projects/linux·blob master · refs

1625 lines · 53.7 KiB

1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Task-based RCU implementations.
4 *
5 * Copyright (C) 2020 Paul E. McKenney
6 */
7
8#ifdef CONFIG_TASKS_RCU_GENERIC
9#include "rcu_segcblist.h"
10
11////////////////////////////////////////////////////////////////////////
12//
13// Generic data structures.
14
15struct rcu_tasks;
16typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
17typedef void (*pregp_func_t)(struct list_head *hop);
18typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
19typedef void (*postscan_func_t)(struct list_head *hop);
20typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
21typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
22
23/**
24 * struct rcu_tasks_percpu - Per-CPU component of definition for a Tasks-RCU-like mechanism.
25 * @cblist: Callback list.
26 * @lock: Lock protecting per-CPU callback list.
27 * @rtp_jiffies: Jiffies counter value for statistics.
28 * @lazy_timer: Timer to unlazify callbacks.
29 * @urgent_gp: Number of additional non-lazy grace periods.
30 * @rtp_n_lock_retries: Rough lock-contention statistic.
31 * @rtp_work: Work queue for invoking callbacks.
32 * @rtp_irq_work: IRQ work queue for deferred wakeups.
33 * @barrier_q_head: RCU callback for barrier operation.
34 * @rtp_blkd_tasks: List of tasks blocked as readers.
35 * @rtp_exit_list: List of tasks in the latter portion of do_exit().
36 * @cpu: CPU number corresponding to this entry.
37 * @index: Index of this CPU in rtpcp_array of the rcu_tasks structure.
38 * @rtpp: Pointer to the rcu_tasks structure.
39 */
40struct rcu_tasks_percpu {
41 struct rcu_segcblist cblist;
42 raw_spinlock_t __private lock;
43 unsigned long rtp_jiffies;
44 unsigned long rtp_n_lock_retries;
45 struct timer_list lazy_timer;
46 unsigned int urgent_gp;
47 struct work_struct rtp_work;
48 struct irq_work rtp_irq_work;
49 struct rcu_head barrier_q_head;
50 struct list_head rtp_blkd_tasks;
51 struct list_head rtp_exit_list;
52 int cpu;
53 int index;
54 struct rcu_tasks *rtpp;
55};
56
57/**
58 * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
59 * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
60 * @cbs_gbl_lock: Lock protecting callback list.
61 * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
62 * @gp_func: This flavor's grace-period-wait function.
63 * @gp_state: Grace period's most recent state transition (debugging).
64 * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
65 * @init_fract: Initial backoff sleep interval.
66 * @gp_jiffies: Time of last @gp_state transition.
67 * @gp_start: Most recent grace-period start in jiffies.
68 * @tasks_gp_seq: Number of grace periods completed since boot in upper bits.
69 * @n_ipis: Number of IPIs sent to encourage grace periods to end.
70 * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
71 * @lazy_jiffies: Number of jiffies to allow callbacks to be lazy.
72 * @pregp_func: This flavor's pre-grace-period function (optional).
73 * @pertask_func: This flavor's per-task scan function (optional).
74 * @postscan_func: This flavor's post-task scan function (optional).
75 * @holdouts_func: This flavor's holdout-list scan function (optional).
76 * @postgp_func: This flavor's post-grace-period function (optional).
77 * @call_func: This flavor's call_rcu()-equivalent function.
78 * @wait_state: Task state for synchronous grace-period waits (default TASK_UNINTERRUPTIBLE).
79 * @rtpcpu: This flavor's rcu_tasks_percpu structure.
80 * @rtpcp_array: Array of pointers to rcu_tasks_percpu structure of CPUs in cpu_possible_mask.
81 * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
82 * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
83 * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
84 * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
85 * @barrier_q_mutex: Serialize barrier operations.
86 * @barrier_q_count: Number of queues being waited on.
87 * @barrier_q_completion: Barrier wait/wakeup mechanism.
88 * @barrier_q_seq: Sequence number for barrier operations.
89 * @barrier_q_start: Most recent barrier start in jiffies.
90 * @name: This flavor's textual name.
91 * @kname: This flavor's kthread name.
92 */
93struct rcu_tasks {
94 struct rcuwait cbs_wait;
95 raw_spinlock_t cbs_gbl_lock;
96 struct mutex tasks_gp_mutex;
97 int gp_state;
98 int gp_sleep;
99 int init_fract;
100 unsigned long gp_jiffies;
101 unsigned long gp_start;
102 unsigned long tasks_gp_seq;
103 unsigned long n_ipis;
104 struct task_struct *kthread_ptr;
105 unsigned long lazy_jiffies;
106 rcu_tasks_gp_func_t gp_func;
107 pregp_func_t pregp_func;
108 pertask_func_t pertask_func;
109 postscan_func_t postscan_func;
110 holdouts_func_t holdouts_func;
111 postgp_func_t postgp_func;
112 call_rcu_func_t call_func;
113 unsigned int wait_state;
114 struct rcu_tasks_percpu __percpu *rtpcpu;
115 struct rcu_tasks_percpu **rtpcp_array;
116 int percpu_enqueue_shift;
117 int percpu_enqueue_lim;
118 int percpu_dequeue_lim;
119 unsigned long percpu_dequeue_gpseq;
120 struct mutex barrier_q_mutex;
121 atomic_t barrier_q_count;
122 struct completion barrier_q_completion;
123 unsigned long barrier_q_seq;
124 unsigned long barrier_q_start;
125 char *name;
126 char *kname;
127};
128
129static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
130
131#define DEFINE_RCU_TASKS(rt_name, gp, call, n) \
132static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = { \
133 .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock), \
134 .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup), \
135}; \
136static struct rcu_tasks rt_name = \
137{ \
138 .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait), \
139 .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock), \
140 .tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex), \
141 .gp_func = gp, \
142 .call_func = call, \
143 .wait_state = TASK_UNINTERRUPTIBLE, \
144 .rtpcpu = &rt_name ## __percpu, \
145 .lazy_jiffies = DIV_ROUND_UP(HZ, 4), \
146 .name = n, \
147 .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS), \
148 .percpu_enqueue_lim = 1, \
149 .percpu_dequeue_lim = 1, \
150 .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex), \
151 .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT, \
152 .kname = #rt_name, \
153}
154
155#ifdef CONFIG_TASKS_RCU
156
157/* Report delay of scan exiting tasklist in rcu_tasks_postscan(). */
158static void tasks_rcu_exit_stall(struct timer_list *unused);
159static DEFINE_TIMER(tasks_rcu_exit_stall_timer, tasks_rcu_exit_stall);
160#endif
161
162/* Control stall timeouts. Disable with <= 0, otherwise jiffies till stall. */
163#define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
164#define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
165static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
166module_param(rcu_task_stall_timeout, int, 0644);
167#define RCU_TASK_STALL_INFO (HZ * 10)
168static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
169module_param(rcu_task_stall_info, int, 0644);
170static int rcu_task_stall_info_mult __read_mostly = 3;
171module_param(rcu_task_stall_info_mult, int, 0444);
172
173static int rcu_task_enqueue_lim __read_mostly = -1;
174module_param(rcu_task_enqueue_lim, int, 0444);
175
176static bool rcu_task_cb_adjust;
177static int rcu_task_contend_lim __read_mostly = 100;
178module_param(rcu_task_contend_lim, int, 0444);
179static int rcu_task_collapse_lim __read_mostly = 10;
180module_param(rcu_task_collapse_lim, int, 0444);
181static int rcu_task_lazy_lim __read_mostly = 32;
182module_param(rcu_task_lazy_lim, int, 0444);
183
184static int rcu_task_cpu_ids;
185
186/* RCU tasks grace-period state for debugging. */
187#define RTGS_INIT 0
188#define RTGS_WAIT_WAIT_CBS 1
189#define RTGS_WAIT_GP 2
190#define RTGS_PRE_WAIT_GP 3
191#define RTGS_SCAN_TASKLIST 4
192#define RTGS_POST_SCAN_TASKLIST 5
193#define RTGS_WAIT_SCAN_HOLDOUTS 6
194#define RTGS_SCAN_HOLDOUTS 7
195#define RTGS_POST_GP 8
196#define RTGS_WAIT_READERS 9
197#define RTGS_INVOKE_CBS 10
198#define RTGS_WAIT_CBS 11
199#ifndef CONFIG_TINY_RCU
200static const char * const rcu_tasks_gp_state_names[] = {
201 "RTGS_INIT",
202 "RTGS_WAIT_WAIT_CBS",
203 "RTGS_WAIT_GP",
204 "RTGS_PRE_WAIT_GP",
205 "RTGS_SCAN_TASKLIST",
206 "RTGS_POST_SCAN_TASKLIST",
207 "RTGS_WAIT_SCAN_HOLDOUTS",
208 "RTGS_SCAN_HOLDOUTS",
209 "RTGS_POST_GP",
210 "RTGS_WAIT_READERS",
211 "RTGS_INVOKE_CBS",
212 "RTGS_WAIT_CBS",
213};
214#endif /* #ifndef CONFIG_TINY_RCU */
215
216////////////////////////////////////////////////////////////////////////
217//
218// Generic code.
219
220static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
221
222/* Record grace-period phase and time. */
223static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
224{
225 rtp->gp_state = newstate;
226 rtp->gp_jiffies = jiffies;
227}
228
229#ifndef CONFIG_TINY_RCU
230/* Return state name. */
231static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
232{
233 int i = data_race(rtp->gp_state); // Let KCSAN detect update races
234 int j = READ_ONCE(i); // Prevent the compiler from reading twice
235
236 if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
237 return "???";
238 return rcu_tasks_gp_state_names[j];
239}
240#endif /* #ifndef CONFIG_TINY_RCU */
241
242// Initialize per-CPU callback lists for the specified flavor of
243// Tasks RCU. Do not enqueue callbacks before this function is invoked.
244static void cblist_init_generic(struct rcu_tasks *rtp)
245{
246 int cpu;
247 int lim;
248 int shift;
249 int maxcpu;
250 int index = 0;
251
252 if (rcu_task_enqueue_lim < 0) {
253 rcu_task_enqueue_lim = 1;
254 rcu_task_cb_adjust = true;
255 } else if (rcu_task_enqueue_lim == 0) {
256 rcu_task_enqueue_lim = 1;
257 }
258 lim = rcu_task_enqueue_lim;
259
260 rtp->rtpcp_array = kzalloc_objs(struct rcu_tasks_percpu *,
261 num_possible_cpus());
262 BUG_ON(!rtp->rtpcp_array);
263
264 for_each_possible_cpu(cpu) {
265 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
266
267 WARN_ON_ONCE(!rtpcp);
268 if (cpu)
269 raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
270 if (rcu_segcblist_empty(&rtpcp->cblist))
271 rcu_segcblist_init(&rtpcp->cblist);
272 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
273 rtpcp->cpu = cpu;
274 rtpcp->rtpp = rtp;
275 rtpcp->index = index;
276 rtp->rtpcp_array[index] = rtpcp;
277 index++;
278 if (!rtpcp->rtp_blkd_tasks.next)
279 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
280 if (!rtpcp->rtp_exit_list.next)
281 INIT_LIST_HEAD(&rtpcp->rtp_exit_list);
282 rtpcp->barrier_q_head.next = &rtpcp->barrier_q_head;
283 maxcpu = cpu;
284 }
285
286 rcu_task_cpu_ids = maxcpu + 1;
287 if (lim > rcu_task_cpu_ids)
288 lim = rcu_task_cpu_ids;
289 shift = ilog2(rcu_task_cpu_ids / lim);
290 if (((rcu_task_cpu_ids - 1) >> shift) >= lim)
291 shift++;
292 rtp->percpu_enqueue_shift = shift;
293 rtp->percpu_dequeue_lim = lim;
294 rtp->percpu_enqueue_lim = lim;
295
296 pr_info("%s: Setting shift to %d and lim to %d rcu_task_cb_adjust=%d rcu_task_cpu_ids=%d.\n",
297 rtp->name, data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim),
298 rcu_task_cb_adjust, rcu_task_cpu_ids);
299}
300
301// Compute wakeup time for lazy callback timer.
302static unsigned long rcu_tasks_lazy_time(struct rcu_tasks *rtp)
303{
304 return jiffies + rtp->lazy_jiffies;
305}
306
307// Timer handler that unlazifies lazy callbacks.
308static void call_rcu_tasks_generic_timer(struct timer_list *tlp)
309{
310 unsigned long flags;
311 bool needwake = false;
312 struct rcu_tasks *rtp;
313 struct rcu_tasks_percpu *rtpcp = timer_container_of(rtpcp, tlp,
314 lazy_timer);
315
316 rtp = rtpcp->rtpp;
317 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
318 if (!rcu_segcblist_empty(&rtpcp->cblist) && rtp->lazy_jiffies) {
319 if (!rtpcp->urgent_gp)
320 rtpcp->urgent_gp = 1;
321 needwake = true;
322 mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
323 }
324 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
325 if (needwake)
326 rcuwait_wake_up(&rtp->cbs_wait);
327}
328
329// IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
330static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
331{
332 struct rcu_tasks *rtp;
333 struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
334
335 rtp = rtpcp->rtpp;
336 rcuwait_wake_up(&rtp->cbs_wait);
337}
338
339// Enqueue a callback for the specified flavor of Tasks RCU.
340static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
341 struct rcu_tasks *rtp)
342{
343 int chosen_cpu;
344 unsigned long flags;
345 bool havekthread = smp_load_acquire(&rtp->kthread_ptr);
346 int ideal_cpu;
347 unsigned long j;
348 bool needadjust = false;
349 bool needwake;
350 struct rcu_tasks_percpu *rtpcp;
351
352 rhp->next = NULL;
353 rhp->func = func;
354 local_irq_save(flags);
355 rcu_read_lock();
356 ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
357 chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
358 WARN_ON_ONCE(chosen_cpu >= rcu_task_cpu_ids);
359 rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
360 if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
361 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
362 j = jiffies;
363 if (rtpcp->rtp_jiffies != j) {
364 rtpcp->rtp_jiffies = j;
365 rtpcp->rtp_n_lock_retries = 0;
366 }
367 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
368 READ_ONCE(rtp->percpu_enqueue_lim) != rcu_task_cpu_ids)
369 needadjust = true; // Defer adjustment to avoid deadlock.
370 }
371 // Queuing callbacks before initialization not yet supported.
372 if (WARN_ON_ONCE(!rcu_segcblist_is_enabled(&rtpcp->cblist)))
373 rcu_segcblist_init(&rtpcp->cblist);
374 needwake = (!havekthread && rcu_segcblist_empty(&rtpcp->cblist)) ||
375 (func == wakeme_after_rcu) ||
376 (rcu_segcblist_n_cbs(&rtpcp->cblist) == rcu_task_lazy_lim);
377 if (havekthread && !needwake && !timer_pending(&rtpcp->lazy_timer)) {
378 if (rtp->lazy_jiffies)
379 mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
380 else
381 needwake = rcu_segcblist_empty(&rtpcp->cblist);
382 }
383 if (needwake)
384 rtpcp->urgent_gp = 3;
385 rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
386 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
387 if (unlikely(needadjust)) {
388 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
389 if (rtp->percpu_enqueue_lim != rcu_task_cpu_ids) {
390 WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
391 WRITE_ONCE(rtp->percpu_dequeue_lim, rcu_task_cpu_ids);
392 smp_store_release(&rtp->percpu_enqueue_lim, rcu_task_cpu_ids);
393 pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
394 }
395 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
396 }
397 rcu_read_unlock();
398 // We can't create the kthread with interrupts disabled because a
399 // scheduler spinlock might be held, so kthread creation is deferred
400 // until core_initcall() time. Similarly, wakeups are deferred using
401 // irq_work in order to avoid potential scheduler-lock-deadlock
402 // lockdep splats.
403 if (needwake && READ_ONCE(rtp->kthread_ptr))
404 irq_work_queue(&rtpcp->rtp_irq_work);
405}
406
407// RCU callback function for rcu_barrier_tasks_generic().
408static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
409{
410 struct rcu_tasks *rtp;
411 struct rcu_tasks_percpu *rtpcp;
412
413 rhp->next = rhp; // Mark the callback as having been invoked.
414 rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
415 rtp = rtpcp->rtpp;
416 if (atomic_dec_and_test(&rtp->barrier_q_count))
417 complete(&rtp->barrier_q_completion);
418}
419
420// Wait for all in-flight callbacks for the specified RCU Tasks flavor.
421// Operates in a manner similar to rcu_barrier().
422static void __maybe_unused rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
423{
424 int cpu;
425 unsigned long flags;
426 struct rcu_tasks_percpu *rtpcp;
427 unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
428
429 mutex_lock(&rtp->barrier_q_mutex);
430 if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
431 smp_mb();
432 mutex_unlock(&rtp->barrier_q_mutex);
433 return;
434 }
435 rtp->barrier_q_start = jiffies;
436 rcu_seq_start(&rtp->barrier_q_seq);
437 init_completion(&rtp->barrier_q_completion);
438 atomic_set(&rtp->barrier_q_count, 2);
439 for_each_possible_cpu(cpu) {
440 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
441 break;
442 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
443 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
444 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
445 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
446 atomic_inc(&rtp->barrier_q_count);
447 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
448 }
449 if (atomic_sub_and_test(2, &rtp->barrier_q_count))
450 complete(&rtp->barrier_q_completion);
451 wait_for_completion(&rtp->barrier_q_completion);
452 rcu_seq_end(&rtp->barrier_q_seq);
453 mutex_unlock(&rtp->barrier_q_mutex);
454}
455
456// Advance callbacks and indicate whether either a grace period or
457// callback invocation is needed.
458static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
459{
460 int cpu;
461 int dequeue_limit;
462 unsigned long flags;
463 bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
464 long n;
465 long ncbs = 0;
466 long ncbsnz = 0;
467 int needgpcb = 0;
468
469 dequeue_limit = smp_load_acquire(&rtp->percpu_dequeue_lim);
470 for (cpu = 0; cpu < dequeue_limit; cpu++) {
471 if (!cpu_possible(cpu))
472 continue;
473 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
474
475 /* Advance and accelerate any new callbacks. */
476 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
477 continue;
478 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
479 // Should we shrink down to a single callback queue?
480 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
481 if (n) {
482 ncbs += n;
483 if (cpu > 0)
484 ncbsnz += n;
485 }
486 srcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
487 (void)srcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
488 if (rtpcp->urgent_gp > 0 && rcu_segcblist_pend_cbs(&rtpcp->cblist)) {
489 if (rtp->lazy_jiffies)
490 rtpcp->urgent_gp--;
491 needgpcb |= 0x3;
492 } else if (rcu_segcblist_empty(&rtpcp->cblist)) {
493 rtpcp->urgent_gp = 0;
494 }
495 if (rcu_segcblist_ready_cbs(&rtpcp->cblist))
496 needgpcb |= 0x1;
497 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
498 }
499
500 // Shrink down to a single callback queue if appropriate.
501 // This is done in two stages: (1) If there are no more than
502 // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
503 // CPU, limit enqueueing to CPU 0. (2) After an RCU grace period,
504 // if there has not been an increase in callbacks, limit dequeuing
505 // to CPU 0. Note the matching RCU read-side critical section in
506 // call_rcu_tasks_generic().
507 if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
508 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
509 if (rtp->percpu_enqueue_lim > 1) {
510 WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(rcu_task_cpu_ids));
511 smp_store_release(&rtp->percpu_enqueue_lim, 1);
512 rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
513 gpdone = false;
514 pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
515 }
516 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
517 }
518 if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
519 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
520 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
521 WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
522 pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
523 }
524 if (rtp->percpu_dequeue_lim == 1) {
525 for (cpu = rtp->percpu_dequeue_lim; cpu < rcu_task_cpu_ids; cpu++) {
526 if (!cpu_possible(cpu))
527 continue;
528 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
529
530 WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
531 }
532 }
533 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
534 }
535
536 return needgpcb;
537}
538
539// Advance callbacks and invoke any that are ready.
540static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
541{
542 int cpuwq;
543 unsigned long flags;
544 int len;
545 int index;
546 struct rcu_head *rhp;
547 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
548 struct rcu_tasks_percpu *rtpcp_next;
549
550 index = rtpcp->index * 2 + 1;
551 if (index < num_possible_cpus()) {
552 rtpcp_next = rtp->rtpcp_array[index];
553 if (rtpcp_next->cpu < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
554 cpuwq = rcu_cpu_beenfullyonline(rtpcp_next->cpu) ? rtpcp_next->cpu : WORK_CPU_UNBOUND;
555 queue_work_on(cpuwq, system_percpu_wq, &rtpcp_next->rtp_work);
556 index++;
557 if (index < num_possible_cpus()) {
558 rtpcp_next = rtp->rtpcp_array[index];
559 if (rtpcp_next->cpu < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
560 cpuwq = rcu_cpu_beenfullyonline(rtpcp_next->cpu) ? rtpcp_next->cpu : WORK_CPU_UNBOUND;
561 queue_work_on(cpuwq, system_percpu_wq, &rtpcp_next->rtp_work);
562 }
563 }
564 }
565 }
566
567 if (rcu_segcblist_empty(&rtpcp->cblist))
568 return;
569 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
570 srcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
571 rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
572 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
573 len = rcl.len;
574 for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
575 debug_rcu_head_callback(rhp);
576 local_bh_disable();
577 rhp->func(rhp);
578 local_bh_enable();
579 cond_resched();
580 }
581 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
582 rcu_segcblist_add_len(&rtpcp->cblist, -len);
583 (void)srcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
584 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
585}
586
587// Workqueue flood to advance callbacks and invoke any that are ready.
588static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
589{
590 struct rcu_tasks *rtp;
591 struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
592
593 rtp = rtpcp->rtpp;
594 rcu_tasks_invoke_cbs(rtp, rtpcp);
595}
596
597// Wait for one grace period.
598static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
599{
600 int needgpcb;
601
602 mutex_lock(&rtp->tasks_gp_mutex);
603
604 // If there were none, wait a bit and start over.
605 if (unlikely(midboot)) {
606 needgpcb = 0x2;
607 } else {
608 mutex_unlock(&rtp->tasks_gp_mutex);
609 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
610 rcuwait_wait_event(&rtp->cbs_wait,
611 (needgpcb = rcu_tasks_need_gpcb(rtp)),
612 TASK_IDLE);
613 mutex_lock(&rtp->tasks_gp_mutex);
614 }
615
616 if (needgpcb & 0x2) {
617 // Wait for one grace period.
618 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
619 rtp->gp_start = jiffies;
620 rcu_seq_start(&rtp->tasks_gp_seq);
621 rtp->gp_func(rtp);
622 rcu_seq_end(&rtp->tasks_gp_seq);
623 }
624
625 // Invoke callbacks.
626 set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
627 rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
628 mutex_unlock(&rtp->tasks_gp_mutex);
629}
630
631// RCU-tasks kthread that detects grace periods and invokes callbacks.
632static int __noreturn rcu_tasks_kthread(void *arg)
633{
634 int cpu;
635 struct rcu_tasks *rtp = arg;
636
637 for_each_possible_cpu(cpu) {
638 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
639
640 timer_setup(&rtpcp->lazy_timer, call_rcu_tasks_generic_timer, 0);
641 rtpcp->urgent_gp = 1;
642 }
643
644 /* Run on housekeeping CPUs by default. Sysadm can move if desired. */
645 housekeeping_affine(current, HK_TYPE_RCU);
646 smp_store_release(&rtp->kthread_ptr, current); // Let GPs start!
647
648 /*
649 * Each pass through the following loop makes one check for
650 * newly arrived callbacks, and, if there are some, waits for
651 * one RCU-tasks grace period and then invokes the callbacks.
652 * This loop is terminated by the system going down. ;-)
653 */
654 for (;;) {
655 // Wait for one grace period and invoke any callbacks
656 // that are ready.
657 rcu_tasks_one_gp(rtp, false);
658
659 // Paranoid sleep to keep this from entering a tight loop.
660 schedule_timeout_idle(rtp->gp_sleep);
661 }
662}
663
664// Wait for a grace period for the specified flavor of Tasks RCU.
665static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
666{
667 /* Complain if the scheduler has not started. */
668 if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
669 "synchronize_%s() called too soon", rtp->name))
670 return;
671
672 // If the grace-period kthread is running, use it.
673 if (READ_ONCE(rtp->kthread_ptr)) {
674 wait_rcu_gp_state(rtp->wait_state, rtp->call_func);
675 return;
676 }
677 rcu_tasks_one_gp(rtp, true);
678}
679
680/* Spawn RCU-tasks grace-period kthread. */
681static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
682{
683 struct task_struct *t;
684
685 t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
686 if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
687 return;
688}
689
690#ifndef CONFIG_TINY_RCU
691
692/*
693 * Print any non-default Tasks RCU settings.
694 */
695static void __init rcu_tasks_bootup_oddness(void)
696{
697#if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
698 int rtsimc;
699
700 if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
701 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
702 rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
703 if (rtsimc != rcu_task_stall_info_mult) {
704 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
705 rcu_task_stall_info_mult = rtsimc;
706 }
707#endif /* #ifdef CONFIG_TASKS_RCU */
708#ifdef CONFIG_TASKS_RCU
709 pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
710#endif /* #ifdef CONFIG_TASKS_RCU */
711#ifdef CONFIG_TASKS_RUDE_RCU
712 pr_info("\tRude variant of Tasks RCU enabled.\n");
713#endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
714#ifdef CONFIG_TASKS_TRACE_RCU
715 pr_info("\tTracing variant of Tasks RCU enabled.\n");
716#endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
717}
718
719/* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
720static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
721{
722 int cpu;
723 bool havecbs = false;
724 bool haveurgent = false;
725 bool haveurgentcbs = false;
726 bool havependtimer = false;
727
728 for_each_possible_cpu(cpu) {
729 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
730
731 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)))
732 havecbs = true;
733 if (data_race(rtpcp->urgent_gp))
734 haveurgent = true;
735 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)) && data_race(rtpcp->urgent_gp))
736 haveurgentcbs = true;
737 if (data_race(timer_pending(&rtpcp->lazy_timer)))
738 havependtimer = true;
739 if (havecbs && haveurgent && haveurgentcbs && havependtimer)
740 break;
741 }
742 pr_info("%s: %s(%d) since %lu g:%lu i:%lu %c%c%c%c%c l:%lu %s\n",
743 rtp->kname,
744 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
745 jiffies - data_race(rtp->gp_jiffies),
746 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
747 data_race(rtp->n_ipis),
748 ".k"[!!data_race(rtp->kthread_ptr)],
749 ".C"[havecbs],
750 ".u"[haveurgent],
751 ".U"[haveurgentcbs],
752 ".P"[havependtimer],
753 rtp->lazy_jiffies,
754 s);
755}
756
757/* Dump out more rcutorture-relevant state common to all RCU-tasks flavors. */
758static void rcu_tasks_torture_stats_print_generic(struct rcu_tasks *rtp, char *tt,
759 char *tf, char *tst)
760{
761 cpumask_var_t cm;
762 int cpu;
763 bool gotcb = false;
764 unsigned long j = jiffies;
765
766 pr_alert("%s%s Tasks%s RCU g%ld gp_start %lu gp_jiffies %lu gp_state %d (%s).\n",
767 tt, tf, tst, data_race(rtp->tasks_gp_seq),
768 j - data_race(rtp->gp_start), j - data_race(rtp->gp_jiffies),
769 data_race(rtp->gp_state), tasks_gp_state_getname(rtp));
770 pr_alert("\tEnqueue shift %d limit %d Dequeue limit %d gpseq %lu.\n",
771 data_race(rtp->percpu_enqueue_shift),
772 data_race(rtp->percpu_enqueue_lim),
773 data_race(rtp->percpu_dequeue_lim),
774 data_race(rtp->percpu_dequeue_gpseq));
775 (void)zalloc_cpumask_var(&cm, GFP_KERNEL);
776 pr_alert("\tCallback counts:");
777 for_each_possible_cpu(cpu) {
778 long n;
779 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
780
781 if (cpumask_available(cm) && !rcu_barrier_cb_is_done(&rtpcp->barrier_q_head))
782 cpumask_set_cpu(cpu, cm);
783 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
784 if (!n)
785 continue;
786 pr_cont(" %d:%ld", cpu, n);
787 gotcb = true;
788 }
789 if (gotcb)
790 pr_cont(".\n");
791 else
792 pr_cont(" (none).\n");
793 pr_alert("\tBarrier seq %lu start %lu count %d holdout CPUs ",
794 data_race(rtp->barrier_q_seq), j - data_race(rtp->barrier_q_start),
795 atomic_read(&rtp->barrier_q_count));
796 if (cpumask_available(cm) && !cpumask_empty(cm))
797 pr_cont(" %*pbl.\n", cpumask_pr_args(cm));
798 else
799 pr_cont("(none).\n");
800 free_cpumask_var(cm);
801}
802
803#endif // #ifndef CONFIG_TINY_RCU
804
805#if defined(CONFIG_TASKS_RCU)
806
807////////////////////////////////////////////////////////////////////////
808//
809// Shared code between task-list-scanning variants of Tasks RCU.
810
811/* Wait for one RCU-tasks grace period. */
812static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
813{
814 struct task_struct *g;
815 int fract;
816 LIST_HEAD(holdouts);
817 unsigned long j;
818 unsigned long lastinfo;
819 unsigned long lastreport;
820 bool reported = false;
821 int rtsi;
822 struct task_struct *t;
823
824 set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
825 rtp->pregp_func(&holdouts);
826
827 /*
828 * There were callbacks, so we need to wait for an RCU-tasks
829 * grace period. Start off by scanning the task list for tasks
830 * that are not already voluntarily blocked. Mark these tasks
831 * and make a list of them in holdouts.
832 */
833 set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
834 if (rtp->pertask_func) {
835 rcu_read_lock();
836 for_each_process_thread(g, t)
837 rtp->pertask_func(t, &holdouts);
838 rcu_read_unlock();
839 }
840
841 set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
842 rtp->postscan_func(&holdouts);
843
844 /*
845 * Each pass through the following loop scans the list of holdout
846 * tasks, removing any that are no longer holdouts. When the list
847 * is empty, we are done.
848 */
849 lastreport = jiffies;
850 lastinfo = lastreport;
851 rtsi = READ_ONCE(rcu_task_stall_info);
852
853 // Start off with initial wait and slowly back off to 1 HZ wait.
854 fract = rtp->init_fract;
855
856 while (!list_empty(&holdouts)) {
857 ktime_t exp;
858 bool firstreport;
859 bool needreport;
860 int rtst;
861
862 // Slowly back off waiting for holdouts
863 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
864 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
865 schedule_timeout_idle(fract);
866 } else {
867 exp = jiffies_to_nsecs(fract);
868 __set_current_state(TASK_IDLE);
869 schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
870 }
871
872 if (fract < HZ)
873 fract++;
874
875 rtst = READ_ONCE(rcu_task_stall_timeout);
876 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
877 if (needreport) {
878 lastreport = jiffies;
879 reported = true;
880 }
881 firstreport = true;
882 WARN_ON(signal_pending(current));
883 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
884 rtp->holdouts_func(&holdouts, needreport, &firstreport);
885
886 // Print pre-stall informational messages if needed.
887 j = jiffies;
888 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
889 lastinfo = j;
890 rtsi = rtsi * rcu_task_stall_info_mult;
891 pr_info("%s: %s grace period number %lu (since boot) is %lu jiffies old.\n",
892 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
893 }
894 }
895
896 set_tasks_gp_state(rtp, RTGS_POST_GP);
897 rtp->postgp_func(rtp);
898}
899
900#endif /* #if defined(CONFIG_TASKS_RCU) */
901
902#ifdef CONFIG_TASKS_RCU
903
904////////////////////////////////////////////////////////////////////////
905//
906// Simple variant of RCU whose quiescent states are voluntary context
907// switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
908// As such, grace periods can take one good long time. There are no
909// read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
910// because this implementation is intended to get the system into a safe
911// state for some of the manipulations involved in tracing and the like.
912// Finally, this implementation does not support high call_rcu_tasks()
913// rates from multiple CPUs. If this is required, per-CPU callback lists
914// will be needed.
915//
916// The implementation uses rcu_tasks_wait_gp(), which relies on function
917// pointers in the rcu_tasks structure. The rcu_spawn_tasks_kthread()
918// function sets these function pointers up so that rcu_tasks_wait_gp()
919// invokes these functions in this order:
920//
921// rcu_tasks_pregp_step():
922// Invokes synchronize_rcu() in order to wait for all in-flight
923// t->on_rq and t->nvcsw transitions to complete. This works because
924// all such transitions are carried out with interrupts disabled.
925// rcu_tasks_pertask(), invoked on every non-idle task:
926// For every runnable non-idle task other than the current one, use
927// get_task_struct() to pin down that task, snapshot that task's
928// number of voluntary context switches, and add that task to the
929// holdout list.
930// rcu_tasks_postscan():
931// Gather per-CPU lists of tasks in do_exit() to ensure that all
932// tasks that were in the process of exiting (and which thus might
933// not know to synchronize with this RCU Tasks grace period) have
934// completed exiting. The synchronize_rcu() in rcu_tasks_postgp()
935// will take care of any tasks stuck in the non-preemptible region
936// of do_exit() following its call to exit_tasks_rcu_finish().
937// check_all_holdout_tasks(), repeatedly until holdout list is empty:
938// Scans the holdout list, attempting to identify a quiescent state
939// for each task on the list. If there is a quiescent state, the
940// corresponding task is removed from the holdout list.
941// rcu_tasks_postgp():
942// Invokes synchronize_rcu() in order to ensure that all prior
943// t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
944// to have happened before the end of this RCU Tasks grace period.
945// Again, this works because all such transitions are carried out
946// with interrupts disabled.
947//
948// For each exiting task, the exit_tasks_rcu_start() and
949// exit_tasks_rcu_finish() functions add and remove, respectively, the
950// current task to a per-CPU list of tasks that rcu_tasks_postscan() must
951// wait on. This is necessary because rcu_tasks_postscan() must wait on
952// tasks that have already been removed from the global list of tasks.
953//
954// Pre-grace-period update-side code is ordered before the grace
955// via the raw_spin_lock.*rcu_node(). Pre-grace-period read-side code
956// is ordered before the grace period via synchronize_rcu() call in
957// rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
958// disabling.
959
960/* Pre-grace-period preparation. */
961static void rcu_tasks_pregp_step(struct list_head *hop)
962{
963 /*
964 * Wait for all pre-existing t->on_rq and t->nvcsw transitions
965 * to complete. Invoking synchronize_rcu() suffices because all
966 * these transitions occur with interrupts disabled. Without this
967 * synchronize_rcu(), a read-side critical section that started
968 * before the grace period might be incorrectly seen as having
969 * started after the grace period.
970 *
971 * This synchronize_rcu() also dispenses with the need for a
972 * memory barrier on the first store to t->rcu_tasks_holdout,
973 * as it forces the store to happen after the beginning of the
974 * grace period.
975 */
976 synchronize_rcu();
977}
978
979/* Check for quiescent states since the pregp's synchronize_rcu() */
980static bool rcu_tasks_is_holdout(struct task_struct *t)
981{
982 int cpu;
983
984 /* Has the task been seen voluntarily sleeping? */
985 if (!READ_ONCE(t->on_rq))
986 return false;
987
988 /*
989 * t->on_rq && !t->se.sched_delayed *could* be considered sleeping but
990 * since it is a spurious state (it will transition into the
991 * traditional blocked state or get woken up without outside
992 * dependencies), not considering it such should only affect timing.
993 *
994 * Be conservative for now and not include it.
995 */
996
997 /*
998 * Idle tasks (or idle injection) within the idle loop are RCU-tasks
999 * quiescent states. But CPU boot code performed by the idle task
1000 * isn't a quiescent state.
1001 */
1002 if (is_idle_task(t))
1003 return false;
1004
1005 cpu = task_cpu(t);
1006
1007 /* Idle tasks on offline CPUs are RCU-tasks quiescent states. */
1008 if (t == idle_task(cpu) && !rcu_cpu_online(cpu))
1009 return false;
1010
1011 return true;
1012}
1013
1014/* Per-task initial processing. */
1015static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
1016{
1017 if (t != current && rcu_tasks_is_holdout(t)) {
1018 get_task_struct(t);
1019 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
1020 WRITE_ONCE(t->rcu_tasks_holdout, true);
1021 list_add(&t->rcu_tasks_holdout_list, hop);
1022 }
1023}
1024
1025void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
1026DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
1027
1028/* Processing between scanning taskslist and draining the holdout list. */
1029static void rcu_tasks_postscan(struct list_head *hop)
1030{
1031 int cpu;
1032 int rtsi = READ_ONCE(rcu_task_stall_info);
1033
1034 if (!IS_ENABLED(CONFIG_TINY_RCU)) {
1035 tasks_rcu_exit_stall_timer.expires = jiffies + rtsi;
1036 add_timer(&tasks_rcu_exit_stall_timer);
1037 }
1038
1039 /*
1040 * Exiting tasks may escape the tasklist scan. Those are vulnerable
1041 * until their final schedule() with TASK_DEAD state. To cope with
1042 * this, divide the fragile exit path part in two intersecting
1043 * read side critical sections:
1044 *
1045 * 1) A task_struct list addition before calling exit_notify(),
1046 * which may remove the task from the tasklist, with the
1047 * removal after the final preempt_disable() call in do_exit().
1048 *
1049 * 2) An _RCU_ read side starting with the final preempt_disable()
1050 * call in do_exit() and ending with the final call to schedule()
1051 * with TASK_DEAD state.
1052 *
1053 * This handles the part 1). And postgp will handle part 2) with a
1054 * call to synchronize_rcu().
1055 */
1056
1057 for_each_possible_cpu(cpu) {
1058 unsigned long j = jiffies + 1;
1059 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, cpu);
1060 struct task_struct *t;
1061 struct task_struct *t1;
1062 struct list_head tmp;
1063
1064 raw_spin_lock_irq_rcu_node(rtpcp);
1065 list_for_each_entry_safe(t, t1, &rtpcp->rtp_exit_list, rcu_tasks_exit_list) {
1066 if (list_empty(&t->rcu_tasks_holdout_list))
1067 rcu_tasks_pertask(t, hop);
1068
1069 // RT kernels need frequent pauses, otherwise
1070 // pause at least once per pair of jiffies.
1071 if (!IS_ENABLED(CONFIG_PREEMPT_RT) && time_before(jiffies, j))
1072 continue;
1073
1074 // Keep our place in the list while pausing.
1075 // Nothing else traverses this list, so adding a
1076 // bare list_head is OK.
1077 list_add(&tmp, &t->rcu_tasks_exit_list);
1078 raw_spin_unlock_irq_rcu_node(rtpcp);
1079 cond_resched(); // For CONFIG_PREEMPT=n kernels
1080 raw_spin_lock_irq_rcu_node(rtpcp);
1081 t1 = list_entry(tmp.next, struct task_struct, rcu_tasks_exit_list);
1082 list_del(&tmp);
1083 j = jiffies + 1;
1084 }
1085 raw_spin_unlock_irq_rcu_node(rtpcp);
1086 }
1087
1088 if (!IS_ENABLED(CONFIG_TINY_RCU))
1089 timer_delete_sync(&tasks_rcu_exit_stall_timer);
1090}
1091
1092/* See if tasks are still holding out, complain if so. */
1093static void check_holdout_task(struct task_struct *t,
1094 bool needreport, bool *firstreport)
1095{
1096 int cpu;
1097
1098 if (!READ_ONCE(t->rcu_tasks_holdout) ||
1099 t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
1100 !rcu_tasks_is_holdout(t) ||
1101 (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
1102 !is_idle_task(t) && READ_ONCE(t->rcu_tasks_idle_cpu) >= 0)) {
1103 WRITE_ONCE(t->rcu_tasks_holdout, false);
1104 list_del_init(&t->rcu_tasks_holdout_list);
1105 put_task_struct(t);
1106 return;
1107 }
1108 rcu_request_urgent_qs_task(t);
1109 if (!needreport)
1110 return;
1111 if (*firstreport) {
1112 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
1113 *firstreport = false;
1114 }
1115 cpu = task_cpu(t);
1116 pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
1117 t, ".I"[is_idle_task(t)],
1118 "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
1119 t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
1120 data_race(t->rcu_tasks_idle_cpu), cpu);
1121 sched_show_task(t);
1122}
1123
1124/* Scan the holdout lists for tasks no longer holding out. */
1125static void check_all_holdout_tasks(struct list_head *hop,
1126 bool needreport, bool *firstreport)
1127{
1128 struct task_struct *t, *t1;
1129
1130 list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
1131 check_holdout_task(t, needreport, firstreport);
1132 cond_resched();
1133 }
1134}
1135
1136/* Finish off the Tasks-RCU grace period. */
1137static void rcu_tasks_postgp(struct rcu_tasks *rtp)
1138{
1139 /*
1140 * Because ->on_rq and ->nvcsw are not guaranteed to have a full
1141 * memory barriers prior to them in the schedule() path, memory
1142 * reordering on other CPUs could cause their RCU-tasks read-side
1143 * critical sections to extend past the end of the grace period.
1144 * However, because these ->nvcsw updates are carried out with
1145 * interrupts disabled, we can use synchronize_rcu() to force the
1146 * needed ordering on all such CPUs.
1147 *
1148 * This synchronize_rcu() also confines all ->rcu_tasks_holdout
1149 * accesses to be within the grace period, avoiding the need for
1150 * memory barriers for ->rcu_tasks_holdout accesses.
1151 *
1152 * In addition, this synchronize_rcu() waits for exiting tasks
1153 * to complete their final preempt_disable() region of execution,
1154 * enforcing the whole region before tasklist removal until
1155 * the final schedule() with TASK_DEAD state to be an RCU TASKS
1156 * read side critical section.
1157 */
1158 synchronize_rcu();
1159}
1160
1161static void tasks_rcu_exit_stall(struct timer_list *unused)
1162{
1163#ifndef CONFIG_TINY_RCU
1164 int rtsi;
1165
1166 rtsi = READ_ONCE(rcu_task_stall_info);
1167 pr_info("%s: %s grace period number %lu (since boot) gp_state: %s is %lu jiffies old.\n",
1168 __func__, rcu_tasks.kname, rcu_tasks.tasks_gp_seq,
1169 tasks_gp_state_getname(&rcu_tasks), jiffies - rcu_tasks.gp_jiffies);
1170 pr_info("Please check any exiting tasks stuck between calls to exit_tasks_rcu_start() and exit_tasks_rcu_finish()\n");
1171 tasks_rcu_exit_stall_timer.expires = jiffies + rtsi;
1172 add_timer(&tasks_rcu_exit_stall_timer);
1173#endif // #ifndef CONFIG_TINY_RCU
1174}
1175
1176/**
1177 * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
1178 * @rhp: structure to be used for queueing the RCU updates.
1179 * @func: actual callback function to be invoked after the grace period
1180 *
1181 * The callback function will be invoked some time after a full grace
1182 * period elapses, in other words after all currently executing rcu-tasks
1183 * read-side critical sections have completed. call_rcu_tasks() assumes
1184 * that the read-side critical sections end at a voluntary context
1185 * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
1186 * or transition to usermode execution. As such, there are no read-side
1187 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1188 * this primitive is intended to determine that all tasks have passed
1189 * through a safe state, not so much for data-structure synchronization.
1190 *
1191 * See the description of call_rcu() for more detailed information on
1192 * memory ordering guarantees.
1193 */
1194void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
1195{
1196 call_rcu_tasks_generic(rhp, func, &rcu_tasks);
1197}
1198EXPORT_SYMBOL_GPL(call_rcu_tasks);
1199
1200/**
1201 * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
1202 *
1203 * Control will return to the caller some time after a full rcu-tasks
1204 * grace period has elapsed, in other words after all currently
1205 * executing rcu-tasks read-side critical sections have elapsed. These
1206 * read-side critical sections are delimited by calls to schedule(),
1207 * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
1208 * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
1209 *
1210 * This is a very specialized primitive, intended only for a few uses in
1211 * tracing and other situations requiring manipulation of function
1212 * preambles and profiling hooks. The synchronize_rcu_tasks() function
1213 * is not (yet) intended for heavy use from multiple CPUs.
1214 *
1215 * See the description of synchronize_rcu() for more detailed information
1216 * on memory ordering guarantees.
1217 */
1218void synchronize_rcu_tasks(void)
1219{
1220 synchronize_rcu_tasks_generic(&rcu_tasks);
1221}
1222EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
1223
1224/**
1225 * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
1226 *
1227 * Although the current implementation is guaranteed to wait, it is not
1228 * obligated to, for example, if there are no pending callbacks.
1229 */
1230void rcu_barrier_tasks(void)
1231{
1232 rcu_barrier_tasks_generic(&rcu_tasks);
1233}
1234EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
1235
1236static int rcu_tasks_lazy_ms = -1;
1237module_param(rcu_tasks_lazy_ms, int, 0444);
1238
1239static int __init rcu_spawn_tasks_kthread(void)
1240{
1241 rcu_tasks.gp_sleep = HZ / 10;
1242 rcu_tasks.init_fract = HZ / 10;
1243 if (rcu_tasks_lazy_ms >= 0)
1244 rcu_tasks.lazy_jiffies = msecs_to_jiffies(rcu_tasks_lazy_ms);
1245 rcu_tasks.pregp_func = rcu_tasks_pregp_step;
1246 rcu_tasks.pertask_func = rcu_tasks_pertask;
1247 rcu_tasks.postscan_func = rcu_tasks_postscan;
1248 rcu_tasks.holdouts_func = check_all_holdout_tasks;
1249 rcu_tasks.postgp_func = rcu_tasks_postgp;
1250 rcu_tasks.wait_state = TASK_IDLE;
1251 rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1252 return 0;
1253}
1254
1255#if !defined(CONFIG_TINY_RCU)
1256void show_rcu_tasks_classic_gp_kthread(void)
1257{
1258 show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1259}
1260EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1261
1262void rcu_tasks_torture_stats_print(char *tt, char *tf)
1263{
1264 rcu_tasks_torture_stats_print_generic(&rcu_tasks, tt, tf, "");
1265}
1266EXPORT_SYMBOL_GPL(rcu_tasks_torture_stats_print);
1267#endif // !defined(CONFIG_TINY_RCU)
1268
1269struct task_struct *get_rcu_tasks_gp_kthread(void)
1270{
1271 return rcu_tasks.kthread_ptr;
1272}
1273EXPORT_SYMBOL_GPL(get_rcu_tasks_gp_kthread);
1274
1275void rcu_tasks_get_gp_data(int *flags, unsigned long *gp_seq)
1276{
1277 *flags = 0;
1278 *gp_seq = rcu_seq_current(&rcu_tasks.tasks_gp_seq);
1279}
1280EXPORT_SYMBOL_GPL(rcu_tasks_get_gp_data);
1281
1282/*
1283 * Protect against tasklist scan blind spot while the task is exiting and
1284 * may be removed from the tasklist. Do this by adding the task to yet
1285 * another list.
1286 *
1287 * Note that the task will remove itself from this list, so there is no
1288 * need for get_task_struct(), except in the case where rcu_tasks_pertask()
1289 * adds it to the holdout list, in which case rcu_tasks_pertask() supplies
1290 * the needed get_task_struct().
1291 */
1292void exit_tasks_rcu_start(void)
1293{
1294 unsigned long flags;
1295 struct rcu_tasks_percpu *rtpcp;
1296 struct task_struct *t = current;
1297
1298 WARN_ON_ONCE(!list_empty(&t->rcu_tasks_exit_list));
1299 preempt_disable();
1300 rtpcp = this_cpu_ptr(rcu_tasks.rtpcpu);
1301 t->rcu_tasks_exit_cpu = smp_processor_id();
1302 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1303 WARN_ON_ONCE(!rtpcp->rtp_exit_list.next);
1304 list_add(&t->rcu_tasks_exit_list, &rtpcp->rtp_exit_list);
1305 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1306 preempt_enable();
1307}
1308
1309/*
1310 * Remove the task from the "yet another list" because do_exit() is now
1311 * non-preemptible, allowing synchronize_rcu() to wait beyond this point.
1312 */
1313void exit_tasks_rcu_finish(void)
1314{
1315 unsigned long flags;
1316 struct rcu_tasks_percpu *rtpcp;
1317 struct task_struct *t = current;
1318
1319 WARN_ON_ONCE(list_empty(&t->rcu_tasks_exit_list));
1320 rtpcp = per_cpu_ptr(rcu_tasks.rtpcpu, t->rcu_tasks_exit_cpu);
1321 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1322 list_del_init(&t->rcu_tasks_exit_list);
1323 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1324}
1325
1326#else /* #ifdef CONFIG_TASKS_RCU */
1327void exit_tasks_rcu_start(void) { }
1328void exit_tasks_rcu_finish(void) { }
1329#endif /* #else #ifdef CONFIG_TASKS_RCU */
1330
1331#ifdef CONFIG_TASKS_RUDE_RCU
1332
1333////////////////////////////////////////////////////////////////////////
1334//
1335// "Rude" variant of Tasks RCU, inspired by Steve Rostedt's
1336// trick of passing an empty function to schedule_on_each_cpu().
1337// This approach provides batching of concurrent calls to the synchronous
1338// synchronize_rcu_tasks_rude() API. This invokes schedule_on_each_cpu()
1339// in order to send IPIs far and wide and induces otherwise unnecessary
1340// context switches on all online CPUs, whether idle or not.
1341//
1342// Callback handling is provided by the rcu_tasks_kthread() function.
1343//
1344// Ordering is provided by the scheduler's context-switch code.
1345
1346// Empty function to allow workqueues to force a context switch.
1347static void rcu_tasks_be_rude(struct work_struct *work)
1348{
1349}
1350
1351// Wait for one rude RCU-tasks grace period.
1352static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1353{
1354 rtp->n_ipis += cpumask_weight(cpu_online_mask);
1355 schedule_on_each_cpu(rcu_tasks_be_rude);
1356}
1357
1358static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1359DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1360 "RCU Tasks Rude");
1361
1362/*
1363 * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1364 * @rhp: structure to be used for queueing the RCU updates.
1365 * @func: actual callback function to be invoked after the grace period
1366 *
1367 * The callback function will be invoked some time after a full grace
1368 * period elapses, in other words after all currently executing rude
1369 * rcu-tasks read-side critical sections have completed. call_rcu_tasks_rude()
1370 * assumes that the read-side critical sections end at context switch,
1371 * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1372 * usermode execution is schedulable). As such, there are no read-side
1373 * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1374 * this primitive is intended to determine that all tasks have passed
1375 * through a safe state, not so much for data-structure synchronization.
1376 *
1377 * See the description of call_rcu() for more detailed information on
1378 * memory ordering guarantees.
1379 *
1380 * This is no longer exported, and is instead reserved for use by
1381 * synchronize_rcu_tasks_rude().
1382 */
1383static void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1384{
1385 call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1386}
1387
1388/**
1389 * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1390 *
1391 * Control will return to the caller some time after a rude rcu-tasks
1392 * grace period has elapsed, in other words after all currently
1393 * executing rude rcu-tasks read-side critical sections have elapsed. These
1394 * read-side critical sections are delimited by calls to schedule(),
1395 * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1396 * context), and (in theory, anyway) cond_resched().
1397 *
1398 * This is a very specialized primitive, intended only for a few uses in
1399 * tracing and other situations requiring manipulation of function preambles
1400 * and profiling hooks. The synchronize_rcu_tasks_rude() function is not
1401 * (yet) intended for heavy use from multiple CPUs.
1402 *
1403 * See the description of synchronize_rcu() for more detailed information
1404 * on memory ordering guarantees.
1405 */
1406void synchronize_rcu_tasks_rude(void)
1407{
1408 if (!IS_ENABLED(CONFIG_ARCH_WANTS_NO_INSTR) || IS_ENABLED(CONFIG_FORCE_TASKS_RUDE_RCU))
1409 synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1410}
1411EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1412
1413static int __init rcu_spawn_tasks_rude_kthread(void)
1414{
1415 rcu_tasks_rude.gp_sleep = HZ / 10;
1416 rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1417 return 0;
1418}
1419
1420#if !defined(CONFIG_TINY_RCU)
1421void show_rcu_tasks_rude_gp_kthread(void)
1422{
1423 show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1424}
1425EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1426
1427void rcu_tasks_rude_torture_stats_print(char *tt, char *tf)
1428{
1429 rcu_tasks_torture_stats_print_generic(&rcu_tasks_rude, tt, tf, "");
1430}
1431EXPORT_SYMBOL_GPL(rcu_tasks_rude_torture_stats_print);
1432#endif // !defined(CONFIG_TINY_RCU)
1433
1434struct task_struct *get_rcu_tasks_rude_gp_kthread(void)
1435{
1436 return rcu_tasks_rude.kthread_ptr;
1437}
1438EXPORT_SYMBOL_GPL(get_rcu_tasks_rude_gp_kthread);
1439
1440void rcu_tasks_rude_get_gp_data(int *flags, unsigned long *gp_seq)
1441{
1442 *flags = 0;
1443 *gp_seq = rcu_seq_current(&rcu_tasks_rude.tasks_gp_seq);
1444}
1445EXPORT_SYMBOL_GPL(rcu_tasks_rude_get_gp_data);
1446
1447#endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1448
1449#ifndef CONFIG_TINY_RCU
1450void show_rcu_tasks_gp_kthreads(void)
1451{
1452 show_rcu_tasks_classic_gp_kthread();
1453 show_rcu_tasks_rude_gp_kthread();
1454}
1455#endif /* #ifndef CONFIG_TINY_RCU */
1456
1457#ifdef CONFIG_PROVE_RCU
1458struct rcu_tasks_test_desc {
1459 struct rcu_head rh;
1460 const char *name;
1461 bool notrun;
1462 unsigned long runstart;
1463 void (*gp_dbg)(void);
1464};
1465
1466static struct rcu_tasks_test_desc tests[] = {
1467 {
1468 .name = "call_rcu_tasks()",
1469 /* If not defined, the test is skipped. */
1470 .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1471 /* Dump rcu tasks status, if test failed. */
1472 .gp_dbg = show_rcu_tasks_classic_gp_kthread
1473 },
1474 {
1475 .name = "call_rcu_tasks_trace()",
1476 /* If not defined, the test is skipped. */
1477 .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1478 }
1479};
1480
1481#if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
1482static void test_rcu_tasks_callback(struct rcu_head *rhp)
1483{
1484 struct rcu_tasks_test_desc *rttd =
1485 container_of(rhp, struct rcu_tasks_test_desc, rh);
1486
1487 pr_info("Callback from %s invoked.\n", rttd->name);
1488
1489 rttd->notrun = false;
1490}
1491#endif // #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
1492
1493static void rcu_tasks_initiate_self_tests(void)
1494{
1495#ifdef CONFIG_TASKS_RCU
1496 pr_info("Running RCU Tasks wait API self tests\n");
1497 tests[0].runstart = jiffies;
1498 synchronize_rcu_tasks();
1499 call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1500#endif
1501
1502#ifdef CONFIG_TASKS_RUDE_RCU
1503 pr_info("Running RCU Tasks Rude wait API self tests\n");
1504 synchronize_rcu_tasks_rude();
1505#endif
1506
1507#ifdef CONFIG_TASKS_TRACE_RCU
1508 pr_info("Running RCU Tasks Trace wait API self tests\n");
1509 tests[1].runstart = jiffies;
1510 synchronize_rcu_tasks_trace();
1511 call_rcu_tasks_trace(&tests[1].rh, test_rcu_tasks_callback);
1512#endif
1513}
1514
1515/*
1516 * Return: 0 - test passed
1517 * 1 - test failed, but have not timed out yet
1518 * -1 - test failed and timed out
1519 */
1520static int rcu_tasks_verify_self_tests(void)
1521{
1522 int ret = 0;
1523 int i;
1524 unsigned long bst = rcu_task_stall_timeout;
1525
1526 if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
1527 bst = RCU_TASK_BOOT_STALL_TIMEOUT;
1528 for (i = 0; i < ARRAY_SIZE(tests); i++) {
1529 while (tests[i].notrun) { // still hanging.
1530 if (time_after(jiffies, tests[i].runstart + bst)) {
1531 pr_err("%s has failed boot-time tests.\n", tests[i].name);
1532 if (tests[i].gp_dbg)
1533 tests[i].gp_dbg();
1534 ret = -1;
1535 break;
1536 }
1537 ret = 1;
1538 break;
1539 }
1540 }
1541 WARN_ON(ret < 0);
1542
1543 return ret;
1544}
1545
1546/*
1547 * Repeat the rcu_tasks_verify_self_tests() call once every second until the
1548 * test passes or has timed out.
1549 */
1550static struct delayed_work rcu_tasks_verify_work;
1551static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
1552{
1553 int ret = rcu_tasks_verify_self_tests();
1554
1555 if (ret <= 0)
1556 return;
1557
1558 /* Test fails but not timed out yet, reschedule another check */
1559 schedule_delayed_work(&rcu_tasks_verify_work, HZ);
1560}
1561
1562static int rcu_tasks_verify_schedule_work(void)
1563{
1564 INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
1565 rcu_tasks_verify_work_fn(NULL);
1566 return 0;
1567}
1568late_initcall(rcu_tasks_verify_schedule_work);
1569#else /* #ifdef CONFIG_PROVE_RCU */
1570static void rcu_tasks_initiate_self_tests(void) { }
1571#endif /* #else #ifdef CONFIG_PROVE_RCU */
1572
1573void __init tasks_cblist_init_generic(void)
1574{
1575 lockdep_assert_irqs_disabled();
1576 WARN_ON(num_online_cpus() > 1);
1577
1578#ifdef CONFIG_TASKS_RCU
1579 cblist_init_generic(&rcu_tasks);
1580#endif
1581
1582#ifdef CONFIG_TASKS_RUDE_RCU
1583 cblist_init_generic(&rcu_tasks_rude);
1584#endif
1585}
1586
1587static int __init rcu_init_tasks_generic(void)
1588{
1589#ifdef CONFIG_TASKS_RCU
1590 rcu_spawn_tasks_kthread();
1591#endif
1592
1593#ifdef CONFIG_TASKS_RUDE_RCU
1594 rcu_spawn_tasks_rude_kthread();
1595#endif
1596
1597 // Run the self-tests.
1598 rcu_tasks_initiate_self_tests();
1599
1600 return 0;
1601}
1602core_initcall(rcu_init_tasks_generic);
1603
1604#else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
1605static inline void rcu_tasks_bootup_oddness(void) {}
1606#endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */
1607
1608#ifdef CONFIG_TASKS_TRACE_RCU
1609
1610////////////////////////////////////////////////////////////////////////
1611//
1612// Tracing variant of Tasks RCU. This variant is designed to be used
1613// to protect tracing hooks, including those of BPF. This variant
1614// is implemented via a straightforward mapping onto SRCU-fast.
1615
1616DEFINE_SRCU_FAST(rcu_tasks_trace_srcu_struct);
1617EXPORT_SYMBOL_GPL(rcu_tasks_trace_srcu_struct);
1618
1619unsigned long rcu_tasks_trace_batches_completed(void)
1620{
1621 return srcu_batches_completed(&rcu_tasks_trace_srcu_struct);
1622}
1623EXPORT_SYMBOL_GPL(rcu_tasks_trace_batches_completed);
1624
1625#endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */