1 /* SPDX-License-Identifier: GPL-2.0+ */
2 /*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions that provide either classic
5 * or preemptible semantics.
6 *
7 * Copyright Red Hat, 2009
8 * Copyright IBM Corporation, 2009
9 *
10 * Author: Ingo Molnar <mingo@elte.hu>
11 * Paul E. McKenney <paulmck@linux.ibm.com>
12 */
13
14 #include "../locking/rtmutex_common.h"
15
16 static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
17 {
18 /*
19 * In order to read the offloaded state of an rdp in a safe
20 * and stable way and prevent from its value to be changed
21 * under us, we must either hold the barrier mutex, the cpu
22 * hotplug lock (read or write) or the nocb lock. Local
23 * non-preemptible reads are also safe. NOCB kthreads and
24 * timers have their own means of synchronization against the
25 * offloaded state updaters.
26 */
27 RCU_LOCKDEP_WARN(
28 !(lockdep_is_held(&rcu_state.barrier_mutex) ||
29 (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
30 rcu_lockdep_is_held_nocb(rdp) ||
31 (!(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible()) &&
32 rdp == this_cpu_ptr(&rcu_data)) ||
33 rcu_current_is_nocb_kthread(rdp)),
34 "Unsafe read of RCU_NOCB offloaded state"
35 );
36
37 return rcu_segcblist_is_offloaded(&rdp->cblist);
38 }
39
40 /*
41 * Check the RCU kernel configuration parameters and print informative
42 * messages about anything out of the ordinary.
43 */
44 static void __init rcu_bootup_announce_oddness(void)
45 {
46 if (IS_ENABLED(CONFIG_RCU_TRACE))
47 pr_info("\tRCU event tracing is enabled.\n");
48 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
49 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
50 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
51 RCU_FANOUT);
52 if (rcu_fanout_exact)
53 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
54 if (IS_ENABLED(CONFIG_PROVE_RCU))
55 pr_info("\tRCU lockdep checking is enabled.\n");
56 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
57 pr_info("\tRCU strict (and thus non-scalable) grace periods are enabled.\n");
58 if (RCU_NUM_LVLS >= 4)
59 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
60 if (RCU_FANOUT_LEAF != 16)
61 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
62 RCU_FANOUT_LEAF);
63 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
64 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
65 rcu_fanout_leaf);
66 if (nr_cpu_ids != NR_CPUS)
67 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
68 #ifdef CONFIG_RCU_BOOST
69 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
70 kthread_prio, CONFIG_RCU_BOOST_DELAY);
71 #endif
72 if (blimit != DEFAULT_RCU_BLIMIT)
73 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
74 if (qhimark != DEFAULT_RCU_QHIMARK)
75 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
76 if (qlowmark != DEFAULT_RCU_QLOMARK)
77 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
78 if (qovld != DEFAULT_RCU_QOVLD)
79 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
80 if (jiffies_till_first_fqs != ULONG_MAX)
81 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
82 if (jiffies_till_next_fqs != ULONG_MAX)
83 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
84 if (jiffies_till_sched_qs != ULONG_MAX)
85 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
86 if (rcu_kick_kthreads)
87 pr_info("\tKick kthreads if too-long grace period.\n");
88 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
89 pr_info("\tRCU callback double-/use-after-free debug is enabled.\n");
90 if (gp_preinit_delay)
91 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
92 if (gp_init_delay)
93 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
94 if (gp_cleanup_delay)
95 pr_info("\tRCU debug GP cleanup slowdown %d jiffies.\n", gp_cleanup_delay);
96 if (nohz_full_patience_delay < 0) {
97 pr_info("\tRCU NOCB CPU patience negative (%d), resetting to zero.\n", nohz_full_patience_delay);
98 nohz_full_patience_delay = 0;
99 } else if (nohz_full_patience_delay > 5 * MSEC_PER_SEC) {
100 pr_info("\tRCU NOCB CPU patience too large (%d), resetting to %ld.\n", nohz_full_patience_delay, 5 * MSEC_PER_SEC);
101 nohz_full_patience_delay = 5 * MSEC_PER_SEC;
102 } else if (nohz_full_patience_delay) {
103 pr_info("\tRCU NOCB CPU patience set to %d milliseconds.\n", nohz_full_patience_delay);
104 }
105 nohz_full_patience_delay_jiffies = msecs_to_jiffies(nohz_full_patience_delay);
106 if (!use_softirq)
107 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
108 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
109 pr_info("\tRCU debug extended QS entry/exit.\n");
110 rcupdate_announce_bootup_oddness();
111 }
112
113 #ifdef CONFIG_PREEMPT_RCU
114
115 static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
116 static void rcu_read_unlock_special(struct task_struct *t);
117
118 /*
119 * Tell them what RCU they are running.
120 */
121 static void __init rcu_bootup_announce(void)
122 {
123 pr_info("Preemptible hierarchical RCU implementation.\n");
124 rcu_bootup_announce_oddness();
125 }
126
127 /* Flags for rcu_preempt_ctxt_queue() decision table. */
128 #define RCU_GP_TASKS 0x8
129 #define RCU_EXP_TASKS 0x4
130 #define RCU_GP_BLKD 0x2
131 #define RCU_EXP_BLKD 0x1
132
133 /*
134 * Queues a task preempted within an RCU-preempt read-side critical
135 * section into the appropriate location within the ->blkd_tasks list,
136 * depending on the states of any ongoing normal and expedited grace
137 * periods. The ->gp_tasks pointer indicates which element the normal
138 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
139 * indicates which element the expedited grace period is waiting on (again,
140 * NULL if none). If a grace period is waiting on a given element in the
141 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
142 * adding a task to the tail of the list blocks any grace period that is
143 * already waiting on one of the elements. In contrast, adding a task
144 * to the head of the list won't block any grace period that is already
145 * waiting on one of the elements.
146 *
147 * This queuing is imprecise, and can sometimes make an ongoing grace
148 * period wait for a task that is not strictly speaking blocking it.
149 * Given the choice, we needlessly block a normal grace period rather than
150 * blocking an expedited grace period.
151 *
152 * Note that an endless sequence of expedited grace periods still cannot
153 * indefinitely postpone a normal grace period. Eventually, all of the
154 * fixed number of preempted tasks blocking the normal grace period that are
155 * not also blocking the expedited grace period will resume and complete
156 * their RCU read-side critical sections. At that point, the ->gp_tasks
157 * pointer will equal the ->exp_tasks pointer, at which point the end of
158 * the corresponding expedited grace period will also be the end of the
159 * normal grace period.
160 */
161 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
162 __releases(rnp->lock) /* But leaves rrupts disabled. */
163 {
164 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
165 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
166 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
167 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
168 struct task_struct *t = current;
169
170 raw_lockdep_assert_held_rcu_node(rnp);
171 WARN_ON_ONCE(rdp->mynode != rnp);
172 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
173 /* RCU better not be waiting on newly onlined CPUs! */
174 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
175 rdp->grpmask);
176
177 /*
178 * Decide where to queue the newly blocked task. In theory,
179 * this could be an if-statement. In practice, when I tried
180 * that, it was quite messy.
181 */
182 switch (blkd_state) {
183 case 0:
184 case RCU_EXP_TASKS:
185 case RCU_EXP_TASKS + RCU_GP_BLKD:
186 case RCU_GP_TASKS:
187 case RCU_GP_TASKS + RCU_EXP_TASKS:
188
189 /*
190 * Blocking neither GP, or first task blocking the normal
191 * GP but not blocking the already-waiting expedited GP.
192 * Queue at the head of the list to avoid unnecessarily
193 * blocking the already-waiting GPs.
194 */
195 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
196 break;
197
198 case RCU_EXP_BLKD:
199 case RCU_GP_BLKD:
200 case RCU_GP_BLKD + RCU_EXP_BLKD:
201 case RCU_GP_TASKS + RCU_EXP_BLKD:
202 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
203 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
204
205 /*
206 * First task arriving that blocks either GP, or first task
207 * arriving that blocks the expedited GP (with the normal
208 * GP already waiting), or a task arriving that blocks
209 * both GPs with both GPs already waiting. Queue at the
210 * tail of the list to avoid any GP waiting on any of the
211 * already queued tasks that are not blocking it.
212 */
213 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
214 break;
215
216 case RCU_EXP_TASKS + RCU_EXP_BLKD:
217 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
218 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
219
220 /*
221 * Second or subsequent task blocking the expedited GP.
222 * The task either does not block the normal GP, or is the
223 * first task blocking the normal GP. Queue just after
224 * the first task blocking the expedited GP.
225 */
226 list_add(&t->rcu_node_entry, rnp->exp_tasks);
227 break;
228
229 case RCU_GP_TASKS + RCU_GP_BLKD:
230 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
231
232 /*
233 * Second or subsequent task blocking the normal GP.
234 * The task does not block the expedited GP. Queue just
235 * after the first task blocking the normal GP.
236 */
237 list_add(&t->rcu_node_entry, rnp->gp_tasks);
238 break;
239
240 default:
241
242 /* Yet another exercise in excessive paranoia. */
243 WARN_ON_ONCE(1);
244 break;
245 }
246
247 /*
248 * We have now queued the task. If it was the first one to
249 * block either grace period, update the ->gp_tasks and/or
250 * ->exp_tasks pointers, respectively, to reference the newly
251 * blocked tasks.
252 */
253 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
254 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
255 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
256 }
257 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
258 WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
259 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
260 !(rnp->qsmask & rdp->grpmask));
261 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
262 !(rnp->expmask & rdp->grpmask));
263 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
264
265 /*
266 * Report the quiescent state for the expedited GP. This expedited
267 * GP should not be able to end until we report, so there should be
268 * no need to check for a subsequent expedited GP. (Though we are
269 * still in a quiescent state in any case.)
270 *
271 * Interrupts are disabled, so ->cpu_no_qs.b.exp cannot change.
272 */
273 if (blkd_state & RCU_EXP_BLKD && rdp->cpu_no_qs.b.exp)
274 rcu_report_exp_rdp(rdp);
275 else
276 WARN_ON_ONCE(rdp->cpu_no_qs.b.exp);
277 }
278
279 /*
280 * Record a preemptible-RCU quiescent state for the specified CPU.
281 * Note that this does not necessarily mean that the task currently running
282 * on the CPU is in a quiescent state: Instead, it means that the current
283 * grace period need not wait on any RCU read-side critical section that
284 * starts later on this CPU. It also means that if the current task is
285 * in an RCU read-side critical section, it has already added itself to
286 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
287 * current task, there might be any number of other tasks blocked while
288 * in an RCU read-side critical section.
289 *
290 * Unlike non-preemptible-RCU, quiescent state reports for expedited
291 * grace periods are handled separately via deferred quiescent states
292 * and context switch events.
293 *
294 * Callers to this function must disable preemption.
295 */
296 static void rcu_qs(void)
297 {
298 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
299 if (__this_cpu_read(rcu_data.cpu_no_qs.b.norm)) {
300 trace_rcu_grace_period(TPS("rcu_preempt"),
301 __this_cpu_read(rcu_data.gp_seq),
302 TPS("cpuqs"));
303 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
304 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
305 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
306 }
307 }
308
309 /*
310 * We have entered the scheduler, and the current task might soon be
311 * context-switched away from. If this task is in an RCU read-side
312 * critical section, we will no longer be able to rely on the CPU to
313 * record that fact, so we enqueue the task on the blkd_tasks list.
314 * The task will dequeue itself when it exits the outermost enclosing
315 * RCU read-side critical section. Therefore, the current grace period
316 * cannot be permitted to complete until the blkd_tasks list entries
317 * predating the current grace period drain, in other words, until
318 * rnp->gp_tasks becomes NULL.
319 *
320 * Caller must disable interrupts.
321 */
322 void rcu_note_context_switch(bool preempt)
323 {
324 struct task_struct *t = current;
325 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
326 struct rcu_node *rnp;
327
328 trace_rcu_utilization(TPS("Start context switch"));
329 lockdep_assert_irqs_disabled();
330 WARN_ONCE(!preempt && rcu_preempt_depth() > 0, "Voluntary context switch within RCU read-side critical section!");
331 if (rcu_preempt_depth() > 0 &&
332 !t->rcu_read_unlock_special.b.blocked) {
333
334 /* Possibly blocking in an RCU read-side critical section. */
335 rnp = rdp->mynode;
336 raw_spin_lock_rcu_node(rnp);
337 t->rcu_read_unlock_special.b.blocked = true;
338 t->rcu_blocked_node = rnp;
339
340 /*
341 * Verify the CPU's sanity, trace the preemption, and
342 * then queue the task as required based on the states
343 * of any ongoing and expedited grace periods.
344 */
345 WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp));
346 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
347 trace_rcu_preempt_task(rcu_state.name,
348 t->pid,
349 (rnp->qsmask & rdp->grpmask)
350 ? rnp->gp_seq
351 : rcu_seq_snap(&rnp->gp_seq));
352 rcu_preempt_ctxt_queue(rnp, rdp);
353 } else {
354 rcu_preempt_deferred_qs(t);
355 }
356
357 /*
358 * Either we were not in an RCU read-side critical section to
359 * begin with, or we have now recorded that critical section
360 * globally. Either way, we can now note a quiescent state
361 * for this CPU. Again, if we were in an RCU read-side critical
362 * section, and if that critical section was blocking the current
363 * grace period, then the fact that the task has been enqueued
364 * means that we continue to block the current grace period.
365 */
366 rcu_qs();
367 if (rdp->cpu_no_qs.b.exp)
368 rcu_report_exp_rdp(rdp);
369 rcu_tasks_qs(current, preempt);
370 trace_rcu_utilization(TPS("End context switch"));
371 }
372 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
373
374 /*
375 * Check for preempted RCU readers blocking the current grace period
376 * for the specified rcu_node structure. If the caller needs a reliable
377 * answer, it must hold the rcu_node's ->lock.
378 */
379 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
380 {
381 return READ_ONCE(rnp->gp_tasks) != NULL;
382 }
383
384 /* limit value for ->rcu_read_lock_nesting. */
385 #define RCU_NEST_PMAX (INT_MAX / 2)
386
387 static void rcu_preempt_read_enter(void)
388 {
389 WRITE_ONCE(current->rcu_read_lock_nesting, READ_ONCE(current->rcu_read_lock_nesting) + 1);
390 }
391
392 static int rcu_preempt_read_exit(void)
393 {
394 int ret = READ_ONCE(current->rcu_read_lock_nesting) - 1;
395
396 WRITE_ONCE(current->rcu_read_lock_nesting, ret);
397 return ret;
398 }
399
400 static void rcu_preempt_depth_set(int val)
401 {
402 WRITE_ONCE(current->rcu_read_lock_nesting, val);
403 }
404
405 /*
406 * Preemptible RCU implementation for rcu_read_lock().
407 * Just increment ->rcu_read_lock_nesting, shared state will be updated
408 * if we block.
409 */
410 void __rcu_read_lock(void)
411 {
412 rcu_preempt_read_enter();
413 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
414 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
415 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
416 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
417 barrier(); /* critical section after entry code. */
418 }
419 EXPORT_SYMBOL_GPL(__rcu_read_lock);
420
421 /*
422 * Preemptible RCU implementation for rcu_read_unlock().
423 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
424 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
425 * invoke rcu_read_unlock_special() to clean up after a context switch
426 * in an RCU read-side critical section and other special cases.
427 */
428 void __rcu_read_unlock(void)
429 {
430 struct task_struct *t = current;
431
432 barrier(); // critical section before exit code.
433 if (rcu_preempt_read_exit() == 0) {
434 barrier(); // critical-section exit before .s check.
435 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
436 rcu_read_unlock_special(t);
437 }
438 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
439 int rrln = rcu_preempt_depth();
440
441 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
442 }
443 }
444 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
445
446 /*
447 * Advance a ->blkd_tasks-list pointer to the next entry, instead
448 * returning NULL if at the end of the list.
449 */
450 static struct list_head *rcu_next_node_entry(struct task_struct *t,
451 struct rcu_node *rnp)
452 {
453 struct list_head *np;
454
455 np = t->rcu_node_entry.next;
456 if (np == &rnp->blkd_tasks)
457 np = NULL;
458 return np;
459 }
460
461 /*
462 * Return true if the specified rcu_node structure has tasks that were
463 * preempted within an RCU read-side critical section.
464 */
465 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
466 {
467 return !list_empty(&rnp->blkd_tasks);
468 }
469
470 /*
471 * Report deferred quiescent states. The deferral time can
472 * be quite short, for example, in the case of the call from
473 * rcu_read_unlock_special().
474 */
475 static notrace void
476 rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
477 {
478 bool empty_exp;
479 bool empty_norm;
480 bool empty_exp_now;
481 struct list_head *np;
482 bool drop_boost_mutex = false;
483 struct rcu_data *rdp;
484 struct rcu_node *rnp;
485 union rcu_special special;
486
487 /*
488 * If RCU core is waiting for this CPU to exit its critical section,
489 * report the fact that it has exited. Because irqs are disabled,
490 * t->rcu_read_unlock_special cannot change.
491 */
492 special = t->rcu_read_unlock_special;
493 rdp = this_cpu_ptr(&rcu_data);
494 if (!special.s && !rdp->cpu_no_qs.b.exp) {
495 local_irq_restore(flags);
496 return;
497 }
498 t->rcu_read_unlock_special.s = 0;
499 if (special.b.need_qs) {
500 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
501 rdp->cpu_no_qs.b.norm = false;
502 rcu_report_qs_rdp(rdp);
503 udelay(rcu_unlock_delay);
504 } else {
505 rcu_qs();
506 }
507 }
508
509 /*
510 * Respond to a request by an expedited grace period for a
511 * quiescent state from this CPU. Note that requests from
512 * tasks are handled when removing the task from the
513 * blocked-tasks list below.
514 */
515 if (rdp->cpu_no_qs.b.exp)
516 rcu_report_exp_rdp(rdp);
517
518 /* Clean up if blocked during RCU read-side critical section. */
519 if (special.b.blocked) {
520
521 /*
522 * Remove this task from the list it blocked on. The task
523 * now remains queued on the rcu_node corresponding to the
524 * CPU it first blocked on, so there is no longer any need
525 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
526 */
527 rnp = t->rcu_blocked_node;
528 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
529 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
530 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
531 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
532 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
533 (!empty_norm || rnp->qsmask));
534 empty_exp = sync_rcu_exp_done(rnp);
535 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
536 np = rcu_next_node_entry(t, rnp);
537 list_del_init(&t->rcu_node_entry);
538 t->rcu_blocked_node = NULL;
539 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
540 rnp->gp_seq, t->pid);
541 if (&t->rcu_node_entry == rnp->gp_tasks)
542 WRITE_ONCE(rnp->gp_tasks, np);
543 if (&t->rcu_node_entry == rnp->exp_tasks)
544 WRITE_ONCE(rnp->exp_tasks, np);
545 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
546 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
547 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx.rtmutex) == t;
548 if (&t->rcu_node_entry == rnp->boost_tasks)
549 WRITE_ONCE(rnp->boost_tasks, np);
550 }
551
552 /*
553 * If this was the last task on the current list, and if
554 * we aren't waiting on any CPUs, report the quiescent state.
555 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
556 * so we must take a snapshot of the expedited state.
557 */
558 empty_exp_now = sync_rcu_exp_done(rnp);
559 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
560 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
561 rnp->gp_seq,
562 0, rnp->qsmask,
563 rnp->level,
564 rnp->grplo,
565 rnp->grphi,
566 !!rnp->gp_tasks);
567 rcu_report_unblock_qs_rnp(rnp, flags);
568 } else {
569 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
570 }
571
572 /*
573 * If this was the last task on the expedited lists,
574 * then we need to report up the rcu_node hierarchy.
575 */
576 if (!empty_exp && empty_exp_now)
577 rcu_report_exp_rnp(rnp, true);
578
579 /* Unboost if we were boosted. */
580 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
581 rt_mutex_futex_unlock(&rnp->boost_mtx.rtmutex);
582 } else {
583 local_irq_restore(flags);
584 }
585 }
586
587 /*
588 * Is a deferred quiescent-state pending, and are we also not in
589 * an RCU read-side critical section? It is the caller's responsibility
590 * to ensure it is otherwise safe to report any deferred quiescent
591 * states. The reason for this is that it is safe to report a
592 * quiescent state during context switch even though preemption
593 * is disabled. This function cannot be expected to understand these
594 * nuances, so the caller must handle them.
595 */
596 static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
597 {
598 return (__this_cpu_read(rcu_data.cpu_no_qs.b.exp) ||
599 READ_ONCE(t->rcu_read_unlock_special.s)) &&
600 rcu_preempt_depth() == 0;
601 }
602
603 /*
604 * Report a deferred quiescent state if needed and safe to do so.
605 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
606 * not being in an RCU read-side critical section. The caller must
607 * evaluate safety in terms of interrupt, softirq, and preemption
608 * disabling.
609 */
610 notrace void rcu_preempt_deferred_qs(struct task_struct *t)
611 {
612 unsigned long flags;
613
614 if (!rcu_preempt_need_deferred_qs(t))
615 return;
616 local_irq_save(flags);
617 rcu_preempt_deferred_qs_irqrestore(t, flags);
618 }
619
620 /*
621 * Minimal handler to give the scheduler a chance to re-evaluate.
622 */
623 static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
624 {
625 struct rcu_data *rdp;
626
627 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
628 rdp->defer_qs_iw_pending = false;
629 }
630
631 /*
632 * Handle special cases during rcu_read_unlock(), such as needing to
633 * notify RCU core processing or task having blocked during the RCU
634 * read-side critical section.
635 */
636 static void rcu_read_unlock_special(struct task_struct *t)
637 {
638 unsigned long flags;
639 bool irqs_were_disabled;
640 bool preempt_bh_were_disabled =
641 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
642
643 /* NMI handlers cannot block and cannot safely manipulate state. */
644 if (in_nmi())
645 return;
646
647 local_irq_save(flags);
648 irqs_were_disabled = irqs_disabled_flags(flags);
649 if (preempt_bh_were_disabled || irqs_were_disabled) {
650 bool expboost; // Expedited GP in flight or possible boosting.
651 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
652 struct rcu_node *rnp = rdp->mynode;
653
654 expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
655 (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
656 (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
657 ((rdp->grpmask & READ_ONCE(rnp->qsmask)) || t->rcu_blocked_node)) ||
658 (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
659 t->rcu_blocked_node);
660 // Need to defer quiescent state until everything is enabled.
661 if (use_softirq && (in_hardirq() || (expboost && !irqs_were_disabled))) {
662 // Using softirq, safe to awaken, and either the
663 // wakeup is free or there is either an expedited
664 // GP in flight or a potential need to deboost.
665 raise_softirq_irqoff(RCU_SOFTIRQ);
666 } else {
667 // Enabling BH or preempt does reschedule, so...
668 // Also if no expediting and no possible deboosting,
669 // slow is OK. Plus nohz_full CPUs eventually get
670 // tick enabled.
671 set_tsk_need_resched(current);
672 set_preempt_need_resched();
673 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
674 expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
675 // Get scheduler to re-evaluate and call hooks.
676 // If !IRQ_WORK, FQS scan will eventually IPI.
677 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
678 IS_ENABLED(CONFIG_PREEMPT_RT))
679 rdp->defer_qs_iw = IRQ_WORK_INIT_HARD(
680 rcu_preempt_deferred_qs_handler);
681 else
682 init_irq_work(&rdp->defer_qs_iw,
683 rcu_preempt_deferred_qs_handler);
684 rdp->defer_qs_iw_pending = true;
685 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
686 }
687 }
688 local_irq_restore(flags);
689 return;
690 }
691 rcu_preempt_deferred_qs_irqrestore(t, flags);
692 }
693
694 /*
695 * Check that the list of blocked tasks for the newly completed grace
696 * period is in fact empty. It is a serious bug to complete a grace
697 * period that still has RCU readers blocked! This function must be
698 * invoked -before- updating this rnp's ->gp_seq.
699 *
700 * Also, if there are blocked tasks on the list, they automatically
701 * block the newly created grace period, so set up ->gp_tasks accordingly.
702 */
703 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
704 {
705 struct task_struct *t;
706
707 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
708 raw_lockdep_assert_held_rcu_node(rnp);
709 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
710 dump_blkd_tasks(rnp, 10);
711 if (rcu_preempt_has_tasks(rnp) &&
712 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
713 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
714 t = container_of(rnp->gp_tasks, struct task_struct,
715 rcu_node_entry);
716 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
717 rnp->gp_seq, t->pid);
718 }
719 WARN_ON_ONCE(rnp->qsmask);
720 }
721
722 /*
723 * Check for a quiescent state from the current CPU, including voluntary
724 * context switches for Tasks RCU. When a task blocks, the task is
725 * recorded in the corresponding CPU's rcu_node structure, which is checked
726 * elsewhere, hence this function need only check for quiescent states
727 * related to the current CPU, not to those related to tasks.
728 */
729 static void rcu_flavor_sched_clock_irq(int user)
730 {
731 struct task_struct *t = current;
732
733 lockdep_assert_irqs_disabled();
734 if (rcu_preempt_depth() > 0 ||
735 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
736 /* No QS, force context switch if deferred. */
737 if (rcu_preempt_need_deferred_qs(t)) {
738 set_tsk_need_resched(t);
739 set_preempt_need_resched();
740 }
741 } else if (rcu_preempt_need_deferred_qs(t)) {
742 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
743 return;
744 } else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
745 rcu_qs(); /* Report immediate QS. */
746 return;
747 }
748
749 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
750 if (rcu_preempt_depth() > 0 &&
751 __this_cpu_read(rcu_data.core_needs_qs) &&
752 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
753 !t->rcu_read_unlock_special.b.need_qs &&
754 time_after(jiffies, rcu_state.gp_start + HZ))
755 t->rcu_read_unlock_special.b.need_qs = true;
756 }
757
758 /*
759 * Check for a task exiting while in a preemptible-RCU read-side
760 * critical section, clean up if so. No need to issue warnings, as
761 * debug_check_no_locks_held() already does this if lockdep is enabled.
762 * Besides, if this function does anything other than just immediately
763 * return, there was a bug of some sort. Spewing warnings from this
764 * function is like as not to simply obscure important prior warnings.
765 */
766 void exit_rcu(void)
767 {
768 struct task_struct *t = current;
769
770 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
771 rcu_preempt_depth_set(1);
772 barrier();
773 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
774 } else if (unlikely(rcu_preempt_depth())) {
775 rcu_preempt_depth_set(1);
776 } else {
777 return;
778 }
779 __rcu_read_unlock();
780 rcu_preempt_deferred_qs(current);
781 }
782
783 /*
784 * Dump the blocked-tasks state, but limit the list dump to the
785 * specified number of elements.
786 */
787 static void
788 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
789 {
790 int cpu;
791 int i;
792 struct list_head *lhp;
793 struct rcu_data *rdp;
794 struct rcu_node *rnp1;
795
796 raw_lockdep_assert_held_rcu_node(rnp);
797 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
798 __func__, rnp->grplo, rnp->grphi, rnp->level,
799 (long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
800 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
801 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
802 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
803 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
804 __func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
805 READ_ONCE(rnp->exp_tasks));
806 pr_info("%s: ->blkd_tasks", __func__);
807 i = 0;
808 list_for_each(lhp, &rnp->blkd_tasks) {
809 pr_cont(" %p", lhp);
810 if (++i >= ncheck)
811 break;
812 }
813 pr_cont("\n");
814 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
815 rdp = per_cpu_ptr(&rcu_data, cpu);
816 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
817 cpu, ".o"[rcu_rdp_cpu_online(rdp)],
818 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_state,
819 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_state);
820 }
821 }
822
823 #else /* #ifdef CONFIG_PREEMPT_RCU */
824
825 /*
826 * If strict grace periods are enabled, and if the calling
827 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
828 * report that quiescent state and, if requested, spin for a bit.
829 */
830 void rcu_read_unlock_strict(void)
831 {
832 struct rcu_data *rdp;
833
834 if (irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
835 return;
836 rdp = this_cpu_ptr(&rcu_data);
837 rdp->cpu_no_qs.b.norm = false;
838 rcu_report_qs_rdp(rdp);
839 udelay(rcu_unlock_delay);
840 }
841 EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
842
843 /*
844 * Tell them what RCU they are running.
845 */
846 static void __init rcu_bootup_announce(void)
847 {
848 pr_info("Hierarchical RCU implementation.\n");
849 rcu_bootup_announce_oddness();
850 }
851
852 /*
853 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
854 * how many quiescent states passed, just if there was at least one since
855 * the start of the grace period, this just sets a flag. The caller must
856 * have disabled preemption.
857 */
858 static void rcu_qs(void)
859 {
860 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
861 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
862 return;
863 trace_rcu_grace_period(TPS("rcu_sched"),
864 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
865 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
866 if (__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
867 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
868 }
869
870 /*
871 * Register an urgently needed quiescent state. If there is an
872 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
873 * dyntick-idle quiescent state visible to other CPUs, which will in
874 * some cases serve for expedited as well as normal grace periods.
875 * Either way, register a lightweight quiescent state.
876 */
877 void rcu_all_qs(void)
878 {
879 unsigned long flags;
880
881 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
882 return;
883 preempt_disable(); // For CONFIG_PREEMPT_COUNT=y kernels
884 /* Load rcu_urgent_qs before other flags. */
885 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
886 preempt_enable();
887 return;
888 }
889 this_cpu_write(rcu_data.rcu_urgent_qs, false);
890 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
891 local_irq_save(flags);
892 rcu_momentary_dyntick_idle();
893 local_irq_restore(flags);
894 }
895 rcu_qs();
896 preempt_enable();
897 }
898 EXPORT_SYMBOL_GPL(rcu_all_qs);
899
900 /*
901 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
902 */
903 void rcu_note_context_switch(bool preempt)
904 {
905 trace_rcu_utilization(TPS("Start context switch"));
906 rcu_qs();
907 /* Load rcu_urgent_qs before other flags. */
908 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
909 goto out;
910 this_cpu_write(rcu_data.rcu_urgent_qs, false);
911 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
912 rcu_momentary_dyntick_idle();
913 out:
914 rcu_tasks_qs(current, preempt);
915 trace_rcu_utilization(TPS("End context switch"));
916 }
917 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
918
919 /*
920 * Because preemptible RCU does not exist, there are never any preempted
921 * RCU readers.
922 */
923 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
924 {
925 return 0;
926 }
927
928 /*
929 * Because there is no preemptible RCU, there can be no readers blocked.
930 */
931 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
932 {
933 return false;
934 }
935
936 /*
937 * Because there is no preemptible RCU, there can be no deferred quiescent
938 * states.
939 */
940 static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
941 {
942 return false;
943 }
944
945 // Except that we do need to respond to a request by an expedited
946 // grace period for a quiescent state from this CPU. Note that in
947 // non-preemptible kernels, there can be no context switches within RCU
948 // read-side critical sections, which in turn means that the leaf rcu_node
949 // structure's blocked-tasks list is always empty. is therefore no need to
950 // actually check it. Instead, a quiescent state from this CPU suffices,
951 // and this function is only called from such a quiescent state.
952 notrace void rcu_preempt_deferred_qs(struct task_struct *t)
953 {
954 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
955
956 if (READ_ONCE(rdp->cpu_no_qs.b.exp))
957 rcu_report_exp_rdp(rdp);
958 }
959
960 /*
961 * Because there is no preemptible RCU, there can be no readers blocked,
962 * so there is no need to check for blocked tasks. So check only for
963 * bogus qsmask values.
964 */
965 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
966 {
967 WARN_ON_ONCE(rnp->qsmask);
968 }
969
970 /*
971 * Check to see if this CPU is in a non-context-switch quiescent state,
972 * namely user mode and idle loop.
973 */
974 static void rcu_flavor_sched_clock_irq(int user)
975 {
976 if (user || rcu_is_cpu_rrupt_from_idle()) {
977
978 /*
979 * Get here if this CPU took its interrupt from user
980 * mode or from the idle loop, and if this is not a
981 * nested interrupt. In this case, the CPU is in
982 * a quiescent state, so note it.
983 *
984 * No memory barrier is required here because rcu_qs()
985 * references only CPU-local variables that other CPUs
986 * neither access nor modify, at least not while the
987 * corresponding CPU is online.
988 */
989 rcu_qs();
990 }
991 }
992
993 /*
994 * Because preemptible RCU does not exist, tasks cannot possibly exit
995 * while in preemptible RCU read-side critical sections.
996 */
997 void exit_rcu(void)
998 {
999 }
1000
1001 /*
1002 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
1003 */
1004 static void
1005 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
1006 {
1007 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
1008 }
1009
1010 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
1011
1012 /*
1013 * If boosting, set rcuc kthreads to realtime priority.
1014 */
1015 static void rcu_cpu_kthread_setup(unsigned int cpu)
1016 {
1017 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1018 #ifdef CONFIG_RCU_BOOST
1019 struct sched_param sp;
1020
1021 sp.sched_priority = kthread_prio;
1022 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1023 #endif /* #ifdef CONFIG_RCU_BOOST */
1024
1025 WRITE_ONCE(rdp->rcuc_activity, jiffies);
1026 }
1027
1028 static bool rcu_is_callbacks_nocb_kthread(struct rcu_data *rdp)
1029 {
1030 #ifdef CONFIG_RCU_NOCB_CPU
1031 return rdp->nocb_cb_kthread == current;
1032 #else
1033 return false;
1034 #endif
1035 }
1036
1037 /*
1038 * Is the current CPU running the RCU-callbacks kthread?
1039 * Caller must have preemption disabled.
1040 */
1041 static bool rcu_is_callbacks_kthread(struct rcu_data *rdp)
1042 {
1043 return rdp->rcu_cpu_kthread_task == current ||
1044 rcu_is_callbacks_nocb_kthread(rdp);
1045 }
1046
1047 #ifdef CONFIG_RCU_BOOST
1048
1049 /*
1050 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1051 * or ->boost_tasks, advancing the pointer to the next task in the
1052 * ->blkd_tasks list.
1053 *
1054 * Note that irqs must be enabled: boosting the task can block.
1055 * Returns 1 if there are more tasks needing to be boosted.
1056 */
1057 static int rcu_boost(struct rcu_node *rnp)
1058 {
1059 unsigned long flags;
1060 struct task_struct *t;
1061 struct list_head *tb;
1062
1063 if (READ_ONCE(rnp->exp_tasks) == NULL &&
1064 READ_ONCE(rnp->boost_tasks) == NULL)
1065 return 0; /* Nothing left to boost. */
1066
1067 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1068
1069 /*
1070 * Recheck under the lock: all tasks in need of boosting
1071 * might exit their RCU read-side critical sections on their own.
1072 */
1073 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1074 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1075 return 0;
1076 }
1077
1078 /*
1079 * Preferentially boost tasks blocking expedited grace periods.
1080 * This cannot starve the normal grace periods because a second
1081 * expedited grace period must boost all blocked tasks, including
1082 * those blocking the pre-existing normal grace period.
1083 */
1084 if (rnp->exp_tasks != NULL)
1085 tb = rnp->exp_tasks;
1086 else
1087 tb = rnp->boost_tasks;
1088
1089 /*
1090 * We boost task t by manufacturing an rt_mutex that appears to
1091 * be held by task t. We leave a pointer to that rt_mutex where
1092 * task t can find it, and task t will release the mutex when it
1093 * exits its outermost RCU read-side critical section. Then
1094 * simply acquiring this artificial rt_mutex will boost task
1095 * t's priority. (Thanks to tglx for suggesting this approach!)
1096 *
1097 * Note that task t must acquire rnp->lock to remove itself from
1098 * the ->blkd_tasks list, which it will do from exit() if from
1099 * nowhere else. We therefore are guaranteed that task t will
1100 * stay around at least until we drop rnp->lock. Note that
1101 * rnp->lock also resolves races between our priority boosting
1102 * and task t's exiting its outermost RCU read-side critical
1103 * section.
1104 */
1105 t = container_of(tb, struct task_struct, rcu_node_entry);
1106 rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
1107 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1108 /* Lock only for side effect: boosts task t's priority. */
1109 rt_mutex_lock(&rnp->boost_mtx);
1110 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1111 rnp->n_boosts++;
1112
1113 return READ_ONCE(rnp->exp_tasks) != NULL ||
1114 READ_ONCE(rnp->boost_tasks) != NULL;
1115 }
1116
1117 /*
1118 * Priority-boosting kthread, one per leaf rcu_node.
1119 */
1120 static int rcu_boost_kthread(void *arg)
1121 {
1122 struct rcu_node *rnp = (struct rcu_node *)arg;
1123 int spincnt = 0;
1124 int more2boost;
1125
1126 trace_rcu_utilization(TPS("Start boost kthread@init"));
1127 for (;;) {
1128 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1129 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1130 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1131 READ_ONCE(rnp->exp_tasks));
1132 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1133 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1134 more2boost = rcu_boost(rnp);
1135 if (more2boost)
1136 spincnt++;
1137 else
1138 spincnt = 0;
1139 if (spincnt > 10) {
1140 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1141 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1142 schedule_timeout_idle(2);
1143 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1144 spincnt = 0;
1145 }
1146 }
1147 /* NOTREACHED */
1148 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1149 return 0;
1150 }
1151
1152 /*
1153 * Check to see if it is time to start boosting RCU readers that are
1154 * blocking the current grace period, and, if so, tell the per-rcu_node
1155 * kthread to start boosting them. If there is an expedited grace
1156 * period in progress, it is always time to boost.
1157 *
1158 * The caller must hold rnp->lock, which this function releases.
1159 * The ->boost_kthread_task is immortal, so we don't need to worry
1160 * about it going away.
1161 */
1162 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1163 __releases(rnp->lock)
1164 {
1165 raw_lockdep_assert_held_rcu_node(rnp);
1166 if (!rnp->boost_kthread_task ||
1167 (!rcu_preempt_blocked_readers_cgp(rnp) && !rnp->exp_tasks)) {
1168 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1169 return;
1170 }
1171 if (rnp->exp_tasks != NULL ||
1172 (rnp->gp_tasks != NULL &&
1173 rnp->boost_tasks == NULL &&
1174 rnp->qsmask == 0 &&
1175 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld ||
1176 IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)))) {
1177 if (rnp->exp_tasks == NULL)
1178 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1179 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1180 rcu_wake_cond(rnp->boost_kthread_task,
1181 READ_ONCE(rnp->boost_kthread_status));
1182 } else {
1183 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1184 }
1185 }
1186
1187 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1188
1189 /*
1190 * Do priority-boost accounting for the start of a new grace period.
1191 */
1192 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1193 {
1194 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1195 }
1196
1197 /*
1198 * Create an RCU-boost kthread for the specified node if one does not
1199 * already exist. We only create this kthread for preemptible RCU.
1200 */
1201 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1202 {
1203 unsigned long flags;
1204 int rnp_index = rnp - rcu_get_root();
1205 struct sched_param sp;
1206 struct task_struct *t;
1207
1208 if (rnp->boost_kthread_task)
1209 return;
1210
1211 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1212 "rcub/%d", rnp_index);
1213 if (WARN_ON_ONCE(IS_ERR(t)))
1214 return;
1215
1216 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1217 rnp->boost_kthread_task = t;
1218 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1219 sp.sched_priority = kthread_prio;
1220 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1221 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1222 }
1223
1224 static struct task_struct *rcu_boost_task(struct rcu_node *rnp)
1225 {
1226 return READ_ONCE(rnp->boost_kthread_task);
1227 }
1228
1229 #else /* #ifdef CONFIG_RCU_BOOST */
1230
1231 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1232 __releases(rnp->lock)
1233 {
1234 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1235 }
1236
1237 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1238 {
1239 }
1240
1241 static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1242 {
1243 }
1244
1245 static struct task_struct *rcu_boost_task(struct rcu_node *rnp)
1246 {
1247 return NULL;
1248 }
1249 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1250
1251 /*
1252 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
1253 * grace-period kthread will do force_quiescent_state() processing?
1254 * The idea is to avoid waking up RCU core processing on such a
1255 * CPU unless the grace period has extended for too long.
1256 *
1257 * This code relies on the fact that all NO_HZ_FULL CPUs are also
1258 * RCU_NOCB_CPU CPUs.
1259 */
1260 static bool rcu_nohz_full_cpu(void)
1261 {
1262 #ifdef CONFIG_NO_HZ_FULL
1263 if (tick_nohz_full_cpu(smp_processor_id()) &&
1264 (!rcu_gp_in_progress() ||
1265 time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
1266 return true;
1267 #endif /* #ifdef CONFIG_NO_HZ_FULL */
1268 return false;
1269 }
1270
1271 /*
1272 * Bind the RCU grace-period kthreads to the housekeeping CPU.
1273 */
1274 static void rcu_bind_gp_kthread(void)
1275 {
1276 if (!tick_nohz_full_enabled())
1277 return;
1278 housekeeping_affine(current, HK_TYPE_RCU);
1279 }
1280
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.