1 /* SPDX-License-Identifier: GPL-2.0+ */
3 * Task-based RCU implementations.
5 * Copyright (C) 2020 Paul E. McKenney
8 #ifdef CONFIG_TASKS_RCU_GENERIC
9 #include "rcu_segcblist.h"
11 ////////////////////////////////////////////////////////////////////////
13 // Generic data structures.
16 typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
17 typedef void (*pregp_func_t)(struct list_head *hop);
18 typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
19 typedef void (*postscan_func_t)(struct list_head *hop);
20 typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
21 typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
24 * struct rcu_tasks_percpu - Per-CPU component of definition for a Tasks-RCU-like mechanism.
25 * @cblist: Callback list.
26 * @lock: Lock protecting per-CPU callback list.
27 * @rtp_jiffies: Jiffies counter value for statistics.
28 * @rtp_n_lock_retries: Rough lock-contention statistic.
29 * @rtp_work: Work queue for invoking callbacks.
30 * @rtp_irq_work: IRQ work queue for deferred wakeups.
31 * @barrier_q_head: RCU callback for barrier operation.
32 * @rtp_blkd_tasks: List of tasks blocked as readers.
33 * @cpu: CPU number corresponding to this entry.
34 * @rtpp: Pointer to the rcu_tasks structure.
36 struct rcu_tasks_percpu {
37 struct rcu_segcblist cblist;
38 raw_spinlock_t __private lock;
39 unsigned long rtp_jiffies;
40 unsigned long rtp_n_lock_retries;
41 struct work_struct rtp_work;
42 struct irq_work rtp_irq_work;
43 struct rcu_head barrier_q_head;
44 struct list_head rtp_blkd_tasks;
46 struct rcu_tasks *rtpp;
50 * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
51 * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
52 * @cbs_gbl_lock: Lock protecting callback list.
53 * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
54 * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
55 * @gp_func: This flavor's grace-period-wait function.
56 * @gp_state: Grace period's most recent state transition (debugging).
57 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
58 * @init_fract: Initial backoff sleep interval.
59 * @gp_jiffies: Time of last @gp_state transition.
60 * @gp_start: Most recent grace-period start in jiffies.
61 * @tasks_gp_seq: Number of grace periods completed since boot.
62 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
63 * @n_ipis_fails: Number of IPI-send failures.
64 * @pregp_func: This flavor's pre-grace-period function (optional).
65 * @pertask_func: This flavor's per-task scan function (optional).
66 * @postscan_func: This flavor's post-task scan function (optional).
67 * @holdouts_func: This flavor's holdout-list scan function (optional).
68 * @postgp_func: This flavor's post-grace-period function (optional).
69 * @call_func: This flavor's call_rcu()-equivalent function.
70 * @rtpcpu: This flavor's rcu_tasks_percpu structure.
71 * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
72 * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
73 * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
74 * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
75 * @barrier_q_mutex: Serialize barrier operations.
76 * @barrier_q_count: Number of queues being waited on.
77 * @barrier_q_completion: Barrier wait/wakeup mechanism.
78 * @barrier_q_seq: Sequence number for barrier operations.
79 * @name: This flavor's textual name.
80 * @kname: This flavor's kthread name.
83 struct rcuwait cbs_wait;
84 raw_spinlock_t cbs_gbl_lock;
85 struct mutex tasks_gp_mutex;
89 unsigned long gp_jiffies;
90 unsigned long gp_start;
91 unsigned long tasks_gp_seq;
93 unsigned long n_ipis_fails;
94 struct task_struct *kthread_ptr;
95 rcu_tasks_gp_func_t gp_func;
96 pregp_func_t pregp_func;
97 pertask_func_t pertask_func;
98 postscan_func_t postscan_func;
99 holdouts_func_t holdouts_func;
100 postgp_func_t postgp_func;
101 call_rcu_func_t call_func;
102 struct rcu_tasks_percpu __percpu *rtpcpu;
103 int percpu_enqueue_shift;
104 int percpu_enqueue_lim;
105 int percpu_dequeue_lim;
106 unsigned long percpu_dequeue_gpseq;
107 struct mutex barrier_q_mutex;
108 atomic_t barrier_q_count;
109 struct completion barrier_q_completion;
110 unsigned long barrier_q_seq;
115 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
117 #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \
118 static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = { \
119 .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock), \
120 .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup), \
122 static struct rcu_tasks rt_name = \
124 .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait), \
125 .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \
126 .tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex), \
129 .rtpcpu = &rt_name ## __percpu, \
131 .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS), \
132 .percpu_enqueue_lim = 1, \
133 .percpu_dequeue_lim = 1, \
134 .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex), \
135 .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT, \
139 /* Track exiting tasks in order to allow them to be waited for. */
140 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
142 /* Avoid IPIing CPUs early in the grace period. */
143 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
144 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
145 module_param(rcu_task_ipi_delay, int, 0644);
147 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
148 #define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
149 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
150 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
151 module_param(rcu_task_stall_timeout, int, 0644);
152 #define RCU_TASK_STALL_INFO (HZ * 10)
153 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
154 module_param(rcu_task_stall_info, int, 0644);
155 static int rcu_task_stall_info_mult __read_mostly = 3;
156 module_param(rcu_task_stall_info_mult, int, 0444);
158 static int rcu_task_enqueue_lim __read_mostly = -1;
159 module_param(rcu_task_enqueue_lim, int, 0444);
161 static bool rcu_task_cb_adjust;
162 static int rcu_task_contend_lim __read_mostly = 100;
163 module_param(rcu_task_contend_lim, int, 0444);
164 static int rcu_task_collapse_lim __read_mostly = 10;
165 module_param(rcu_task_collapse_lim, int, 0444);
167 /* RCU tasks grace-period state for debugging. */
169 #define RTGS_WAIT_WAIT_CBS 1
170 #define RTGS_WAIT_GP 2
171 #define RTGS_PRE_WAIT_GP 3
172 #define RTGS_SCAN_TASKLIST 4
173 #define RTGS_POST_SCAN_TASKLIST 5
174 #define RTGS_WAIT_SCAN_HOLDOUTS 6
175 #define RTGS_SCAN_HOLDOUTS 7
176 #define RTGS_POST_GP 8
177 #define RTGS_WAIT_READERS 9
178 #define RTGS_INVOKE_CBS 10
179 #define RTGS_WAIT_CBS 11
180 #ifndef CONFIG_TINY_RCU
181 static const char * const rcu_tasks_gp_state_names[] = {
183 "RTGS_WAIT_WAIT_CBS",
186 "RTGS_SCAN_TASKLIST",
187 "RTGS_POST_SCAN_TASKLIST",
188 "RTGS_WAIT_SCAN_HOLDOUTS",
189 "RTGS_SCAN_HOLDOUTS",
195 #endif /* #ifndef CONFIG_TINY_RCU */
197 ////////////////////////////////////////////////////////////////////////
201 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
203 /* Record grace-period phase and time. */
204 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
206 rtp->gp_state = newstate;
207 rtp->gp_jiffies = jiffies;
210 #ifndef CONFIG_TINY_RCU
211 /* Return state name. */
212 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
214 int i = data_race(rtp->gp_state); // Let KCSAN detect update races
215 int j = READ_ONCE(i); // Prevent the compiler from reading twice
217 if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
219 return rcu_tasks_gp_state_names[j];
221 #endif /* #ifndef CONFIG_TINY_RCU */
223 // Initialize per-CPU callback lists for the specified flavor of
225 static void cblist_init_generic(struct rcu_tasks *rtp)
232 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
233 if (rcu_task_enqueue_lim < 0) {
234 rcu_task_enqueue_lim = 1;
235 rcu_task_cb_adjust = true;
236 pr_info("%s: Setting adjustable number of callback queues.\n", __func__);
237 } else if (rcu_task_enqueue_lim == 0) {
238 rcu_task_enqueue_lim = 1;
240 lim = rcu_task_enqueue_lim;
242 if (lim > nr_cpu_ids)
244 shift = ilog2(nr_cpu_ids / lim);
245 if (((nr_cpu_ids - 1) >> shift) >= lim)
247 WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
248 WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
249 smp_store_release(&rtp->percpu_enqueue_lim, lim);
250 for_each_possible_cpu(cpu) {
251 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
253 WARN_ON_ONCE(!rtpcp);
255 raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
256 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
257 if (rcu_segcblist_empty(&rtpcp->cblist))
258 rcu_segcblist_init(&rtpcp->cblist);
259 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
262 if (!rtpcp->rtp_blkd_tasks.next)
263 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
264 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
266 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
267 pr_info("%s: Setting shift to %d and lim to %d.\n", __func__, data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim));
270 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
271 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
273 struct rcu_tasks *rtp;
274 struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
277 rcuwait_wake_up(&rtp->cbs_wait);
280 // Enqueue a callback for the specified flavor of Tasks RCU.
281 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
282 struct rcu_tasks *rtp)
288 bool needadjust = false;
290 struct rcu_tasks_percpu *rtpcp;
294 local_irq_save(flags);
296 ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
297 chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
298 rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
299 if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
300 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
302 if (rtpcp->rtp_jiffies != j) {
303 rtpcp->rtp_jiffies = j;
304 rtpcp->rtp_n_lock_retries = 0;
306 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
307 READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
308 needadjust = true; // Defer adjustment to avoid deadlock.
310 if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
311 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
312 cblist_init_generic(rtp);
313 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
315 needwake = rcu_segcblist_empty(&rtpcp->cblist);
316 rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
317 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
318 if (unlikely(needadjust)) {
319 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
320 if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
321 WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
322 WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
323 smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
324 pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
326 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
329 /* We can't create the thread unless interrupts are enabled. */
330 if (needwake && READ_ONCE(rtp->kthread_ptr))
331 irq_work_queue(&rtpcp->rtp_irq_work);
334 // RCU callback function for rcu_barrier_tasks_generic().
335 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
337 struct rcu_tasks *rtp;
338 struct rcu_tasks_percpu *rtpcp;
340 rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
342 if (atomic_dec_and_test(&rtp->barrier_q_count))
343 complete(&rtp->barrier_q_completion);
346 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
347 // Operates in a manner similar to rcu_barrier().
348 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
352 struct rcu_tasks_percpu *rtpcp;
353 unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
355 mutex_lock(&rtp->barrier_q_mutex);
356 if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
358 mutex_unlock(&rtp->barrier_q_mutex);
361 rcu_seq_start(&rtp->barrier_q_seq);
362 init_completion(&rtp->barrier_q_completion);
363 atomic_set(&rtp->barrier_q_count, 2);
364 for_each_possible_cpu(cpu) {
365 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
367 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
368 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
369 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
370 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
371 atomic_inc(&rtp->barrier_q_count);
372 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
374 if (atomic_sub_and_test(2, &rtp->barrier_q_count))
375 complete(&rtp->barrier_q_completion);
376 wait_for_completion(&rtp->barrier_q_completion);
377 rcu_seq_end(&rtp->barrier_q_seq);
378 mutex_unlock(&rtp->barrier_q_mutex);
381 // Advance callbacks and indicate whether either a grace period or
382 // callback invocation is needed.
383 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
387 bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
393 for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
394 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
396 /* Advance and accelerate any new callbacks. */
397 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
399 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
400 // Should we shrink down to a single callback queue?
401 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
407 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
408 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
409 if (rcu_segcblist_pend_cbs(&rtpcp->cblist))
411 if (!rcu_segcblist_empty(&rtpcp->cblist))
413 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
416 // Shrink down to a single callback queue if appropriate.
417 // This is done in two stages: (1) If there are no more than
418 // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
419 // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period,
420 // if there has not been an increase in callbacks, limit dequeuing
421 // to CPU 0. Note the matching RCU read-side critical section in
422 // call_rcu_tasks_generic().
423 if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
424 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
425 if (rtp->percpu_enqueue_lim > 1) {
426 WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
427 smp_store_release(&rtp->percpu_enqueue_lim, 1);
428 rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
430 pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
432 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
434 if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
435 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
436 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
437 WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
438 pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
440 if (rtp->percpu_dequeue_lim == 1) {
441 for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
442 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
444 WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
447 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
453 // Advance callbacks and invoke any that are ready.
454 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
461 struct rcu_head *rhp;
462 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
463 struct rcu_tasks_percpu *rtpcp_next;
466 cpunext = cpu * 2 + 1;
467 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
468 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
469 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
470 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
472 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
473 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
474 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
475 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
479 if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
481 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
482 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
483 rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
484 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
486 for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
492 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
493 rcu_segcblist_add_len(&rtpcp->cblist, -len);
494 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
495 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
498 // Workqueue flood to advance callbacks and invoke any that are ready.
499 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
501 struct rcu_tasks *rtp;
502 struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
505 rcu_tasks_invoke_cbs(rtp, rtpcp);
508 // Wait for one grace period.
509 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
513 mutex_lock(&rtp->tasks_gp_mutex);
515 // If there were none, wait a bit and start over.
516 if (unlikely(midboot)) {
519 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
520 rcuwait_wait_event(&rtp->cbs_wait,
521 (needgpcb = rcu_tasks_need_gpcb(rtp)),
525 if (needgpcb & 0x2) {
526 // Wait for one grace period.
527 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
528 rtp->gp_start = jiffies;
529 rcu_seq_start(&rtp->tasks_gp_seq);
531 rcu_seq_end(&rtp->tasks_gp_seq);
535 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
536 rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
537 mutex_unlock(&rtp->tasks_gp_mutex);
540 // RCU-tasks kthread that detects grace periods and invokes callbacks.
541 static int __noreturn rcu_tasks_kthread(void *arg)
543 struct rcu_tasks *rtp = arg;
545 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
546 housekeeping_affine(current, HK_TYPE_RCU);
547 WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
550 * Each pass through the following loop makes one check for
551 * newly arrived callbacks, and, if there are some, waits for
552 * one RCU-tasks grace period and then invokes the callbacks.
553 * This loop is terminated by the system going down. ;-)
556 // Wait for one grace period and invoke any callbacks
558 rcu_tasks_one_gp(rtp, false);
560 // Paranoid sleep to keep this from entering a tight loop.
561 schedule_timeout_idle(rtp->gp_sleep);
565 // Wait for a grace period for the specified flavor of Tasks RCU.
566 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
568 /* Complain if the scheduler has not started. */
569 if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
570 "synchronize_%s() called too soon", rtp->name))
573 // If the grace-period kthread is running, use it.
574 if (READ_ONCE(rtp->kthread_ptr)) {
575 wait_rcu_gp(rtp->call_func);
578 rcu_tasks_one_gp(rtp, true);
581 /* Spawn RCU-tasks grace-period kthread. */
582 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
584 struct task_struct *t;
586 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
587 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
589 smp_mb(); /* Ensure others see full kthread. */
592 #ifndef CONFIG_TINY_RCU
595 * Print any non-default Tasks RCU settings.
597 static void __init rcu_tasks_bootup_oddness(void)
599 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
602 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
603 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
604 rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
605 if (rtsimc != rcu_task_stall_info_mult) {
606 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
607 rcu_task_stall_info_mult = rtsimc;
609 #endif /* #ifdef CONFIG_TASKS_RCU */
610 #ifdef CONFIG_TASKS_RCU
611 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
612 #endif /* #ifdef CONFIG_TASKS_RCU */
613 #ifdef CONFIG_TASKS_RUDE_RCU
614 pr_info("\tRude variant of Tasks RCU enabled.\n");
615 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
616 #ifdef CONFIG_TASKS_TRACE_RCU
617 pr_info("\tTracing variant of Tasks RCU enabled.\n");
618 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
621 #endif /* #ifndef CONFIG_TINY_RCU */
623 #ifndef CONFIG_TINY_RCU
624 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
625 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
628 bool havecbs = false;
630 for_each_possible_cpu(cpu) {
631 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
633 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist))) {
638 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
640 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
641 jiffies - data_race(rtp->gp_jiffies),
642 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
643 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
644 ".k"[!!data_race(rtp->kthread_ptr)],
648 #endif // #ifndef CONFIG_TINY_RCU
650 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
652 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
654 ////////////////////////////////////////////////////////////////////////
656 // Shared code between task-list-scanning variants of Tasks RCU.
658 /* Wait for one RCU-tasks grace period. */
659 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
661 struct task_struct *g;
665 unsigned long lastinfo;
666 unsigned long lastreport;
667 bool reported = false;
669 struct task_struct *t;
671 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
672 rtp->pregp_func(&holdouts);
675 * There were callbacks, so we need to wait for an RCU-tasks
676 * grace period. Start off by scanning the task list for tasks
677 * that are not already voluntarily blocked. Mark these tasks
678 * and make a list of them in holdouts.
680 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
681 if (rtp->pertask_func) {
683 for_each_process_thread(g, t)
684 rtp->pertask_func(t, &holdouts);
688 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
689 rtp->postscan_func(&holdouts);
692 * Each pass through the following loop scans the list of holdout
693 * tasks, removing any that are no longer holdouts. When the list
694 * is empty, we are done.
696 lastreport = jiffies;
697 lastinfo = lastreport;
698 rtsi = READ_ONCE(rcu_task_stall_info);
700 // Start off with initial wait and slowly back off to 1 HZ wait.
701 fract = rtp->init_fract;
703 while (!list_empty(&holdouts)) {
709 // Slowly back off waiting for holdouts
710 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
711 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
712 schedule_timeout_idle(fract);
714 exp = jiffies_to_nsecs(fract);
715 __set_current_state(TASK_IDLE);
716 schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
722 rtst = READ_ONCE(rcu_task_stall_timeout);
723 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
725 lastreport = jiffies;
729 WARN_ON(signal_pending(current));
730 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
731 rtp->holdouts_func(&holdouts, needreport, &firstreport);
733 // Print pre-stall informational messages if needed.
735 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
737 rtsi = rtsi * rcu_task_stall_info_mult;
738 pr_info("%s: %s grace period %lu is %lu jiffies old.\n",
739 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
743 set_tasks_gp_state(rtp, RTGS_POST_GP);
744 rtp->postgp_func(rtp);
747 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
749 #ifdef CONFIG_TASKS_RCU
751 ////////////////////////////////////////////////////////////////////////
753 // Simple variant of RCU whose quiescent states are voluntary context
754 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
755 // As such, grace periods can take one good long time. There are no
756 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
757 // because this implementation is intended to get the system into a safe
758 // state for some of the manipulations involved in tracing and the like.
759 // Finally, this implementation does not support high call_rcu_tasks()
760 // rates from multiple CPUs. If this is required, per-CPU callback lists
763 // The implementation uses rcu_tasks_wait_gp(), which relies on function
764 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
765 // function sets these function pointers up so that rcu_tasks_wait_gp()
766 // invokes these functions in this order:
768 // rcu_tasks_pregp_step():
769 // Invokes synchronize_rcu() in order to wait for all in-flight
770 // t->on_rq and t->nvcsw transitions to complete. This works because
771 // all such transitions are carried out with interrupts disabled.
772 // rcu_tasks_pertask(), invoked on every non-idle task:
773 // For every runnable non-idle task other than the current one, use
774 // get_task_struct() to pin down that task, snapshot that task's
775 // number of voluntary context switches, and add that task to the
777 // rcu_tasks_postscan():
778 // Invoke synchronize_srcu() to ensure that all tasks that were
779 // in the process of exiting (and which thus might not know to
780 // synchronize with this RCU Tasks grace period) have completed
782 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
783 // Scans the holdout list, attempting to identify a quiescent state
784 // for each task on the list. If there is a quiescent state, the
785 // corresponding task is removed from the holdout list.
786 // rcu_tasks_postgp():
787 // Invokes synchronize_rcu() in order to ensure that all prior
788 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
789 // to have happened before the end of this RCU Tasks grace period.
790 // Again, this works because all such transitions are carried out
791 // with interrupts disabled.
793 // For each exiting task, the exit_tasks_rcu_start() and
794 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
795 // read-side critical sections waited for by rcu_tasks_postscan().
797 // Pre-grace-period update-side code is ordered before the grace
798 // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code
799 // is ordered before the grace period via synchronize_rcu() call in
800 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
803 /* Pre-grace-period preparation. */
804 static void rcu_tasks_pregp_step(struct list_head *hop)
807 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
808 * to complete. Invoking synchronize_rcu() suffices because all
809 * these transitions occur with interrupts disabled. Without this
810 * synchronize_rcu(), a read-side critical section that started
811 * before the grace period might be incorrectly seen as having
812 * started after the grace period.
814 * This synchronize_rcu() also dispenses with the need for a
815 * memory barrier on the first store to t->rcu_tasks_holdout,
816 * as it forces the store to happen after the beginning of the
822 /* Per-task initial processing. */
823 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
825 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
827 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
828 WRITE_ONCE(t->rcu_tasks_holdout, true);
829 list_add(&t->rcu_tasks_holdout_list, hop);
833 /* Processing between scanning taskslist and draining the holdout list. */
834 static void rcu_tasks_postscan(struct list_head *hop)
837 * Exiting tasks may escape the tasklist scan. Those are vulnerable
838 * until their final schedule() with TASK_DEAD state. To cope with
839 * this, divide the fragile exit path part in two intersecting
840 * read side critical sections:
842 * 1) An _SRCU_ read side starting before calling exit_notify(),
843 * which may remove the task from the tasklist, and ending after
844 * the final preempt_disable() call in do_exit().
846 * 2) An _RCU_ read side starting with the final preempt_disable()
847 * call in do_exit() and ending with the final call to schedule()
848 * with TASK_DEAD state.
850 * This handles the part 1). And postgp will handle part 2) with a
851 * call to synchronize_rcu().
853 synchronize_srcu(&tasks_rcu_exit_srcu);
856 /* See if tasks are still holding out, complain if so. */
857 static void check_holdout_task(struct task_struct *t,
858 bool needreport, bool *firstreport)
862 if (!READ_ONCE(t->rcu_tasks_holdout) ||
863 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
864 !READ_ONCE(t->on_rq) ||
865 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
866 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
867 WRITE_ONCE(t->rcu_tasks_holdout, false);
868 list_del_init(&t->rcu_tasks_holdout_list);
872 rcu_request_urgent_qs_task(t);
876 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
877 *firstreport = false;
880 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
881 t, ".I"[is_idle_task(t)],
882 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
883 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
884 t->rcu_tasks_idle_cpu, cpu);
888 /* Scan the holdout lists for tasks no longer holding out. */
889 static void check_all_holdout_tasks(struct list_head *hop,
890 bool needreport, bool *firstreport)
892 struct task_struct *t, *t1;
894 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
895 check_holdout_task(t, needreport, firstreport);
900 /* Finish off the Tasks-RCU grace period. */
901 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
904 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
905 * memory barriers prior to them in the schedule() path, memory
906 * reordering on other CPUs could cause their RCU-tasks read-side
907 * critical sections to extend past the end of the grace period.
908 * However, because these ->nvcsw updates are carried out with
909 * interrupts disabled, we can use synchronize_rcu() to force the
910 * needed ordering on all such CPUs.
912 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
913 * accesses to be within the grace period, avoiding the need for
914 * memory barriers for ->rcu_tasks_holdout accesses.
916 * In addition, this synchronize_rcu() waits for exiting tasks
917 * to complete their final preempt_disable() region of execution,
918 * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
919 * enforcing the whole region before tasklist removal until
920 * the final schedule() with TASK_DEAD state to be an RCU TASKS
921 * read side critical section.
926 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
927 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
930 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
931 * @rhp: structure to be used for queueing the RCU updates.
932 * @func: actual callback function to be invoked after the grace period
934 * The callback function will be invoked some time after a full grace
935 * period elapses, in other words after all currently executing RCU
936 * read-side critical sections have completed. call_rcu_tasks() assumes
937 * that the read-side critical sections end at a voluntary context
938 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
939 * or transition to usermode execution. As such, there are no read-side
940 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
941 * this primitive is intended to determine that all tasks have passed
942 * through a safe state, not so much for data-structure synchronization.
944 * See the description of call_rcu() for more detailed information on
945 * memory ordering guarantees.
947 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
949 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
951 EXPORT_SYMBOL_GPL(call_rcu_tasks);
954 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
956 * Control will return to the caller some time after a full rcu-tasks
957 * grace period has elapsed, in other words after all currently
958 * executing rcu-tasks read-side critical sections have elapsed. These
959 * read-side critical sections are delimited by calls to schedule(),
960 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
961 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
963 * This is a very specialized primitive, intended only for a few uses in
964 * tracing and other situations requiring manipulation of function
965 * preambles and profiling hooks. The synchronize_rcu_tasks() function
966 * is not (yet) intended for heavy use from multiple CPUs.
968 * See the description of synchronize_rcu() for more detailed information
969 * on memory ordering guarantees.
971 void synchronize_rcu_tasks(void)
973 synchronize_rcu_tasks_generic(&rcu_tasks);
975 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
978 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
980 * Although the current implementation is guaranteed to wait, it is not
981 * obligated to, for example, if there are no pending callbacks.
983 void rcu_barrier_tasks(void)
985 rcu_barrier_tasks_generic(&rcu_tasks);
987 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
989 static int __init rcu_spawn_tasks_kthread(void)
991 cblist_init_generic(&rcu_tasks);
992 rcu_tasks.gp_sleep = HZ / 10;
993 rcu_tasks.init_fract = HZ / 10;
994 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
995 rcu_tasks.pertask_func = rcu_tasks_pertask;
996 rcu_tasks.postscan_func = rcu_tasks_postscan;
997 rcu_tasks.holdouts_func = check_all_holdout_tasks;
998 rcu_tasks.postgp_func = rcu_tasks_postgp;
999 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1003 #if !defined(CONFIG_TINY_RCU)
1004 void show_rcu_tasks_classic_gp_kthread(void)
1006 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1008 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1009 #endif // !defined(CONFIG_TINY_RCU)
1012 * Contribute to protect against tasklist scan blind spot while the
1013 * task is exiting and may be removed from the tasklist. See
1014 * corresponding synchronize_srcu() for further details.
1016 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
1018 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
1022 * Contribute to protect against tasklist scan blind spot while the
1023 * task is exiting and may be removed from the tasklist. See
1024 * corresponding synchronize_srcu() for further details.
1026 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
1028 struct task_struct *t = current;
1030 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1034 * Contribute to protect against tasklist scan blind spot while the
1035 * task is exiting and may be removed from the tasklist. See
1036 * corresponding synchronize_srcu() for further details.
1038 void exit_tasks_rcu_finish(void)
1040 exit_tasks_rcu_stop();
1041 exit_tasks_rcu_finish_trace(current);
1044 #else /* #ifdef CONFIG_TASKS_RCU */
1045 void exit_tasks_rcu_start(void) { }
1046 void exit_tasks_rcu_stop(void) { }
1047 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1048 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1050 #ifdef CONFIG_TASKS_RUDE_RCU
1052 ////////////////////////////////////////////////////////////////////////
1054 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1055 // passing an empty function to schedule_on_each_cpu(). This approach
1056 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1057 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1058 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1059 // and induces otherwise unnecessary context switches on all online CPUs,
1060 // whether idle or not.
1062 // Callback handling is provided by the rcu_tasks_kthread() function.
1064 // Ordering is provided by the scheduler's context-switch code.
1066 // Empty function to allow workqueues to force a context switch.
1067 static void rcu_tasks_be_rude(struct work_struct *work)
1071 // Wait for one rude RCU-tasks grace period.
1072 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1074 rtp->n_ipis += cpumask_weight(cpu_online_mask);
1075 schedule_on_each_cpu(rcu_tasks_be_rude);
1078 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1079 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1083 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1084 * @rhp: structure to be used for queueing the RCU updates.
1085 * @func: actual callback function to be invoked after the grace period
1087 * The callback function will be invoked some time after a full grace
1088 * period elapses, in other words after all currently executing RCU
1089 * read-side critical sections have completed. call_rcu_tasks_rude()
1090 * assumes that the read-side critical sections end at context switch,
1091 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1092 * usermode execution is schedulable). As such, there are no read-side
1093 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1094 * this primitive is intended to determine that all tasks have passed
1095 * through a safe state, not so much for data-structure synchronization.
1097 * See the description of call_rcu() for more detailed information on
1098 * memory ordering guarantees.
1100 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1102 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1104 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1107 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1109 * Control will return to the caller some time after a rude rcu-tasks
1110 * grace period has elapsed, in other words after all currently
1111 * executing rcu-tasks read-side critical sections have elapsed. These
1112 * read-side critical sections are delimited by calls to schedule(),
1113 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1114 * context), and (in theory, anyway) cond_resched().
1116 * This is a very specialized primitive, intended only for a few uses in
1117 * tracing and other situations requiring manipulation of function preambles
1118 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
1119 * (yet) intended for heavy use from multiple CPUs.
1121 * See the description of synchronize_rcu() for more detailed information
1122 * on memory ordering guarantees.
1124 void synchronize_rcu_tasks_rude(void)
1126 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1128 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1131 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1133 * Although the current implementation is guaranteed to wait, it is not
1134 * obligated to, for example, if there are no pending callbacks.
1136 void rcu_barrier_tasks_rude(void)
1138 rcu_barrier_tasks_generic(&rcu_tasks_rude);
1140 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1142 static int __init rcu_spawn_tasks_rude_kthread(void)
1144 cblist_init_generic(&rcu_tasks_rude);
1145 rcu_tasks_rude.gp_sleep = HZ / 10;
1146 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1150 #if !defined(CONFIG_TINY_RCU)
1151 void show_rcu_tasks_rude_gp_kthread(void)
1153 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1155 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1156 #endif // !defined(CONFIG_TINY_RCU)
1157 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1159 ////////////////////////////////////////////////////////////////////////
1161 // Tracing variant of Tasks RCU. This variant is designed to be used
1162 // to protect tracing hooks, including those of BPF. This variant
1165 // 1. Has explicit read-side markers to allow finite grace periods
1166 // in the face of in-kernel loops for PREEMPT=n builds.
1168 // 2. Protects code in the idle loop, exception entry/exit, and
1169 // CPU-hotplug code paths, similar to the capabilities of SRCU.
1171 // 3. Avoids expensive read-side instructions, having overhead similar
1172 // to that of Preemptible RCU.
1174 // There are of course downsides. For example, the grace-period code
1175 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1176 // in nohz_full userspace. If needed, these downsides can be at least
1177 // partially remedied.
1179 // Perhaps most important, this variant of RCU does not affect the vanilla
1180 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
1181 // readers can operate from idle, offline, and exception entry/exit in no
1182 // way allows rcu_preempt and rcu_sched readers to also do so.
1184 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1185 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
1186 // function sets these function pointers up so that rcu_tasks_wait_gp()
1187 // invokes these functions in this order:
1189 // rcu_tasks_trace_pregp_step():
1190 // Disables CPU hotplug, adds all currently executing tasks to the
1191 // holdout list, then checks the state of all tasks that blocked
1192 // or were preempted within their current RCU Tasks Trace read-side
1193 // critical section, adding them to the holdout list if appropriate.
1194 // Finally, this function re-enables CPU hotplug.
1195 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1196 // rcu_tasks_trace_postscan():
1197 // Invokes synchronize_rcu() to wait for late-stage exiting tasks
1198 // to finish exiting.
1199 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1200 // Scans the holdout list, attempting to identify a quiescent state
1201 // for each task on the list. If there is a quiescent state, the
1202 // corresponding task is removed from the holdout list. Once this
1203 // list is empty, the grace period has completed.
1204 // rcu_tasks_trace_postgp():
1205 // Provides the needed full memory barrier and does debug checks.
1207 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1209 // Pre-grace-period update-side code is ordered before the grace period
1210 // via the ->cbs_lock and barriers in rcu_tasks_kthread(). Pre-grace-period
1211 // read-side code is ordered before the grace period by atomic operations
1212 // on .b.need_qs flag of each task involved in this process, or by scheduler
1213 // context-switch ordering (for locked-down non-running readers).
1215 // The lockdep state must be outside of #ifdef to be useful.
1216 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1217 static struct lock_class_key rcu_lock_trace_key;
1218 struct lockdep_map rcu_trace_lock_map =
1219 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1220 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1221 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1223 #ifdef CONFIG_TASKS_TRACE_RCU
1225 // Record outstanding IPIs to each CPU. No point in sending two...
1226 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1228 // The number of detections of task quiescent state relying on
1229 // heavyweight readers executing explicit memory barriers.
1230 static unsigned long n_heavy_reader_attempts;
1231 static unsigned long n_heavy_reader_updates;
1232 static unsigned long n_heavy_reader_ofl_updates;
1233 static unsigned long n_trc_holdouts;
1235 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1236 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1239 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1240 static u8 rcu_ld_need_qs(struct task_struct *t)
1242 smp_mb(); // Enforce full grace-period ordering.
1243 return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1246 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1247 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1249 smp_store_release(&t->trc_reader_special.b.need_qs, v);
1250 smp_mb(); // Enforce full grace-period ordering.
1254 * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1255 * the four-byte operand-size restriction of some platforms.
1256 * Returns the old value, which is often ignored.
1258 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1260 union rcu_special ret;
1261 union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1262 union rcu_special trs_new = trs_old;
1264 if (trs_old.b.need_qs != old)
1265 return trs_old.b.need_qs;
1266 trs_new.b.need_qs = new;
1267 ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1268 return ret.b.need_qs;
1270 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1273 * If we are the last reader, signal the grace-period kthread.
1274 * Also remove from the per-CPU list of blocked tasks.
1276 void rcu_read_unlock_trace_special(struct task_struct *t)
1278 unsigned long flags;
1279 struct rcu_tasks_percpu *rtpcp;
1280 union rcu_special trs;
1282 // Open-coded full-word version of rcu_ld_need_qs().
1283 smp_mb(); // Enforce full grace-period ordering.
1284 trs = smp_load_acquire(&t->trc_reader_special);
1286 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1287 smp_mb(); // Pairs with update-side barriers.
1288 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1289 if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1290 u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1291 TRC_NEED_QS_CHECKED);
1293 WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1295 if (trs.b.blocked) {
1296 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1297 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1298 list_del_init(&t->trc_blkd_node);
1299 WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1300 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1302 WRITE_ONCE(t->trc_reader_nesting, 0);
1304 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1306 /* Add a newly blocked reader task to its CPU's list. */
1307 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1309 unsigned long flags;
1310 struct rcu_tasks_percpu *rtpcp;
1312 local_irq_save(flags);
1313 rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1314 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1315 t->trc_blkd_cpu = smp_processor_id();
1316 if (!rtpcp->rtp_blkd_tasks.next)
1317 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1318 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1319 WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1320 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1322 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1324 /* Add a task to the holdout list, if it is not already on the list. */
1325 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1327 if (list_empty(&t->trc_holdout_list)) {
1329 list_add(&t->trc_holdout_list, bhp);
1334 /* Remove a task from the holdout list, if it is in fact present. */
1335 static void trc_del_holdout(struct task_struct *t)
1337 if (!list_empty(&t->trc_holdout_list)) {
1338 list_del_init(&t->trc_holdout_list);
1344 /* IPI handler to check task state. */
1345 static void trc_read_check_handler(void *t_in)
1348 struct task_struct *t = current;
1349 struct task_struct *texp = t_in;
1351 // If the task is no longer running on this CPU, leave.
1352 if (unlikely(texp != t))
1353 goto reset_ipi; // Already on holdout list, so will check later.
1355 // If the task is not in a read-side critical section, and
1356 // if this is the last reader, awaken the grace-period kthread.
1357 nesting = READ_ONCE(t->trc_reader_nesting);
1358 if (likely(!nesting)) {
1359 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1362 // If we are racing with an rcu_read_unlock_trace(), try again later.
1363 if (unlikely(nesting < 0))
1366 // Get here if the task is in a read-side critical section.
1367 // Set its state so that it will update state for the grace-period
1368 // kthread upon exit from that critical section.
1369 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1372 // Allow future IPIs to be sent on CPU and for task.
1373 // Also order this IPI handler against any later manipulations of
1374 // the intended task.
1375 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1376 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1379 /* Callback function for scheduler to check locked-down task. */
1380 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1382 struct list_head *bhp = bhp_in;
1383 int cpu = task_cpu(t);
1385 bool ofl = cpu_is_offline(cpu);
1387 if (task_curr(t) && !ofl) {
1388 // If no chance of heavyweight readers, do it the hard way.
1389 if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1392 // If heavyweight readers are enabled on the remote task,
1393 // we can inspect its state despite its currently running.
1394 // However, we cannot safely change its state.
1395 n_heavy_reader_attempts++;
1396 // Check for "running" idle tasks on offline CPUs.
1397 if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1398 return -EINVAL; // No quiescent state, do it the hard way.
1399 n_heavy_reader_updates++;
1402 // The task is not running, so C-language access is safe.
1403 nesting = t->trc_reader_nesting;
1404 WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1405 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1406 n_heavy_reader_ofl_updates++;
1409 // If not exiting a read-side critical section, mark as checked
1410 // so that the grace-period kthread will remove it from the
1413 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1414 return 0; // In QS, so done.
1417 return -EINVAL; // Reader transitioning, try again later.
1419 // The task is in a read-side critical section, so set up its
1420 // state so that it will update state upon exit from that critical
1422 if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1423 trc_add_holdout(t, bhp);
1427 /* Attempt to extract the state for the specified task. */
1428 static void trc_wait_for_one_reader(struct task_struct *t,
1429 struct list_head *bhp)
1433 // If a previous IPI is still in flight, let it complete.
1434 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1437 // The current task had better be in a quiescent state.
1439 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1440 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1444 // Attempt to nail down the task for inspection.
1446 if (!task_call_func(t, trc_inspect_reader, bhp)) {
1452 // If this task is not yet on the holdout list, then we are in
1453 // an RCU read-side critical section. Otherwise, the invocation of
1454 // trc_add_holdout() that added it to the list did the necessary
1455 // get_task_struct(). Either way, the task cannot be freed out
1456 // from under this code.
1458 // If currently running, send an IPI, either way, add to list.
1459 trc_add_holdout(t, bhp);
1461 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1462 // The task is currently running, so try IPIing it.
1465 // If there is already an IPI outstanding, let it happen.
1466 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1469 per_cpu(trc_ipi_to_cpu, cpu) = true;
1470 t->trc_ipi_to_cpu = cpu;
1471 rcu_tasks_trace.n_ipis++;
1472 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1473 // Just in case there is some other reason for
1474 // failure than the target CPU being offline.
1475 WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n",
1477 rcu_tasks_trace.n_ipis_fails++;
1478 per_cpu(trc_ipi_to_cpu, cpu) = false;
1479 t->trc_ipi_to_cpu = -1;
1485 * Initialize for first-round processing for the specified task.
1486 * Return false if task is NULL or already taken care of, true otherwise.
1488 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1490 // During early boot when there is only the one boot CPU, there
1491 // is no idle task for the other CPUs. Also, the grace-period
1492 // kthread is always in a quiescent state. In addition, just return
1493 // if this task is already on the list.
1494 if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1497 rcu_st_need_qs(t, 0);
1498 t->trc_ipi_to_cpu = -1;
1502 /* Do first-round processing for the specified task. */
1503 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1505 if (rcu_tasks_trace_pertask_prep(t, true))
1506 trc_wait_for_one_reader(t, hop);
1509 /* Initialize for a new RCU-tasks-trace grace period. */
1510 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1512 LIST_HEAD(blkd_tasks);
1514 unsigned long flags;
1515 struct rcu_tasks_percpu *rtpcp;
1516 struct task_struct *t;
1518 // There shouldn't be any old IPIs, but...
1519 for_each_possible_cpu(cpu)
1520 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1522 // Disable CPU hotplug across the CPU scan for the benefit of
1523 // any IPIs that might be needed. This also waits for all readers
1524 // in CPU-hotplug code paths.
1527 // These rcu_tasks_trace_pertask_prep() calls are serialized to
1528 // allow safe access to the hop list.
1529 for_each_online_cpu(cpu) {
1531 t = cpu_curr_snapshot(cpu);
1532 if (rcu_tasks_trace_pertask_prep(t, true))
1533 trc_add_holdout(t, hop);
1535 cond_resched_tasks_rcu_qs();
1538 // Only after all running tasks have been accounted for is it
1539 // safe to take care of the tasks that have blocked within their
1540 // current RCU tasks trace read-side critical section.
1541 for_each_possible_cpu(cpu) {
1542 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1543 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1544 list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1545 while (!list_empty(&blkd_tasks)) {
1547 t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1548 list_del_init(&t->trc_blkd_node);
1549 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1550 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1551 rcu_tasks_trace_pertask(t, hop);
1553 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1555 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1556 cond_resched_tasks_rcu_qs();
1559 // Re-enable CPU hotplug now that the holdout list is populated.
1564 * Do intermediate processing between task and holdout scans.
1566 static void rcu_tasks_trace_postscan(struct list_head *hop)
1568 // Wait for late-stage exiting tasks to finish exiting.
1569 // These might have passed the call to exit_tasks_rcu_finish().
1571 // Any tasks that exit after this point will set
1572 // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1575 /* Communicate task state back to the RCU tasks trace stall warning request. */
1576 struct trc_stall_chk_rdr {
1582 static int trc_check_slow_task(struct task_struct *t, void *arg)
1584 struct trc_stall_chk_rdr *trc_rdrp = arg;
1586 if (task_curr(t) && cpu_online(task_cpu(t)))
1587 return false; // It is running, so decline to inspect it.
1588 trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1589 trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1590 trc_rdrp->needqs = rcu_ld_need_qs(t);
1594 /* Show the state of a task stalling the current RCU tasks trace GP. */
1595 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1598 struct trc_stall_chk_rdr trc_rdr;
1599 bool is_idle_tsk = is_idle_task(t);
1602 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1603 *firstreport = false;
1606 if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1607 pr_alert("P%d: %c%c\n",
1609 ".I"[t->trc_ipi_to_cpu >= 0],
1612 pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1614 ".I"[trc_rdr.ipi_to_cpu >= 0],
1616 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1617 ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1619 " !CN"[trc_rdr.needqs & 0x3],
1620 " ?"[trc_rdr.needqs > 0x3],
1621 cpu, cpu_online(cpu) ? "" : "(offline)");
1625 /* List stalled IPIs for RCU tasks trace. */
1626 static void show_stalled_ipi_trace(void)
1630 for_each_possible_cpu(cpu)
1631 if (per_cpu(trc_ipi_to_cpu, cpu))
1632 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1635 /* Do one scan of the holdout list. */
1636 static void check_all_holdout_tasks_trace(struct list_head *hop,
1637 bool needreport, bool *firstreport)
1639 struct task_struct *g, *t;
1641 // Disable CPU hotplug across the holdout list scan for IPIs.
1644 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1645 // If safe and needed, try to check the current task.
1646 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1647 !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1648 trc_wait_for_one_reader(t, hop);
1650 // If check succeeded, remove this task from the list.
1651 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1652 rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1654 else if (needreport)
1655 show_stalled_task_trace(t, firstreport);
1656 cond_resched_tasks_rcu_qs();
1659 // Re-enable CPU hotplug now that the holdout list scan has completed.
1664 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1665 show_stalled_ipi_trace();
1669 static void rcu_tasks_trace_empty_fn(void *unused)
1673 /* Wait for grace period to complete and provide ordering. */
1674 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1678 // Wait for any lingering IPI handlers to complete. Note that
1679 // if a CPU has gone offline or transitioned to userspace in the
1680 // meantime, all IPI handlers should have been drained beforehand.
1681 // Yes, this assumes that CPUs process IPIs in order. If that ever
1682 // changes, there will need to be a recheck and/or timed wait.
1683 for_each_online_cpu(cpu)
1684 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1685 smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1687 smp_mb(); // Caller's code must be ordered after wakeup.
1688 // Pairs with pretty much every ordering primitive.
1691 /* Report any needed quiescent state for this exiting task. */
1692 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1694 union rcu_special trs = READ_ONCE(t->trc_reader_special);
1696 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1697 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1698 if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1699 rcu_read_unlock_trace_special(t);
1701 WRITE_ONCE(t->trc_reader_nesting, 0);
1705 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1706 * @rhp: structure to be used for queueing the RCU updates.
1707 * @func: actual callback function to be invoked after the grace period
1709 * The callback function will be invoked some time after a trace rcu-tasks
1710 * grace period elapses, in other words after all currently executing
1711 * trace rcu-tasks read-side critical sections have completed. These
1712 * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1713 * and rcu_read_unlock_trace().
1715 * See the description of call_rcu() for more detailed information on
1716 * memory ordering guarantees.
1718 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1720 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1722 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1725 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1727 * Control will return to the caller some time after a trace rcu-tasks
1728 * grace period has elapsed, in other words after all currently executing
1729 * trace rcu-tasks read-side critical sections have elapsed. These read-side
1730 * critical sections are delimited by calls to rcu_read_lock_trace()
1731 * and rcu_read_unlock_trace().
1733 * This is a very specialized primitive, intended only for a few uses in
1734 * tracing and other situations requiring manipulation of function preambles
1735 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1736 * (yet) intended for heavy use from multiple CPUs.
1738 * See the description of synchronize_rcu() for more detailed information
1739 * on memory ordering guarantees.
1741 void synchronize_rcu_tasks_trace(void)
1743 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1744 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1746 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1749 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1751 * Although the current implementation is guaranteed to wait, it is not
1752 * obligated to, for example, if there are no pending callbacks.
1754 void rcu_barrier_tasks_trace(void)
1756 rcu_barrier_tasks_generic(&rcu_tasks_trace);
1758 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1760 static int __init rcu_spawn_tasks_trace_kthread(void)
1762 cblist_init_generic(&rcu_tasks_trace);
1763 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1764 rcu_tasks_trace.gp_sleep = HZ / 10;
1765 rcu_tasks_trace.init_fract = HZ / 10;
1767 rcu_tasks_trace.gp_sleep = HZ / 200;
1768 if (rcu_tasks_trace.gp_sleep <= 0)
1769 rcu_tasks_trace.gp_sleep = 1;
1770 rcu_tasks_trace.init_fract = HZ / 200;
1771 if (rcu_tasks_trace.init_fract <= 0)
1772 rcu_tasks_trace.init_fract = 1;
1774 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1775 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1776 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1777 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1778 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1782 #if !defined(CONFIG_TINY_RCU)
1783 void show_rcu_tasks_trace_gp_kthread(void)
1787 sprintf(buf, "N%lu h:%lu/%lu/%lu",
1788 data_race(n_trc_holdouts),
1789 data_race(n_heavy_reader_ofl_updates),
1790 data_race(n_heavy_reader_updates),
1791 data_race(n_heavy_reader_attempts));
1792 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1794 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1795 #endif // !defined(CONFIG_TINY_RCU)
1797 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1798 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1799 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1801 #ifndef CONFIG_TINY_RCU
1802 void show_rcu_tasks_gp_kthreads(void)
1804 show_rcu_tasks_classic_gp_kthread();
1805 show_rcu_tasks_rude_gp_kthread();
1806 show_rcu_tasks_trace_gp_kthread();
1808 #endif /* #ifndef CONFIG_TINY_RCU */
1810 #ifdef CONFIG_PROVE_RCU
1811 struct rcu_tasks_test_desc {
1815 unsigned long runstart;
1818 static struct rcu_tasks_test_desc tests[] = {
1820 .name = "call_rcu_tasks()",
1821 /* If not defined, the test is skipped. */
1822 .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1825 .name = "call_rcu_tasks_rude()",
1826 /* If not defined, the test is skipped. */
1827 .notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1830 .name = "call_rcu_tasks_trace()",
1831 /* If not defined, the test is skipped. */
1832 .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1836 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1838 struct rcu_tasks_test_desc *rttd =
1839 container_of(rhp, struct rcu_tasks_test_desc, rh);
1841 pr_info("Callback from %s invoked.\n", rttd->name);
1843 rttd->notrun = false;
1846 static void rcu_tasks_initiate_self_tests(void)
1848 unsigned long j = jiffies;
1850 pr_info("Running RCU-tasks wait API self tests\n");
1851 #ifdef CONFIG_TASKS_RCU
1852 tests[0].runstart = j;
1853 synchronize_rcu_tasks();
1854 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1857 #ifdef CONFIG_TASKS_RUDE_RCU
1858 tests[1].runstart = j;
1859 synchronize_rcu_tasks_rude();
1860 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1863 #ifdef CONFIG_TASKS_TRACE_RCU
1864 tests[2].runstart = j;
1865 synchronize_rcu_tasks_trace();
1866 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1871 * Return: 0 - test passed
1872 * 1 - test failed, but have not timed out yet
1873 * -1 - test failed and timed out
1875 static int rcu_tasks_verify_self_tests(void)
1879 unsigned long bst = rcu_task_stall_timeout;
1881 if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
1882 bst = RCU_TASK_BOOT_STALL_TIMEOUT;
1883 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1884 while (tests[i].notrun) { // still hanging.
1885 if (time_after(jiffies, tests[i].runstart + bst)) {
1886 pr_err("%s has failed boot-time tests.\n", tests[i].name);
1900 * Repeat the rcu_tasks_verify_self_tests() call once every second until the
1901 * test passes or has timed out.
1903 static struct delayed_work rcu_tasks_verify_work;
1904 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
1906 int ret = rcu_tasks_verify_self_tests();
1911 /* Test fails but not timed out yet, reschedule another check */
1912 schedule_delayed_work(&rcu_tasks_verify_work, HZ);
1915 static int rcu_tasks_verify_schedule_work(void)
1917 INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
1918 rcu_tasks_verify_work_fn(NULL);
1921 late_initcall(rcu_tasks_verify_schedule_work);
1922 #else /* #ifdef CONFIG_PROVE_RCU */
1923 static void rcu_tasks_initiate_self_tests(void) { }
1924 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1926 void __init rcu_init_tasks_generic(void)
1928 #ifdef CONFIG_TASKS_RCU
1929 rcu_spawn_tasks_kthread();
1932 #ifdef CONFIG_TASKS_RUDE_RCU
1933 rcu_spawn_tasks_rude_kthread();
1936 #ifdef CONFIG_TASKS_TRACE_RCU
1937 rcu_spawn_tasks_trace_kthread();
1940 // Run the self-tests.
1941 rcu_tasks_initiate_self_tests();
1944 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1945 static inline void rcu_tasks_bootup_oddness(void) {}
1946 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */