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 #ifdef CONFIG_TASKS_RCU
140 /* Track exiting tasks in order to allow them to be waited for. */
141 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
144 #ifdef CONFIG_TASKS_RCU
145 /* Report delay in synchronize_srcu() completion in rcu_tasks_postscan(). */
146 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused);
147 static DEFINE_TIMER(tasks_rcu_exit_srcu_stall_timer, tasks_rcu_exit_srcu_stall);
150 /* Avoid IPIing CPUs early in the grace period. */
151 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
152 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
153 module_param(rcu_task_ipi_delay, int, 0644);
155 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
156 #define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
157 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
158 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
159 module_param(rcu_task_stall_timeout, int, 0644);
160 #define RCU_TASK_STALL_INFO (HZ * 10)
161 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
162 module_param(rcu_task_stall_info, int, 0644);
163 static int rcu_task_stall_info_mult __read_mostly = 3;
164 module_param(rcu_task_stall_info_mult, int, 0444);
166 static int rcu_task_enqueue_lim __read_mostly = -1;
167 module_param(rcu_task_enqueue_lim, int, 0444);
169 static bool rcu_task_cb_adjust;
170 static int rcu_task_contend_lim __read_mostly = 100;
171 module_param(rcu_task_contend_lim, int, 0444);
172 static int rcu_task_collapse_lim __read_mostly = 10;
173 module_param(rcu_task_collapse_lim, int, 0444);
175 /* RCU tasks grace-period state for debugging. */
177 #define RTGS_WAIT_WAIT_CBS 1
178 #define RTGS_WAIT_GP 2
179 #define RTGS_PRE_WAIT_GP 3
180 #define RTGS_SCAN_TASKLIST 4
181 #define RTGS_POST_SCAN_TASKLIST 5
182 #define RTGS_WAIT_SCAN_HOLDOUTS 6
183 #define RTGS_SCAN_HOLDOUTS 7
184 #define RTGS_POST_GP 8
185 #define RTGS_WAIT_READERS 9
186 #define RTGS_INVOKE_CBS 10
187 #define RTGS_WAIT_CBS 11
188 #ifndef CONFIG_TINY_RCU
189 static const char * const rcu_tasks_gp_state_names[] = {
191 "RTGS_WAIT_WAIT_CBS",
194 "RTGS_SCAN_TASKLIST",
195 "RTGS_POST_SCAN_TASKLIST",
196 "RTGS_WAIT_SCAN_HOLDOUTS",
197 "RTGS_SCAN_HOLDOUTS",
203 #endif /* #ifndef CONFIG_TINY_RCU */
205 ////////////////////////////////////////////////////////////////////////
209 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
211 /* Record grace-period phase and time. */
212 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
214 rtp->gp_state = newstate;
215 rtp->gp_jiffies = jiffies;
218 #ifndef CONFIG_TINY_RCU
219 /* Return state name. */
220 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
222 int i = data_race(rtp->gp_state); // Let KCSAN detect update races
223 int j = READ_ONCE(i); // Prevent the compiler from reading twice
225 if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
227 return rcu_tasks_gp_state_names[j];
229 #endif /* #ifndef CONFIG_TINY_RCU */
231 // Initialize per-CPU callback lists for the specified flavor of
233 static void cblist_init_generic(struct rcu_tasks *rtp)
240 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
241 if (rcu_task_enqueue_lim < 0) {
242 rcu_task_enqueue_lim = 1;
243 rcu_task_cb_adjust = true;
244 pr_info("%s: Setting adjustable number of callback queues.\n", __func__);
245 } else if (rcu_task_enqueue_lim == 0) {
246 rcu_task_enqueue_lim = 1;
248 lim = rcu_task_enqueue_lim;
250 if (lim > nr_cpu_ids)
252 shift = ilog2(nr_cpu_ids / lim);
253 if (((nr_cpu_ids - 1) >> shift) >= lim)
255 WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
256 WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
257 smp_store_release(&rtp->percpu_enqueue_lim, lim);
258 for_each_possible_cpu(cpu) {
259 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
261 WARN_ON_ONCE(!rtpcp);
263 raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
264 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
265 if (rcu_segcblist_empty(&rtpcp->cblist))
266 rcu_segcblist_init(&rtpcp->cblist);
267 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
270 if (!rtpcp->rtp_blkd_tasks.next)
271 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
272 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
274 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
275 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));
278 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
279 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
281 struct rcu_tasks *rtp;
282 struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
285 rcuwait_wake_up(&rtp->cbs_wait);
288 // Enqueue a callback for the specified flavor of Tasks RCU.
289 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
290 struct rcu_tasks *rtp)
296 bool needadjust = false;
298 struct rcu_tasks_percpu *rtpcp;
302 local_irq_save(flags);
304 ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
305 chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
306 rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
307 if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
308 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
310 if (rtpcp->rtp_jiffies != j) {
311 rtpcp->rtp_jiffies = j;
312 rtpcp->rtp_n_lock_retries = 0;
314 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
315 READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
316 needadjust = true; // Defer adjustment to avoid deadlock.
318 if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
319 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
320 cblist_init_generic(rtp);
321 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
323 needwake = rcu_segcblist_empty(&rtpcp->cblist);
324 rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
325 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
326 if (unlikely(needadjust)) {
327 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
328 if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
329 WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
330 WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
331 smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
332 pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
334 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
337 /* We can't create the thread unless interrupts are enabled. */
338 if (needwake && READ_ONCE(rtp->kthread_ptr))
339 irq_work_queue(&rtpcp->rtp_irq_work);
342 // RCU callback function for rcu_barrier_tasks_generic().
343 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
345 struct rcu_tasks *rtp;
346 struct rcu_tasks_percpu *rtpcp;
348 rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
350 if (atomic_dec_and_test(&rtp->barrier_q_count))
351 complete(&rtp->barrier_q_completion);
354 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
355 // Operates in a manner similar to rcu_barrier().
356 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
360 struct rcu_tasks_percpu *rtpcp;
361 unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
363 mutex_lock(&rtp->barrier_q_mutex);
364 if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
366 mutex_unlock(&rtp->barrier_q_mutex);
369 rcu_seq_start(&rtp->barrier_q_seq);
370 init_completion(&rtp->barrier_q_completion);
371 atomic_set(&rtp->barrier_q_count, 2);
372 for_each_possible_cpu(cpu) {
373 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
375 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
376 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
377 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
378 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
379 atomic_inc(&rtp->barrier_q_count);
380 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
382 if (atomic_sub_and_test(2, &rtp->barrier_q_count))
383 complete(&rtp->barrier_q_completion);
384 wait_for_completion(&rtp->barrier_q_completion);
385 rcu_seq_end(&rtp->barrier_q_seq);
386 mutex_unlock(&rtp->barrier_q_mutex);
389 // Advance callbacks and indicate whether either a grace period or
390 // callback invocation is needed.
391 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
395 bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
401 for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
402 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
404 /* Advance and accelerate any new callbacks. */
405 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
407 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
408 // Should we shrink down to a single callback queue?
409 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
415 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
416 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
417 if (rcu_segcblist_pend_cbs(&rtpcp->cblist))
419 if (!rcu_segcblist_empty(&rtpcp->cblist))
421 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
424 // Shrink down to a single callback queue if appropriate.
425 // This is done in two stages: (1) If there are no more than
426 // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
427 // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period,
428 // if there has not been an increase in callbacks, limit dequeuing
429 // to CPU 0. Note the matching RCU read-side critical section in
430 // call_rcu_tasks_generic().
431 if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
432 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
433 if (rtp->percpu_enqueue_lim > 1) {
434 WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
435 smp_store_release(&rtp->percpu_enqueue_lim, 1);
436 rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
438 pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
440 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
442 if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
443 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
444 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
445 WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
446 pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
448 if (rtp->percpu_dequeue_lim == 1) {
449 for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
450 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
452 WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
455 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
461 // Advance callbacks and invoke any that are ready.
462 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
468 struct rcu_head *rhp;
469 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
470 struct rcu_tasks_percpu *rtpcp_next;
473 cpunext = cpu * 2 + 1;
474 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
475 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
476 queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
478 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
479 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
480 queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
484 if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
486 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
487 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
488 rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
489 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
491 for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
497 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
498 rcu_segcblist_add_len(&rtpcp->cblist, -len);
499 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
500 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
503 // Workqueue flood to advance callbacks and invoke any that are ready.
504 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
506 struct rcu_tasks *rtp;
507 struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
510 rcu_tasks_invoke_cbs(rtp, rtpcp);
513 // Wait for one grace period.
514 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
518 mutex_lock(&rtp->tasks_gp_mutex);
520 // If there were none, wait a bit and start over.
521 if (unlikely(midboot)) {
524 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
525 rcuwait_wait_event(&rtp->cbs_wait,
526 (needgpcb = rcu_tasks_need_gpcb(rtp)),
530 if (needgpcb & 0x2) {
531 // Wait for one grace period.
532 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
533 rtp->gp_start = jiffies;
534 rcu_seq_start(&rtp->tasks_gp_seq);
536 rcu_seq_end(&rtp->tasks_gp_seq);
540 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
541 rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
542 mutex_unlock(&rtp->tasks_gp_mutex);
545 // RCU-tasks kthread that detects grace periods and invokes callbacks.
546 static int __noreturn rcu_tasks_kthread(void *arg)
548 struct rcu_tasks *rtp = arg;
550 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
551 housekeeping_affine(current, HK_TYPE_RCU);
552 WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
555 * Each pass through the following loop makes one check for
556 * newly arrived callbacks, and, if there are some, waits for
557 * one RCU-tasks grace period and then invokes the callbacks.
558 * This loop is terminated by the system going down. ;-)
561 // Wait for one grace period and invoke any callbacks
563 rcu_tasks_one_gp(rtp, false);
565 // Paranoid sleep to keep this from entering a tight loop.
566 schedule_timeout_idle(rtp->gp_sleep);
570 // Wait for a grace period for the specified flavor of Tasks RCU.
571 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
573 /* Complain if the scheduler has not started. */
574 if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
575 "synchronize_%s() called too soon", rtp->name))
578 // If the grace-period kthread is running, use it.
579 if (READ_ONCE(rtp->kthread_ptr)) {
580 wait_rcu_gp(rtp->call_func);
583 rcu_tasks_one_gp(rtp, true);
586 /* Spawn RCU-tasks grace-period kthread. */
587 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
589 struct task_struct *t;
591 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
592 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
594 smp_mb(); /* Ensure others see full kthread. */
597 #ifndef CONFIG_TINY_RCU
600 * Print any non-default Tasks RCU settings.
602 static void __init rcu_tasks_bootup_oddness(void)
604 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
607 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
608 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
609 rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
610 if (rtsimc != rcu_task_stall_info_mult) {
611 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
612 rcu_task_stall_info_mult = rtsimc;
614 #endif /* #ifdef CONFIG_TASKS_RCU */
615 #ifdef CONFIG_TASKS_RCU
616 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
617 #endif /* #ifdef CONFIG_TASKS_RCU */
618 #ifdef CONFIG_TASKS_RUDE_RCU
619 pr_info("\tRude variant of Tasks RCU enabled.\n");
620 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
621 #ifdef CONFIG_TASKS_TRACE_RCU
622 pr_info("\tTracing variant of Tasks RCU enabled.\n");
623 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
626 #endif /* #ifndef CONFIG_TINY_RCU */
628 #ifndef CONFIG_TINY_RCU
629 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
630 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
633 bool havecbs = false;
635 for_each_possible_cpu(cpu) {
636 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
638 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist))) {
643 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
645 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
646 jiffies - data_race(rtp->gp_jiffies),
647 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
648 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
649 ".k"[!!data_race(rtp->kthread_ptr)],
653 #endif // #ifndef CONFIG_TINY_RCU
655 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
657 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
659 ////////////////////////////////////////////////////////////////////////
661 // Shared code between task-list-scanning variants of Tasks RCU.
663 /* Wait for one RCU-tasks grace period. */
664 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
666 struct task_struct *g;
670 unsigned long lastinfo;
671 unsigned long lastreport;
672 bool reported = false;
674 struct task_struct *t;
676 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
677 rtp->pregp_func(&holdouts);
680 * There were callbacks, so we need to wait for an RCU-tasks
681 * grace period. Start off by scanning the task list for tasks
682 * that are not already voluntarily blocked. Mark these tasks
683 * and make a list of them in holdouts.
685 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
686 if (rtp->pertask_func) {
688 for_each_process_thread(g, t)
689 rtp->pertask_func(t, &holdouts);
693 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
694 rtp->postscan_func(&holdouts);
697 * Each pass through the following loop scans the list of holdout
698 * tasks, removing any that are no longer holdouts. When the list
699 * is empty, we are done.
701 lastreport = jiffies;
702 lastinfo = lastreport;
703 rtsi = READ_ONCE(rcu_task_stall_info);
705 // Start off with initial wait and slowly back off to 1 HZ wait.
706 fract = rtp->init_fract;
708 while (!list_empty(&holdouts)) {
714 // Slowly back off waiting for holdouts
715 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
716 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
717 schedule_timeout_idle(fract);
719 exp = jiffies_to_nsecs(fract);
720 __set_current_state(TASK_IDLE);
721 schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
727 rtst = READ_ONCE(rcu_task_stall_timeout);
728 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
730 lastreport = jiffies;
734 WARN_ON(signal_pending(current));
735 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
736 rtp->holdouts_func(&holdouts, needreport, &firstreport);
738 // Print pre-stall informational messages if needed.
740 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
742 rtsi = rtsi * rcu_task_stall_info_mult;
743 pr_info("%s: %s grace period number %lu (since boot) is %lu jiffies old.\n",
744 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
748 set_tasks_gp_state(rtp, RTGS_POST_GP);
749 rtp->postgp_func(rtp);
752 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
754 #ifdef CONFIG_TASKS_RCU
756 ////////////////////////////////////////////////////////////////////////
758 // Simple variant of RCU whose quiescent states are voluntary context
759 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
760 // As such, grace periods can take one good long time. There are no
761 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
762 // because this implementation is intended to get the system into a safe
763 // state for some of the manipulations involved in tracing and the like.
764 // Finally, this implementation does not support high call_rcu_tasks()
765 // rates from multiple CPUs. If this is required, per-CPU callback lists
768 // The implementation uses rcu_tasks_wait_gp(), which relies on function
769 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
770 // function sets these function pointers up so that rcu_tasks_wait_gp()
771 // invokes these functions in this order:
773 // rcu_tasks_pregp_step():
774 // Invokes synchronize_rcu() in order to wait for all in-flight
775 // t->on_rq and t->nvcsw transitions to complete. This works because
776 // all such transitions are carried out with interrupts disabled.
777 // rcu_tasks_pertask(), invoked on every non-idle task:
778 // For every runnable non-idle task other than the current one, use
779 // get_task_struct() to pin down that task, snapshot that task's
780 // number of voluntary context switches, and add that task to the
782 // rcu_tasks_postscan():
783 // Invoke synchronize_srcu() to ensure that all tasks that were
784 // in the process of exiting (and which thus might not know to
785 // synchronize with this RCU Tasks grace period) have completed
787 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
788 // Scans the holdout list, attempting to identify a quiescent state
789 // for each task on the list. If there is a quiescent state, the
790 // corresponding task is removed from the holdout list.
791 // rcu_tasks_postgp():
792 // Invokes synchronize_rcu() in order to ensure that all prior
793 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
794 // to have happened before the end of this RCU Tasks grace period.
795 // Again, this works because all such transitions are carried out
796 // with interrupts disabled.
798 // For each exiting task, the exit_tasks_rcu_start() and
799 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
800 // read-side critical sections waited for by rcu_tasks_postscan().
802 // Pre-grace-period update-side code is ordered before the grace
803 // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code
804 // is ordered before the grace period via synchronize_rcu() call in
805 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
808 /* Pre-grace-period preparation. */
809 static void rcu_tasks_pregp_step(struct list_head *hop)
812 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
813 * to complete. Invoking synchronize_rcu() suffices because all
814 * these transitions occur with interrupts disabled. Without this
815 * synchronize_rcu(), a read-side critical section that started
816 * before the grace period might be incorrectly seen as having
817 * started after the grace period.
819 * This synchronize_rcu() also dispenses with the need for a
820 * memory barrier on the first store to t->rcu_tasks_holdout,
821 * as it forces the store to happen after the beginning of the
827 /* Per-task initial processing. */
828 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
830 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
832 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
833 WRITE_ONCE(t->rcu_tasks_holdout, true);
834 list_add(&t->rcu_tasks_holdout_list, hop);
838 /* Processing between scanning taskslist and draining the holdout list. */
839 static void rcu_tasks_postscan(struct list_head *hop)
841 int rtsi = READ_ONCE(rcu_task_stall_info);
843 if (!IS_ENABLED(CONFIG_TINY_RCU)) {
844 tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
845 add_timer(&tasks_rcu_exit_srcu_stall_timer);
849 * Exiting tasks may escape the tasklist scan. Those are vulnerable
850 * until their final schedule() with TASK_DEAD state. To cope with
851 * this, divide the fragile exit path part in two intersecting
852 * read side critical sections:
854 * 1) An _SRCU_ read side starting before calling exit_notify(),
855 * which may remove the task from the tasklist, and ending after
856 * the final preempt_disable() call in do_exit().
858 * 2) An _RCU_ read side starting with the final preempt_disable()
859 * call in do_exit() and ending with the final call to schedule()
860 * with TASK_DEAD state.
862 * This handles the part 1). And postgp will handle part 2) with a
863 * call to synchronize_rcu().
865 synchronize_srcu(&tasks_rcu_exit_srcu);
867 if (!IS_ENABLED(CONFIG_TINY_RCU))
868 del_timer_sync(&tasks_rcu_exit_srcu_stall_timer);
871 /* See if tasks are still holding out, complain if so. */
872 static void check_holdout_task(struct task_struct *t,
873 bool needreport, bool *firstreport)
877 if (!READ_ONCE(t->rcu_tasks_holdout) ||
878 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
879 !READ_ONCE(t->on_rq) ||
880 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
881 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
882 WRITE_ONCE(t->rcu_tasks_holdout, false);
883 list_del_init(&t->rcu_tasks_holdout_list);
887 rcu_request_urgent_qs_task(t);
891 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
892 *firstreport = false;
895 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
896 t, ".I"[is_idle_task(t)],
897 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
898 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
899 t->rcu_tasks_idle_cpu, cpu);
903 /* Scan the holdout lists for tasks no longer holding out. */
904 static void check_all_holdout_tasks(struct list_head *hop,
905 bool needreport, bool *firstreport)
907 struct task_struct *t, *t1;
909 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
910 check_holdout_task(t, needreport, firstreport);
915 /* Finish off the Tasks-RCU grace period. */
916 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
919 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
920 * memory barriers prior to them in the schedule() path, memory
921 * reordering on other CPUs could cause their RCU-tasks read-side
922 * critical sections to extend past the end of the grace period.
923 * However, because these ->nvcsw updates are carried out with
924 * interrupts disabled, we can use synchronize_rcu() to force the
925 * needed ordering on all such CPUs.
927 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
928 * accesses to be within the grace period, avoiding the need for
929 * memory barriers for ->rcu_tasks_holdout accesses.
931 * In addition, this synchronize_rcu() waits for exiting tasks
932 * to complete their final preempt_disable() region of execution,
933 * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
934 * enforcing the whole region before tasklist removal until
935 * the final schedule() with TASK_DEAD state to be an RCU TASKS
936 * read side critical section.
941 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
942 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
944 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused)
946 #ifndef CONFIG_TINY_RCU
949 rtsi = READ_ONCE(rcu_task_stall_info);
950 pr_info("%s: %s grace period number %lu (since boot) gp_state: %s is %lu jiffies old.\n",
951 __func__, rcu_tasks.kname, rcu_tasks.tasks_gp_seq,
952 tasks_gp_state_getname(&rcu_tasks), jiffies - rcu_tasks.gp_jiffies);
953 pr_info("Please check any exiting tasks stuck between calls to exit_tasks_rcu_start() and exit_tasks_rcu_finish()\n");
954 tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
955 add_timer(&tasks_rcu_exit_srcu_stall_timer);
956 #endif // #ifndef CONFIG_TINY_RCU
960 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
961 * @rhp: structure to be used for queueing the RCU updates.
962 * @func: actual callback function to be invoked after the grace period
964 * The callback function will be invoked some time after a full grace
965 * period elapses, in other words after all currently executing RCU
966 * read-side critical sections have completed. call_rcu_tasks() assumes
967 * that the read-side critical sections end at a voluntary context
968 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
969 * or transition to usermode execution. As such, there are no read-side
970 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
971 * this primitive is intended to determine that all tasks have passed
972 * through a safe state, not so much for data-structure synchronization.
974 * See the description of call_rcu() for more detailed information on
975 * memory ordering guarantees.
977 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
979 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
981 EXPORT_SYMBOL_GPL(call_rcu_tasks);
984 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
986 * Control will return to the caller some time after a full rcu-tasks
987 * grace period has elapsed, in other words after all currently
988 * executing rcu-tasks read-side critical sections have elapsed. These
989 * read-side critical sections are delimited by calls to schedule(),
990 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
991 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
993 * This is a very specialized primitive, intended only for a few uses in
994 * tracing and other situations requiring manipulation of function
995 * preambles and profiling hooks. The synchronize_rcu_tasks() function
996 * is not (yet) intended for heavy use from multiple CPUs.
998 * See the description of synchronize_rcu() for more detailed information
999 * on memory ordering guarantees.
1001 void synchronize_rcu_tasks(void)
1003 synchronize_rcu_tasks_generic(&rcu_tasks);
1005 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
1008 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
1010 * Although the current implementation is guaranteed to wait, it is not
1011 * obligated to, for example, if there are no pending callbacks.
1013 void rcu_barrier_tasks(void)
1015 rcu_barrier_tasks_generic(&rcu_tasks);
1017 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
1019 static int __init rcu_spawn_tasks_kthread(void)
1021 cblist_init_generic(&rcu_tasks);
1022 rcu_tasks.gp_sleep = HZ / 10;
1023 rcu_tasks.init_fract = HZ / 10;
1024 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
1025 rcu_tasks.pertask_func = rcu_tasks_pertask;
1026 rcu_tasks.postscan_func = rcu_tasks_postscan;
1027 rcu_tasks.holdouts_func = check_all_holdout_tasks;
1028 rcu_tasks.postgp_func = rcu_tasks_postgp;
1029 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1033 #if !defined(CONFIG_TINY_RCU)
1034 void show_rcu_tasks_classic_gp_kthread(void)
1036 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1038 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1039 #endif // !defined(CONFIG_TINY_RCU)
1042 * Contribute to protect against tasklist scan blind spot while the
1043 * task is exiting and may be removed from the tasklist. See
1044 * corresponding synchronize_srcu() for further details.
1046 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
1048 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
1052 * Contribute to protect against tasklist scan blind spot while the
1053 * task is exiting and may be removed from the tasklist. See
1054 * corresponding synchronize_srcu() for further details.
1056 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
1058 struct task_struct *t = current;
1060 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1064 * Contribute to protect against tasklist scan blind spot while the
1065 * task is exiting and may be removed from the tasklist. See
1066 * corresponding synchronize_srcu() for further details.
1068 void exit_tasks_rcu_finish(void)
1070 exit_tasks_rcu_stop();
1071 exit_tasks_rcu_finish_trace(current);
1074 #else /* #ifdef CONFIG_TASKS_RCU */
1075 void exit_tasks_rcu_start(void) { }
1076 void exit_tasks_rcu_stop(void) { }
1077 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1078 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1080 #ifdef CONFIG_TASKS_RUDE_RCU
1082 ////////////////////////////////////////////////////////////////////////
1084 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1085 // passing an empty function to schedule_on_each_cpu(). This approach
1086 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1087 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1088 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1089 // and induces otherwise unnecessary context switches on all online CPUs,
1090 // whether idle or not.
1092 // Callback handling is provided by the rcu_tasks_kthread() function.
1094 // Ordering is provided by the scheduler's context-switch code.
1096 // Empty function to allow workqueues to force a context switch.
1097 static void rcu_tasks_be_rude(struct work_struct *work)
1101 // Wait for one rude RCU-tasks grace period.
1102 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1104 rtp->n_ipis += cpumask_weight(cpu_online_mask);
1105 schedule_on_each_cpu(rcu_tasks_be_rude);
1108 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1109 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1113 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1114 * @rhp: structure to be used for queueing the RCU updates.
1115 * @func: actual callback function to be invoked after the grace period
1117 * The callback function will be invoked some time after a full grace
1118 * period elapses, in other words after all currently executing RCU
1119 * read-side critical sections have completed. call_rcu_tasks_rude()
1120 * assumes that the read-side critical sections end at context switch,
1121 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1122 * usermode execution is schedulable). As such, there are no read-side
1123 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1124 * this primitive is intended to determine that all tasks have passed
1125 * through a safe state, not so much for data-structure synchronization.
1127 * See the description of call_rcu() for more detailed information on
1128 * memory ordering guarantees.
1130 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1132 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1134 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1137 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1139 * Control will return to the caller some time after a rude rcu-tasks
1140 * grace period has elapsed, in other words after all currently
1141 * executing rcu-tasks read-side critical sections have elapsed. These
1142 * read-side critical sections are delimited by calls to schedule(),
1143 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1144 * context), and (in theory, anyway) cond_resched().
1146 * This is a very specialized primitive, intended only for a few uses in
1147 * tracing and other situations requiring manipulation of function preambles
1148 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
1149 * (yet) intended for heavy use from multiple CPUs.
1151 * See the description of synchronize_rcu() for more detailed information
1152 * on memory ordering guarantees.
1154 void synchronize_rcu_tasks_rude(void)
1156 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1158 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1161 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1163 * Although the current implementation is guaranteed to wait, it is not
1164 * obligated to, for example, if there are no pending callbacks.
1166 void rcu_barrier_tasks_rude(void)
1168 rcu_barrier_tasks_generic(&rcu_tasks_rude);
1170 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1172 static int __init rcu_spawn_tasks_rude_kthread(void)
1174 cblist_init_generic(&rcu_tasks_rude);
1175 rcu_tasks_rude.gp_sleep = HZ / 10;
1176 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1180 #if !defined(CONFIG_TINY_RCU)
1181 void show_rcu_tasks_rude_gp_kthread(void)
1183 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1185 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1186 #endif // !defined(CONFIG_TINY_RCU)
1187 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1189 ////////////////////////////////////////////////////////////////////////
1191 // Tracing variant of Tasks RCU. This variant is designed to be used
1192 // to protect tracing hooks, including those of BPF. This variant
1195 // 1. Has explicit read-side markers to allow finite grace periods
1196 // in the face of in-kernel loops for PREEMPT=n builds.
1198 // 2. Protects code in the idle loop, exception entry/exit, and
1199 // CPU-hotplug code paths, similar to the capabilities of SRCU.
1201 // 3. Avoids expensive read-side instructions, having overhead similar
1202 // to that of Preemptible RCU.
1204 // There are of course downsides. For example, the grace-period code
1205 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1206 // in nohz_full userspace. If needed, these downsides can be at least
1207 // partially remedied.
1209 // Perhaps most important, this variant of RCU does not affect the vanilla
1210 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
1211 // readers can operate from idle, offline, and exception entry/exit in no
1212 // way allows rcu_preempt and rcu_sched readers to also do so.
1214 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1215 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
1216 // function sets these function pointers up so that rcu_tasks_wait_gp()
1217 // invokes these functions in this order:
1219 // rcu_tasks_trace_pregp_step():
1220 // Disables CPU hotplug, adds all currently executing tasks to the
1221 // holdout list, then checks the state of all tasks that blocked
1222 // or were preempted within their current RCU Tasks Trace read-side
1223 // critical section, adding them to the holdout list if appropriate.
1224 // Finally, this function re-enables CPU hotplug.
1225 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1226 // rcu_tasks_trace_postscan():
1227 // Invokes synchronize_rcu() to wait for late-stage exiting tasks
1228 // to finish exiting.
1229 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1230 // Scans the holdout list, attempting to identify a quiescent state
1231 // for each task on the list. If there is a quiescent state, the
1232 // corresponding task is removed from the holdout list. Once this
1233 // list is empty, the grace period has completed.
1234 // rcu_tasks_trace_postgp():
1235 // Provides the needed full memory barrier and does debug checks.
1237 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1239 // Pre-grace-period update-side code is ordered before the grace period
1240 // via the ->cbs_lock and barriers in rcu_tasks_kthread(). Pre-grace-period
1241 // read-side code is ordered before the grace period by atomic operations
1242 // on .b.need_qs flag of each task involved in this process, or by scheduler
1243 // context-switch ordering (for locked-down non-running readers).
1245 // The lockdep state must be outside of #ifdef to be useful.
1246 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1247 static struct lock_class_key rcu_lock_trace_key;
1248 struct lockdep_map rcu_trace_lock_map =
1249 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1250 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1251 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1253 #ifdef CONFIG_TASKS_TRACE_RCU
1255 // Record outstanding IPIs to each CPU. No point in sending two...
1256 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1258 // The number of detections of task quiescent state relying on
1259 // heavyweight readers executing explicit memory barriers.
1260 static unsigned long n_heavy_reader_attempts;
1261 static unsigned long n_heavy_reader_updates;
1262 static unsigned long n_heavy_reader_ofl_updates;
1263 static unsigned long n_trc_holdouts;
1265 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1266 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1269 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1270 static u8 rcu_ld_need_qs(struct task_struct *t)
1272 smp_mb(); // Enforce full grace-period ordering.
1273 return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1276 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1277 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1279 smp_store_release(&t->trc_reader_special.b.need_qs, v);
1280 smp_mb(); // Enforce full grace-period ordering.
1284 * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1285 * the four-byte operand-size restriction of some platforms.
1286 * Returns the old value, which is often ignored.
1288 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1290 union rcu_special ret;
1291 union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1292 union rcu_special trs_new = trs_old;
1294 if (trs_old.b.need_qs != old)
1295 return trs_old.b.need_qs;
1296 trs_new.b.need_qs = new;
1297 ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1298 return ret.b.need_qs;
1300 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1303 * If we are the last reader, signal the grace-period kthread.
1304 * Also remove from the per-CPU list of blocked tasks.
1306 void rcu_read_unlock_trace_special(struct task_struct *t)
1308 unsigned long flags;
1309 struct rcu_tasks_percpu *rtpcp;
1310 union rcu_special trs;
1312 // Open-coded full-word version of rcu_ld_need_qs().
1313 smp_mb(); // Enforce full grace-period ordering.
1314 trs = smp_load_acquire(&t->trc_reader_special);
1316 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1317 smp_mb(); // Pairs with update-side barriers.
1318 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1319 if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1320 u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1321 TRC_NEED_QS_CHECKED);
1323 WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1325 if (trs.b.blocked) {
1326 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1327 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1328 list_del_init(&t->trc_blkd_node);
1329 WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1330 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1332 WRITE_ONCE(t->trc_reader_nesting, 0);
1334 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1336 /* Add a newly blocked reader task to its CPU's list. */
1337 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1339 unsigned long flags;
1340 struct rcu_tasks_percpu *rtpcp;
1342 local_irq_save(flags);
1343 rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1344 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1345 t->trc_blkd_cpu = smp_processor_id();
1346 if (!rtpcp->rtp_blkd_tasks.next)
1347 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1348 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1349 WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1350 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1352 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1354 /* Add a task to the holdout list, if it is not already on the list. */
1355 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1357 if (list_empty(&t->trc_holdout_list)) {
1359 list_add(&t->trc_holdout_list, bhp);
1364 /* Remove a task from the holdout list, if it is in fact present. */
1365 static void trc_del_holdout(struct task_struct *t)
1367 if (!list_empty(&t->trc_holdout_list)) {
1368 list_del_init(&t->trc_holdout_list);
1374 /* IPI handler to check task state. */
1375 static void trc_read_check_handler(void *t_in)
1378 struct task_struct *t = current;
1379 struct task_struct *texp = t_in;
1381 // If the task is no longer running on this CPU, leave.
1382 if (unlikely(texp != t))
1383 goto reset_ipi; // Already on holdout list, so will check later.
1385 // If the task is not in a read-side critical section, and
1386 // if this is the last reader, awaken the grace-period kthread.
1387 nesting = READ_ONCE(t->trc_reader_nesting);
1388 if (likely(!nesting)) {
1389 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1392 // If we are racing with an rcu_read_unlock_trace(), try again later.
1393 if (unlikely(nesting < 0))
1396 // Get here if the task is in a read-side critical section.
1397 // Set its state so that it will update state for the grace-period
1398 // kthread upon exit from that critical section.
1399 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1402 // Allow future IPIs to be sent on CPU and for task.
1403 // Also order this IPI handler against any later manipulations of
1404 // the intended task.
1405 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1406 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1409 /* Callback function for scheduler to check locked-down task. */
1410 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1412 struct list_head *bhp = bhp_in;
1413 int cpu = task_cpu(t);
1415 bool ofl = cpu_is_offline(cpu);
1417 if (task_curr(t) && !ofl) {
1418 // If no chance of heavyweight readers, do it the hard way.
1419 if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1422 // If heavyweight readers are enabled on the remote task,
1423 // we can inspect its state despite its currently running.
1424 // However, we cannot safely change its state.
1425 n_heavy_reader_attempts++;
1426 // Check for "running" idle tasks on offline CPUs.
1427 if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1428 return -EINVAL; // No quiescent state, do it the hard way.
1429 n_heavy_reader_updates++;
1432 // The task is not running, so C-language access is safe.
1433 nesting = t->trc_reader_nesting;
1434 WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1435 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1436 n_heavy_reader_ofl_updates++;
1439 // If not exiting a read-side critical section, mark as checked
1440 // so that the grace-period kthread will remove it from the
1443 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1444 return 0; // In QS, so done.
1447 return -EINVAL; // Reader transitioning, try again later.
1449 // The task is in a read-side critical section, so set up its
1450 // state so that it will update state upon exit from that critical
1452 if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1453 trc_add_holdout(t, bhp);
1457 /* Attempt to extract the state for the specified task. */
1458 static void trc_wait_for_one_reader(struct task_struct *t,
1459 struct list_head *bhp)
1463 // If a previous IPI is still in flight, let it complete.
1464 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1467 // The current task had better be in a quiescent state.
1469 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1470 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1474 // Attempt to nail down the task for inspection.
1476 if (!task_call_func(t, trc_inspect_reader, bhp)) {
1482 // If this task is not yet on the holdout list, then we are in
1483 // an RCU read-side critical section. Otherwise, the invocation of
1484 // trc_add_holdout() that added it to the list did the necessary
1485 // get_task_struct(). Either way, the task cannot be freed out
1486 // from under this code.
1488 // If currently running, send an IPI, either way, add to list.
1489 trc_add_holdout(t, bhp);
1491 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1492 // The task is currently running, so try IPIing it.
1495 // If there is already an IPI outstanding, let it happen.
1496 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1499 per_cpu(trc_ipi_to_cpu, cpu) = true;
1500 t->trc_ipi_to_cpu = cpu;
1501 rcu_tasks_trace.n_ipis++;
1502 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1503 // Just in case there is some other reason for
1504 // failure than the target CPU being offline.
1505 WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n",
1507 rcu_tasks_trace.n_ipis_fails++;
1508 per_cpu(trc_ipi_to_cpu, cpu) = false;
1509 t->trc_ipi_to_cpu = -1;
1515 * Initialize for first-round processing for the specified task.
1516 * Return false if task is NULL or already taken care of, true otherwise.
1518 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1520 // During early boot when there is only the one boot CPU, there
1521 // is no idle task for the other CPUs. Also, the grace-period
1522 // kthread is always in a quiescent state. In addition, just return
1523 // if this task is already on the list.
1524 if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1527 rcu_st_need_qs(t, 0);
1528 t->trc_ipi_to_cpu = -1;
1532 /* Do first-round processing for the specified task. */
1533 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1535 if (rcu_tasks_trace_pertask_prep(t, true))
1536 trc_wait_for_one_reader(t, hop);
1539 /* Initialize for a new RCU-tasks-trace grace period. */
1540 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1542 LIST_HEAD(blkd_tasks);
1544 unsigned long flags;
1545 struct rcu_tasks_percpu *rtpcp;
1546 struct task_struct *t;
1548 // There shouldn't be any old IPIs, but...
1549 for_each_possible_cpu(cpu)
1550 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1552 // Disable CPU hotplug across the CPU scan for the benefit of
1553 // any IPIs that might be needed. This also waits for all readers
1554 // in CPU-hotplug code paths.
1557 // These rcu_tasks_trace_pertask_prep() calls are serialized to
1558 // allow safe access to the hop list.
1559 for_each_online_cpu(cpu) {
1561 t = cpu_curr_snapshot(cpu);
1562 if (rcu_tasks_trace_pertask_prep(t, true))
1563 trc_add_holdout(t, hop);
1565 cond_resched_tasks_rcu_qs();
1568 // Only after all running tasks have been accounted for is it
1569 // safe to take care of the tasks that have blocked within their
1570 // current RCU tasks trace read-side critical section.
1571 for_each_possible_cpu(cpu) {
1572 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1573 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1574 list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1575 while (!list_empty(&blkd_tasks)) {
1577 t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1578 list_del_init(&t->trc_blkd_node);
1579 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1580 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1581 rcu_tasks_trace_pertask(t, hop);
1583 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1585 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1586 cond_resched_tasks_rcu_qs();
1589 // Re-enable CPU hotplug now that the holdout list is populated.
1594 * Do intermediate processing between task and holdout scans.
1596 static void rcu_tasks_trace_postscan(struct list_head *hop)
1598 // Wait for late-stage exiting tasks to finish exiting.
1599 // These might have passed the call to exit_tasks_rcu_finish().
1601 // If you remove the following line, update rcu_trace_implies_rcu_gp()!!!
1603 // Any tasks that exit after this point will set
1604 // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1607 /* Communicate task state back to the RCU tasks trace stall warning request. */
1608 struct trc_stall_chk_rdr {
1614 static int trc_check_slow_task(struct task_struct *t, void *arg)
1616 struct trc_stall_chk_rdr *trc_rdrp = arg;
1618 if (task_curr(t) && cpu_online(task_cpu(t)))
1619 return false; // It is running, so decline to inspect it.
1620 trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1621 trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1622 trc_rdrp->needqs = rcu_ld_need_qs(t);
1626 /* Show the state of a task stalling the current RCU tasks trace GP. */
1627 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1630 struct trc_stall_chk_rdr trc_rdr;
1631 bool is_idle_tsk = is_idle_task(t);
1634 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1635 *firstreport = false;
1638 if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1639 pr_alert("P%d: %c%c\n",
1641 ".I"[t->trc_ipi_to_cpu >= 0],
1644 pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1646 ".I"[trc_rdr.ipi_to_cpu >= 0],
1648 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1649 ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1651 " !CN"[trc_rdr.needqs & 0x3],
1652 " ?"[trc_rdr.needqs > 0x3],
1653 cpu, cpu_online(cpu) ? "" : "(offline)");
1657 /* List stalled IPIs for RCU tasks trace. */
1658 static void show_stalled_ipi_trace(void)
1662 for_each_possible_cpu(cpu)
1663 if (per_cpu(trc_ipi_to_cpu, cpu))
1664 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1667 /* Do one scan of the holdout list. */
1668 static void check_all_holdout_tasks_trace(struct list_head *hop,
1669 bool needreport, bool *firstreport)
1671 struct task_struct *g, *t;
1673 // Disable CPU hotplug across the holdout list scan for IPIs.
1676 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1677 // If safe and needed, try to check the current task.
1678 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1679 !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1680 trc_wait_for_one_reader(t, hop);
1682 // If check succeeded, remove this task from the list.
1683 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1684 rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1686 else if (needreport)
1687 show_stalled_task_trace(t, firstreport);
1688 cond_resched_tasks_rcu_qs();
1691 // Re-enable CPU hotplug now that the holdout list scan has completed.
1696 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1697 show_stalled_ipi_trace();
1701 static void rcu_tasks_trace_empty_fn(void *unused)
1705 /* Wait for grace period to complete and provide ordering. */
1706 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1710 // Wait for any lingering IPI handlers to complete. Note that
1711 // if a CPU has gone offline or transitioned to userspace in the
1712 // meantime, all IPI handlers should have been drained beforehand.
1713 // Yes, this assumes that CPUs process IPIs in order. If that ever
1714 // changes, there will need to be a recheck and/or timed wait.
1715 for_each_online_cpu(cpu)
1716 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1717 smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1719 smp_mb(); // Caller's code must be ordered after wakeup.
1720 // Pairs with pretty much every ordering primitive.
1723 /* Report any needed quiescent state for this exiting task. */
1724 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1726 union rcu_special trs = READ_ONCE(t->trc_reader_special);
1728 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1729 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1730 if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1731 rcu_read_unlock_trace_special(t);
1733 WRITE_ONCE(t->trc_reader_nesting, 0);
1737 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1738 * @rhp: structure to be used for queueing the RCU updates.
1739 * @func: actual callback function to be invoked after the grace period
1741 * The callback function will be invoked some time after a trace rcu-tasks
1742 * grace period elapses, in other words after all currently executing
1743 * trace rcu-tasks read-side critical sections have completed. These
1744 * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1745 * and rcu_read_unlock_trace().
1747 * See the description of call_rcu() for more detailed information on
1748 * memory ordering guarantees.
1750 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1752 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1754 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1757 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1759 * Control will return to the caller some time after a trace rcu-tasks
1760 * grace period has elapsed, in other words after all currently executing
1761 * trace rcu-tasks read-side critical sections have elapsed. These read-side
1762 * critical sections are delimited by calls to rcu_read_lock_trace()
1763 * and rcu_read_unlock_trace().
1765 * This is a very specialized primitive, intended only for a few uses in
1766 * tracing and other situations requiring manipulation of function preambles
1767 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1768 * (yet) intended for heavy use from multiple CPUs.
1770 * See the description of synchronize_rcu() for more detailed information
1771 * on memory ordering guarantees.
1773 void synchronize_rcu_tasks_trace(void)
1775 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1776 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1778 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1781 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1783 * Although the current implementation is guaranteed to wait, it is not
1784 * obligated to, for example, if there are no pending callbacks.
1786 void rcu_barrier_tasks_trace(void)
1788 rcu_barrier_tasks_generic(&rcu_tasks_trace);
1790 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1792 static int __init rcu_spawn_tasks_trace_kthread(void)
1794 cblist_init_generic(&rcu_tasks_trace);
1795 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1796 rcu_tasks_trace.gp_sleep = HZ / 10;
1797 rcu_tasks_trace.init_fract = HZ / 10;
1799 rcu_tasks_trace.gp_sleep = HZ / 200;
1800 if (rcu_tasks_trace.gp_sleep <= 0)
1801 rcu_tasks_trace.gp_sleep = 1;
1802 rcu_tasks_trace.init_fract = HZ / 200;
1803 if (rcu_tasks_trace.init_fract <= 0)
1804 rcu_tasks_trace.init_fract = 1;
1806 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1807 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1808 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1809 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1810 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1814 #if !defined(CONFIG_TINY_RCU)
1815 void show_rcu_tasks_trace_gp_kthread(void)
1819 sprintf(buf, "N%lu h:%lu/%lu/%lu",
1820 data_race(n_trc_holdouts),
1821 data_race(n_heavy_reader_ofl_updates),
1822 data_race(n_heavy_reader_updates),
1823 data_race(n_heavy_reader_attempts));
1824 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1826 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1827 #endif // !defined(CONFIG_TINY_RCU)
1829 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1830 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1831 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1833 #ifndef CONFIG_TINY_RCU
1834 void show_rcu_tasks_gp_kthreads(void)
1836 show_rcu_tasks_classic_gp_kthread();
1837 show_rcu_tasks_rude_gp_kthread();
1838 show_rcu_tasks_trace_gp_kthread();
1840 #endif /* #ifndef CONFIG_TINY_RCU */
1842 #ifdef CONFIG_PROVE_RCU
1843 struct rcu_tasks_test_desc {
1847 unsigned long runstart;
1850 static struct rcu_tasks_test_desc tests[] = {
1852 .name = "call_rcu_tasks()",
1853 /* If not defined, the test is skipped. */
1854 .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1857 .name = "call_rcu_tasks_rude()",
1858 /* If not defined, the test is skipped. */
1859 .notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1862 .name = "call_rcu_tasks_trace()",
1863 /* If not defined, the test is skipped. */
1864 .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1868 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1870 struct rcu_tasks_test_desc *rttd =
1871 container_of(rhp, struct rcu_tasks_test_desc, rh);
1873 pr_info("Callback from %s invoked.\n", rttd->name);
1875 rttd->notrun = false;
1878 static void rcu_tasks_initiate_self_tests(void)
1880 pr_info("Running RCU-tasks wait API self tests\n");
1881 #ifdef CONFIG_TASKS_RCU
1882 tests[0].runstart = jiffies;
1883 synchronize_rcu_tasks();
1884 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1887 #ifdef CONFIG_TASKS_RUDE_RCU
1888 tests[1].runstart = jiffies;
1889 synchronize_rcu_tasks_rude();
1890 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1893 #ifdef CONFIG_TASKS_TRACE_RCU
1894 tests[2].runstart = jiffies;
1895 synchronize_rcu_tasks_trace();
1896 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1901 * Return: 0 - test passed
1902 * 1 - test failed, but have not timed out yet
1903 * -1 - test failed and timed out
1905 static int rcu_tasks_verify_self_tests(void)
1909 unsigned long bst = rcu_task_stall_timeout;
1911 if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
1912 bst = RCU_TASK_BOOT_STALL_TIMEOUT;
1913 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1914 while (tests[i].notrun) { // still hanging.
1915 if (time_after(jiffies, tests[i].runstart + bst)) {
1916 pr_err("%s has failed boot-time tests.\n", tests[i].name);
1930 * Repeat the rcu_tasks_verify_self_tests() call once every second until the
1931 * test passes or has timed out.
1933 static struct delayed_work rcu_tasks_verify_work;
1934 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
1936 int ret = rcu_tasks_verify_self_tests();
1941 /* Test fails but not timed out yet, reschedule another check */
1942 schedule_delayed_work(&rcu_tasks_verify_work, HZ);
1945 static int rcu_tasks_verify_schedule_work(void)
1947 INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
1948 rcu_tasks_verify_work_fn(NULL);
1951 late_initcall(rcu_tasks_verify_schedule_work);
1952 #else /* #ifdef CONFIG_PROVE_RCU */
1953 static void rcu_tasks_initiate_self_tests(void) { }
1954 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1956 void __init rcu_init_tasks_generic(void)
1958 #ifdef CONFIG_TASKS_RCU
1959 rcu_spawn_tasks_kthread();
1962 #ifdef CONFIG_TASKS_RUDE_RCU
1963 rcu_spawn_tasks_rude_kthread();
1966 #ifdef CONFIG_TASKS_TRACE_RCU
1967 rcu_spawn_tasks_trace_kthread();
1970 // Run the self-tests.
1971 rcu_tasks_initiate_self_tests();
1974 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1975 static inline void rcu_tasks_bootup_oddness(void) {}
1976 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */