Merge tag 'vfs-6.7.iomap' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs
[linux-2.6-microblaze.git] / kernel / rcu / tasks.h
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
15 struct rcu_tasks;
16 typedef void (*rcu_tasks_gp_func_t)(struct rcu_tasks *rtp);
17 typedef void (*pregp_func_t)(struct list_head *hop);
18 typedef void (*pertask_func_t)(struct task_struct *t, struct list_head *hop);
19 typedef void (*postscan_func_t)(struct list_head *hop);
20 typedef void (*holdouts_func_t)(struct list_head *hop, bool ndrpt, bool *frptp);
21 typedef void (*postgp_func_t)(struct rcu_tasks *rtp);
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  * @cpu: CPU number corresponding to this entry.
36  * @rtpp: Pointer to the rcu_tasks structure.
37  */
38 struct rcu_tasks_percpu {
39         struct rcu_segcblist cblist;
40         raw_spinlock_t __private lock;
41         unsigned long rtp_jiffies;
42         unsigned long rtp_n_lock_retries;
43         struct timer_list lazy_timer;
44         unsigned int urgent_gp;
45         struct work_struct rtp_work;
46         struct irq_work rtp_irq_work;
47         struct rcu_head barrier_q_head;
48         struct list_head rtp_blkd_tasks;
49         int cpu;
50         struct rcu_tasks *rtpp;
51 };
52
53 /**
54  * struct rcu_tasks - Definition for a Tasks-RCU-like mechanism.
55  * @cbs_wait: RCU wait allowing a new callback to get kthread's attention.
56  * @cbs_gbl_lock: Lock protecting callback list.
57  * @tasks_gp_mutex: Mutex protecting grace period, needed during mid-boot dead zone.
58  * @gp_func: This flavor's grace-period-wait function.
59  * @gp_state: Grace period's most recent state transition (debugging).
60  * @gp_sleep: Per-grace-period sleep to prevent CPU-bound looping.
61  * @init_fract: Initial backoff sleep interval.
62  * @gp_jiffies: Time of last @gp_state transition.
63  * @gp_start: Most recent grace-period start in jiffies.
64  * @tasks_gp_seq: Number of grace periods completed since boot.
65  * @n_ipis: Number of IPIs sent to encourage grace periods to end.
66  * @n_ipis_fails: Number of IPI-send failures.
67  * @kthread_ptr: This flavor's grace-period/callback-invocation kthread.
68  * @lazy_jiffies: Number of jiffies to allow callbacks to be lazy.
69  * @pregp_func: This flavor's pre-grace-period function (optional).
70  * @pertask_func: This flavor's per-task scan function (optional).
71  * @postscan_func: This flavor's post-task scan function (optional).
72  * @holdouts_func: This flavor's holdout-list scan function (optional).
73  * @postgp_func: This flavor's post-grace-period function (optional).
74  * @call_func: This flavor's call_rcu()-equivalent function.
75  * @rtpcpu: This flavor's rcu_tasks_percpu structure.
76  * @percpu_enqueue_shift: Shift down CPU ID this much when enqueuing callbacks.
77  * @percpu_enqueue_lim: Number of per-CPU callback queues in use for enqueuing.
78  * @percpu_dequeue_lim: Number of per-CPU callback queues in use for dequeuing.
79  * @percpu_dequeue_gpseq: RCU grace-period number to propagate enqueue limit to dequeuers.
80  * @barrier_q_mutex: Serialize barrier operations.
81  * @barrier_q_count: Number of queues being waited on.
82  * @barrier_q_completion: Barrier wait/wakeup mechanism.
83  * @barrier_q_seq: Sequence number for barrier operations.
84  * @name: This flavor's textual name.
85  * @kname: This flavor's kthread name.
86  */
87 struct rcu_tasks {
88         struct rcuwait cbs_wait;
89         raw_spinlock_t cbs_gbl_lock;
90         struct mutex tasks_gp_mutex;
91         int gp_state;
92         int gp_sleep;
93         int init_fract;
94         unsigned long gp_jiffies;
95         unsigned long gp_start;
96         unsigned long tasks_gp_seq;
97         unsigned long n_ipis;
98         unsigned long n_ipis_fails;
99         struct task_struct *kthread_ptr;
100         unsigned long lazy_jiffies;
101         rcu_tasks_gp_func_t gp_func;
102         pregp_func_t pregp_func;
103         pertask_func_t pertask_func;
104         postscan_func_t postscan_func;
105         holdouts_func_t holdouts_func;
106         postgp_func_t postgp_func;
107         call_rcu_func_t call_func;
108         struct rcu_tasks_percpu __percpu *rtpcpu;
109         int percpu_enqueue_shift;
110         int percpu_enqueue_lim;
111         int percpu_dequeue_lim;
112         unsigned long percpu_dequeue_gpseq;
113         struct mutex barrier_q_mutex;
114         atomic_t barrier_q_count;
115         struct completion barrier_q_completion;
116         unsigned long barrier_q_seq;
117         char *name;
118         char *kname;
119 };
120
121 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp);
122
123 #define DEFINE_RCU_TASKS(rt_name, gp, call, n)                                          \
124 static DEFINE_PER_CPU(struct rcu_tasks_percpu, rt_name ## __percpu) = {                 \
125         .lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name ## __percpu.cbs_pcpu_lock),            \
126         .rtp_irq_work = IRQ_WORK_INIT_HARD(call_rcu_tasks_iw_wakeup),                   \
127 };                                                                                      \
128 static struct rcu_tasks rt_name =                                                       \
129 {                                                                                       \
130         .cbs_wait = __RCUWAIT_INITIALIZER(rt_name.wait),                                \
131         .cbs_gbl_lock = __RAW_SPIN_LOCK_UNLOCKED(rt_name.cbs_gbl_lock),                 \
132         .tasks_gp_mutex = __MUTEX_INITIALIZER(rt_name.tasks_gp_mutex),                  \
133         .gp_func = gp,                                                                  \
134         .call_func = call,                                                              \
135         .rtpcpu = &rt_name ## __percpu,                                                 \
136         .lazy_jiffies = DIV_ROUND_UP(HZ, 4),                                            \
137         .name = n,                                                                      \
138         .percpu_enqueue_shift = order_base_2(CONFIG_NR_CPUS),                           \
139         .percpu_enqueue_lim = 1,                                                        \
140         .percpu_dequeue_lim = 1,                                                        \
141         .barrier_q_mutex = __MUTEX_INITIALIZER(rt_name.barrier_q_mutex),                \
142         .barrier_q_seq = (0UL - 50UL) << RCU_SEQ_CTR_SHIFT,                             \
143         .kname = #rt_name,                                                              \
144 }
145
146 #ifdef CONFIG_TASKS_RCU
147 /* Track exiting tasks in order to allow them to be waited for. */
148 DEFINE_STATIC_SRCU(tasks_rcu_exit_srcu);
149
150 /* Report delay in synchronize_srcu() completion in rcu_tasks_postscan(). */
151 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused);
152 static DEFINE_TIMER(tasks_rcu_exit_srcu_stall_timer, tasks_rcu_exit_srcu_stall);
153 #endif
154
155 /* Avoid IPIing CPUs early in the grace period. */
156 #define RCU_TASK_IPI_DELAY (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) ? HZ / 2 : 0)
157 static int rcu_task_ipi_delay __read_mostly = RCU_TASK_IPI_DELAY;
158 module_param(rcu_task_ipi_delay, int, 0644);
159
160 /* Control stall timeouts.  Disable with <= 0, otherwise jiffies till stall. */
161 #define RCU_TASK_BOOT_STALL_TIMEOUT (HZ * 30)
162 #define RCU_TASK_STALL_TIMEOUT (HZ * 60 * 10)
163 static int rcu_task_stall_timeout __read_mostly = RCU_TASK_STALL_TIMEOUT;
164 module_param(rcu_task_stall_timeout, int, 0644);
165 #define RCU_TASK_STALL_INFO (HZ * 10)
166 static int rcu_task_stall_info __read_mostly = RCU_TASK_STALL_INFO;
167 module_param(rcu_task_stall_info, int, 0644);
168 static int rcu_task_stall_info_mult __read_mostly = 3;
169 module_param(rcu_task_stall_info_mult, int, 0444);
170
171 static int rcu_task_enqueue_lim __read_mostly = -1;
172 module_param(rcu_task_enqueue_lim, int, 0444);
173
174 static bool rcu_task_cb_adjust;
175 static int rcu_task_contend_lim __read_mostly = 100;
176 module_param(rcu_task_contend_lim, int, 0444);
177 static int rcu_task_collapse_lim __read_mostly = 10;
178 module_param(rcu_task_collapse_lim, int, 0444);
179 static int rcu_task_lazy_lim __read_mostly = 32;
180 module_param(rcu_task_lazy_lim, int, 0444);
181
182 /* RCU tasks grace-period state for debugging. */
183 #define RTGS_INIT                0
184 #define RTGS_WAIT_WAIT_CBS       1
185 #define RTGS_WAIT_GP             2
186 #define RTGS_PRE_WAIT_GP         3
187 #define RTGS_SCAN_TASKLIST       4
188 #define RTGS_POST_SCAN_TASKLIST  5
189 #define RTGS_WAIT_SCAN_HOLDOUTS  6
190 #define RTGS_SCAN_HOLDOUTS       7
191 #define RTGS_POST_GP             8
192 #define RTGS_WAIT_READERS        9
193 #define RTGS_INVOKE_CBS         10
194 #define RTGS_WAIT_CBS           11
195 #ifndef CONFIG_TINY_RCU
196 static const char * const rcu_tasks_gp_state_names[] = {
197         "RTGS_INIT",
198         "RTGS_WAIT_WAIT_CBS",
199         "RTGS_WAIT_GP",
200         "RTGS_PRE_WAIT_GP",
201         "RTGS_SCAN_TASKLIST",
202         "RTGS_POST_SCAN_TASKLIST",
203         "RTGS_WAIT_SCAN_HOLDOUTS",
204         "RTGS_SCAN_HOLDOUTS",
205         "RTGS_POST_GP",
206         "RTGS_WAIT_READERS",
207         "RTGS_INVOKE_CBS",
208         "RTGS_WAIT_CBS",
209 };
210 #endif /* #ifndef CONFIG_TINY_RCU */
211
212 ////////////////////////////////////////////////////////////////////////
213 //
214 // Generic code.
215
216 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp);
217
218 /* Record grace-period phase and time. */
219 static void set_tasks_gp_state(struct rcu_tasks *rtp, int newstate)
220 {
221         rtp->gp_state = newstate;
222         rtp->gp_jiffies = jiffies;
223 }
224
225 #ifndef CONFIG_TINY_RCU
226 /* Return state name. */
227 static const char *tasks_gp_state_getname(struct rcu_tasks *rtp)
228 {
229         int i = data_race(rtp->gp_state); // Let KCSAN detect update races
230         int j = READ_ONCE(i); // Prevent the compiler from reading twice
231
232         if (j >= ARRAY_SIZE(rcu_tasks_gp_state_names))
233                 return "???";
234         return rcu_tasks_gp_state_names[j];
235 }
236 #endif /* #ifndef CONFIG_TINY_RCU */
237
238 // Initialize per-CPU callback lists for the specified flavor of
239 // Tasks RCU.  Do not enqueue callbacks before this function is invoked.
240 static void cblist_init_generic(struct rcu_tasks *rtp)
241 {
242         int cpu;
243         unsigned long flags;
244         int lim;
245         int shift;
246
247         if (rcu_task_enqueue_lim < 0) {
248                 rcu_task_enqueue_lim = 1;
249                 rcu_task_cb_adjust = true;
250         } else if (rcu_task_enqueue_lim == 0) {
251                 rcu_task_enqueue_lim = 1;
252         }
253         lim = rcu_task_enqueue_lim;
254
255         if (lim > nr_cpu_ids)
256                 lim = nr_cpu_ids;
257         shift = ilog2(nr_cpu_ids / lim);
258         if (((nr_cpu_ids - 1) >> shift) >= lim)
259                 shift++;
260         WRITE_ONCE(rtp->percpu_enqueue_shift, shift);
261         WRITE_ONCE(rtp->percpu_dequeue_lim, lim);
262         smp_store_release(&rtp->percpu_enqueue_lim, lim);
263         for_each_possible_cpu(cpu) {
264                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
265
266                 WARN_ON_ONCE(!rtpcp);
267                 if (cpu)
268                         raw_spin_lock_init(&ACCESS_PRIVATE(rtpcp, lock));
269                 local_irq_save(flags);  // serialize initialization
270                 if (rcu_segcblist_empty(&rtpcp->cblist))
271                         rcu_segcblist_init(&rtpcp->cblist);
272                 local_irq_restore(flags);
273                 INIT_WORK(&rtpcp->rtp_work, rcu_tasks_invoke_cbs_wq);
274                 rtpcp->cpu = cpu;
275                 rtpcp->rtpp = rtp;
276                 if (!rtpcp->rtp_blkd_tasks.next)
277                         INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
278         }
279
280         pr_info("%s: Setting shift to %d and lim to %d rcu_task_cb_adjust=%d.\n", rtp->name,
281                         data_race(rtp->percpu_enqueue_shift), data_race(rtp->percpu_enqueue_lim), rcu_task_cb_adjust);
282 }
283
284 // Compute wakeup time for lazy callback timer.
285 static unsigned long rcu_tasks_lazy_time(struct rcu_tasks *rtp)
286 {
287         return jiffies + rtp->lazy_jiffies;
288 }
289
290 // Timer handler that unlazifies lazy callbacks.
291 static void call_rcu_tasks_generic_timer(struct timer_list *tlp)
292 {
293         unsigned long flags;
294         bool needwake = false;
295         struct rcu_tasks *rtp;
296         struct rcu_tasks_percpu *rtpcp = from_timer(rtpcp, tlp, lazy_timer);
297
298         rtp = rtpcp->rtpp;
299         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
300         if (!rcu_segcblist_empty(&rtpcp->cblist) && rtp->lazy_jiffies) {
301                 if (!rtpcp->urgent_gp)
302                         rtpcp->urgent_gp = 1;
303                 needwake = true;
304                 mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
305         }
306         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
307         if (needwake)
308                 rcuwait_wake_up(&rtp->cbs_wait);
309 }
310
311 // IRQ-work handler that does deferred wakeup for call_rcu_tasks_generic().
312 static void call_rcu_tasks_iw_wakeup(struct irq_work *iwp)
313 {
314         struct rcu_tasks *rtp;
315         struct rcu_tasks_percpu *rtpcp = container_of(iwp, struct rcu_tasks_percpu, rtp_irq_work);
316
317         rtp = rtpcp->rtpp;
318         rcuwait_wake_up(&rtp->cbs_wait);
319 }
320
321 // Enqueue a callback for the specified flavor of Tasks RCU.
322 static void call_rcu_tasks_generic(struct rcu_head *rhp, rcu_callback_t func,
323                                    struct rcu_tasks *rtp)
324 {
325         int chosen_cpu;
326         unsigned long flags;
327         bool havekthread = smp_load_acquire(&rtp->kthread_ptr);
328         int ideal_cpu;
329         unsigned long j;
330         bool needadjust = false;
331         bool needwake;
332         struct rcu_tasks_percpu *rtpcp;
333
334         rhp->next = NULL;
335         rhp->func = func;
336         local_irq_save(flags);
337         rcu_read_lock();
338         ideal_cpu = smp_processor_id() >> READ_ONCE(rtp->percpu_enqueue_shift);
339         chosen_cpu = cpumask_next(ideal_cpu - 1, cpu_possible_mask);
340         rtpcp = per_cpu_ptr(rtp->rtpcpu, chosen_cpu);
341         if (!raw_spin_trylock_rcu_node(rtpcp)) { // irqs already disabled.
342                 raw_spin_lock_rcu_node(rtpcp); // irqs already disabled.
343                 j = jiffies;
344                 if (rtpcp->rtp_jiffies != j) {
345                         rtpcp->rtp_jiffies = j;
346                         rtpcp->rtp_n_lock_retries = 0;
347                 }
348                 if (rcu_task_cb_adjust && ++rtpcp->rtp_n_lock_retries > rcu_task_contend_lim &&
349                     READ_ONCE(rtp->percpu_enqueue_lim) != nr_cpu_ids)
350                         needadjust = true;  // Defer adjustment to avoid deadlock.
351         }
352         // Queuing callbacks before initialization not yet supported.
353         if (WARN_ON_ONCE(!rcu_segcblist_is_enabled(&rtpcp->cblist)))
354                 rcu_segcblist_init(&rtpcp->cblist);
355         needwake = (func == wakeme_after_rcu) ||
356                    (rcu_segcblist_n_cbs(&rtpcp->cblist) == rcu_task_lazy_lim);
357         if (havekthread && !needwake && !timer_pending(&rtpcp->lazy_timer)) {
358                 if (rtp->lazy_jiffies)
359                         mod_timer(&rtpcp->lazy_timer, rcu_tasks_lazy_time(rtp));
360                 else
361                         needwake = rcu_segcblist_empty(&rtpcp->cblist);
362         }
363         if (needwake)
364                 rtpcp->urgent_gp = 3;
365         rcu_segcblist_enqueue(&rtpcp->cblist, rhp);
366         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
367         if (unlikely(needadjust)) {
368                 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
369                 if (rtp->percpu_enqueue_lim != nr_cpu_ids) {
370                         WRITE_ONCE(rtp->percpu_enqueue_shift, 0);
371                         WRITE_ONCE(rtp->percpu_dequeue_lim, nr_cpu_ids);
372                         smp_store_release(&rtp->percpu_enqueue_lim, nr_cpu_ids);
373                         pr_info("Switching %s to per-CPU callback queuing.\n", rtp->name);
374                 }
375                 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
376         }
377         rcu_read_unlock();
378         /* We can't create the thread unless interrupts are enabled. */
379         if (needwake && READ_ONCE(rtp->kthread_ptr))
380                 irq_work_queue(&rtpcp->rtp_irq_work);
381 }
382
383 // RCU callback function for rcu_barrier_tasks_generic().
384 static void rcu_barrier_tasks_generic_cb(struct rcu_head *rhp)
385 {
386         struct rcu_tasks *rtp;
387         struct rcu_tasks_percpu *rtpcp;
388
389         rtpcp = container_of(rhp, struct rcu_tasks_percpu, barrier_q_head);
390         rtp = rtpcp->rtpp;
391         if (atomic_dec_and_test(&rtp->barrier_q_count))
392                 complete(&rtp->barrier_q_completion);
393 }
394
395 // Wait for all in-flight callbacks for the specified RCU Tasks flavor.
396 // Operates in a manner similar to rcu_barrier().
397 static void rcu_barrier_tasks_generic(struct rcu_tasks *rtp)
398 {
399         int cpu;
400         unsigned long flags;
401         struct rcu_tasks_percpu *rtpcp;
402         unsigned long s = rcu_seq_snap(&rtp->barrier_q_seq);
403
404         mutex_lock(&rtp->barrier_q_mutex);
405         if (rcu_seq_done(&rtp->barrier_q_seq, s)) {
406                 smp_mb();
407                 mutex_unlock(&rtp->barrier_q_mutex);
408                 return;
409         }
410         rcu_seq_start(&rtp->barrier_q_seq);
411         init_completion(&rtp->barrier_q_completion);
412         atomic_set(&rtp->barrier_q_count, 2);
413         for_each_possible_cpu(cpu) {
414                 if (cpu >= smp_load_acquire(&rtp->percpu_dequeue_lim))
415                         break;
416                 rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
417                 rtpcp->barrier_q_head.func = rcu_barrier_tasks_generic_cb;
418                 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
419                 if (rcu_segcblist_entrain(&rtpcp->cblist, &rtpcp->barrier_q_head))
420                         atomic_inc(&rtp->barrier_q_count);
421                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
422         }
423         if (atomic_sub_and_test(2, &rtp->barrier_q_count))
424                 complete(&rtp->barrier_q_completion);
425         wait_for_completion(&rtp->barrier_q_completion);
426         rcu_seq_end(&rtp->barrier_q_seq);
427         mutex_unlock(&rtp->barrier_q_mutex);
428 }
429
430 // Advance callbacks and indicate whether either a grace period or
431 // callback invocation is needed.
432 static int rcu_tasks_need_gpcb(struct rcu_tasks *rtp)
433 {
434         int cpu;
435         int dequeue_limit;
436         unsigned long flags;
437         bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
438         long n;
439         long ncbs = 0;
440         long ncbsnz = 0;
441         int needgpcb = 0;
442
443         dequeue_limit = smp_load_acquire(&rtp->percpu_dequeue_lim);
444         for (cpu = 0; cpu < dequeue_limit; cpu++) {
445                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
446
447                 /* Advance and accelerate any new callbacks. */
448                 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
449                         continue;
450                 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
451                 // Should we shrink down to a single callback queue?
452                 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
453                 if (n) {
454                         ncbs += n;
455                         if (cpu > 0)
456                                 ncbsnz += n;
457                 }
458                 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
459                 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
460                 if (rtpcp->urgent_gp > 0 && rcu_segcblist_pend_cbs(&rtpcp->cblist)) {
461                         if (rtp->lazy_jiffies)
462                                 rtpcp->urgent_gp--;
463                         needgpcb |= 0x3;
464                 } else if (rcu_segcblist_empty(&rtpcp->cblist)) {
465                         rtpcp->urgent_gp = 0;
466                 }
467                 if (rcu_segcblist_ready_cbs(&rtpcp->cblist))
468                         needgpcb |= 0x1;
469                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
470         }
471
472         // Shrink down to a single callback queue if appropriate.
473         // This is done in two stages: (1) If there are no more than
474         // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
475         // CPU, limit enqueueing to CPU 0.  (2) After an RCU grace period,
476         // if there has not been an increase in callbacks, limit dequeuing
477         // to CPU 0.  Note the matching RCU read-side critical section in
478         // call_rcu_tasks_generic().
479         if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
480                 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
481                 if (rtp->percpu_enqueue_lim > 1) {
482                         WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
483                         smp_store_release(&rtp->percpu_enqueue_lim, 1);
484                         rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
485                         gpdone = false;
486                         pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
487                 }
488                 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
489         }
490         if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
491                 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
492                 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
493                         WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
494                         pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
495                 }
496                 if (rtp->percpu_dequeue_lim == 1) {
497                         for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
498                                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
499
500                                 WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
501                         }
502                 }
503                 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
504         }
505
506         return needgpcb;
507 }
508
509 // Advance callbacks and invoke any that are ready.
510 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
511 {
512         int cpu;
513         int cpunext;
514         int cpuwq;
515         unsigned long flags;
516         int len;
517         struct rcu_head *rhp;
518         struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
519         struct rcu_tasks_percpu *rtpcp_next;
520
521         cpu = rtpcp->cpu;
522         cpunext = cpu * 2 + 1;
523         if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
524                 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
525                 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
526                 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
527                 cpunext++;
528                 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
529                         rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
530                         cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
531                         queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
532                 }
533         }
534
535         if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
536                 return;
537         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
538         rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
539         rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
540         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
541         len = rcl.len;
542         for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
543                 debug_rcu_head_callback(rhp);
544                 local_bh_disable();
545                 rhp->func(rhp);
546                 local_bh_enable();
547                 cond_resched();
548         }
549         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
550         rcu_segcblist_add_len(&rtpcp->cblist, -len);
551         (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
552         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
553 }
554
555 // Workqueue flood to advance callbacks and invoke any that are ready.
556 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
557 {
558         struct rcu_tasks *rtp;
559         struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
560
561         rtp = rtpcp->rtpp;
562         rcu_tasks_invoke_cbs(rtp, rtpcp);
563 }
564
565 // Wait for one grace period.
566 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
567 {
568         int needgpcb;
569
570         mutex_lock(&rtp->tasks_gp_mutex);
571
572         // If there were none, wait a bit and start over.
573         if (unlikely(midboot)) {
574                 needgpcb = 0x2;
575         } else {
576                 mutex_unlock(&rtp->tasks_gp_mutex);
577                 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
578                 rcuwait_wait_event(&rtp->cbs_wait,
579                                    (needgpcb = rcu_tasks_need_gpcb(rtp)),
580                                    TASK_IDLE);
581                 mutex_lock(&rtp->tasks_gp_mutex);
582         }
583
584         if (needgpcb & 0x2) {
585                 // Wait for one grace period.
586                 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
587                 rtp->gp_start = jiffies;
588                 rcu_seq_start(&rtp->tasks_gp_seq);
589                 rtp->gp_func(rtp);
590                 rcu_seq_end(&rtp->tasks_gp_seq);
591         }
592
593         // Invoke callbacks.
594         set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
595         rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
596         mutex_unlock(&rtp->tasks_gp_mutex);
597 }
598
599 // RCU-tasks kthread that detects grace periods and invokes callbacks.
600 static int __noreturn rcu_tasks_kthread(void *arg)
601 {
602         int cpu;
603         struct rcu_tasks *rtp = arg;
604
605         for_each_possible_cpu(cpu) {
606                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
607
608                 timer_setup(&rtpcp->lazy_timer, call_rcu_tasks_generic_timer, 0);
609                 rtpcp->urgent_gp = 1;
610         }
611
612         /* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
613         housekeeping_affine(current, HK_TYPE_RCU);
614         smp_store_release(&rtp->kthread_ptr, current); // Let GPs start!
615
616         /*
617          * Each pass through the following loop makes one check for
618          * newly arrived callbacks, and, if there are some, waits for
619          * one RCU-tasks grace period and then invokes the callbacks.
620          * This loop is terminated by the system going down.  ;-)
621          */
622         for (;;) {
623                 // Wait for one grace period and invoke any callbacks
624                 // that are ready.
625                 rcu_tasks_one_gp(rtp, false);
626
627                 // Paranoid sleep to keep this from entering a tight loop.
628                 schedule_timeout_idle(rtp->gp_sleep);
629         }
630 }
631
632 // Wait for a grace period for the specified flavor of Tasks RCU.
633 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
634 {
635         /* Complain if the scheduler has not started.  */
636         if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
637                          "synchronize_%s() called too soon", rtp->name))
638                 return;
639
640         // If the grace-period kthread is running, use it.
641         if (READ_ONCE(rtp->kthread_ptr)) {
642                 wait_rcu_gp(rtp->call_func);
643                 return;
644         }
645         rcu_tasks_one_gp(rtp, true);
646 }
647
648 /* Spawn RCU-tasks grace-period kthread. */
649 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
650 {
651         struct task_struct *t;
652
653         t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
654         if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
655                 return;
656         smp_mb(); /* Ensure others see full kthread. */
657 }
658
659 #ifndef CONFIG_TINY_RCU
660
661 /*
662  * Print any non-default Tasks RCU settings.
663  */
664 static void __init rcu_tasks_bootup_oddness(void)
665 {
666 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
667         int rtsimc;
668
669         if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
670                 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
671         rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
672         if (rtsimc != rcu_task_stall_info_mult) {
673                 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
674                 rcu_task_stall_info_mult = rtsimc;
675         }
676 #endif /* #ifdef CONFIG_TASKS_RCU */
677 #ifdef CONFIG_TASKS_RCU
678         pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
679 #endif /* #ifdef CONFIG_TASKS_RCU */
680 #ifdef CONFIG_TASKS_RUDE_RCU
681         pr_info("\tRude variant of Tasks RCU enabled.\n");
682 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
683 #ifdef CONFIG_TASKS_TRACE_RCU
684         pr_info("\tTracing variant of Tasks RCU enabled.\n");
685 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
686 }
687
688 #endif /* #ifndef CONFIG_TINY_RCU */
689
690 #ifndef CONFIG_TINY_RCU
691 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
692 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
693 {
694         int cpu;
695         bool havecbs = false;
696         bool haveurgent = false;
697         bool haveurgentcbs = false;
698
699         for_each_possible_cpu(cpu) {
700                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
701
702                 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)))
703                         havecbs = true;
704                 if (data_race(rtpcp->urgent_gp))
705                         haveurgent = true;
706                 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)) && data_race(rtpcp->urgent_gp))
707                         haveurgentcbs = true;
708                 if (havecbs && haveurgent && haveurgentcbs)
709                         break;
710         }
711         pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c%c%c l:%lu %s\n",
712                 rtp->kname,
713                 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
714                 jiffies - data_race(rtp->gp_jiffies),
715                 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
716                 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
717                 ".k"[!!data_race(rtp->kthread_ptr)],
718                 ".C"[havecbs],
719                 ".u"[haveurgent],
720                 ".U"[haveurgentcbs],
721                 rtp->lazy_jiffies,
722                 s);
723 }
724 #endif // #ifndef CONFIG_TINY_RCU
725
726 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
727
728 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
729
730 ////////////////////////////////////////////////////////////////////////
731 //
732 // Shared code between task-list-scanning variants of Tasks RCU.
733
734 /* Wait for one RCU-tasks grace period. */
735 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
736 {
737         struct task_struct *g;
738         int fract;
739         LIST_HEAD(holdouts);
740         unsigned long j;
741         unsigned long lastinfo;
742         unsigned long lastreport;
743         bool reported = false;
744         int rtsi;
745         struct task_struct *t;
746
747         set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
748         rtp->pregp_func(&holdouts);
749
750         /*
751          * There were callbacks, so we need to wait for an RCU-tasks
752          * grace period.  Start off by scanning the task list for tasks
753          * that are not already voluntarily blocked.  Mark these tasks
754          * and make a list of them in holdouts.
755          */
756         set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
757         if (rtp->pertask_func) {
758                 rcu_read_lock();
759                 for_each_process_thread(g, t)
760                         rtp->pertask_func(t, &holdouts);
761                 rcu_read_unlock();
762         }
763
764         set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
765         rtp->postscan_func(&holdouts);
766
767         /*
768          * Each pass through the following loop scans the list of holdout
769          * tasks, removing any that are no longer holdouts.  When the list
770          * is empty, we are done.
771          */
772         lastreport = jiffies;
773         lastinfo = lastreport;
774         rtsi = READ_ONCE(rcu_task_stall_info);
775
776         // Start off with initial wait and slowly back off to 1 HZ wait.
777         fract = rtp->init_fract;
778
779         while (!list_empty(&holdouts)) {
780                 ktime_t exp;
781                 bool firstreport;
782                 bool needreport;
783                 int rtst;
784
785                 // Slowly back off waiting for holdouts
786                 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
787                 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
788                         schedule_timeout_idle(fract);
789                 } else {
790                         exp = jiffies_to_nsecs(fract);
791                         __set_current_state(TASK_IDLE);
792                         schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
793                 }
794
795                 if (fract < HZ)
796                         fract++;
797
798                 rtst = READ_ONCE(rcu_task_stall_timeout);
799                 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
800                 if (needreport) {
801                         lastreport = jiffies;
802                         reported = true;
803                 }
804                 firstreport = true;
805                 WARN_ON(signal_pending(current));
806                 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
807                 rtp->holdouts_func(&holdouts, needreport, &firstreport);
808
809                 // Print pre-stall informational messages if needed.
810                 j = jiffies;
811                 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
812                         lastinfo = j;
813                         rtsi = rtsi * rcu_task_stall_info_mult;
814                         pr_info("%s: %s grace period number %lu (since boot) is %lu jiffies old.\n",
815                                 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
816                 }
817         }
818
819         set_tasks_gp_state(rtp, RTGS_POST_GP);
820         rtp->postgp_func(rtp);
821 }
822
823 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
824
825 #ifdef CONFIG_TASKS_RCU
826
827 ////////////////////////////////////////////////////////////////////////
828 //
829 // Simple variant of RCU whose quiescent states are voluntary context
830 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
831 // As such, grace periods can take one good long time.  There are no
832 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
833 // because this implementation is intended to get the system into a safe
834 // state for some of the manipulations involved in tracing and the like.
835 // Finally, this implementation does not support high call_rcu_tasks()
836 // rates from multiple CPUs.  If this is required, per-CPU callback lists
837 // will be needed.
838 //
839 // The implementation uses rcu_tasks_wait_gp(), which relies on function
840 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_kthread()
841 // function sets these function pointers up so that rcu_tasks_wait_gp()
842 // invokes these functions in this order:
843 //
844 // rcu_tasks_pregp_step():
845 //      Invokes synchronize_rcu() in order to wait for all in-flight
846 //      t->on_rq and t->nvcsw transitions to complete.  This works because
847 //      all such transitions are carried out with interrupts disabled.
848 // rcu_tasks_pertask(), invoked on every non-idle task:
849 //      For every runnable non-idle task other than the current one, use
850 //      get_task_struct() to pin down that task, snapshot that task's
851 //      number of voluntary context switches, and add that task to the
852 //      holdout list.
853 // rcu_tasks_postscan():
854 //      Invoke synchronize_srcu() to ensure that all tasks that were
855 //      in the process of exiting (and which thus might not know to
856 //      synchronize with this RCU Tasks grace period) have completed
857 //      exiting.
858 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
859 //      Scans the holdout list, attempting to identify a quiescent state
860 //      for each task on the list.  If there is a quiescent state, the
861 //      corresponding task is removed from the holdout list.
862 // rcu_tasks_postgp():
863 //      Invokes synchronize_rcu() in order to ensure that all prior
864 //      t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
865 //      to have happened before the end of this RCU Tasks grace period.
866 //      Again, this works because all such transitions are carried out
867 //      with interrupts disabled.
868 //
869 // For each exiting task, the exit_tasks_rcu_start() and
870 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
871 // read-side critical sections waited for by rcu_tasks_postscan().
872 //
873 // Pre-grace-period update-side code is ordered before the grace
874 // via the raw_spin_lock.*rcu_node().  Pre-grace-period read-side code
875 // is ordered before the grace period via synchronize_rcu() call in
876 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
877 // disabling.
878
879 /* Pre-grace-period preparation. */
880 static void rcu_tasks_pregp_step(struct list_head *hop)
881 {
882         /*
883          * Wait for all pre-existing t->on_rq and t->nvcsw transitions
884          * to complete.  Invoking synchronize_rcu() suffices because all
885          * these transitions occur with interrupts disabled.  Without this
886          * synchronize_rcu(), a read-side critical section that started
887          * before the grace period might be incorrectly seen as having
888          * started after the grace period.
889          *
890          * This synchronize_rcu() also dispenses with the need for a
891          * memory barrier on the first store to t->rcu_tasks_holdout,
892          * as it forces the store to happen after the beginning of the
893          * grace period.
894          */
895         synchronize_rcu();
896 }
897
898 /* Per-task initial processing. */
899 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
900 {
901         if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
902                 get_task_struct(t);
903                 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
904                 WRITE_ONCE(t->rcu_tasks_holdout, true);
905                 list_add(&t->rcu_tasks_holdout_list, hop);
906         }
907 }
908
909 /* Processing between scanning taskslist and draining the holdout list. */
910 static void rcu_tasks_postscan(struct list_head *hop)
911 {
912         int rtsi = READ_ONCE(rcu_task_stall_info);
913
914         if (!IS_ENABLED(CONFIG_TINY_RCU)) {
915                 tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
916                 add_timer(&tasks_rcu_exit_srcu_stall_timer);
917         }
918
919         /*
920          * Exiting tasks may escape the tasklist scan. Those are vulnerable
921          * until their final schedule() with TASK_DEAD state. To cope with
922          * this, divide the fragile exit path part in two intersecting
923          * read side critical sections:
924          *
925          * 1) An _SRCU_ read side starting before calling exit_notify(),
926          *    which may remove the task from the tasklist, and ending after
927          *    the final preempt_disable() call in do_exit().
928          *
929          * 2) An _RCU_ read side starting with the final preempt_disable()
930          *    call in do_exit() and ending with the final call to schedule()
931          *    with TASK_DEAD state.
932          *
933          * This handles the part 1). And postgp will handle part 2) with a
934          * call to synchronize_rcu().
935          */
936         synchronize_srcu(&tasks_rcu_exit_srcu);
937
938         if (!IS_ENABLED(CONFIG_TINY_RCU))
939                 del_timer_sync(&tasks_rcu_exit_srcu_stall_timer);
940 }
941
942 /* See if tasks are still holding out, complain if so. */
943 static void check_holdout_task(struct task_struct *t,
944                                bool needreport, bool *firstreport)
945 {
946         int cpu;
947
948         if (!READ_ONCE(t->rcu_tasks_holdout) ||
949             t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
950             !READ_ONCE(t->on_rq) ||
951             (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
952              !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
953                 WRITE_ONCE(t->rcu_tasks_holdout, false);
954                 list_del_init(&t->rcu_tasks_holdout_list);
955                 put_task_struct(t);
956                 return;
957         }
958         rcu_request_urgent_qs_task(t);
959         if (!needreport)
960                 return;
961         if (*firstreport) {
962                 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
963                 *firstreport = false;
964         }
965         cpu = task_cpu(t);
966         pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
967                  t, ".I"[is_idle_task(t)],
968                  "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
969                  t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
970                  t->rcu_tasks_idle_cpu, cpu);
971         sched_show_task(t);
972 }
973
974 /* Scan the holdout lists for tasks no longer holding out. */
975 static void check_all_holdout_tasks(struct list_head *hop,
976                                     bool needreport, bool *firstreport)
977 {
978         struct task_struct *t, *t1;
979
980         list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
981                 check_holdout_task(t, needreport, firstreport);
982                 cond_resched();
983         }
984 }
985
986 /* Finish off the Tasks-RCU grace period. */
987 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
988 {
989         /*
990          * Because ->on_rq and ->nvcsw are not guaranteed to have a full
991          * memory barriers prior to them in the schedule() path, memory
992          * reordering on other CPUs could cause their RCU-tasks read-side
993          * critical sections to extend past the end of the grace period.
994          * However, because these ->nvcsw updates are carried out with
995          * interrupts disabled, we can use synchronize_rcu() to force the
996          * needed ordering on all such CPUs.
997          *
998          * This synchronize_rcu() also confines all ->rcu_tasks_holdout
999          * accesses to be within the grace period, avoiding the need for
1000          * memory barriers for ->rcu_tasks_holdout accesses.
1001          *
1002          * In addition, this synchronize_rcu() waits for exiting tasks
1003          * to complete their final preempt_disable() region of execution,
1004          * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
1005          * enforcing the whole region before tasklist removal until
1006          * the final schedule() with TASK_DEAD state to be an RCU TASKS
1007          * read side critical section.
1008          */
1009         synchronize_rcu();
1010 }
1011
1012 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
1013 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
1014
1015 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused)
1016 {
1017 #ifndef CONFIG_TINY_RCU
1018         int rtsi;
1019
1020         rtsi = READ_ONCE(rcu_task_stall_info);
1021         pr_info("%s: %s grace period number %lu (since boot) gp_state: %s is %lu jiffies old.\n",
1022                 __func__, rcu_tasks.kname, rcu_tasks.tasks_gp_seq,
1023                 tasks_gp_state_getname(&rcu_tasks), jiffies - rcu_tasks.gp_jiffies);
1024         pr_info("Please check any exiting tasks stuck between calls to exit_tasks_rcu_start() and exit_tasks_rcu_finish()\n");
1025         tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
1026         add_timer(&tasks_rcu_exit_srcu_stall_timer);
1027 #endif // #ifndef CONFIG_TINY_RCU
1028 }
1029
1030 /**
1031  * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
1032  * @rhp: structure to be used for queueing the RCU updates.
1033  * @func: actual callback function to be invoked after the grace period
1034  *
1035  * The callback function will be invoked some time after a full grace
1036  * period elapses, in other words after all currently executing RCU
1037  * read-side critical sections have completed. call_rcu_tasks() assumes
1038  * that the read-side critical sections end at a voluntary context
1039  * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
1040  * or transition to usermode execution.  As such, there are no read-side
1041  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1042  * this primitive is intended to determine that all tasks have passed
1043  * through a safe state, not so much for data-structure synchronization.
1044  *
1045  * See the description of call_rcu() for more detailed information on
1046  * memory ordering guarantees.
1047  */
1048 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
1049 {
1050         call_rcu_tasks_generic(rhp, func, &rcu_tasks);
1051 }
1052 EXPORT_SYMBOL_GPL(call_rcu_tasks);
1053
1054 /**
1055  * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
1056  *
1057  * Control will return to the caller some time after a full rcu-tasks
1058  * grace period has elapsed, in other words after all currently
1059  * executing rcu-tasks read-side critical sections have elapsed.  These
1060  * read-side critical sections are delimited by calls to schedule(),
1061  * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
1062  * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
1063  *
1064  * This is a very specialized primitive, intended only for a few uses in
1065  * tracing and other situations requiring manipulation of function
1066  * preambles and profiling hooks.  The synchronize_rcu_tasks() function
1067  * is not (yet) intended for heavy use from multiple CPUs.
1068  *
1069  * See the description of synchronize_rcu() for more detailed information
1070  * on memory ordering guarantees.
1071  */
1072 void synchronize_rcu_tasks(void)
1073 {
1074         synchronize_rcu_tasks_generic(&rcu_tasks);
1075 }
1076 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
1077
1078 /**
1079  * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
1080  *
1081  * Although the current implementation is guaranteed to wait, it is not
1082  * obligated to, for example, if there are no pending callbacks.
1083  */
1084 void rcu_barrier_tasks(void)
1085 {
1086         rcu_barrier_tasks_generic(&rcu_tasks);
1087 }
1088 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
1089
1090 static int rcu_tasks_lazy_ms = -1;
1091 module_param(rcu_tasks_lazy_ms, int, 0444);
1092
1093 static int __init rcu_spawn_tasks_kthread(void)
1094 {
1095         cblist_init_generic(&rcu_tasks);
1096         rcu_tasks.gp_sleep = HZ / 10;
1097         rcu_tasks.init_fract = HZ / 10;
1098         if (rcu_tasks_lazy_ms >= 0)
1099                 rcu_tasks.lazy_jiffies = msecs_to_jiffies(rcu_tasks_lazy_ms);
1100         rcu_tasks.pregp_func = rcu_tasks_pregp_step;
1101         rcu_tasks.pertask_func = rcu_tasks_pertask;
1102         rcu_tasks.postscan_func = rcu_tasks_postscan;
1103         rcu_tasks.holdouts_func = check_all_holdout_tasks;
1104         rcu_tasks.postgp_func = rcu_tasks_postgp;
1105         rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1106         return 0;
1107 }
1108
1109 #if !defined(CONFIG_TINY_RCU)
1110 void show_rcu_tasks_classic_gp_kthread(void)
1111 {
1112         show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1113 }
1114 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1115 #endif // !defined(CONFIG_TINY_RCU)
1116
1117 struct task_struct *get_rcu_tasks_gp_kthread(void)
1118 {
1119         return rcu_tasks.kthread_ptr;
1120 }
1121 EXPORT_SYMBOL_GPL(get_rcu_tasks_gp_kthread);
1122
1123 /*
1124  * Contribute to protect against tasklist scan blind spot while the
1125  * task is exiting and may be removed from the tasklist. See
1126  * corresponding synchronize_srcu() for further details.
1127  */
1128 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
1129 {
1130         current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
1131 }
1132
1133 /*
1134  * Contribute to protect against tasklist scan blind spot while the
1135  * task is exiting and may be removed from the tasklist. See
1136  * corresponding synchronize_srcu() for further details.
1137  */
1138 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
1139 {
1140         struct task_struct *t = current;
1141
1142         __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1143 }
1144
1145 /*
1146  * Contribute to protect against tasklist scan blind spot while the
1147  * task is exiting and may be removed from the tasklist. See
1148  * corresponding synchronize_srcu() for further details.
1149  */
1150 void exit_tasks_rcu_finish(void)
1151 {
1152         exit_tasks_rcu_stop();
1153         exit_tasks_rcu_finish_trace(current);
1154 }
1155
1156 #else /* #ifdef CONFIG_TASKS_RCU */
1157 void exit_tasks_rcu_start(void) { }
1158 void exit_tasks_rcu_stop(void) { }
1159 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1160 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1161
1162 #ifdef CONFIG_TASKS_RUDE_RCU
1163
1164 ////////////////////////////////////////////////////////////////////////
1165 //
1166 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1167 // passing an empty function to schedule_on_each_cpu().  This approach
1168 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1169 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1170 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1171 // and induces otherwise unnecessary context switches on all online CPUs,
1172 // whether idle or not.
1173 //
1174 // Callback handling is provided by the rcu_tasks_kthread() function.
1175 //
1176 // Ordering is provided by the scheduler's context-switch code.
1177
1178 // Empty function to allow workqueues to force a context switch.
1179 static void rcu_tasks_be_rude(struct work_struct *work)
1180 {
1181 }
1182
1183 // Wait for one rude RCU-tasks grace period.
1184 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1185 {
1186         rtp->n_ipis += cpumask_weight(cpu_online_mask);
1187         schedule_on_each_cpu(rcu_tasks_be_rude);
1188 }
1189
1190 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1191 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1192                  "RCU Tasks Rude");
1193
1194 /**
1195  * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1196  * @rhp: structure to be used for queueing the RCU updates.
1197  * @func: actual callback function to be invoked after the grace period
1198  *
1199  * The callback function will be invoked some time after a full grace
1200  * period elapses, in other words after all currently executing RCU
1201  * read-side critical sections have completed. call_rcu_tasks_rude()
1202  * assumes that the read-side critical sections end at context switch,
1203  * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1204  * usermode execution is schedulable). As such, there are no read-side
1205  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1206  * this primitive is intended to determine that all tasks have passed
1207  * through a safe state, not so much for data-structure synchronization.
1208  *
1209  * See the description of call_rcu() for more detailed information on
1210  * memory ordering guarantees.
1211  */
1212 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1213 {
1214         call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1215 }
1216 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1217
1218 /**
1219  * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1220  *
1221  * Control will return to the caller some time after a rude rcu-tasks
1222  * grace period has elapsed, in other words after all currently
1223  * executing rcu-tasks read-side critical sections have elapsed.  These
1224  * read-side critical sections are delimited by calls to schedule(),
1225  * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1226  * context), and (in theory, anyway) cond_resched().
1227  *
1228  * This is a very specialized primitive, intended only for a few uses in
1229  * tracing and other situations requiring manipulation of function preambles
1230  * and profiling hooks.  The synchronize_rcu_tasks_rude() function is not
1231  * (yet) intended for heavy use from multiple CPUs.
1232  *
1233  * See the description of synchronize_rcu() for more detailed information
1234  * on memory ordering guarantees.
1235  */
1236 void synchronize_rcu_tasks_rude(void)
1237 {
1238         synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1239 }
1240 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1241
1242 /**
1243  * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1244  *
1245  * Although the current implementation is guaranteed to wait, it is not
1246  * obligated to, for example, if there are no pending callbacks.
1247  */
1248 void rcu_barrier_tasks_rude(void)
1249 {
1250         rcu_barrier_tasks_generic(&rcu_tasks_rude);
1251 }
1252 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1253
1254 int rcu_tasks_rude_lazy_ms = -1;
1255 module_param(rcu_tasks_rude_lazy_ms, int, 0444);
1256
1257 static int __init rcu_spawn_tasks_rude_kthread(void)
1258 {
1259         cblist_init_generic(&rcu_tasks_rude);
1260         rcu_tasks_rude.gp_sleep = HZ / 10;
1261         if (rcu_tasks_rude_lazy_ms >= 0)
1262                 rcu_tasks_rude.lazy_jiffies = msecs_to_jiffies(rcu_tasks_rude_lazy_ms);
1263         rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1264         return 0;
1265 }
1266
1267 #if !defined(CONFIG_TINY_RCU)
1268 void show_rcu_tasks_rude_gp_kthread(void)
1269 {
1270         show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1271 }
1272 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1273 #endif // !defined(CONFIG_TINY_RCU)
1274
1275 struct task_struct *get_rcu_tasks_rude_gp_kthread(void)
1276 {
1277         return rcu_tasks_rude.kthread_ptr;
1278 }
1279 EXPORT_SYMBOL_GPL(get_rcu_tasks_rude_gp_kthread);
1280
1281 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1282
1283 ////////////////////////////////////////////////////////////////////////
1284 //
1285 // Tracing variant of Tasks RCU.  This variant is designed to be used
1286 // to protect tracing hooks, including those of BPF.  This variant
1287 // therefore:
1288 //
1289 // 1.   Has explicit read-side markers to allow finite grace periods
1290 //      in the face of in-kernel loops for PREEMPT=n builds.
1291 //
1292 // 2.   Protects code in the idle loop, exception entry/exit, and
1293 //      CPU-hotplug code paths, similar to the capabilities of SRCU.
1294 //
1295 // 3.   Avoids expensive read-side instructions, having overhead similar
1296 //      to that of Preemptible RCU.
1297 //
1298 // There are of course downsides.  For example, the grace-period code
1299 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1300 // in nohz_full userspace.  If needed, these downsides can be at least
1301 // partially remedied.
1302 //
1303 // Perhaps most important, this variant of RCU does not affect the vanilla
1304 // flavors, rcu_preempt and rcu_sched.  The fact that RCU Tasks Trace
1305 // readers can operate from idle, offline, and exception entry/exit in no
1306 // way allows rcu_preempt and rcu_sched readers to also do so.
1307 //
1308 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1309 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_trace_kthread()
1310 // function sets these function pointers up so that rcu_tasks_wait_gp()
1311 // invokes these functions in this order:
1312 //
1313 // rcu_tasks_trace_pregp_step():
1314 //      Disables CPU hotplug, adds all currently executing tasks to the
1315 //      holdout list, then checks the state of all tasks that blocked
1316 //      or were preempted within their current RCU Tasks Trace read-side
1317 //      critical section, adding them to the holdout list if appropriate.
1318 //      Finally, this function re-enables CPU hotplug.
1319 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1320 // rcu_tasks_trace_postscan():
1321 //      Invokes synchronize_rcu() to wait for late-stage exiting tasks
1322 //      to finish exiting.
1323 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1324 //      Scans the holdout list, attempting to identify a quiescent state
1325 //      for each task on the list.  If there is a quiescent state, the
1326 //      corresponding task is removed from the holdout list.  Once this
1327 //      list is empty, the grace period has completed.
1328 // rcu_tasks_trace_postgp():
1329 //      Provides the needed full memory barrier and does debug checks.
1330 //
1331 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1332 //
1333 // Pre-grace-period update-side code is ordered before the grace period
1334 // via the ->cbs_lock and barriers in rcu_tasks_kthread().  Pre-grace-period
1335 // read-side code is ordered before the grace period by atomic operations
1336 // on .b.need_qs flag of each task involved in this process, or by scheduler
1337 // context-switch ordering (for locked-down non-running readers).
1338
1339 // The lockdep state must be outside of #ifdef to be useful.
1340 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1341 static struct lock_class_key rcu_lock_trace_key;
1342 struct lockdep_map rcu_trace_lock_map =
1343         STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1344 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1345 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1346
1347 #ifdef CONFIG_TASKS_TRACE_RCU
1348
1349 // Record outstanding IPIs to each CPU.  No point in sending two...
1350 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1351
1352 // The number of detections of task quiescent state relying on
1353 // heavyweight readers executing explicit memory barriers.
1354 static unsigned long n_heavy_reader_attempts;
1355 static unsigned long n_heavy_reader_updates;
1356 static unsigned long n_heavy_reader_ofl_updates;
1357 static unsigned long n_trc_holdouts;
1358
1359 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1360 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1361                  "RCU Tasks Trace");
1362
1363 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1364 static u8 rcu_ld_need_qs(struct task_struct *t)
1365 {
1366         smp_mb(); // Enforce full grace-period ordering.
1367         return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1368 }
1369
1370 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1371 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1372 {
1373         smp_store_release(&t->trc_reader_special.b.need_qs, v);
1374         smp_mb(); // Enforce full grace-period ordering.
1375 }
1376
1377 /*
1378  * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1379  * the four-byte operand-size restriction of some platforms.
1380  * Returns the old value, which is often ignored.
1381  */
1382 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1383 {
1384         union rcu_special ret;
1385         union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1386         union rcu_special trs_new = trs_old;
1387
1388         if (trs_old.b.need_qs != old)
1389                 return trs_old.b.need_qs;
1390         trs_new.b.need_qs = new;
1391         ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1392         return ret.b.need_qs;
1393 }
1394 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1395
1396 /*
1397  * If we are the last reader, signal the grace-period kthread.
1398  * Also remove from the per-CPU list of blocked tasks.
1399  */
1400 void rcu_read_unlock_trace_special(struct task_struct *t)
1401 {
1402         unsigned long flags;
1403         struct rcu_tasks_percpu *rtpcp;
1404         union rcu_special trs;
1405
1406         // Open-coded full-word version of rcu_ld_need_qs().
1407         smp_mb(); // Enforce full grace-period ordering.
1408         trs = smp_load_acquire(&t->trc_reader_special);
1409
1410         if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1411                 smp_mb(); // Pairs with update-side barriers.
1412         // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1413         if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1414                 u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1415                                                        TRC_NEED_QS_CHECKED);
1416
1417                 WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1418         }
1419         if (trs.b.blocked) {
1420                 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1421                 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1422                 list_del_init(&t->trc_blkd_node);
1423                 WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1424                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1425         }
1426         WRITE_ONCE(t->trc_reader_nesting, 0);
1427 }
1428 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1429
1430 /* Add a newly blocked reader task to its CPU's list. */
1431 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1432 {
1433         unsigned long flags;
1434         struct rcu_tasks_percpu *rtpcp;
1435
1436         local_irq_save(flags);
1437         rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1438         raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1439         t->trc_blkd_cpu = smp_processor_id();
1440         if (!rtpcp->rtp_blkd_tasks.next)
1441                 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1442         list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1443         WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1444         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1445 }
1446 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1447
1448 /* Add a task to the holdout list, if it is not already on the list. */
1449 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1450 {
1451         if (list_empty(&t->trc_holdout_list)) {
1452                 get_task_struct(t);
1453                 list_add(&t->trc_holdout_list, bhp);
1454                 n_trc_holdouts++;
1455         }
1456 }
1457
1458 /* Remove a task from the holdout list, if it is in fact present. */
1459 static void trc_del_holdout(struct task_struct *t)
1460 {
1461         if (!list_empty(&t->trc_holdout_list)) {
1462                 list_del_init(&t->trc_holdout_list);
1463                 put_task_struct(t);
1464                 n_trc_holdouts--;
1465         }
1466 }
1467
1468 /* IPI handler to check task state. */
1469 static void trc_read_check_handler(void *t_in)
1470 {
1471         int nesting;
1472         struct task_struct *t = current;
1473         struct task_struct *texp = t_in;
1474
1475         // If the task is no longer running on this CPU, leave.
1476         if (unlikely(texp != t))
1477                 goto reset_ipi; // Already on holdout list, so will check later.
1478
1479         // If the task is not in a read-side critical section, and
1480         // if this is the last reader, awaken the grace-period kthread.
1481         nesting = READ_ONCE(t->trc_reader_nesting);
1482         if (likely(!nesting)) {
1483                 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1484                 goto reset_ipi;
1485         }
1486         // If we are racing with an rcu_read_unlock_trace(), try again later.
1487         if (unlikely(nesting < 0))
1488                 goto reset_ipi;
1489
1490         // Get here if the task is in a read-side critical section.
1491         // Set its state so that it will update state for the grace-period
1492         // kthread upon exit from that critical section.
1493         rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1494
1495 reset_ipi:
1496         // Allow future IPIs to be sent on CPU and for task.
1497         // Also order this IPI handler against any later manipulations of
1498         // the intended task.
1499         smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1500         smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1501 }
1502
1503 /* Callback function for scheduler to check locked-down task.  */
1504 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1505 {
1506         struct list_head *bhp = bhp_in;
1507         int cpu = task_cpu(t);
1508         int nesting;
1509         bool ofl = cpu_is_offline(cpu);
1510
1511         if (task_curr(t) && !ofl) {
1512                 // If no chance of heavyweight readers, do it the hard way.
1513                 if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1514                         return -EINVAL;
1515
1516                 // If heavyweight readers are enabled on the remote task,
1517                 // we can inspect its state despite its currently running.
1518                 // However, we cannot safely change its state.
1519                 n_heavy_reader_attempts++;
1520                 // Check for "running" idle tasks on offline CPUs.
1521                 if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1522                         return -EINVAL; // No quiescent state, do it the hard way.
1523                 n_heavy_reader_updates++;
1524                 nesting = 0;
1525         } else {
1526                 // The task is not running, so C-language access is safe.
1527                 nesting = t->trc_reader_nesting;
1528                 WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1529                 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1530                         n_heavy_reader_ofl_updates++;
1531         }
1532
1533         // If not exiting a read-side critical section, mark as checked
1534         // so that the grace-period kthread will remove it from the
1535         // holdout list.
1536         if (!nesting) {
1537                 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1538                 return 0;  // In QS, so done.
1539         }
1540         if (nesting < 0)
1541                 return -EINVAL; // Reader transitioning, try again later.
1542
1543         // The task is in a read-side critical section, so set up its
1544         // state so that it will update state upon exit from that critical
1545         // section.
1546         if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1547                 trc_add_holdout(t, bhp);
1548         return 0;
1549 }
1550
1551 /* Attempt to extract the state for the specified task. */
1552 static void trc_wait_for_one_reader(struct task_struct *t,
1553                                     struct list_head *bhp)
1554 {
1555         int cpu;
1556
1557         // If a previous IPI is still in flight, let it complete.
1558         if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1559                 return;
1560
1561         // The current task had better be in a quiescent state.
1562         if (t == current) {
1563                 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1564                 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1565                 return;
1566         }
1567
1568         // Attempt to nail down the task for inspection.
1569         get_task_struct(t);
1570         if (!task_call_func(t, trc_inspect_reader, bhp)) {
1571                 put_task_struct(t);
1572                 return;
1573         }
1574         put_task_struct(t);
1575
1576         // If this task is not yet on the holdout list, then we are in
1577         // an RCU read-side critical section.  Otherwise, the invocation of
1578         // trc_add_holdout() that added it to the list did the necessary
1579         // get_task_struct().  Either way, the task cannot be freed out
1580         // from under this code.
1581
1582         // If currently running, send an IPI, either way, add to list.
1583         trc_add_holdout(t, bhp);
1584         if (task_curr(t) &&
1585             time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1586                 // The task is currently running, so try IPIing it.
1587                 cpu = task_cpu(t);
1588
1589                 // If there is already an IPI outstanding, let it happen.
1590                 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1591                         return;
1592
1593                 per_cpu(trc_ipi_to_cpu, cpu) = true;
1594                 t->trc_ipi_to_cpu = cpu;
1595                 rcu_tasks_trace.n_ipis++;
1596                 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1597                         // Just in case there is some other reason for
1598                         // failure than the target CPU being offline.
1599                         WARN_ONCE(1, "%s():  smp_call_function_single() failed for CPU: %d\n",
1600                                   __func__, cpu);
1601                         rcu_tasks_trace.n_ipis_fails++;
1602                         per_cpu(trc_ipi_to_cpu, cpu) = false;
1603                         t->trc_ipi_to_cpu = -1;
1604                 }
1605         }
1606 }
1607
1608 /*
1609  * Initialize for first-round processing for the specified task.
1610  * Return false if task is NULL or already taken care of, true otherwise.
1611  */
1612 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1613 {
1614         // During early boot when there is only the one boot CPU, there
1615         // is no idle task for the other CPUs.  Also, the grace-period
1616         // kthread is always in a quiescent state.  In addition, just return
1617         // if this task is already on the list.
1618         if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1619                 return false;
1620
1621         rcu_st_need_qs(t, 0);
1622         t->trc_ipi_to_cpu = -1;
1623         return true;
1624 }
1625
1626 /* Do first-round processing for the specified task. */
1627 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1628 {
1629         if (rcu_tasks_trace_pertask_prep(t, true))
1630                 trc_wait_for_one_reader(t, hop);
1631 }
1632
1633 /* Initialize for a new RCU-tasks-trace grace period. */
1634 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1635 {
1636         LIST_HEAD(blkd_tasks);
1637         int cpu;
1638         unsigned long flags;
1639         struct rcu_tasks_percpu *rtpcp;
1640         struct task_struct *t;
1641
1642         // There shouldn't be any old IPIs, but...
1643         for_each_possible_cpu(cpu)
1644                 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1645
1646         // Disable CPU hotplug across the CPU scan for the benefit of
1647         // any IPIs that might be needed.  This also waits for all readers
1648         // in CPU-hotplug code paths.
1649         cpus_read_lock();
1650
1651         // These rcu_tasks_trace_pertask_prep() calls are serialized to
1652         // allow safe access to the hop list.
1653         for_each_online_cpu(cpu) {
1654                 rcu_read_lock();
1655                 t = cpu_curr_snapshot(cpu);
1656                 if (rcu_tasks_trace_pertask_prep(t, true))
1657                         trc_add_holdout(t, hop);
1658                 rcu_read_unlock();
1659                 cond_resched_tasks_rcu_qs();
1660         }
1661
1662         // Only after all running tasks have been accounted for is it
1663         // safe to take care of the tasks that have blocked within their
1664         // current RCU tasks trace read-side critical section.
1665         for_each_possible_cpu(cpu) {
1666                 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1667                 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1668                 list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1669                 while (!list_empty(&blkd_tasks)) {
1670                         rcu_read_lock();
1671                         t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1672                         list_del_init(&t->trc_blkd_node);
1673                         list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1674                         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1675                         rcu_tasks_trace_pertask(t, hop);
1676                         rcu_read_unlock();
1677                         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1678                 }
1679                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1680                 cond_resched_tasks_rcu_qs();
1681         }
1682
1683         // Re-enable CPU hotplug now that the holdout list is populated.
1684         cpus_read_unlock();
1685 }
1686
1687 /*
1688  * Do intermediate processing between task and holdout scans.
1689  */
1690 static void rcu_tasks_trace_postscan(struct list_head *hop)
1691 {
1692         // Wait for late-stage exiting tasks to finish exiting.
1693         // These might have passed the call to exit_tasks_rcu_finish().
1694
1695         // If you remove the following line, update rcu_trace_implies_rcu_gp()!!!
1696         synchronize_rcu();
1697         // Any tasks that exit after this point will set
1698         // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1699 }
1700
1701 /* Communicate task state back to the RCU tasks trace stall warning request. */
1702 struct trc_stall_chk_rdr {
1703         int nesting;
1704         int ipi_to_cpu;
1705         u8 needqs;
1706 };
1707
1708 static int trc_check_slow_task(struct task_struct *t, void *arg)
1709 {
1710         struct trc_stall_chk_rdr *trc_rdrp = arg;
1711
1712         if (task_curr(t) && cpu_online(task_cpu(t)))
1713                 return false; // It is running, so decline to inspect it.
1714         trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1715         trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1716         trc_rdrp->needqs = rcu_ld_need_qs(t);
1717         return true;
1718 }
1719
1720 /* Show the state of a task stalling the current RCU tasks trace GP. */
1721 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1722 {
1723         int cpu;
1724         struct trc_stall_chk_rdr trc_rdr;
1725         bool is_idle_tsk = is_idle_task(t);
1726
1727         if (*firstreport) {
1728                 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1729                 *firstreport = false;
1730         }
1731         cpu = task_cpu(t);
1732         if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1733                 pr_alert("P%d: %c%c\n",
1734                          t->pid,
1735                          ".I"[t->trc_ipi_to_cpu >= 0],
1736                          ".i"[is_idle_tsk]);
1737         else
1738                 pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1739                          t->pid,
1740                          ".I"[trc_rdr.ipi_to_cpu >= 0],
1741                          ".i"[is_idle_tsk],
1742                          ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1743                          ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1744                          trc_rdr.nesting,
1745                          " !CN"[trc_rdr.needqs & 0x3],
1746                          " ?"[trc_rdr.needqs > 0x3],
1747                          cpu, cpu_online(cpu) ? "" : "(offline)");
1748         sched_show_task(t);
1749 }
1750
1751 /* List stalled IPIs for RCU tasks trace. */
1752 static void show_stalled_ipi_trace(void)
1753 {
1754         int cpu;
1755
1756         for_each_possible_cpu(cpu)
1757                 if (per_cpu(trc_ipi_to_cpu, cpu))
1758                         pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1759 }
1760
1761 /* Do one scan of the holdout list. */
1762 static void check_all_holdout_tasks_trace(struct list_head *hop,
1763                                           bool needreport, bool *firstreport)
1764 {
1765         struct task_struct *g, *t;
1766
1767         // Disable CPU hotplug across the holdout list scan for IPIs.
1768         cpus_read_lock();
1769
1770         list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1771                 // If safe and needed, try to check the current task.
1772                 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1773                     !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1774                         trc_wait_for_one_reader(t, hop);
1775
1776                 // If check succeeded, remove this task from the list.
1777                 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1778                     rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1779                         trc_del_holdout(t);
1780                 else if (needreport)
1781                         show_stalled_task_trace(t, firstreport);
1782                 cond_resched_tasks_rcu_qs();
1783         }
1784
1785         // Re-enable CPU hotplug now that the holdout list scan has completed.
1786         cpus_read_unlock();
1787
1788         if (needreport) {
1789                 if (*firstreport)
1790                         pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1791                 show_stalled_ipi_trace();
1792         }
1793 }
1794
1795 static void rcu_tasks_trace_empty_fn(void *unused)
1796 {
1797 }
1798
1799 /* Wait for grace period to complete and provide ordering. */
1800 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1801 {
1802         int cpu;
1803
1804         // Wait for any lingering IPI handlers to complete.  Note that
1805         // if a CPU has gone offline or transitioned to userspace in the
1806         // meantime, all IPI handlers should have been drained beforehand.
1807         // Yes, this assumes that CPUs process IPIs in order.  If that ever
1808         // changes, there will need to be a recheck and/or timed wait.
1809         for_each_online_cpu(cpu)
1810                 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1811                         smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1812
1813         smp_mb(); // Caller's code must be ordered after wakeup.
1814                   // Pairs with pretty much every ordering primitive.
1815 }
1816
1817 /* Report any needed quiescent state for this exiting task. */
1818 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1819 {
1820         union rcu_special trs = READ_ONCE(t->trc_reader_special);
1821
1822         rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1823         WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1824         if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1825                 rcu_read_unlock_trace_special(t);
1826         else
1827                 WRITE_ONCE(t->trc_reader_nesting, 0);
1828 }
1829
1830 /**
1831  * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1832  * @rhp: structure to be used for queueing the RCU updates.
1833  * @func: actual callback function to be invoked after the grace period
1834  *
1835  * The callback function will be invoked some time after a trace rcu-tasks
1836  * grace period elapses, in other words after all currently executing
1837  * trace rcu-tasks read-side critical sections have completed. These
1838  * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1839  * and rcu_read_unlock_trace().
1840  *
1841  * See the description of call_rcu() for more detailed information on
1842  * memory ordering guarantees.
1843  */
1844 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1845 {
1846         call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1847 }
1848 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1849
1850 /**
1851  * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1852  *
1853  * Control will return to the caller some time after a trace rcu-tasks
1854  * grace period has elapsed, in other words after all currently executing
1855  * trace rcu-tasks read-side critical sections have elapsed. These read-side
1856  * critical sections are delimited by calls to rcu_read_lock_trace()
1857  * and rcu_read_unlock_trace().
1858  *
1859  * This is a very specialized primitive, intended only for a few uses in
1860  * tracing and other situations requiring manipulation of function preambles
1861  * and profiling hooks.  The synchronize_rcu_tasks_trace() function is not
1862  * (yet) intended for heavy use from multiple CPUs.
1863  *
1864  * See the description of synchronize_rcu() for more detailed information
1865  * on memory ordering guarantees.
1866  */
1867 void synchronize_rcu_tasks_trace(void)
1868 {
1869         RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1870         synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1871 }
1872 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1873
1874 /**
1875  * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1876  *
1877  * Although the current implementation is guaranteed to wait, it is not
1878  * obligated to, for example, if there are no pending callbacks.
1879  */
1880 void rcu_barrier_tasks_trace(void)
1881 {
1882         rcu_barrier_tasks_generic(&rcu_tasks_trace);
1883 }
1884 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1885
1886 int rcu_tasks_trace_lazy_ms = -1;
1887 module_param(rcu_tasks_trace_lazy_ms, int, 0444);
1888
1889 static int __init rcu_spawn_tasks_trace_kthread(void)
1890 {
1891         cblist_init_generic(&rcu_tasks_trace);
1892         if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1893                 rcu_tasks_trace.gp_sleep = HZ / 10;
1894                 rcu_tasks_trace.init_fract = HZ / 10;
1895         } else {
1896                 rcu_tasks_trace.gp_sleep = HZ / 200;
1897                 if (rcu_tasks_trace.gp_sleep <= 0)
1898                         rcu_tasks_trace.gp_sleep = 1;
1899                 rcu_tasks_trace.init_fract = HZ / 200;
1900                 if (rcu_tasks_trace.init_fract <= 0)
1901                         rcu_tasks_trace.init_fract = 1;
1902         }
1903         if (rcu_tasks_trace_lazy_ms >= 0)
1904                 rcu_tasks_trace.lazy_jiffies = msecs_to_jiffies(rcu_tasks_trace_lazy_ms);
1905         rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1906         rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1907         rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1908         rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1909         rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1910         return 0;
1911 }
1912
1913 #if !defined(CONFIG_TINY_RCU)
1914 void show_rcu_tasks_trace_gp_kthread(void)
1915 {
1916         char buf[64];
1917
1918         sprintf(buf, "N%lu h:%lu/%lu/%lu",
1919                 data_race(n_trc_holdouts),
1920                 data_race(n_heavy_reader_ofl_updates),
1921                 data_race(n_heavy_reader_updates),
1922                 data_race(n_heavy_reader_attempts));
1923         show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1924 }
1925 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1926 #endif // !defined(CONFIG_TINY_RCU)
1927
1928 struct task_struct *get_rcu_tasks_trace_gp_kthread(void)
1929 {
1930         return rcu_tasks_trace.kthread_ptr;
1931 }
1932 EXPORT_SYMBOL_GPL(get_rcu_tasks_trace_gp_kthread);
1933
1934 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1935 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1936 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1937
1938 #ifndef CONFIG_TINY_RCU
1939 void show_rcu_tasks_gp_kthreads(void)
1940 {
1941         show_rcu_tasks_classic_gp_kthread();
1942         show_rcu_tasks_rude_gp_kthread();
1943         show_rcu_tasks_trace_gp_kthread();
1944 }
1945 #endif /* #ifndef CONFIG_TINY_RCU */
1946
1947 #ifdef CONFIG_PROVE_RCU
1948 struct rcu_tasks_test_desc {
1949         struct rcu_head rh;
1950         const char *name;
1951         bool notrun;
1952         unsigned long runstart;
1953 };
1954
1955 static struct rcu_tasks_test_desc tests[] = {
1956         {
1957                 .name = "call_rcu_tasks()",
1958                 /* If not defined, the test is skipped. */
1959                 .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1960         },
1961         {
1962                 .name = "call_rcu_tasks_rude()",
1963                 /* If not defined, the test is skipped. */
1964                 .notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1965         },
1966         {
1967                 .name = "call_rcu_tasks_trace()",
1968                 /* If not defined, the test is skipped. */
1969                 .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1970         }
1971 };
1972
1973 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1974 {
1975         struct rcu_tasks_test_desc *rttd =
1976                 container_of(rhp, struct rcu_tasks_test_desc, rh);
1977
1978         pr_info("Callback from %s invoked.\n", rttd->name);
1979
1980         rttd->notrun = false;
1981 }
1982
1983 static void rcu_tasks_initiate_self_tests(void)
1984 {
1985 #ifdef CONFIG_TASKS_RCU
1986         pr_info("Running RCU Tasks wait API self tests\n");
1987         tests[0].runstart = jiffies;
1988         synchronize_rcu_tasks();
1989         call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1990 #endif
1991
1992 #ifdef CONFIG_TASKS_RUDE_RCU
1993         pr_info("Running RCU Tasks Rude wait API self tests\n");
1994         tests[1].runstart = jiffies;
1995         synchronize_rcu_tasks_rude();
1996         call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1997 #endif
1998
1999 #ifdef CONFIG_TASKS_TRACE_RCU
2000         pr_info("Running RCU Tasks Trace wait API self tests\n");
2001         tests[2].runstart = jiffies;
2002         synchronize_rcu_tasks_trace();
2003         call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
2004 #endif
2005 }
2006
2007 /*
2008  * Return:  0 - test passed
2009  *          1 - test failed, but have not timed out yet
2010  *         -1 - test failed and timed out
2011  */
2012 static int rcu_tasks_verify_self_tests(void)
2013 {
2014         int ret = 0;
2015         int i;
2016         unsigned long bst = rcu_task_stall_timeout;
2017
2018         if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
2019                 bst = RCU_TASK_BOOT_STALL_TIMEOUT;
2020         for (i = 0; i < ARRAY_SIZE(tests); i++) {
2021                 while (tests[i].notrun) {               // still hanging.
2022                         if (time_after(jiffies, tests[i].runstart + bst)) {
2023                                 pr_err("%s has failed boot-time tests.\n", tests[i].name);
2024                                 ret = -1;
2025                                 break;
2026                         }
2027                         ret = 1;
2028                         break;
2029                 }
2030         }
2031         WARN_ON(ret < 0);
2032
2033         return ret;
2034 }
2035
2036 /*
2037  * Repeat the rcu_tasks_verify_self_tests() call once every second until the
2038  * test passes or has timed out.
2039  */
2040 static struct delayed_work rcu_tasks_verify_work;
2041 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
2042 {
2043         int ret = rcu_tasks_verify_self_tests();
2044
2045         if (ret <= 0)
2046                 return;
2047
2048         /* Test fails but not timed out yet, reschedule another check */
2049         schedule_delayed_work(&rcu_tasks_verify_work, HZ);
2050 }
2051
2052 static int rcu_tasks_verify_schedule_work(void)
2053 {
2054         INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
2055         rcu_tasks_verify_work_fn(NULL);
2056         return 0;
2057 }
2058 late_initcall(rcu_tasks_verify_schedule_work);
2059 #else /* #ifdef CONFIG_PROVE_RCU */
2060 static void rcu_tasks_initiate_self_tests(void) { }
2061 #endif /* #else #ifdef CONFIG_PROVE_RCU */
2062
2063 void __init rcu_init_tasks_generic(void)
2064 {
2065 #ifdef CONFIG_TASKS_RCU
2066         rcu_spawn_tasks_kthread();
2067 #endif
2068
2069 #ifdef CONFIG_TASKS_RUDE_RCU
2070         rcu_spawn_tasks_rude_kthread();
2071 #endif
2072
2073 #ifdef CONFIG_TASKS_TRACE_RCU
2074         rcu_spawn_tasks_trace_kthread();
2075 #endif
2076
2077         // Run the self-tests.
2078         rcu_tasks_initiate_self_tests();
2079 }
2080
2081 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
2082 static inline void rcu_tasks_bootup_oddness(void) {}
2083 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */