1 <!-- DO NOT HAND EDIT. -->
2 <!-- Instead, edit Data-Structures.htmlx and run 'sh htmlqqz.sh Data-Structures' -->
3 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN"
4 "http://www.w3.org/TR/html4/loose.dtd">
6 <head><title>A Tour Through TREE_RCU's Data Structures [LWN.net]</title>
7 <meta HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
9 <p>January 27, 2016</p>
10 <p>This article was contributed by Paul E. McKenney</p>
14 This document describes RCU's major data structures and their relationship
18 <li> <a href="#Data-Structure Relationships">
19 Data-Structure Relationships</a>
20 <li> <a href="#The rcu_state Structure">
21 The <tt>rcu_state</tt> Structure</a>
22 <li> <a href="#The rcu_node Structure">
23 The <tt>rcu_node</tt> Structure</a>
24 <li> <a href="#The rcu_data Structure">
25 The <tt>rcu_data</tt> Structure</a>
26 <li> <a href="#The rcu_dynticks Structure">
27 The <tt>rcu_dynticks</tt> Structure</a>
28 <li> <a href="#The rcu_head Structure">
29 The <tt>rcu_head</tt> Structure</a>
30 <li> <a href="#RCU-Specific Fields in the task_struct Structure">
31 RCU-Specific Fields in the <tt>task_struct</tt> Structure</a>
32 <li> <a href="#Accessor Functions">
33 Accessor Functions</a>
36 At the end we have the
37 <a href="#Answers to Quick Quizzes">answers to the quick quizzes</a>.
39 <h3><a name="Data-Structure Relationships">Data-Structure Relationships</a></h3>
41 <p>RCU is for all intents and purposes a large state machine, and its
42 data structures maintain the state in such a way as to allow RCU readers
43 to execute extremely quickly, while also processing the RCU grace periods
44 requested by updaters in an efficient and extremely scalable fashion.
45 The efficiency and scalability of RCU updaters is provided primarily
46 by a combining tree, as shown below:
48 </p><p><img src="BigTreeClassicRCU.svg" alt="BigTreeClassicRCU.svg" width="30%">
50 </p><p>This diagram shows an enclosing <tt>rcu_state</tt> structure
51 containing a tree of <tt>rcu_node</tt> structures.
52 Each leaf node of the <tt>rcu_node</tt> tree has up to 16
53 <tt>rcu_data</tt> structures associated with it, so that there
54 are <tt>NR_CPUS</tt> number of <tt>rcu_data</tt> structures,
55 one for each possible CPU.
56 This structure is adjusted at boot time, if needed, to handle the
57 common case where <tt>nr_cpu_ids</tt> is much less than
59 For example, a number of Linux distributions set <tt>NR_CPUs=4096</tt>,
60 which results in a three-level <tt>rcu_node</tt> tree.
61 If the actual hardware has only 16 CPUs, RCU will adjust itself
62 at boot time, resulting in an <tt>rcu_node</tt> tree with only a single node.
64 </p><p>The purpose of this combining tree is to allow per-CPU events
65 such as quiescent states, dyntick-idle transitions,
66 and CPU hotplug operations to be processed efficiently
68 Quiescent states are recorded by the per-CPU <tt>rcu_data</tt> structures,
69 and other events are recorded by the leaf-level <tt>rcu_node</tt>
71 All of these events are combined at each level of the tree until finally
72 grace periods are completed at the tree's root <tt>rcu_node</tt>
74 A grace period can be completed at the root once every CPU
75 (or, in the case of <tt>CONFIG_TREE_PREEMPT_RCU</tt>, task)
76 has passed through a quiescent state.
77 Once a grace period has completed, record of that fact is propagated
80 </p><p>As can be seen from the diagram, on a 64-bit system
81 a two-level tree with 64 leaves can accommodate 1,024 CPUs, with a fanout
82 of 64 at the root and a fanout of 16 at the leaves.
84 <p><a name="Quick Quiz 1"><b>Quick Quiz 1</b>:</a>
85 Why isn't the fanout at the leaves also 64?
86 <br><a href="#qq1answer">Answer</a>
88 </p><p>If your system has more than 1,024 CPUs (or more than 512 CPUs on
89 a 32-bit system), then RCU will automatically add more levels to the
91 For example, if you are crazy enough to build a 64-bit system with 65,536
92 CPUs, RCU would configure the <tt>rcu_node</tt> tree as follows:
94 </p><p><img src="HugeTreeClassicRCU.svg" alt="HugeTreeClassicRCU.svg" width="50%">
96 </p><p>RCU currently permits up to a four-level tree, which on a 64-bit system
97 accommodates up to 4,194,304 CPUs, though only a mere 524,288 CPUs for
99 On the other hand, you can set <tt>CONFIG_RCU_FANOUT</tt> to be
100 as small as 2 if you wish, which would permit only 16 CPUs, which
101 is useful for testing.
103 </p><p>The Linux kernel actually supports multiple flavors of RCU
104 running concurrently, so RCU builds separate data structures for each
106 For example, for <tt>CONFIG_TREE_RCU=y</tt> kernels, RCU provides
107 rcu_sched and rcu_bh, as shown below:
109 </p><p><img src="BigTreeClassicRCUBH.svg" alt="BigTreeClassicRCUBH.svg" width="33%">
111 </p><p>Energy efficiency is increasingly important, and for that
112 reason the Linux kernel provides <tt>CONFIG_NO_HZ_IDLE</tt>, which
113 turns off the scheduling-clock interrupts on idle CPUs, which in
114 turn allows those CPUs to attain deeper sleep states and to consume
116 CPUs whose scheduling-clock interrupts have been turned off are
117 said to be in <i>dyntick-idle mode</i>.
118 RCU must handle dyntick-idle CPUs specially
119 because RCU would otherwise wake up each CPU on every grace period,
120 which would defeat the whole purpose of <tt>CONFIG_NO_HZ_IDLE</tt>.
121 RCU uses the <tt>rcu_dynticks</tt> structure to track
122 which CPUs are in dyntick idle mode, as shown below:
124 </p><p><img src="BigTreeClassicRCUBHdyntick.svg" alt="BigTreeClassicRCUBHdyntick.svg" width="33%">
126 </p><p>However, if a CPU is in dyntick-idle mode, it is in that mode
127 for all flavors of RCU.
128 Therefore, a single <tt>rcu_dynticks</tt> structure is allocated per
129 CPU, and all of a given CPU's <tt>rcu_data</tt> structures share
130 that <tt>rcu_dynticks</tt>, as shown in the figure.
132 </p><p>Kernels built with <tt>CONFIG_TREE_PREEMPT_RCU</tt> support
133 rcu_preempt in addition to rcu_sched and rcu_bh, as shown below:
135 </p><p><img src="BigTreePreemptRCUBHdyntick.svg" alt="BigTreePreemptRCUBHdyntick.svg" width="35%">
137 </p><p>RCU updaters wait for normal grace periods by registering
138 RCU callbacks, either directly via <tt>call_rcu()</tt> and
139 friends (namely <tt>call_rcu_bh()</tt> and <tt>call_rcu_sched()</tt>),
140 there being a separate interface per flavor of RCU)
141 or indirectly via <tt>synchronize_rcu()</tt> and friends.
142 RCU callbacks are represented by <tt>rcu_head</tt> structures,
143 which are queued on <tt>rcu_data</tt> structures while they are
144 waiting for a grace period to elapse, as shown in the following figure:
146 </p><p><img src="BigTreePreemptRCUBHdyntickCB.svg" alt="BigTreePreemptRCUBHdyntickCB.svg" width="40%">
148 </p><p>This figure shows how <tt>TREE_RCU</tt>'s and
149 <tt>TREE_PREEMPT_RCU</tt>'s major data structures are related.
150 Lesser data structures will be introduced with the algorithms that
153 </p><p>Note that each of the data structures in the above figure has
154 its own synchronization:
157 <li> Each <tt>rcu_state</tt> structures has a lock and a mutex,
158 and some fields are protected by the corresponding root
159 <tt>rcu_node</tt> structure's lock.
160 <li> Each <tt>rcu_node</tt> structure has a spinlock.
161 <li> The fields in <tt>rcu_data</tt> are private to the corresponding
162 CPU, although a few can be read and written by other CPUs.
163 <li> Similarly, the fields in <tt>rcu_dynticks</tt> are private
164 to the corresponding CPU, although a few can be read by
168 <p>It is important to note that different data structures can have
169 very different ideas about the state of RCU at any given time.
170 For but one example, awareness of the start or end of a given RCU
171 grace period propagates slowly through the data structures.
172 This slow propagation is absolutely necessary for RCU to have good
173 read-side performance.
174 If this balkanized implementation seems foreign to you, one useful
175 trick is to consider each instance of these data structures to be
176 a different person, each having the usual slightly different
179 </p><p>The general role of each of these data structures is as
183 <li> <tt>rcu_state</tt>:
184 This structure forms the interconnection between the
185 <tt>rcu_node</tt> and <tt>rcu_data</tt> structures,
186 tracks grace periods, serves as short-term repository
187 for callbacks orphaned by CPU-hotplug events,
188 maintains <tt>rcu_barrier()</tt> state,
189 tracks expedited grace-period state,
190 and maintains state used to force quiescent states when
191 grace periods extend too long,
192 <li> <tt>rcu_node</tt>: This structure forms the combining
193 tree that propagates quiescent-state
194 information from the leaves to the root, and also propagates
195 grace-period information from the root to the leaves.
196 It provides local copies of the grace-period state in order
197 to allow this information to be accessed in a synchronized
198 manner without suffering the scalability limitations that
199 would otherwise be imposed by global locking.
200 In <tt>CONFIG_TREE_PREEMPT_RCU</tt> kernels, it manages the lists
201 of tasks that have blocked while in their current
202 RCU read-side critical section.
203 In <tt>CONFIG_TREE_PREEMPT_RCU</tt> with
204 <tt>CONFIG_RCU_BOOST</tt>, it manages the
205 per-<tt>rcu_node</tt> priority-boosting
206 kernel threads (kthreads) and state.
207 Finally, it records CPU-hotplug state in order to determine
208 which CPUs should be ignored during a given grace period.
209 <li> <tt>rcu_data</tt>: This per-CPU structure is the
210 focus of quiescent-state detection and RCU callback queuing.
211 It also tracks its relationship to the corresponding leaf
212 <tt>rcu_node</tt> structure to allow more-efficient
213 propagation of quiescent states up the <tt>rcu_node</tt>
215 Like the <tt>rcu_node</tt> structure, it provides a local
216 copy of the grace-period information to allow for-free
218 access to this information from the corresponding CPU.
219 Finally, this structure records past dyntick-idle state
220 for the corresponding CPU and also tracks statistics.
221 <li> <tt>rcu_dynticks</tt>:
222 This per-CPU structure tracks the current dyntick-idle
223 state for the corresponding CPU.
224 Unlike the other three structures, the <tt>rcu_dynticks</tt>
225 structure is not replicated per RCU flavor.
226 <li> <tt>rcu_head</tt>:
227 This structure represents RCU callbacks, and is the
228 only structure allocated and managed by RCU users.
229 The <tt>rcu_head</tt> structure is normally embedded
230 within the RCU-protected data structure.
233 <p>If all you wanted from this article was a general notion of how
234 RCU's data structures are related, you are done.
235 Otherwise, each of the following sections give more details on
236 the <tt>rcu_state</tt>, <tt>rcu_node</tt>, <tt>rcu_data</tt>,
237 and <tt>rcu_dynticks</tt> data structures.
239 <h3><a name="The rcu_state Structure">
240 The <tt>rcu_state</tt> Structure</a></h3>
242 <p>The <tt>rcu_state</tt> structure is the base structure that
243 represents a flavor of RCU.
244 This structure forms the interconnection between the
245 <tt>rcu_node</tt> and <tt>rcu_data</tt> structures,
246 tracks grace periods, contains the lock used to
247 synchronize with CPU-hotplug events,
248 and maintains state used to force quiescent states when
249 grace periods extend too long,
251 </p><p>A few of the <tt>rcu_state</tt> structure's fields are discussed,
252 singly and in groups, in the following sections.
253 The more specialized fields are covered in the discussion of their
256 <h5>Relationship to rcu_node and rcu_data Structures</h5>
258 This portion of the <tt>rcu_state</tt> structure is declared
262 1 struct rcu_node node[NUM_RCU_NODES];
263 2 struct rcu_node *level[NUM_RCU_LVLS + 1];
264 3 struct rcu_data __percpu *rda;
267 <p><a name="Quick Quiz 2"><b>Quick Quiz 2</b>:</a>
269 You said that the <tt>rcu_node</tt> structures formed a tree,
270 but they are declared as a flat array!
272 <br><a href="#qq2answer">Answer</a>
274 </p><p>The <tt>rcu_node</tt> tree is embedded into the
275 <tt>->node[]</tt> array as shown in the following figure:
277 </p><p><img src="TreeMapping.svg" alt="TreeMapping.svg" width="40%">
279 </p><p>One interesting consequence of this mapping is that a
280 breadth-first traversal of the tree is implemented as a simple
281 linear scan of the array, which is in fact what the
282 <tt>rcu_for_each_node_breadth_first()</tt> macro does.
283 This macro is used at the beginning and ends of grace periods.
285 </p><p>Each entry of the <tt>->level</tt> array references
286 the first <tt>rcu_node</tt> structure on the corresponding level
287 of the tree, for example, as shown below:
289 </p><p><img src="TreeMappingLevel.svg" alt="TreeMappingLevel.svg" width="40%">
291 </p><p>The zero<sup>th</sup> element of the array references the root
292 <tt>rcu_node</tt> structure, the first element references the
293 first child of the root <tt>rcu_node</tt>, and finally the second
294 element references the first leaf <tt>rcu_node</tt> structure.
296 <p><a name="Quick Quiz 3"><b>Quick Quiz 3</b>:</a>
297 Given that this array represents a tree, why can't the diagram that
298 includes the <tt>->level</tt> array be planar?
299 <br><a href="#qq3answer">Answer</a>
301 </p><p>Finally, the <tt>->rda</tt> field references a per-CPU
302 pointer to the corresponding CPU's <tt>rcu_data</tt> structure.
304 </p><p>All of these fields are constant once initialization is complete,
305 and therefore need no protection.
307 <h5>Grace-Period Tracking</h5>
309 <p>This portion of the <tt>rcu_state</tt> structure is declared
313 1 unsigned long gpnum;
314 2 unsigned long completed;
317 <p>RCU grace periods are numbered, and
318 the <tt>->gpnum</tt> field contains the number of the grace
319 period that started most recently.
320 The <tt>->completed</tt> field contains the number of the
321 grace period that completed most recently.
322 If the two fields are equal, the RCU grace period that most recently
323 started has already completed, and therefore the corresponding
324 flavor of RCU is idle.
325 If <tt>->gpnum</tt> is one greater than <tt>->completed</tt>,
326 then <tt>->gpnum</tt> gives the number of the current RCU
327 grace period, which has not yet completed.
328 Any other combination of values indicates that something is broken.
329 These two fields are protected by the root <tt>rcu_node</tt>'s
330 <tt>->lock</tt> field.
332 </p><p>There are <tt>->gpnum</tt> and <tt>->completed</tt> fields
333 in the <tt>rcu_node</tt> and <tt>rcu_data</tt> structures
335 The fields in the <tt>rcu_state</tt> structure represent the
336 most current values, and those of the other structures are compared
337 in order to detect the start of a new grace period in a distributed
339 The values flow from <tt>rcu_state</tt> to <tt>rcu_node</tt>
340 (down the tree from the root to the leaves) to <tt>rcu_data</tt>.
342 <h5>Miscellaneous</h5>
344 <p>This portion of the <tt>rcu_state</tt> structure is declared
348 1 unsigned long gp_max;
353 <p>The <tt>->gp_max</tt> field tracks the duration of the longest
354 grace period in jiffies.
355 It is protected by the root <tt>rcu_node</tt>'s <tt>->lock</tt>.
357 <p>The <tt>->name</tt> field points to the name of the RCU flavor
358 (for example, “rcu_sched”), and is constant.
359 The <tt>->abbr</tt> field contains a one-character abbreviation,
360 for example, “s” for RCU-sched.
362 <h3><a name="The rcu_node Structure">
363 The <tt>rcu_node</tt> Structure</a></h3>
365 <p>The <tt>rcu_node</tt> structures form the combining
366 tree that propagates quiescent-state
367 information from the leaves to the root and also that propagates
368 grace-period information from the root down to the leaves.
369 They provides local copies of the grace-period state in order
370 to allow this information to be accessed in a synchronized
371 manner without suffering the scalability limitations that
372 would otherwise be imposed by global locking.
373 In <tt>CONFIG_TREE_PREEMPT_RCU</tt> kernels, they manage the lists
374 of tasks that have blocked while in their current
375 RCU read-side critical section.
376 In <tt>CONFIG_TREE_PREEMPT_RCU</tt> with
377 <tt>CONFIG_RCU_BOOST</tt>, they manage the
378 per-<tt>rcu_node</tt> priority-boosting
379 kernel threads (kthreads) and state.
380 Finally, they record CPU-hotplug state in order to determine
381 which CPUs should be ignored during a given grace period.
383 </p><p>The <tt>rcu_node</tt> structure's fields are discussed,
384 singly and in groups, in the following sections.
386 <h5>Connection to Combining Tree</h5>
388 <p>This portion of the <tt>rcu_node</tt> structure is declared
392 1 struct rcu_node *parent;
395 4 unsigned long grpmask;
400 <p>The <tt>->parent</tt> pointer references the <tt>rcu_node</tt>
401 one level up in the tree, and is <tt>NULL</tt> for the root
403 The RCU implementation makes heavy use of this field to push quiescent
405 The <tt>->level</tt> field gives the level in the tree, with
406 the root being at level zero, its children at level one, and so on.
407 The <tt>->grpnum</tt> field gives this node's position within
408 the children of its parent, so this number can range between 0 and 31
409 on 32-bit systems and between 0 and 63 on 64-bit systems.
410 The <tt>->level</tt> and <tt>->grpnum</tt> fields are
411 used only during initialization and for tracing.
412 The <tt>->grpmask</tt> field is the bitmask counterpart of
413 <tt>->grpnum</tt>, and therefore always has exactly one bit set.
414 This mask is used to clear the bit corresponding to this <tt>rcu_node</tt>
415 structure in its parent's bitmasks, which are described later.
416 Finally, the <tt>->grplo</tt> and <tt>->grphi</tt> fields
417 contain the lowest and highest numbered CPU served by this
418 <tt>rcu_node</tt> structure, respectively.
420 </p><p>All of these fields are constant, and thus do not require any
423 <h5>Synchronization</h5>
425 <p>This field of the <tt>rcu_node</tt> structure is declared
429 1 raw_spinlock_t lock;
432 <p>This field is used to protect the remaining fields in this structure,
433 unless otherwise stated.
434 That said, all of the fields in this structure can be accessed without
435 locking for tracing purposes.
436 Yes, this can result in confusing traces, but better some tracing confusion
437 than to be heisenbugged out of existence.
439 <h5>Grace-Period Tracking</h5>
441 <p>This portion of the <tt>rcu_node</tt> structure is declared
445 1 unsigned long gpnum;
446 2 unsigned long completed;
449 <p>These fields are the counterparts of the fields of the same name in
450 the <tt>rcu_state</tt> structure.
451 They each may lag up to one behind their <tt>rcu_state</tt>
453 If a given <tt>rcu_node</tt> structure's <tt>->gpnum</tt> and
454 <tt>->complete</tt> fields are equal, then this <tt>rcu_node</tt>
455 structure believes that RCU is idle.
456 Otherwise, as with the <tt>rcu_state</tt> structure,
457 the <tt>->gpnum</tt> field will be one greater than the
458 <tt>->complete</tt> fields, with <tt>->gpnum</tt>
459 indicating which grace period this <tt>rcu_node</tt> believes
460 is still being waited for.
462 </p><p>The <tt>>gpnum</tt> field of each <tt>rcu_node</tt>
463 structure is updated at the beginning
464 of each grace period, and the <tt>->completed</tt> fields are
465 updated at the end of each grace period.
467 <h5>Quiescent-State Tracking</h5>
469 <p>These fields manage the propagation of quiescent states up the
472 </p><p>This portion of the <tt>rcu_node</tt> structure has fields
476 1 unsigned long qsmask;
477 2 unsigned long expmask;
478 3 unsigned long qsmaskinit;
479 4 unsigned long expmaskinit;
482 <p>The <tt>->qsmask</tt> field tracks which of this
483 <tt>rcu_node</tt> structure's children still need to report
484 quiescent states for the current normal grace period.
485 Such children will have a value of 1 in their corresponding bit.
486 Note that the leaf <tt>rcu_node</tt> structures should be
487 thought of as having <tt>rcu_data</tt> structures as their
489 Similarly, the <tt>->expmask</tt> field tracks which
490 of this <tt>rcu_node</tt> structure's children still need to report
491 quiescent states for the current expedited grace period.
492 An expedited grace period has
493 the same conceptual properties as a normal grace period, but the
494 expedited implementation accepts extreme CPU overhead to obtain
495 much lower grace-period latency, for example, consuming a few
496 tens of microseconds worth of CPU time to reduce grace-period
497 duration from milliseconds to tens of microseconds.
498 The <tt>->qsmaskinit</tt> field tracks which of this
499 <tt>rcu_node</tt> structure's children cover for at least
501 This mask is used to initialize <tt>->qsmask</tt>,
502 and <tt>->expmaskinit</tt> is used to initialize
503 <tt>->expmask</tt> and the beginning of the
504 normal and expedited grace periods, respectively.
506 <p><a name="Quick Quiz 4"><b>Quick Quiz 4</b>:</a>
507 Why are these bitmasks protected by locking?
508 Come on, haven't you heard of atomic instructions???
509 <br><a href="#qq4answer">Answer</a>
511 <h5>Blocked-Task Management</h5>
513 <p><tt>TREE_PREEMPT_RCU</tt> allows tasks to be preempted in the
514 midst of their RCU read-side critical sections, and these tasks
515 must be tracked explicitly.
516 The details of exactly why and how they are tracked will be covered
517 in a separate article on RCU read-side processing.
518 For now, it is enough to know that the <tt>rcu_node</tt>
519 structure tracks them.
522 1 struct list_head blkd_tasks;
523 2 struct list_head *gp_tasks;
524 3 struct list_head *exp_tasks;
525 4 bool wait_blkd_tasks;
528 <p>The <tt>->blkd_tasks</tt> field is a list header for
529 the list of blocked and preempted tasks.
530 As tasks undergo context switches within RCU read-side critical
531 sections, their <tt>task_struct</tt> structures are enqueued
532 (via the <tt>task_struct</tt>'s <tt>->rcu_node_entry</tt>
533 field) onto the head of the <tt>->blkd_tasks</tt> list for the
534 leaf <tt>rcu_node</tt> structure corresponding to the CPU
535 on which the outgoing context switch executed.
536 As these tasks later exit their RCU read-side critical sections,
537 they remove themselves from the list.
538 This list is therefore in reverse time order, so that if one of the tasks
539 is blocking the current grace period, all subsequent tasks must
540 also be blocking that same grace period.
541 Therefore, a single pointer into this list suffices to track
542 all tasks blocking a given grace period.
543 That pointer is stored in <tt>->gp_tasks</tt> for normal
544 grace periods and in <tt>->exp_tasks</tt> for expedited
546 These last two fields are <tt>NULL</tt> if either there is
547 no grace period in flight or if there are no blocked tasks
548 preventing that grace period from completing.
549 If either of these two pointers is referencing a task that
550 removes itself from the <tt>->blkd_tasks</tt> list,
551 then that task must advance the pointer to the next task on
552 the list, or set the pointer to <tt>NULL</tt> if there
553 are no subsequent tasks on the list.
555 </p><p>For example, suppose that tasks T1, T2, and T3 are
556 all hard-affinitied to the largest-numbered CPU in the system.
557 Then if task T1 blocked in an RCU read-side
558 critical section, then an expedited grace period started,
559 then task T2 blocked in an RCU read-side critical section,
560 then a normal grace period started, and finally task 3 blocked
561 in an RCU read-side critical section, then the state of the
562 last leaf <tt>rcu_node</tt> structure's blocked-task list
563 would be as shown below:
565 </p><p><img src="blkd_task.svg" alt="blkd_task.svg" width="60%">
567 </p><p>Task T1 is blocking both grace periods, task T2 is
568 blocking only the normal grace period, and task T3 is blocking
569 neither grace period.
570 Note that these tasks will not remove themselves from this list
571 immediately upon resuming execution.
572 They will instead remain on the list until they execute the outermost
573 <tt>rcu_read_unlock()</tt> that ends their RCU read-side critical
577 The <tt>->wait_blkd_tasks</tt> field indicates whether or not
578 the current grace period is waiting on a blocked task.
580 <h5>Sizing the <tt>rcu_node</tt> Array</h5>
582 <p>The <tt>rcu_node</tt> array is sized via a series of
583 C-preprocessor expressions as follows:
586 1 #ifdef CONFIG_RCU_FANOUT
587 2 #define RCU_FANOUT CONFIG_RCU_FANOUT
589 4 # ifdef CONFIG_64BIT
590 5 # define RCU_FANOUT 64
592 7 # define RCU_FANOUT 32
596 11 #ifdef CONFIG_RCU_FANOUT_LEAF
597 12 #define RCU_FANOUT_LEAF CONFIG_RCU_FANOUT_LEAF
599 14 # ifdef CONFIG_64BIT
600 15 # define RCU_FANOUT_LEAF 64
602 17 # define RCU_FANOUT_LEAF 32
606 21 #define RCU_FANOUT_1 (RCU_FANOUT_LEAF)
607 22 #define RCU_FANOUT_2 (RCU_FANOUT_1 * RCU_FANOUT)
608 23 #define RCU_FANOUT_3 (RCU_FANOUT_2 * RCU_FANOUT)
609 24 #define RCU_FANOUT_4 (RCU_FANOUT_3 * RCU_FANOUT)
611 26 #if NR_CPUS <= RCU_FANOUT_1
612 27 # define RCU_NUM_LVLS 1
613 28 # define NUM_RCU_LVL_0 1
614 29 # define NUM_RCU_NODES NUM_RCU_LVL_0
615 30 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0 }
616 31 # define RCU_NODE_NAME_INIT { "rcu_node_0" }
617 32 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0" }
618 33 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0" }
619 34 #elif NR_CPUS <= RCU_FANOUT_2
620 35 # define RCU_NUM_LVLS 2
621 36 # define NUM_RCU_LVL_0 1
622 37 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
623 38 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1)
624 39 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1 }
625 40 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1" }
626 41 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1" }
627 42 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1" }
628 43 #elif NR_CPUS <= RCU_FANOUT_3
629 44 # define RCU_NUM_LVLS 3
630 45 # define NUM_RCU_LVL_0 1
631 46 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
632 47 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
633 48 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2)
634 49 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1, NUM_RCU_LVL_2 }
635 50 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1", "rcu_node_2" }
636 51 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1", "rcu_node_fqs_2" }
637 52 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1", "rcu_node_exp_2" }
638 53 #elif NR_CPUS <= RCU_FANOUT_4
639 54 # define RCU_NUM_LVLS 4
640 55 # define NUM_RCU_LVL_0 1
641 56 # define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_3)
642 57 # define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
643 58 # define NUM_RCU_LVL_3 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
644 59 # define NUM_RCU_NODES (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2 + NUM_RCU_LVL_3)
645 60 # define NUM_RCU_LVL_INIT { NUM_RCU_LVL_0, NUM_RCU_LVL_1, NUM_RCU_LVL_2, NUM_RCU_LVL_3 }
646 61 # define RCU_NODE_NAME_INIT { "rcu_node_0", "rcu_node_1", "rcu_node_2", "rcu_node_3" }
647 62 # define RCU_FQS_NAME_INIT { "rcu_node_fqs_0", "rcu_node_fqs_1", "rcu_node_fqs_2", "rcu_node_fqs_3" }
648 63 # define RCU_EXP_NAME_INIT { "rcu_node_exp_0", "rcu_node_exp_1", "rcu_node_exp_2", "rcu_node_exp_3" }
650 65 # error "CONFIG_RCU_FANOUT insufficient for NR_CPUS"
654 <p>The maximum number of levels in the <tt>rcu_node</tt> structure
655 is currently limited to four, as specified by lines 21-24
656 and the structure of the subsequent “if” statement.
657 For 32-bit systems, this allows 16*32*32*32=524,288 CPUs, which
658 should be sufficient for the next few years at least.
659 For 64-bit systems, 16*64*64*64=4,194,304 CPUs is allowed, which
660 should see us through the next decade or so.
661 This four-level tree also allows kernels built with
662 <tt>CONFIG_RCU_FANOUT=8</tt> to support up to 4096 CPUs,
663 which might be useful in very large systems having eight CPUs per
664 socket (but please note that no one has yet shown any measurable
665 performance degradation due to misaligned socket and <tt>rcu_node</tt>
667 In addition, building kernels with a full four levels of <tt>rcu_node</tt>
668 tree permits better testing of RCU's combining-tree code.
670 </p><p>The <tt>RCU_FANOUT</tt> symbol controls how many children
671 are permitted at each non-leaf level of the <tt>rcu_node</tt> tree.
672 If the <tt>CONFIG_RCU_FANOUT</tt> Kconfig option is not specified,
673 it is set based on the word size of the system, which is also
676 </p><p>The <tt>RCU_FANOUT_LEAF</tt> symbol controls how many CPUs are
677 handled by each leaf <tt>rcu_node</tt> structure.
678 Experience has shown that allowing a given leaf <tt>rcu_node</tt>
679 structure to handle 64 CPUs, as permitted by the number of bits in
680 the <tt>->qsmask</tt> field on a 64-bit system, results in
681 excessive contention for the leaf <tt>rcu_node</tt> structures'
682 <tt>->lock</tt> fields.
683 The number of CPUs per leaf <tt>rcu_node</tt> structure is therefore
684 limited to 16 given the default value of <tt>CONFIG_RCU_FANOUT_LEAF</tt>.
685 If <tt>CONFIG_RCU_FANOUT_LEAF</tt> is unspecified, the value
686 selected is based on the word size of the system, just as for
687 <tt>CONFIG_RCU_FANOUT</tt>.
688 Lines 11-19 perform this computation.
690 </p><p>Lines 21-24 compute the maximum number of CPUs supported by
691 a single-level (which contains a single <tt>rcu_node</tt> structure),
692 two-level, three-level, and four-level <tt>rcu_node</tt> tree,
693 respectively, given the fanout specified by <tt>RCU_FANOUT</tt>
694 and <tt>RCU_FANOUT_LEAF</tt>.
695 These numbers of CPUs are retained in the
696 <tt>RCU_FANOUT_1</tt>,
697 <tt>RCU_FANOUT_2</tt>,
698 <tt>RCU_FANOUT_3</tt>, and
699 <tt>RCU_FANOUT_4</tt>
700 C-preprocessor variables, respectively.
702 </p><p>These variables are used to control the C-preprocessor <tt>#if</tt>
703 statement spanning lines 26-66 that computes the number of
704 <tt>rcu_node</tt> structures required for each level of the tree,
705 as well as the number of levels required.
706 The number of levels is placed in the <tt>NUM_RCU_LVLS</tt>
707 C-preprocessor variable by lines 27, 35, 44, and 54.
708 The number of <tt>rcu_node</tt> structures for the topmost level
709 of the tree is always exactly one, and this value is unconditionally
710 placed into <tt>NUM_RCU_LVL_0</tt> by lines 28, 36, 45, and 55.
711 The rest of the levels (if any) of the <tt>rcu_node</tt> tree
712 are computed by dividing the maximum number of CPUs by the
713 fanout supported by the number of levels from the current level down,
714 rounding up. This computation is performed by lines 37,
715 46-47, and 56-58.
716 Lines 31-33, 40-42, 50-52, and 62-63 create initializers
717 for lockdep lock-class names.
718 Finally, lines 64-66 produce an error if the maximum number of
719 CPUs is too large for the specified fanout.
721 <h3><a name="The rcu_data Structure">
722 The <tt>rcu_data</tt> Structure</a></h3>
724 <p>The <tt>rcu_data</tt> maintains the per-CPU state for the
725 corresponding flavor of RCU.
726 The fields in this structure may be accessed only from the corresponding
727 CPU (and from tracing) unless otherwise stated.
728 This structure is the
729 focus of quiescent-state detection and RCU callback queuing.
730 It also tracks its relationship to the corresponding leaf
731 <tt>rcu_node</tt> structure to allow more-efficient
732 propagation of quiescent states up the <tt>rcu_node</tt>
734 Like the <tt>rcu_node</tt> structure, it provides a local
735 copy of the grace-period information to allow for-free
737 access to this information from the corresponding CPU.
738 Finally, this structure records past dyntick-idle state
739 for the corresponding CPU and also tracks statistics.
741 </p><p>The <tt>rcu_data</tt> structure's fields are discussed,
742 singly and in groups, in the following sections.
744 <h5>Connection to Other Data Structures</h5>
746 <p>This portion of the <tt>rcu_data</tt> structure is declared
751 2 struct rcu_state *rsp;
752 3 struct rcu_node *mynode;
753 4 struct rcu_dynticks *dynticks;
754 5 unsigned long grpmask;
758 <p>The <tt>->cpu</tt> field contains the number of the
759 corresponding CPU, the <tt>->rsp</tt> pointer references
760 the corresponding <tt>rcu_state</tt> structure (and is most frequently
761 used to locate the name of the corresponding flavor of RCU for tracing),
762 and the <tt>->mynode</tt> field references the corresponding
763 <tt>rcu_node</tt> structure.
764 The <tt>->mynode</tt> is used to propagate quiescent states
765 up the combining tree.
766 <p>The <tt>->dynticks</tt> pointer references the
767 <tt>rcu_dynticks</tt> structure corresponding to this
769 Recall that a single per-CPU instance of the <tt>rcu_dynticks</tt>
770 structure is shared among all flavors of RCU.
771 These first four fields are constant and therefore require not
774 </p><p>The <tt>->grpmask</tt> field indicates the bit in
775 the <tt>->mynode->qsmask</tt> corresponding to this
776 <tt>rcu_data</tt> structure, and is also used when propagating
778 The <tt>->beenonline</tt> flag is set whenever the corresponding
779 CPU comes online, which means that the debugfs tracing need not dump
780 out any <tt>rcu_data</tt> structure for which this flag is not set.
782 <h5>Quiescent-State and Grace-Period Tracking</h5>
784 <p>This portion of the <tt>rcu_data</tt> structure is declared
788 1 unsigned long completed;
789 2 unsigned long gpnum;
791 4 bool core_needs_qs;
793 6 unsigned long rcu_qs_ctr_snap;
796 <p>The <tt>completed</tt> and <tt>gpnum</tt>
797 fields are the counterparts of the fields of the same name
798 in the <tt>rcu_state</tt> and <tt>rcu_node</tt> structures.
799 They may each lag up to one behind their <tt>rcu_node</tt>
800 counterparts, but in <tt>CONFIG_NO_HZ_IDLE</tt> and
801 <tt>CONFIG_NO_HZ_FULL</tt> kernels can lag
802 arbitrarily far behind for CPUs in dyntick-idle mode (but these counters
803 will catch up upon exit from dyntick-idle mode).
804 If a given <tt>rcu_data</tt> structure's <tt>->gpnum</tt> and
805 <tt>->complete</tt> fields are equal, then this <tt>rcu_data</tt>
806 structure believes that RCU is idle.
807 Otherwise, as with the <tt>rcu_state</tt> and <tt>rcu_node</tt>
809 the <tt>->gpnum</tt> field will be one greater than the
810 <tt>->complete</tt> fields, with <tt>->gpnum</tt>
811 indicating which grace period this <tt>rcu_data</tt> believes
812 is still being waited for.
814 <p><a name="Quick Quiz 5"><b>Quick Quiz 5</b>:</a>
815 All this replication of the grace period numbers can only cause
817 Why not just keep a global pair of counters and be done with it???
818 <br><a href="#qq5answer">Answer</a>
820 </p><p>The <tt>->cpu_no_qs</tt> flag indicates that the
821 CPU has not yet passed through a quiescent state,
822 while the <tt>->core_needs_qs</tt> flag indicates that the
823 RCU core needs a quiescent state from the corresponding CPU.
824 The <tt>->gpwrap</tt> field indicates that the corresponding
825 CPU has remained idle for so long that the <tt>completed</tt>
826 and <tt>gpnum</tt> counters are in danger of overflow, which
827 will cause the CPU to disregard the values of its counters on
828 its next exit from idle.
829 Finally, the <tt>rcu_qs_ctr_snap</tt> field is used to detect
830 cases where a given operation has resulted in a quiescent state
831 for all flavors of RCU, for example, <tt>cond_resched_rcu_qs()</tt>.
833 <h5>RCU Callback Handling</h5>
835 <p>In the absence of CPU-hotplug events, RCU callbacks are invoked by
836 the same CPU that registered them.
837 This is strictly a cache-locality optimization: callbacks can and
838 do get invoked on CPUs other than the one that registered them.
839 After all, if the CPU that registered a given callback has gone
840 offline before the callback can be invoked, there really is no other
843 </p><p>This portion of the <tt>rcu_data</tt> structure is declared
847 1 struct rcu_head *nxtlist;
848 2 struct rcu_head **nxttail[RCU_NEXT_SIZE];
849 3 unsigned long nxtcompleted[RCU_NEXT_SIZE];
852 6 long qlen_last_fqs_check;
853 7 unsigned long n_force_qs_snap;
854 8 unsigned long n_cbs_invoked;
855 9 unsigned long n_cbs_orphaned;
856 10 unsigned long n_cbs_adopted;
860 <p>The <tt>->nxtlist</tt> pointer and the
861 <tt>->nxttail[]</tt> array form a four-segment list with
862 older callbacks near the head and newer ones near the tail.
863 Each segment contains callbacks with the corresponding relationship
864 to the current grace period.
865 The pointer out of the end of each of the four segments is referenced
866 by the element of the <tt>->nxttail[]</tt> array indexed by
867 <tt>RCU_DONE_TAIL</tt> (for callbacks handled by a prior grace period),
868 <tt>RCU_WAIT_TAIL</tt> (for callbacks waiting on the current grace period),
869 <tt>RCU_NEXT_READY_TAIL</tt> (for callbacks that will wait on the next
871 <tt>RCU_NEXT_TAIL</tt> (for callbacks that are not yet associated
872 with a specific grace period)
873 respectively, as shown in the following figure.
875 </p><p><img src="nxtlist.svg" alt="nxtlist.svg" width="40%">
877 </p><p>In this figure, the <tt>->nxtlist</tt> pointer references the
879 RCU callback in the list.
880 The <tt>->nxttail[RCU_DONE_TAIL]</tt> array element references
881 the <tt>->nxtlist</tt> pointer itself, indicating that none
882 of the callbacks is ready to invoke.
883 The <tt>->nxttail[RCU_WAIT_TAIL]</tt> array element references callback
884 CB 2's <tt>->next</tt> pointer, which indicates that
885 CB 1 and CB 2 are both waiting on the current grace period.
886 The <tt>->nxttail[RCU_NEXT_READY_TAIL]</tt> array element
887 references the same RCU callback that <tt>->nxttail[RCU_WAIT_TAIL]</tt>
888 does, which indicates that there are no callbacks waiting on the next
890 The <tt>->nxttail[RCU_NEXT_TAIL]</tt> array element references
891 CB 4's <tt>->next</tt> pointer, indicating that all the
892 remaining RCU callbacks have not yet been assigned to an RCU grace
894 Note that the <tt>->nxttail[RCU_NEXT_TAIL]</tt> array element
895 always references the last RCU callback's <tt>->next</tt> pointer
896 unless the callback list is empty, in which case it references
897 the <tt>->nxtlist</tt> pointer.
899 </p><p>CPUs advance their callbacks from the
900 <tt>RCU_NEXT_TAIL</tt> to the <tt>RCU_NEXT_READY_TAIL</tt> to the
901 <tt>RCU_WAIT_TAIL</tt> to the <tt>RCU_DONE_TAIL</tt> list segments
902 as grace periods advance.
903 The CPU advances the callbacks in its <tt>rcu_data</tt> structure
904 whenever it notices that another RCU grace period has completed.
905 The CPU detects the completion of an RCU grace period by noticing
906 that the value of its <tt>rcu_data</tt> structure's
907 <tt>->completed</tt> field differs from that of its leaf
908 <tt>rcu_node</tt> structure.
909 Recall that each <tt>rcu_node</tt> structure's
910 <tt>->completed</tt> field is updated at the end of each
913 </p><p>The <tt>->nxtcompleted[]</tt> array records grace-period
914 numbers corresponding to the list segments.
915 This allows CPUs that go idle for extended periods to determine
916 which of their callbacks are ready to be invoked after reawakening.
918 </p><p>The <tt>->qlen</tt> counter contains the number of
919 callbacks in <tt>->nxtlist</tt>, and the
920 <tt>->qlen_lazy</tt> contains the number of those callbacks that
921 are known to only free memory, and whose invocation can therefore
923 The <tt>->qlen_last_fqs_check</tt> and
924 <tt>->n_force_qs_snap</tt> coordinate the forcing of quiescent
925 states from <tt>call_rcu()</tt> and friends when callback
926 lists grow excessively long.
928 </p><p>The <tt>->n_cbs_invoked</tt>,
929 <tt>->n_cbs_orphaned</tt>, and <tt>->n_cbs_adopted</tt>
930 fields count the number of callbacks invoked,
931 sent to other CPUs when this CPU goes offline,
932 and received from other CPUs when those other CPUs go offline.
933 Finally, the <tt>->blimit</tt> counter is the maximum number of
934 RCU callbacks that may be invoked at a given time.
936 <h5>Dyntick-Idle Handling</h5>
938 <p>This portion of the <tt>rcu_data</tt> structure is declared
943 2 unsigned long dynticks_fqs;
946 The <tt>->dynticks_snap</tt> field is used to take a snapshot
947 of the corresponding CPU's dyntick-idle state when forcing
948 quiescent states, and is therefore accessed from other CPUs.
949 Finally, the <tt>->dynticks_fqs</tt> field is used to
950 count the number of times this CPU is determined to be in
951 dyntick-idle state, and is used for tracing and debugging purposes.
953 <h3><a name="The rcu_dynticks Structure">
954 The <tt>rcu_dynticks</tt> Structure</a></h3>
956 <p>The <tt>rcu_dynticks</tt> maintains the per-CPU dyntick-idle state
957 for the corresponding CPU.
958 Unlike the other structures, <tt>rcu_dynticks</tt> is not
959 replicated over the different flavors of RCU.
960 The fields in this structure may be accessed only from the corresponding
961 CPU (and from tracing) unless otherwise stated.
962 Its fields are as follows:
965 1 int dynticks_nesting;
966 2 int dynticks_nmi_nesting;
970 <p>The <tt>->dynticks_nesting</tt> field counts the
971 nesting depth of normal interrupts.
972 In addition, this counter is incremented when exiting dyntick-idle
973 mode and decremented when entering it.
974 This counter can therefore be thought of as counting the number
975 of reasons why this CPU cannot be permitted to enter dyntick-idle
976 mode, aside from non-maskable interrupts (NMIs).
977 NMIs are counted by the <tt>->dynticks_nmi_nesting</tt>
978 field, except that NMIs that interrupt non-dyntick-idle execution
981 </p><p>Finally, the <tt>->dynticks</tt> field counts the corresponding
982 CPU's transitions to and from dyntick-idle mode, so that this counter
983 has an even value when the CPU is in dyntick-idle mode and an odd
986 <p><a name="Quick Quiz 6"><b>Quick Quiz 6</b>:</a>
987 Why not just count all NMIs?
988 Wouldn't that be simpler and less error prone?
989 <br><a href="#qq6answer">Answer</a>
991 </p><p>Additional fields are present for some special-purpose
992 builds, and are discussed separately.
994 <h3><a name="The rcu_head Structure">
995 The <tt>rcu_head</tt> Structure</a></h3>
997 <p>Each <tt>rcu_head</tt> structure represents an RCU callback.
998 These structures are normally embedded within RCU-protected data
999 structures whose algorithms use asynchronous grace periods.
1000 In contrast, when using algorithms that block waiting for RCU grace periods,
1001 RCU users need not provide <tt>rcu_head</tt> structures.
1003 </p><p>The <tt>rcu_head</tt> structure has fields as follows:
1006 1 struct rcu_head *next;
1007 2 void (*func)(struct rcu_head *head);
1010 <p>The <tt>->next</tt> field is used
1011 to link the <tt>rcu_head</tt> structures together in the
1012 lists within the <tt>rcu_data</tt> structures.
1013 The <tt>->func</tt> field is a pointer to the function
1014 to be called when the callback is ready to be invoked, and
1015 this function is passed a pointer to the <tt>rcu_head</tt>
1017 However, <tt>kfree_rcu()</tt> uses the <tt>->func</tt>
1018 field to record the offset of the <tt>rcu_head</tt>
1019 structure within the enclosing RCU-protected data structure.
1021 </p><p>Both of these fields are used internally by RCU.
1022 From the viewpoint of RCU users, this structure is an
1023 opaque “cookie”.
1025 <p><a name="Quick Quiz 7"><b>Quick Quiz 7</b>:</a>
1026 Given that the callback function <tt>->func</tt>
1027 is passed a pointer to the <tt>rcu_head</tt> structure,
1028 how is that function supposed to find the beginning of the
1029 enclosing RCU-protected data structure?
1030 <br><a href="#qq7answer">Answer</a>
1032 <h3><a name="RCU-Specific Fields in the task_struct Structure">
1033 RCU-Specific Fields in the <tt>task_struct</tt> Structure</a></h3>
1035 <p>The <tt>CONFIG_TREE_PREEMPT_RCU</tt> implementation uses some
1036 additional fields in the <tt>task_struct</tt> structure:
1039 1 #ifdef CONFIG_PREEMPT_RCU
1040 2 int rcu_read_lock_nesting;
1041 3 union rcu_special rcu_read_unlock_special;
1042 4 struct list_head rcu_node_entry;
1043 5 struct rcu_node *rcu_blocked_node;
1044 6 #endif /* #ifdef CONFIG_PREEMPT_RCU */
1045 7 #ifdef CONFIG_TASKS_RCU
1046 8 unsigned long rcu_tasks_nvcsw;
1047 9 bool rcu_tasks_holdout;
1048 10 struct list_head rcu_tasks_holdout_list;
1049 11 int rcu_tasks_idle_cpu;
1050 12 #endif /* #ifdef CONFIG_TASKS_RCU */
1053 <p>The <tt>->rcu_read_lock_nesting</tt> field records the
1054 nesting level for RCU read-side critical sections, and
1055 the <tt>->rcu_read_unlock_special</tt> field is a bitmask
1056 that records special conditions that require <tt>rcu_read_unlock()</tt>
1057 to do additional work.
1058 The <tt>->rcu_node_entry</tt> field is used to form lists of
1059 tasks that have blocked within preemptible-RCU read-side critical
1060 sections and the <tt>->rcu_blocked_node</tt> field references
1061 the <tt>rcu_node</tt> structure whose list this task is a member of,
1062 or <tt>NULL</tt> if it is not blocked within a preemptible-RCU
1063 read-side critical section.
1065 <p>The <tt>->rcu_tasks_nvcsw</tt> field tracks the number of
1066 voluntary context switches that this task had undergone at the
1067 beginning of the current tasks-RCU grace period,
1068 <tt>->rcu_tasks_holdout</tt> is set if the current tasks-RCU
1069 grace period is waiting on this task, <tt>->rcu_tasks_holdout_list</tt>
1070 is a list element enqueuing this task on the holdout list,
1071 and <tt>->rcu_tasks_idle_cpu</tt> tracks which CPU this
1072 idle task is running, but only if the task is currently running,
1073 that is, if the CPU is currently idle.
1075 <p><a name="Quick Quiz 8"><b>Quick Quiz 8</b>:</a>
1076 Why is <tt>->rcu_boosted</tt> required, given that there is
1077 a <tt>RCU_READ_UNLOCK_BOOSTED</tt> bit in
1078 <tt>->rcu_read_unlock_special</tt>?
1079 <br><a href="#qq8answer">Answer</a>
1081 <h3><a name="Accessor Functions">
1082 Accessor Functions</a></h3>
1084 <p>The following listing shows the
1085 <tt>rcu_get_root()</tt>, <tt>rcu_for_each_node_breadth_first</tt>,
1086 <tt>rcu_for_each_nonleaf_node_breadth_first()</tt>, and
1087 <tt>rcu_for_each_leaf_node()</tt> function and macros:
1090 1 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
1092 3 return &rsp->node[0];
1095 6 #define rcu_for_each_node_breadth_first(rsp, rnp) \
1096 7 for ((rnp) = &(rsp)->node[0]; \
1097 8 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
1099 10 #define rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) \
1100 11 for ((rnp) = &(rsp)->node[0]; \
1101 12 (rnp) < (rsp)->level[NUM_RCU_LVLS - 1]; (rnp)++)
1103 14 #define rcu_for_each_leaf_node(rsp, rnp) \
1104 15 for ((rnp) = (rsp)->level[NUM_RCU_LVLS - 1]; \
1105 16 (rnp) < &(rsp)->node[NUM_RCU_NODES]; (rnp)++)
1108 <p>The <tt>rcu_get_root()</tt> simply returns a pointer to the
1109 first element of the specified <tt>rcu_state</tt> structure's
1110 <tt>->node[]</tt> array, which is the root <tt>rcu_node</tt>
1113 </p><p>As noted earlier, the <tt>rcu_for_each_node_breadth_first()</tt>
1114 macro takes advantage of the layout of the <tt>rcu_node</tt>
1115 structures in the <tt>rcu_state</tt> structure's
1116 <tt>->node[]</tt> array, performing a breadth-first traversal by
1117 simply traversing the array in order.
1118 The <tt>rcu_for_each_nonleaf_node_breadth_first()</tt> macro operates
1119 similarly, but traverses only the first part of the array, thus excluding
1120 the leaf <tt>rcu_node</tt> structures.
1121 Finally, the <tt>rcu_for_each_leaf_node()</tt> macro traverses only
1122 the last part of the array, thus traversing only the leaf
1123 <tt>rcu_node</tt> structures.
1125 <p><a name="Quick Quiz 9"><b>Quick Quiz 9</b>:</a>
1126 What do <tt>rcu_for_each_nonleaf_node_breadth_first()</tt> and
1127 <tt>rcu_for_each_leaf_node()</tt> do if the <tt>rcu_node</tt> tree
1128 contains only a single node?
1129 <br><a href="#qq9answer">Answer</a>
1131 <h3><a name="Summary">
1134 So each flavor of RCU is represented by an <tt>rcu_state</tt> structure,
1135 which contains a combining tree of <tt>rcu_node</tt> and
1136 <tt>rcu_data</tt> structures.
1137 Finally, in <tt>CONFIG_NO_HZ_IDLE</tt> kernels, each CPU's dyntick-idle
1138 state is tracked by an <tt>rcu_dynticks</tt> structure.
1140 If you made it this far, you are well prepared to read the code
1141 walkthroughs in the other articles in this series.
1143 <h3><a name="Acknowledgments">
1144 Acknowledgments</a></h3>
1146 I owe thanks to Cyrill Gorcunov, Mathieu Desnoyers, Dhaval Giani, Paul
1147 Turner, Abhishek Srivastava, Matt Kowalczyk, and Serge Hallyn
1148 for helping me get this document into a more human-readable state.
1150 <h3><a name="Legal Statement">
1151 Legal Statement</a></h3>
1153 <p>This work represents the view of the author and does not necessarily
1154 represent the view of IBM.
1156 </p><p>Linux is a registered trademark of Linus Torvalds.
1158 </p><p>Other company, product, and service names may be trademarks or
1159 service marks of others.
1162 <h3><a name="Answers to Quick Quizzes">
1163 Answers to Quick Quizzes</a></h3>
1165 <a name="qq1answer"></a>
1166 <p><b>Quick Quiz 1</b>:
1167 Why isn't the fanout at the leaves also 64?
1170 </p><p><b>Answer</b>:
1171 Because there are more types of events that affect the leaf-level
1172 <tt>rcu_node</tt> structures than further up the tree.
1173 Therefore, if the leaf <tt>rcu_node</tt> structures have
1174 fanout of 64, the contention on these structures' <tt>->structures</tt>
1176 Experimentation on a wide variety of systems has shown that a fanout
1177 of 16 works well for the leaves of the <tt>rcu_node</tt> tree.
1179 </p><p>Of course, further experience with systems having hundreds or
1180 thousands of CPUs may demonstrate that the fanout for the non-leaf
1181 <tt>rcu_node</tt> structures must also be reduced.
1182 Such reduction can be easily carried out when and if it proves necessary.
1183 In the meantime, if you are using such a system and running into
1184 contention problems on the non-leaf <tt>rcu_node</tt> structures,
1185 you may use the <tt>CONFIG_RCU_FANOUT</tt> kernel configuration
1186 parameter to reduce the non-leaf fanout as needed.
1188 </p><p>Kernels built for systems with strong NUMA characteristics might
1189 also need to adjust <tt>CONFIG_RCU_FANOUT</tt> so that the
1190 domains of the <tt>rcu_node</tt> structures align with hardware
1192 However, there has thus far been no need for this.
1195 </p><p><a href="#Quick%20Quiz%201"><b>Back to Quick Quiz 1</b>.</a>
1197 <a name="qq2answer"></a>
1198 <p><b>Quick Quiz 2</b>:
1200 You said that the <tt>rcu_node</tt> structures formed a tree,
1201 but they are declared as a flat array!
1205 </p><p><b>Answer</b>:
1206 The tree is laid out in the array.
1207 The first node In the array is the head, the next set of nodes in the
1208 array are children of the head node, and so on until the last set of
1209 nodes in the array are the leaves.
1211 </p><p>See the following diagrams to see how this works.
1214 </p><p><a href="#Quick%20Quiz%202"><b>Back to Quick Quiz 2</b>.</a>
1216 <a name="qq3answer"></a>
1217 <p><b>Quick Quiz 3</b>:
1218 Given that this array represents a tree, why can't the diagram that
1219 includes the <tt>->level</tt> array be planar?
1222 </p><p><b>Answer</b>:
1223 It can be planar, it is just that it looks uglier that way.
1224 But don't take my word for it, draw it yourself!
1226 </p><p>But if you draw the tree to be tree-shaped rather than
1227 array-shaped, it is easy to draw a planar representation:
1229 </p><p><img src="TreeLevel.svg" alt="TreeLevel.svg" width="60%">
1232 </p><p><a href="#Quick%20Quiz%203"><b>Back to Quick Quiz 3</b>.</a>
1234 <a name="qq4answer"></a>
1235 <p><b>Quick Quiz 4</b>:
1236 Why are these bitmasks protected by locking?
1237 Come on, haven't you heard of atomic instructions???
1240 </p><p><b>Answer</b>:
1241 Lockless grace-period computation! Such a tantalizing possibility!
1243 </p><p>But consider the following sequence of events:
1246 <li> CPU 0 has been in dyntick-idle mode for quite
1248 When it wakes up, it notices that the current RCU
1249 grace period needs it to report in, so it sets a
1250 flag where the scheduling clock interrupt will find it.
1251 <li> Meanwhile, CPU 1 is running <tt>force_quiescent_state()</tt>,
1252 and notices that CPU 0 has been in dyntick idle mode,
1253 which qualifies as an extended quiescent state.
1254 <li> CPU 0's scheduling clock interrupt fires in the
1255 middle of an RCU read-side critical section, and notices
1256 that the RCU core needs something, so commences RCU softirq
1258 <li> CPU 0's softirq handler executes and is just about ready
1259 to report its quiescent state up the <tt>rcu_node</tt>
1261 <li> But CPU 1 beats it to the punch, completing the current
1262 grace period and starting a new one.
1263 <li> CPU 0 now reports its quiescent state for the wrong
1265 That grace period might now end before the RCU read-side
1267 If that happens, disaster will ensue.
1270 <p>So the locking is absolutely required in order to coordinate clearing
1271 of the bits with the grace-period numbers in <tt>->gpnum</tt>
1272 and <tt>->completed</tt>.
1275 </p><p><a href="#Quick%20Quiz%204"><b>Back to Quick Quiz 4</b>.</a>
1277 <a name="qq5answer"></a>
1278 <p><b>Quick Quiz 5</b>:
1279 All this replication of the grace period numbers can only cause
1281 Why not just keep a global pair of counters and be done with it???
1284 </p><p><b>Answer</b>:
1285 Because if there was only a single global pair of grace-period numbers,
1286 there would need to be a single global lock to allow safely accessing
1288 And if we are not going to have a single global lock, we need to carefully
1289 manage the numbers on a per-node basis.
1290 Recall from the answer to a previous Quick Quiz that the consequences
1291 of applying a previously sampled quiescent state to the wrong
1292 grace period are quite severe.
1295 </p><p><a href="#Quick%20Quiz%205"><b>Back to Quick Quiz 5</b>.</a>
1297 <a name="qq6answer"></a>
1298 <p><b>Quick Quiz 6</b>:
1299 Why not just count all NMIs?
1300 Wouldn't that be simpler and less error prone?
1303 </p><p><b>Answer</b>:
1304 It seems simpler only until you think hard about how to go about
1305 updating the <tt>rcu_dynticks</tt> structure's
1306 <tt>->dynticks</tt> field.
1309 </p><p><a href="#Quick%20Quiz%206"><b>Back to Quick Quiz 6</b>.</a>
1311 <a name="qq7answer"></a>
1312 <p><b>Quick Quiz 7</b>:
1313 Given that the callback function <tt>->func</tt>
1314 is passed a pointer to the <tt>rcu_head</tt> structure,
1315 how is that function supposed to find the beginning of the
1316 enclosing RCU-protected data structure?
1319 </p><p><b>Answer</b>:
1320 In actual practice, there is a separate callback function per
1321 type of RCU-protected data structure.
1322 The callback function can therefore use the <tt>container_of()</tt>
1323 macro in the Linux kernel (or other pointer-manipulation facilities
1324 in other software environments) to find the beginning of the
1325 enclosing structure.
1328 </p><p><a href="#Quick%20Quiz%207"><b>Back to Quick Quiz 7</b>.</a>
1330 <a name="qq8answer"></a>
1331 <p><b>Quick Quiz 8</b>:
1332 Why is <tt>->rcu_boosted</tt> required, given that there is
1333 a <tt>RCU_READ_UNLOCK_BOOSTED</tt> bit in
1334 <tt>->rcu_read_unlock_special</tt>?
1337 </p><p><b>Answer</b>:
1338 The <tt>->rcu_read_unlock_special</tt> field may only be
1339 updated by the task itself.
1340 By definition, RCU priority boosting must be carried out by some
1342 This other task cannot safely update the boosted task's
1343 <tt>->rcu_read_unlock_special</tt> field without the use of
1344 expensive atomic instructions.
1345 The <tt>->rcu_boosted</tt> field is therefore used by the
1346 boosting task to let the boosted task know that it has been boosted.
1347 The boosted task makes use of the
1348 <tt>RCU_READ_UNLOCK_BOOSTED</tt> bit in
1349 <tt>->rcu_read_unlock_special</tt>
1350 when deboosting itself.
1353 </p><p><a href="#Quick%20Quiz%208"><b>Back to Quick Quiz 8</b>.</a>
1355 <a name="qq9answer"></a>
1356 <p><b>Quick Quiz 9</b>:
1357 What do <tt>rcu_for_each_nonleaf_node_breadth_first()</tt> and
1358 <tt>rcu_for_each_leaf_node()</tt> do if the <tt>rcu_node</tt> tree
1359 contains only a single node?
1362 </p><p><b>Answer</b>:
1363 In the single-node case,
1364 <tt>rcu_for_each_nonleaf_node_breadth_first()</tt> is a no-op
1365 and <tt>rcu_for_each_leaf_node()</tt> traverses the single node.
1368 </p><p><a href="#Quick%20Quiz%209"><b>Back to Quick Quiz 9</b>.</a>