Merge tag 'ceph-for-6.6-rc1' of https://github.com/ceph/ceph-client
[platform/kernel/linux-starfive.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         unsigned long flags;
436         bool gpdone = poll_state_synchronize_rcu(rtp->percpu_dequeue_gpseq);
437         long n;
438         long ncbs = 0;
439         long ncbsnz = 0;
440         int needgpcb = 0;
441
442         for (cpu = 0; cpu < smp_load_acquire(&rtp->percpu_dequeue_lim); cpu++) {
443                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
444
445                 /* Advance and accelerate any new callbacks. */
446                 if (!rcu_segcblist_n_cbs(&rtpcp->cblist))
447                         continue;
448                 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
449                 // Should we shrink down to a single callback queue?
450                 n = rcu_segcblist_n_cbs(&rtpcp->cblist);
451                 if (n) {
452                         ncbs += n;
453                         if (cpu > 0)
454                                 ncbsnz += n;
455                 }
456                 rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
457                 (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
458                 if (rtpcp->urgent_gp > 0 && rcu_segcblist_pend_cbs(&rtpcp->cblist)) {
459                         if (rtp->lazy_jiffies)
460                                 rtpcp->urgent_gp--;
461                         needgpcb |= 0x3;
462                 } else if (rcu_segcblist_empty(&rtpcp->cblist)) {
463                         rtpcp->urgent_gp = 0;
464                 }
465                 if (rcu_segcblist_ready_cbs(&rtpcp->cblist))
466                         needgpcb |= 0x1;
467                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
468         }
469
470         // Shrink down to a single callback queue if appropriate.
471         // This is done in two stages: (1) If there are no more than
472         // rcu_task_collapse_lim callbacks on CPU 0 and none on any other
473         // CPU, limit enqueueing to CPU 0.  (2) After an RCU grace period,
474         // if there has not been an increase in callbacks, limit dequeuing
475         // to CPU 0.  Note the matching RCU read-side critical section in
476         // call_rcu_tasks_generic().
477         if (rcu_task_cb_adjust && ncbs <= rcu_task_collapse_lim) {
478                 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
479                 if (rtp->percpu_enqueue_lim > 1) {
480                         WRITE_ONCE(rtp->percpu_enqueue_shift, order_base_2(nr_cpu_ids));
481                         smp_store_release(&rtp->percpu_enqueue_lim, 1);
482                         rtp->percpu_dequeue_gpseq = get_state_synchronize_rcu();
483                         gpdone = false;
484                         pr_info("Starting switch %s to CPU-0 callback queuing.\n", rtp->name);
485                 }
486                 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
487         }
488         if (rcu_task_cb_adjust && !ncbsnz && gpdone) {
489                 raw_spin_lock_irqsave(&rtp->cbs_gbl_lock, flags);
490                 if (rtp->percpu_enqueue_lim < rtp->percpu_dequeue_lim) {
491                         WRITE_ONCE(rtp->percpu_dequeue_lim, 1);
492                         pr_info("Completing switch %s to CPU-0 callback queuing.\n", rtp->name);
493                 }
494                 if (rtp->percpu_dequeue_lim == 1) {
495                         for (cpu = rtp->percpu_dequeue_lim; cpu < nr_cpu_ids; cpu++) {
496                                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
497
498                                 WARN_ON_ONCE(rcu_segcblist_n_cbs(&rtpcp->cblist));
499                         }
500                 }
501                 raw_spin_unlock_irqrestore(&rtp->cbs_gbl_lock, flags);
502         }
503
504         return needgpcb;
505 }
506
507 // Advance callbacks and invoke any that are ready.
508 static void rcu_tasks_invoke_cbs(struct rcu_tasks *rtp, struct rcu_tasks_percpu *rtpcp)
509 {
510         int cpu;
511         int cpunext;
512         int cpuwq;
513         unsigned long flags;
514         int len;
515         struct rcu_head *rhp;
516         struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
517         struct rcu_tasks_percpu *rtpcp_next;
518
519         cpu = rtpcp->cpu;
520         cpunext = cpu * 2 + 1;
521         if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
522                 rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
523                 cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
524                 queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
525                 cpunext++;
526                 if (cpunext < smp_load_acquire(&rtp->percpu_dequeue_lim)) {
527                         rtpcp_next = per_cpu_ptr(rtp->rtpcpu, cpunext);
528                         cpuwq = rcu_cpu_beenfullyonline(cpunext) ? cpunext : WORK_CPU_UNBOUND;
529                         queue_work_on(cpuwq, system_wq, &rtpcp_next->rtp_work);
530                 }
531         }
532
533         if (rcu_segcblist_empty(&rtpcp->cblist) || !cpu_possible(cpu))
534                 return;
535         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
536         rcu_segcblist_advance(&rtpcp->cblist, rcu_seq_current(&rtp->tasks_gp_seq));
537         rcu_segcblist_extract_done_cbs(&rtpcp->cblist, &rcl);
538         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
539         len = rcl.len;
540         for (rhp = rcu_cblist_dequeue(&rcl); rhp; rhp = rcu_cblist_dequeue(&rcl)) {
541                 local_bh_disable();
542                 rhp->func(rhp);
543                 local_bh_enable();
544                 cond_resched();
545         }
546         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
547         rcu_segcblist_add_len(&rtpcp->cblist, -len);
548         (void)rcu_segcblist_accelerate(&rtpcp->cblist, rcu_seq_snap(&rtp->tasks_gp_seq));
549         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
550 }
551
552 // Workqueue flood to advance callbacks and invoke any that are ready.
553 static void rcu_tasks_invoke_cbs_wq(struct work_struct *wp)
554 {
555         struct rcu_tasks *rtp;
556         struct rcu_tasks_percpu *rtpcp = container_of(wp, struct rcu_tasks_percpu, rtp_work);
557
558         rtp = rtpcp->rtpp;
559         rcu_tasks_invoke_cbs(rtp, rtpcp);
560 }
561
562 // Wait for one grace period.
563 static void rcu_tasks_one_gp(struct rcu_tasks *rtp, bool midboot)
564 {
565         int needgpcb;
566
567         mutex_lock(&rtp->tasks_gp_mutex);
568
569         // If there were none, wait a bit and start over.
570         if (unlikely(midboot)) {
571                 needgpcb = 0x2;
572         } else {
573                 mutex_unlock(&rtp->tasks_gp_mutex);
574                 set_tasks_gp_state(rtp, RTGS_WAIT_CBS);
575                 rcuwait_wait_event(&rtp->cbs_wait,
576                                    (needgpcb = rcu_tasks_need_gpcb(rtp)),
577                                    TASK_IDLE);
578                 mutex_lock(&rtp->tasks_gp_mutex);
579         }
580
581         if (needgpcb & 0x2) {
582                 // Wait for one grace period.
583                 set_tasks_gp_state(rtp, RTGS_WAIT_GP);
584                 rtp->gp_start = jiffies;
585                 rcu_seq_start(&rtp->tasks_gp_seq);
586                 rtp->gp_func(rtp);
587                 rcu_seq_end(&rtp->tasks_gp_seq);
588         }
589
590         // Invoke callbacks.
591         set_tasks_gp_state(rtp, RTGS_INVOKE_CBS);
592         rcu_tasks_invoke_cbs(rtp, per_cpu_ptr(rtp->rtpcpu, 0));
593         mutex_unlock(&rtp->tasks_gp_mutex);
594 }
595
596 // RCU-tasks kthread that detects grace periods and invokes callbacks.
597 static int __noreturn rcu_tasks_kthread(void *arg)
598 {
599         int cpu;
600         struct rcu_tasks *rtp = arg;
601
602         for_each_possible_cpu(cpu) {
603                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
604
605                 timer_setup(&rtpcp->lazy_timer, call_rcu_tasks_generic_timer, 0);
606                 rtpcp->urgent_gp = 1;
607         }
608
609         /* Run on housekeeping CPUs by default.  Sysadm can move if desired. */
610         housekeeping_affine(current, HK_TYPE_RCU);
611         smp_store_release(&rtp->kthread_ptr, current); // Let GPs start!
612
613         /*
614          * Each pass through the following loop makes one check for
615          * newly arrived callbacks, and, if there are some, waits for
616          * one RCU-tasks grace period and then invokes the callbacks.
617          * This loop is terminated by the system going down.  ;-)
618          */
619         for (;;) {
620                 // Wait for one grace period and invoke any callbacks
621                 // that are ready.
622                 rcu_tasks_one_gp(rtp, false);
623
624                 // Paranoid sleep to keep this from entering a tight loop.
625                 schedule_timeout_idle(rtp->gp_sleep);
626         }
627 }
628
629 // Wait for a grace period for the specified flavor of Tasks RCU.
630 static void synchronize_rcu_tasks_generic(struct rcu_tasks *rtp)
631 {
632         /* Complain if the scheduler has not started.  */
633         if (WARN_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INACTIVE,
634                          "synchronize_%s() called too soon", rtp->name))
635                 return;
636
637         // If the grace-period kthread is running, use it.
638         if (READ_ONCE(rtp->kthread_ptr)) {
639                 wait_rcu_gp(rtp->call_func);
640                 return;
641         }
642         rcu_tasks_one_gp(rtp, true);
643 }
644
645 /* Spawn RCU-tasks grace-period kthread. */
646 static void __init rcu_spawn_tasks_kthread_generic(struct rcu_tasks *rtp)
647 {
648         struct task_struct *t;
649
650         t = kthread_run(rcu_tasks_kthread, rtp, "%s_kthread", rtp->kname);
651         if (WARN_ONCE(IS_ERR(t), "%s: Could not start %s grace-period kthread, OOM is now expected behavior\n", __func__, rtp->name))
652                 return;
653         smp_mb(); /* Ensure others see full kthread. */
654 }
655
656 #ifndef CONFIG_TINY_RCU
657
658 /*
659  * Print any non-default Tasks RCU settings.
660  */
661 static void __init rcu_tasks_bootup_oddness(void)
662 {
663 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
664         int rtsimc;
665
666         if (rcu_task_stall_timeout != RCU_TASK_STALL_TIMEOUT)
667                 pr_info("\tTasks-RCU CPU stall warnings timeout set to %d (rcu_task_stall_timeout).\n", rcu_task_stall_timeout);
668         rtsimc = clamp(rcu_task_stall_info_mult, 1, 10);
669         if (rtsimc != rcu_task_stall_info_mult) {
670                 pr_info("\tTasks-RCU CPU stall info multiplier clamped to %d (rcu_task_stall_info_mult).\n", rtsimc);
671                 rcu_task_stall_info_mult = rtsimc;
672         }
673 #endif /* #ifdef CONFIG_TASKS_RCU */
674 #ifdef CONFIG_TASKS_RCU
675         pr_info("\tTrampoline variant of Tasks RCU enabled.\n");
676 #endif /* #ifdef CONFIG_TASKS_RCU */
677 #ifdef CONFIG_TASKS_RUDE_RCU
678         pr_info("\tRude variant of Tasks RCU enabled.\n");
679 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
680 #ifdef CONFIG_TASKS_TRACE_RCU
681         pr_info("\tTracing variant of Tasks RCU enabled.\n");
682 #endif /* #ifdef CONFIG_TASKS_TRACE_RCU */
683 }
684
685 #endif /* #ifndef CONFIG_TINY_RCU */
686
687 #ifndef CONFIG_TINY_RCU
688 /* Dump out rcutorture-relevant state common to all RCU-tasks flavors. */
689 static void show_rcu_tasks_generic_gp_kthread(struct rcu_tasks *rtp, char *s)
690 {
691         int cpu;
692         bool havecbs = false;
693         bool haveurgent = false;
694         bool haveurgentcbs = false;
695
696         for_each_possible_cpu(cpu) {
697                 struct rcu_tasks_percpu *rtpcp = per_cpu_ptr(rtp->rtpcpu, cpu);
698
699                 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)))
700                         havecbs = true;
701                 if (data_race(rtpcp->urgent_gp))
702                         haveurgent = true;
703                 if (!data_race(rcu_segcblist_empty(&rtpcp->cblist)) && data_race(rtpcp->urgent_gp))
704                         haveurgentcbs = true;
705                 if (havecbs && haveurgent && haveurgentcbs)
706                         break;
707         }
708         pr_info("%s: %s(%d) since %lu g:%lu i:%lu/%lu %c%c%c%c l:%lu %s\n",
709                 rtp->kname,
710                 tasks_gp_state_getname(rtp), data_race(rtp->gp_state),
711                 jiffies - data_race(rtp->gp_jiffies),
712                 data_race(rcu_seq_current(&rtp->tasks_gp_seq)),
713                 data_race(rtp->n_ipis_fails), data_race(rtp->n_ipis),
714                 ".k"[!!data_race(rtp->kthread_ptr)],
715                 ".C"[havecbs],
716                 ".u"[haveurgent],
717                 ".U"[haveurgentcbs],
718                 rtp->lazy_jiffies,
719                 s);
720 }
721 #endif // #ifndef CONFIG_TINY_RCU
722
723 static void exit_tasks_rcu_finish_trace(struct task_struct *t);
724
725 #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU)
726
727 ////////////////////////////////////////////////////////////////////////
728 //
729 // Shared code between task-list-scanning variants of Tasks RCU.
730
731 /* Wait for one RCU-tasks grace period. */
732 static void rcu_tasks_wait_gp(struct rcu_tasks *rtp)
733 {
734         struct task_struct *g;
735         int fract;
736         LIST_HEAD(holdouts);
737         unsigned long j;
738         unsigned long lastinfo;
739         unsigned long lastreport;
740         bool reported = false;
741         int rtsi;
742         struct task_struct *t;
743
744         set_tasks_gp_state(rtp, RTGS_PRE_WAIT_GP);
745         rtp->pregp_func(&holdouts);
746
747         /*
748          * There were callbacks, so we need to wait for an RCU-tasks
749          * grace period.  Start off by scanning the task list for tasks
750          * that are not already voluntarily blocked.  Mark these tasks
751          * and make a list of them in holdouts.
752          */
753         set_tasks_gp_state(rtp, RTGS_SCAN_TASKLIST);
754         if (rtp->pertask_func) {
755                 rcu_read_lock();
756                 for_each_process_thread(g, t)
757                         rtp->pertask_func(t, &holdouts);
758                 rcu_read_unlock();
759         }
760
761         set_tasks_gp_state(rtp, RTGS_POST_SCAN_TASKLIST);
762         rtp->postscan_func(&holdouts);
763
764         /*
765          * Each pass through the following loop scans the list of holdout
766          * tasks, removing any that are no longer holdouts.  When the list
767          * is empty, we are done.
768          */
769         lastreport = jiffies;
770         lastinfo = lastreport;
771         rtsi = READ_ONCE(rcu_task_stall_info);
772
773         // Start off with initial wait and slowly back off to 1 HZ wait.
774         fract = rtp->init_fract;
775
776         while (!list_empty(&holdouts)) {
777                 ktime_t exp;
778                 bool firstreport;
779                 bool needreport;
780                 int rtst;
781
782                 // Slowly back off waiting for holdouts
783                 set_tasks_gp_state(rtp, RTGS_WAIT_SCAN_HOLDOUTS);
784                 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
785                         schedule_timeout_idle(fract);
786                 } else {
787                         exp = jiffies_to_nsecs(fract);
788                         __set_current_state(TASK_IDLE);
789                         schedule_hrtimeout_range(&exp, jiffies_to_nsecs(HZ / 2), HRTIMER_MODE_REL_HARD);
790                 }
791
792                 if (fract < HZ)
793                         fract++;
794
795                 rtst = READ_ONCE(rcu_task_stall_timeout);
796                 needreport = rtst > 0 && time_after(jiffies, lastreport + rtst);
797                 if (needreport) {
798                         lastreport = jiffies;
799                         reported = true;
800                 }
801                 firstreport = true;
802                 WARN_ON(signal_pending(current));
803                 set_tasks_gp_state(rtp, RTGS_SCAN_HOLDOUTS);
804                 rtp->holdouts_func(&holdouts, needreport, &firstreport);
805
806                 // Print pre-stall informational messages if needed.
807                 j = jiffies;
808                 if (rtsi > 0 && !reported && time_after(j, lastinfo + rtsi)) {
809                         lastinfo = j;
810                         rtsi = rtsi * rcu_task_stall_info_mult;
811                         pr_info("%s: %s grace period number %lu (since boot) is %lu jiffies old.\n",
812                                 __func__, rtp->kname, rtp->tasks_gp_seq, j - rtp->gp_start);
813                 }
814         }
815
816         set_tasks_gp_state(rtp, RTGS_POST_GP);
817         rtp->postgp_func(rtp);
818 }
819
820 #endif /* #if defined(CONFIG_TASKS_RCU) || defined(CONFIG_TASKS_TRACE_RCU) */
821
822 #ifdef CONFIG_TASKS_RCU
823
824 ////////////////////////////////////////////////////////////////////////
825 //
826 // Simple variant of RCU whose quiescent states are voluntary context
827 // switch, cond_resched_tasks_rcu_qs(), user-space execution, and idle.
828 // As such, grace periods can take one good long time.  There are no
829 // read-side primitives similar to rcu_read_lock() and rcu_read_unlock()
830 // because this implementation is intended to get the system into a safe
831 // state for some of the manipulations involved in tracing and the like.
832 // Finally, this implementation does not support high call_rcu_tasks()
833 // rates from multiple CPUs.  If this is required, per-CPU callback lists
834 // will be needed.
835 //
836 // The implementation uses rcu_tasks_wait_gp(), which relies on function
837 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_kthread()
838 // function sets these function pointers up so that rcu_tasks_wait_gp()
839 // invokes these functions in this order:
840 //
841 // rcu_tasks_pregp_step():
842 //      Invokes synchronize_rcu() in order to wait for all in-flight
843 //      t->on_rq and t->nvcsw transitions to complete.  This works because
844 //      all such transitions are carried out with interrupts disabled.
845 // rcu_tasks_pertask(), invoked on every non-idle task:
846 //      For every runnable non-idle task other than the current one, use
847 //      get_task_struct() to pin down that task, snapshot that task's
848 //      number of voluntary context switches, and add that task to the
849 //      holdout list.
850 // rcu_tasks_postscan():
851 //      Invoke synchronize_srcu() to ensure that all tasks that were
852 //      in the process of exiting (and which thus might not know to
853 //      synchronize with this RCU Tasks grace period) have completed
854 //      exiting.
855 // check_all_holdout_tasks(), repeatedly until holdout list is empty:
856 //      Scans the holdout list, attempting to identify a quiescent state
857 //      for each task on the list.  If there is a quiescent state, the
858 //      corresponding task is removed from the holdout list.
859 // rcu_tasks_postgp():
860 //      Invokes synchronize_rcu() in order to ensure that all prior
861 //      t->on_rq and t->nvcsw transitions are seen by all CPUs and tasks
862 //      to have happened before the end of this RCU Tasks grace period.
863 //      Again, this works because all such transitions are carried out
864 //      with interrupts disabled.
865 //
866 // For each exiting task, the exit_tasks_rcu_start() and
867 // exit_tasks_rcu_finish() functions begin and end, respectively, the SRCU
868 // read-side critical sections waited for by rcu_tasks_postscan().
869 //
870 // Pre-grace-period update-side code is ordered before the grace
871 // via the raw_spin_lock.*rcu_node().  Pre-grace-period read-side code
872 // is ordered before the grace period via synchronize_rcu() call in
873 // rcu_tasks_pregp_step() and by the scheduler's locks and interrupt
874 // disabling.
875
876 /* Pre-grace-period preparation. */
877 static void rcu_tasks_pregp_step(struct list_head *hop)
878 {
879         /*
880          * Wait for all pre-existing t->on_rq and t->nvcsw transitions
881          * to complete.  Invoking synchronize_rcu() suffices because all
882          * these transitions occur with interrupts disabled.  Without this
883          * synchronize_rcu(), a read-side critical section that started
884          * before the grace period might be incorrectly seen as having
885          * started after the grace period.
886          *
887          * This synchronize_rcu() also dispenses with the need for a
888          * memory barrier on the first store to t->rcu_tasks_holdout,
889          * as it forces the store to happen after the beginning of the
890          * grace period.
891          */
892         synchronize_rcu();
893 }
894
895 /* Per-task initial processing. */
896 static void rcu_tasks_pertask(struct task_struct *t, struct list_head *hop)
897 {
898         if (t != current && READ_ONCE(t->on_rq) && !is_idle_task(t)) {
899                 get_task_struct(t);
900                 t->rcu_tasks_nvcsw = READ_ONCE(t->nvcsw);
901                 WRITE_ONCE(t->rcu_tasks_holdout, true);
902                 list_add(&t->rcu_tasks_holdout_list, hop);
903         }
904 }
905
906 /* Processing between scanning taskslist and draining the holdout list. */
907 static void rcu_tasks_postscan(struct list_head *hop)
908 {
909         int rtsi = READ_ONCE(rcu_task_stall_info);
910
911         if (!IS_ENABLED(CONFIG_TINY_RCU)) {
912                 tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
913                 add_timer(&tasks_rcu_exit_srcu_stall_timer);
914         }
915
916         /*
917          * Exiting tasks may escape the tasklist scan. Those are vulnerable
918          * until their final schedule() with TASK_DEAD state. To cope with
919          * this, divide the fragile exit path part in two intersecting
920          * read side critical sections:
921          *
922          * 1) An _SRCU_ read side starting before calling exit_notify(),
923          *    which may remove the task from the tasklist, and ending after
924          *    the final preempt_disable() call in do_exit().
925          *
926          * 2) An _RCU_ read side starting with the final preempt_disable()
927          *    call in do_exit() and ending with the final call to schedule()
928          *    with TASK_DEAD state.
929          *
930          * This handles the part 1). And postgp will handle part 2) with a
931          * call to synchronize_rcu().
932          */
933         synchronize_srcu(&tasks_rcu_exit_srcu);
934
935         if (!IS_ENABLED(CONFIG_TINY_RCU))
936                 del_timer_sync(&tasks_rcu_exit_srcu_stall_timer);
937 }
938
939 /* See if tasks are still holding out, complain if so. */
940 static void check_holdout_task(struct task_struct *t,
941                                bool needreport, bool *firstreport)
942 {
943         int cpu;
944
945         if (!READ_ONCE(t->rcu_tasks_holdout) ||
946             t->rcu_tasks_nvcsw != READ_ONCE(t->nvcsw) ||
947             !READ_ONCE(t->on_rq) ||
948             (IS_ENABLED(CONFIG_NO_HZ_FULL) &&
949              !is_idle_task(t) && t->rcu_tasks_idle_cpu >= 0)) {
950                 WRITE_ONCE(t->rcu_tasks_holdout, false);
951                 list_del_init(&t->rcu_tasks_holdout_list);
952                 put_task_struct(t);
953                 return;
954         }
955         rcu_request_urgent_qs_task(t);
956         if (!needreport)
957                 return;
958         if (*firstreport) {
959                 pr_err("INFO: rcu_tasks detected stalls on tasks:\n");
960                 *firstreport = false;
961         }
962         cpu = task_cpu(t);
963         pr_alert("%p: %c%c nvcsw: %lu/%lu holdout: %d idle_cpu: %d/%d\n",
964                  t, ".I"[is_idle_task(t)],
965                  "N."[cpu < 0 || !tick_nohz_full_cpu(cpu)],
966                  t->rcu_tasks_nvcsw, t->nvcsw, t->rcu_tasks_holdout,
967                  t->rcu_tasks_idle_cpu, cpu);
968         sched_show_task(t);
969 }
970
971 /* Scan the holdout lists for tasks no longer holding out. */
972 static void check_all_holdout_tasks(struct list_head *hop,
973                                     bool needreport, bool *firstreport)
974 {
975         struct task_struct *t, *t1;
976
977         list_for_each_entry_safe(t, t1, hop, rcu_tasks_holdout_list) {
978                 check_holdout_task(t, needreport, firstreport);
979                 cond_resched();
980         }
981 }
982
983 /* Finish off the Tasks-RCU grace period. */
984 static void rcu_tasks_postgp(struct rcu_tasks *rtp)
985 {
986         /*
987          * Because ->on_rq and ->nvcsw are not guaranteed to have a full
988          * memory barriers prior to them in the schedule() path, memory
989          * reordering on other CPUs could cause their RCU-tasks read-side
990          * critical sections to extend past the end of the grace period.
991          * However, because these ->nvcsw updates are carried out with
992          * interrupts disabled, we can use synchronize_rcu() to force the
993          * needed ordering on all such CPUs.
994          *
995          * This synchronize_rcu() also confines all ->rcu_tasks_holdout
996          * accesses to be within the grace period, avoiding the need for
997          * memory barriers for ->rcu_tasks_holdout accesses.
998          *
999          * In addition, this synchronize_rcu() waits for exiting tasks
1000          * to complete their final preempt_disable() region of execution,
1001          * cleaning up after synchronize_srcu(&tasks_rcu_exit_srcu),
1002          * enforcing the whole region before tasklist removal until
1003          * the final schedule() with TASK_DEAD state to be an RCU TASKS
1004          * read side critical section.
1005          */
1006         synchronize_rcu();
1007 }
1008
1009 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func);
1010 DEFINE_RCU_TASKS(rcu_tasks, rcu_tasks_wait_gp, call_rcu_tasks, "RCU Tasks");
1011
1012 static void tasks_rcu_exit_srcu_stall(struct timer_list *unused)
1013 {
1014 #ifndef CONFIG_TINY_RCU
1015         int rtsi;
1016
1017         rtsi = READ_ONCE(rcu_task_stall_info);
1018         pr_info("%s: %s grace period number %lu (since boot) gp_state: %s is %lu jiffies old.\n",
1019                 __func__, rcu_tasks.kname, rcu_tasks.tasks_gp_seq,
1020                 tasks_gp_state_getname(&rcu_tasks), jiffies - rcu_tasks.gp_jiffies);
1021         pr_info("Please check any exiting tasks stuck between calls to exit_tasks_rcu_start() and exit_tasks_rcu_finish()\n");
1022         tasks_rcu_exit_srcu_stall_timer.expires = jiffies + rtsi;
1023         add_timer(&tasks_rcu_exit_srcu_stall_timer);
1024 #endif // #ifndef CONFIG_TINY_RCU
1025 }
1026
1027 /**
1028  * call_rcu_tasks() - Queue an RCU for invocation task-based grace period
1029  * @rhp: structure to be used for queueing the RCU updates.
1030  * @func: actual callback function to be invoked after the grace period
1031  *
1032  * The callback function will be invoked some time after a full grace
1033  * period elapses, in other words after all currently executing RCU
1034  * read-side critical sections have completed. call_rcu_tasks() assumes
1035  * that the read-side critical sections end at a voluntary context
1036  * switch (not a preemption!), cond_resched_tasks_rcu_qs(), entry into idle,
1037  * or transition to usermode execution.  As such, there are no read-side
1038  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1039  * this primitive is intended to determine that all tasks have passed
1040  * through a safe state, not so much for data-structure synchronization.
1041  *
1042  * See the description of call_rcu() for more detailed information on
1043  * memory ordering guarantees.
1044  */
1045 void call_rcu_tasks(struct rcu_head *rhp, rcu_callback_t func)
1046 {
1047         call_rcu_tasks_generic(rhp, func, &rcu_tasks);
1048 }
1049 EXPORT_SYMBOL_GPL(call_rcu_tasks);
1050
1051 /**
1052  * synchronize_rcu_tasks - wait until an rcu-tasks grace period has elapsed.
1053  *
1054  * Control will return to the caller some time after a full rcu-tasks
1055  * grace period has elapsed, in other words after all currently
1056  * executing rcu-tasks read-side critical sections have elapsed.  These
1057  * read-side critical sections are delimited by calls to schedule(),
1058  * cond_resched_tasks_rcu_qs(), idle execution, userspace execution, calls
1059  * to synchronize_rcu_tasks(), and (in theory, anyway) cond_resched().
1060  *
1061  * This is a very specialized primitive, intended only for a few uses in
1062  * tracing and other situations requiring manipulation of function
1063  * preambles and profiling hooks.  The synchronize_rcu_tasks() function
1064  * is not (yet) intended for heavy use from multiple CPUs.
1065  *
1066  * See the description of synchronize_rcu() for more detailed information
1067  * on memory ordering guarantees.
1068  */
1069 void synchronize_rcu_tasks(void)
1070 {
1071         synchronize_rcu_tasks_generic(&rcu_tasks);
1072 }
1073 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks);
1074
1075 /**
1076  * rcu_barrier_tasks - Wait for in-flight call_rcu_tasks() callbacks.
1077  *
1078  * Although the current implementation is guaranteed to wait, it is not
1079  * obligated to, for example, if there are no pending callbacks.
1080  */
1081 void rcu_barrier_tasks(void)
1082 {
1083         rcu_barrier_tasks_generic(&rcu_tasks);
1084 }
1085 EXPORT_SYMBOL_GPL(rcu_barrier_tasks);
1086
1087 int rcu_tasks_lazy_ms = -1;
1088 module_param(rcu_tasks_lazy_ms, int, 0444);
1089
1090 static int __init rcu_spawn_tasks_kthread(void)
1091 {
1092         cblist_init_generic(&rcu_tasks);
1093         rcu_tasks.gp_sleep = HZ / 10;
1094         rcu_tasks.init_fract = HZ / 10;
1095         if (rcu_tasks_lazy_ms >= 0)
1096                 rcu_tasks.lazy_jiffies = msecs_to_jiffies(rcu_tasks_lazy_ms);
1097         rcu_tasks.pregp_func = rcu_tasks_pregp_step;
1098         rcu_tasks.pertask_func = rcu_tasks_pertask;
1099         rcu_tasks.postscan_func = rcu_tasks_postscan;
1100         rcu_tasks.holdouts_func = check_all_holdout_tasks;
1101         rcu_tasks.postgp_func = rcu_tasks_postgp;
1102         rcu_spawn_tasks_kthread_generic(&rcu_tasks);
1103         return 0;
1104 }
1105
1106 #if !defined(CONFIG_TINY_RCU)
1107 void show_rcu_tasks_classic_gp_kthread(void)
1108 {
1109         show_rcu_tasks_generic_gp_kthread(&rcu_tasks, "");
1110 }
1111 EXPORT_SYMBOL_GPL(show_rcu_tasks_classic_gp_kthread);
1112 #endif // !defined(CONFIG_TINY_RCU)
1113
1114 struct task_struct *get_rcu_tasks_gp_kthread(void)
1115 {
1116         return rcu_tasks.kthread_ptr;
1117 }
1118 EXPORT_SYMBOL_GPL(get_rcu_tasks_gp_kthread);
1119
1120 /*
1121  * Contribute to protect against tasklist scan blind spot while the
1122  * task is exiting and may be removed from the tasklist. See
1123  * corresponding synchronize_srcu() for further details.
1124  */
1125 void exit_tasks_rcu_start(void) __acquires(&tasks_rcu_exit_srcu)
1126 {
1127         current->rcu_tasks_idx = __srcu_read_lock(&tasks_rcu_exit_srcu);
1128 }
1129
1130 /*
1131  * Contribute to protect against tasklist scan blind spot while the
1132  * task is exiting and may be removed from the tasklist. See
1133  * corresponding synchronize_srcu() for further details.
1134  */
1135 void exit_tasks_rcu_stop(void) __releases(&tasks_rcu_exit_srcu)
1136 {
1137         struct task_struct *t = current;
1138
1139         __srcu_read_unlock(&tasks_rcu_exit_srcu, t->rcu_tasks_idx);
1140 }
1141
1142 /*
1143  * Contribute to protect against tasklist scan blind spot while the
1144  * task is exiting and may be removed from the tasklist. See
1145  * corresponding synchronize_srcu() for further details.
1146  */
1147 void exit_tasks_rcu_finish(void)
1148 {
1149         exit_tasks_rcu_stop();
1150         exit_tasks_rcu_finish_trace(current);
1151 }
1152
1153 #else /* #ifdef CONFIG_TASKS_RCU */
1154 void exit_tasks_rcu_start(void) { }
1155 void exit_tasks_rcu_stop(void) { }
1156 void exit_tasks_rcu_finish(void) { exit_tasks_rcu_finish_trace(current); }
1157 #endif /* #else #ifdef CONFIG_TASKS_RCU */
1158
1159 #ifdef CONFIG_TASKS_RUDE_RCU
1160
1161 ////////////////////////////////////////////////////////////////////////
1162 //
1163 // "Rude" variant of Tasks RCU, inspired by Steve Rostedt's trick of
1164 // passing an empty function to schedule_on_each_cpu().  This approach
1165 // provides an asynchronous call_rcu_tasks_rude() API and batching of
1166 // concurrent calls to the synchronous synchronize_rcu_tasks_rude() API.
1167 // This invokes schedule_on_each_cpu() in order to send IPIs far and wide
1168 // and induces otherwise unnecessary context switches on all online CPUs,
1169 // whether idle or not.
1170 //
1171 // Callback handling is provided by the rcu_tasks_kthread() function.
1172 //
1173 // Ordering is provided by the scheduler's context-switch code.
1174
1175 // Empty function to allow workqueues to force a context switch.
1176 static void rcu_tasks_be_rude(struct work_struct *work)
1177 {
1178 }
1179
1180 // Wait for one rude RCU-tasks grace period.
1181 static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
1182 {
1183         rtp->n_ipis += cpumask_weight(cpu_online_mask);
1184         schedule_on_each_cpu(rcu_tasks_be_rude);
1185 }
1186
1187 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
1188 DEFINE_RCU_TASKS(rcu_tasks_rude, rcu_tasks_rude_wait_gp, call_rcu_tasks_rude,
1189                  "RCU Tasks Rude");
1190
1191 /**
1192  * call_rcu_tasks_rude() - Queue a callback rude task-based grace period
1193  * @rhp: structure to be used for queueing the RCU updates.
1194  * @func: actual callback function to be invoked after the grace period
1195  *
1196  * The callback function will be invoked some time after a full grace
1197  * period elapses, in other words after all currently executing RCU
1198  * read-side critical sections have completed. call_rcu_tasks_rude()
1199  * assumes that the read-side critical sections end at context switch,
1200  * cond_resched_tasks_rcu_qs(), or transition to usermode execution (as
1201  * usermode execution is schedulable). As such, there are no read-side
1202  * primitives analogous to rcu_read_lock() and rcu_read_unlock() because
1203  * this primitive is intended to determine that all tasks have passed
1204  * through a safe state, not so much for data-structure synchronization.
1205  *
1206  * See the description of call_rcu() for more detailed information on
1207  * memory ordering guarantees.
1208  */
1209 void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func)
1210 {
1211         call_rcu_tasks_generic(rhp, func, &rcu_tasks_rude);
1212 }
1213 EXPORT_SYMBOL_GPL(call_rcu_tasks_rude);
1214
1215 /**
1216  * synchronize_rcu_tasks_rude - wait for a rude rcu-tasks grace period
1217  *
1218  * Control will return to the caller some time after a rude rcu-tasks
1219  * grace period has elapsed, in other words after all currently
1220  * executing rcu-tasks read-side critical sections have elapsed.  These
1221  * read-side critical sections are delimited by calls to schedule(),
1222  * cond_resched_tasks_rcu_qs(), userspace execution (which is a schedulable
1223  * context), and (in theory, anyway) cond_resched().
1224  *
1225  * This is a very specialized primitive, intended only for a few uses in
1226  * tracing and other situations requiring manipulation of function preambles
1227  * and profiling hooks.  The synchronize_rcu_tasks_rude() function is not
1228  * (yet) intended for heavy use from multiple CPUs.
1229  *
1230  * See the description of synchronize_rcu() for more detailed information
1231  * on memory ordering guarantees.
1232  */
1233 void synchronize_rcu_tasks_rude(void)
1234 {
1235         synchronize_rcu_tasks_generic(&rcu_tasks_rude);
1236 }
1237 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_rude);
1238
1239 /**
1240  * rcu_barrier_tasks_rude - Wait for in-flight call_rcu_tasks_rude() callbacks.
1241  *
1242  * Although the current implementation is guaranteed to wait, it is not
1243  * obligated to, for example, if there are no pending callbacks.
1244  */
1245 void rcu_barrier_tasks_rude(void)
1246 {
1247         rcu_barrier_tasks_generic(&rcu_tasks_rude);
1248 }
1249 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_rude);
1250
1251 int rcu_tasks_rude_lazy_ms = -1;
1252 module_param(rcu_tasks_rude_lazy_ms, int, 0444);
1253
1254 static int __init rcu_spawn_tasks_rude_kthread(void)
1255 {
1256         cblist_init_generic(&rcu_tasks_rude);
1257         rcu_tasks_rude.gp_sleep = HZ / 10;
1258         if (rcu_tasks_rude_lazy_ms >= 0)
1259                 rcu_tasks_rude.lazy_jiffies = msecs_to_jiffies(rcu_tasks_rude_lazy_ms);
1260         rcu_spawn_tasks_kthread_generic(&rcu_tasks_rude);
1261         return 0;
1262 }
1263
1264 #if !defined(CONFIG_TINY_RCU)
1265 void show_rcu_tasks_rude_gp_kthread(void)
1266 {
1267         show_rcu_tasks_generic_gp_kthread(&rcu_tasks_rude, "");
1268 }
1269 EXPORT_SYMBOL_GPL(show_rcu_tasks_rude_gp_kthread);
1270 #endif // !defined(CONFIG_TINY_RCU)
1271
1272 struct task_struct *get_rcu_tasks_rude_gp_kthread(void)
1273 {
1274         return rcu_tasks_rude.kthread_ptr;
1275 }
1276 EXPORT_SYMBOL_GPL(get_rcu_tasks_rude_gp_kthread);
1277
1278 #endif /* #ifdef CONFIG_TASKS_RUDE_RCU */
1279
1280 ////////////////////////////////////////////////////////////////////////
1281 //
1282 // Tracing variant of Tasks RCU.  This variant is designed to be used
1283 // to protect tracing hooks, including those of BPF.  This variant
1284 // therefore:
1285 //
1286 // 1.   Has explicit read-side markers to allow finite grace periods
1287 //      in the face of in-kernel loops for PREEMPT=n builds.
1288 //
1289 // 2.   Protects code in the idle loop, exception entry/exit, and
1290 //      CPU-hotplug code paths, similar to the capabilities of SRCU.
1291 //
1292 // 3.   Avoids expensive read-side instructions, having overhead similar
1293 //      to that of Preemptible RCU.
1294 //
1295 // There are of course downsides.  For example, the grace-period code
1296 // can send IPIs to CPUs, even when those CPUs are in the idle loop or
1297 // in nohz_full userspace.  If needed, these downsides can be at least
1298 // partially remedied.
1299 //
1300 // Perhaps most important, this variant of RCU does not affect the vanilla
1301 // flavors, rcu_preempt and rcu_sched.  The fact that RCU Tasks Trace
1302 // readers can operate from idle, offline, and exception entry/exit in no
1303 // way allows rcu_preempt and rcu_sched readers to also do so.
1304 //
1305 // The implementation uses rcu_tasks_wait_gp(), which relies on function
1306 // pointers in the rcu_tasks structure.  The rcu_spawn_tasks_trace_kthread()
1307 // function sets these function pointers up so that rcu_tasks_wait_gp()
1308 // invokes these functions in this order:
1309 //
1310 // rcu_tasks_trace_pregp_step():
1311 //      Disables CPU hotplug, adds all currently executing tasks to the
1312 //      holdout list, then checks the state of all tasks that blocked
1313 //      or were preempted within their current RCU Tasks Trace read-side
1314 //      critical section, adding them to the holdout list if appropriate.
1315 //      Finally, this function re-enables CPU hotplug.
1316 // The ->pertask_func() pointer is NULL, so there is no per-task processing.
1317 // rcu_tasks_trace_postscan():
1318 //      Invokes synchronize_rcu() to wait for late-stage exiting tasks
1319 //      to finish exiting.
1320 // check_all_holdout_tasks_trace(), repeatedly until holdout list is empty:
1321 //      Scans the holdout list, attempting to identify a quiescent state
1322 //      for each task on the list.  If there is a quiescent state, the
1323 //      corresponding task is removed from the holdout list.  Once this
1324 //      list is empty, the grace period has completed.
1325 // rcu_tasks_trace_postgp():
1326 //      Provides the needed full memory barrier and does debug checks.
1327 //
1328 // The exit_tasks_rcu_finish_trace() synchronizes with exiting tasks.
1329 //
1330 // Pre-grace-period update-side code is ordered before the grace period
1331 // via the ->cbs_lock and barriers in rcu_tasks_kthread().  Pre-grace-period
1332 // read-side code is ordered before the grace period by atomic operations
1333 // on .b.need_qs flag of each task involved in this process, or by scheduler
1334 // context-switch ordering (for locked-down non-running readers).
1335
1336 // The lockdep state must be outside of #ifdef to be useful.
1337 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1338 static struct lock_class_key rcu_lock_trace_key;
1339 struct lockdep_map rcu_trace_lock_map =
1340         STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_trace", &rcu_lock_trace_key);
1341 EXPORT_SYMBOL_GPL(rcu_trace_lock_map);
1342 #endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
1343
1344 #ifdef CONFIG_TASKS_TRACE_RCU
1345
1346 // Record outstanding IPIs to each CPU.  No point in sending two...
1347 static DEFINE_PER_CPU(bool, trc_ipi_to_cpu);
1348
1349 // The number of detections of task quiescent state relying on
1350 // heavyweight readers executing explicit memory barriers.
1351 static unsigned long n_heavy_reader_attempts;
1352 static unsigned long n_heavy_reader_updates;
1353 static unsigned long n_heavy_reader_ofl_updates;
1354 static unsigned long n_trc_holdouts;
1355
1356 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func);
1357 DEFINE_RCU_TASKS(rcu_tasks_trace, rcu_tasks_wait_gp, call_rcu_tasks_trace,
1358                  "RCU Tasks Trace");
1359
1360 /* Load from ->trc_reader_special.b.need_qs with proper ordering. */
1361 static u8 rcu_ld_need_qs(struct task_struct *t)
1362 {
1363         smp_mb(); // Enforce full grace-period ordering.
1364         return smp_load_acquire(&t->trc_reader_special.b.need_qs);
1365 }
1366
1367 /* Store to ->trc_reader_special.b.need_qs with proper ordering. */
1368 static void rcu_st_need_qs(struct task_struct *t, u8 v)
1369 {
1370         smp_store_release(&t->trc_reader_special.b.need_qs, v);
1371         smp_mb(); // Enforce full grace-period ordering.
1372 }
1373
1374 /*
1375  * Do a cmpxchg() on ->trc_reader_special.b.need_qs, allowing for
1376  * the four-byte operand-size restriction of some platforms.
1377  * Returns the old value, which is often ignored.
1378  */
1379 u8 rcu_trc_cmpxchg_need_qs(struct task_struct *t, u8 old, u8 new)
1380 {
1381         union rcu_special ret;
1382         union rcu_special trs_old = READ_ONCE(t->trc_reader_special);
1383         union rcu_special trs_new = trs_old;
1384
1385         if (trs_old.b.need_qs != old)
1386                 return trs_old.b.need_qs;
1387         trs_new.b.need_qs = new;
1388         ret.s = cmpxchg(&t->trc_reader_special.s, trs_old.s, trs_new.s);
1389         return ret.b.need_qs;
1390 }
1391 EXPORT_SYMBOL_GPL(rcu_trc_cmpxchg_need_qs);
1392
1393 /*
1394  * If we are the last reader, signal the grace-period kthread.
1395  * Also remove from the per-CPU list of blocked tasks.
1396  */
1397 void rcu_read_unlock_trace_special(struct task_struct *t)
1398 {
1399         unsigned long flags;
1400         struct rcu_tasks_percpu *rtpcp;
1401         union rcu_special trs;
1402
1403         // Open-coded full-word version of rcu_ld_need_qs().
1404         smp_mb(); // Enforce full grace-period ordering.
1405         trs = smp_load_acquire(&t->trc_reader_special);
1406
1407         if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && t->trc_reader_special.b.need_mb)
1408                 smp_mb(); // Pairs with update-side barriers.
1409         // Update .need_qs before ->trc_reader_nesting for irq/NMI handlers.
1410         if (trs.b.need_qs == (TRC_NEED_QS_CHECKED | TRC_NEED_QS)) {
1411                 u8 result = rcu_trc_cmpxchg_need_qs(t, TRC_NEED_QS_CHECKED | TRC_NEED_QS,
1412                                                        TRC_NEED_QS_CHECKED);
1413
1414                 WARN_ONCE(result != trs.b.need_qs, "%s: result = %d", __func__, result);
1415         }
1416         if (trs.b.blocked) {
1417                 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, t->trc_blkd_cpu);
1418                 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1419                 list_del_init(&t->trc_blkd_node);
1420                 WRITE_ONCE(t->trc_reader_special.b.blocked, false);
1421                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1422         }
1423         WRITE_ONCE(t->trc_reader_nesting, 0);
1424 }
1425 EXPORT_SYMBOL_GPL(rcu_read_unlock_trace_special);
1426
1427 /* Add a newly blocked reader task to its CPU's list. */
1428 void rcu_tasks_trace_qs_blkd(struct task_struct *t)
1429 {
1430         unsigned long flags;
1431         struct rcu_tasks_percpu *rtpcp;
1432
1433         local_irq_save(flags);
1434         rtpcp = this_cpu_ptr(rcu_tasks_trace.rtpcpu);
1435         raw_spin_lock_rcu_node(rtpcp); // irqs already disabled
1436         t->trc_blkd_cpu = smp_processor_id();
1437         if (!rtpcp->rtp_blkd_tasks.next)
1438                 INIT_LIST_HEAD(&rtpcp->rtp_blkd_tasks);
1439         list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1440         WRITE_ONCE(t->trc_reader_special.b.blocked, true);
1441         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1442 }
1443 EXPORT_SYMBOL_GPL(rcu_tasks_trace_qs_blkd);
1444
1445 /* Add a task to the holdout list, if it is not already on the list. */
1446 static void trc_add_holdout(struct task_struct *t, struct list_head *bhp)
1447 {
1448         if (list_empty(&t->trc_holdout_list)) {
1449                 get_task_struct(t);
1450                 list_add(&t->trc_holdout_list, bhp);
1451                 n_trc_holdouts++;
1452         }
1453 }
1454
1455 /* Remove a task from the holdout list, if it is in fact present. */
1456 static void trc_del_holdout(struct task_struct *t)
1457 {
1458         if (!list_empty(&t->trc_holdout_list)) {
1459                 list_del_init(&t->trc_holdout_list);
1460                 put_task_struct(t);
1461                 n_trc_holdouts--;
1462         }
1463 }
1464
1465 /* IPI handler to check task state. */
1466 static void trc_read_check_handler(void *t_in)
1467 {
1468         int nesting;
1469         struct task_struct *t = current;
1470         struct task_struct *texp = t_in;
1471
1472         // If the task is no longer running on this CPU, leave.
1473         if (unlikely(texp != t))
1474                 goto reset_ipi; // Already on holdout list, so will check later.
1475
1476         // If the task is not in a read-side critical section, and
1477         // if this is the last reader, awaken the grace-period kthread.
1478         nesting = READ_ONCE(t->trc_reader_nesting);
1479         if (likely(!nesting)) {
1480                 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1481                 goto reset_ipi;
1482         }
1483         // If we are racing with an rcu_read_unlock_trace(), try again later.
1484         if (unlikely(nesting < 0))
1485                 goto reset_ipi;
1486
1487         // Get here if the task is in a read-side critical section.
1488         // Set its state so that it will update state for the grace-period
1489         // kthread upon exit from that critical section.
1490         rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED);
1491
1492 reset_ipi:
1493         // Allow future IPIs to be sent on CPU and for task.
1494         // Also order this IPI handler against any later manipulations of
1495         // the intended task.
1496         smp_store_release(per_cpu_ptr(&trc_ipi_to_cpu, smp_processor_id()), false); // ^^^
1497         smp_store_release(&texp->trc_ipi_to_cpu, -1); // ^^^
1498 }
1499
1500 /* Callback function for scheduler to check locked-down task.  */
1501 static int trc_inspect_reader(struct task_struct *t, void *bhp_in)
1502 {
1503         struct list_head *bhp = bhp_in;
1504         int cpu = task_cpu(t);
1505         int nesting;
1506         bool ofl = cpu_is_offline(cpu);
1507
1508         if (task_curr(t) && !ofl) {
1509                 // If no chance of heavyweight readers, do it the hard way.
1510                 if (!IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
1511                         return -EINVAL;
1512
1513                 // If heavyweight readers are enabled on the remote task,
1514                 // we can inspect its state despite its currently running.
1515                 // However, we cannot safely change its state.
1516                 n_heavy_reader_attempts++;
1517                 // Check for "running" idle tasks on offline CPUs.
1518                 if (!rcu_dynticks_zero_in_eqs(cpu, &t->trc_reader_nesting))
1519                         return -EINVAL; // No quiescent state, do it the hard way.
1520                 n_heavy_reader_updates++;
1521                 nesting = 0;
1522         } else {
1523                 // The task is not running, so C-language access is safe.
1524                 nesting = t->trc_reader_nesting;
1525                 WARN_ON_ONCE(ofl && task_curr(t) && !is_idle_task(t));
1526                 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB) && ofl)
1527                         n_heavy_reader_ofl_updates++;
1528         }
1529
1530         // If not exiting a read-side critical section, mark as checked
1531         // so that the grace-period kthread will remove it from the
1532         // holdout list.
1533         if (!nesting) {
1534                 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1535                 return 0;  // In QS, so done.
1536         }
1537         if (nesting < 0)
1538                 return -EINVAL; // Reader transitioning, try again later.
1539
1540         // The task is in a read-side critical section, so set up its
1541         // state so that it will update state upon exit from that critical
1542         // section.
1543         if (!rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS | TRC_NEED_QS_CHECKED))
1544                 trc_add_holdout(t, bhp);
1545         return 0;
1546 }
1547
1548 /* Attempt to extract the state for the specified task. */
1549 static void trc_wait_for_one_reader(struct task_struct *t,
1550                                     struct list_head *bhp)
1551 {
1552         int cpu;
1553
1554         // If a previous IPI is still in flight, let it complete.
1555         if (smp_load_acquire(&t->trc_ipi_to_cpu) != -1) // Order IPI
1556                 return;
1557
1558         // The current task had better be in a quiescent state.
1559         if (t == current) {
1560                 rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1561                 WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1562                 return;
1563         }
1564
1565         // Attempt to nail down the task for inspection.
1566         get_task_struct(t);
1567         if (!task_call_func(t, trc_inspect_reader, bhp)) {
1568                 put_task_struct(t);
1569                 return;
1570         }
1571         put_task_struct(t);
1572
1573         // If this task is not yet on the holdout list, then we are in
1574         // an RCU read-side critical section.  Otherwise, the invocation of
1575         // trc_add_holdout() that added it to the list did the necessary
1576         // get_task_struct().  Either way, the task cannot be freed out
1577         // from under this code.
1578
1579         // If currently running, send an IPI, either way, add to list.
1580         trc_add_holdout(t, bhp);
1581         if (task_curr(t) &&
1582             time_after(jiffies + 1, rcu_tasks_trace.gp_start + rcu_task_ipi_delay)) {
1583                 // The task is currently running, so try IPIing it.
1584                 cpu = task_cpu(t);
1585
1586                 // If there is already an IPI outstanding, let it happen.
1587                 if (per_cpu(trc_ipi_to_cpu, cpu) || t->trc_ipi_to_cpu >= 0)
1588                         return;
1589
1590                 per_cpu(trc_ipi_to_cpu, cpu) = true;
1591                 t->trc_ipi_to_cpu = cpu;
1592                 rcu_tasks_trace.n_ipis++;
1593                 if (smp_call_function_single(cpu, trc_read_check_handler, t, 0)) {
1594                         // Just in case there is some other reason for
1595                         // failure than the target CPU being offline.
1596                         WARN_ONCE(1, "%s():  smp_call_function_single() failed for CPU: %d\n",
1597                                   __func__, cpu);
1598                         rcu_tasks_trace.n_ipis_fails++;
1599                         per_cpu(trc_ipi_to_cpu, cpu) = false;
1600                         t->trc_ipi_to_cpu = -1;
1601                 }
1602         }
1603 }
1604
1605 /*
1606  * Initialize for first-round processing for the specified task.
1607  * Return false if task is NULL or already taken care of, true otherwise.
1608  */
1609 static bool rcu_tasks_trace_pertask_prep(struct task_struct *t, bool notself)
1610 {
1611         // During early boot when there is only the one boot CPU, there
1612         // is no idle task for the other CPUs.  Also, the grace-period
1613         // kthread is always in a quiescent state.  In addition, just return
1614         // if this task is already on the list.
1615         if (unlikely(t == NULL) || (t == current && notself) || !list_empty(&t->trc_holdout_list))
1616                 return false;
1617
1618         rcu_st_need_qs(t, 0);
1619         t->trc_ipi_to_cpu = -1;
1620         return true;
1621 }
1622
1623 /* Do first-round processing for the specified task. */
1624 static void rcu_tasks_trace_pertask(struct task_struct *t, struct list_head *hop)
1625 {
1626         if (rcu_tasks_trace_pertask_prep(t, true))
1627                 trc_wait_for_one_reader(t, hop);
1628 }
1629
1630 /* Initialize for a new RCU-tasks-trace grace period. */
1631 static void rcu_tasks_trace_pregp_step(struct list_head *hop)
1632 {
1633         LIST_HEAD(blkd_tasks);
1634         int cpu;
1635         unsigned long flags;
1636         struct rcu_tasks_percpu *rtpcp;
1637         struct task_struct *t;
1638
1639         // There shouldn't be any old IPIs, but...
1640         for_each_possible_cpu(cpu)
1641                 WARN_ON_ONCE(per_cpu(trc_ipi_to_cpu, cpu));
1642
1643         // Disable CPU hotplug across the CPU scan for the benefit of
1644         // any IPIs that might be needed.  This also waits for all readers
1645         // in CPU-hotplug code paths.
1646         cpus_read_lock();
1647
1648         // These rcu_tasks_trace_pertask_prep() calls are serialized to
1649         // allow safe access to the hop list.
1650         for_each_online_cpu(cpu) {
1651                 rcu_read_lock();
1652                 t = cpu_curr_snapshot(cpu);
1653                 if (rcu_tasks_trace_pertask_prep(t, true))
1654                         trc_add_holdout(t, hop);
1655                 rcu_read_unlock();
1656                 cond_resched_tasks_rcu_qs();
1657         }
1658
1659         // Only after all running tasks have been accounted for is it
1660         // safe to take care of the tasks that have blocked within their
1661         // current RCU tasks trace read-side critical section.
1662         for_each_possible_cpu(cpu) {
1663                 rtpcp = per_cpu_ptr(rcu_tasks_trace.rtpcpu, cpu);
1664                 raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1665                 list_splice_init(&rtpcp->rtp_blkd_tasks, &blkd_tasks);
1666                 while (!list_empty(&blkd_tasks)) {
1667                         rcu_read_lock();
1668                         t = list_first_entry(&blkd_tasks, struct task_struct, trc_blkd_node);
1669                         list_del_init(&t->trc_blkd_node);
1670                         list_add(&t->trc_blkd_node, &rtpcp->rtp_blkd_tasks);
1671                         raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1672                         rcu_tasks_trace_pertask(t, hop);
1673                         rcu_read_unlock();
1674                         raw_spin_lock_irqsave_rcu_node(rtpcp, flags);
1675                 }
1676                 raw_spin_unlock_irqrestore_rcu_node(rtpcp, flags);
1677                 cond_resched_tasks_rcu_qs();
1678         }
1679
1680         // Re-enable CPU hotplug now that the holdout list is populated.
1681         cpus_read_unlock();
1682 }
1683
1684 /*
1685  * Do intermediate processing between task and holdout scans.
1686  */
1687 static void rcu_tasks_trace_postscan(struct list_head *hop)
1688 {
1689         // Wait for late-stage exiting tasks to finish exiting.
1690         // These might have passed the call to exit_tasks_rcu_finish().
1691
1692         // If you remove the following line, update rcu_trace_implies_rcu_gp()!!!
1693         synchronize_rcu();
1694         // Any tasks that exit after this point will set
1695         // TRC_NEED_QS_CHECKED in ->trc_reader_special.b.need_qs.
1696 }
1697
1698 /* Communicate task state back to the RCU tasks trace stall warning request. */
1699 struct trc_stall_chk_rdr {
1700         int nesting;
1701         int ipi_to_cpu;
1702         u8 needqs;
1703 };
1704
1705 static int trc_check_slow_task(struct task_struct *t, void *arg)
1706 {
1707         struct trc_stall_chk_rdr *trc_rdrp = arg;
1708
1709         if (task_curr(t) && cpu_online(task_cpu(t)))
1710                 return false; // It is running, so decline to inspect it.
1711         trc_rdrp->nesting = READ_ONCE(t->trc_reader_nesting);
1712         trc_rdrp->ipi_to_cpu = READ_ONCE(t->trc_ipi_to_cpu);
1713         trc_rdrp->needqs = rcu_ld_need_qs(t);
1714         return true;
1715 }
1716
1717 /* Show the state of a task stalling the current RCU tasks trace GP. */
1718 static void show_stalled_task_trace(struct task_struct *t, bool *firstreport)
1719 {
1720         int cpu;
1721         struct trc_stall_chk_rdr trc_rdr;
1722         bool is_idle_tsk = is_idle_task(t);
1723
1724         if (*firstreport) {
1725                 pr_err("INFO: rcu_tasks_trace detected stalls on tasks:\n");
1726                 *firstreport = false;
1727         }
1728         cpu = task_cpu(t);
1729         if (!task_call_func(t, trc_check_slow_task, &trc_rdr))
1730                 pr_alert("P%d: %c%c\n",
1731                          t->pid,
1732                          ".I"[t->trc_ipi_to_cpu >= 0],
1733                          ".i"[is_idle_tsk]);
1734         else
1735                 pr_alert("P%d: %c%c%c%c nesting: %d%c%c cpu: %d%s\n",
1736                          t->pid,
1737                          ".I"[trc_rdr.ipi_to_cpu >= 0],
1738                          ".i"[is_idle_tsk],
1739                          ".N"[cpu >= 0 && tick_nohz_full_cpu(cpu)],
1740                          ".B"[!!data_race(t->trc_reader_special.b.blocked)],
1741                          trc_rdr.nesting,
1742                          " !CN"[trc_rdr.needqs & 0x3],
1743                          " ?"[trc_rdr.needqs > 0x3],
1744                          cpu, cpu_online(cpu) ? "" : "(offline)");
1745         sched_show_task(t);
1746 }
1747
1748 /* List stalled IPIs for RCU tasks trace. */
1749 static void show_stalled_ipi_trace(void)
1750 {
1751         int cpu;
1752
1753         for_each_possible_cpu(cpu)
1754                 if (per_cpu(trc_ipi_to_cpu, cpu))
1755                         pr_alert("\tIPI outstanding to CPU %d\n", cpu);
1756 }
1757
1758 /* Do one scan of the holdout list. */
1759 static void check_all_holdout_tasks_trace(struct list_head *hop,
1760                                           bool needreport, bool *firstreport)
1761 {
1762         struct task_struct *g, *t;
1763
1764         // Disable CPU hotplug across the holdout list scan for IPIs.
1765         cpus_read_lock();
1766
1767         list_for_each_entry_safe(t, g, hop, trc_holdout_list) {
1768                 // If safe and needed, try to check the current task.
1769                 if (READ_ONCE(t->trc_ipi_to_cpu) == -1 &&
1770                     !(rcu_ld_need_qs(t) & TRC_NEED_QS_CHECKED))
1771                         trc_wait_for_one_reader(t, hop);
1772
1773                 // If check succeeded, remove this task from the list.
1774                 if (smp_load_acquire(&t->trc_ipi_to_cpu) == -1 &&
1775                     rcu_ld_need_qs(t) == TRC_NEED_QS_CHECKED)
1776                         trc_del_holdout(t);
1777                 else if (needreport)
1778                         show_stalled_task_trace(t, firstreport);
1779                 cond_resched_tasks_rcu_qs();
1780         }
1781
1782         // Re-enable CPU hotplug now that the holdout list scan has completed.
1783         cpus_read_unlock();
1784
1785         if (needreport) {
1786                 if (*firstreport)
1787                         pr_err("INFO: rcu_tasks_trace detected stalls? (Late IPI?)\n");
1788                 show_stalled_ipi_trace();
1789         }
1790 }
1791
1792 static void rcu_tasks_trace_empty_fn(void *unused)
1793 {
1794 }
1795
1796 /* Wait for grace period to complete and provide ordering. */
1797 static void rcu_tasks_trace_postgp(struct rcu_tasks *rtp)
1798 {
1799         int cpu;
1800
1801         // Wait for any lingering IPI handlers to complete.  Note that
1802         // if a CPU has gone offline or transitioned to userspace in the
1803         // meantime, all IPI handlers should have been drained beforehand.
1804         // Yes, this assumes that CPUs process IPIs in order.  If that ever
1805         // changes, there will need to be a recheck and/or timed wait.
1806         for_each_online_cpu(cpu)
1807                 if (WARN_ON_ONCE(smp_load_acquire(per_cpu_ptr(&trc_ipi_to_cpu, cpu))))
1808                         smp_call_function_single(cpu, rcu_tasks_trace_empty_fn, NULL, 1);
1809
1810         smp_mb(); // Caller's code must be ordered after wakeup.
1811                   // Pairs with pretty much every ordering primitive.
1812 }
1813
1814 /* Report any needed quiescent state for this exiting task. */
1815 static void exit_tasks_rcu_finish_trace(struct task_struct *t)
1816 {
1817         union rcu_special trs = READ_ONCE(t->trc_reader_special);
1818
1819         rcu_trc_cmpxchg_need_qs(t, 0, TRC_NEED_QS_CHECKED);
1820         WARN_ON_ONCE(READ_ONCE(t->trc_reader_nesting));
1821         if (WARN_ON_ONCE(rcu_ld_need_qs(t) & TRC_NEED_QS || trs.b.blocked))
1822                 rcu_read_unlock_trace_special(t);
1823         else
1824                 WRITE_ONCE(t->trc_reader_nesting, 0);
1825 }
1826
1827 /**
1828  * call_rcu_tasks_trace() - Queue a callback trace task-based grace period
1829  * @rhp: structure to be used for queueing the RCU updates.
1830  * @func: actual callback function to be invoked after the grace period
1831  *
1832  * The callback function will be invoked some time after a trace rcu-tasks
1833  * grace period elapses, in other words after all currently executing
1834  * trace rcu-tasks read-side critical sections have completed. These
1835  * read-side critical sections are delimited by calls to rcu_read_lock_trace()
1836  * and rcu_read_unlock_trace().
1837  *
1838  * See the description of call_rcu() for more detailed information on
1839  * memory ordering guarantees.
1840  */
1841 void call_rcu_tasks_trace(struct rcu_head *rhp, rcu_callback_t func)
1842 {
1843         call_rcu_tasks_generic(rhp, func, &rcu_tasks_trace);
1844 }
1845 EXPORT_SYMBOL_GPL(call_rcu_tasks_trace);
1846
1847 /**
1848  * synchronize_rcu_tasks_trace - wait for a trace rcu-tasks grace period
1849  *
1850  * Control will return to the caller some time after a trace rcu-tasks
1851  * grace period has elapsed, in other words after all currently executing
1852  * trace rcu-tasks read-side critical sections have elapsed. These read-side
1853  * critical sections are delimited by calls to rcu_read_lock_trace()
1854  * and rcu_read_unlock_trace().
1855  *
1856  * This is a very specialized primitive, intended only for a few uses in
1857  * tracing and other situations requiring manipulation of function preambles
1858  * and profiling hooks.  The synchronize_rcu_tasks_trace() function is not
1859  * (yet) intended for heavy use from multiple CPUs.
1860  *
1861  * See the description of synchronize_rcu() for more detailed information
1862  * on memory ordering guarantees.
1863  */
1864 void synchronize_rcu_tasks_trace(void)
1865 {
1866         RCU_LOCKDEP_WARN(lock_is_held(&rcu_trace_lock_map), "Illegal synchronize_rcu_tasks_trace() in RCU Tasks Trace read-side critical section");
1867         synchronize_rcu_tasks_generic(&rcu_tasks_trace);
1868 }
1869 EXPORT_SYMBOL_GPL(synchronize_rcu_tasks_trace);
1870
1871 /**
1872  * rcu_barrier_tasks_trace - Wait for in-flight call_rcu_tasks_trace() callbacks.
1873  *
1874  * Although the current implementation is guaranteed to wait, it is not
1875  * obligated to, for example, if there are no pending callbacks.
1876  */
1877 void rcu_barrier_tasks_trace(void)
1878 {
1879         rcu_barrier_tasks_generic(&rcu_tasks_trace);
1880 }
1881 EXPORT_SYMBOL_GPL(rcu_barrier_tasks_trace);
1882
1883 int rcu_tasks_trace_lazy_ms = -1;
1884 module_param(rcu_tasks_trace_lazy_ms, int, 0444);
1885
1886 static int __init rcu_spawn_tasks_trace_kthread(void)
1887 {
1888         cblist_init_generic(&rcu_tasks_trace);
1889         if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB)) {
1890                 rcu_tasks_trace.gp_sleep = HZ / 10;
1891                 rcu_tasks_trace.init_fract = HZ / 10;
1892         } else {
1893                 rcu_tasks_trace.gp_sleep = HZ / 200;
1894                 if (rcu_tasks_trace.gp_sleep <= 0)
1895                         rcu_tasks_trace.gp_sleep = 1;
1896                 rcu_tasks_trace.init_fract = HZ / 200;
1897                 if (rcu_tasks_trace.init_fract <= 0)
1898                         rcu_tasks_trace.init_fract = 1;
1899         }
1900         if (rcu_tasks_trace_lazy_ms >= 0)
1901                 rcu_tasks_trace.lazy_jiffies = msecs_to_jiffies(rcu_tasks_trace_lazy_ms);
1902         rcu_tasks_trace.pregp_func = rcu_tasks_trace_pregp_step;
1903         rcu_tasks_trace.postscan_func = rcu_tasks_trace_postscan;
1904         rcu_tasks_trace.holdouts_func = check_all_holdout_tasks_trace;
1905         rcu_tasks_trace.postgp_func = rcu_tasks_trace_postgp;
1906         rcu_spawn_tasks_kthread_generic(&rcu_tasks_trace);
1907         return 0;
1908 }
1909
1910 #if !defined(CONFIG_TINY_RCU)
1911 void show_rcu_tasks_trace_gp_kthread(void)
1912 {
1913         char buf[64];
1914
1915         sprintf(buf, "N%lu h:%lu/%lu/%lu",
1916                 data_race(n_trc_holdouts),
1917                 data_race(n_heavy_reader_ofl_updates),
1918                 data_race(n_heavy_reader_updates),
1919                 data_race(n_heavy_reader_attempts));
1920         show_rcu_tasks_generic_gp_kthread(&rcu_tasks_trace, buf);
1921 }
1922 EXPORT_SYMBOL_GPL(show_rcu_tasks_trace_gp_kthread);
1923 #endif // !defined(CONFIG_TINY_RCU)
1924
1925 struct task_struct *get_rcu_tasks_trace_gp_kthread(void)
1926 {
1927         return rcu_tasks_trace.kthread_ptr;
1928 }
1929 EXPORT_SYMBOL_GPL(get_rcu_tasks_trace_gp_kthread);
1930
1931 #else /* #ifdef CONFIG_TASKS_TRACE_RCU */
1932 static void exit_tasks_rcu_finish_trace(struct task_struct *t) { }
1933 #endif /* #else #ifdef CONFIG_TASKS_TRACE_RCU */
1934
1935 #ifndef CONFIG_TINY_RCU
1936 void show_rcu_tasks_gp_kthreads(void)
1937 {
1938         show_rcu_tasks_classic_gp_kthread();
1939         show_rcu_tasks_rude_gp_kthread();
1940         show_rcu_tasks_trace_gp_kthread();
1941 }
1942 #endif /* #ifndef CONFIG_TINY_RCU */
1943
1944 #ifdef CONFIG_PROVE_RCU
1945 struct rcu_tasks_test_desc {
1946         struct rcu_head rh;
1947         const char *name;
1948         bool notrun;
1949         unsigned long runstart;
1950 };
1951
1952 static struct rcu_tasks_test_desc tests[] = {
1953         {
1954                 .name = "call_rcu_tasks()",
1955                 /* If not defined, the test is skipped. */
1956                 .notrun = IS_ENABLED(CONFIG_TASKS_RCU),
1957         },
1958         {
1959                 .name = "call_rcu_tasks_rude()",
1960                 /* If not defined, the test is skipped. */
1961                 .notrun = IS_ENABLED(CONFIG_TASKS_RUDE_RCU),
1962         },
1963         {
1964                 .name = "call_rcu_tasks_trace()",
1965                 /* If not defined, the test is skipped. */
1966                 .notrun = IS_ENABLED(CONFIG_TASKS_TRACE_RCU)
1967         }
1968 };
1969
1970 static void test_rcu_tasks_callback(struct rcu_head *rhp)
1971 {
1972         struct rcu_tasks_test_desc *rttd =
1973                 container_of(rhp, struct rcu_tasks_test_desc, rh);
1974
1975         pr_info("Callback from %s invoked.\n", rttd->name);
1976
1977         rttd->notrun = false;
1978 }
1979
1980 static void rcu_tasks_initiate_self_tests(void)
1981 {
1982         pr_info("Running RCU-tasks wait API self tests\n");
1983 #ifdef CONFIG_TASKS_RCU
1984         tests[0].runstart = jiffies;
1985         synchronize_rcu_tasks();
1986         call_rcu_tasks(&tests[0].rh, test_rcu_tasks_callback);
1987 #endif
1988
1989 #ifdef CONFIG_TASKS_RUDE_RCU
1990         tests[1].runstart = jiffies;
1991         synchronize_rcu_tasks_rude();
1992         call_rcu_tasks_rude(&tests[1].rh, test_rcu_tasks_callback);
1993 #endif
1994
1995 #ifdef CONFIG_TASKS_TRACE_RCU
1996         tests[2].runstart = jiffies;
1997         synchronize_rcu_tasks_trace();
1998         call_rcu_tasks_trace(&tests[2].rh, test_rcu_tasks_callback);
1999 #endif
2000 }
2001
2002 /*
2003  * Return:  0 - test passed
2004  *          1 - test failed, but have not timed out yet
2005  *         -1 - test failed and timed out
2006  */
2007 static int rcu_tasks_verify_self_tests(void)
2008 {
2009         int ret = 0;
2010         int i;
2011         unsigned long bst = rcu_task_stall_timeout;
2012
2013         if (bst <= 0 || bst > RCU_TASK_BOOT_STALL_TIMEOUT)
2014                 bst = RCU_TASK_BOOT_STALL_TIMEOUT;
2015         for (i = 0; i < ARRAY_SIZE(tests); i++) {
2016                 while (tests[i].notrun) {               // still hanging.
2017                         if (time_after(jiffies, tests[i].runstart + bst)) {
2018                                 pr_err("%s has failed boot-time tests.\n", tests[i].name);
2019                                 ret = -1;
2020                                 break;
2021                         }
2022                         ret = 1;
2023                         break;
2024                 }
2025         }
2026         WARN_ON(ret < 0);
2027
2028         return ret;
2029 }
2030
2031 /*
2032  * Repeat the rcu_tasks_verify_self_tests() call once every second until the
2033  * test passes or has timed out.
2034  */
2035 static struct delayed_work rcu_tasks_verify_work;
2036 static void rcu_tasks_verify_work_fn(struct work_struct *work __maybe_unused)
2037 {
2038         int ret = rcu_tasks_verify_self_tests();
2039
2040         if (ret <= 0)
2041                 return;
2042
2043         /* Test fails but not timed out yet, reschedule another check */
2044         schedule_delayed_work(&rcu_tasks_verify_work, HZ);
2045 }
2046
2047 static int rcu_tasks_verify_schedule_work(void)
2048 {
2049         INIT_DELAYED_WORK(&rcu_tasks_verify_work, rcu_tasks_verify_work_fn);
2050         rcu_tasks_verify_work_fn(NULL);
2051         return 0;
2052 }
2053 late_initcall(rcu_tasks_verify_schedule_work);
2054 #else /* #ifdef CONFIG_PROVE_RCU */
2055 static void rcu_tasks_initiate_self_tests(void) { }
2056 #endif /* #else #ifdef CONFIG_PROVE_RCU */
2057
2058 void __init rcu_init_tasks_generic(void)
2059 {
2060 #ifdef CONFIG_TASKS_RCU
2061         rcu_spawn_tasks_kthread();
2062 #endif
2063
2064 #ifdef CONFIG_TASKS_RUDE_RCU
2065         rcu_spawn_tasks_rude_kthread();
2066 #endif
2067
2068 #ifdef CONFIG_TASKS_TRACE_RCU
2069         rcu_spawn_tasks_trace_kthread();
2070 #endif
2071
2072         // Run the self-tests.
2073         rcu_tasks_initiate_self_tests();
2074 }
2075
2076 #else /* #ifdef CONFIG_TASKS_RCU_GENERIC */
2077 static inline void rcu_tasks_bootup_oddness(void) {}
2078 #endif /* #else #ifdef CONFIG_TASKS_RCU_GENERIC */