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 * @lazy_timer: Timer to unlazify callbacks.
29 * @urgent_gp: Number of additional non-lazy grace periods.
30 * @rtp_n_lock_retries: Rough lock-contention statistic.
31 * @rtp_work: Work queue for invoking callbacks.
32 * @rtp_irq_work: IRQ work queue for deferred wakeups.
33 * @barrier_q_head: RCU callback for barrier operation.
34 * @rtp_blkd_tasks: List of tasks blocked as readers.
35 * @cpu: CPU number corresponding to this entry.
36 * @rtpp: Pointer to the rcu_tasks structure.
38 struct rcu_tasks_percpu {
39 struct rcu_segcblist cblist;
40 raw_spinlock_t __private lock;
41 unsigned long rtp_jiffies;
42 unsigned long rtp_n_lock_retries;
43 struct timer_list lazy_timer;
44 unsigned int urgent_gp;
45 struct work_struct rtp_work;
46 struct irq_work rtp_irq_work;
47 struct rcu_head barrier_q_head;
48 struct list_head rtp_blkd_tasks;
50 struct rcu_tasks *rtpp;
54 * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
55 * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
56 * @cbs_gbl_lock: Lock protecting callback list.
57 * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
58 * @gp_func: This flavor's grace-period-wait function.
59 * @gp_state: Grace period's most recent state transition (debugging).
60 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
61 * @init_fract: Initial backoff sleep interval.
62 * @gp_jiffies: Time of last @gp_state transition.
63 * @gp_start: Most recent grace-period start in jiffies.
64 * @tasks_gp_seq: Number of grace periods completed since boot.
65 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
66 * @n_ipis_fails: Number of IPI-send failures.
67 * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
68 * @lazy_jiffies: Number of jiffies to allow callbacks to be lazy.
69 * @pregp_func: This flavor's pre-grace-period function (optional).
70 * @pertask_func: This flavor's per-task scan function (optional).
71 * @postscan_func: This flavor's post-task scan function (optional).
72 * @holdouts_func: This flavor's holdout-list scan function (optional).
73 * @postgp_func: This flavor's post-grace-period function (optional).
74 * @call_func: This flavor's call_rcu()-equivalent function.
75 * @rtpcpu: This flavor's rcu_tasks_percpu structure.
76 * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
77 * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
78 * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
79 * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
80 * @barrier_q_mutex: Serialize barrier operations.
81 * @barrier_q_count: Number of queues being waited on.
82 * @barrier_q_completion: Barrier wait/wakeup mechanism.
83 * @barrier_q_seq: Sequence number for barrier operations.
84 * @name: This flavor's textual name.
85 * @kname: This flavor's kthread name.
88 struct rcuwait cbs_wait;
89 raw_spinlock_t cbs_gbl_lock;
90 struct mutex tasks_gp_mutex;
94 unsigned long gp_jiffies;
95 unsigned long gp_start;
96 unsigned long tasks_gp_seq;
98 unsigned long n_ipis_fails;
99 struct task_struct *kthread_ptr;
100 unsigned long lazy_jiffies;
101 rcu_tasks_gp_func_t gp_func;
102 pregp_func_t pregp_func;
103 pertask_func_t pertask_func;
104 postscan_func_t postscan_func;
105 holdouts_func_t holdouts_func;
106 postgp_func_t postgp_func;
107 call_rcu_func_t call_func;
108 struct rcu_tasks_percpu __percpu *rtpcpu;
109 int percpu_enqueue_shift;
110 int percpu_enqueue_lim;
111 int percpu_dequeue_lim;
112 unsigned long percpu_dequeue_gpseq;
113 struct mutex barrier_q_mutex;
114 atomic_t barrier_q_count;
115 struct completion barrier_q_completion;
116 unsigned long barrier_q_seq;
121 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
123 #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \
124 static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = { \
125 .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock), \
126 .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup), \
128 static struct rcu_tasks rt_name = \
130 .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait), \
131 .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \
132 .tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex), \
135 .rtpcpu = &rt_name ## __percpu, \
136 .lazy_jiffies = DIV_ROUND_UP(HZ, 4), \
138 .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS), \
139 .percpu_enqueue_lim = 1, \
140 .percpu_dequeue_lim = 1, \
141 .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex), \
142 .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT, \
146 #ifdef CONFIG_TASKS_RCU
147 /* Track exiting tasks in order to allow them to be waited for. */
148 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
150 /* Report delay in synchronize_srcu() completion in rcu_tasks_postscan(). */
151 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused);
152 static DEFINE_TIMER(tasks_rcu_exit_srcu_stall_timer, tasks_rcu_exit_srcu_stall);
155 /* Avoid IPIing CPUs early in the grace period. */
156 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
157 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
158 module_param(rcu_task_ipi_delay, int, 0644);
160 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
161 #define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
162 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
163 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
164 module_param(rcu_task_stall_timeout, int, 0644);
165 #define RCU_TASK_STALL_INFO (HZ * 10)
166 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
167 module_param(rcu_task_stall_info, int, 0644);
168 static int rcu_task_stall_info_mult __read_mostly = 3;
169 module_param(rcu_task_stall_info_mult, int, 0444);
171 static int rcu_task_enqueue_lim __read_mostly = -1;
172 module_param(rcu_task_enqueue_lim, int, 0444);
174 static bool rcu_task_cb_adjust;
175 static int rcu_task_contend_lim __read_mostly = 100;
176 module_param(rcu_task_contend_lim, int, 0444);
177 static int rcu_task_collapse_lim __read_mostly = 10;
178 module_param(rcu_task_collapse_lim, int, 0444);
179 static int rcu_task_lazy_lim __read_mostly = 32;
180 module_param(rcu_task_lazy_lim, int, 0444);
182 /* RCU tasks grace-period state for debugging. */
184 #define RTGS_WAIT_WAIT_CBS 1
185 #define RTGS_WAIT_GP 2
186 #define RTGS_PRE_WAIT_GP 3
187 #define RTGS_SCAN_TASKLIST 4
188 #define RTGS_POST_SCAN_TASKLIST 5
189 #define RTGS_WAIT_SCAN_HOLDOUTS 6
190 #define RTGS_SCAN_HOLDOUTS 7
191 #define RTGS_POST_GP 8
192 #define RTGS_WAIT_READERS 9
193 #define RTGS_INVOKE_CBS 10
194 #define RTGS_WAIT_CBS 11
195 #ifndef CONFIG_TINY_RCU
196 static const char * const rcu_tasks_gp_state_names[] = {
198 "RTGS_WAIT_WAIT_CBS",
201 "RTGS_SCAN_TASKLIST",
202 "RTGS_POST_SCAN_TASKLIST",
203 "RTGS_WAIT_SCAN_HOLDOUTS",
204 "RTGS_SCAN_HOLDOUTS",
210 #endif /* #ifndef CONFIG_TINY_RCU */
212 ////////////////////////////////////////////////////////////////////////
216 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
218 /* Record grace-period phase and time. */
219 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
221 rtp->gp_state = newstate;
222 rtp->gp_jiffies = jiffies;
225 #ifndef CONFIG_TINY_RCU
226 /* Return state name. */
227 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
229 int i = data_race(rtp->gp_state); // Let KCSAN detect update races
230 int j = READ_ONCE(i); // Prevent the compiler from reading twice
232 if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
234 return rcu_tasks_gp_state_names[j];
236 #endif /* #ifndef CONFIG_TINY_RCU */
238 // Initialize per-CPU callback lists for the specified flavor of
239 // Tasks RCU. Do not enqueue callbacks before this function is invoked.
240 static void cblist_init_generic(struct rcu_tasks *rtp)
247 if (rcu_task_enqueue_lim < 0) {
248 rcu_task_enqueue_lim = 1;
249 rcu_task_cb_adjust = true;
250 } else if (rcu_task_enqueue_lim == 0) {
251 rcu_task_enqueue_lim = 1;
253 lim = rcu_task_enqueue_lim;
255 if (lim > nr_cpu_ids)
257 shift = ilog2(nr_cpu_ids / lim);
258 if (((nr_cpu_ids - 1) >> shift) >= lim)
260 WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
261 WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
262 smp_store_release(&rtp->percpu_enqueue_lim, lim);
263 for_each_possible_cpu(cpu) {
264 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
266 WARN_ON_ONCE(!rtpcp);
268 raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
269 local_irq_save(flags); // serialize initialization
270 if (rcu_segcblist_empty(&rtpcp->cblist))
271 rcu_segcblist_init(&rtpcp->cblist);
272 local_irq_restore(flags);
273 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
276 if (!rtpcp->rtp_blkd_tasks.next)
277 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
280 pr_info("%s: Setting shift to %d and lim to %d rcu_task_cb_adjust=%d.\n", rtp->name,
281 data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim), rcu_task_cb_adjust);
284 // Compute wakeup time for lazy callback timer.
285 static unsigned long rcu_tasks_lazy_time(struct rcu_tasks *rtp)
287 return jiffies + rtp->lazy_jiffies;
290 // Timer handler that unlazifies lazy callbacks.
291 static void call_rcu_tasks_generic_timer(struct timer_list *tlp)
294 bool needwake = false;
295 struct rcu_tasks *rtp;
296 struct rcu_tasks_percpu *rtpcp = from_timer(rtpcp, tlp, lazy_timer);
299 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
300 if (!rcu_segcblist_empty(&rtpcp->cblist) && rtp->lazy_jiffies) {
301 if (!rtpcp->urgent_gp)
302 rtpcp->urgent_gp = 1;
304 mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
306 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
308 rcuwait_wake_up(&rtp->cbs_wait);
311 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
312 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
314 struct rcu_tasks *rtp;
315 struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
318 rcuwait_wake_up(&rtp->cbs_wait);
321 // Enqueue a callback for the specified flavor of Tasks RCU.
322 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
323 struct rcu_tasks *rtp)
327 bool havekthread = smp_load_acquire(&rtp->kthread_ptr);
330 bool needadjust = false;
332 struct rcu_tasks_percpu *rtpcp;
336 local_irq_save(flags);
338 ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
339 chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
340 rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
341 if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
342 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
344 if (rtpcp->rtp_jiffies != j) {
345 rtpcp->rtp_jiffies = j;
346 rtpcp->rtp_n_lock_retries = 0;
348 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
349 READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
350 needadjust = true; // Defer adjustment to avoid deadlock.
352 // Queuing callbacks before initialization not yet supported.
353 if (WARN_ON_ONCE(!rcu_segcblist_is_enabled(&rtpcp->cblist)))
354 rcu_segcblist_init(&rtpcp->cblist);
355 needwake = (func == wakeme_after_rcu) ||
356 (rcu_segcblist_n_cbs(&rtpcp->cblist) == rcu_task_lazy_lim);
357 if (havekthread && !needwake && !timer_pending(&rtpcp->lazy_timer)) {
358 if (rtp->lazy_jiffies)
359 mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
361 needwake = rcu_segcblist_empty(&rtpcp->cblist);
364 rtpcp->urgent_gp = 3;
365 rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
366 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
367 if (unlikely(needadjust)) {
368 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
369 if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
370 WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
371 WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
372 smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
373 pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
375 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
378 /* We can't create the thread unless interrupts are enabled. */
379 if (needwake && READ_ONCE(rtp->kthread_ptr))
380 irq_work_queue(&rtpcp->rtp_irq_work);
383 // RCU callback function for rcu_barrier_tasks_generic().
384 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
386 struct rcu_tasks *rtp;
387 struct rcu_tasks_percpu *rtpcp;
389 rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
391 if (atomic_dec_and_test(&rtp->barrier_q_count))
392 complete(&rtp->barrier_q_completion);
395 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
396 // Operates in a manner similar to rcu_barrier().
397 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
401 struct rcu_tasks_percpu *rtpcp;
402 unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
404 mutex_lock(&rtp->barrier_q_mutex);
405 if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
407 mutex_unlock(&rtp->barrier_q_mutex);
410 rcu_seq_start(&rtp->barrier_q_seq);
411 init_completion(&rtp->barrier_q_completion);
412 atomic_set(&rtp->barrier_q_count, 2);
413 for_each_possible_cpu(cpu) {
414 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
416 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
417 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
418 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
419 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
420 atomic_inc(&rtp->barrier_q_count);
421 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
423 if (atomic_sub_and_test(2, &rtp->barrier_q_count))
424 complete(&rtp->barrier_q_completion);
425 wait_for_completion(&rtp->barrier_q_completion);
426 rcu_seq_end(&rtp->barrier_q_seq);
427 mutex_unlock(&rtp->barrier_q_mutex);
430 // Advance callbacks and indicate whether either a grace period or
431 // callback invocation is needed.
432 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
436 bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
442 for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
443 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
445 /* Advance and accelerate any new callbacks. */
446 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
448 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
449 // Should we shrink down to a single callback queue?
450 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
456 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
457 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
458 if (rtpcp->urgent_gp > 0 && rcu_segcblist_pend_cbs(&rtpcp->cblist)) {
459 if (rtp->lazy_jiffies)
462 } else if (rcu_segcblist_empty(&rtpcp->cblist)) {
463 rtpcp->urgent_gp = 0;
465 if (rcu_segcblist_ready_cbs(&rtpcp->cblist))
467 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
470 // Shrink down to a single callback queue if appropriate.
471 // This is done in two stages: (1) If there are no more than
472 // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
473 // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period,
474 // if there has not been an increase in callbacks, limit dequeuing
475 // to CPU 0. Note the matching RCU read-side critical section in
476 // call_rcu_tasks_generic().
477 if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
478 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
479 if (rtp->percpu_enqueue_lim > 1) {
480 WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
481 smp_store_release(&rtp->percpu_enqueue_lim, 1);
482 rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
484 pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
486 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
488 if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
489 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
490 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
491 WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
492 pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
494 if (rtp->percpu_dequeue_lim == 1) {
495 for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
496 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
498 WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
501 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
507 // Advance callbacks and invoke any that are ready.
508 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
515 struct rcu_head *rhp;
516 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
517 struct rcu_tasks_percpu *rtpcp_next;
520 cpunext = cpu * 2 + 1;
521 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
522 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
523 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
524 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
526 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
527 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
528 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
529 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
533 if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
535 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
536 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
537 rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
538 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
540 for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
546 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
547 rcu_segcblist_add_len(&rtpcp->cblist, -len);
548 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
549 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
552 // Workqueue flood to advance callbacks and invoke any that are ready.
553 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
555 struct rcu_tasks *rtp;
556 struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
559 rcu_tasks_invoke_cbs(rtp, rtpcp);
562 // Wait for one grace period.
563 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
567 mutex_lock(&rtp->tasks_gp_mutex);
569 // If there were none, wait a bit and start over.
570 if (unlikely(midboot)) {
573 mutex_unlock(&rtp->tasks_gp_mutex);
574 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
575 rcuwait_wait_event(&rtp->cbs_wait,
576 (needgpcb = rcu_tasks_need_gpcb(rtp)),
578 mutex_lock(&rtp->tasks_gp_mutex);
581 if (needgpcb & 0x2) {
582 // Wait for one grace period.
583 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
584 rtp->gp_start = jiffies;
585 rcu_seq_start(&rtp->tasks_gp_seq);
587 rcu_seq_end(&rtp->tasks_gp_seq);
591 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
592 rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
593 mutex_unlock(&rtp->tasks_gp_mutex);
596 // RCU-tasks kthread that detects grace periods and invokes callbacks.
597 static int __noreturn rcu_tasks_kthread(void *arg)
600 struct rcu_tasks *rtp = arg;
602 for_each_possible_cpu(cpu) {
603 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
605 timer_setup(&rtpcp->lazy_timer, call_rcu_tasks_generic_timer, 0);
606 rtpcp->urgent_gp = 1;
609 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
610 housekeeping_affine(current, HK_TYPE_RCU);
611 smp_store_release(&rtp->kthread_ptr, current); // Let GPs start!
614 * Each pass through the following loop makes one check for
615 * newly arrived callbacks, and, if there are some, waits for
616 * one RCU-tasks grace period and then invokes the callbacks.
617 * This loop is terminated by the system going down. ;-)
620 // Wait for one grace period and invoke any callbacks
622 rcu_tasks_one_gp(rtp, false);
624 // Paranoid sleep to keep this from entering a tight loop.
625 schedule_timeout_idle(rtp->gp_sleep);
629 // Wait for a grace period for the specified flavor of Tasks RCU.
630 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
632 /* Complain if the scheduler has not started. */
633 if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
634 "synchronize_%s() called too soon", rtp->name))
637 // If the grace-period kthread is running, use it.
638 if (READ_ONCE(rtp->kthread_ptr)) {
639 wait_rcu_gp(rtp->call_func);
642 rcu_tasks_one_gp(rtp, true);
645 /* Spawn RCU-tasks grace-period kthread. */
646 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
648 struct task_struct *t;
650 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
651 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
653 smp_mb(); /* Ensure others see full kthread. */
656 #ifndef CONFIG_TINY_RCU
659 * Print any non-default Tasks RCU settings.
661 static void __init rcu_tasks_bootup_oddness(void)
663 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
666 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
667 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
668 rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
669 if (rtsimc != rcu_task_stall_info_mult) {
670 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
671 rcu_task_stall_info_mult = rtsimc;
673 #endif /* #ifdef CONFIG_TASKS_RCU */
674 #ifdef CONFIG_TASKS_RCU
675 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
676 #endif /* #ifdef CONFIG_TASKS_RCU */
677 #ifdef CONFIG_TASKS_RUDE_RCU
678 pr_info("\tRude variant of Tasks RCU enabled.\n");
679 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
680 #ifdef CONFIG_TASKS_TRACE_RCU
681 pr_info("\tTracing variant of Tasks RCU enabled.\n");
682 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
685 #endif /* #ifndef CONFIG_TINY_RCU */
687 #ifndef CONFIG_TINY_RCU
688 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
689 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
692 bool havecbs = false;
693 bool haveurgent = false;
694 bool haveurgentcbs = false;
696 for_each_possible_cpu(cpu) {
697 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
699 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)))
701 if (data_race(rtpcp->urgent_gp))
703 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)) && data_race(rtpcp->urgent_gp))
704 haveurgentcbs = true;
705 if (havecbs && haveurgent && haveurgentcbs)
708 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c%c%c l:%lu %s\n",
710 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
711 jiffies - data_race(rtp->gp_jiffies),
712 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
713 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
714 ".k"[!!data_race(rtp->kthread_ptr)],
721 #endif // #ifndef CONFIG_TINY_RCU
723 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
725 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
727 ////////////////////////////////////////////////////////////////////////
729 // Shared code between task-list-scanning variants of Tasks RCU.
731 /* Wait for one RCU-tasks grace period. */
732 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
734 struct task_struct *g;
738 unsigned long lastinfo;
739 unsigned long lastreport;
740 bool reported = false;
742 struct task_struct *t;
744 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
745 rtp->pregp_func(&holdouts);
748 * There were callbacks, so we need to wait for an RCU-tasks
749 * grace period. Start off by scanning the task list for tasks
750 * that are not already voluntarily blocked. Mark these tasks
751 * and make a list of them in holdouts.
753 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
754 if (rtp->pertask_func) {
756 for_each_process_thread(g, t)
757 rtp->pertask_func(t, &holdouts);
761 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
762 rtp->postscan_func(&holdouts);
765 * Each pass through the following loop scans the list of holdout
766 * tasks, removing any that are no longer holdouts. When the list
767 * is empty, we are done.
769 lastreport = jiffies;
770 lastinfo = lastreport;
771 rtsi = READ_ONCE(rcu_task_stall_info);
773 // Start off with initial wait and slowly back off to 1 HZ wait.
774 fract = rtp->init_fract;
776 while (!list_empty(&holdouts)) {
782 // Slowly back off waiting for holdouts
783 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
784 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
785 schedule_timeout_idle(fract);
787 exp = jiffies_to_nsecs(fract);
788 __set_current_state(TASK_IDLE);
789 schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
795 rtst = READ_ONCE(rcu_task_stall_timeout);
796 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
798 lastreport = jiffies;
802 WARN_ON(signal_pending(current));
803 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
804 rtp->holdouts_func(&holdouts, needreport, &firstreport);
806 // Print pre-stall informational messages if needed.
808 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
810 rtsi = rtsi * rcu_task_stall_info_mult;
811 pr_info("%s: %s grace period number %lu (since boot) is %lu jiffies old.\n",
812 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
816 set_tasks_gp_state(rtp, RTGS_POST_GP);
817 rtp->postgp_func(rtp);
820 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
822 #ifdef CONFIG_TASKS_RCU
824 ////////////////////////////////////////////////////////////////////////
826 // Simple variant of RCU whose quiescent states are voluntary context
827 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
828 // As such, grace periods can take one good long time. There are no
829 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
830 // because this implementation is intended to get the system into a safe
831 // state for some of the manipulations involved in tracing and the like.
832 // Finally, this implementation does not support high call_rcu_tasks()
833 // rates from multiple CPUs. If this is required, per-CPU callback lists
836 // The implementation uses rcu_tasks_wait_gp(), which relies on function
837 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
838 // function sets these function pointers up so that rcu_tasks_wait_gp()
839 // invokes these functions in this order:
841 // rcu_tasks_pregp_step():
842 // Invokes synchronize_rcu() in order to wait for all in-flight
843 // t->on_rq and t->nvcsw transitions to complete. This works because
844 // all such transitions are carried out with interrupts disabled.
845 // rcu_tasks_pertask(), invoked on every non-idle task:
846 // For every runnable non-idle task other than the current one, use
847 // get_task_struct() to pin down that task, snapshot that task's
848 // number of voluntary context switches, and add that task to the
850 // rcu_tasks_postscan():
851 // Invoke synchronize_srcu() to ensure that all tasks that were
852 // in the process of exiting (and which thus might not know to
853 // synchronize with this RCU Tasks grace period) have completed
855 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
856 // Scans the holdout list, attempting to identify a quiescent state
857 // for each task on the list. If there is a quiescent state, the
858 // corresponding task is removed from the holdout list.
859 // rcu_tasks_postgp():
860 // Invokes synchronize_rcu() in order to ensure that all prior
861 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
862 // to have happened before the end of this RCU Tasks grace period.
863 // Again, this works because all such transitions are carried out
864 // with interrupts disabled.
866 // For each exiting task, the exit_tasks_rcu_start() and
867 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
868 // read-side critical sections waited for by rcu_tasks_postscan().
870 // Pre-grace-period update-side code is ordered before the grace
871 // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code
872 // is ordered before the grace period via synchronize_rcu() call in
873 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
876 /* Pre-grace-period preparation. */
877 static void rcu_tasks_pregp_step(struct list_head *hop)
880 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
881 * to complete. Invoking synchronize_rcu() suffices because all
882 * these transitions occur with interrupts disabled. Without this
883 * synchronize_rcu(), a read-side critical section that started
884 * before the grace period might be incorrectly seen as having
885 * started after the grace period.
887 * This synchronize_rcu() also dispenses with the need for a
888 * memory barrier on the first store to t->rcu_tasks_holdout,
889 * as it forces the store to happen after the beginning of the
895 /* Check for quiescent states since the pregp's synchronize_rcu() */
896 static bool rcu_tasks_is_holdout(struct task_struct *t)
900 /* Has the task been seen voluntarily sleeping? */
901 if (!READ_ONCE(t->on_rq))
905 * Idle tasks (or idle injection) within the idle loop are RCU-tasks
906 * quiescent states. But CPU boot code performed by the idle task
907 * isn't a quiescent state.
914 /* Idle tasks on offline CPUs are RCU-tasks quiescent states. */
915 if (t == idle_task(cpu) && !rcu_cpu_online(cpu))
921 /* Per-task initial processing. */
922 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
924 if (t != current && rcu_tasks_is_holdout(t)) {
926 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
927 WRITE_ONCE(t->rcu_tasks_holdout, true);
928 list_add(&t->rcu_tasks_holdout_list, hop);
932 /* Processing between scanning taskslist and draining the holdout list. */
933 static void rcu_tasks_postscan(struct list_head *hop)
935 int rtsi = READ_ONCE(rcu_task_stall_info);
937 if (!IS_ENABLED(CONFIG_TINY_RCU)) {
938 tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
939 add_timer(&tasks_rcu_exit_srcu_stall_timer);
943 * Exiting tasks may escape the tasklist scan. Those are vulnerable
944 * until their final schedule() with TASK_DEAD state. To cope with
945 * this, divide the fragile exit path part in two intersecting
946 * read side critical sections:
948 * 1) An _SRCU_ read side starting before calling exit_notify(),
949 * which may remove the task from the tasklist, and ending after
950 * the final preempt_disable() call in do_exit().
952 * 2) An _RCU_ read side starting with the final preempt_disable()
953 * call in do_exit() and ending with the final call to schedule()
954 * with TASK_DEAD state.
956 * This handles the part 1). And postgp will handle part 2) with a
957 * call to synchronize_rcu().
959 synchronize_srcu(&tasks_rcu_exit_srcu);
961 if (!IS_ENABLED(CONFIG_TINY_RCU))
962 del_timer_sync(&tasks_rcu_exit_srcu_stall_timer);
965 /* See if tasks are still holding out, complain if so. */
966 static void check_holdout_task(struct task_struct *t,
967 bool needreport, bool *firstreport)
971 if (!READ_ONCE(t->rcu_tasks_holdout) ||
972 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
973 !rcu_tasks_is_holdout(t) ||
974 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
975 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
976 WRITE_ONCE(t->rcu_tasks_holdout, false);
977 list_del_init(&t->rcu_tasks_holdout_list);
981 rcu_request_urgent_qs_task(t);
985 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
986 *firstreport = false;
989 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
990 t, ".I"[is_idle_task(t)],
991 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
992 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
993 t->rcu_tasks_idle_cpu, cpu);
997 /* Scan the holdout lists for tasks no longer holding out. */
998 static void check_all_holdout_tasks(struct list_head *hop,
999 bool needreport, bool *firstreport)
1001 struct task_struct *t, *t1;
1003 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
1004 check_holdout_task(t, needreport, firstreport);
1009 /* Finish off the Tasks-RCU grace period. */
1010 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
1013 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
1014 * memory barriers prior to them in the schedule() path, memory
1015 * reordering on other CPUs could cause their RCU-tasks read-side
1016 * critical sections to extend past the end of the grace period.
1017 * However, because these ->nvcsw updates are carried out with
1018 * interrupts disabled, we can use synchronize_rcu() to force the
1019 * needed ordering on all such CPUs.
1021 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
1022 * accesses to be within the grace period, avoiding the need for
1023 * memory barriers for ->rcu_tasks_holdout accesses.
1025 * In addition, this synchronize_rcu() waits for exiting tasks
1026 * to complete their final preempt_disable() region of execution,
1027 * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
1028 * enforcing the whole region before tasklist removal until
1029 * the final schedule() with TASK_DEAD state to be an RCU TASKS
1030 * read side critical section.
1035 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
1036 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
1038 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused)
1040 #ifndef CONFIG_TINY_RCU
1043 rtsi = READ_ONCE(rcu_task_stall_info);
1044 pr_info("%s: %s grace period number %lu (since boot) gp_state: %s is %lu jiffies old.\n",
1045 __func__, rcu_tasks.kname, rcu_tasks.tasks_gp_seq,
1046 tasks_gp_state_getname(&rcu_tasks), jiffies - rcu_tasks.gp_jiffies);
1047 pr_info("Please check any exiting tasks stuck between calls to exit_tasks_rcu_start() and exit_tasks_rcu_finish()\n");
1048 tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
1049 add_timer(&tasks_rcu_exit_srcu_stall_timer);
1050 #endif // #ifndef CONFIG_TINY_RCU
1054 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
1055 * @rhp: structure to be used for queueing the RCU updates.
1056 * @func: actual callback function to be invoked after the grace period
1058 * The callback function will be invoked some time after a full grace
1059 * period elapses, in other words after all currently executing RCU
1060 * read-side critical sections have completed. call_rcu_tasks() assumes
1061 * that the read-side critical sections end at a voluntary context
1062 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
1063 * or transition to usermode execution. As such, there are no read-side
1064 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1065 * this primitive is intended to determine that all tasks have passed
1066 * through a safe state, not so much for data-structure synchronization.
1068 * See the description of call_rcu() for more detailed information on
1069 * memory ordering guarantees.
1071 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
1073 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
1075 EXPORT_SYMBOL_GPL(call_rcu_tasks);
1078 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
1080 * Control will return to the caller some time after a full rcu-tasks
1081 * grace period has elapsed, in other words after all currently
1082 * executing rcu-tasks read-side critical sections have elapsed. These
1083 * read-side critical sections are delimited by calls to schedule(),
1084 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
1085 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
1087 * This is a very specialized primitive, intended only for a few uses in
1088 * tracing and other situations requiring manipulation of function
1089 * preambles and profiling hooks. The synchronize_rcu_tasks() function
1090 * is not (yet) intended for heavy use from multiple CPUs.
1092 * See the description of synchronize_rcu() for more detailed information
1093 * on memory ordering guarantees.
1095 void synchronize_rcu_tasks(void)
1097 synchronize_rcu_tasks_generic(&rcu_tasks);
1099 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
1102 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
1104 * Although the current implementation is guaranteed to wait, it is not
1105 * obligated to, for example, if there are no pending callbacks.
1107 void rcu_barrier_tasks(void)
1109 rcu_barrier_tasks_generic(&rcu_tasks);
1111 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
1113 int rcu_tasks_lazy_ms = -1;
1114 module_param(rcu_tasks_lazy_ms, int, 0444);
1116 static int __init rcu_spawn_tasks_kthread(void)
1118 cblist_init_generic(&rcu_tasks);
1119 rcu_tasks.gp_sleep = HZ / 10;
1120 rcu_tasks.init_fract = HZ / 10;
1121 if (rcu_tasks_lazy_ms >= 0)
1122 rcu_tasks.lazy_jiffies = msecs_to_jiffies(rcu_tasks_lazy_ms);
1123 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
1124 rcu_tasks.pertask_func = rcu_tasks_pertask;
1125 rcu_tasks.postscan_func = rcu_tasks_postscan;
1126 rcu_tasks.holdouts_func = check_all_holdout_tasks;
1127 rcu_tasks.postgp_func = rcu_tasks_postgp;
1128 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1132 #if !defined(CONFIG_TINY_RCU)
1133 void show_rcu_tasks_classic_gp_kthread(void)
1135 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1137 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1138 #endif // !defined(CONFIG_TINY_RCU)
1140 struct task_struct *get_rcu_tasks_gp_kthread(void)
1142 return rcu_tasks.kthread_ptr;
1144 EXPORT_SYMBOL_GPL(get_rcu_tasks_gp_kthread);
1147 * Contribute to protect against tasklist scan blind spot while the
1148 * task is exiting and may be removed from the tasklist. See
1149 * corresponding synchronize_srcu() for further details.
1151 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
1153 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
1157 * Contribute to protect against tasklist scan blind spot while the
1158 * task is exiting and may be removed from the tasklist. See
1159 * corresponding synchronize_srcu() for further details.
1161 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
1163 struct task_struct *t = current;
1165 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1169 * Contribute to protect against tasklist scan blind spot while the
1170 * task is exiting and may be removed from the tasklist. See
1171 * corresponding synchronize_srcu() for further details.
1173 void exit_tasks_rcu_finish(void)
1175 exit_tasks_rcu_stop();
1176 exit_tasks_rcu_finish_trace(current);
1179 #else /* #ifdef CONFIG_TASKS_RCU */
1180 void exit_tasks_rcu_start(void) { }
1181 void exit_tasks_rcu_stop(void) { }
1182 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1183 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1185 #ifdef CONFIG_TASKS_RUDE_RCU
1187 ////////////////////////////////////////////////////////////////////////
1189 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1190 // passing an empty function to schedule_on_each_cpu(). This approach
1191 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1192 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1193 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1194 // and induces otherwise unnecessary context switches on all online CPUs,
1195 // whether idle or not.
1197 // Callback handling is provided by the rcu_tasks_kthread() function.
1199 // Ordering is provided by the scheduler's context-switch code.
1201 // Empty function to allow workqueues to force a context switch.
1202 static void rcu_tasks_be_rude(struct work_struct *work)
1206 // Wait for one rude RCU-tasks grace period.
1207 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1209 rtp->n_ipis += cpumask_weight(cpu_online_mask);
1210 schedule_on_each_cpu(rcu_tasks_be_rude);
1213 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1214 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1218 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1219 * @rhp: structure to be used for queueing the RCU updates.
1220 * @func: actual callback function to be invoked after the grace period
1222 * The callback function will be invoked some time after a full grace
1223 * period elapses, in other words after all currently executing RCU
1224 * read-side critical sections have completed. call_rcu_tasks_rude()
1225 * assumes that the read-side critical sections end at context switch,
1226 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1227 * usermode execution is schedulable). As such, there are no read-side
1228 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1229 * this primitive is intended to determine that all tasks have passed
1230 * through a safe state, not so much for data-structure synchronization.
1232 * See the description of call_rcu() for more detailed information on
1233 * memory ordering guarantees.
1235 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1237 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1239 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1242 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1244 * Control will return to the caller some time after a rude rcu-tasks
1245 * grace period has elapsed, in other words after all currently
1246 * executing rcu-tasks read-side critical sections have elapsed. These
1247 * read-side critical sections are delimited by calls to schedule(),
1248 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1249 * context), and (in theory, anyway) cond_resched().
1251 * This is a very specialized primitive, intended only for a few uses in
1252 * tracing and other situations requiring manipulation of function preambles
1253 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
1254 * (yet) intended for heavy use from multiple CPUs.
1256 * See the description of synchronize_rcu() for more detailed information
1257 * on memory ordering guarantees.
1259 void synchronize_rcu_tasks_rude(void)
1261 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1263 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1266 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1268 * Although the current implementation is guaranteed to wait, it is not
1269 * obligated to, for example, if there are no pending callbacks.
1271 void rcu_barrier_tasks_rude(void)
1273 rcu_barrier_tasks_generic(&rcu_tasks_rude);
1275 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1277 int rcu_tasks_rude_lazy_ms = -1;
1278 module_param(rcu_tasks_rude_lazy_ms, int, 0444);
1280 static int __init rcu_spawn_tasks_rude_kthread(void)
1282 cblist_init_generic(&rcu_tasks_rude);
1283 rcu_tasks_rude.gp_sleep = HZ / 10;
1284 if (rcu_tasks_rude_lazy_ms >= 0)
1285 rcu_tasks_rude.lazy_jiffies = msecs_to_jiffies(rcu_tasks_rude_lazy_ms);
1286 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1290 #if !defined(CONFIG_TINY_RCU)
1291 void show_rcu_tasks_rude_gp_kthread(void)
1293 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1295 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1296 #endif // !defined(CONFIG_TINY_RCU)
1298 struct task_struct *get_rcu_tasks_rude_gp_kthread(void)
1300 return rcu_tasks_rude.kthread_ptr;
1302 EXPORT_SYMBOL_GPL(get_rcu_tasks_rude_gp_kthread);
1304 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1306 ////////////////////////////////////////////////////////////////////////
1308 // Tracing variant of Tasks RCU. This variant is designed to be used
1309 // to protect tracing hooks, including those of BPF. This variant
1312 // 1. Has explicit read-side markers to allow finite grace periods
1313 // in the face of in-kernel loops for PREEMPT=n builds.
1315 // 2. Protects code in the idle loop, exception entry/exit, and
1316 // CPU-hotplug code paths, similar to the capabilities of SRCU.
1318 // 3. Avoids expensive read-side instructions, having overhead similar
1319 // to that of Preemptible RCU.
1321 // There are of course downsides. For example, the grace-period code
1322 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1323 // in nohz_full userspace. If needed, these downsides can be at least
1324 // partially remedied.
1326 // Perhaps most important, this variant of RCU does not affect the vanilla
1327 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
1328 // readers can operate from idle, offline, and exception entry/exit in no
1329 // way allows rcu_preempt and rcu_sched readers to also do so.
1331 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1332 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
1333 // function sets these function pointers up so that rcu_tasks_wait_gp()
1334 // invokes these functions in this order:
1336 // rcu_tasks_trace_pregp_step():
1337 // Disables CPU hotplug, adds all currently executing tasks to the
1338 // holdout list, then checks the state of all tasks that blocked
1339 // or were preempted within their current RCU Tasks Trace read-side
1340 // critical section, adding them to the holdout list if appropriate.
1341 // Finally, this function re-enables CPU hotplug.
1342 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1343 // rcu_tasks_trace_postscan():
1344 // Invokes synchronize_rcu() to wait for late-stage exiting tasks
1345 // to finish exiting.
1346 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1347 // Scans the holdout list, attempting to identify a quiescent state
1348 // for each task on the list. If there is a quiescent state, the
1349 // corresponding task is removed from the holdout list. Once this
1350 // list is empty, the grace period has completed.
1351 // rcu_tasks_trace_postgp():
1352 // Provides the needed full memory barrier and does debug checks.
1354 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1356 // Pre-grace-period update-side code is ordered before the grace period
1357 // via the ->cbs_lock and barriers in rcu_tasks_kthread(). Pre-grace-period
1358 // read-side code is ordered before the grace period by atomic operations
1359 // on .b.need_qs flag of each task involved in this process, or by scheduler
1360 // context-switch ordering (for locked-down non-running readers).
1362 // The lockdep state must be outside of #ifdef to be useful.
1363 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1364 static struct lock_class_key rcu_lock_trace_key;
1365 struct lockdep_map rcu_trace_lock_map =
1366 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1367 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1368 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1370 #ifdef CONFIG_TASKS_TRACE_RCU
1372 // Record outstanding IPIs to each CPU. No point in sending two...
1373 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1375 // The number of detections of task quiescent state relying on
1376 // heavyweight readers executing explicit memory barriers.
1377 static unsigned long n_heavy_reader_attempts;
1378 static unsigned long n_heavy_reader_updates;
1379 static unsigned long n_heavy_reader_ofl_updates;
1380 static unsigned long n_trc_holdouts;
1382 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1383 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1386 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1387 static u8 rcu_ld_need_qs(struct task_struct *t)
1389 smp_mb(); // Enforce full grace-period ordering.
1390 return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1393 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1394 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1396 smp_store_release(&t->trc_reader_special.b.need_qs, v);
1397 smp_mb(); // Enforce full grace-period ordering.
1401 * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1402 * the four-byte operand-size restriction of some platforms.
1403 * Returns the old value, which is often ignored.
1405 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1407 union rcu_special ret;
1408 union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1409 union rcu_special trs_new = trs_old;
1411 if (trs_old.b.need_qs != old)
1412 return trs_old.b.need_qs;
1413 trs_new.b.need_qs = new;
1414 ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1415 return ret.b.need_qs;
1417 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1420 * If we are the last reader, signal the grace-period kthread.
1421 * Also remove from the per-CPU list of blocked tasks.
1423 void rcu_read_unlock_trace_special(struct task_struct *t)
1425 unsigned long flags;
1426 struct rcu_tasks_percpu *rtpcp;
1427 union rcu_special trs;
1429 // Open-coded full-word version of rcu_ld_need_qs().
1430 smp_mb(); // Enforce full grace-period ordering.
1431 trs = smp_load_acquire(&t->trc_reader_special);
1433 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1434 smp_mb(); // Pairs with update-side barriers.
1435 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1436 if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1437 u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1438 TRC_NEED_QS_CHECKED);
1440 WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1442 if (trs.b.blocked) {
1443 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1444 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1445 list_del_init(&t->trc_blkd_node);
1446 WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1447 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1449 WRITE_ONCE(t->trc_reader_nesting, 0);
1451 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1453 /* Add a newly blocked reader task to its CPU's list. */
1454 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1456 unsigned long flags;
1457 struct rcu_tasks_percpu *rtpcp;
1459 local_irq_save(flags);
1460 rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1461 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1462 t->trc_blkd_cpu = smp_processor_id();
1463 if (!rtpcp->rtp_blkd_tasks.next)
1464 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1465 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1466 WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1467 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1469 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1471 /* Add a task to the holdout list, if it is not already on the list. */
1472 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1474 if (list_empty(&t->trc_holdout_list)) {
1476 list_add(&t->trc_holdout_list, bhp);
1481 /* Remove a task from the holdout list, if it is in fact present. */
1482 static void trc_del_holdout(struct task_struct *t)
1484 if (!list_empty(&t->trc_holdout_list)) {
1485 list_del_init(&t->trc_holdout_list);
1491 /* IPI handler to check task state. */
1492 static void trc_read_check_handler(void *t_in)
1495 struct task_struct *t = current;
1496 struct task_struct *texp = t_in;
1498 // If the task is no longer running on this CPU, leave.
1499 if (unlikely(texp != t))
1500 goto reset_ipi; // Already on holdout list, so will check later.
1502 // If the task is not in a read-side critical section, and
1503 // if this is the last reader, awaken the grace-period kthread.
1504 nesting = READ_ONCE(t->trc_reader_nesting);
1505 if (likely(!nesting)) {
1506 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1509 // If we are racing with an rcu_read_unlock_trace(), try again later.
1510 if (unlikely(nesting < 0))
1513 // Get here if the task is in a read-side critical section.
1514 // Set its state so that it will update state for the grace-period
1515 // kthread upon exit from that critical section.
1516 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1519 // Allow future IPIs to be sent on CPU and for task.
1520 // Also order this IPI handler against any later manipulations of
1521 // the intended task.
1522 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1523 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1526 /* Callback function for scheduler to check locked-down task. */
1527 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1529 struct list_head *bhp = bhp_in;
1530 int cpu = task_cpu(t);
1532 bool ofl = cpu_is_offline(cpu);
1534 if (task_curr(t) && !ofl) {
1535 // If no chance of heavyweight readers, do it the hard way.
1536 if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1539 // If heavyweight readers are enabled on the remote task,
1540 // we can inspect its state despite its currently running.
1541 // However, we cannot safely change its state.
1542 n_heavy_reader_attempts++;
1543 // Check for "running" idle tasks on offline CPUs.
1544 if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1545 return -EINVAL; // No quiescent state, do it the hard way.
1546 n_heavy_reader_updates++;
1549 // The task is not running, so C-language access is safe.
1550 nesting = t->trc_reader_nesting;
1551 WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1552 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1553 n_heavy_reader_ofl_updates++;
1556 // If not exiting a read-side critical section, mark as checked
1557 // so that the grace-period kthread will remove it from the
1560 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1561 return 0; // In QS, so done.
1564 return -EINVAL; // Reader transitioning, try again later.
1566 // The task is in a read-side critical section, so set up its
1567 // state so that it will update state upon exit from that critical
1569 if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1570 trc_add_holdout(t, bhp);
1574 /* Attempt to extract the state for the specified task. */
1575 static void trc_wait_for_one_reader(struct task_struct *t,
1576 struct list_head *bhp)
1580 // If a previous IPI is still in flight, let it complete.
1581 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1584 // The current task had better be in a quiescent state.
1586 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1587 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1591 // Attempt to nail down the task for inspection.
1593 if (!task_call_func(t, trc_inspect_reader, bhp)) {
1599 // If this task is not yet on the holdout list, then we are in
1600 // an RCU read-side critical section. Otherwise, the invocation of
1601 // trc_add_holdout() that added it to the list did the necessary
1602 // get_task_struct(). Either way, the task cannot be freed out
1603 // from under this code.
1605 // If currently running, send an IPI, either way, add to list.
1606 trc_add_holdout(t, bhp);
1608 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1609 // The task is currently running, so try IPIing it.
1612 // If there is already an IPI outstanding, let it happen.
1613 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1616 per_cpu(trc_ipi_to_cpu, cpu) = true;
1617 t->trc_ipi_to_cpu = cpu;
1618 rcu_tasks_trace.n_ipis++;
1619 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1620 // Just in case there is some other reason for
1621 // failure than the target CPU being offline.
1622 WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n",
1624 rcu_tasks_trace.n_ipis_fails++;
1625 per_cpu(trc_ipi_to_cpu, cpu) = false;
1626 t->trc_ipi_to_cpu = -1;
1632 * Initialize for first-round processing for the specified task.
1633 * Return false if task is NULL or already taken care of, true otherwise.
1635 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1637 // During early boot when there is only the one boot CPU, there
1638 // is no idle task for the other CPUs. Also, the grace-period
1639 // kthread is always in a quiescent state. In addition, just return
1640 // if this task is already on the list.
1641 if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1644 rcu_st_need_qs(t, 0);
1645 t->trc_ipi_to_cpu = -1;
1649 /* Do first-round processing for the specified task. */
1650 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1652 if (rcu_tasks_trace_pertask_prep(t, true))
1653 trc_wait_for_one_reader(t, hop);
1656 /* Initialize for a new RCU-tasks-trace grace period. */
1657 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1659 LIST_HEAD(blkd_tasks);
1661 unsigned long flags;
1662 struct rcu_tasks_percpu *rtpcp;
1663 struct task_struct *t;
1665 // There shouldn't be any old IPIs, but...
1666 for_each_possible_cpu(cpu)
1667 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1669 // Disable CPU hotplug across the CPU scan for the benefit of
1670 // any IPIs that might be needed. This also waits for all readers
1671 // in CPU-hotplug code paths.
1674 // These rcu_tasks_trace_pertask_prep() calls are serialized to
1675 // allow safe access to the hop list.
1676 for_each_online_cpu(cpu) {
1678 t = cpu_curr_snapshot(cpu);
1679 if (rcu_tasks_trace_pertask_prep(t, true))
1680 trc_add_holdout(t, hop);
1682 cond_resched_tasks_rcu_qs();
1685 // Only after all running tasks have been accounted for is it
1686 // safe to take care of the tasks that have blocked within their
1687 // current RCU tasks trace read-side critical section.
1688 for_each_possible_cpu(cpu) {
1689 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1690 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1691 list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1692 while (!list_empty(&blkd_tasks)) {
1694 t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1695 list_del_init(&t->trc_blkd_node);
1696 list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1697 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1698 rcu_tasks_trace_pertask(t, hop);
1700 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1702 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1703 cond_resched_tasks_rcu_qs();
1706 // Re-enable CPU hotplug now that the holdout list is populated.
1711 * Do intermediate processing between task and holdout scans.
1713 static void rcu_tasks_trace_postscan(struct list_head *hop)
1715 // Wait for late-stage exiting tasks to finish exiting.
1716 // These might have passed the call to exit_tasks_rcu_finish().
1718 // If you remove the following line, update rcu_trace_implies_rcu_gp()!!!
1720 // Any tasks that exit after this point will set
1721 // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1724 /* Communicate task state back to the RCU tasks trace stall warning request. */
1725 struct trc_stall_chk_rdr {
1731 static int trc_check_slow_task(struct task_struct *t, void *arg)
1733 struct trc_stall_chk_rdr *trc_rdrp = arg;
1735 if (task_curr(t) && cpu_online(task_cpu(t)))
1736 return false; // It is running, so decline to inspect it.
1737 trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1738 trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1739 trc_rdrp->needqs = rcu_ld_need_qs(t);
1743 /* Show the state of a task stalling the current RCU tasks trace GP. */
1744 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1747 struct trc_stall_chk_rdr trc_rdr;
1748 bool is_idle_tsk = is_idle_task(t);
1751 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1752 *firstreport = false;
1755 if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1756 pr_alert("P%d: %c%c\n",
1758 ".I"[t->trc_ipi_to_cpu >= 0],
1761 pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1763 ".I"[trc_rdr.ipi_to_cpu >= 0],
1765 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1766 ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1768 " !CN"[trc_rdr.needqs & 0x3],
1769 " ?"[trc_rdr.needqs > 0x3],
1770 cpu, cpu_online(cpu) ? "" : "(offline)");
1774 /* List stalled IPIs for RCU tasks trace. */
1775 static void show_stalled_ipi_trace(void)
1779 for_each_possible_cpu(cpu)
1780 if (per_cpu(trc_ipi_to_cpu, cpu))
1781 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1784 /* Do one scan of the holdout list. */
1785 static void check_all_holdout_tasks_trace(struct list_head *hop,
1786 bool needreport, bool *firstreport)
1788 struct task_struct *g, *t;
1790 // Disable CPU hotplug across the holdout list scan for IPIs.
1793 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1794 // If safe and needed, try to check the current task.
1795 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1796 !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1797 trc_wait_for_one_reader(t, hop);
1799 // If check succeeded, remove this task from the list.
1800 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1801 rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1803 else if (needreport)
1804 show_stalled_task_trace(t, firstreport);
1805 cond_resched_tasks_rcu_qs();
1808 // Re-enable CPU hotplug now that the holdout list scan has completed.
1813 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1814 show_stalled_ipi_trace();
1818 static void rcu_tasks_trace_empty_fn(void *unused)
1822 /* Wait for grace period to complete and provide ordering. */
1823 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1827 // Wait for any lingering IPI handlers to complete. Note that
1828 // if a CPU has gone offline or transitioned to userspace in the
1829 // meantime, all IPI handlers should have been drained beforehand.
1830 // Yes, this assumes that CPUs process IPIs in order. If that ever
1831 // changes, there will need to be a recheck and/or timed wait.
1832 for_each_online_cpu(cpu)
1833 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1834 smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1836 smp_mb(); // Caller's code must be ordered after wakeup.
1837 // Pairs with pretty much every ordering primitive.
1840 /* Report any needed quiescent state for this exiting task. */
1841 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1843 union rcu_special trs = READ_ONCE(t->trc_reader_special);
1845 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1846 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1847 if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1848 rcu_read_unlock_trace_special(t);
1850 WRITE_ONCE(t->trc_reader_nesting, 0);
1854 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1855 * @rhp: structure to be used for queueing the RCU updates.
1856 * @func: actual callback function to be invoked after the grace period
1858 * The callback function will be invoked some time after a trace rcu-tasks
1859 * grace period elapses, in other words after all currently executing
1860 * trace rcu-tasks read-side critical sections have completed. These
1861 * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1862 * and rcu_read_unlock_trace().
1864 * See the description of call_rcu() for more detailed information on
1865 * memory ordering guarantees.
1867 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1869 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1871 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1874 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1876 * Control will return to the caller some time after a trace rcu-tasks
1877 * grace period has elapsed, in other words after all currently executing
1878 * trace rcu-tasks read-side critical sections have elapsed. These read-side
1879 * critical sections are delimited by calls to rcu_read_lock_trace()
1880 * and rcu_read_unlock_trace().
1882 * This is a very specialized primitive, intended only for a few uses in
1883 * tracing and other situations requiring manipulation of function preambles
1884 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1885 * (yet) intended for heavy use from multiple CPUs.
1887 * See the description of synchronize_rcu() for more detailed information
1888 * on memory ordering guarantees.
1890 void synchronize_rcu_tasks_trace(void)
1892 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1893 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1895 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1898 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1900 * Although the current implementation is guaranteed to wait, it is not
1901 * obligated to, for example, if there are no pending callbacks.
1903 void rcu_barrier_tasks_trace(void)
1905 rcu_barrier_tasks_generic(&rcu_tasks_trace);
1907 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1909 int rcu_tasks_trace_lazy_ms = -1;
1910 module_param(rcu_tasks_trace_lazy_ms, int, 0444);
1912 static int __init rcu_spawn_tasks_trace_kthread(void)
1914 cblist_init_generic(&rcu_tasks_trace);
1915 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1916 rcu_tasks_trace.gp_sleep = HZ / 10;
1917 rcu_tasks_trace.init_fract = HZ / 10;
1919 rcu_tasks_trace.gp_sleep = HZ / 200;
1920 if (rcu_tasks_trace.gp_sleep <= 0)
1921 rcu_tasks_trace.gp_sleep = 1;
1922 rcu_tasks_trace.init_fract = HZ / 200;
1923 if (rcu_tasks_trace.init_fract <= 0)
1924 rcu_tasks_trace.init_fract = 1;
1926 if (rcu_tasks_trace_lazy_ms >= 0)
1927 rcu_tasks_trace.lazy_jiffies = msecs_to_jiffies(rcu_tasks_trace_lazy_ms);
1928 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1929 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1930 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1931 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1932 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1936 #if !defined(CONFIG_TINY_RCU)
1937 void show_rcu_tasks_trace_gp_kthread(void)
1941 sprintf(buf, "N%lu h:%lu/%lu/%lu",
1942 data_race(n_trc_holdouts),
1943 data_race(n_heavy_reader_ofl_updates),
1944 data_race(n_heavy_reader_updates),
1945 data_race(n_heavy_reader_attempts));
1946 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1948 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1949 #endif // !defined(CONFIG_TINY_RCU)
1951 struct task_struct *get_rcu_tasks_trace_gp_kthread(void)
1953 return rcu_tasks_trace.kthread_ptr;
1955 EXPORT_SYMBOL_GPL(get_rcu_tasks_trace_gp_kthread);
1957 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1958 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1959 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1961 #ifndef CONFIG_TINY_RCU
1962 void show_rcu_tasks_gp_kthreads(void)
1964 show_rcu_tasks_classic_gp_kthread();
1965 show_rcu_tasks_rude_gp_kthread();
1966 show_rcu_tasks_trace_gp_kthread();
1968 #endif /* #ifndef CONFIG_TINY_RCU */
1970 #ifdef CONFIG_PROVE_RCU
1971 struct rcu_tasks_test_desc {
1975 unsigned long runstart;
1978 static struct rcu_tasks_test_desc tests[] = {
1980 .name = "call_rcu_tasks()",
1981 /* If not defined, the test is skipped. */
1982 .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1985 .name = "call_rcu_tasks_rude()",
1986 /* If not defined, the test is skipped. */
1987 .notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1990 .name = "call_rcu_tasks_trace()",
1991 /* If not defined, the test is skipped. */
1992 .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1996 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1998 struct rcu_tasks_test_desc *rttd =
1999 container_of(rhp, struct rcu_tasks_test_desc, rh);
2001 pr_info("Callback from %s invoked.\n", rttd->name);
2003 rttd->notrun = false;
2006 static void rcu_tasks_initiate_self_tests(void)
2008 pr_info("Running RCU-tasks wait API self tests\n");
2009 #ifdef CONFIG_TASKS_RCU
2010 tests[0].runstart = jiffies;
2011 synchronize_rcu_tasks();
2012 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
2015 #ifdef CONFIG_TASKS_RUDE_RCU
2016 tests[1].runstart = jiffies;
2017 synchronize_rcu_tasks_rude();
2018 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
2021 #ifdef CONFIG_TASKS_TRACE_RCU
2022 tests[2].runstart = jiffies;
2023 synchronize_rcu_tasks_trace();
2024 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
2029 * Return: 0 - test passed
2030 * 1 - test failed, but have not timed out yet
2031 * -1 - test failed and timed out
2033 static int rcu_tasks_verify_self_tests(void)
2037 unsigned long bst = rcu_task_stall_timeout;
2039 if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
2040 bst = RCU_TASK_BOOT_STALL_TIMEOUT;
2041 for (i = 0; i < ARRAY_SIZE(tests); i++) {
2042 while (tests[i].notrun) { // still hanging.
2043 if (time_after(jiffies, tests[i].runstart + bst)) {
2044 pr_err("%s has failed boot-time tests.\n", tests[i].name);
2058 * Repeat the rcu_tasks_verify_self_tests() call once every second until the
2059 * test passes or has timed out.
2061 static struct delayed_work rcu_tasks_verify_work;
2062 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
2064 int ret = rcu_tasks_verify_self_tests();
2069 /* Test fails but not timed out yet, reschedule another check */
2070 schedule_delayed_work(&rcu_tasks_verify_work, HZ);
2073 static int rcu_tasks_verify_schedule_work(void)
2075 INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
2076 rcu_tasks_verify_work_fn(NULL);
2079 late_initcall(rcu_tasks_verify_schedule_work);
2080 #else /* #ifdef CONFIG_PROVE_RCU */
2081 static void rcu_tasks_initiate_self_tests(void) { }
2082 #endif /* #else #ifdef CONFIG_PROVE_RCU */
2084 void __init rcu_init_tasks_generic(void)
2086 #ifdef CONFIG_TASKS_RCU
2087 rcu_spawn_tasks_kthread();
2090 #ifdef CONFIG_TASKS_RUDE_RCU
2091 rcu_spawn_tasks_rude_kthread();
2094 #ifdef CONFIG_TASKS_TRACE_RCU
2095 rcu_spawn_tasks_trace_kthread();
2098 // Run the self-tests.
2099 rcu_tasks_initiate_self_tests();
2102 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
2103 static inline void rcu_tasks_bootup_oddness(void) {}
2104 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */