th></tr> 1436<tr><td bgcolor="#ffffff"><font color="ffffff"> 1437 These are forbidden within Linux-kernel RCU read-side critical 1438 sections because it is not legal to place a quiescent state 1439 (in this case, voluntary context switch) within an RCU read-side 1440 critical section. 1441 However, sleeping locks may be used within userspace RCU read-side 1442 critical sections, and also within Linux-kernel sleepable RCU 1443 <a href="#Sleepable RCU"><font color="ffffff">(SRCU)</font></a> 1444 read-side critical sections. 1445 In addition, the -rt patchset turns spinlocks into a 1446 sleeping locks so that the corresponding critical sections 1447 can be preempted, which also means that these sleeplockified 1448 spinlocks (but not other sleeping locks!) may be acquire within 1449 -rt-Linux-kernel RCU read-side critical sections. 1450 </font> 1451 1452 <p><font color="ffffff"> 1453 Note that it <i>is</i> legal for a normal RCU read-side 1454 critical section to conditionally acquire a sleeping locks 1455 (as in <tt>mutex_trylock()</tt>), but only as long as it does 1456 not loop indefinitely attempting to conditionally acquire that 1457 sleeping locks. 1458 The key point is that things like <tt>mutex_trylock()</tt> 1459 either return with the mutex held, or return an error indication if 1460 the mutex was not immediately available. 1461 Either way, <tt>mutex_trylock()</tt> returns immediately without 1462 sleeping. 1463</font></td></tr> 1464<tr><td> </td></tr> 1465</table> 1466 1467<p> 1468It often comes as a surprise that many algorithms do not require a 1469consistent view of data, but many can function in that mode, 1470with network routing being the poster child. 1471Internet routing algorithms take significant time to propagate 1472updates, so that by the time an update arrives at a given system, 1473that system has been sending network traffic the wrong way for 1474a considerable length of time. 1475Having a few threads continue to send traffic the wrong way for a 1476few more milliseconds is clearly not a problem: In the worst case, 1477TCP retransmissions will eventually get the data where it needs to go. 1478In general, when tracking the state of the universe outside of the 1479computer, some level of inconsistency must be tolerated due to 1480speed-of-light delays if nothing else. 1481 1482<p> 1483Furthermore, uncertainty about external state is inherent in many cases. 1484For example, a pair of veterinarians might use heartbeat to determine 1485whether or not a given cat was alive. 1486But how long should they wait after the last heartbeat to decide that 1487the cat is in fact dead? 1488Waiting less than 400 milliseconds makes no sense because this would 1489mean that a relaxed cat would be considered to cycle between death 1490and life more than 100 times per minute. 1491Moreover, just as with human beings, a cat's heart might stop for 1492some period of time, so the exact wait period is a judgment call. 1493One of our pair of veterinarians might wait 30 seconds before pronouncing 1494the cat dead, while the other might insist on waiting a full minute. 1495The two veterinarians would then disagree on the state of the cat during 1496the final 30 seconds of the minute following the last heartbeat. 1497 1498<p> 1499Interestingly enough, this same situation applies to hardware. 1500When push comes to shove, how do we tell whether or not some 1501external server has failed? 1502We send messages to it periodically, and declare it failed if we 1503don't receive a response within a given period of time. 1504Policy decisions can usually tolerate short 1505periods of inconsistency. 1506The policy was decided some time ago, and is only now being put into 1507effect, so a few milliseconds of delay is normally inconsequential. 1508 1509<p> 1510However, there are algorithms that absolutely must see consistent data. 1511For example, the translation between a user-level SystemV semaphore 1512ID to the corresponding in-kernel data structure is protected by RCU, 1513but it is absolutely forbidden to update a semaphore that has just been 1514removed. 1515In the Linux kernel, this need for consistency is accommodated by acquiring 1516spinlocks located in the in-kernel data structure from within 1517the RCU read-side critical section, and this is indicated by the 1518green box in the figure above. 1519Many other techniques may be used, and are in fact used within the 1520Linux kernel. 1521 1522<p> 1523In short, RCU is not required to maintain consistency, and other 1524mechanisms may be used in concert with RCU when consistency is required. 1525RCU's specialization allows it to do its job extremely well, and its 1526ability to interoperate with other synchronization mechanisms allows 1527the right mix of synchronization tools to be used for a given job. 1528 1529<h3><a name="Performance and Scalability">Performance and Scalability</a></h3> 1530 1531<p> 1532Energy efficiency is a critical component of performance today, 1533and Linux-kernel RCU implementations must therefore avoid unnecessarily 1534awakening idle CPUs. 1535I cannot claim that this requirement was premeditated. 1536In fact, I learned of it during a telephone conversation in which I 1537was given “frank and open” feedback on the importance 1538of energy efficiency in battery-powered systems and on specific 1539energy-efficiency shortcomings of the Linux-kernel RCU implementation. 1540In my experience, the battery-powered embedded community will consider 1541any unnecessary wakeups to be extremely unfriendly acts. 1542So much so that mere Linux-kernel-mailing-list posts are 1543insufficient to vent their ire. 1544 1545<p> 1546Memory consumption is not particularly important for in most 1547situations, and has become decreasingly 1548so as memory sizes have expanded and memory 1549costs have plummeted. 1550However, as I learned from Matt Mackall's 1551<a href="http://elinux.org/Linux_Tiny-FAQ">bloatwatch</a> 1552efforts, memory footprint is critically important on single-CPU systems with 1553non-preemptible (<tt>CONFIG_PREEMPT=n</tt>) kernels, and thus 1554<a href="https://lkml.kernel.org/g/20090113221724.GA15307@linux.vnet.ibm.com">tiny RCU</a> 1555was born. 1556Josh Triplett has since taken over the small-memory banner with his 1557<a href="https://tiny.wiki.kernel.org/">Linux kernel tinification</a> 1558project, which resulted in 1559<a href="#Sleepable RCU">SRCU</a> 1560becoming optional for those kernels not needing it. 1561 1562<p> 1563The remaining performance requirements are, for the most part, 1564unsurprising. 1565For example, in keeping with RCU's read-side specialization, 1566<tt>rcu_dereference()</tt> should have negligible overhead (for 1567example, suppression of a few minor compiler optimizations). 1568Similarly, in non-preemptible environments, <tt>rcu_read_lock()</tt> and 1569<tt>rcu_read_unlock()</tt> should have exactly zero overhead. 1570 1571<p> 1572In preemptible environments, in the case where the RCU read-side 1573critical section was not preempted (as will be the case for the 1574highest-priority real-time process), <tt>rcu_read_lock()</tt> and 1575<tt>rcu_read_unlock()</tt> should have minimal overhead. 1576In particular, they should not contain atomic read-modify-write 1577operations, memory-barrier instructions, preemption disabling, 1578interrupt disabling, or backwards branches. 1579However, in the case where the RCU read-side critical section was preempted, 1580<tt>rcu_read_unlock()</tt> may acquire spinlocks and disable interrupts. 1581This is why it is better to nest an RCU read-side critical section 1582within a preempt-disable region than vice versa, at least in cases 1583where that critical section is short enough to avoid unduly degrading 1584real-time latencies. 1585 1586<p> 1587The <tt>synchronize_rcu()</tt> grace-period-wait primitive is 1588optimized for throughput. 1589It may therefore incur several milliseconds of latency in addition to 1590the duration of the longest RCU read-side critical section. 1591On the other hand, multiple concurrent invocations of 1592<tt>synchronize_rcu()</tt> are required to use batching optimizations 1593so that they can be satisfied by a single underlying grace-period-wait 1594operation. 1595For example, in the Linux kernel, it is not unusual for a single 1596grace-period-wait operation to serve more than 1597<a href="https://www.usenix.org/conference/2004-usenix-annual-technical-conference/making-rcu-safe-deep-sub-millisecond-response">1,000 separate invocations</a> 1598of <tt>synchronize_rcu()</tt>, thus amortizing the per-invocation 1599overhead down to nearly zero. 1600However, the grace-period optimization is also required to avoid 1601measurable degradation of real-time scheduling and interrupt latencies. 1602 1603<p> 1604In some cases, the multi-millisecond <tt>synchronize_rcu()</tt> 1605latencies are unacceptable. 1606In these cases, <tt>synchronize_rcu_expedited()</tt> may be used 1607instead, reducing the grace-period latency down to a few tens of 1608microseconds on small systems, at least in cases where the RCU read-side 1609critical sections are short. 1610There are currently no special latency requirements for 1611<tt>synchronize_rcu_expedited()</tt> on large systems, but, 1612consistent with the empirical nature of the RCU specification, 1613that is subject to change. 1614However, there most definitely are scalability requirements: 1615A storm of <tt>synchronize_rcu_expedited()</tt> invocations on 4096 1616CPUs should at least make reasonable forward progress. 1617In return for its shorter latencies, <tt>synchronize_rcu_expedited()</tt> 1618is permitted to impose modest degradation of real-time latency 1619on non-idle online CPUs. 1620Here, “modest” means roughly the same latency 1621degradation as a scheduling-clock interrupt. 1622 1623<p> 1624There are a number of situations where even 1625<tt>synchronize_rcu_expedited()</tt>'s reduced grace-period 1626latency is unacceptable. 1627In these situations, the asynchronous <tt>call_rcu()</tt> can be 1628used in place of <tt>synchronize_rcu()</tt> as follows: 1629 1630<blockquote> 1631<pre> 1632 1 struct foo { 1633 2 int a; 1634 3 int b; 1635 4 struct rcu_head rh; 1636 5 }; 1637 6 1638 7 static void remove_gp_cb(struct rcu_head *rhp) 1639 8 { 1640 9 struct foo *p = container_of(rhp, struct foo, rh); 164110 164211 kfree(p); 164312 } 164413 164514 bool remove_gp_asynchronous(void) 164615 { 164716 struct foo *p; 164817 164918 spin_lock(&gp_lock); 165019 p = rcu_dereference(gp); 165120 if (!p) { 165221 spin_unlock(&gp_lock); 165322 return false; 165423 } 165524 rcu_assign_pointer(gp, NULL); 165625 call_rcu(&p->rh, remove_gp_cb); 165726 spin_unlock(&gp_lock); 165827 return true; 165928 } 1660</pre> 1661</blockquote> 1662 1663<p> 1664A definition of <tt>struct foo</tt> is finally needed, and appears 1665on lines 1-5. 1666The function <tt>remove_gp_cb()</tt> is passed to <tt>call_rcu()</tt> 1667on line 25, and will be invoked after the end of a subsequent 1668grace period. 1669This gets the same effect as <tt>remove_gp_synchronous()</tt>, 1670but without forcing the updater to wait for a grace period to elapse. 1671The <tt>call_rcu()</tt> function may be used in a number of 1672situations where neither <tt>synchronize_rcu()</tt> nor 1673<tt>synchronize_rcu_expedited()</tt> would be legal, 1674including within preempt-disable code, <tt>local_bh_disable()</tt> code, 1675interrupt-disable code, and interrupt handlers. 1676However, even <tt>call_rcu()</tt> is illegal within NMI handlers 1677and from idle and offline CPUs. 1678The callback function (<tt>remove_gp_cb()</tt> in this case) will be 1679executed within softirq (software interrupt) environment within the 1680Linux kernel, 1681either within a real softirq handler or under the protection 1682of <tt>local_bh_disable()</tt>. 1683In both the Linux kernel and in userspace, it is bad practice to 1684write an RCU callback function that takes too long. 1685Long-running operations should be relegated to separate threads or 1686(in the Linux kernel) workqueues. 1687 1688<table> 1689<tr><th> </th></tr> 1690<tr><th align="left">Quick Quiz:</th></tr> 1691<tr><td> 1692 Why does line 19 use <tt>rcu_access_pointer()</tt>? 1693 After all, <tt>call_rcu()</tt> on line 25 stores into the 1694 structure, which would interact badly with concurrent insertions. 1695 Doesn't this mean that <tt>rcu_dereference()</tt> is required? 1696</td></tr> 1697<tr><th align="left">Answer:</th></tr> 1698<tr><td bgcolor="#ffffff"><font color="ffffff"> 1699 Presumably the <tt>->gp_lock</tt> acquired on line 18 excludes 1700 any changes, including any insertions that <tt>rcu_dereference()</tt> 1701 would protect against. 1702 Therefore, any insertions will be delayed until after 1703 <tt>->gp_lock</tt> 1704 is released on line 25, which in turn means that 1705 <tt>rcu_access_pointer()</tt> suffices. 1706</font></td></tr> 1707<tr><td> </td></tr> 1708</table> 1709 1710<p> 1711However, all that <tt>remove_gp_cb()</tt> is doing is 1712invoking <tt>kfree()</tt> on the data element. 1713This is a common idiom, and is supported by <tt>kfree_rcu()</tt>, 1714which allows “fire and forget” operation as shown below: 1715 1716<blockquote> 1717<pre> 1718 1 struct foo { 1719 2 int a; 1720 3 int b; 1721 4 struct rcu_head rh; 1722 5 }; 1723 6 1724 7 bool remove_gp_faf(void) 1725 8 { 1726 9 struct foo *p; 172710 172811 spin_lock(&gp_lock); 172912 p = rcu_dereference(gp); 173013 if (!p) { 173114 spin_unlock(&gp_lock); 173215 return false; 173316 } 173417 rcu_assign_pointer(gp, NULL); 173518 kfree_rcu(p, rh); 173619 spin_unlock(&gp_lock); 173720 return true; 173821 } 1739</pre> 1740</blockquote> 1741 1742<p> 1743Note that <tt>remove_gp_faf()</tt> simply invokes 1744<tt>kfree_rcu()</tt> and proceeds, without any need to pay any 1745further attention to the subsequent grace period and <tt>kfree()</tt>. 1746It is permissible to invoke <tt>kfree_rcu()</tt> from the same 1747environments as for <tt>call_rcu()</tt>. 1748Interestingly enough, DYNIX/ptx had the equivalents of 1749<tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>, but not 1750<tt>synchronize_rcu()</tt>. 1751This was due to the fact that RCU was not heavily used within DYNIX/ptx, 1752so the very few places that needed something like 1753<tt>synchronize_rcu()</tt> simply open-coded it. 1754 1755<table> 1756<tr><th> </th></tr> 1757<tr><th align="left">Quick Quiz:</th></tr> 1758<tr><td> 1759 Earlier it was claimed that <tt>call_rcu()</tt> and 1760 <tt>kfree_rcu()</tt> allowed updaters to avoid being blocked 1761 by readers. 1762 But how can that be correct, given that the invocation of the callback 1763 and the freeing of the memory (respectively) must still wait for 1764 a grace period to elapse? 1765</td></tr> 1766<tr><th align="left">Answer:</th></tr> 1767<tr><td bgcolor="#ffffff"><font color="ffffff"> 1768 We could define things this way, but keep in mind that this sort of 1769 definition would say that updates in garbage-collected languages 1770 cannot complete until the next time the garbage collector runs, 1771 which does not seem at all reasonable. 1772 The key point is that in most cases, an updater using either 1773 <tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> can proceed to the 1774 next update as soon as it has invoked <tt>call_rcu()</tt> or 1775 <tt>kfree_rcu()</tt>, without having to wait for a subsequent 1776 grace period. 1777</font></td></tr> 1778<tr><td> </td></tr> 1779</table> 1780 1781<p> 1782But what if the updater must wait for the completion of code to be 1783executed after the end of the grace period, but has other tasks 1784that can be carried out in the meantime? 1785The polling-style <tt>get_state_synchronize_rcu()</tt> and 1786<tt>cond_synchronize_rcu()</tt> functions may be used for this 1787purpose, as shown below: 1788 1789<blockquote> 1790<pre> 1791 1 bool remove_gp_poll(void) 1792 2 { 1793 3 struct foo *p; 1794 4 unsigned long s; 1795 5 1796 6 spin_lock(&gp_lock); 1797 7 p = rcu_access_pointer(gp); 1798 8 if (!p) { 1799 9 spin_unlock(&gp_lock); 180010 return false; 180111 } 180212 rcu_assign_pointer(gp, NULL); 180313 spin_unlock(&gp_lock); 180414 s = get_state_synchronize_rcu(); 180515 do_something_while_waiting(); 180616 cond_synchronize_rcu(s); 180717 kfree(p); 180818 return true; 180919 } 1810</pre> 1811</blockquote> 1812 1813<p> 1814On line 14, <tt>get_state_synchronize_rcu()</tt> obtains a 1815“cookie” from RCU, 1816then line 15 carries out other tasks, 1817and finally, line 16 returns immediately if a grace period has 1818elapsed in the meantime, but otherwise waits as required. 1819The need for <tt>get_state_synchronize_rcu</tt> and 1820<tt>cond_synchronize_rcu()</tt> has appeared quite recently, 1821so it is too early to tell whether they will stand the test of time. 1822 1823<p> 1824RCU thus provides a range of tools to allow updaters to strike the 1825required tradeoff between latency, flexibility and CPU overhead. 1826 1827<h3><a name="Composability">Composability</a></h3> 1828 1829<p> 1830Composability has received much attention in recent years, perhaps in part 1831due to the collision of multicore hardware with object-oriented techniques 1832designed in single-threaded environments for single-threaded use. 1833And in theory, RCU read-side critical sections may be composed, and in 1834fact may be nested arbitrarily deeply. 1835In practice, as with all real-world implementations of composable 1836constructs, there are limitations. 1837 1838<p> 1839Implementations of RCU for which <tt>rcu_read_lock()</tt> 1840and <tt>rcu_read_unlock()</tt> generate no code, such as 1841Linux-kernel RCU when <tt>CONFIG_PREEMPT=n</tt>, can be 1842nested arbitrarily deeply. 1843After all, there is no overhead. 1844Except that if all these instances of <tt>rcu_read_lock()</tt> 1845and <tt>rcu_read_unlock()</tt> are visible to the compiler, 1846compilation will eventually fail due to exhausting memory, 1847mass storage, or user patience, whichever comes first. 1848If the nesting is not visible to the compiler, as is the case with 1849mutually recursive functions each in its own translation unit, 1850stack overflow will result. 1851If the nesting takes the form of loops, perhaps in the guise of tail 1852recursion, either the control variable 1853will overflow or (in the Linux kernel) you will get an RCU CPU stall warning. 1854Nevertheless, this class of RCU implementations is one 1855of the most composable constructs in existence. 1856 1857<p> 1858RCU implementations that explicitly track nesting depth 1859are limited by the nesting-depth counter. 1860For example, the Linux kernel's preemptible RCU limits nesting to 1861<tt>INT_MAX</tt>. 1862This should suffice for almost all practical purposes. 1863That said, a consecutive pair of RCU read-side critical sections 1864between which there is an operation that waits for a grace period 1865cannot be enclosed in another RCU read-side critical section. 1866This is because it is not legal to wait for a grace period within 1867an RCU read-side critical section: To do so would result either 1868in deadlock or 1869in RCU implicitly splitting the enclosing RCU read-side critical 1870section, neither of which is conducive to a long-lived and prosperous 1871kernel. 1872 1873<p> 1874It is worth noting that RCU is not alone in limiting composability. 1875For example, many transactional-memory implementations prohibit 1876composing a pair of transactions separated by an irrevocable 1877operation (for example, a network receive operation). 1878For another example, lock-based critical sections can be composed 1879surprisingly freely, but only if deadlock is avoided. 1880 1881<p> 1882In short, although RCU read-side critical sections are highly composable, 1883care is required in some situations, just as is the case for any other 1884composable synchronization mechanism. 1885 1886<h3><a name="Corner Cases">Corner Cases</a></h3> 1887 1888<p> 1889A given RCU workload might have an endless and intense stream of 1890RCU read-side critical sections, perhaps even so intense that there 1891was never a point in time during which there was not at least one 1892RCU read-side critical section in flight. 1893RCU cannot allow this situation to block grace periods: As long as 1894all the RCU read-side critical sections are finite, grace periods 1895must also be finite. 1896 1897<p> 1898That said, preemptible RCU implementations could potentially result 1899in RCU read-side critical sections being preempted for long durations, 1900which has the effect of creating a long-duration RCU read-side 1901critical section. 1902This situation can arise only in heavily loaded systems, but systems using 1903real-time priorities are of course more vulnerable. 1904Therefore, RCU priority boosting is provided to help deal with this 1905case. 1906That said, the exact requirements on RCU priority boosting will likely 1907evolve as more experience accumulates. 1908 1909<p> 1910Other workloads might have very high update rates. 1911Although one can argue that such workloads should instead use 1912something other than RCU, the fact remains that RCU must 1913handle such workloads gracefully. 1914This requirement is another factor driving batching of grace periods, 1915but it is also the driving force behind the checks for large numbers 1916of queued RCU callbacks in the <tt>call_rcu()</tt> code path. 1917Finally, high update rates should not delay RCU read-side critical 1918sections, although some small read-side delays can occur when using 1919<tt>synchronize_rcu_expedited()</tt>, courtesy of this function's use 1920of <tt>smp_call_function_single()</tt>. 1921 1922<p> 1923Although all three of these corner cases were understood in the early 19241990s, a simple user-level test consisting of <tt>close(open(path))</tt> 1925in a tight loop 1926in the early 2000s suddenly provided a much deeper appreciation of the 1927high-update-rate corner case. 1928This test also motivated addition of some RCU code to react to high update 1929rates, for example, if a given CPU finds itself with more than 10,000 1930RCU callbacks queued, it will cause RCU to take evasive action by 1931more aggressively starting grace periods and more aggressively forcing 1932completion of grace-period processing. 1933This evasive action causes the grace period to complete more quickly, 1934but at the cost of restricting RCU's batching optimizations, thus 1935increasing the CPU overhead incurred by that grace period. 1936 1937<h2><a name="Software-Engineering Requirements"> 1938Software-Engineering Requirements</a></h2> 1939 1940<p> 1941Between Murphy's Law and “To err is human”, it is necessary to 1942guard against mishaps and misuse: 1943 1944<ol> 1945<li> It is all too easy to forget to use <tt>rcu_read_lock()</tt> 1946 everywhere that it is needed, so kernels built with 1947 <tt>CONFIG_PROVE_RCU=y</tt> will splat if 1948 <tt>rcu_dereference()</tt> is used outside of an 1949 RCU read-side critical section. 1950 Update-side code can use <tt>rcu_dereference_protected()</tt>, 1951 which takes a 1952 <a href="https://lwn.net/Articles/371986/">lockdep expression</a> 1953 to indicate what is providing the protection. 1954 If the indicated protection is not provided, a lockdep splat 1955 is emitted. 1956 1957 <p> 1958 Code shared between readers and updaters can use 1959 <tt>rcu_dereference_check()</tt>, which also takes a 1960 lockdep expression, and emits a lockdep splat if neither 1961 <tt>rcu_read_lock()</tt> nor the indicated protection 1962 is in place. 1963 In addition, <tt>rcu_dereference_raw()</tt> is used in those 1964 (hopefully rare) cases where the required protection cannot 1965 be easily described. 1966 Finally, <tt>rcu_read_lock_held()</tt> is provided to 1967 allow a function to verify that it has been invoked within 1968 an RCU read-side critical section. 1969 I was made aware of this set of requirements shortly after Thomas 1970 Gleixner audited a number of RCU uses. 1971<li> A given function might wish to check for RCU-related preconditions 1972 upon entry, before using any other RCU API. 1973 The <tt>rcu_lockdep_assert()</tt> does this job, 1974 asserting the expression in kernels having lockdep enabled 1975 and doing nothing otherwise. 1976<li> It is also easy to forget to use <tt>rcu_assign_pointer()</tt> 1977 and <tt>rcu_dereference()</tt>, perhaps (incorrectly) 1978 substituting a simple assignment. 1979 To catch this sort of error, a given RCU-protected pointer may be 1980 tagged with <tt>__rcu</tt>, after which sparse 1981 will complain about simple-assignment accesses to that pointer. 1982 Arnd Bergmann made me aware of this requirement, and also 1983 supplied the needed 1984 <a href="https://lwn.net/Articles/376011/">patch series</a>. 1985<li> Kernels built with <tt>CONFIG_DEBUG_OBJECTS_RCU_HEAD=y</tt> 1986 will splat if a data element is passed to <tt>call_rcu()</tt> 1987 twice in a row, without a grace period in between. 1988 (This error is similar to a double free.) 1989 The corresponding <tt>rcu_head</tt> structures that are 1990 dynamically allocated are automatically tracked, but 1991 <tt>rcu_head</tt> structures allocated on the stack 1992 must be initialized with <tt>init_rcu_head_on_stack()</tt> 1993 and cleaned up with <tt>destroy_rcu_head_on_stack()</tt>. 1994 Similarly, statically allocated non-stack <tt>rcu_head</tt> 1995 structures must be initialized with <tt>init_rcu_head()</tt> 1996 and cleaned up with <tt>destroy_rcu_head()</tt>. 1997 Mathieu Desnoyers made me aware of this requirement, and also 1998 supplied the needed 1999 <a href="https://lkml.kernel.org/g/20100319013024.GA28456@Krystal">patch</a>. 2000<li> An infinite loop in an RCU read-side critical section will 2001 eventually trigger an RCU CPU stall warning splat, with 2002 the duration of “eventually” being controlled by the 2003 <tt>RCU_CPU_STALL_TIMEOUT</tt> <tt>Kconfig</tt> option, or, 2004 alternatively, by the 2005 <tt>rcupdate.rcu_cpu_stall_timeout</tt> boot/sysfs 2006 parameter. 2007 However, RCU is not obligated to produce this splat 2008 unless there is a grace period waiting on that particular 2009 RCU read-side critical section. 2010 <p> 2011 Some extreme workloads might intentionally delay 2012 RCU grace periods, and systems running those workloads can 2013 be booted with <tt>rcupdate.rcu_cpu_stall_suppress</tt> 2014 to suppress the splats. 2015 This kernel parameter may also be set via <tt>sysfs</tt>. 2016 Furthermore, RCU CPU stall warnings are counter-productive 2017 during sysrq dumps and during panics. 2018 RCU therefore supplies the <tt>rcu_sysrq_start()</tt> and 2019 <tt>rcu_sysrq_end()</tt> API members to be called before 2020 and after long sysrq dumps. 2021 RCU also supplies the <tt>rcu_panic()</tt> notifier that is 2022 automatically invoked at the beginning of a panic to suppress 2023 further RCU CPU stall warnings. 2024 2025 <p> 2026 This requirement made itself known in the early 1990s, pretty 2027 much the first time that it was necessary to debug a CPU stall. 2028 That said, the initial implementation in DYNIX/ptx was quite 2029 generic in comparison with that of Linux. 2030<li> Although it would be very good to detect pointers leaking out 2031 of RCU read-side critical sections, there is currently no 2032 good way of doing this. 2033 One complication is the need to distinguish between pointers 2034 leaking and pointers that have been handed off from RCU to 2035 some other synchronization mechanism, for example, reference 2036 counting. 2037<li> In kernels built with <tt>CONFIG_RCU_TRACE=y</tt>, RCU-related 2038 information is provided via event tracing. 2039<li> Open-coded use of <tt>rcu_assign_pointer()</tt> and 2040 <tt>rcu_dereference()</tt> to create typical linked 2041 data structures can be surprisingly error-prone. 2042 Therefore, RCU-protected 2043 <a href="https://lwn.net/Articles/609973/#RCU List APIs">linked lists</a> 2044 and, more recently, RCU-protected 2045 <a href="https://lwn.net/Articles/612100/">hash tables</a> 2046 are available. 2047 Many other special-purpose RCU-protected data structures are 2048 available in the Linux kernel and the userspace RCU library. 2049<li> Some linked structures are created at compile time, but still 2050 require <tt>__rcu</tt> checking. 2051 The <tt>RCU_POINTER_INITIALIZER()</tt> macro serves this 2052 purpose. 2053<li> It is not necessary to use <tt>rcu_assign_pointer()</tt> 2054 when creating linked structures that are to be published via 2055 a single external pointer. 2056 The <tt>RCU_INIT_POINTER()</tt> macro is provided for 2057 this task and also for assigning <tt>NULL</tt> pointers 2058 at runtime. 2059</ol> 2060 2061<p> 2062This not a hard-and-fast list: RCU's diagnostic capabilities will 2063continue to be guided by the number and type of usage bugs found 2064in real-world RCU usage. 2065 2066<h2><a name="Linux Kernel Complications">Linux Kernel Complications</a></h2> 2067 2068<p> 2069The Linux kernel provides an interesting environment for all kinds of 2070software, including RCU. 2071Some of the relevant points of interest are as follows: 2072 2073<ol> 2074<li> <a href="#Configuration">Configuration</a>. 2075<li> <a href="#Firmware Interface">Firmware Interface</a>. 2076<li> <a href="#Early Boot">Early Boot</a>. 2077<li> <a href="#Interrupts and NMIs"> 2078 Interrupts and non-maskable interrupts (NMIs)</a>. 2079<li> <a href="#Loadable Modules">Loadable Modules</a>. 2080<li> <a href="#Hotplug CPU">Hotplug CPU</a>. 2081<li> <a href="#Scheduler and RCU">Scheduler and RCU</a>. 2082<li> <a href="#Tracing and RCU">Tracing and RCU</a>. 2083<li> <a href="#Energy Efficiency">Energy Efficiency</a>. 2084<li> <a href="#Scheduling-Clock Interrupts and RCU"> 2085 Scheduling-Clock Interrupts and RCU</a>. 2086<li> <a href="#Memory Efficiency">Memory Efficiency</a>. 2087<li> <a href="#Performance, Scalability, Response Time, and Reliability"> 2088 Performance, Scalability, Response Time, and Reliability</a>. 2089</ol> 2090 2091<p> 2092This list is probably incomplete, but it does give a feel for the 2093most notable Linux-kernel complications. 2094Each of the following sections covers one of the above topics. 2095 2096<h3><a name="Configuration">Configuration</a></h3> 2097 2098<p> 2099RCU's goal is automatic configuration, so that almost nobody 2100needs to worry about RCU's <tt>Kconfig</tt> options. 2101And for almost all users, RCU does in fact work well 2102“out of the box.” 2103 2104<p> 2105However, there are specialized use cases that are handled by 2106kernel boot parameters and <tt>Kconfig</tt> options. 2107Unfortunately, the <tt>Kconfig</tt> system will explicitly ask users 2108about new <tt>Kconfig</tt> options, which requires almost all of them 2109be hidden behind a <tt>CONFIG_RCU_EXPERT</tt> <tt>Kconfig</tt> option. 2110 2111<p> 2112This all should be quite obvious, but the fact remains that 2113Linus Torvalds recently had to 2114<a href="https://lkml.kernel.org/g/CA+55aFy4wcCwaL4okTs8wXhGZ5h-ibecy_Meg9C4MNQrUnwMcg@mail.gmail.com">remind</a> 2115me of this requirement. 2116 2117<h3><a name="Firmware Interface">Firmware Interface</a></h3> 2118 2119<p> 2120In many cases, kernel obtains information about the system from the 2121firmware, and sometimes things are lost in translation. 2122Or the translation is accurate, but the original message is bogus. 2123 2124<p> 2125For example, some systems' firmware overreports the number of CPUs, 2126sometimes by a large factor. 2127If RCU naively believed the firmware, as it used to do, 2128it would create too many per-CPU kthreads. 2129Although the resulting system will still run correctly, the extra 2130kthreads needlessly consume memory and can cause confusion 2131when they show up in <tt>ps</tt> listings. 2132 2133<p> 2134RCU must therefore wait for a given CPU to actually come online before 2135it can allow itself to believe that the CPU actually exists. 2136The resulting “ghost CPUs” (which are never going to 2137come online) cause a number of 2138<a href="https://paulmck.livejournal.com/37494.html">interesting complications</a>. 2139 2140<h3><a name="Early Boot">Early Boot</a></h3> 2141 2142<p> 2143The Linux kernel's boot sequence is an interesting process, 2144and RCU is used early, even before <tt>rcu_init()</tt> 2145is invoked. 2146In fact, a number of RCU's primitives can be used as soon as the 2147initial task's <tt>task_struct</tt> is available and the 2148boot CPU's per-CPU variables are set up. 2149The read-side primitives (<tt>rcu_read_lock()</tt>, 2150<tt>rcu_read_unlock()</tt>, <tt>rcu_dereference()</tt>, 2151and <tt>rcu_access_pointer()</tt>) will operate normally very early on, 2152as will <tt>rcu_assign_pointer()</tt>. 2153 2154<p> 2155Although <tt>call_rcu()</tt> may be invoked at any 2156time during boot, callbacks are not guaranteed to be invoked until after 2157all of RCU's kthreads have been spawned, which occurs at 2158<tt>early_initcall()</tt> time. 2159This delay in callback invocation is due to the fact that RCU does not 2160invoke callbacks until it is fully initialized, and this full initialization 2161cannot occur until after the scheduler has initialized itself to the 2162point where RCU can spawn and run its kthreads. 2163In theory, it would be possible to invoke callbacks earlier, 2164however, this is not a panacea because there would be severe restrictions 2165on what operations those callbacks could invoke. 2166 2167<p> 2168Perhaps surprisingly, <tt>synchronize_rcu()</tt>, 2169<a href="#Bottom-Half Flavor"><tt>synchronize_rcu_bh()</tt></a> 2170(<a href="#Bottom-Half Flavor">discussed below</a>), 2171<a href="#Sched Flavor"><tt>synchronize_sched()</tt></a>, 2172<tt>synchronize_rcu_expedited()</tt>, 2173<tt>synchronize_rcu_bh_expedited()</tt>, and 2174<tt>synchronize_sched_expedited()</tt> 2175will all operate normally 2176during very early boot, the reason being that there is only one CPU 2177and preemption is disabled. 2178This means that the call <tt>synchronize_rcu()</tt> (or friends) 2179itself is a quiescent 2180state and thus a grace period, so the early-boot implementation can 2181be a no-op. 2182 2183<p> 2184However, once the scheduler has spawned its first kthread, this early 2185boot trick fails for <tt>synchronize_rcu()</tt> (as well as for 2186<tt>synchronize_rcu_expedited()</tt>) in <tt>CONFIG_PREEMPT=y</tt> 2187kernels. 2188The reason is that an RCU read-side critical section might be preempted, 2189which means that a subsequent <tt>synchronize_rcu()</tt> really does have 2190to wait for something, as opposed to simply returning immediately. 2191Unfortunately, <tt>synchronize_rcu()</tt> can't do this until all of 2192its kthreads are spawned, which doesn't happen until some time during 2193<tt>early_initcalls()</tt> time. 2194But this is no excuse: RCU is nevertheless required to correctly handle 2195synchronous grace periods during this time period. 2196Once all of its kthreads are up and running, RCU starts running 2197normally. 2198 2199<table> 2200<tr><th> </th></tr> 2201<tr><th align="left">Quick Quiz:</th></tr> 2202<tr><td> 2203 How can RCU possibly handle grace periods before all of its 2204 kthreads have been spawned??? 2205</td></tr> 2206<tr><th align="left">Answer:</th></tr> 2207<tr><td bgcolor="#ffffff"><font color="ffffff"> 2208 Very carefully! 2209 </font> 2210 2211 <p><font color="ffffff"> 2212 During the “dead zone” between the time that the 2213 scheduler spawns the first task and the time that all of RCU's 2214 kthreads have been spawned, all synchronous grace periods are 2215 handled by the expedited grace-period mechanism. 2216 At runtime, this expedited mechanism relies on workqueues, but 2217 during the dead zone the requesting task itself drives the 2218 desired expedited grace period. 2219 Because dead-zone execution takes place within task context, 2220 everything works. 2221 Once the dead zone ends, expedited grace periods go back to 2222 using workqueues, as is required to avoid problems that would 2223 otherwise occur when a user task received a POSIX signal while 2224 driving an expedited grace period. 2225 </font> 2226 2227 <p><font color="ffffff"> 2228 And yes, this does mean that it is unhelpful to send POSIX 2229 signals to random tasks between the time that the scheduler 2230 spawns its first kthread and the time that RCU's kthreads 2231 have all been spawned. 2232 If there ever turns out to be a good reason for sending POSIX 2233 signals during that time, appropriate adjustments will be made. 2234 (If it turns out that POSIX signals are sent during this time for 2235 no good reason, other adjustments will be made, appropriate 2236 or otherwise.) 2237</font></td></tr> 2238<tr><td> </td></tr> 2239</table> 2240 2241<p> 2242I learned of these boot-time requirements as a result of a series of 2243system hangs. 2244 2245<h3><a name="Interrupts and NMIs">Interrupts and NMIs</a></h3> 2246 2247<p> 2248The Linux kernel has interrupts, and RCU read-side critical sections are 2249legal within interrupt handlers and within interrupt-disabled regions 2250of code, as are invocations of <tt>call_rcu()</tt>. 2251 2252<p> 2253Some Linux-kernel architectures can enter an interrupt handler from 2254non-idle process context, and then just never leave it, instead stealthily 2255transitioning back to process context. 2256This trick is sometimes used to invoke system calls from inside the kernel. 2257These “half-interrupts” mean that RCU has to be very careful 2258about how it counts interrupt nesting levels. 2259I learned of this requirement the hard way during a rewrite 2260of RCU's dyntick-idle code. 2261 2262<p> 2263The Linux kernel has non-maskable interrupts (NMIs), and 2264RCU read-side critical sections are legal within NMI handlers. 2265Thankfully, RCU update-side primitives, including 2266<tt>call_rcu()</tt>, are prohibited within NMI handlers. 2267 2268<p> 2269The name notwithstanding, some Linux-kernel architectures 2270can have nested NMIs, which RCU must handle correctly. 2271Andy Lutomirski 2272<a href="https://lkml.kernel.org/g/CALCETrXLq1y7e_dKFPgou-FKHB6Pu-r8+t-6Ds+8=va7anBWDA@mail.gmail.com">surprised me</a> 2273with this requirement; 2274he also kindly surprised me with 2275<a href="https://lkml.kernel.org/g/CALCETrXSY9JpW3uE6H8WYk81sg56qasA2aqmjMPsq5dOtzso=g@mail.gmail.com">an algorithm</a> 2276that meets this requirement. 2277 2278<h3><a name="Loadable Modules">Loadable Modules</a></h3> 2279 2280<p> 2281The Linux kernel has loadable modules, and these modules can 2282also be unloaded. 2283After a given module has been unloaded, any attempt to call 2284one of its functions results in a segmentation fault. 2285The module-unload functions must therefore cancel any 2286delayed calls to loadable-module functions, for example, 2287any outstanding <tt>mod_timer()</tt> must be dealt with 2288via <tt>del_timer_sync()</tt> or similar. 2289 2290<p> 2291Unfortunately, there is no way to cancel an RCU callback; 2292once you invoke <tt>call_rcu()</tt>, the callback function is 2293going to eventually be invoked, unless the system goes down first. 2294Because it is normally considered socially irresponsible to crash the system 2295in response to a module unload request, we need some other way 2296to deal with in-flight RCU callbacks. 2297 2298<p> 2299RCU therefore provides 2300<tt><a href="https://lwn.net/Articles/217484/">rcu_barrier()</a></tt>, 2301which waits until all in-flight RCU callbacks have been invoked. 2302If a module uses <tt>call_rcu()</tt>, its exit function should therefore 2303prevent any future invocation of <tt>call_rcu()</tt>, then invoke 2304<tt>rcu_barrier()</tt>. 2305In theory, the underlying module-unload code could invoke 2306<tt>rcu_barrier()</tt> unconditionally, but in practice this would 2307incur unacceptable latencies. 2308 2309<p> 2310Nikita Danilov noted this requirement for an analogous filesystem-unmount 2311situation, and Dipankar Sarma incorporated <tt>rcu_barrier()</tt> into RCU. 2312The need for <tt>rcu_barrier()</tt> for module unloading became 2313apparent later. 2314 2315<p> 2316<b>Important note</b>: The <tt>rcu_barrier()</tt> function is not, 2317repeat, <i>not</i>, obligated to wait for a grace period. 2318It is instead only required to wait for RCU callbacks that have 2319already been posted. 2320Therefore, if there are no RCU callbacks posted anywhere in the system, 2321<tt>rcu_barrier()</tt> is within its rights to return immediately. 2322Even if there are callbacks posted, <tt>rcu_barrier()</tt> does not 2323necessarily need to wait for a grace period. 2324 2325<table> 2326<tr><th> </th></tr> 2327<tr><th align="left">Quick Quiz:</th></tr> 2328<tr><td> 2329 Wait a minute! 2330 Each RCU callbacks must wait for a grace period to complete, 2331 and <tt>rcu_barrier()</tt> must wait for each pre-existing 2332 callback to be invoked. 2333 Doesn't <tt>rcu_barrier()</tt> therefore need to wait for 2334 a full grace period if there is even one callback posted anywhere 2335 in the system? 2336</td></tr> 2337<tr><th align="left">Answer:</th></tr> 2338<tr><td bgcolor="#ffffff"><font color="ffffff"> 2339 Absolutely not!!! 2340 </font> 2341 2342 <p><font color="ffffff"> 2343 Yes, each RCU callbacks must wait for a grace period to complete, 2344 but it might well be partly (or even completely) finished waiting 2345 by the time <tt>rcu_barrier()</tt> is invoked. 2346 In that case, <tt>rcu_barrier()</tt> need only wait for the 2347 remaining portion of the grace period to elapse. 2348 So even if there are quite a few callbacks posted, 2349 <tt>rcu_barrier()</tt> might well return quite quickly. 2350 </font> 2351 2352 <p><font color="ffffff"> 2353 So if you need to wait for a grace period as well as for all 2354 pre-existing callbacks, you will need to invoke both 2355 <tt>synchronize_rcu()</tt> and <tt>rcu_barrier()</tt>. 2356 If latency is a concern, you can always use workqueues 2357 to invoke them concurrently. 2358</font></td></tr> 2359<tr><td> </td></tr> 2360</table> 2361 2362<h3><a name="Hotplug CPU">Hotplug CPU</a></h3> 2363 2364<p> 2365The Linux kernel supports CPU hotplug, which means that CPUs 2366can come and go. 2367It is of course illegal to use any RCU API member from an offline CPU, 2368with the exception of <a href="#Sleepable RCU">SRCU</a> read-side 2369critical sections. 2370This requirement was present from day one in DYNIX/ptx, but 2371on the other hand, the Linux kernel's CPU-hotplug implementation 2372is “interesting.” 2373 2374<p> 2375The Linux-kernel CPU-hotplug implementation has notifiers that 2376are used to allow the various kernel subsystems (including RCU) 2377to respond appropriately to a given CPU-hotplug operation. 2378Most RCU operations may be invoked from CPU-hotplug notifiers, 2379including even synchronous grace-period operations such as 2380<tt>synchronize_rcu()</tt> and <tt>synchronize_rcu_expedited()</tt>. 2381 2382<p> 2383However, all-callback-wait operations such as 2384<tt>rcu_barrier()</tt> are also not supported, due to the 2385fact that there are phases of CPU-hotplug operations where 2386the outgoing CPU's callbacks will not be invoked until after 2387the CPU-hotplug operation ends, which could also result in deadlock. 2388Furthermore, <tt>rcu_barrier()</tt> blocks CPU-hotplug operations 2389during its execution, which results in another type of deadlock 2390when invoked from a CPU-hotplug notifier. 2391 2392<h3><a name="Scheduler and RCU">Scheduler and RCU</a></h3> 2393 2394<p> 2395RCU depends on the scheduler, and the scheduler uses RCU to 2396protect some of its data structures. 2397This means the scheduler is forbidden from acquiring 2398the runqueue locks and the priority-inheritance locks 2399in the middle of an outermost RCU read-side critical section unless either 2400(1) it releases them before exiting that same 2401RCU read-side critical section, or 2402(2) interrupts are disabled across 2403that entire RCU read-side critical section. 2404This same prohibition also applies (recursively!) to any lock that is acquired 2405while holding any lock to which this prohibition applies. 2406Adhering to this rule prevents preemptible RCU from invoking 2407<tt>rcu_read_unlock_special()</tt> while either runqueue or 2408priority-inheritance locks are held, thus avoiding deadlock. 2409 2410<p> 2411Prior to v4.4, it was only necessary to disable preemption across 2412RCU read-side critical sections that acquired scheduler locks. 2413In v4.4, expedited grace periods started using IPIs, and these 2414IPIs could force a <tt>rcu_read_unlock()</tt> to take the slowpath. 2415Therefore, this expedited-grace-period change required disabling of 2416interrupts, not just preemption. 2417 2418<p> 2419For RCU's part, the preemptible-RCU <tt>rcu_read_unlock()</tt> 2420implementation must be written carefully to avoid similar deadlocks. 2421In particular, <tt>rcu_read_unlock()</tt> must tolerate an 2422interrupt where the interrupt handler invokes both 2423<tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>. 2424This possibility requires <tt>rcu_read_unlock()</tt> to use 2425negative nesting levels to avoid destructive recursion via 2426interrupt handler's use of RCU. 2427 2428<p> 2429This pair of mutual scheduler-RCU requirements came as a 2430<a href="https://lwn.net/Articles/453002/">complete surprise</a>. 2431 2432<p> 2433As noted above, RCU makes use of kthreads, and it is necessary to 2434avoid excessive CPU-time accumulation by these kthreads. 2435This requirement was no surprise, but RCU's violation of it 2436when running context-switch-heavy workloads when built with 2437<tt>CONFIG_NO_HZ_FULL=y</tt> 2438<a href="http://www.rdrop.com/users/paulmck/scalability/paper/BareMetal.2015.01.15b.pdf">did come as a surprise [PDF]</a>. 2439RCU has made good progress towards meeting this requirement, even 2440for context-switch-have <tt>CONFIG_NO_HZ_FULL=y</tt> workloads, 2441but there is room for further improvement. 2442 2443<h3><a name="Tracing and RCU">Tracing and RCU</a></h3> 2444 2445<p> 2446It is possible to use tracing on RCU code, but tracing itself 2447uses RCU. 2448For this reason, <tt>rcu_dereference_raw_notrace()</tt> 2449is provided for use by tracing, which avoids the destructive 2450recursion that could otherwise ensue. 2451This API is also used by virtualization in some architectures, 2452where RCU readers execute in environments in which tracing 2453cannot be used. 2454The tracing folks both located the requirement and provided the 2455needed fix, so this surprise requirement was relatively painless. 2456 2457<h3><a name="Energy Efficiency">Energy Efficiency</a></h3> 2458 2459<p> 2460Interrupting idle CPUs is considered socially unacceptable, 2461especially by people with battery-powered embedded systems. 2462RCU therefore conserves energy by detecting which CPUs are 2463idle, including tracking CPUs that have been interrupted from idle. 2464This is a large part of the energy-efficiency requirement, 2465so I learned of this via an irate phone call. 2466 2467<p> 2468Because RCU avoids interrupting idle CPUs, it is illegal to 2469execute an RCU read-side critical section on an idle CPU. 2470(Kernels built with <tt>CONFIG_PROVE_RCU=y</tt> will splat 2471if you try it.) 2472The <tt>RCU_NONIDLE()</tt> macro and <tt>_rcuidle</tt> 2473event tracing is provided to work around this restriction. 2474In addition, <tt>rcu_is_watching()</tt> may be used to 2475test whether or not it is currently legal to run RCU read-side 2476critical sections on this CPU. 2477I learned of the need for diagnostics on the one hand 2478and <tt>RCU_NONIDLE()</tt> on the other while inspecting 2479idle-loop code. 2480Steven Rostedt supplied <tt>_rcuidle</tt> event tracing, 2481which is used quite heavily in the idle loop. 2482However, there are some restrictions on the code placed within 2483<tt>RCU_NONIDLE()</tt>: 2484 2485<ol> 2486<li> Blocking is prohibited. 2487 In practice, this is not a serious restriction given that idle 2488 tasks are prohibited from blocking to begin with. 2489<li> Although nesting <tt>RCU_NONIDLE()</tt> is permitted, they cannot 2490 nest indefinitely deeply. 2491 However, given that they can be nested on the order of a million 2492 deep, even on 32-bit systems, this should not be a serious 2493 restriction. 2494 This nesting limit would probably be reached long after the 2495 compiler OOMed or the stack overflowed. 2496<li> Any code path that enters <tt>RCU_NONIDLE()</tt> must sequence 2497 out of that same <tt>RCU_NONIDLE()</tt>. 2498 For example, the following is grossly illegal: 2499 2500 <blockquote> 2501 <pre> 2502 1 RCU_NONIDLE({ 2503 2 do_something(); 2504 3 goto bad_idea; /* BUG!!! */ 2505 4 do_something_else();}); 2506 5 bad_idea: 2507 </pre> 2508 </blockquote> 2509 2510 <p> 2511 It is just as illegal to transfer control into the middle of 2512 <tt>RCU_NONIDLE()</tt>'s argument. 2513 Yes, in theory, you could transfer in as long as you also 2514 transferred out, but in practice you could also expect to get sharply 2515 worded review comments. 2516</ol> 2517 2518<p> 2519It is similarly socially unacceptable to interrupt an 2520<tt>nohz_full</tt> CPU running in userspace. 2521RCU must therefore track <tt>nohz_full</tt> userspace 2522execution. 2523RCU must therefore be able to sample state at two points in 2524time, and be able to determine whether or not some other CPU spent 2525any time idle and/or executing in userspace. 2526 2527<p> 2528These energy-efficiency requirements have proven quite difficult to 2529understand and to meet, for example, there have been more than five 2530clean-sheet rewrites of RCU's energy-efficiency code, the last of 2531which was finally able to demonstrate 2532<a href="http://www.rdrop.com/users/paulmck/realtime/paper/AMPenergy.2013.04.19a.pdf">real energy savings running on real hardware [PDF]</a>. 2533As noted earlier, 2534I learned of many of these requirements via angry phone calls: 2535Flaming me on the Linux-kernel mailing list was apparently not 2536sufficient to fully vent their ire at RCU's energy-efficiency bugs! 2537 2538<h3><a name="Scheduling-Clock Interrupts and RCU"> 2539Scheduling-Clock Interrupts and RCU</a></h3> 2540 2541<p> 2542The kernel transitions between in-kernel non-idle execution, userspace 2543execution, and the idle loop. 2544Depending on kernel configuration, RCU handles these states differently: 2545 2546<table border=3> 2547<tr><th><tt>HZ</tt> Kconfig</th> 2548 <th>In-Kernel</th> 2549 <th>Usermode</th> 2550 <th>Idle</th></tr> 2551<tr><th align="left"><tt>HZ_PERIODIC</tt></th> 2552 <td>Can rely on scheduling-clock interrupt.</td> 2553 <td>Can rely on scheduling-clock interrupt and its 2554 detection of interrupt from usermode.</td> 2555 <td>Can rely on RCU's dyntick-idle detection.</td></tr> 2556<tr><th align="left"><tt>NO_HZ_IDLE</tt></th> 2557 <td>Can rely on scheduling-clock interrupt.</td> 2558 <td>Can rely on scheduling-clock interrupt and its 2559 detection of interrupt from usermode.</td> 2560 <td>Can rely on RCU's dyntick-idle detection.</td></tr> 2561<tr><th align="left"><tt>NO_HZ_FULL</tt></th> 2562 <td>Can only sometimes rely on scheduling-clock interrupt. 2563 In other cases, it is necessary to bound kernel execution 2564 times and/or use IPIs.</td> 2565 <td>Can rely on RCU's dyntick-idle detection.</td> 2566 <td>Can rely on RCU's dyntick-idle detection.</td></tr> 2567</table> 2568 2569<table> 2570<tr><th> </th></tr> 2571<tr><th align="left">Quick Quiz:</th></tr> 2572<tr><td> 2573 Why can't <tt>NO_HZ_FULL</tt> in-kernel execution rely on the 2574 scheduling-clock interrupt, just like <tt>HZ_PERIODIC</tt> 2575 and <tt>NO_HZ_IDLE</tt> do? 2576</td></tr> 2577<tr><th align="left">Answer:</th></tr> 2578<tr><td bgcolor="#ffffff"><font color="ffffff"> 2579 Because, as a performance optimization, <tt>NO_HZ_FULL</tt> 2580 does not necessarily re-enable the scheduling-clock interrupt 2581 on entry to each and every system call. 2582</font></td></tr> 2583<tr><td> </td></tr> 2584</table> 2585 2586<p> 2587However, RCU must be reliably informed as to whether any given 2588CPU is currently in the idle loop, and, for <tt>NO_HZ_FULL</tt>, 2589also whether that CPU is executing in usermode, as discussed 2590<a href="#Energy Efficiency">earlier</a>. 2591It also requires that the scheduling-clock interrupt be enabled when 2592RCU needs it to be: 2593 2594<ol> 2595<li> If a CPU is either idle or executing in usermode, and RCU believes 2596 it is non-idle, the scheduling-clock tick had better be running. 2597 Otherwise, you will get RCU CPU stall warnings. Or at best, 2598 very long (11-second) grace periods, with a pointless IPI waking 2599 the CPU from time to time. 2600<li> If a CPU is in a portion of the kernel that executes RCU read-side 2601 critical sections, and RCU believes this CPU to be idle, you will get 2602 random memory corruption. <b>DON'T DO THIS!!!</b> 2603 2604 <br>This is one reason to test with lockdep, which will complain 2605 about this sort of thing. 2606<li> If a CPU is in a portion of the kernel that is absolutely 2607 positively no-joking guaranteed to never execute any RCU read-side 2608 critical sections, and RCU believes this CPU to to be idle, 2609 no problem. This sort of thing is used by some architectures 2610 for light-weight exception handlers, which can then avoid the 2611 overhead of <tt>rcu_irq_enter()</tt> and <tt>rcu_irq_exit()</tt> 2612 at exception entry and exit, respectively. 2613 Some go further and avoid the entireties of <tt>irq_enter()</tt> 2614 and <tt>irq_exit()</tt>. 2615 2616 <br>Just make very sure you are running some of your tests with 2617 <tt>CONFIG_PROVE_RCU=y</tt>, just in case one of your code paths 2618 was in fact joking about not doing RCU read-side critical sections. 2619<li> If a CPU is executing in the kernel with the scheduling-clock 2620 interrupt disabled and RCU believes this CPU to be non-idle, 2621 and if the CPU goes idle (from an RCU perspective) every few 2622 jiffies, no problem. It is usually OK for there to be the 2623 occasional gap between idle periods of up to a second or so. 2624 2625 <br>If the gap grows too long, you get RCU CPU stall warnings. 2626<li> If a CPU is either idle or executing in usermode, and RCU believes 2627 it to be idle, of course no problem. 2628<li> If a CPU is executing in the kernel, the kernel code 2629 path is passing through quiescent states at a reasonable 2630 frequency (preferably about once per few jiffies, but the 2631 occasional excursion to a second or so is usually OK) and the 2632 scheduling-clock interrupt is enabled, of course no problem. 2633 2634 <br>If the gap between a successive pair of quiescent states grows 2635 too long, you get RCU CPU stall warnings. 2636</ol> 2637 2638<table> 2639<tr><th> </th></tr> 2640<tr><th align="left">Quick Quiz:</th></tr> 2641<tr><td> 2642 But what if my driver has a hardware interrupt handler 2643 that can run for many seconds? 2644 I cannot invoke <tt>schedule()</tt> from an hardware 2645 interrupt handler, after all! 2646</td></tr> 2647<tr><th align="left">Answer:</th></tr> 2648<tr><td bgcolor="#ffffff"><font color="ffffff"> 2649 One approach is to do <tt>rcu_irq_exit();rcu_irq_enter();</tt> 2650 every so often. 2651 But given that long-running interrupt handlers can cause 2652 other problems, not least for response time, shouldn't you 2653 work to keep your interrupt handler's runtime within reasonable 2654 bounds? 2655</font></td></tr> 2656<tr><td> </td></tr> 2657</table> 2658 2659<p> 2660But as long as RCU is properly informed of kernel state transitions between 2661in-kernel execution, usermode execution, and idle, and as long as the 2662scheduling-clock interrupt is enabled when RCU needs it to be, you 2663can rest assured that the bugs you encounter will be in some other 2664part of RCU or some other part of the kernel! 2665 2666<h3><a name="Memory Efficiency">Memory Efficiency</a></h3> 2667 2668<p> 2669Although small-memory non-realtime systems can simply use Tiny RCU, 2670code size is only one aspect of memory efficiency. 2671Another aspect is the size of the <tt>rcu_head</tt> structure 2672used by <tt>call_rcu()</tt> and <tt>kfree_rcu()</tt>. 2673Although this structure contains nothing more than a pair of pointers, 2674it does appear in many RCU-protected data structures, including 2675some that are size critical. 2676The <tt>page</tt> structure is a case in point, as evidenced by 2677the many occurrences of the <tt>union</tt> keyword within that structure. 2678 2679<p> 2680This need for memory efficiency is one reason that RCU uses hand-crafted 2681singly linked lists to track the <tt>rcu_head</tt> structures that 2682are waiting for a grace period to elapse. 2683It is also the reason why <tt>rcu_head</tt> structures do not contain 2684debug information, such as fields tracking the file and line of the 2685<tt>call_rcu()</tt> or <tt>kfree_rcu()</tt> that posted them. 2686Although this information might appear in debug-only kernel builds at some 2687point, in the meantime, the <tt>->func</tt> field will often provide 2688the needed debug information. 2689 2690<p> 2691However, in some cases, the need for memory efficiency leads to even 2692more extreme measures. 2693Returning to the <tt>page</tt> structure, the <tt>rcu_head</tt> field 2694shares storage with a great many other structures that are used at 2695various points in the corresponding page's lifetime. 2696In order to correctly resolve certain 2697<a href="https://lkml.kernel.org/g/1439976106-137226-1-git-send-email-kirill.shutemov@linux.intel.com">race conditions</a>, 2698the Linux kernel's memory-management subsystem needs a particular bit 2699to remain zero during all phases of grace-period processing, 2700and that bit happens to map to the bottom bit of the 2701<tt>rcu_head</tt> structure's <tt>->next</tt> field. 2702RCU makes this guarantee as long as <tt>call_rcu()</tt> 2703is used to post the callback, as opposed to <tt>kfree_rcu()</tt> 2704or some future “lazy” 2705variant of <tt>call_rcu()</tt> that might one day be created for 2706energy-efficiency purposes. 2707 2708<p> 2709That said, there are limits. 2710RCU requires that the <tt>rcu_head</tt> structure be aligned to a 2711two-byte boundary, and passing a misaligned <tt>rcu_head</tt> 2712structure to one of the <tt>call_rcu()</tt> family of functions 2713will result in a splat. 2714It is therefore necessary to exercise caution when packing 2715structures containing fields of type <tt>rcu_head</tt>. 2716Why not a four-byte or even eight-byte alignment requirement? 2717Because the m68k architecture provides only two-byte alignment, 2718and thus acts as alignment's least common denominator. 2719 2720<p> 2721The reason for reserving the bottom bit of pointers to 2722<tt>rcu_head</tt> structures is to leave the door open to 2723“lazy” callbacks whose invocations can safely be deferred. 2724Deferring invocation could potentially have energy-efficiency 2725benefits, but only if the rate of non-lazy callbacks decreases 2726significantly for some important workload. 2727In the meantime, reserving the bottom bit keeps this option open 2728in case it one day becomes useful. 2729 2730<h3><a name="Performance, Scalability, Response Time, and Reliability"> 2731Performance, Scalability, Response Time, and Reliability</a></h3> 2732 2733<p> 2734Expanding on the 2735<a href="#Performance and Scalability">earlier discussion</a>, 2736RCU is used heavily by hot code paths in performance-critical 2737portions of the Linux kernel's networking, security, virtualization, 2738and scheduling code paths. 2739RCU must therefore use efficient implementations, especially in its 2740read-side primitives. 2741To that end, it would be good if preemptible RCU's implementation 2742of <tt>rcu_read_lock()</tt> could be inlined, however, doing 2743this requires resolving <tt>#include</tt> issues with the 2744<tt>task_struct</tt> structure. 2745 2746<p> 2747The Linux kernel supports hardware configurations with up to 27484096 CPUs, which means that RCU must be extremely scalable. 2749Algorithms that involve frequent acquisitions of global locks or 2750frequent atomic operations on global variables simply cannot be 2751tolerated within the RCU implementation. 2752RCU therefore makes heavy use of a combining tree based on the 2753<tt>rcu_node</tt> structure. 2754RCU is required to tolerate all CPUs continuously invoking any 2755combination of RCU's runtime primitives with minimal per-operation 2756overhead. 2757In fact, in many cases, increasing load must <i>decrease</i> the 2758per-operation overhead, witness the batching optimizations for 2759<tt>synchronize_rcu()</tt>, <tt>call_rcu()</tt>, 2760<tt>synchronize_rcu_expedited()</tt>, and <tt>rcu_barrier()</tt>. 2761As a general rule, RCU must cheerfully accept whatever the 2762rest of the Linux kernel decides to throw at it. 2763 2764<p> 2765The Linux kernel is used for real-time workloads, especially 2766in conjunction with the 2767<a href="https://rt.wiki.kernel.org/index.php/Main_Page">-rt patchset</a>. 2768The real-time-latency response requirements are such that the 2769traditional approach of disabling preemption across RCU 2770read-side critical sections is inappropriate. 2771Kernels built with <tt>CONFIG_PREEMPT=y</tt> therefore 2772use an RCU implementation that allows RCU read-side critical 2773sections to be preempted. 2774This requirement made its presence known after users made it 2775clear that an earlier 2776<a href="https://lwn.net/Articles/107930/">real-time patch</a> 2777did not meet their needs, in conjunction with some 2778<a href="https://lkml.kernel.org/g/20050318002026.GA2693@us.ibm.com">RCU issues</a> 2779encountered by a very early version of the -rt patchset. 2780 2781<p> 2782In addition, RCU must make do with a sub-100-microsecond real-time latency 2783budget. 2784In fact, on smaller systems with the -rt patchset, the Linux kernel 2785provides sub-20-microsecond real-time latencies for the whole kernel, 2786including RCU. 2787RCU's scalability and latency must therefore be sufficient for 2788these sorts of configurations. 2789To my surprise, the sub-100-microsecond real-time latency budget 2790<a href="http://www.rdrop.com/users/paulmck/realtime/paper/bigrt.2013.01.31a.LCA.pdf"> 2791applies to even the largest systems [PDF]</a>, 2792up to and including systems with 4096 CPUs. 2793This real-time requirement motivated the grace-period kthread, which 2794also simplified handling of a number of race conditions. 2795 2796<p> 2797RCU must avoid degrading real-time response for CPU-bound threads, whether 2798executing in usermode (which is one use case for 2799<tt>CONFIG_NO_HZ_FULL=y</tt>) or in the kernel. 2800That said, CPU-bound loops in the kernel must execute 2801<tt>cond_resched()</tt> at least once per few tens of milliseconds 2802in order to avoid receiving an IPI from RCU. 2803 2804<p> 2805Finally, RCU's status as a synchronization primitive means that 2806any RCU failure can result in arbitrary memory corruption that can be 2807extremely difficult to debug. 2808This means that RCU must be extremely reliable, which in 2809practice also means that RCU must have an aggressive stress-test 2810suite. 2811This stress-test suite is called <tt>rcutorture</tt>. 2812 2813<p> 2814Although the need for <tt>rcutorture</tt> was no surprise, 2815the current immense popularity of the Linux kernel is posing 2816interesting—and perhaps unprecedented—validation 2817challenges. 2818To see this, keep in mind that there are well over one billion 2819instances of the Linux kernel running today, given Android 2820smartphones, Linux-powered televisions, and servers. 2821This number can be expected to increase sharply with the advent of 2822the celebrated Internet of Things. 2823 2824<p> 2825Suppose that RCU contains a race condition that manifests on average 2826once per million years of runtime. 2827This bug will be occurring about three times per <i>day</i> across 2828the installed base. 2829RCU could simply hide behind hardware error rates, given that no one 2830should really expect their smartphone to last for a million years. 2831However, anyone taking too much comfort from this thought should 2832consider the fact that in most jurisdictions, a successful multi-year 2833test of a given mechanism, which might include a Linux kernel, 2834suffices for a number of types of safety-critical certifications. 2835In fact, rumor has it that the Linux kernel is already being used 2836in production for safety-critical applications. 2837I don't know about you, but I would feel quite bad if a bug in RCU 2838killed someone. 2839Which might explain my recent focus on validation and verification. 2840 2841<h2><a name="Other RCU Flavors">Other RCU Flavors</a></h2> 2842 2843<p> 2844One of the more surprising things about RCU is that there are now 2845no fewer than five <i>flavors</i>, or API families. 2846In addition, the primary flavor that has been the sole focus up to 2847this point has two different implementations, non-preemptible and 2848preemptible. 2849The other four flavors are listed below, with requirements for each 2850described in a separate section. 2851 2852<ol> 2853<li> <a href="#Bottom-Half Flavor">Bottom-Half Flavor</a> 2854<li> <a href="#Sched Flavor">Sched Flavor</a> 2855<li> <a href="#Sleepable RCU">Sleepable RCU</a> 2856<li> <a href="#Tasks RCU">Tasks RCU</a> 2857<li> <a href="#Waiting for Multiple Grace Periods"> 2858 Waiting for Multiple Grace Periods</a> 2859</ol> 2860 2861<h3><a name="Bottom-Half Flavor">Bottom-Half Flavor</a></h3> 2862 2863<p> 2864The softirq-disable (AKA “bottom-half”, 2865hence the “_bh” abbreviations) 2866flavor of RCU, or <i>RCU-bh</i>, was developed by 2867Dipankar Sarma to provide a flavor of RCU that could withstand the 2868network-based denial-of-service attacks researched by Robert 2869Olsson. 2870These attacks placed so much networking load on the system 2871that some of the CPUs never exited softirq execution, 2872which in turn prevented those CPUs from ever executing a context switch, 2873which, in the RCU implementation of that time, prevented grace periods 2874from ever ending. 2875The result was an out-of-memory condition and a system hang. 2876 2877<p> 2878The solution was the creation of RCU-bh, which does 2879<tt>local_bh_disable()</tt> 2880across its read-side critical sections, and which uses the transition 2881from one type of softirq processing to another as a quiescent state 2882in addition to context switch, idle, user mode, and offline. 2883This means that RCU-bh grace periods can complete even when some of 2884the CPUs execute in softirq indefinitely, thus allowing algorithms 2885based on RCU-bh to withstand network-based denial-of-service attacks. 2886 2887<p> 2888Because 2889<tt>rcu_read_lock_bh()</tt> and <tt>rcu_read_unlock_bh()</tt> 2890disable and re-enable softirq handlers, any attempt to start a softirq 2891handlers during the 2892RCU-bh read-side critical section will be deferred. 2893In this case, <tt>rcu_read_unlock_bh()</tt> 2894will invoke softirq processing, which can take considerable time. 2895One can of course argue that this softirq overhead should be associated 2896with the code following the RCU-bh read-side critical section rather 2897than <tt>rcu_read_unlock_bh()</tt>, but the fact 2898is that most profiling tools cannot be expected to make this sort 2899of fine distinction. 2900For example, suppose that a three-millisecond-long RCU-bh read-side 2901critical section executes during a time of heavy networking load. 2902There will very likely be an attempt to invoke at least one softirq 2903handler during that three milliseconds, but any such invocation will 2904be delayed until the time of the <tt>rcu_read_unlock_bh()</tt>. 2905This can of course make it appear at first glance as if 2906<tt>rcu_read_unlock_bh()</tt> was executing very slowly. 2907 2908<p> 2909The 2910<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-bh API</a> 2911includes 2912<tt>rcu_read_lock_bh()</tt>, 2913<tt>rcu_read_unlock_bh()</tt>, 2914<tt>rcu_dereference_bh()</tt>, 2915<tt>rcu_dereference_bh_check()</tt>, 2916<tt>synchronize_rcu_bh()</tt>, 2917<tt>synchronize_rcu_bh_expedited()</tt>, 2918<tt>call_rcu_bh()</tt>, 2919<tt>rcu_barrier_bh()</tt>, and 2920<tt>rcu_read_lock_bh_held()</tt>. 2921 2922<h3><a name="Sched Flavor">Sched Flavor</a></h3> 2923 2924<p> 2925Before preemptible RCU, waiting for an RCU grace period had the 2926side effect of also waiting for all pre-existing interrupt 2927and NMI handlers. 2928However, there are legitimate preemptible-RCU implementations that 2929do not have this property, given that any point in the code outside 2930of an RCU read-side critical section can be a quiescent state. 2931Therefore, <i>RCU-sched</i> was created, which follows “classic” 2932RCU in that an RCU-sched grace period waits for for pre-existing 2933interrupt and NMI handlers. 2934In kernels built with <tt>CONFIG_PREEMPT=n</tt>, the RCU and RCU-sched 2935APIs have identical implementations, while kernels built with 2936<tt>CONFIG_PREEMPT=y</tt> provide a separate implementation for each. 2937 2938<p> 2939Note well that in <tt>CONFIG_PREEMPT=y</tt> kernels, 2940<tt>rcu_read_lock_sched()</tt> and <tt>rcu_read_unlock_sched()</tt> 2941disable and re-enable preemption, respectively. 2942This means that if there was a preemption attempt during the 2943RCU-sched read-side critical section, <tt>rcu_read_unlock_sched()</tt> 2944will enter the scheduler, with all the latency and overhead entailed. 2945Just as with <tt>rcu_read_unlock_bh()</tt>, this can make it look 2946as if <tt>rcu_read_unlock_sched()</tt> was executing very slowly. 2947However, the highest-priority task won't be preempted, so that task 2948will enjoy low-overhead <tt>rcu_read_unlock_sched()</tt> invocations. 2949 2950<p> 2951The 2952<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">RCU-sched API</a> 2953includes 2954<tt>rcu_read_lock_sched()</tt>, 2955<tt>rcu_read_unlock_sched()</tt>, 2956<tt>rcu_read_lock_sched_notrace()</tt>, 2957<tt>rcu_read_unlock_sched_notrace()</tt>, 2958<tt>rcu_dereference_sched()</tt>, 2959<tt>rcu_dereference_sched_check()</tt>, 2960<tt>synchronize_sched()</tt>, 2961<tt>synchronize_rcu_sched_expedited()</tt>, 2962<tt>call_rcu_sched()</tt>, 2963<tt>rcu_barrier_sched()</tt>, and 2964<tt>rcu_read_lock_sched_held()</tt>. 2965However, anything that disables preemption also marks an RCU-sched 2966read-side critical section, including 2967<tt>preempt_disable()</tt> and <tt>preempt_enable()</tt>, 2968<tt>local_irq_save()</tt> and <tt>local_irq_restore()</tt>, 2969and so on. 2970 2971<h3><a name="Sleepable RCU">Sleepable RCU</a></h3> 2972 2973<p> 2974For well over a decade, someone saying “I need to block within 2975an RCU read-side critical section” was a reliable indication 2976that this someone did not understand RCU. 2977After all, if you are always blocking in an RCU read-side critical 2978section, you can probably afford to use a higher-overhead synchronization 2979mechanism. 2980However, that changed with the advent of the Linux kernel's notifiers, 2981whose RCU read-side critical 2982sections almost never sleep, but sometimes need to. 2983This resulted in the introduction of 2984<a href="https://lwn.net/Articles/202847/">sleepable RCU</a>, 2985or <i>SRCU</i>. 2986 2987<p> 2988SRCU allows different domains to be defined, with each such domain 2989defined by an instance of an <tt>srcu_struct</tt> structure. 2990A pointer to this structure must be passed in to each SRCU function, 2991for example, <tt>synchronize_srcu(&ss)</tt>, where 2992<tt>ss</tt> is the <tt>srcu_struct</tt> structure. 2993The key benefit of these domains is that a slow SRCU reader in one 2994domain does not delay an SRCU grace period in some other domain. 2995That said, one consequence of these domains is that read-side code 2996must pass a “cookie” from <tt>srcu_read_lock()</tt> 2997to <tt>srcu_read_unlock()</tt>, for example, as follows: 2998 2999<blockquote> 3000<pre> 3001 1 int idx; 3002 2 3003 3 idx = srcu_read_lock(&ss); 3004 4 do_something(); 3005 5 srcu_read_unlock(&ss, idx); 3006</pre> 3007</blockquote> 3008 3009<p> 3010As noted above, it is legal to block within SRCU read-side critical sections, 3011however, with great power comes great responsibility. 3012If you block forever in one of a given domain's SRCU read-side critical 3013sections, then that domain's grace periods will also be blocked forever. 3014Of course, one good way to block forever is to deadlock, which can 3015happen if any operation in a given domain's SRCU read-side critical 3016section can block waiting, either directly or indirectly, for that domain's 3017grace period to elapse. 3018For example, this results in a self-deadlock: 3019 3020<blockquote> 3021<pre> 3022 1 int idx; 3023 2 3024 3 idx = srcu_read_lock(&ss); 3025 4 do_something(); 3026 5 synchronize_srcu(&ss); 3027 6 srcu_read_unlock(&ss, idx); 3028</pre> 3029</blockquote> 3030 3031<p> 3032However, if line 5 acquired a mutex that was held across 3033a <tt>synchronize_srcu()</tt> for domain <tt>ss</tt>, 3034deadlock would still be possible. 3035Furthermore, if line 5 acquired a mutex that was held across 3036a <tt>synchronize_srcu()</tt> for some other domain <tt>ss1</tt>, 3037and if an <tt>ss1</tt>-domain SRCU read-side critical section 3038acquired another mutex that was held across as <tt>ss</tt>-domain 3039<tt>synchronize_srcu()</tt>, 3040deadlock would again be possible. 3041Such a deadlock cycle could extend across an arbitrarily large number 3042of different SRCU domains. 3043Again, with great power comes great responsibility. 3044 3045<p> 3046Unlike the other RCU flavors, SRCU read-side critical sections can 3047run on idle and even offline CPUs. 3048This ability requires that <tt>srcu_read_lock()</tt> and 3049<tt>srcu_read_unlock()</tt> contain memory barriers, which means 3050that SRCU readers will run a bit slower than would RCU readers. 3051It also motivates the <tt>smp_mb__after_srcu_read_unlock()</tt> 3052API, which, in combination with <tt>srcu_read_unlock()</tt>, 3053guarantees a full memory barrier. 3054 3055<p> 3056Also unlike other RCU flavors, SRCU's callbacks-wait function 3057<tt>srcu_barrier()</tt> may be invoked from CPU-hotplug notifiers, 3058though this is not necessarily a good idea. 3059The reason that this is possible is that SRCU is insensitive 3060to whether or not a CPU is online, which means that <tt>srcu_barrier()</tt> 3061need not exclude CPU-hotplug operations. 3062 3063<p> 3064SRCU also differs from other RCU flavors in that SRCU's expedited and 3065non-expedited grace periods are implemented by the same mechanism. 3066This means that in the current SRCU implementation, expediting a 3067future grace period has the side effect of expediting all prior 3068grace periods that have not yet completed. 3069(But please note that this is a property of the current implementation, 3070not necessarily of future implementations.) 3071In addition, if SRCU has been idle for longer than the interval 3072specified by the <tt>srcutree.exp_holdoff</tt> kernel boot parameter 3073(25 microseconds by default), 3074and if a <tt>synchronize_srcu()</tt> invocation ends this idle period, 3075that invocation will be automatically expedited. 3076 3077<p> 3078As of v4.12, SRCU's callbacks are maintained per-CPU, eliminating 3079a locking bottleneck present in prior kernel versions. 3080Although this will allow users to put much heavier stress on 3081<tt>call_srcu()</tt>, it is important to note that SRCU does not 3082yet take any special steps to deal with callback flooding. 3083So if you are posting (say) 10,000 SRCU callbacks per second per CPU, 3084you are probably totally OK, but if you intend to post (say) 1,000,000 3085SRCU callbacks per second per CPU, please run some tests first. 3086SRCU just might need a few adjustment to deal with that sort of load. 3087Of course, your mileage may vary based on the speed of your CPUs and 3088the size of your memory. 3089 3090<p> 3091The 3092<a href="https://lwn.net/Articles/609973/#RCU Per-Flavor API Table">SRCU API</a> 3093includes 3094<tt>srcu_read_lock()</tt>, 3095<tt>srcu_read_unlock()</tt>, 3096<tt>srcu_dereference()</tt>, 3097<tt>srcu_dereference_check()</tt>, 3098<tt>synchronize_srcu()</tt>, 3099<tt>synchronize_srcu_expedited()</tt>, 3100<tt>call_srcu()</tt>, 3101<tt>srcu_barrier()</tt>, and 3102<tt>srcu_read_lock_held()</tt>. 3103It also includes 3104<tt>DEFINE_SRCU()</tt>, 3105<tt>DEFINE_STATIC_SRCU()</tt>, and 3106<tt>init_srcu_struct()</tt> 3107APIs for defining and initializing <tt>srcu_struct</tt> structures. 3108 3109<h3><a name="Tasks RCU">Tasks RCU</a></h3> 3110 3111<p> 3112Some forms of tracing use “trampolines” to handle the 3113binary rewriting required to install different types of probes. 3114It would be good to be able to free old trampolines, which sounds 3115like a job for some form of RCU. 3116However, because it is necessary to be able to install a trace 3117anywhere in the code, it is not possible to use read-side markers 3118such as <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt>. 3119In addition, it does not work to have these markers in the trampoline 3120itself, because there would need to be instructions following 3121<tt>rcu_read_unlock()</tt>. 3122Although <tt>synchronize_rcu()</tt> would guarantee that execution 3123reached the <tt>rcu_read_unlock()</tt>, it would not be able to 3124guarantee that execution had completely left the trampoline. 3125 3126<p> 3127The solution, in the form of 3128<a href="https://lwn.net/Articles/607117/"><i>Tasks RCU</i></a>, 3129is to have implicit 3130read-side critical sections that are delimited by voluntary context 3131switches, that is, calls to <tt>schedule()</tt>, 3132<tt>cond_resched()</tt>, and 3133<tt>synchronize_rcu_tasks()</tt>. 3134In addition, transitions to and from userspace execution also delimit 3135tasks-RCU read-side critical sections. 3136 3137<p> 3138The tasks-RCU API is quite compact, consisting only of 3139<tt>call_rcu_tasks()</tt>, 3140<tt>synchronize_rcu_tasks()</tt>, and 3141<tt>rcu_barrier_tasks()</tt>. 3142 3143<h3><a name="Waiting for Multiple Grace Periods"> 3144Waiting for Multiple Grace Periods</a></h3> 3145 3146<p> 3147Perhaps you have an RCU protected data structure that is accessed from 3148RCU read-side critical sections, from softirq handlers, and from 3149hardware interrupt handlers. 3150That is three flavors of RCU, the normal flavor, the bottom-half flavor, 3151and the sched flavor. 3152How to wait for a compound grace period? 3153 3154<p> 3155The best approach is usually to “just say no!” and 3156insert <tt>rcu_read_lock()</tt> and <tt>rcu_read_unlock()</tt> 3157around each RCU read-side critical section, regardless of what 3158environment it happens to be in. 3159But suppose that some of the RCU read-side critical sections are 3160on extremely hot code paths, and that use of <tt>CONFIG_PREEMPT=n</tt> 3161is not a viable option, so that <tt>rcu_read_lock()</tt> and 3162<tt>rcu_read_unlock()</tt> are not free. 3163What then? 3164 3165<p> 3166You <i>could</i> wait on all three grace periods in succession, as follows: 3167 3168<blockquote> 3169<pre> 3170 1 synchronize_rcu(); 3171 2 synchronize_rcu_bh(); 3172 3 synchronize_sched(); 3173</pre> 3174</blockquote> 3175 3176<p> 3177This works, but triples the update-side latency penalty. 3178In cases where this is not acceptable, <tt>synchronize_rcu_mult()</tt> 3179may be used to wait on all three flavors of grace period concurrently: 3180 3181<blockquote> 3182<pre> 3183 1 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched); 3184</pre> 3185</blockquote> 3186 3187<p> 3188But what if it is necessary to also wait on SRCU? 3189This can be done as follows: 3190 3191<blockquote> 3192<pre> 3193 1 static void call_my_srcu(struct rcu_head *head, 3194 2 void (*func)(struct rcu_head *head)) 3195 3 { 3196 4 call_srcu(&my_srcu, head, func); 3197 5 } 3198 6 3199 7 synchronize_rcu_mult(call_rcu, call_rcu_bh, call_rcu_sched, call_my_srcu); 3200</pre> 3201</blockquote> 3202 3203<p> 3204If you needed to wait on multiple different flavors of SRCU 3205(but why???), you would need to create a wrapper function resembling 3206<tt>call_my_srcu()</tt> for each SRCU flavor. 3207 3208<table> 3209<tr><th> </th></tr> 3210<tr><th align="left">Quick Quiz:</th></tr> 3211<tr><td> 3212 But what if I need to wait for multiple RCU flavors, but I also need 3213 the grace periods to be expedited? 3214</td></tr> 3215<tr><th align="left">Answer:</th></tr> 3216<tr><td bgcolor="#ffffff"><font color="ffffff"> 3217 If you are using expedited grace periods, there should be less penalty 3218 for waiting on them in succession. 3219 But if that is nevertheless a problem, you can use workqueues 3220 or multiple kthreads to wait on the various expedited grace 3221 periods concurrently. 3222</font></td></tr> 3223<tr><td> </td></tr> 3224</table> 3225 3226<p> 3227Again, it is usually better to adjust the RCU read-side critical sections 3228to use a single flavor of RCU, but when this is not feasible, you can use 3229<tt>synchronize_rcu_mult()</tt>. 3230 3231<h2><a name="Possible Future Changes">Possible Future Changes</a></h2> 3232 3233<p> 3234One of the tricks that RCU uses to attain update-side scalability is 3235to increase grace-period latency with increasing numbers of CPUs. 3236If this becomes a serious problem, it will be necessary to rework the 3237grace-period state machine so as to avoid the need for the additional 3238latency. 3239 3240<p> 3241Expedited grace periods scan the CPUs, so their latency and overhead 3242increases with increasing numbers of CPUs. 3243If this becomes a serious problem on large systems, it will be necessary 3244to do some redesign to avoid this scalability problem. 3245 3246<p> 3247RCU disables CPU hotplug in a few places, perhaps most notably in the 3248<tt>rcu_barrier()</tt> operations. 3249If there is a strong reason to use <tt>rcu_barrier()</tt> in CPU-hotplug 3250notifiers, it will be necessary to avoid disabling CPU hotplug. 3251This would introduce some complexity, so there had better be a <i>very</i> 3252good reason. 3253 3254<p> 3255The tradeoff between grace-period latency on the one hand and interruptions 3256of other CPUs on the other hand may need to be re-examined. 3257The desire is of course for zero grace-period latency as well as zero 3258interprocessor interrupts undertaken during an expedited grace period 3259operation. 3260While this ideal is unlikely to be achievable, it is quite possible that 3261further improvements can be made. 3262 3263<p> 3264The multiprocessor implementations of RCU use a combining tree that 3265groups CPUs so as to reduce lock contention and increase cache locality. 3266However, this combining tree does not spread its memory across NUMA 3267nodes nor does it align the CPU groups with hardware features such 3268as sockets or cores. 3269Such spreading and alignment is currently believed to be unnecessary 3270because the hotpath read-side primitives do not access the combining 3271tree, nor does <tt>call_rcu()</tt> in the common case. 3272If you believe that your architecture needs such spreading and alignment, 3273then your architecture should also benefit from the 3274<tt>rcutree.rcu_fanout_leaf</tt> boot parameter, which can be set 3275to the number of CPUs in a socket, NUMA node, or whatever. 3276If the number of CPUs is too large, use a fraction of the number of 3277CPUs. 3278If the number of CPUs is a large prime number, well, that certainly 3279is an “interesting” architectural choice! 3280More flexible arrangements might be considered, but only if 3281<tt>rcutree.rcu_fanout_leaf</tt> has proven inadequate, and only 3282if the inadequacy has been demonstrated by a carefully run and 3283realistic system-level workload. 3284 3285<p> 3286Please note that arrangements that require RCU to remap CPU numbers will 3287require extremely good demonstration of need and full exploration of 3288alternatives. 3289 3290<p> 3291There is an embarrassingly large number of flavors of RCU, and this 3292number has been increasing over time. 3293Perhaps it will be possible to combine some at some future date. 3294 3295<p> 3296RCU's various kthreads are reasonably recent additions. 3297It is quite likely that adjustments will be required to more gracefully 3298handle extreme loads. 3299It might also be necessary to be able to relate CPU utilization by 3300RCU's kthreads and softirq handlers to the code that instigated this 3301CPU utilization. 3302For example, RCU callback overhead might be charged back to the 3303originating <tt>call_rcu()</tt> instance, though probably not 3304in production kernels. 3305 3306<h2><a name="Summary">Summary</a></h2> 3307 3308<p> 3309This document has presented more than two decade's worth of RCU 3310requirements. 3311Given that the requirements keep changing, this will not be the last 3312word on this subject, but at least it serves to get an important 3313subset of the requirements set forth. 3314 3315<h2><a name="Acknowledgments">Acknowledgments</a></h2> 3316 3317I am grateful to Steven Rostedt, Lai Jiangshan, Ingo Molnar, 3318Oleg Nesterov, Borislav Petkov, Peter Zijlstra, Boqun Feng, and 3319Andy Lutomirski for their help in rendering 3320this article human readable, and to Michelle Rankin for her support 3321of this effort. 3322Other contributions are acknowledged in the Linux kernel's git archive. 3323 3324</body></html> 3325
