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)(void);
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 * @cpu: CPU number corresponding to this entry.
33 * @rtpp: Pointer to the rcu_tasks structure.
35 struct rcu_tasks_percpu {
36 struct rcu_segcblist cblist;
37 raw_spinlock_t __private lock;
38 unsigned long rtp_jiffies;
39 unsigned long rtp_n_lock_retries;
40 struct work_struct rtp_work;
41 struct irq_work rtp_irq_work;
42 struct rcu_head barrier_q_head;
44 struct rcu_tasks *rtpp;
48 * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
49 * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
50 * @cbs_gbl_lock: Lock protecting callback list.
51 * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
52 * @gp_func: This flavor's grace-period-wait function.
53 * @gp_state: Grace period's most recent state transition (debugging).
54 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
55 * @init_fract: Initial backoff sleep interval.
56 * @gp_jiffies: Time of last @gp_state transition.
57 * @gp_start: Most recent grace-period start in jiffies.
58 * @tasks_gp_seq: Number of grace periods completed since boot.
59 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
60 * @n_ipis_fails: Number of IPI-send failures.
61 * @pregp_func: This flavor's pre-grace-period function (optional).
62 * @pertask_func: This flavor's per-task scan function (optional).
63 * @postscan_func: This flavor's post-task scan function (optional).
64 * @holdouts_func: This flavor's holdout-list scan function (optional).
65 * @postgp_func: This flavor's post-grace-period function (optional).
66 * @call_func: This flavor's call_rcu()-equivalent function.
67 * @rtpcpu: This flavor's rcu_tasks_percpu structure.
68 * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
69 * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
70 * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
71 * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
72 * @barrier_q_mutex: Serialize barrier operations.
73 * @barrier_q_count: Number of queues being waited on.
74 * @barrier_q_completion: Barrier wait/wakeup mechanism.
75 * @barrier_q_seq: Sequence number for barrier operations.
76 * @name: This flavor's textual name.
77 * @kname: This flavor's kthread name.
80 struct rcuwait cbs_wait;
81 raw_spinlock_t cbs_gbl_lock;
85 unsigned long gp_jiffies;
86 unsigned long gp_start;
87 unsigned long tasks_gp_seq;
89 unsigned long n_ipis_fails;
90 struct task_struct *kthread_ptr;
91 rcu_tasks_gp_func_t gp_func;
92 pregp_func_t pregp_func;
93 pertask_func_t pertask_func;
94 postscan_func_t postscan_func;
95 holdouts_func_t holdouts_func;
96 postgp_func_t postgp_func;
97 call_rcu_func_t call_func;
98 struct rcu_tasks_percpu __percpu *rtpcpu;
99 int percpu_enqueue_shift;
100 int percpu_enqueue_lim;
101 int percpu_dequeue_lim;
102 unsigned long percpu_dequeue_gpseq;
103 struct mutex barrier_q_mutex;
104 atomic_t barrier_q_count;
105 struct completion barrier_q_completion;
106 unsigned long barrier_q_seq;
111 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
113 #define DEFINE_RCU_TASKS(rt_name, gp, call, n) \
114 static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = { \
115 .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock), \
116 .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup), \
118 static struct rcu_tasks rt_name = \
120 .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait), \
121 .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \
124 .rtpcpu = &rt_name ## __percpu, \
126 .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS), \
127 .percpu_enqueue_lim = 1, \
128 .percpu_dequeue_lim = 1, \
129 .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex), \
130 .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT, \
134 /* Track exiting tasks in order to allow them to be waited for. */
135 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
137 /* Avoid IPIing CPUs early in the grace period. */
138 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
139 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
140 module_param(rcu_task_ipi_delay, int, 0644);
142 /* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
143 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
144 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
145 module_param(rcu_task_stall_timeout, int, 0644);
146 #define RCU_TASK_STALL_INFO (HZ * 10)
147 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
148 module_param(rcu_task_stall_info, int, 0644);
149 static int rcu_task_stall_info_mult __read_mostly = 3;
150 module_param(rcu_task_stall_info_mult, int, 0444);
152 static int rcu_task_enqueue_lim __read_mostly = -1;
153 module_param(rcu_task_enqueue_lim, int, 0444);
155 static bool rcu_task_cb_adjust;
156 static int rcu_task_contend_lim __read_mostly = 100;
157 module_param(rcu_task_contend_lim, int, 0444);
158 static int rcu_task_collapse_lim __read_mostly = 10;
159 module_param(rcu_task_collapse_lim, int, 0444);
161 /* RCU tasks grace-period state for debugging. */
163 #define RTGS_WAIT_WAIT_CBS 1
164 #define RTGS_WAIT_GP 2
165 #define RTGS_PRE_WAIT_GP 3
166 #define RTGS_SCAN_TASKLIST 4
167 #define RTGS_POST_SCAN_TASKLIST 5
168 #define RTGS_WAIT_SCAN_HOLDOUTS 6
169 #define RTGS_SCAN_HOLDOUTS 7
170 #define RTGS_POST_GP 8
171 #define RTGS_WAIT_READERS 9
172 #define RTGS_INVOKE_CBS 10
173 #define RTGS_WAIT_CBS 11
174 #ifndef CONFIG_TINY_RCU
175 static const char * const rcu_tasks_gp_state_names[] = {
177 "RTGS_WAIT_WAIT_CBS",
180 "RTGS_SCAN_TASKLIST",
181 "RTGS_POST_SCAN_TASKLIST",
182 "RTGS_WAIT_SCAN_HOLDOUTS",
183 "RTGS_SCAN_HOLDOUTS",
189 #endif /* #ifndef CONFIG_TINY_RCU */
191 ////////////////////////////////////////////////////////////////////////
195 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
197 /* Record grace-period phase and time. */
198 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
200 rtp->gp_state = newstate;
201 rtp->gp_jiffies = jiffies;
204 #ifndef CONFIG_TINY_RCU
205 /* Return state name. */
206 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
208 int i = data_race(rtp->gp_state); // Let KCSAN detect update races
209 int j = READ_ONCE(i); // Prevent the compiler from reading twice
211 if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
213 return rcu_tasks_gp_state_names[j];
215 #endif /* #ifndef CONFIG_TINY_RCU */
217 // Initialize per-CPU callback lists for the specified flavor of
219 static void cblist_init_generic(struct rcu_tasks *rtp)
226 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
227 if (rcu_task_enqueue_lim < 0) {
228 rcu_task_enqueue_lim = 1;
229 rcu_task_cb_adjust = true;
230 pr_info("%s: Setting adjustable number of callback queues.\n", __func__);
231 } else if (rcu_task_enqueue_lim == 0) {
232 rcu_task_enqueue_lim = 1;
234 lim = rcu_task_enqueue_lim;
236 if (lim > nr_cpu_ids)
238 shift = ilog2(nr_cpu_ids / lim);
239 if (((nr_cpu_ids - 1) >> shift) >= lim)
241 WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
242 WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
243 smp_store_release(&rtp->percpu_enqueue_lim, lim);
244 for_each_possible_cpu(cpu) {
245 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
247 WARN_ON_ONCE(!rtpcp);
249 raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
250 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
251 if (rcu_segcblist_empty(&rtpcp->cblist))
252 rcu_segcblist_init(&rtpcp->cblist);
253 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
256 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
258 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
259 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));
262 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
263 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
265 struct rcu_tasks *rtp;
266 struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
269 rcuwait_wake_up(&rtp->cbs_wait);
272 // Enqueue a callback for the specified flavor of Tasks RCU.
273 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
274 struct rcu_tasks *rtp)
280 bool needadjust = false;
282 struct rcu_tasks_percpu *rtpcp;
286 local_irq_save(flags);
288 ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
289 chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
290 rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
291 if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
292 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
294 if (rtpcp->rtp_jiffies != j) {
295 rtpcp->rtp_jiffies = j;
296 rtpcp->rtp_n_lock_retries = 0;
298 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
299 READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
300 needadjust = true; // Defer adjustment to avoid deadlock.
302 if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
303 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
304 cblist_init_generic(rtp);
305 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
307 needwake = rcu_segcblist_empty(&rtpcp->cblist);
308 rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
309 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
310 if (unlikely(needadjust)) {
311 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
312 if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
313 WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
314 WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
315 smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
316 pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
318 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
321 /* We can't create the thread unless interrupts are enabled. */
322 if (needwake && READ_ONCE(rtp->kthread_ptr))
323 irq_work_queue(&rtpcp->rtp_irq_work);
326 // Wait for a grace period for the specified flavor of Tasks RCU.
327 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
329 /* Complain if the scheduler has not started. */
330 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
331 "synchronize_rcu_tasks called too soon");
333 /* Wait for the grace period. */
334 wait_rcu_gp(rtp->call_func);
337 // RCU callback function for rcu_barrier_tasks_generic().
338 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
340 struct rcu_tasks *rtp;
341 struct rcu_tasks_percpu *rtpcp;
343 rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
345 if (atomic_dec_and_test(&rtp->barrier_q_count))
346 complete(&rtp->barrier_q_completion);
349 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
350 // Operates in a manner similar to rcu_barrier().
351 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
355 struct rcu_tasks_percpu *rtpcp;
356 unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
358 mutex_lock(&rtp->barrier_q_mutex);
359 if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
361 mutex_unlock(&rtp->barrier_q_mutex);
364 rcu_seq_start(&rtp->barrier_q_seq);
365 init_completion(&rtp->barrier_q_completion);
366 atomic_set(&rtp->barrier_q_count, 2);
367 for_each_possible_cpu(cpu) {
368 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
370 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
371 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
372 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
373 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
374 atomic_inc(&rtp->barrier_q_count);
375 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
377 if (atomic_sub_and_test(2, &rtp->barrier_q_count))
378 complete(&rtp->barrier_q_completion);
379 wait_for_completion(&rtp->barrier_q_completion);
380 rcu_seq_end(&rtp->barrier_q_seq);
381 mutex_unlock(&rtp->barrier_q_mutex);
384 // Advance callbacks and indicate whether either a grace period or
385 // callback invocation is needed.
386 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
395 for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
396 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
398 /* Advance and accelerate any new callbacks. */
399 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
401 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
402 // Should we shrink down to a single callback queue?
403 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
409 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
410 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
411 if (rcu_segcblist_pend_cbs(&rtpcp->cblist))
413 if (!rcu_segcblist_empty(&rtpcp->cblist))
415 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
418 // Shrink down to a single callback queue if appropriate.
419 // This is done in two stages: (1) If there are no more than
420 // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
421 // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period,
422 // if there has not been an increase in callbacks, limit dequeuing
423 // to CPU 0. Note the matching RCU read-side critical section in
424 // call_rcu_tasks_generic().
425 if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
426 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
427 if (rtp->percpu_enqueue_lim > 1) {
428 WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
429 smp_store_release(&rtp->percpu_enqueue_lim, 1);
430 rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
431 pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
433 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
435 if (rcu_task_cb_adjust && !ncbsnz &&
436 poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq)) {
437 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
438 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
439 WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
440 pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
442 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
448 // Advance callbacks and invoke any that are ready.
449 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
455 struct rcu_head *rhp;
456 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
457 struct rcu_tasks_percpu *rtpcp_next;
460 cpunext = cpu * 2 + 1;
461 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
462 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
463 queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
465 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
466 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
467 queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
471 if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
473 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
474 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
475 rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
476 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
478 for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
484 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
485 rcu_segcblist_add_len(&rtpcp->cblist, -len);
486 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
487 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
490 // Workqueue flood to advance callbacks and invoke any that are ready.
491 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
493 struct rcu_tasks *rtp;
494 struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
497 rcu_tasks_invoke_cbs(rtp, rtpcp);
500 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
501 static int __noreturn rcu_tasks_kthread(void *arg)
504 struct rcu_tasks *rtp = arg;
506 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
507 housekeeping_affine(current, HK_TYPE_RCU);
508 WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
511 * Each pass through the following loop makes one check for
512 * newly arrived callbacks, and, if there are some, waits for
513 * one RCU-tasks grace period and then invokes the callbacks.
514 * This loop is terminated by the system going down. ;-)
517 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
519 /* If there were none, wait a bit and start over. */
520 rcuwait_wait_event(&rtp->cbs_wait,
521 (needgpcb = rcu_tasks_need_gpcb(rtp)),
524 if (needgpcb & 0x2) {
525 // Wait for one grace period.
526 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
527 rtp->gp_start = jiffies;
528 rcu_seq_start(&rtp->tasks_gp_seq);
530 rcu_seq_end(&rtp->tasks_gp_seq);
533 /* Invoke callbacks. */
534 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
535 rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
537 /* Paranoid sleep to keep this from entering a tight loop */
538 schedule_timeout_idle(rtp->gp_sleep);
542 /* Spawn RCU-tasks grace-period kthread. */
543 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
545 struct task_struct *t;
547 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
548 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
550 smp_mb(); /* Ensure others see full kthread. */
553 #ifndef CONFIG_TINY_RCU
556 * Print any non-default Tasks RCU settings.
558 static void __init rcu_tasks_bootup_oddness(void)
560 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
563 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
564 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
565 rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
566 if (rtsimc != rcu_task_stall_info_mult) {
567 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
568 rcu_task_stall_info_mult = rtsimc;
570 #endif /* #ifdef CONFIG_TASKS_RCU */
571 #ifdef CONFIG_TASKS_RCU
572 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
573 #endif /* #ifdef CONFIG_TASKS_RCU */
574 #ifdef CONFIG_TASKS_RUDE_RCU
575 pr_info("\tRude variant of Tasks RCU enabled.\n");
576 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
577 #ifdef CONFIG_TASKS_TRACE_RCU
578 pr_info("\tTracing variant of Tasks RCU enabled.\n");
579 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
582 #endif /* #ifndef CONFIG_TINY_RCU */
584 #ifndef CONFIG_TINY_RCU
585 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
586 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
589 bool havecbs = false;
591 for_each_possible_cpu(cpu) {
592 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
594 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist))) {
599 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
601 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
602 jiffies - data_race(rtp->gp_jiffies),
603 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
604 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
605 ".k"[!!data_race(rtp->kthread_ptr)],
609 #endif // #ifndef CONFIG_TINY_RCU
611 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
613 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
615 ////////////////////////////////////////////////////////////////////////
617 // Shared code between task-list-scanning variants of Tasks RCU.
619 /* Wait for one RCU-tasks grace period. */
620 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
622 struct task_struct *g;
626 unsigned long lastinfo;
627 unsigned long lastreport;
628 bool reported = false;
630 struct task_struct *t;
632 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
636 * There were callbacks, so we need to wait for an RCU-tasks
637 * grace period. Start off by scanning the task list for tasks
638 * that are not already voluntarily blocked. Mark these tasks
639 * and make a list of them in holdouts.
641 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
643 for_each_process_thread(g, t)
644 rtp->pertask_func(t, &holdouts);
647 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
648 rtp->postscan_func(&holdouts);
651 * Each pass through the following loop scans the list of holdout
652 * tasks, removing any that are no longer holdouts. When the list
653 * is empty, we are done.
655 lastreport = jiffies;
656 lastinfo = lastreport;
657 rtsi = READ_ONCE(rcu_task_stall_info);
659 // Start off with initial wait and slowly back off to 1 HZ wait.
660 fract = rtp->init_fract;
662 while (!list_empty(&holdouts)) {
668 // Slowly back off waiting for holdouts
669 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
670 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
671 schedule_timeout_idle(fract);
673 exp = jiffies_to_nsecs(fract);
674 __set_current_state(TASK_IDLE);
675 schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
681 rtst = READ_ONCE(rcu_task_stall_timeout);
682 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
684 lastreport = jiffies;
688 WARN_ON(signal_pending(current));
689 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
690 rtp->holdouts_func(&holdouts, needreport, &firstreport);
692 // Print pre-stall informational messages if needed.
694 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
696 rtsi = rtsi * rcu_task_stall_info_mult;
697 pr_info("%s: %s grace period %lu is %lu jiffies old.\n",
698 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
702 set_tasks_gp_state(rtp, RTGS_POST_GP);
703 rtp->postgp_func(rtp);
706 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
708 #ifdef CONFIG_TASKS_RCU
710 ////////////////////////////////////////////////////////////////////////
712 // Simple variant of RCU whose quiescent states are voluntary context
713 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
714 // As such, grace periods can take one good long time. There are no
715 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
716 // because this implementation is intended to get the system into a safe
717 // state for some of the manipulations involved in tracing and the like.
718 // Finally, this implementation does not support high call_rcu_tasks()
719 // rates from multiple CPUs. If this is required, per-CPU callback lists
722 // The implementation uses rcu_tasks_wait_gp(), which relies on function
723 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
724 // function sets these function pointers up so that rcu_tasks_wait_gp()
725 // invokes these functions in this order:
727 // rcu_tasks_pregp_step():
728 // Invokes synchronize_rcu() in order to wait for all in-flight
729 // t->on_rq and t->nvcsw transitions to complete. This works because
730 // all such transitions are carried out with interrupts disabled.
731 // rcu_tasks_pertask(), invoked on every non-idle task:
732 // For every runnable non-idle task other than the current one, use
733 // get_task_struct() to pin down that task, snapshot that task's
734 // number of voluntary context switches, and add that task to the
736 // rcu_tasks_postscan():
737 // Invoke synchronize_srcu() to ensure that all tasks that were
738 // in the process of exiting (and which thus might not know to
739 // synchronize with this RCU Tasks grace period) have completed
741 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
742 // Scans the holdout list, attempting to identify a quiescent state
743 // for each task on the list. If there is a quiescent state, the
744 // corresponding task is removed from the holdout list.
745 // rcu_tasks_postgp():
746 // Invokes synchronize_rcu() in order to ensure that all prior
747 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
748 // to have happened before the end of this RCU Tasks grace period.
749 // Again, this works because all such transitions are carried out
750 // with interrupts disabled.
752 // For each exiting task, the exit_tasks_rcu_start() and
753 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
754 // read-side critical sections waited for by rcu_tasks_postscan().
756 // Pre-grace-period update-side code is ordered before the grace
757 // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code
758 // is ordered before the grace period via synchronize_rcu() call in
759 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
762 /* Pre-grace-period preparation. */
763 static void rcu_tasks_pregp_step(void)
766 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
767 * to complete. Invoking synchronize_rcu() suffices because all
768 * these transitions occur with interrupts disabled. Without this
769 * synchronize_rcu(), a read-side critical section that started
770 * before the grace period might be incorrectly seen as having
771 * started after the grace period.
773 * This synchronize_rcu() also dispenses with the need for a
774 * memory barrier on the first store to t->rcu_tasks_holdout,
775 * as it forces the store to happen after the beginning of the
781 /* Per-task initial processing. */
782 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
784 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
786 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
787 WRITE_ONCE(t->rcu_tasks_holdout, true);
788 list_add(&t->rcu_tasks_holdout_list, hop);
792 /* Processing between scanning taskslist and draining the holdout list. */
793 static void rcu_tasks_postscan(struct list_head *hop)
796 * Wait for tasks that are in the process of exiting. This
797 * does only part of the job, ensuring that all tasks that were
798 * previously exiting reach the point where they have disabled
799 * preemption, allowing the later synchronize_rcu() to finish
802 synchronize_srcu(&tasks_rcu_exit_srcu);
805 /* See if tasks are still holding out, complain if so. */
806 static void check_holdout_task(struct task_struct *t,
807 bool needreport, bool *firstreport)
811 if (!READ_ONCE(t->rcu_tasks_holdout) ||
812 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
813 !READ_ONCE(t->on_rq) ||
814 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
815 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
816 WRITE_ONCE(t->rcu_tasks_holdout, false);
817 list_del_init(&t->rcu_tasks_holdout_list);
821 rcu_request_urgent_qs_task(t);
825 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
826 *firstreport = false;
829 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
830 t, ".I"[is_idle_task(t)],
831 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
832 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
833 t->rcu_tasks_idle_cpu, cpu);
837 /* Scan the holdout lists for tasks no longer holding out. */
838 static void check_all_holdout_tasks(struct list_head *hop,
839 bool needreport, bool *firstreport)
841 struct task_struct *t, *t1;
843 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
844 check_holdout_task(t, needreport, firstreport);
849 /* Finish off the Tasks-RCU grace period. */
850 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
853 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
854 * memory barriers prior to them in the schedule() path, memory
855 * reordering on other CPUs could cause their RCU-tasks read-side
856 * critical sections to extend past the end of the grace period.
857 * However, because these ->nvcsw updates are carried out with
858 * interrupts disabled, we can use synchronize_rcu() to force the
859 * needed ordering on all such CPUs.
861 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
862 * accesses to be within the grace period, avoiding the need for
863 * memory barriers for ->rcu_tasks_holdout accesses.
865 * In addition, this synchronize_rcu() waits for exiting tasks
866 * to complete their final preempt_disable() region of execution,
867 * cleaning up after the synchronize_srcu() above.
872 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
873 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
876 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
877 * @rhp: structure to be used for queueing the RCU updates.
878 * @func: actual callback function to be invoked after the grace period
880 * The callback function will be invoked some time after a full grace
881 * period elapses, in other words after all currently executing RCU
882 * read-side critical sections have completed. call_rcu_tasks() assumes
883 * that the read-side critical sections end at a voluntary context
884 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
885 * or transition to usermode execution. As such, there are no read-side
886 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
887 * this primitive is intended to determine that all tasks have passed
888 * through a safe state, not so much for data-structure synchronization.
890 * See the description of call_rcu() for more detailed information on
891 * memory ordering guarantees.
893 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
895 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
897 EXPORT_SYMBOL_GPL(call_rcu_tasks);
900 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
902 * Control will return to the caller some time after a full rcu-tasks
903 * grace period has elapsed, in other words after all currently
904 * executing rcu-tasks read-side critical sections have elapsed. These
905 * read-side critical sections are delimited by calls to schedule(),
906 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
907 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
909 * This is a very specialized primitive, intended only for a few uses in
910 * tracing and other situations requiring manipulation of function
911 * preambles and profiling hooks. The synchronize_rcu_tasks() function
912 * is not (yet) intended for heavy use from multiple CPUs.
914 * See the description of synchronize_rcu() for more detailed information
915 * on memory ordering guarantees.
917 void synchronize_rcu_tasks(void)
919 synchronize_rcu_tasks_generic(&rcu_tasks);
921 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
924 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
926 * Although the current implementation is guaranteed to wait, it is not
927 * obligated to, for example, if there are no pending callbacks.
929 void rcu_barrier_tasks(void)
931 rcu_barrier_tasks_generic(&rcu_tasks);
933 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
935 static int __init rcu_spawn_tasks_kthread(void)
937 cblist_init_generic(&rcu_tasks);
938 rcu_tasks.gp_sleep = HZ / 10;
939 rcu_tasks.init_fract = HZ / 10;
940 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
941 rcu_tasks.pertask_func = rcu_tasks_pertask;
942 rcu_tasks.postscan_func = rcu_tasks_postscan;
943 rcu_tasks.holdouts_func = check_all_holdout_tasks;
944 rcu_tasks.postgp_func = rcu_tasks_postgp;
945 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
949 #if !defined(CONFIG_TINY_RCU)
950 void show_rcu_tasks_classic_gp_kthread(void)
952 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
954 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
955 #endif // !defined(CONFIG_TINY_RCU)
957 /* Do the srcu_read_lock() for the above synchronize_srcu(). */
958 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
961 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
965 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */
966 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu)
968 struct task_struct *t = current;
971 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
973 exit_tasks_rcu_finish_trace(t);
976 #else /* #ifdef CONFIG_TASKS_RCU */
977 void exit_tasks_rcu_start(void) { }
978 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
979 #endif /* #else #ifdef CONFIG_TASKS_RCU */
981 #ifdef CONFIG_TASKS_RUDE_RCU
983 ////////////////////////////////////////////////////////////////////////
985 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
986 // passing an empty function to schedule_on_each_cpu(). This approach
987 // provides an asynchronous call_rcu_tasks_rude() API and batching of
988 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
989 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
990 // and induces otherwise unnecessary context switches on all online CPUs,
991 // whether idle or not.
993 // Callback handling is provided by the rcu_tasks_kthread() function.
995 // Ordering is provided by the scheduler's context-switch code.
997 // Empty function to allow workqueues to force a context switch.
998 static void rcu_tasks_be_rude(struct work_struct *work)
1002 // Wait for one rude RCU-tasks grace period.
1003 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1005 if (num_online_cpus() <= 1)
1006 return; // Fastpath for only one CPU.
1008 rtp->n_ipis += cpumask_weight(cpu_online_mask);
1009 schedule_on_each_cpu(rcu_tasks_be_rude);
1012 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1013 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1017 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1018 * @rhp: structure to be used for queueing the RCU updates.
1019 * @func: actual callback function to be invoked after the grace period
1021 * The callback function will be invoked some time after a full grace
1022 * period elapses, in other words after all currently executing RCU
1023 * read-side critical sections have completed. call_rcu_tasks_rude()
1024 * assumes that the read-side critical sections end at context switch,
1025 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1026 * usermode execution is schedulable). As such, there are no read-side
1027 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1028 * this primitive is intended to determine that all tasks have passed
1029 * through a safe state, not so much for data-structure synchronization.
1031 * See the description of call_rcu() for more detailed information on
1032 * memory ordering guarantees.
1034 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1036 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1038 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1041 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1043 * Control will return to the caller some time after a rude rcu-tasks
1044 * grace period has elapsed, in other words after all currently
1045 * executing rcu-tasks read-side critical sections have elapsed. These
1046 * read-side critical sections are delimited by calls to schedule(),
1047 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1048 * context), and (in theory, anyway) cond_resched().
1050 * This is a very specialized primitive, intended only for a few uses in
1051 * tracing and other situations requiring manipulation of function preambles
1052 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
1053 * (yet) intended for heavy use from multiple CPUs.
1055 * See the description of synchronize_rcu() for more detailed information
1056 * on memory ordering guarantees.
1058 void synchronize_rcu_tasks_rude(void)
1060 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1062 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1065 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1067 * Although the current implementation is guaranteed to wait, it is not
1068 * obligated to, for example, if there are no pending callbacks.
1070 void rcu_barrier_tasks_rude(void)
1072 rcu_barrier_tasks_generic(&rcu_tasks_rude);
1074 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1076 static int __init rcu_spawn_tasks_rude_kthread(void)
1078 cblist_init_generic(&rcu_tasks_rude);
1079 rcu_tasks_rude.gp_sleep = HZ / 10;
1080 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1084 #if !defined(CONFIG_TINY_RCU)
1085 void show_rcu_tasks_rude_gp_kthread(void)
1087 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1089 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1090 #endif // !defined(CONFIG_TINY_RCU)
1091 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1093 ////////////////////////////////////////////////////////////////////////
1095 // Tracing variant of Tasks RCU. This variant is designed to be used
1096 // to protect tracing hooks, including those of BPF. This variant
1099 // 1. Has explicit read-side markers to allow finite grace periods
1100 // in the face of in-kernel loops for PREEMPT=n builds.
1102 // 2. Protects code in the idle loop, exception entry/exit, and
1103 // CPU-hotplug code paths, similar to the capabilities of SRCU.
1105 // 3. Avoids expensive read-side instructions, having overhead similar
1106 // to that of Preemptible RCU.
1108 // There are of course downsides. The grace-period code can send IPIs to
1109 // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
1110 // It is necessary to scan the full tasklist, much as for Tasks RCU. There
1111 // is a single callback queue guarded by a single lock, again, much as for
1112 // Tasks RCU. If needed, these downsides can be at least partially remedied.
1114 // Perhaps most important, this variant of RCU does not affect the vanilla
1115 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
1116 // readers can operate from idle, offline, and exception entry/exit in no
1117 // way allows rcu_preempt and rcu_sched readers to also do so.
1119 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1120 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
1121 // function sets these function pointers up so that rcu_tasks_wait_gp()
1122 // invokes these functions in this order:
1124 // rcu_tasks_trace_pregp_step():
1125 // Initialize the count of readers and block CPU-hotplug operations.
1126 // rcu_tasks_trace_pertask(), invoked on every non-idle task:
1127 // Initialize per-task state and attempt to identify an immediate
1128 // quiescent state for that task, or, failing that, attempt to
1129 // set that task's .need_qs flag so that task's next outermost
1130 // rcu_read_unlock_trace() will report the quiescent state (in which
1131 // case the count of readers is incremented). If both attempts fail,
1132 // the task is added to a "holdout" list. Note that IPIs are used
1133 // to invoke trc_read_check_handler() in the context of running tasks
1134 // in order to avoid ordering overhead on common-case shared-variable
1136 // rcu_tasks_trace_postscan():
1137 // Initialize state and attempt to identify an immediate quiescent
1138 // state as above (but only for idle tasks), unblock CPU-hotplug
1139 // operations, and wait for an RCU grace period to avoid races with
1140 // tasks that are in the process of exiting.
1141 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1142 // Scans the holdout list, attempting to identify a quiescent state
1143 // for each task on the list. If there is a quiescent state, the
1144 // corresponding task is removed from the holdout list.
1145 // rcu_tasks_trace_postgp():
1146 // Wait for the count of readers do drop to zero, reporting any stalls.
1147 // Also execute full memory barriers to maintain ordering with code
1148 // executing after the grace period.
1150 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1152 // Pre-grace-period update-side code is ordered before the grace
1153 // period via the ->cbs_lock and barriers in rcu_tasks_kthread().
1154 // Pre-grace-period read-side code is ordered before the grace period by
1155 // atomic_dec_and_test() of the count of readers (for IPIed readers) and by
1156 // scheduler context-switch ordering (for locked-down non-running readers).
1158 // The lockdep state must be outside of #ifdef to be useful.
1159 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1160 static struct lock_class_key rcu_lock_trace_key;
1161 struct lockdep_map rcu_trace_lock_map =
1162 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1163 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1164 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1166 #ifdef CONFIG_TASKS_TRACE_RCU
1168 static atomic_t trc_n_readers_need_end; // Number of waited-for readers.
1169 static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks.
1171 // Record outstanding IPIs to each CPU. No point in sending two...
1172 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1174 // The number of detections of task quiescent state relying on
1175 // heavyweight readers executing explicit memory barriers.
1176 static unsigned long n_heavy_reader_attempts;
1177 static unsigned long n_heavy_reader_updates;
1178 static unsigned long n_heavy_reader_ofl_updates;
1180 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1181 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1185 * This irq_work handler allows rcu_read_unlock_trace() to be invoked
1186 * while the scheduler locks are held.
1188 static void rcu_read_unlock_iw(struct irq_work *iwp)
1192 static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw);
1194 /* If we are the last reader, wake up the grace-period kthread. */
1195 void rcu_read_unlock_trace_special(struct task_struct *t)
1197 int nq = READ_ONCE(t->trc_reader_special.b.need_qs);
1199 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) &&
1200 t->trc_reader_special.b.need_mb)
1201 smp_mb(); // Pairs with update-side barriers.
1202 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1204 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
1205 WRITE_ONCE(t->trc_reader_nesting, 0);
1206 if (nq && atomic_dec_and_test(&trc_n_readers_need_end))
1207 irq_work_queue(&rcu_tasks_trace_iw);
1209 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1211 /* Add a task to the holdout list, if it is not already on the list. */
1212 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1214 if (list_empty(&t->trc_holdout_list)) {
1216 list_add(&t->trc_holdout_list, bhp);
1220 /* Remove a task from the holdout list, if it is in fact present. */
1221 static void trc_del_holdout(struct task_struct *t)
1223 if (!list_empty(&t->trc_holdout_list)) {
1224 list_del_init(&t->trc_holdout_list);
1229 /* IPI handler to check task state. */
1230 static void trc_read_check_handler(void *t_in)
1232 struct task_struct *t = current;
1233 struct task_struct *texp = t_in;
1235 // If the task is no longer running on this CPU, leave.
1236 if (unlikely(texp != t)) {
1237 goto reset_ipi; // Already on holdout list, so will check later.
1240 // If the task is not in a read-side critical section, and
1241 // if this is the last reader, awaken the grace-period kthread.
1242 if (likely(!READ_ONCE(t->trc_reader_nesting))) {
1243 WRITE_ONCE(t->trc_reader_checked, true);
1246 // If we are racing with an rcu_read_unlock_trace(), try again later.
1247 if (unlikely(READ_ONCE(t->trc_reader_nesting) < 0))
1249 WRITE_ONCE(t->trc_reader_checked, true);
1251 // Get here if the task is in a read-side critical section. Set
1252 // its state so that it will awaken the grace-period kthread upon
1253 // exit from that critical section.
1254 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
1255 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
1256 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
1259 // Allow future IPIs to be sent on CPU and for task.
1260 // Also order this IPI handler against any later manipulations of
1261 // the intended task.
1262 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1263 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1266 /* Callback function for scheduler to check locked-down task. */
1267 static int trc_inspect_reader(struct task_struct *t, void *arg)
1269 int cpu = task_cpu(t);
1271 bool ofl = cpu_is_offline(cpu);
1274 WARN_ON_ONCE(ofl && !is_idle_task(t));
1276 // If no chance of heavyweight readers, do it the hard way.
1277 if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1280 // If heavyweight readers are enabled on the remote task,
1281 // we can inspect its state despite its currently running.
1282 // However, we cannot safely change its state.
1283 n_heavy_reader_attempts++;
1284 if (!ofl && // Check for "running" idle tasks on offline CPUs.
1285 !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1286 return -EINVAL; // No quiescent state, do it the hard way.
1287 n_heavy_reader_updates++;
1289 n_heavy_reader_ofl_updates++;
1292 // The task is not running, so C-language access is safe.
1293 nesting = t->trc_reader_nesting;
1296 // If not exiting a read-side critical section, mark as checked
1297 // so that the grace-period kthread will remove it from the
1299 t->trc_reader_checked = nesting >= 0;
1301 return nesting ? -EINVAL : 0; // If in QS, done, otherwise try again later.
1303 // The task is in a read-side critical section, so set up its
1304 // state so that it will awaken the grace-period kthread upon exit
1305 // from that critical section.
1306 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
1307 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
1308 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
1312 /* Attempt to extract the state for the specified task. */
1313 static void trc_wait_for_one_reader(struct task_struct *t,
1314 struct list_head *bhp)
1318 // If a previous IPI is still in flight, let it complete.
1319 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1322 // The current task had better be in a quiescent state.
1324 t->trc_reader_checked = true;
1325 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1329 // Attempt to nail down the task for inspection.
1331 if (!task_call_func(t, trc_inspect_reader, NULL)) {
1337 // If this task is not yet on the holdout list, then we are in
1338 // an RCU read-side critical section. Otherwise, the invocation of
1339 // trc_add_holdout() that added it to the list did the necessary
1340 // get_task_struct(). Either way, the task cannot be freed out
1341 // from under this code.
1343 // If currently running, send an IPI, either way, add to list.
1344 trc_add_holdout(t, bhp);
1346 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1347 // The task is currently running, so try IPIing it.
1350 // If there is already an IPI outstanding, let it happen.
1351 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1354 per_cpu(trc_ipi_to_cpu, cpu) = true;
1355 t->trc_ipi_to_cpu = cpu;
1356 rcu_tasks_trace.n_ipis++;
1357 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1358 // Just in case there is some other reason for
1359 // failure than the target CPU being offline.
1360 WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n",
1362 rcu_tasks_trace.n_ipis_fails++;
1363 per_cpu(trc_ipi_to_cpu, cpu) = false;
1364 t->trc_ipi_to_cpu = -1;
1369 /* Initialize for a new RCU-tasks-trace grace period. */
1370 static void rcu_tasks_trace_pregp_step(void)
1374 // Allow for fast-acting IPIs.
1375 atomic_set(&trc_n_readers_need_end, 1);
1377 // There shouldn't be any old IPIs, but...
1378 for_each_possible_cpu(cpu)
1379 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1381 // Disable CPU hotplug across the tasklist scan.
1382 // This also waits for all readers in CPU-hotplug code paths.
1386 /* Do first-round processing for the specified task. */
1387 static void rcu_tasks_trace_pertask(struct task_struct *t,
1388 struct list_head *hop)
1390 // During early boot when there is only the one boot CPU, there
1391 // is no idle task for the other CPUs. Just return.
1392 if (unlikely(t == NULL))
1395 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
1396 WRITE_ONCE(t->trc_reader_checked, false);
1397 t->trc_ipi_to_cpu = -1;
1398 trc_wait_for_one_reader(t, hop);
1402 * Do intermediate processing between task and holdout scans and
1403 * pick up the idle tasks.
1405 static void rcu_tasks_trace_postscan(struct list_head *hop)
1409 for_each_possible_cpu(cpu)
1410 rcu_tasks_trace_pertask(idle_task(cpu), hop);
1412 // Re-enable CPU hotplug now that the tasklist scan has completed.
1415 // Wait for late-stage exiting tasks to finish exiting.
1416 // These might have passed the call to exit_tasks_rcu_finish().
1418 // Any tasks that exit after this point will set ->trc_reader_checked.
1421 /* Communicate task state back to the RCU tasks trace stall warning request. */
1422 struct trc_stall_chk_rdr {
1428 static int trc_check_slow_task(struct task_struct *t, void *arg)
1430 struct trc_stall_chk_rdr *trc_rdrp = arg;
1433 return false; // It is running, so decline to inspect it.
1434 trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1435 trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1436 trc_rdrp->needqs = READ_ONCE(t->trc_reader_special.b.need_qs);
1440 /* Show the state of a task stalling the current RCU tasks trace GP. */
1441 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1444 struct trc_stall_chk_rdr trc_rdr;
1445 bool is_idle_tsk = is_idle_task(t);
1448 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1449 *firstreport = false;
1452 if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1453 pr_alert("P%d: %c\n",
1457 pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1459 ".I"[trc_rdr.ipi_to_cpu >= 0],
1461 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1463 " N"[!!trc_rdr.needqs],
1468 /* List stalled IPIs for RCU tasks trace. */
1469 static void show_stalled_ipi_trace(void)
1473 for_each_possible_cpu(cpu)
1474 if (per_cpu(trc_ipi_to_cpu, cpu))
1475 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1478 /* Do one scan of the holdout list. */
1479 static void check_all_holdout_tasks_trace(struct list_head *hop,
1480 bool needreport, bool *firstreport)
1482 struct task_struct *g, *t;
1484 // Disable CPU hotplug across the holdout list scan.
1487 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1488 // If safe and needed, try to check the current task.
1489 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1490 !READ_ONCE(t->trc_reader_checked))
1491 trc_wait_for_one_reader(t, hop);
1493 // If check succeeded, remove this task from the list.
1494 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1495 READ_ONCE(t->trc_reader_checked))
1497 else if (needreport)
1498 show_stalled_task_trace(t, firstreport);
1501 // Re-enable CPU hotplug now that the holdout list scan has completed.
1506 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1507 show_stalled_ipi_trace();
1511 static void rcu_tasks_trace_empty_fn(void *unused)
1515 /* Wait for grace period to complete and provide ordering. */
1516 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1520 struct task_struct *g, *t;
1521 LIST_HEAD(holdouts);
1524 // Wait for any lingering IPI handlers to complete. Note that
1525 // if a CPU has gone offline or transitioned to userspace in the
1526 // meantime, all IPI handlers should have been drained beforehand.
1527 // Yes, this assumes that CPUs process IPIs in order. If that ever
1528 // changes, there will need to be a recheck and/or timed wait.
1529 for_each_online_cpu(cpu)
1530 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1531 smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1533 // Remove the safety count.
1534 smp_mb__before_atomic(); // Order vs. earlier atomics
1535 atomic_dec(&trc_n_readers_need_end);
1536 smp_mb__after_atomic(); // Order vs. later atomics
1538 // Wait for readers.
1539 set_tasks_gp_state(rtp, RTGS_WAIT_READERS);
1541 ret = wait_event_idle_exclusive_timeout(
1543 atomic_read(&trc_n_readers_need_end) == 0,
1544 READ_ONCE(rcu_task_stall_timeout));
1546 break; // Count reached zero.
1547 // Stall warning time, so make a list of the offenders.
1549 for_each_process_thread(g, t)
1550 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1551 trc_add_holdout(t, &holdouts);
1554 list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) {
1555 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1556 show_stalled_task_trace(t, &firstreport);
1557 trc_del_holdout(t); // Release task_struct reference.
1560 pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1561 show_stalled_ipi_trace();
1562 pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end));
1564 smp_mb(); // Caller's code must be ordered after wakeup.
1565 // Pairs with pretty much every ordering primitive.
1568 /* Report any needed quiescent state for this exiting task. */
1569 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1571 WRITE_ONCE(t->trc_reader_checked, true);
1572 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1573 WRITE_ONCE(t->trc_reader_nesting, 0);
1574 if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)))
1575 rcu_read_unlock_trace_special(t);
1579 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1580 * @rhp: structure to be used for queueing the RCU updates.
1581 * @func: actual callback function to be invoked after the grace period
1583 * The callback function will be invoked some time after a trace rcu-tasks
1584 * grace period elapses, in other words after all currently executing
1585 * trace rcu-tasks read-side critical sections have completed. These
1586 * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1587 * and rcu_read_unlock_trace().
1589 * See the description of call_rcu() for more detailed information on
1590 * memory ordering guarantees.
1592 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1594 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1596 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1599 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1601 * Control will return to the caller some time after a trace rcu-tasks
1602 * grace period has elapsed, in other words after all currently executing
1603 * trace rcu-tasks read-side critical sections have elapsed. These read-side
1604 * critical sections are delimited by calls to rcu_read_lock_trace()
1605 * and rcu_read_unlock_trace().
1607 * This is a very specialized primitive, intended only for a few uses in
1608 * tracing and other situations requiring manipulation of function preambles
1609 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1610 * (yet) intended for heavy use from multiple CPUs.
1612 * See the description of synchronize_rcu() for more detailed information
1613 * on memory ordering guarantees.
1615 void synchronize_rcu_tasks_trace(void)
1617 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1618 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1620 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1623 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1625 * Although the current implementation is guaranteed to wait, it is not
1626 * obligated to, for example, if there are no pending callbacks.
1628 void rcu_barrier_tasks_trace(void)
1630 rcu_barrier_tasks_generic(&rcu_tasks_trace);
1632 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1634 static int __init rcu_spawn_tasks_trace_kthread(void)
1636 cblist_init_generic(&rcu_tasks_trace);
1637 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1638 rcu_tasks_trace.gp_sleep = HZ / 10;
1639 rcu_tasks_trace.init_fract = HZ / 10;
1641 rcu_tasks_trace.gp_sleep = HZ / 200;
1642 if (rcu_tasks_trace.gp_sleep <= 0)
1643 rcu_tasks_trace.gp_sleep = 1;
1644 rcu_tasks_trace.init_fract = HZ / 200;
1645 if (rcu_tasks_trace.init_fract <= 0)
1646 rcu_tasks_trace.init_fract = 1;
1648 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1649 rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask;
1650 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1651 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1652 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1653 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1657 #if !defined(CONFIG_TINY_RCU)
1658 void show_rcu_tasks_trace_gp_kthread(void)
1662 sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end),
1663 data_race(n_heavy_reader_ofl_updates),
1664 data_race(n_heavy_reader_updates),
1665 data_race(n_heavy_reader_attempts));
1666 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1668 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1669 #endif // !defined(CONFIG_TINY_RCU)
1671 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1672 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1673 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1675 #ifndef CONFIG_TINY_RCU
1676 void show_rcu_tasks_gp_kthreads(void)
1678 show_rcu_tasks_classic_gp_kthread();
1679 show_rcu_tasks_rude_gp_kthread();
1680 show_rcu_tasks_trace_gp_kthread();
1682 #endif /* #ifndef CONFIG_TINY_RCU */
1684 #ifdef CONFIG_PROVE_RCU
1685 struct rcu_tasks_test_desc {
1691 static struct rcu_tasks_test_desc tests[] = {
1693 .name = "call_rcu_tasks()",
1694 /* If not defined, the test is skipped. */
1695 .notrun = !IS_ENABLED(CONFIG_TASKS_RCU),
1698 .name = "call_rcu_tasks_rude()",
1699 /* If not defined, the test is skipped. */
1700 .notrun = !IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1703 .name = "call_rcu_tasks_trace()",
1704 /* If not defined, the test is skipped. */
1705 .notrun = !IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1709 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1711 struct rcu_tasks_test_desc *rttd =
1712 container_of(rhp, struct rcu_tasks_test_desc, rh);
1714 pr_info("Callback from %s invoked.\n", rttd->name);
1716 rttd->notrun = true;
1719 static void rcu_tasks_initiate_self_tests(void)
1721 pr_info("Running RCU-tasks wait API self tests\n");
1722 #ifdef CONFIG_TASKS_RCU
1723 synchronize_rcu_tasks();
1724 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1727 #ifdef CONFIG_TASKS_RUDE_RCU
1728 synchronize_rcu_tasks_rude();
1729 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1732 #ifdef CONFIG_TASKS_TRACE_RCU
1733 synchronize_rcu_tasks_trace();
1734 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1738 static int rcu_tasks_verify_self_tests(void)
1743 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1744 if (!tests[i].notrun) { // still hanging.
1745 pr_err("%s has been failed.\n", tests[i].name);
1755 late_initcall(rcu_tasks_verify_self_tests);
1756 #else /* #ifdef CONFIG_PROVE_RCU */
1757 static void rcu_tasks_initiate_self_tests(void) { }
1758 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1760 void __init rcu_init_tasks_generic(void)
1762 #ifdef CONFIG_TASKS_RCU
1763 rcu_spawn_tasks_kthread();
1766 #ifdef CONFIG_TASKS_RUDE_RCU
1767 rcu_spawn_tasks_rude_kthread();
1770 #ifdef CONFIG_TASKS_TRACE_RCU
1771 rcu_spawn_tasks_trace_kthread();
1774 // Run the self-tests.
1775 rcu_tasks_initiate_self_tests();
1778 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1779 static inline void rcu_tasks_bootup_oddness(void) {}
1780 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */