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_wq: Wait queue allowing 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 wait_queue_head cbs_wq;
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(call_rcu_tasks_iw_wakeup), \
118 static struct rcu_tasks rt_name = \
120 .cbs_wq = __WAIT_QUEUE_HEAD_INITIALIZER(rt_name.cbs_wq), \
121 .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \
124 .rtpcpu = &rt_name ## __percpu, \
126 .percpu_enqueue_shift = ilog2(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);
147 static int rcu_task_enqueue_lim __read_mostly = -1;
148 module_param(rcu_task_enqueue_lim, int, 0444);
150 static bool rcu_task_cb_adjust;
151 static int rcu_task_contend_lim __read_mostly = 100;
152 module_param(rcu_task_contend_lim, int, 0444);
153 static int rcu_task_collapse_lim __read_mostly = 10;
154 module_param(rcu_task_collapse_lim, int, 0444);
156 /* RCU tasks grace-period state for debugging. */
158 #define RTGS_WAIT_WAIT_CBS 1
159 #define RTGS_WAIT_GP 2
160 #define RTGS_PRE_WAIT_GP 3
161 #define RTGS_SCAN_TASKLIST 4
162 #define RTGS_POST_SCAN_TASKLIST 5
163 #define RTGS_WAIT_SCAN_HOLDOUTS 6
164 #define RTGS_SCAN_HOLDOUTS 7
165 #define RTGS_POST_GP 8
166 #define RTGS_WAIT_READERS 9
167 #define RTGS_INVOKE_CBS 10
168 #define RTGS_WAIT_CBS 11
169 #ifndef CONFIG_TINY_RCU
170 static const char * const rcu_tasks_gp_state_names[] = {
172 "RTGS_WAIT_WAIT_CBS",
175 "RTGS_SCAN_TASKLIST",
176 "RTGS_POST_SCAN_TASKLIST",
177 "RTGS_WAIT_SCAN_HOLDOUTS",
178 "RTGS_SCAN_HOLDOUTS",
184 #endif /* #ifndef CONFIG_TINY_RCU */
186 ////////////////////////////////////////////////////////////////////////
190 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
192 /* Record grace-period phase and time. */
193 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
195 rtp->gp_state = newstate;
196 rtp->gp_jiffies = jiffies;
199 #ifndef CONFIG_TINY_RCU
200 /* Return state name. */
201 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
203 int i = data_race(rtp->gp_state); // Let KCSAN detect update races
204 int j = READ_ONCE(i); // Prevent the compiler from reading twice
206 if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
208 return rcu_tasks_gp_state_names[j];
210 #endif /* #ifndef CONFIG_TINY_RCU */
212 // Initialize per-CPU callback lists for the specified flavor of
214 static void cblist_init_generic(struct rcu_tasks *rtp)
220 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
221 if (rcu_task_enqueue_lim < 0) {
222 rcu_task_enqueue_lim = 1;
223 rcu_task_cb_adjust = true;
224 pr_info("%s: Setting adjustable number of callback queues.\n", __func__);
225 } else if (rcu_task_enqueue_lim == 0) {
226 rcu_task_enqueue_lim = 1;
228 lim = rcu_task_enqueue_lim;
230 if (lim > nr_cpu_ids)
232 WRITE_ONCE(rtp->percpu_enqueue_shift, ilog2(nr_cpu_ids / lim));
233 WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
234 smp_store_release(&rtp->percpu_enqueue_lim, lim);
235 for_each_possible_cpu(cpu) {
236 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
238 WARN_ON_ONCE(!rtpcp);
240 raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
241 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
242 if (rcu_segcblist_empty(&rtpcp->cblist))
243 rcu_segcblist_init(&rtpcp->cblist);
244 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
247 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
249 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
250 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));
253 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
254 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
256 struct rcu_tasks *rtp;
257 struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
260 wake_up(&rtp->cbs_wq);
263 // Enqueue a callback for the specified flavor of Tasks RCU.
264 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
265 struct rcu_tasks *rtp)
269 bool needadjust = false;
271 struct rcu_tasks_percpu *rtpcp;
275 local_irq_save(flags);
277 rtpcp = per_cpu_ptr(rtp->rtpcpu,
278 smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift));
279 if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
280 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
282 if (rtpcp->rtp_jiffies != j) {
283 rtpcp->rtp_jiffies = j;
284 rtpcp->rtp_n_lock_retries = 0;
286 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
287 READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
288 needadjust = true; // Defer adjustment to avoid deadlock.
290 if (!rcu_segcblist_is_enabled(&rtpcp->cblist)) {
291 raw_spin_unlock_rcu_node(rtpcp); // irqs remain disabled.
292 cblist_init_generic(rtp);
293 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
295 needwake = rcu_segcblist_empty(&rtpcp->cblist);
296 rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
297 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
298 if (unlikely(needadjust)) {
299 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
300 if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
301 WRITE_ONCE(rtp->percpu_enqueue_shift, ilog2(nr_cpu_ids));
302 WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
303 smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
304 pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
306 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
309 /* We can't create the thread unless interrupts are enabled. */
310 if (needwake && READ_ONCE(rtp->kthread_ptr))
311 irq_work_queue(&rtpcp->rtp_irq_work);
314 // Wait for a grace period for the specified flavor of Tasks RCU.
315 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
317 /* Complain if the scheduler has not started. */
318 RCU_LOCKDEP_WARN(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
319 "synchronize_rcu_tasks called too soon");
321 /* Wait for the grace period. */
322 wait_rcu_gp(rtp->call_func);
325 // RCU callback function for rcu_barrier_tasks_generic().
326 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
328 struct rcu_tasks *rtp;
329 struct rcu_tasks_percpu *rtpcp;
331 rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
333 if (atomic_dec_and_test(&rtp->barrier_q_count))
334 complete(&rtp->barrier_q_completion);
337 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
338 // Operates in a manner similar to rcu_barrier().
339 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
343 struct rcu_tasks_percpu *rtpcp;
344 unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
346 mutex_lock(&rtp->barrier_q_mutex);
347 if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
349 mutex_unlock(&rtp->barrier_q_mutex);
352 rcu_seq_start(&rtp->barrier_q_seq);
353 init_completion(&rtp->barrier_q_completion);
354 atomic_set(&rtp->barrier_q_count, 2);
355 for_each_possible_cpu(cpu) {
356 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
358 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
359 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
360 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
361 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
362 atomic_inc(&rtp->barrier_q_count);
363 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
365 if (atomic_sub_and_test(2, &rtp->barrier_q_count))
366 complete(&rtp->barrier_q_completion);
367 wait_for_completion(&rtp->barrier_q_completion);
368 rcu_seq_end(&rtp->barrier_q_seq);
369 mutex_unlock(&rtp->barrier_q_mutex);
372 // Advance callbacks and indicate whether either a grace period or
373 // callback invocation is needed.
374 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
383 for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
384 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
386 /* Advance and accelerate any new callbacks. */
387 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
389 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
390 // Should we shrink down to a single callback queue?
391 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
397 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
398 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
399 if (rcu_segcblist_pend_cbs(&rtpcp->cblist))
401 if (!rcu_segcblist_empty(&rtpcp->cblist))
403 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
406 // Shrink down to a single callback queue if appropriate.
407 // This is done in two stages: (1) If there are no more than
408 // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
409 // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period,
410 // if there has not been an increase in callbacks, limit dequeuing
411 // to CPU 0. Note the matching RCU read-side critical section in
412 // call_rcu_tasks_generic().
413 if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
414 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
415 if (rtp->percpu_enqueue_lim > 1) {
416 WRITE_ONCE(rtp->percpu_enqueue_shift, ilog2(nr_cpu_ids));
417 smp_store_release(&rtp->percpu_enqueue_lim, 1);
418 rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
419 pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
421 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
423 if (rcu_task_cb_adjust && !ncbsnz &&
424 poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq)) {
425 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
426 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
427 WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
428 pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
430 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
436 // Advance callbacks and invoke any that are ready.
437 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
443 struct rcu_head *rhp;
444 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
445 struct rcu_tasks_percpu *rtpcp_next;
448 cpunext = cpu * 2 + 1;
449 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
450 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
451 queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
453 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
454 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
455 queue_work_on(cpunext, system_wq, &rtpcp_next->rtp_work);
459 if (rcu_segcblist_empty(&rtpcp->cblist))
461 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
462 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
463 rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
464 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
466 for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
472 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
473 rcu_segcblist_add_len(&rtpcp->cblist, -len);
474 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
475 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
478 // Workqueue flood to advance callbacks and invoke any that are ready.
479 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
481 struct rcu_tasks *rtp;
482 struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
485 rcu_tasks_invoke_cbs(rtp, rtpcp);
488 /* RCU-tasks kthread that detects grace periods and invokes callbacks. */
489 static int __noreturn rcu_tasks_kthread(void *arg)
492 struct rcu_tasks *rtp = arg;
494 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
495 housekeeping_affine(current, HK_TYPE_RCU);
496 WRITE_ONCE(rtp->kthread_ptr, current); // Let GPs start!
499 * Each pass through the following loop makes one check for
500 * newly arrived callbacks, and, if there are some, waits for
501 * one RCU-tasks grace period and then invokes the callbacks.
502 * This loop is terminated by the system going down. ;-)
505 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
507 /* If there were none, wait a bit and start over. */
508 wait_event_idle(rtp->cbs_wq, (needgpcb = rcu_tasks_need_gpcb(rtp)));
510 if (needgpcb & 0x2) {
511 // Wait for one grace period.
512 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
513 rtp->gp_start = jiffies;
514 rcu_seq_start(&rtp->tasks_gp_seq);
516 rcu_seq_end(&rtp->tasks_gp_seq);
519 /* Invoke callbacks. */
520 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
521 rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
523 /* Paranoid sleep to keep this from entering a tight loop */
524 schedule_timeout_idle(rtp->gp_sleep);
528 /* Spawn RCU-tasks grace-period kthread. */
529 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
531 struct task_struct *t;
533 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
534 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
536 smp_mb(); /* Ensure others see full kthread. */
539 #ifndef CONFIG_TINY_RCU
542 * Print any non-default Tasks RCU settings.
544 static void __init rcu_tasks_bootup_oddness(void)
546 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
547 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
548 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
549 #endif /* #ifdef CONFIG_TASKS_RCU */
550 #ifdef CONFIG_TASKS_RCU
551 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
552 #endif /* #ifdef CONFIG_TASKS_RCU */
553 #ifdef CONFIG_TASKS_RUDE_RCU
554 pr_info("\tRude variant of Tasks RCU enabled.\n");
555 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
556 #ifdef CONFIG_TASKS_TRACE_RCU
557 pr_info("\tTracing variant of Tasks RCU enabled.\n");
558 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
561 #endif /* #ifndef CONFIG_TINY_RCU */
563 #ifndef CONFIG_TINY_RCU
564 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
565 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
567 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, 0); // for_each...
568 pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c %s\n",
570 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
571 jiffies - data_race(rtp->gp_jiffies),
572 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
573 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
574 ".k"[!!data_race(rtp->kthread_ptr)],
575 ".C"[!data_race(rcu_segcblist_empty(&rtpcp->cblist))],
578 #endif // #ifndef CONFIG_TINY_RCU
580 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
582 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
584 ////////////////////////////////////////////////////////////////////////
586 // Shared code between task-list-scanning variants of Tasks RCU.
588 /* Wait for one RCU-tasks grace period. */
589 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
591 struct task_struct *g, *t;
592 unsigned long lastreport;
596 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
600 * There were callbacks, so we need to wait for an RCU-tasks
601 * grace period. Start off by scanning the task list for tasks
602 * that are not already voluntarily blocked. Mark these tasks
603 * and make a list of them in holdouts.
605 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
607 for_each_process_thread(g, t)
608 rtp->pertask_func(t, &holdouts);
611 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
612 rtp->postscan_func(&holdouts);
615 * Each pass through the following loop scans the list of holdout
616 * tasks, removing any that are no longer holdouts. When the list
617 * is empty, we are done.
619 lastreport = jiffies;
621 // Start off with initial wait and slowly back off to 1 HZ wait.
622 fract = rtp->init_fract;
624 while (!list_empty(&holdouts)) {
629 /* Slowly back off waiting for holdouts */
630 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
631 schedule_timeout_idle(fract);
636 rtst = READ_ONCE(rcu_task_stall_timeout);
637 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
639 lastreport = jiffies;
641 WARN_ON(signal_pending(current));
642 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
643 rtp->holdouts_func(&holdouts, needreport, &firstreport);
646 set_tasks_gp_state(rtp, RTGS_POST_GP);
647 rtp->postgp_func(rtp);
650 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
652 #ifdef CONFIG_TASKS_RCU
654 ////////////////////////////////////////////////////////////////////////
656 // Simple variant of RCU whose quiescent states are voluntary context
657 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
658 // As such, grace periods can take one good long time. There are no
659 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
660 // because this implementation is intended to get the system into a safe
661 // state for some of the manipulations involved in tracing and the like.
662 // Finally, this implementation does not support high call_rcu_tasks()
663 // rates from multiple CPUs. If this is required, per-CPU callback lists
666 // The implementation uses rcu_tasks_wait_gp(), which relies on function
667 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
668 // function sets these function pointers up so that rcu_tasks_wait_gp()
669 // invokes these functions in this order:
671 // rcu_tasks_pregp_step():
672 // Invokes synchronize_rcu() in order to wait for all in-flight
673 // t->on_rq and t->nvcsw transitions to complete. This works because
674 // all such transitions are carried out with interrupts disabled.
675 // rcu_tasks_pertask(), invoked on every non-idle task:
676 // For every runnable non-idle task other than the current one, use
677 // get_task_struct() to pin down that task, snapshot that task's
678 // number of voluntary context switches, and add that task to the
680 // rcu_tasks_postscan():
681 // Invoke synchronize_srcu() to ensure that all tasks that were
682 // in the process of exiting (and which thus might not know to
683 // synchronize with this RCU Tasks grace period) have completed
685 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
686 // Scans the holdout list, attempting to identify a quiescent state
687 // for each task on the list. If there is a quiescent state, the
688 // corresponding task is removed from the holdout list.
689 // rcu_tasks_postgp():
690 // Invokes synchronize_rcu() in order to ensure that all prior
691 // t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
692 // to have happened before the end of this RCU Tasks grace period.
693 // Again, this works because all such transitions are carried out
694 // with interrupts disabled.
696 // For each exiting task, the exit_tasks_rcu_start() and
697 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
698 // read-side critical sections waited for by rcu_tasks_postscan().
700 // Pre-grace-period update-side code is ordered before the grace
701 // via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code
702 // is ordered before the grace period via synchronize_rcu() call in
703 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
706 /* Pre-grace-period preparation. */
707 static void rcu_tasks_pregp_step(void)
710 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
711 * to complete. Invoking synchronize_rcu() suffices because all
712 * these transitions occur with interrupts disabled. Without this
713 * synchronize_rcu(), a read-side critical section that started
714 * before the grace period might be incorrectly seen as having
715 * started after the grace period.
717 * This synchronize_rcu() also dispenses with the need for a
718 * memory barrier on the first store to t->rcu_tasks_holdout,
719 * as it forces the store to happen after the beginning of the
725 /* Per-task initial processing. */
726 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
728 if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
730 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
731 WRITE_ONCE(t->rcu_tasks_holdout, true);
732 list_add(&t->rcu_tasks_holdout_list, hop);
736 /* Processing between scanning taskslist and draining the holdout list. */
737 static void rcu_tasks_postscan(struct list_head *hop)
740 * Wait for tasks that are in the process of exiting. This
741 * does only part of the job, ensuring that all tasks that were
742 * previously exiting reach the point where they have disabled
743 * preemption, allowing the later synchronize_rcu() to finish
746 synchronize_srcu(&tasks_rcu_exit_srcu);
749 /* See if tasks are still holding out, complain if so. */
750 static void check_holdout_task(struct task_struct *t,
751 bool needreport, bool *firstreport)
755 if (!READ_ONCE(t->rcu_tasks_holdout) ||
756 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
757 !READ_ONCE(t->on_rq) ||
758 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
759 !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
760 WRITE_ONCE(t->rcu_tasks_holdout, false);
761 list_del_init(&t->rcu_tasks_holdout_list);
765 rcu_request_urgent_qs_task(t);
769 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
770 *firstreport = false;
773 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
774 t, ".I"[is_idle_task(t)],
775 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
776 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
777 t->rcu_tasks_idle_cpu, cpu);
781 /* Scan the holdout lists for tasks no longer holding out. */
782 static void check_all_holdout_tasks(struct list_head *hop,
783 bool needreport, bool *firstreport)
785 struct task_struct *t, *t1;
787 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
788 check_holdout_task(t, needreport, firstreport);
793 /* Finish off the Tasks-RCU grace period. */
794 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
797 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
798 * memory barriers prior to them in the schedule() path, memory
799 * reordering on other CPUs could cause their RCU-tasks read-side
800 * critical sections to extend past the end of the grace period.
801 * However, because these ->nvcsw updates are carried out with
802 * interrupts disabled, we can use synchronize_rcu() to force the
803 * needed ordering on all such CPUs.
805 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
806 * accesses to be within the grace period, avoiding the need for
807 * memory barriers for ->rcu_tasks_holdout accesses.
809 * In addition, this synchronize_rcu() waits for exiting tasks
810 * to complete their final preempt_disable() region of execution,
811 * cleaning up after the synchronize_srcu() above.
816 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
817 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
820 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
821 * @rhp: structure to be used for queueing the RCU updates.
822 * @func: actual callback function to be invoked after the grace period
824 * The callback function will be invoked some time after a full grace
825 * period elapses, in other words after all currently executing RCU
826 * read-side critical sections have completed. call_rcu_tasks() assumes
827 * that the read-side critical sections end at a voluntary context
828 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
829 * or transition to usermode execution. As such, there are no read-side
830 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
831 * this primitive is intended to determine that all tasks have passed
832 * through a safe state, not so much for data-structure synchronization.
834 * See the description of call_rcu() for more detailed information on
835 * memory ordering guarantees.
837 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
839 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
841 EXPORT_SYMBOL_GPL(call_rcu_tasks);
844 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
846 * Control will return to the caller some time after a full rcu-tasks
847 * grace period has elapsed, in other words after all currently
848 * executing rcu-tasks read-side critical sections have elapsed. These
849 * read-side critical sections are delimited by calls to schedule(),
850 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
851 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
853 * This is a very specialized primitive, intended only for a few uses in
854 * tracing and other situations requiring manipulation of function
855 * preambles and profiling hooks. The synchronize_rcu_tasks() function
856 * is not (yet) intended for heavy use from multiple CPUs.
858 * See the description of synchronize_rcu() for more detailed information
859 * on memory ordering guarantees.
861 void synchronize_rcu_tasks(void)
863 synchronize_rcu_tasks_generic(&rcu_tasks);
865 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
868 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
870 * Although the current implementation is guaranteed to wait, it is not
871 * obligated to, for example, if there are no pending callbacks.
873 void rcu_barrier_tasks(void)
875 rcu_barrier_tasks_generic(&rcu_tasks);
877 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
879 static int __init rcu_spawn_tasks_kthread(void)
881 cblist_init_generic(&rcu_tasks);
882 rcu_tasks.gp_sleep = HZ / 10;
883 rcu_tasks.init_fract = HZ / 10;
884 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
885 rcu_tasks.pertask_func = rcu_tasks_pertask;
886 rcu_tasks.postscan_func = rcu_tasks_postscan;
887 rcu_tasks.holdouts_func = check_all_holdout_tasks;
888 rcu_tasks.postgp_func = rcu_tasks_postgp;
889 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
893 #if !defined(CONFIG_TINY_RCU)
894 void show_rcu_tasks_classic_gp_kthread(void)
896 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
898 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
899 #endif // !defined(CONFIG_TINY_RCU)
901 /* Do the srcu_read_lock() for the above synchronize_srcu(). */
902 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
905 current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
909 /* Do the srcu_read_unlock() for the above synchronize_srcu(). */
910 void exit_tasks_rcu_finish(void) __releases(&tasks_rcu_exit_srcu)
912 struct task_struct *t = current;
915 __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
917 exit_tasks_rcu_finish_trace(t);
920 #else /* #ifdef CONFIG_TASKS_RCU */
921 void exit_tasks_rcu_start(void) { }
922 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
923 #endif /* #else #ifdef CONFIG_TASKS_RCU */
925 #ifdef CONFIG_TASKS_RUDE_RCU
927 ////////////////////////////////////////////////////////////////////////
929 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
930 // passing an empty function to schedule_on_each_cpu(). This approach
931 // provides an asynchronous call_rcu_tasks_rude() API and batching of
932 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
933 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
934 // and induces otherwise unnecessary context switches on all online CPUs,
935 // whether idle or not.
937 // Callback handling is provided by the rcu_tasks_kthread() function.
939 // Ordering is provided by the scheduler's context-switch code.
941 // Empty function to allow workqueues to force a context switch.
942 static void rcu_tasks_be_rude(struct work_struct *work)
946 // Wait for one rude RCU-tasks grace period.
947 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
949 rtp->n_ipis += cpumask_weight(cpu_online_mask);
950 schedule_on_each_cpu(rcu_tasks_be_rude);
953 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
954 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
958 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
959 * @rhp: structure to be used for queueing the RCU updates.
960 * @func: actual callback function to be invoked after the grace period
962 * The callback function will be invoked some time after a full grace
963 * period elapses, in other words after all currently executing RCU
964 * read-side critical sections have completed. call_rcu_tasks_rude()
965 * assumes that the read-side critical sections end at context switch,
966 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
967 * usermode execution is schedulable). As such, there are no read-side
968 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
969 * this primitive is intended to determine that all tasks have passed
970 * through a safe state, not so much for data-structure synchronization.
972 * See the description of call_rcu() for more detailed information on
973 * memory ordering guarantees.
975 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
977 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
979 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
982 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
984 * Control will return to the caller some time after a rude rcu-tasks
985 * grace period has elapsed, in other words after all currently
986 * executing rcu-tasks read-side critical sections have elapsed. These
987 * read-side critical sections are delimited by calls to schedule(),
988 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
989 * context), and (in theory, anyway) cond_resched().
991 * This is a very specialized primitive, intended only for a few uses in
992 * tracing and other situations requiring manipulation of function preambles
993 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
994 * (yet) intended for heavy use from multiple CPUs.
996 * See the description of synchronize_rcu() for more detailed information
997 * on memory ordering guarantees.
999 void synchronize_rcu_tasks_rude(void)
1001 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1003 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1006 * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1008 * Although the current implementation is guaranteed to wait, it is not
1009 * obligated to, for example, if there are no pending callbacks.
1011 void rcu_barrier_tasks_rude(void)
1013 rcu_barrier_tasks_generic(&rcu_tasks_rude);
1015 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1017 static int __init rcu_spawn_tasks_rude_kthread(void)
1019 cblist_init_generic(&rcu_tasks_rude);
1020 rcu_tasks_rude.gp_sleep = HZ / 10;
1021 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1025 #if !defined(CONFIG_TINY_RCU)
1026 void show_rcu_tasks_rude_gp_kthread(void)
1028 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1030 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1031 #endif // !defined(CONFIG_TINY_RCU)
1032 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1034 ////////////////////////////////////////////////////////////////////////
1036 // Tracing variant of Tasks RCU. This variant is designed to be used
1037 // to protect tracing hooks, including those of BPF. This variant
1040 // 1. Has explicit read-side markers to allow finite grace periods
1041 // in the face of in-kernel loops for PREEMPT=n builds.
1043 // 2. Protects code in the idle loop, exception entry/exit, and
1044 // CPU-hotplug code paths, similar to the capabilities of SRCU.
1046 // 3. Avoids expensive read-side instructions, having overhead similar
1047 // to that of Preemptible RCU.
1049 // There are of course downsides. The grace-period code can send IPIs to
1050 // CPUs, even when those CPUs are in the idle loop or in nohz_full userspace.
1051 // It is necessary to scan the full tasklist, much as for Tasks RCU. There
1052 // is a single callback queue guarded by a single lock, again, much as for
1053 // Tasks RCU. If needed, these downsides can be at least partially remedied.
1055 // Perhaps most important, this variant of RCU does not affect the vanilla
1056 // flavors, rcu_preempt and rcu_sched. The fact that RCU Tasks Trace
1057 // readers can operate from idle, offline, and exception entry/exit in no
1058 // way allows rcu_preempt and rcu_sched readers to also do so.
1060 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1061 // pointers in the rcu_tasks structure. The rcu_spawn_tasks_trace_kthread()
1062 // function sets these function pointers up so that rcu_tasks_wait_gp()
1063 // invokes these functions in this order:
1065 // rcu_tasks_trace_pregp_step():
1066 // Initialize the count of readers and block CPU-hotplug operations.
1067 // rcu_tasks_trace_pertask(), invoked on every non-idle task:
1068 // Initialize per-task state and attempt to identify an immediate
1069 // quiescent state for that task, or, failing that, attempt to
1070 // set that task's .need_qs flag so that task's next outermost
1071 // rcu_read_unlock_trace() will report the quiescent state (in which
1072 // case the count of readers is incremented). If both attempts fail,
1073 // the task is added to a "holdout" list. Note that IPIs are used
1074 // to invoke trc_read_check_handler() in the context of running tasks
1075 // in order to avoid ordering overhead on common-case shared-variable
1077 // rcu_tasks_trace_postscan():
1078 // Initialize state and attempt to identify an immediate quiescent
1079 // state as above (but only for idle tasks), unblock CPU-hotplug
1080 // operations, and wait for an RCU grace period to avoid races with
1081 // tasks that are in the process of exiting.
1082 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1083 // Scans the holdout list, attempting to identify a quiescent state
1084 // for each task on the list. If there is a quiescent state, the
1085 // corresponding task is removed from the holdout list.
1086 // rcu_tasks_trace_postgp():
1087 // Wait for the count of readers do drop to zero, reporting any stalls.
1088 // Also execute full memory barriers to maintain ordering with code
1089 // executing after the grace period.
1091 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1093 // Pre-grace-period update-side code is ordered before the grace
1094 // period via the ->cbs_lock and barriers in rcu_tasks_kthread().
1095 // Pre-grace-period read-side code is ordered before the grace period by
1096 // atomic_dec_and_test() of the count of readers (for IPIed readers) and by
1097 // scheduler context-switch ordering (for locked-down non-running readers).
1099 // The lockdep state must be outside of #ifdef to be useful.
1100 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1101 static struct lock_class_key rcu_lock_trace_key;
1102 struct lockdep_map rcu_trace_lock_map =
1103 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1104 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1105 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1107 #ifdef CONFIG_TASKS_TRACE_RCU
1109 static atomic_t trc_n_readers_need_end; // Number of waited-for readers.
1110 static DECLARE_WAIT_QUEUE_HEAD(trc_wait); // List of holdout tasks.
1112 // Record outstanding IPIs to each CPU. No point in sending two...
1113 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1115 // The number of detections of task quiescent state relying on
1116 // heavyweight readers executing explicit memory barriers.
1117 static unsigned long n_heavy_reader_attempts;
1118 static unsigned long n_heavy_reader_updates;
1119 static unsigned long n_heavy_reader_ofl_updates;
1121 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1122 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1126 * This irq_work handler allows rcu_read_unlock_trace() to be invoked
1127 * while the scheduler locks are held.
1129 static void rcu_read_unlock_iw(struct irq_work *iwp)
1133 static DEFINE_IRQ_WORK(rcu_tasks_trace_iw, rcu_read_unlock_iw);
1135 /* If we are the last reader, wake up the grace-period kthread. */
1136 void rcu_read_unlock_trace_special(struct task_struct *t)
1138 int nq = READ_ONCE(t->trc_reader_special.b.need_qs);
1140 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) &&
1141 t->trc_reader_special.b.need_mb)
1142 smp_mb(); // Pairs with update-side barriers.
1143 // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1145 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
1146 WRITE_ONCE(t->trc_reader_nesting, 0);
1147 if (nq && atomic_dec_and_test(&trc_n_readers_need_end))
1148 irq_work_queue(&rcu_tasks_trace_iw);
1150 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1152 /* Add a task to the holdout list, if it is not already on the list. */
1153 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1155 if (list_empty(&t->trc_holdout_list)) {
1157 list_add(&t->trc_holdout_list, bhp);
1161 /* Remove a task from the holdout list, if it is in fact present. */
1162 static void trc_del_holdout(struct task_struct *t)
1164 if (!list_empty(&t->trc_holdout_list)) {
1165 list_del_init(&t->trc_holdout_list);
1170 /* IPI handler to check task state. */
1171 static void trc_read_check_handler(void *t_in)
1173 struct task_struct *t = current;
1174 struct task_struct *texp = t_in;
1176 // If the task is no longer running on this CPU, leave.
1177 if (unlikely(texp != t)) {
1178 goto reset_ipi; // Already on holdout list, so will check later.
1181 // If the task is not in a read-side critical section, and
1182 // if this is the last reader, awaken the grace-period kthread.
1183 if (likely(!READ_ONCE(t->trc_reader_nesting))) {
1184 WRITE_ONCE(t->trc_reader_checked, true);
1187 // If we are racing with an rcu_read_unlock_trace(), try again later.
1188 if (unlikely(READ_ONCE(t->trc_reader_nesting) < 0))
1190 WRITE_ONCE(t->trc_reader_checked, true);
1192 // Get here if the task is in a read-side critical section. Set
1193 // its state so that it will awaken the grace-period kthread upon
1194 // exit from that critical section.
1195 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
1196 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
1197 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
1200 // Allow future IPIs to be sent on CPU and for task.
1201 // Also order this IPI handler against any later manipulations of
1202 // the intended task.
1203 smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1204 smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1207 /* Callback function for scheduler to check locked-down task. */
1208 static int trc_inspect_reader(struct task_struct *t, void *arg)
1210 int cpu = task_cpu(t);
1212 bool ofl = cpu_is_offline(cpu);
1215 WARN_ON_ONCE(ofl && !is_idle_task(t));
1217 // If no chance of heavyweight readers, do it the hard way.
1218 if (!ofl && !IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1221 // If heavyweight readers are enabled on the remote task,
1222 // we can inspect its state despite its currently running.
1223 // However, we cannot safely change its state.
1224 n_heavy_reader_attempts++;
1225 if (!ofl && // Check for "running" idle tasks on offline CPUs.
1226 !rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1227 return -EINVAL; // No quiescent state, do it the hard way.
1228 n_heavy_reader_updates++;
1230 n_heavy_reader_ofl_updates++;
1233 // The task is not running, so C-language access is safe.
1234 nesting = t->trc_reader_nesting;
1237 // If not exiting a read-side critical section, mark as checked
1238 // so that the grace-period kthread will remove it from the
1240 t->trc_reader_checked = nesting >= 0;
1242 return nesting ? -EINVAL : 0; // If in QS, done, otherwise try again later.
1244 // The task is in a read-side critical section, so set up its
1245 // state so that it will awaken the grace-period kthread upon exit
1246 // from that critical section.
1247 atomic_inc(&trc_n_readers_need_end); // One more to wait on.
1248 WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs));
1249 WRITE_ONCE(t->trc_reader_special.b.need_qs, true);
1253 /* Attempt to extract the state for the specified task. */
1254 static void trc_wait_for_one_reader(struct task_struct *t,
1255 struct list_head *bhp)
1259 // If a previous IPI is still in flight, let it complete.
1260 if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1263 // The current task had better be in a quiescent state.
1265 t->trc_reader_checked = true;
1266 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1270 // Attempt to nail down the task for inspection.
1272 if (!task_call_func(t, trc_inspect_reader, NULL)) {
1278 // If this task is not yet on the holdout list, then we are in
1279 // an RCU read-side critical section. Otherwise, the invocation of
1280 // trc_add_holdout() that added it to the list did the necessary
1281 // get_task_struct(). Either way, the task cannot be freed out
1282 // from under this code.
1284 // If currently running, send an IPI, either way, add to list.
1285 trc_add_holdout(t, bhp);
1287 time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1288 // The task is currently running, so try IPIing it.
1291 // If there is already an IPI outstanding, let it happen.
1292 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1295 per_cpu(trc_ipi_to_cpu, cpu) = true;
1296 t->trc_ipi_to_cpu = cpu;
1297 rcu_tasks_trace.n_ipis++;
1298 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1299 // Just in case there is some other reason for
1300 // failure than the target CPU being offline.
1301 WARN_ONCE(1, "%s(): smp_call_function_single() failed for CPU: %d\n",
1303 rcu_tasks_trace.n_ipis_fails++;
1304 per_cpu(trc_ipi_to_cpu, cpu) = false;
1305 t->trc_ipi_to_cpu = -1;
1310 /* Initialize for a new RCU-tasks-trace grace period. */
1311 static void rcu_tasks_trace_pregp_step(void)
1315 // Allow for fast-acting IPIs.
1316 atomic_set(&trc_n_readers_need_end, 1);
1318 // There shouldn't be any old IPIs, but...
1319 for_each_possible_cpu(cpu)
1320 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1322 // Disable CPU hotplug across the tasklist scan.
1323 // This also waits for all readers in CPU-hotplug code paths.
1327 /* Do first-round processing for the specified task. */
1328 static void rcu_tasks_trace_pertask(struct task_struct *t,
1329 struct list_head *hop)
1331 // During early boot when there is only the one boot CPU, there
1332 // is no idle task for the other CPUs. Just return.
1333 if (unlikely(t == NULL))
1336 WRITE_ONCE(t->trc_reader_special.b.need_qs, false);
1337 WRITE_ONCE(t->trc_reader_checked, false);
1338 t->trc_ipi_to_cpu = -1;
1339 trc_wait_for_one_reader(t, hop);
1343 * Do intermediate processing between task and holdout scans and
1344 * pick up the idle tasks.
1346 static void rcu_tasks_trace_postscan(struct list_head *hop)
1350 for_each_possible_cpu(cpu)
1351 rcu_tasks_trace_pertask(idle_task(cpu), hop);
1353 // Re-enable CPU hotplug now that the tasklist scan has completed.
1356 // Wait for late-stage exiting tasks to finish exiting.
1357 // These might have passed the call to exit_tasks_rcu_finish().
1359 // Any tasks that exit after this point will set ->trc_reader_checked.
1362 /* Communicate task state back to the RCU tasks trace stall warning request. */
1363 struct trc_stall_chk_rdr {
1369 static int trc_check_slow_task(struct task_struct *t, void *arg)
1371 struct trc_stall_chk_rdr *trc_rdrp = arg;
1374 return false; // It is running, so decline to inspect it.
1375 trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1376 trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1377 trc_rdrp->needqs = READ_ONCE(t->trc_reader_special.b.need_qs);
1381 /* Show the state of a task stalling the current RCU tasks trace GP. */
1382 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1385 struct trc_stall_chk_rdr trc_rdr;
1386 bool is_idle_tsk = is_idle_task(t);
1389 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1390 *firstreport = false;
1393 if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1394 pr_alert("P%d: %c\n",
1398 pr_alert("P%d: %c%c%c nesting: %d%c cpu: %d\n",
1400 ".I"[trc_rdr.ipi_to_cpu >= 0],
1402 ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1404 " N"[!!trc_rdr.needqs],
1409 /* List stalled IPIs for RCU tasks trace. */
1410 static void show_stalled_ipi_trace(void)
1414 for_each_possible_cpu(cpu)
1415 if (per_cpu(trc_ipi_to_cpu, cpu))
1416 pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1419 /* Do one scan of the holdout list. */
1420 static void check_all_holdout_tasks_trace(struct list_head *hop,
1421 bool needreport, bool *firstreport)
1423 struct task_struct *g, *t;
1425 // Disable CPU hotplug across the holdout list scan.
1428 list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1429 // If safe and needed, try to check the current task.
1430 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1431 !READ_ONCE(t->trc_reader_checked))
1432 trc_wait_for_one_reader(t, hop);
1434 // If check succeeded, remove this task from the list.
1435 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1436 READ_ONCE(t->trc_reader_checked))
1438 else if (needreport)
1439 show_stalled_task_trace(t, firstreport);
1442 // Re-enable CPU hotplug now that the holdout list scan has completed.
1447 pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1448 show_stalled_ipi_trace();
1452 static void rcu_tasks_trace_empty_fn(void *unused)
1456 /* Wait for grace period to complete and provide ordering. */
1457 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1461 struct task_struct *g, *t;
1462 LIST_HEAD(holdouts);
1465 // Wait for any lingering IPI handlers to complete. Note that
1466 // if a CPU has gone offline or transitioned to userspace in the
1467 // meantime, all IPI handlers should have been drained beforehand.
1468 // Yes, this assumes that CPUs process IPIs in order. If that ever
1469 // changes, there will need to be a recheck and/or timed wait.
1470 for_each_online_cpu(cpu)
1471 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1472 smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1474 // Remove the safety count.
1475 smp_mb__before_atomic(); // Order vs. earlier atomics
1476 atomic_dec(&trc_n_readers_need_end);
1477 smp_mb__after_atomic(); // Order vs. later atomics
1479 // Wait for readers.
1480 set_tasks_gp_state(rtp, RTGS_WAIT_READERS);
1482 ret = wait_event_idle_exclusive_timeout(
1484 atomic_read(&trc_n_readers_need_end) == 0,
1485 READ_ONCE(rcu_task_stall_timeout));
1487 break; // Count reached zero.
1488 // Stall warning time, so make a list of the offenders.
1490 for_each_process_thread(g, t)
1491 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1492 trc_add_holdout(t, &holdouts);
1495 list_for_each_entry_safe(t, g, &holdouts, trc_holdout_list) {
1496 if (READ_ONCE(t->trc_reader_special.b.need_qs))
1497 show_stalled_task_trace(t, &firstreport);
1498 trc_del_holdout(t); // Release task_struct reference.
1501 pr_err("INFO: rcu_tasks_trace detected stalls? (Counter/taskslist mismatch?)\n");
1502 show_stalled_ipi_trace();
1503 pr_err("\t%d holdouts\n", atomic_read(&trc_n_readers_need_end));
1505 smp_mb(); // Caller's code must be ordered after wakeup.
1506 // Pairs with pretty much every ordering primitive.
1509 /* Report any needed quiescent state for this exiting task. */
1510 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1512 WRITE_ONCE(t->trc_reader_checked, true);
1513 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1514 WRITE_ONCE(t->trc_reader_nesting, 0);
1515 if (WARN_ON_ONCE(READ_ONCE(t->trc_reader_special.b.need_qs)))
1516 rcu_read_unlock_trace_special(t);
1520 * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1521 * @rhp: structure to be used for queueing the RCU updates.
1522 * @func: actual callback function to be invoked after the grace period
1524 * The callback function will be invoked some time after a trace rcu-tasks
1525 * grace period elapses, in other words after all currently executing
1526 * trace rcu-tasks read-side critical sections have completed. These
1527 * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1528 * and rcu_read_unlock_trace().
1530 * See the description of call_rcu() for more detailed information on
1531 * memory ordering guarantees.
1533 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1535 call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1537 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1540 * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1542 * Control will return to the caller some time after a trace rcu-tasks
1543 * grace period has elapsed, in other words after all currently executing
1544 * trace rcu-tasks read-side critical sections have elapsed. These read-side
1545 * critical sections are delimited by calls to rcu_read_lock_trace()
1546 * and rcu_read_unlock_trace().
1548 * This is a very specialized primitive, intended only for a few uses in
1549 * tracing and other situations requiring manipulation of function preambles
1550 * and profiling hooks. The synchronize_rcu_tasks_trace() function is not
1551 * (yet) intended for heavy use from multiple CPUs.
1553 * See the description of synchronize_rcu() for more detailed information
1554 * on memory ordering guarantees.
1556 void synchronize_rcu_tasks_trace(void)
1558 RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1559 synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1561 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1564 * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1566 * Although the current implementation is guaranteed to wait, it is not
1567 * obligated to, for example, if there are no pending callbacks.
1569 void rcu_barrier_tasks_trace(void)
1571 rcu_barrier_tasks_generic(&rcu_tasks_trace);
1573 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1575 static int __init rcu_spawn_tasks_trace_kthread(void)
1577 cblist_init_generic(&rcu_tasks_trace);
1578 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1579 rcu_tasks_trace.gp_sleep = HZ / 10;
1580 rcu_tasks_trace.init_fract = HZ / 10;
1582 rcu_tasks_trace.gp_sleep = HZ / 200;
1583 if (rcu_tasks_trace.gp_sleep <= 0)
1584 rcu_tasks_trace.gp_sleep = 1;
1585 rcu_tasks_trace.init_fract = HZ / 200;
1586 if (rcu_tasks_trace.init_fract <= 0)
1587 rcu_tasks_trace.init_fract = 1;
1589 rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1590 rcu_tasks_trace.pertask_func = rcu_tasks_trace_pertask;
1591 rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1592 rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1593 rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1594 rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1598 #if !defined(CONFIG_TINY_RCU)
1599 void show_rcu_tasks_trace_gp_kthread(void)
1603 sprintf(buf, "N%d h:%lu/%lu/%lu", atomic_read(&trc_n_readers_need_end),
1604 data_race(n_heavy_reader_ofl_updates),
1605 data_race(n_heavy_reader_updates),
1606 data_race(n_heavy_reader_attempts));
1607 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1609 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1610 #endif // !defined(CONFIG_TINY_RCU)
1612 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1613 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1614 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1616 #ifndef CONFIG_TINY_RCU
1617 void show_rcu_tasks_gp_kthreads(void)
1619 show_rcu_tasks_classic_gp_kthread();
1620 show_rcu_tasks_rude_gp_kthread();
1621 show_rcu_tasks_trace_gp_kthread();
1623 #endif /* #ifndef CONFIG_TINY_RCU */
1625 #ifdef CONFIG_PROVE_RCU
1626 struct rcu_tasks_test_desc {
1632 static struct rcu_tasks_test_desc tests[] = {
1634 .name = "call_rcu_tasks()",
1635 /* If not defined, the test is skipped. */
1636 .notrun = !IS_ENABLED(CONFIG_TASKS_RCU),
1639 .name = "call_rcu_tasks_rude()",
1640 /* If not defined, the test is skipped. */
1641 .notrun = !IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1644 .name = "call_rcu_tasks_trace()",
1645 /* If not defined, the test is skipped. */
1646 .notrun = !IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1650 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1652 struct rcu_tasks_test_desc *rttd =
1653 container_of(rhp, struct rcu_tasks_test_desc, rh);
1655 pr_info("Callback from %s invoked.\n", rttd->name);
1657 rttd->notrun = true;
1660 static void rcu_tasks_initiate_self_tests(void)
1662 pr_info("Running RCU-tasks wait API self tests\n");
1663 #ifdef CONFIG_TASKS_RCU
1664 synchronize_rcu_tasks();
1665 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1668 #ifdef CONFIG_TASKS_RUDE_RCU
1669 synchronize_rcu_tasks_rude();
1670 call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1673 #ifdef CONFIG_TASKS_TRACE_RCU
1674 synchronize_rcu_tasks_trace();
1675 call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1679 static int rcu_tasks_verify_self_tests(void)
1684 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1685 if (!tests[i].notrun) { // still hanging.
1686 pr_err("%s has been failed.\n", tests[i].name);
1696 late_initcall(rcu_tasks_verify_self_tests);
1697 #else /* #ifdef CONFIG_PROVE_RCU */
1698 static void rcu_tasks_initiate_self_tests(void) { }
1699 #endif /* #else #ifdef CONFIG_PROVE_RCU */
1701 void __init rcu_init_tasks_generic(void)
1703 #ifdef CONFIG_TASKS_RCU
1704 rcu_spawn_tasks_kthread();
1707 #ifdef CONFIG_TASKS_RUDE_RCU
1708 rcu_spawn_tasks_rude_kthread();
1711 #ifdef CONFIG_TASKS_TRACE_RCU
1712 rcu_spawn_tasks_trace_kthread();
1715 // Run the self-tests.
1716 rcu_tasks_initiate_self_tests();
1719 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1720 static inline void rcu_tasks_bootup_oddness(void) {}
1721 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */